Merge pull request #15664 from akien-mga/thirdparty

Bugfix updates to various thirdparty libraries

122 files changed, 57775 insertions, 0 deletions

diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuBroadphaseInterface.h b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuBroadphaseInterface.h
new file mode 100644
index 0000000000..0ed8aa8232
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuBroadphaseInterface.h
@@ -0,0 +1,44 @@
+
+#ifndef B3_GPU_BROADPHASE_INTERFACE_H
+#define B3_GPU_BROADPHASE_INTERFACE_H
+
+#include "Bullet3OpenCL/Initialize/b3OpenCLInclude.h"
+#include "Bullet3Common/b3Vector3.h"
+#include "b3SapAabb.h"
+#include "Bullet3Common/shared/b3Int2.h"
+#include "Bullet3Common/shared/b3Int4.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
+
+class b3GpuBroadphaseInterface
+{
+public:
+
+	typedef class b3GpuBroadphaseInterface* (CreateFunc)(cl_context ctx,cl_device_id device, cl_command_queue  q);
+
+	virtual ~b3GpuBroadphaseInterface()
+	{
+	}
+
+	virtual void createProxy(const b3Vector3& aabbMin,  const b3Vector3& aabbMax, int userPtr , int collisionFilterGroup, int collisionFilterMask)=0;
+	virtual void createLargeProxy(const b3Vector3& aabbMin,  const b3Vector3& aabbMax, int userPtr , int collisionFilterGroup, int collisionFilterMask)=0;
+
+	virtual void  calculateOverlappingPairs(int maxPairs)=0;
+	virtual void  calculateOverlappingPairsHost(int maxPairs)=0;
+
+	//call writeAabbsToGpu after done making all changes (createProxy etc)
+	virtual void writeAabbsToGpu()=0;
+
+	virtual cl_mem	getAabbBufferWS()=0;
+	virtual int	getNumOverlap()=0;
+	virtual cl_mem	getOverlappingPairBuffer()=0;
+
+	virtual b3OpenCLArray<b3SapAabb>&	getAllAabbsGPU()=0;
+	virtual b3AlignedObjectArray<b3SapAabb>&	getAllAabbsCPU()=0;
+	
+	virtual b3OpenCLArray<b3Int4>& getOverlappingPairsGPU() = 0;
+	virtual b3OpenCLArray<int>& getSmallAabbIndicesGPU() = 0;
+	virtual b3OpenCLArray<int>& getLargeAabbIndicesGPU() = 0;
+
+};
+
+#endif //B3_GPU_BROADPHASE_INTERFACE_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuGridBroadphase.cpp b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuGridBroadphase.cpp
new file mode 100644
index 0000000000..74d0c8056c
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuGridBroadphase.cpp
@@ -0,0 +1,385 @@
+
+#include "b3GpuGridBroadphase.h"
+#include "Bullet3Geometry/b3AabbUtil.h"
+#include "kernels/gridBroadphaseKernels.h"
+#include "kernels/sapKernels.h"
+//#include "kernels/gridBroadphase.cl"
+
+
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h"
+
+
+
+#define B3_BROADPHASE_SAP_PATH "src/Bullet3OpenCL/BroadphaseCollision/kernels/sap.cl"
+#define B3_GRID_BROADPHASE_PATH "src/Bullet3OpenCL/BroadphaseCollision/kernels/gridBroadphase.cl"
+
+cl_kernel kCalcHashAABB;
+cl_kernel kClearCellStart;
+cl_kernel kFindCellStart;
+cl_kernel kFindOverlappingPairs;
+cl_kernel m_copyAabbsKernel;
+cl_kernel m_sap2Kernel;
+
+
+
+
+
+//int maxPairsPerBody = 64;
+int maxBodiesPerCell = 256;//??
+
+b3GpuGridBroadphase::b3GpuGridBroadphase(cl_context ctx,cl_device_id device, cl_command_queue  q )
+:m_context(ctx),
+m_device(device),
+m_queue(q),
+m_allAabbsGPU1(ctx,q),
+m_smallAabbsMappingGPU(ctx,q),
+m_largeAabbsMappingGPU(ctx,q),
+m_gpuPairs(ctx,q),
+
+m_hashGpu(ctx,q),
+
+m_cellStartGpu(ctx,q),
+m_paramsGPU(ctx,q)
+{
+
+	
+	b3Vector3 gridSize = b3MakeVector3(3,3,3);
+	b3Vector3 invGridSize = b3MakeVector3(1.f/gridSize[0],1.f/gridSize[1],1.f/gridSize[2]);
+
+	m_paramsCPU.m_gridSize[0] = 128;
+	m_paramsCPU.m_gridSize[1] = 128;
+	m_paramsCPU.m_gridSize[2] = 128;
+	m_paramsCPU.m_gridSize[3] = maxBodiesPerCell;
+	m_paramsCPU.setMaxBodiesPerCell(maxBodiesPerCell);
+	m_paramsCPU.m_invCellSize[0] = invGridSize[0];
+	m_paramsCPU.m_invCellSize[1] = invGridSize[1];
+	m_paramsCPU.m_invCellSize[2] = invGridSize[2];
+	m_paramsCPU.m_invCellSize[3] = 0.f;
+	m_paramsGPU.push_back(m_paramsCPU);
+
+	cl_int errNum=0;
+
+	{
+		const char* sapSrc = sapCL;
+		cl_program sapProg = b3OpenCLUtils::compileCLProgramFromString(m_context,m_device,sapSrc,&errNum,"",B3_BROADPHASE_SAP_PATH);
+		b3Assert(errNum==CL_SUCCESS);
+		m_copyAabbsKernel= b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,sapSrc, "copyAabbsKernel",&errNum,sapProg );
+		m_sap2Kernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,sapSrc, "computePairsKernelTwoArrays",&errNum,sapProg );
+		b3Assert(errNum==CL_SUCCESS);
+	}
+
+	{
+		
+		cl_program gridProg = b3OpenCLUtils::compileCLProgramFromString(m_context,m_device,gridBroadphaseCL,&errNum,"",B3_GRID_BROADPHASE_PATH);
+		b3Assert(errNum==CL_SUCCESS);
+
+		kCalcHashAABB = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,gridBroadphaseCL, "kCalcHashAABB",&errNum,gridProg);
+		b3Assert(errNum==CL_SUCCESS);
+	
+		kClearCellStart = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,gridBroadphaseCL, "kClearCellStart",&errNum,gridProg);
+		b3Assert(errNum==CL_SUCCESS);
+
+		kFindCellStart = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,gridBroadphaseCL, "kFindCellStart",&errNum,gridProg);
+		b3Assert(errNum==CL_SUCCESS);
+
+	
+		kFindOverlappingPairs = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,gridBroadphaseCL, "kFindOverlappingPairs",&errNum,gridProg);
+		b3Assert(errNum==CL_SUCCESS);
+
+		
+		
+		
+	}
+
+	m_sorter = new b3RadixSort32CL(m_context,m_device,m_queue);
+
+}
+b3GpuGridBroadphase::~b3GpuGridBroadphase()
+{
+	clReleaseKernel( kCalcHashAABB);
+	clReleaseKernel( kClearCellStart);
+	clReleaseKernel( kFindCellStart);
+	clReleaseKernel( kFindOverlappingPairs);
+	clReleaseKernel( m_sap2Kernel);
+	clReleaseKernel( m_copyAabbsKernel);
+	
+	
+	
+	delete m_sorter;
+}
+
+
+
+void b3GpuGridBroadphase::createProxy(const b3Vector3& aabbMin,  const b3Vector3& aabbMax, int userPtr , int collisionFilterGroup, int collisionFilterMask)
+{
+	b3SapAabb aabb;
+	aabb.m_minVec = aabbMin;
+	aabb.m_maxVec = aabbMax;
+	aabb.m_minIndices[3] = userPtr;
+	aabb.m_signedMaxIndices[3] = m_allAabbsCPU1.size();//NOT userPtr;
+	m_smallAabbsMappingCPU.push_back(m_allAabbsCPU1.size());
+
+	m_allAabbsCPU1.push_back(aabb);
+
+}
+void b3GpuGridBroadphase::createLargeProxy(const b3Vector3& aabbMin,  const b3Vector3& aabbMax, int userPtr , int collisionFilterGroup, int collisionFilterMask)
+{
+	b3SapAabb aabb;
+	aabb.m_minVec = aabbMin;
+	aabb.m_maxVec = aabbMax;
+	aabb.m_minIndices[3] = userPtr;
+	aabb.m_signedMaxIndices[3] = m_allAabbsCPU1.size();//NOT userPtr;
+	m_largeAabbsMappingCPU.push_back(m_allAabbsCPU1.size());
+
+	m_allAabbsCPU1.push_back(aabb);
+}
+
+void  b3GpuGridBroadphase::calculateOverlappingPairs(int maxPairs)
+{
+	B3_PROFILE("b3GpuGridBroadphase::calculateOverlappingPairs");
+	
+
+	if (0)
+	{
+		calculateOverlappingPairsHost(maxPairs);
+	/*
+		b3AlignedObjectArray<b3Int4> cpuPairs;
+		m_gpuPairs.copyToHost(cpuPairs);
+		printf("host m_gpuPairs.size()=%d\n",m_gpuPairs.size());
+		for (int i=0;i<m_gpuPairs.size();i++)
+		{
+			printf("host pair %d = %d,%d\n",i,cpuPairs[i].x,cpuPairs[i].y);
+		}
+		*/
+		return;
+	}
+	
+	
+
+
+	
+	int numSmallAabbs = m_smallAabbsMappingGPU.size();
+
+	b3OpenCLArray<int> pairCount(m_context,m_queue);
+	pairCount.push_back(0);
+	m_gpuPairs.resize(maxPairs);//numSmallAabbs*maxPairsPerBody);
+
+	{
+		int numLargeAabbs = m_largeAabbsMappingGPU.size();
+		if (numLargeAabbs && numSmallAabbs)
+		{
+			B3_PROFILE("sap2Kernel");
+			b3BufferInfoCL bInfo[] = { 
+				b3BufferInfoCL( m_allAabbsGPU1.getBufferCL() ),
+				b3BufferInfoCL( m_largeAabbsMappingGPU.getBufferCL() ),
+				b3BufferInfoCL( m_smallAabbsMappingGPU.getBufferCL() ), 
+				b3BufferInfoCL( m_gpuPairs.getBufferCL() ), 
+				b3BufferInfoCL(pairCount.getBufferCL())};
+			b3LauncherCL launcher(m_queue, m_sap2Kernel,"m_sap2Kernel");
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(   numLargeAabbs  );
+			launcher.setConst( numSmallAabbs);
+			launcher.setConst( 0  );//axis is not used
+			launcher.setConst( maxPairs  );
+	//@todo: use actual maximum work item sizes of the device instead of hardcoded values
+			launcher.launch2D( numLargeAabbs, numSmallAabbs,4,64);
+                
+			int numPairs = pairCount.at(0);
+			
+			if (numPairs >maxPairs)
+			{
+				b3Error("Error running out of pairs: numPairs = %d, maxPairs = %d.\n", numPairs, maxPairs);
+				numPairs =maxPairs;
+			}
+		}
+	}
+
+
+
+
+	if (numSmallAabbs)
+	{
+		B3_PROFILE("gridKernel");
+		m_hashGpu.resize(numSmallAabbs);
+		{
+			B3_PROFILE("kCalcHashAABB");
+			b3LauncherCL launch(m_queue,kCalcHashAABB,"kCalcHashAABB");
+			launch.setConst(numSmallAabbs);
+			launch.setBuffer(m_allAabbsGPU1.getBufferCL());
+			launch.setBuffer(m_smallAabbsMappingGPU.getBufferCL());
+			launch.setBuffer(m_hashGpu.getBufferCL());
+			launch.setBuffer(this->m_paramsGPU.getBufferCL());
+			launch.launch1D(numSmallAabbs);
+		}
+
+		m_sorter->execute(m_hashGpu);
+		
+		int numCells = this->m_paramsCPU.m_gridSize[0]*this->m_paramsCPU.m_gridSize[1]*this->m_paramsCPU.m_gridSize[2];
+		m_cellStartGpu.resize(numCells);
+		//b3AlignedObjectArray<int >			cellStartCpu;
+				
+		
+		{
+			B3_PROFILE("kClearCellStart");
+			b3LauncherCL launch(m_queue,kClearCellStart,"kClearCellStart");
+			launch.setConst(numCells);
+			launch.setBuffer(m_cellStartGpu.getBufferCL());
+			launch.launch1D(numCells);
+			//m_cellStartGpu.copyToHost(cellStartCpu);
+			//printf("??\n");
+
+		}
+
+
+		{
+			B3_PROFILE("kFindCellStart");
+			b3LauncherCL launch(m_queue,kFindCellStart,"kFindCellStart");
+			launch.setConst(numSmallAabbs);
+			launch.setBuffer(m_hashGpu.getBufferCL());
+			launch.setBuffer(m_cellStartGpu.getBufferCL());
+			launch.launch1D(numSmallAabbs);
+			//m_cellStartGpu.copyToHost(cellStartCpu);
+			//printf("??\n");
+
+		}
+		
+		{
+			B3_PROFILE("kFindOverlappingPairs");
+			
+			
+			b3LauncherCL launch(m_queue,kFindOverlappingPairs,"kFindOverlappingPairs");
+			launch.setConst(numSmallAabbs);
+			launch.setBuffer(m_allAabbsGPU1.getBufferCL());
+			launch.setBuffer(m_smallAabbsMappingGPU.getBufferCL());
+			launch.setBuffer(m_hashGpu.getBufferCL());
+			launch.setBuffer(m_cellStartGpu.getBufferCL());
+			
+			launch.setBuffer(m_paramsGPU.getBufferCL());
+			//launch.setBuffer(0);
+			launch.setBuffer(pairCount.getBufferCL());
+			launch.setBuffer(m_gpuPairs.getBufferCL());
+			
+			launch.setConst(maxPairs);
+			launch.launch1D(numSmallAabbs);
+			
+
+			int numPairs = pairCount.at(0);
+			if (numPairs >maxPairs)
+			{
+				b3Error("Error running out of pairs: numPairs = %d, maxPairs = %d.\n", numPairs, maxPairs);
+				numPairs =maxPairs;
+			}
+			
+			m_gpuPairs.resize(numPairs);
+	
+			if (0)
+			{
+				b3AlignedObjectArray<b3Int4> pairsCpu;
+				m_gpuPairs.copyToHost(pairsCpu);
+
+				int sz = m_gpuPairs.size();
+				printf("m_gpuPairs.size()=%d\n",sz);
+				for (int i=0;i<m_gpuPairs.size();i++)
+				{
+					printf("pair %d = %d,%d\n",i,pairsCpu[i].x,pairsCpu[i].y);
+				}
+
+				printf("?!?\n");
+			}
+			
+		}
+	
+
+	}
+
+	
+
+
+
+	//calculateOverlappingPairsHost(maxPairs);
+}
+void  b3GpuGridBroadphase::calculateOverlappingPairsHost(int maxPairs)
+{
+
+	m_hostPairs.resize(0);
+	m_allAabbsGPU1.copyToHost(m_allAabbsCPU1);
+	for (int i=0;i<m_allAabbsCPU1.size();i++)
+	{
+		for (int j=i+1;j<m_allAabbsCPU1.size();j++)
+		{
+			if (b3TestAabbAgainstAabb2(m_allAabbsCPU1[i].m_minVec, m_allAabbsCPU1[i].m_maxVec,
+				m_allAabbsCPU1[j].m_minVec,m_allAabbsCPU1[j].m_maxVec))
+			{
+				b3Int4 pair;
+				int a = m_allAabbsCPU1[j].m_minIndices[3];
+				int b = m_allAabbsCPU1[i].m_minIndices[3];
+				if (a<=b)
+				{
+					pair.x = a; 
+					pair.y = b;//store the original index in the unsorted aabb array
+				} else
+				{
+					pair.x = b;
+					pair.y = a;//store the original index in the unsorted aabb array
+				}
+					
+				if (m_hostPairs.size()<maxPairs)
+				{
+					m_hostPairs.push_back(pair);
+				}
+			}
+		}
+	}
+
+
+	m_gpuPairs.copyFromHost(m_hostPairs);
+
+
+}
+
+	//call writeAabbsToGpu after done making all changes (createProxy etc)
+void b3GpuGridBroadphase::writeAabbsToGpu()
+{
+	m_allAabbsGPU1.copyFromHost(m_allAabbsCPU1);
+	m_smallAabbsMappingGPU.copyFromHost(m_smallAabbsMappingCPU);
+	m_largeAabbsMappingGPU.copyFromHost(m_largeAabbsMappingCPU);
+
+}
+
+cl_mem	b3GpuGridBroadphase::getAabbBufferWS()
+{
+	return this->m_allAabbsGPU1.getBufferCL();
+}
+int	b3GpuGridBroadphase::getNumOverlap()
+{
+	return m_gpuPairs.size();
+}
+cl_mem	b3GpuGridBroadphase::getOverlappingPairBuffer()
+{
+	return m_gpuPairs.getBufferCL();
+}
+
+b3OpenCLArray<b3SapAabb>&	b3GpuGridBroadphase::getAllAabbsGPU()
+{
+	return m_allAabbsGPU1;
+}
+
+b3AlignedObjectArray<b3SapAabb>&	b3GpuGridBroadphase::getAllAabbsCPU()
+{
+	return m_allAabbsCPU1;
+}
+
+b3OpenCLArray<b3Int4>& b3GpuGridBroadphase::getOverlappingPairsGPU()
+{
+	return m_gpuPairs;
+}
+b3OpenCLArray<int>& b3GpuGridBroadphase::getSmallAabbIndicesGPU()
+{
+	return m_smallAabbsMappingGPU;
+}
+b3OpenCLArray<int>& b3GpuGridBroadphase::getLargeAabbIndicesGPU()
+{
+	return m_largeAabbsMappingGPU;
+}
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuGridBroadphase.h b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuGridBroadphase.h
new file mode 100644
index 0000000000..ec18c9f716
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuGridBroadphase.h
@@ -0,0 +1,88 @@
+#ifndef B3_GPU_GRID_BROADPHASE_H
+#define B3_GPU_GRID_BROADPHASE_H
+
+#include "b3GpuBroadphaseInterface.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h"
+
+struct b3ParamsGridBroadphaseCL
+{
+
+	float m_invCellSize[4];
+	int   m_gridSize[4];
+
+	int	getMaxBodiesPerCell() const
+	{
+		return m_gridSize[3];
+	}
+
+	void setMaxBodiesPerCell(int maxOverlap) 
+	{
+		m_gridSize[3] = maxOverlap;
+	}
+};
+
+
+class b3GpuGridBroadphase : public b3GpuBroadphaseInterface
+{
+protected:
+	cl_context				m_context;
+	cl_device_id			m_device;
+	cl_command_queue		m_queue;
+
+	b3OpenCLArray<b3SapAabb>	m_allAabbsGPU1;
+	b3AlignedObjectArray<b3SapAabb>	m_allAabbsCPU1;
+
+	b3OpenCLArray<int>	m_smallAabbsMappingGPU;
+	b3AlignedObjectArray<int> m_smallAabbsMappingCPU;
+
+	b3OpenCLArray<int>	m_largeAabbsMappingGPU;
+	b3AlignedObjectArray<int> m_largeAabbsMappingCPU;
+
+	b3AlignedObjectArray<b3Int4> m_hostPairs;
+	b3OpenCLArray<b3Int4>			m_gpuPairs;
+
+	b3OpenCLArray<b3SortData>			m_hashGpu;
+	b3OpenCLArray<int>			m_cellStartGpu;
+	
+
+	b3ParamsGridBroadphaseCL		m_paramsCPU;
+	b3OpenCLArray<b3ParamsGridBroadphaseCL>		m_paramsGPU;
+
+	class b3RadixSort32CL*			m_sorter;
+
+public:
+
+	b3GpuGridBroadphase(cl_context ctx,cl_device_id device, cl_command_queue  q );
+	virtual ~b3GpuGridBroadphase();
+
+	static b3GpuBroadphaseInterface* CreateFunc(cl_context ctx,cl_device_id device, cl_command_queue  q)
+	{
+		return new b3GpuGridBroadphase(ctx,device,q);
+	}
+
+	
+
+
+	virtual void createProxy(const b3Vector3& aabbMin,  const b3Vector3& aabbMax, int userPtr , int collisionFilterGroup, int collisionFilterMask);
+	virtual void createLargeProxy(const b3Vector3& aabbMin,  const b3Vector3& aabbMax, int userPtr , int collisionFilterGroup, int collisionFilterMask);
+
+	virtual void  calculateOverlappingPairs(int maxPairs);
+	virtual void  calculateOverlappingPairsHost(int maxPairs);
+
+	//call writeAabbsToGpu after done making all changes (createProxy etc)
+	virtual void writeAabbsToGpu();
+
+	virtual cl_mem	getAabbBufferWS();
+	virtual int	getNumOverlap();
+	virtual cl_mem	getOverlappingPairBuffer();
+
+	virtual b3OpenCLArray<b3SapAabb>&	getAllAabbsGPU();
+	virtual b3AlignedObjectArray<b3SapAabb>&	getAllAabbsCPU();
+	
+	virtual b3OpenCLArray<b3Int4>& getOverlappingPairsGPU();
+	virtual b3OpenCLArray<int>& getSmallAabbIndicesGPU();
+	virtual b3OpenCLArray<int>& getLargeAabbIndicesGPU();
+
+};
+
+#endif //B3_GPU_GRID_BROADPHASE_H
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvh.cpp b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvh.cpp
new file mode 100644
index 0000000000..641df9eb12
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvh.cpp
@@ -0,0 +1,577 @@
+/*
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose,
+including commercial applications, and to alter it and redistribute it freely,
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Initial Author Jackson Lee, 2014
+
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h"
+
+#include "b3GpuParallelLinearBvh.h"
+
+b3GpuParallelLinearBvh::b3GpuParallelLinearBvh(cl_context context, cl_device_id device, cl_command_queue queue) :
+	m_queue(queue),
+	m_radixSorter(context, device, queue),
+	
+	m_rootNodeIndex(context, queue),
+	m_maxDistanceFromRoot(context, queue),
+	m_temp(context, queue),
+	
+	m_internalNodeAabbs(context, queue),
+	m_internalNodeLeafIndexRanges(context, queue),
+	m_internalNodeChildNodes(context, queue),
+	m_internalNodeParentNodes(context, queue),
+	
+	m_commonPrefixes(context, queue),
+	m_commonPrefixLengths(context, queue),
+	m_distanceFromRoot(context, queue),
+	
+	m_leafNodeParentNodes(context, queue),
+	m_mortonCodesAndAabbIndicies(context, queue),
+	m_mergedAabb(context, queue),
+	m_leafNodeAabbs(context, queue),
+	
+	m_largeAabbs(context, queue)
+{
+	m_rootNodeIndex.resize(1);
+	m_maxDistanceFromRoot.resize(1);
+	m_temp.resize(1);
+	
+	//
+	const char CL_PROGRAM_PATH[] = "src/Bullet3OpenCL/BroadphaseCollision/kernels/parallelLinearBvh.cl";
+	
+	const char* kernelSource = parallelLinearBvhCL;	//parallelLinearBvhCL.h
+	cl_int error;
+	char* additionalMacros = 0;
+	m_parallelLinearBvhProgram = b3OpenCLUtils::compileCLProgramFromString(context, device, kernelSource, &error, additionalMacros, CL_PROGRAM_PATH);
+	b3Assert(m_parallelLinearBvhProgram);
+	
+	m_separateAabbsKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "separateAabbs", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_separateAabbsKernel);
+	m_findAllNodesMergedAabbKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "findAllNodesMergedAabb", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_findAllNodesMergedAabbKernel);
+	m_assignMortonCodesAndAabbIndiciesKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "assignMortonCodesAndAabbIndicies", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_assignMortonCodesAndAabbIndiciesKernel);
+	
+	m_computeAdjacentPairCommonPrefixKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "computeAdjacentPairCommonPrefix", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_computeAdjacentPairCommonPrefixKernel);
+	m_buildBinaryRadixTreeLeafNodesKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "buildBinaryRadixTreeLeafNodes", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_buildBinaryRadixTreeLeafNodesKernel);
+	m_buildBinaryRadixTreeInternalNodesKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "buildBinaryRadixTreeInternalNodes", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_buildBinaryRadixTreeInternalNodesKernel);
+	m_findDistanceFromRootKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "findDistanceFromRoot", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_findDistanceFromRootKernel);
+	m_buildBinaryRadixTreeAabbsRecursiveKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "buildBinaryRadixTreeAabbsRecursive", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_buildBinaryRadixTreeAabbsRecursiveKernel);
+	
+	m_findLeafIndexRangesKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "findLeafIndexRanges", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_findLeafIndexRangesKernel);
+	
+	m_plbvhCalculateOverlappingPairsKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "plbvhCalculateOverlappingPairs", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_plbvhCalculateOverlappingPairsKernel);
+	m_plbvhRayTraverseKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "plbvhRayTraverse", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_plbvhRayTraverseKernel);
+	m_plbvhLargeAabbAabbTestKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "plbvhLargeAabbAabbTest", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_plbvhLargeAabbAabbTestKernel);
+	m_plbvhLargeAabbRayTestKernel = b3OpenCLUtils::compileCLKernelFromString( context, device, kernelSource, "plbvhLargeAabbRayTest", &error, m_parallelLinearBvhProgram, additionalMacros );
+	b3Assert(m_plbvhLargeAabbRayTestKernel);
+}
+
+b3GpuParallelLinearBvh::~b3GpuParallelLinearBvh() 
+{
+	clReleaseKernel(m_separateAabbsKernel);
+	clReleaseKernel(m_findAllNodesMergedAabbKernel);
+	clReleaseKernel(m_assignMortonCodesAndAabbIndiciesKernel);
+	
+	clReleaseKernel(m_computeAdjacentPairCommonPrefixKernel);
+	clReleaseKernel(m_buildBinaryRadixTreeLeafNodesKernel);
+	clReleaseKernel(m_buildBinaryRadixTreeInternalNodesKernel);
+	clReleaseKernel(m_findDistanceFromRootKernel);
+	clReleaseKernel(m_buildBinaryRadixTreeAabbsRecursiveKernel);
+	
+	clReleaseKernel(m_findLeafIndexRangesKernel);
+	
+	clReleaseKernel(m_plbvhCalculateOverlappingPairsKernel);
+	clReleaseKernel(m_plbvhRayTraverseKernel);
+	clReleaseKernel(m_plbvhLargeAabbAabbTestKernel);
+	clReleaseKernel(m_plbvhLargeAabbRayTestKernel);
+	
+	clReleaseProgram(m_parallelLinearBvhProgram);
+}
+
+void b3GpuParallelLinearBvh::build(const b3OpenCLArray<b3SapAabb>& worldSpaceAabbs, const b3OpenCLArray<int>& smallAabbIndices, 
+									const b3OpenCLArray<int>& largeAabbIndices)
+{
+	B3_PROFILE("b3ParallelLinearBvh::build()");
+	
+	int numLargeAabbs = largeAabbIndices.size();
+	int numSmallAabbs = smallAabbIndices.size();
+	
+	//Since all AABBs(both large and small) are input as a contiguous array, 
+	//with 2 additional arrays used to indicate the indices of large and small AABBs,
+	//it is necessary to separate the AABBs so that the large AABBs will not degrade the quality of the BVH.
+	{
+		B3_PROFILE("Separate large and small AABBs");
+		
+		m_largeAabbs.resize(numLargeAabbs);
+		m_leafNodeAabbs.resize(numSmallAabbs);
+		
+		//Write large AABBs into m_largeAabbs
+		{
+			b3BufferInfoCL bufferInfo[] = 
+			{
+				b3BufferInfoCL( worldSpaceAabbs.getBufferCL() ),
+				b3BufferInfoCL( largeAabbIndices.getBufferCL() ),
+				
+				b3BufferInfoCL( m_largeAabbs.getBufferCL() )
+			};
+			
+			b3LauncherCL launcher(m_queue, m_separateAabbsKernel, "m_separateAabbsKernel");
+			launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(numLargeAabbs);
+			
+			launcher.launch1D(numLargeAabbs);
+		}
+		
+		//Write small AABBs into m_leafNodeAabbs
+		{
+			b3BufferInfoCL bufferInfo[] = 
+			{
+				b3BufferInfoCL( worldSpaceAabbs.getBufferCL() ),
+				b3BufferInfoCL( smallAabbIndices.getBufferCL() ),
+				
+				b3BufferInfoCL( m_leafNodeAabbs.getBufferCL() )
+			};
+			
+			b3LauncherCL launcher(m_queue, m_separateAabbsKernel, "m_separateAabbsKernel");
+			launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(numSmallAabbs);
+			
+			launcher.launch1D(numSmallAabbs);
+		}
+		
+		clFinish(m_queue);
+	}
+	
+	//
+	int numLeaves = numSmallAabbs;	//Number of leaves in the BVH == Number of rigid bodies with small AABBs
+	int numInternalNodes = numLeaves - 1;
+	
+	if(numLeaves < 2)
+	{
+		//Number of leaf nodes is checked in calculateOverlappingPairs() and testRaysAgainstBvhAabbs(),
+		//so it does not matter if numLeaves == 0 and rootNodeIndex == -1
+		int rootNodeIndex = numLeaves - 1;
+		m_rootNodeIndex.copyFromHostPointer(&rootNodeIndex, 1);
+		
+		//Since the AABBs need to be rearranged(sorted) for the BVH construction algorithm,
+		//m_mortonCodesAndAabbIndicies.m_value is used to map a sorted AABB index to the unsorted AABB index
+		//instead of directly moving the AABBs. It needs to be set for the ray cast traversal kernel to work.
+		//( m_mortonCodesAndAabbIndicies[].m_value == unsorted index == index of m_leafNodeAabbs )
+		if(numLeaves == 1)
+		{
+			b3SortData leaf;
+			leaf.m_value = 0;		//1 leaf so index is always 0; leaf.m_key does not need to be set
+			
+			m_mortonCodesAndAabbIndicies.resize(1);
+			m_mortonCodesAndAabbIndicies.copyFromHostPointer(&leaf, 1);
+		}
+		
+		return;
+	}
+	
+	//
+	{
+		m_internalNodeAabbs.resize(numInternalNodes);
+		m_internalNodeLeafIndexRanges.resize(numInternalNodes);
+		m_internalNodeChildNodes.resize(numInternalNodes);
+		m_internalNodeParentNodes.resize(numInternalNodes);
+
+		m_commonPrefixes.resize(numInternalNodes);
+		m_commonPrefixLengths.resize(numInternalNodes);
+		m_distanceFromRoot.resize(numInternalNodes);
+	
+		m_leafNodeParentNodes.resize(numLeaves);
+		m_mortonCodesAndAabbIndicies.resize(numLeaves);
+		m_mergedAabb.resize(numLeaves);
+	}
+	
+	//Find the merged AABB of all small AABBs; this is used to define the size of 
+	//each cell in the virtual grid for the next kernel(2^10 cells in each dimension).
+	{
+		B3_PROFILE("Find AABB of merged nodes");
+	
+		m_mergedAabb.copyFromOpenCLArray(m_leafNodeAabbs);	//Need to make a copy since the kernel modifies the array
+			
+		for(int numAabbsNeedingMerge = numLeaves; numAabbsNeedingMerge >= 2; 
+				numAabbsNeedingMerge = numAabbsNeedingMerge / 2 + numAabbsNeedingMerge % 2)
+		{
+			b3BufferInfoCL bufferInfo[] = 
+			{
+				b3BufferInfoCL( m_mergedAabb.getBufferCL() )		//Resulting AABB is stored in m_mergedAabb[0]
+			};
+			
+			b3LauncherCL launcher(m_queue, m_findAllNodesMergedAabbKernel, "m_findAllNodesMergedAabbKernel");
+			launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(numAabbsNeedingMerge);
+			
+			launcher.launch1D(numAabbsNeedingMerge);
+		}
+		
+		clFinish(m_queue);
+	}
+	
+	//Insert the center of the AABBs into a virtual grid,
+	//then convert the discrete grid coordinates into a morton code
+	//For each element in m_mortonCodesAndAabbIndicies, set
+	//	m_key == morton code (value to sort by)
+	//	m_value == small AABB index
+	{
+		B3_PROFILE("Assign morton codes");
+	
+		b3BufferInfoCL bufferInfo[] = 
+		{
+			b3BufferInfoCL( m_leafNodeAabbs.getBufferCL() ),
+			b3BufferInfoCL( m_mergedAabb.getBufferCL() ),
+			b3BufferInfoCL( m_mortonCodesAndAabbIndicies.getBufferCL() )
+		};
+		
+		b3LauncherCL launcher(m_queue, m_assignMortonCodesAndAabbIndiciesKernel, "m_assignMortonCodesAndAabbIndiciesKernel");
+		launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(numLeaves);
+		
+		launcher.launch1D(numLeaves);
+		clFinish(m_queue);
+	}
+	
+	//
+	{
+		B3_PROFILE("Sort leaves by morton codes");
+	
+		m_radixSorter.execute(m_mortonCodesAndAabbIndicies);
+		clFinish(m_queue);
+	}
+	
+	//
+	constructBinaryRadixTree();
+	
+	
+	//Since it is a sorted binary radix tree, each internal node contains a contiguous subset of leaf node indices.
+	//The root node contains leaf node indices in the range [0, numLeafNodes - 1].
+	//The child nodes of each node split their parent's index range into 2 contiguous halves.
+	//
+	//For example, if the root has indices [0, 31], its children might partition that range into [0, 11] and [12, 31].
+	//The next level in the tree could then split those ranges into [0, 2], [3, 11], [12, 22], and [23, 31].
+	//
+	//This property can be used for optimizing calculateOverlappingPairs(), to avoid testing each AABB pair twice
+	{
+		B3_PROFILE("m_findLeafIndexRangesKernel");
+	
+		b3BufferInfoCL bufferInfo[] = 
+		{
+			b3BufferInfoCL( m_internalNodeChildNodes.getBufferCL() ),
+			b3BufferInfoCL( m_internalNodeLeafIndexRanges.getBufferCL() )
+		};
+		
+		b3LauncherCL launcher(m_queue, m_findLeafIndexRangesKernel, "m_findLeafIndexRangesKernel");
+		launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(numInternalNodes);
+		
+		launcher.launch1D(numInternalNodes);
+		clFinish(m_queue);
+	}
+}
+
+void b3GpuParallelLinearBvh::calculateOverlappingPairs(b3OpenCLArray<b3Int4>& out_overlappingPairs)
+{
+	int maxPairs = out_overlappingPairs.size();
+	b3OpenCLArray<int>& numPairsGpu = m_temp;
+	
+	int reset = 0;
+	numPairsGpu.copyFromHostPointer(&reset, 1);
+	
+	//
+	if( m_leafNodeAabbs.size() > 1 )
+	{
+		B3_PROFILE("PLBVH small-small AABB test");
+	
+		int numQueryAabbs = m_leafNodeAabbs.size();
+		
+		b3BufferInfoCL bufferInfo[] = 
+		{
+			b3BufferInfoCL( m_leafNodeAabbs.getBufferCL() ),
+			
+			b3BufferInfoCL( m_rootNodeIndex.getBufferCL() ),
+			b3BufferInfoCL( m_internalNodeChildNodes.getBufferCL() ),
+			b3BufferInfoCL( m_internalNodeAabbs.getBufferCL() ),
+			b3BufferInfoCL( m_internalNodeLeafIndexRanges.getBufferCL() ),
+			b3BufferInfoCL( m_mortonCodesAndAabbIndicies.getBufferCL() ),
+			
+			b3BufferInfoCL( numPairsGpu.getBufferCL() ),
+			b3BufferInfoCL( out_overlappingPairs.getBufferCL() )
+		};
+		
+		b3LauncherCL launcher(m_queue, m_plbvhCalculateOverlappingPairsKernel, "m_plbvhCalculateOverlappingPairsKernel");
+		launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(maxPairs);
+		launcher.setConst(numQueryAabbs);
+		
+		launcher.launch1D(numQueryAabbs);
+		clFinish(m_queue);
+	}
+	
+	int numLargeAabbRigids = m_largeAabbs.size();
+	if( numLargeAabbRigids > 0 && m_leafNodeAabbs.size() > 0 )
+	{
+		B3_PROFILE("PLBVH large-small AABB test");
+	
+		int numQueryAabbs = m_leafNodeAabbs.size();
+		
+		b3BufferInfoCL bufferInfo[] = 
+		{
+			b3BufferInfoCL( m_leafNodeAabbs.getBufferCL() ),
+			b3BufferInfoCL( m_largeAabbs.getBufferCL() ),
+			
+			b3BufferInfoCL( numPairsGpu.getBufferCL() ),
+			b3BufferInfoCL( out_overlappingPairs.getBufferCL() )
+		};
+		
+		b3LauncherCL launcher(m_queue, m_plbvhLargeAabbAabbTestKernel, "m_plbvhLargeAabbAabbTestKernel");
+		launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(maxPairs);
+		launcher.setConst(numLargeAabbRigids);
+		launcher.setConst(numQueryAabbs);
+		
+		launcher.launch1D(numQueryAabbs);
+		clFinish(m_queue);
+	}
+	
+	
+	//
+	int numPairs = -1;
+	numPairsGpu.copyToHostPointer(&numPairs, 1);
+	if(numPairs > maxPairs)
+	{
+		b3Error("Error running out of pairs: numPairs = %d, maxPairs = %d.\n", numPairs, maxPairs);
+		numPairs = maxPairs;
+		numPairsGpu.copyFromHostPointer(&maxPairs, 1);
+	}
+	
+	out_overlappingPairs.resize(numPairs);
+}
+
+
+void b3GpuParallelLinearBvh::testRaysAgainstBvhAabbs(const b3OpenCLArray<b3RayInfo>& rays, 
+							b3OpenCLArray<int>& out_numRayRigidPairs, b3OpenCLArray<b3Int2>& out_rayRigidPairs)
+{
+	B3_PROFILE("PLBVH testRaysAgainstBvhAabbs()");
+	
+	int numRays = rays.size();
+	int maxRayRigidPairs = out_rayRigidPairs.size();
+	
+	int reset = 0;
+	out_numRayRigidPairs.copyFromHostPointer(&reset, 1);
+	
+	//
+	if( m_leafNodeAabbs.size() > 0 )
+	{
+		B3_PROFILE("PLBVH ray test small AABB");
+	
+		b3BufferInfoCL bufferInfo[] = 
+		{
+			b3BufferInfoCL( m_leafNodeAabbs.getBufferCL() ),
+			
+			b3BufferInfoCL( m_rootNodeIndex.getBufferCL() ),
+			b3BufferInfoCL( m_internalNodeChildNodes.getBufferCL() ),
+			b3BufferInfoCL( m_internalNodeAabbs.getBufferCL() ),
+			b3BufferInfoCL( m_internalNodeLeafIndexRanges.getBufferCL() ),
+			b3BufferInfoCL( m_mortonCodesAndAabbIndicies.getBufferCL() ),
+			
+			b3BufferInfoCL( rays.getBufferCL() ),
+			
+			b3BufferInfoCL( out_numRayRigidPairs.getBufferCL() ),
+			b3BufferInfoCL( out_rayRigidPairs.getBufferCL() )
+		};
+		
+		b3LauncherCL launcher(m_queue, m_plbvhRayTraverseKernel, "m_plbvhRayTraverseKernel");
+		launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(maxRayRigidPairs);
+		launcher.setConst(numRays);
+		
+		launcher.launch1D(numRays);
+		clFinish(m_queue);
+	}
+	
+	int numLargeAabbRigids = m_largeAabbs.size();
+	if(numLargeAabbRigids > 0)
+	{
+		B3_PROFILE("PLBVH ray test large AABB");
+		
+		b3BufferInfoCL bufferInfo[] = 
+		{
+			b3BufferInfoCL( m_largeAabbs.getBufferCL() ),
+			b3BufferInfoCL( rays.getBufferCL() ),
+			
+			b3BufferInfoCL( out_numRayRigidPairs.getBufferCL() ),
+			b3BufferInfoCL( out_rayRigidPairs.getBufferCL() )
+		};
+		
+		b3LauncherCL launcher(m_queue, m_plbvhLargeAabbRayTestKernel, "m_plbvhLargeAabbRayTestKernel");
+		launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(numLargeAabbRigids);
+		launcher.setConst(maxRayRigidPairs);
+		launcher.setConst(numRays);
+		
+		launcher.launch1D(numRays);
+		clFinish(m_queue);
+	}
+	
+	//
+	int numRayRigidPairs = -1;
+	out_numRayRigidPairs.copyToHostPointer(&numRayRigidPairs, 1);
+	
+	if(numRayRigidPairs > maxRayRigidPairs)
+		b3Error("Error running out of rayRigid pairs: numRayRigidPairs = %d, maxRayRigidPairs = %d.\n", numRayRigidPairs, maxRayRigidPairs);
+	
+}
+
+void b3GpuParallelLinearBvh::constructBinaryRadixTree()
+{
+	B3_PROFILE("b3GpuParallelLinearBvh::constructBinaryRadixTree()");
+	
+	int numLeaves = m_leafNodeAabbs.size();
+	int numInternalNodes = numLeaves - 1;
+	
+	//Each internal node is placed in between 2 leaf nodes.
+	//By using this arrangement and computing the common prefix between
+	//these 2 adjacent leaf nodes, it is possible to quickly construct a binary radix tree.
+	{
+		B3_PROFILE("m_computeAdjacentPairCommonPrefixKernel");
+		
+		b3BufferInfoCL bufferInfo[] = 
+		{
+			b3BufferInfoCL( m_mortonCodesAndAabbIndicies.getBufferCL() ),
+			b3BufferInfoCL( m_commonPrefixes.getBufferCL() ),
+			b3BufferInfoCL( m_commonPrefixLengths.getBufferCL() )
+		};
+		
+		b3LauncherCL launcher(m_queue, m_computeAdjacentPairCommonPrefixKernel, "m_computeAdjacentPairCommonPrefixKernel");
+		launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(numInternalNodes);
+		
+		launcher.launch1D(numInternalNodes);
+		clFinish(m_queue);
+	}
+	
+	//For each leaf node, select its parent node by 
+	//comparing the 2 nearest internal nodes and assign child node indices
+	{
+		B3_PROFILE("m_buildBinaryRadixTreeLeafNodesKernel");
+		
+		b3BufferInfoCL bufferInfo[] = 
+		{
+			b3BufferInfoCL( m_commonPrefixLengths.getBufferCL() ),
+			b3BufferInfoCL( m_leafNodeParentNodes.getBufferCL() ),
+			b3BufferInfoCL( m_internalNodeChildNodes.getBufferCL() )
+		};
+		
+		b3LauncherCL launcher(m_queue, m_buildBinaryRadixTreeLeafNodesKernel, "m_buildBinaryRadixTreeLeafNodesKernel");
+		launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(numLeaves);
+		
+		launcher.launch1D(numLeaves);
+		clFinish(m_queue);
+	}
+	
+	//For each internal node, perform 2 binary searches among the other internal nodes
+	//to its left and right to find its potential parent nodes and assign child node indices
+	{
+		B3_PROFILE("m_buildBinaryRadixTreeInternalNodesKernel");
+		
+		b3BufferInfoCL bufferInfo[] = 
+		{
+			b3BufferInfoCL( m_commonPrefixes.getBufferCL() ),
+			b3BufferInfoCL( m_commonPrefixLengths.getBufferCL() ),
+			b3BufferInfoCL( m_internalNodeChildNodes.getBufferCL() ),
+			b3BufferInfoCL( m_internalNodeParentNodes.getBufferCL() ),
+			b3BufferInfoCL( m_rootNodeIndex.getBufferCL() )
+		};
+		
+		b3LauncherCL launcher(m_queue, m_buildBinaryRadixTreeInternalNodesKernel, "m_buildBinaryRadixTreeInternalNodesKernel");
+		launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(numInternalNodes);
+		
+		launcher.launch1D(numInternalNodes);
+		clFinish(m_queue);
+	}
+	
+	//Find the number of nodes seperating each internal node and the root node
+	//so that the AABBs can be set using the next kernel.
+	//Also determine the maximum number of nodes separating an internal node and the root node.
+	{
+		B3_PROFILE("m_findDistanceFromRootKernel");
+	
+		b3BufferInfoCL bufferInfo[] = 
+		{
+			b3BufferInfoCL( m_rootNodeIndex.getBufferCL() ),
+			b3BufferInfoCL( m_internalNodeParentNodes.getBufferCL() ),
+			b3BufferInfoCL( m_maxDistanceFromRoot.getBufferCL() ),
+			b3BufferInfoCL( m_distanceFromRoot.getBufferCL() )
+		};
+		
+		b3LauncherCL launcher(m_queue, m_findDistanceFromRootKernel, "m_findDistanceFromRootKernel");
+		launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(numInternalNodes);
+		
+		launcher.launch1D(numInternalNodes);
+		clFinish(m_queue);
+	}
+	
+	//Starting from the internal nodes nearest to the leaf nodes, recursively move up
+	//the tree towards the root to set the AABBs of each internal node; each internal node
+	//checks its children and merges their AABBs
+	{
+		B3_PROFILE("m_buildBinaryRadixTreeAabbsRecursiveKernel");
+		
+		int maxDistanceFromRoot = -1;
+		{
+			B3_PROFILE("copy maxDistanceFromRoot to CPU");
+			m_maxDistanceFromRoot.copyToHostPointer(&maxDistanceFromRoot, 1);
+			clFinish(m_queue);
+		}
+		
+		for(int distanceFromRoot = maxDistanceFromRoot; distanceFromRoot >= 0; --distanceFromRoot)
+		{
+			b3BufferInfoCL bufferInfo[] = 
+			{
+				b3BufferInfoCL( m_distanceFromRoot.getBufferCL() ),
+				b3BufferInfoCL( m_mortonCodesAndAabbIndicies.getBufferCL() ),
+				b3BufferInfoCL( m_internalNodeChildNodes.getBufferCL() ),
+				b3BufferInfoCL( m_leafNodeAabbs.getBufferCL() ),
+				b3BufferInfoCL( m_internalNodeAabbs.getBufferCL() )
+			};
+			
+			b3LauncherCL launcher(m_queue, m_buildBinaryRadixTreeAabbsRecursiveKernel, "m_buildBinaryRadixTreeAabbsRecursiveKernel");
+			launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(maxDistanceFromRoot);
+			launcher.setConst(distanceFromRoot);
+			launcher.setConst(numInternalNodes);
+			
+			//It may seem inefficent to launch a thread for each internal node when a
+			//much smaller number of nodes is actually processed, but this is actually
+			//faster than determining the exact nodes that are ready to merge their child AABBs. 
+			launcher.launch1D(numInternalNodes);
+		}
+		
+		clFinish(m_queue);
+	}
+}
+
+	
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvh.h b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvh.h
new file mode 100644
index 0000000000..effe617b7b
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvh.h
@@ -0,0 +1,125 @@
+/*
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose,
+including commercial applications, and to alter it and redistribute it freely,
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Initial Author Jackson Lee, 2014
+
+#ifndef B3_GPU_PARALLEL_LINEAR_BVH_H
+#define B3_GPU_PARALLEL_LINEAR_BVH_H
+
+//#include "Bullet3Collision/BroadPhaseCollision/shared/b3Aabb.h"
+#include "Bullet3OpenCL/BroadphaseCollision/b3SapAabb.h"
+#include "Bullet3Common/shared/b3Int2.h"
+#include "Bullet3Common/shared/b3Int4.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3RaycastInfo.h"
+
+#include "Bullet3OpenCL/ParallelPrimitives/b3FillCL.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3PrefixScanCL.h"
+
+#include "Bullet3OpenCL/BroadphaseCollision/kernels/parallelLinearBvhKernels.h"
+
+#define b3Int64 cl_long
+
+///@brief GPU Parallel Linearized Bounding Volume Heirarchy(LBVH) that is reconstructed every frame
+///@remarks
+///See presentation in docs/b3GpuParallelLinearBvh.pdf for algorithm details.
+///@par
+///Related papers: \n
+///"Fast BVH Construction on GPUs" [Lauterbach et al. 2009] \n
+///"Maximizing Parallelism in the Construction of BVHs, Octrees, and k-d trees" [Karras 2012] \n
+///@par
+///The basic algorithm for building the BVH as presented in [Lauterbach et al. 2009] consists of 4 stages:
+/// - [fully parallel] Assign morton codes for each AABB using its center (after quantizing the AABB centers into a virtual grid) 
+/// - [fully parallel] Sort morton codes
+/// - [somewhat parallel] Build binary radix tree (assign parent/child pointers for internal nodes of the BVH) 
+/// - [somewhat parallel] Set internal node AABBs 
+///@par
+///[Karras 2012] improves on the algorithm by introducing fully parallel methods for the last 2 stages.
+///The BVH implementation here shares many concepts with [Karras 2012], but a different method is used for constructing the tree.
+///Instead of searching for the child nodes of each internal node, we search for the parent node of each node.
+///Additionally, a non-atomic traversal that starts from the leaf nodes and moves towards the root node is used to set the AABBs.
+class b3GpuParallelLinearBvh
+{
+	cl_command_queue m_queue;
+	
+	cl_program m_parallelLinearBvhProgram;
+	
+	cl_kernel m_separateAabbsKernel;
+	cl_kernel m_findAllNodesMergedAabbKernel;
+	cl_kernel m_assignMortonCodesAndAabbIndiciesKernel;
+	
+	//Binary radix tree construction kernels
+	cl_kernel m_computeAdjacentPairCommonPrefixKernel;
+	cl_kernel m_buildBinaryRadixTreeLeafNodesKernel;
+	cl_kernel m_buildBinaryRadixTreeInternalNodesKernel;
+	cl_kernel m_findDistanceFromRootKernel;
+	cl_kernel m_buildBinaryRadixTreeAabbsRecursiveKernel;
+	
+	cl_kernel m_findLeafIndexRangesKernel;
+	
+	//Traversal kernels
+	cl_kernel m_plbvhCalculateOverlappingPairsKernel;
+	cl_kernel m_plbvhRayTraverseKernel;
+	cl_kernel m_plbvhLargeAabbAabbTestKernel;
+	cl_kernel m_plbvhLargeAabbRayTestKernel;
+	
+	b3RadixSort32CL m_radixSorter;
+	
+	//1 element
+	b3OpenCLArray<int> m_rootNodeIndex;							//Most significant bit(0x80000000) is set to indicate internal node
+	b3OpenCLArray<int> m_maxDistanceFromRoot;					//Max number of internal nodes between an internal node and the root node
+	b3OpenCLArray<int> m_temp;									//Used to hold the number of pairs in calculateOverlappingPairs()
+	
+	//1 element per internal node (number_of_internal_nodes == number_of_leaves - 1)
+	b3OpenCLArray<b3SapAabb> m_internalNodeAabbs;
+	b3OpenCLArray<b3Int2> m_internalNodeLeafIndexRanges;		//x == min leaf index, y == max leaf index
+	b3OpenCLArray<b3Int2> m_internalNodeChildNodes;				//x == left child, y == right child; msb(0x80000000) is set to indicate internal node
+	b3OpenCLArray<int> m_internalNodeParentNodes;				//For parent node index, msb(0x80000000) is not set since it is always internal
+	
+	//1 element per internal node; for binary radix tree construction
+	b3OpenCLArray<b3Int64> m_commonPrefixes;
+	b3OpenCLArray<int> m_commonPrefixLengths;
+	b3OpenCLArray<int> m_distanceFromRoot;						//Number of internal nodes between this node and the root
+	
+	//1 element per leaf node (leaf nodes only include small AABBs)
+	b3OpenCLArray<int> m_leafNodeParentNodes;					//For parent node index, msb(0x80000000) is not set since it is always internal
+	b3OpenCLArray<b3SortData> m_mortonCodesAndAabbIndicies;		//m_key == morton code, m_value == aabb index in m_leafNodeAabbs
+	b3OpenCLArray<b3SapAabb> m_mergedAabb;						//m_mergedAabb[0] contains the merged AABB of all leaf nodes
+	b3OpenCLArray<b3SapAabb> m_leafNodeAabbs;					//Contains only small AABBs
+	
+	//1 element per large AABB, which is not stored in the BVH
+	b3OpenCLArray<b3SapAabb> m_largeAabbs;
+	
+public:
+	b3GpuParallelLinearBvh(cl_context context, cl_device_id device, cl_command_queue queue);
+	virtual ~b3GpuParallelLinearBvh();
+	
+	///Must be called before any other function
+	void build(const b3OpenCLArray<b3SapAabb>& worldSpaceAabbs, const b3OpenCLArray<int>& smallAabbIndices, 
+				const b3OpenCLArray<int>& largeAabbIndices);
+	
+	///calculateOverlappingPairs() uses the worldSpaceAabbs parameter of b3GpuParallelLinearBvh::build() as the query AABBs.
+	///@param out_overlappingPairs The size() of this array is used to determine the max number of pairs.
+	///If the number of overlapping pairs is < out_overlappingPairs.size(), out_overlappingPairs is resized.
+	void calculateOverlappingPairs(b3OpenCLArray<b3Int4>& out_overlappingPairs);
+	
+	///@param out_numRigidRayPairs Array of length 1; contains the number of detected ray-rigid AABB intersections;
+	///this value may be greater than out_rayRigidPairs.size() if out_rayRigidPairs is not large enough.
+	///@param out_rayRigidPairs Contains an array of rays intersecting rigid AABBs; x == ray index, y == rigid body index.
+	///If the size of this array is insufficient to hold all ray-rigid AABB intersections, additional intersections are discarded.
+	void testRaysAgainstBvhAabbs(const b3OpenCLArray<b3RayInfo>& rays, 
+								b3OpenCLArray<int>& out_numRayRigidPairs, b3OpenCLArray<b3Int2>& out_rayRigidPairs);
+								
+private:
+	void constructBinaryRadixTree();
+};
+
+#endif
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvhBroadphase.cpp b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvhBroadphase.cpp
new file mode 100644
index 0000000000..d2618024ac
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvhBroadphase.cpp
@@ -0,0 +1,80 @@
+/*
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose,
+including commercial applications, and to alter it and redistribute it freely,
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Initial Author Jackson Lee, 2014
+
+#include "b3GpuParallelLinearBvhBroadphase.h"
+
+b3GpuParallelLinearBvhBroadphase::b3GpuParallelLinearBvhBroadphase(cl_context context, cl_device_id device, cl_command_queue queue) : 
+	m_plbvh(context, device, queue),
+	
+	m_overlappingPairsGpu(context, queue),
+	
+	m_aabbsGpu(context, queue),
+	m_smallAabbsMappingGpu(context, queue),
+	m_largeAabbsMappingGpu(context, queue)
+{
+}
+
+void b3GpuParallelLinearBvhBroadphase::createProxy(const b3Vector3& aabbMin, const b3Vector3& aabbMax, int userPtr,  int collisionFilterGroup,  int collisionFilterMask)
+{
+	int newAabbIndex = m_aabbsCpu.size();
+
+	b3SapAabb aabb;
+	aabb.m_minVec = aabbMin;
+	aabb.m_maxVec = aabbMax;
+	
+	aabb.m_minIndices[3] = userPtr;
+	aabb.m_signedMaxIndices[3] = newAabbIndex;
+	
+	m_smallAabbsMappingCpu.push_back(newAabbIndex);
+	
+	m_aabbsCpu.push_back(aabb);
+}
+void b3GpuParallelLinearBvhBroadphase::createLargeProxy(const b3Vector3& aabbMin, const b3Vector3& aabbMax, int userPtr,  int collisionFilterGroup,  int collisionFilterMask)
+{
+	int newAabbIndex = m_aabbsCpu.size();
+
+	b3SapAabb aabb;
+	aabb.m_minVec = aabbMin;
+	aabb.m_maxVec = aabbMax;
+	
+	aabb.m_minIndices[3] = userPtr;
+	aabb.m_signedMaxIndices[3] = newAabbIndex;
+	
+	m_largeAabbsMappingCpu.push_back(newAabbIndex);
+	
+	m_aabbsCpu.push_back(aabb);
+}
+
+void b3GpuParallelLinearBvhBroadphase::calculateOverlappingPairs(int maxPairs)
+{
+	//Reconstruct BVH
+	m_plbvh.build(m_aabbsGpu, m_smallAabbsMappingGpu, m_largeAabbsMappingGpu);
+	
+	//
+	m_overlappingPairsGpu.resize(maxPairs);
+	m_plbvh.calculateOverlappingPairs(m_overlappingPairsGpu);
+}
+void b3GpuParallelLinearBvhBroadphase::calculateOverlappingPairsHost(int maxPairs)
+{
+	b3Assert(0);	//CPU version not implemented
+}
+
+void b3GpuParallelLinearBvhBroadphase::writeAabbsToGpu() 
+{ 
+	m_aabbsGpu.copyFromHost(m_aabbsCpu); 
+	m_smallAabbsMappingGpu.copyFromHost(m_smallAabbsMappingCpu);
+	m_largeAabbsMappingGpu.copyFromHost(m_largeAabbsMappingCpu);
+}
+	
+	
+	
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvhBroadphase.h b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvhBroadphase.h
new file mode 100644
index 0000000000..e518500637
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvhBroadphase.h
@@ -0,0 +1,66 @@
+/*
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose,
+including commercial applications, and to alter it and redistribute it freely,
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Initial Author Jackson Lee, 2014
+
+#ifndef B3_GPU_PARALLEL_LINEAR_BVH_BROADPHASE_H
+#define B3_GPU_PARALLEL_LINEAR_BVH_BROADPHASE_H
+
+#include "Bullet3OpenCL/BroadphaseCollision/b3GpuBroadphaseInterface.h"
+
+#include "b3GpuParallelLinearBvh.h"
+
+class b3GpuParallelLinearBvhBroadphase : public b3GpuBroadphaseInterface
+{
+	b3GpuParallelLinearBvh m_plbvh;
+	
+	b3OpenCLArray<b3Int4> m_overlappingPairsGpu;
+	
+	b3OpenCLArray<b3SapAabb> m_aabbsGpu;
+	b3OpenCLArray<int> m_smallAabbsMappingGpu;
+	b3OpenCLArray<int> m_largeAabbsMappingGpu;
+	
+	b3AlignedObjectArray<b3SapAabb> m_aabbsCpu;
+	b3AlignedObjectArray<int> m_smallAabbsMappingCpu;
+	b3AlignedObjectArray<int> m_largeAabbsMappingCpu;
+	
+public:
+	b3GpuParallelLinearBvhBroadphase(cl_context context, cl_device_id device, cl_command_queue queue);
+	virtual ~b3GpuParallelLinearBvhBroadphase() {}
+
+	virtual void createProxy(const b3Vector3& aabbMin, const b3Vector3& aabbMax, int userPtr,  int collisionFilterGroup,  int collisionFilterMask);
+	virtual void createLargeProxy(const b3Vector3& aabbMin, const b3Vector3& aabbMax, int userPtr,  int collisionFilterGroup,  int collisionFilterMask);
+	
+	virtual void calculateOverlappingPairs(int maxPairs);
+	virtual void calculateOverlappingPairsHost(int maxPairs);
+
+	//call writeAabbsToGpu after done making all changes (createProxy etc)
+	virtual void writeAabbsToGpu();
+	
+	virtual int	getNumOverlap() { return m_overlappingPairsGpu.size(); }
+	virtual cl_mem getOverlappingPairBuffer() { return m_overlappingPairsGpu.getBufferCL(); }
+
+	virtual cl_mem getAabbBufferWS() { return m_aabbsGpu.getBufferCL(); }
+	virtual b3OpenCLArray<b3SapAabb>& getAllAabbsGPU() { return m_aabbsGpu; }
+	
+	virtual b3OpenCLArray<b3Int4>& getOverlappingPairsGPU() { return m_overlappingPairsGpu; }
+	virtual b3OpenCLArray<int>& getSmallAabbIndicesGPU() { return m_smallAabbsMappingGpu; }
+	virtual b3OpenCLArray<int>& getLargeAabbIndicesGPU() { return m_largeAabbsMappingGpu; }
+	
+	virtual b3AlignedObjectArray<b3SapAabb>& getAllAabbsCPU() { return m_aabbsCpu; }
+	
+	static b3GpuBroadphaseInterface* CreateFunc(cl_context context, cl_device_id device, cl_command_queue queue)
+	{
+		return new b3GpuParallelLinearBvhBroadphase(context, device, queue);
+	}
+};
+
+#endif
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuSapBroadphase.cpp b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuSapBroadphase.cpp
new file mode 100644
index 0000000000..c45fbbdcaa
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuSapBroadphase.cpp
@@ -0,0 +1,1366 @@
+
+bool searchIncremental3dSapOnGpu = true;
+#include <limits.h>
+#include "b3GpuSapBroadphase.h"
+#include "Bullet3Common/b3Vector3.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3PrefixScanFloat4CL.h"
+
+
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "kernels/sapKernels.h"
+
+#include "Bullet3Common/b3MinMax.h"
+
+#define B3_BROADPHASE_SAP_PATH "src/Bullet3OpenCL/BroadphaseCollision/kernels/sap.cl"
+
+/*
+	
+ 
+	
+	
+	
+ 
+	b3OpenCLArray<int> m_pairCount;
+ 
+ 
+	b3OpenCLArray<b3SapAabb>	m_allAabbsGPU;
+	b3AlignedObjectArray<b3SapAabb>	m_allAabbsCPU;
+ 
+	virtual b3OpenCLArray<b3SapAabb>&	getAllAabbsGPU()
+	{
+ return m_allAabbsGPU;
+	}
+	virtual b3AlignedObjectArray<b3SapAabb>&	getAllAabbsCPU()
+	{
+ return m_allAabbsCPU;
+	}
+ 
+	b3OpenCLArray<b3Vector3>	m_sum;
+	b3OpenCLArray<b3Vector3>	m_sum2;
+	b3OpenCLArray<b3Vector3>	m_dst;
+ 
+	b3OpenCLArray<int>	m_smallAabbsMappingGPU;
+	b3AlignedObjectArray<int> m_smallAabbsMappingCPU;
+ 
+	b3OpenCLArray<int>	m_largeAabbsMappingGPU;
+	b3AlignedObjectArray<int> m_largeAabbsMappingCPU;
+ 
+	
+	b3OpenCLArray<b3Int4>		m_overlappingPairs;
+ 
+	//temporary gpu work memory
+	b3OpenCLArray<b3SortData>	m_gpuSmallSortData;
+	b3OpenCLArray<b3SapAabb>	m_gpuSmallSortedAabbs;
+ 
+	class b3PrefixScanFloat4CL*		m_prefixScanFloat4;
+ */
+
+b3GpuSapBroadphase::b3GpuSapBroadphase(cl_context ctx,cl_device_id device, cl_command_queue  q , b3GpuSapKernelType kernelType)
+:m_context(ctx),
+m_device(device),
+m_queue(q),
+
+m_objectMinMaxIndexGPUaxis0(ctx,q),
+m_objectMinMaxIndexGPUaxis1(ctx,q),
+m_objectMinMaxIndexGPUaxis2(ctx,q),
+m_objectMinMaxIndexGPUaxis0prev(ctx,q),
+m_objectMinMaxIndexGPUaxis1prev(ctx,q),
+m_objectMinMaxIndexGPUaxis2prev(ctx,q),
+m_sortedAxisGPU0(ctx,q),
+m_sortedAxisGPU1(ctx,q),
+m_sortedAxisGPU2(ctx,q),
+m_sortedAxisGPU0prev(ctx,q),
+m_sortedAxisGPU1prev(ctx,q),
+m_sortedAxisGPU2prev(ctx,q),
+m_addedHostPairsGPU(ctx,q),
+m_removedHostPairsGPU(ctx,q),
+m_addedCountGPU(ctx,q),
+m_removedCountGPU(ctx,q),
+m_currentBuffer(-1),
+m_pairCount(ctx,q),
+m_allAabbsGPU(ctx,q),
+m_sum(ctx,q),
+m_sum2(ctx,q),
+m_dst(ctx,q),
+m_smallAabbsMappingGPU(ctx,q),
+m_largeAabbsMappingGPU(ctx,q),
+m_overlappingPairs(ctx,q),
+m_gpuSmallSortData(ctx,q),
+m_gpuSmallSortedAabbs(ctx,q)
+{
+	const char* sapSrc = sapCL;
+    
+    
+	cl_int errNum=0;
+
+	b3Assert(m_context);
+	b3Assert(m_device);
+	cl_program sapProg = b3OpenCLUtils::compileCLProgramFromString(m_context,m_device,sapSrc,&errNum,"",B3_BROADPHASE_SAP_PATH);
+	b3Assert(errNum==CL_SUCCESS);
+
+
+	b3Assert(errNum==CL_SUCCESS);
+#ifndef __APPLE__
+	m_prefixScanFloat4 = new b3PrefixScanFloat4CL(m_context,m_device,m_queue);
+#else
+	m_prefixScanFloat4 = 0;
+#endif
+	m_sapKernel = 0;
+	
+	switch (kernelType)
+	{
+		case B3_GPU_SAP_KERNEL_BRUTE_FORCE_CPU:
+		{
+			m_sapKernel=0;
+			break;
+		}
+		case 		B3_GPU_SAP_KERNEL_BRUTE_FORCE_GPU:
+		{
+			m_sapKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,sapSrc, "computePairsKernelBruteForce",&errNum,sapProg );
+			break;
+		}
+
+		case B3_GPU_SAP_KERNEL_ORIGINAL:
+		{
+			m_sapKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,sapSrc, "computePairsKernelOriginal",&errNum,sapProg );
+			break;
+		}
+		case B3_GPU_SAP_KERNEL_BARRIER:
+		{
+			m_sapKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,sapSrc, "computePairsKernelBarrier",&errNum,sapProg );
+			break;
+		}
+		case B3_GPU_SAP_KERNEL_LOCAL_SHARED_MEMORY:
+		{
+			m_sapKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,sapSrc, "computePairsKernelLocalSharedMemory",&errNum,sapProg );
+			break;
+		}
+
+		default:
+		{
+			m_sapKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,sapSrc, "computePairsKernelLocalSharedMemory",&errNum,sapProg );
+			b3Error("Unknown 3D GPU SAP provided, fallback to computePairsKernelLocalSharedMemory");
+		}
+	};
+	
+		
+	
+	m_sap2Kernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,sapSrc, "computePairsKernelTwoArrays",&errNum,sapProg );
+	b3Assert(errNum==CL_SUCCESS);
+
+	m_prepareSumVarianceKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,sapSrc, "prepareSumVarianceKernel",&errNum,sapProg );
+	b3Assert(errNum==CL_SUCCESS);
+
+		
+	m_flipFloatKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,sapSrc, "flipFloatKernel",&errNum,sapProg );
+
+	m_copyAabbsKernel= b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,sapSrc, "copyAabbsKernel",&errNum,sapProg );
+
+	m_scatterKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,sapSrc, "scatterKernel",&errNum,sapProg );
+
+	m_sorter = new b3RadixSort32CL(m_context,m_device,m_queue);
+}
+
+b3GpuSapBroadphase::~b3GpuSapBroadphase()
+{
+	delete m_sorter;
+	delete m_prefixScanFloat4;
+
+	clReleaseKernel(m_scatterKernel);
+	clReleaseKernel(m_flipFloatKernel);
+	clReleaseKernel(m_copyAabbsKernel);
+	clReleaseKernel(m_sapKernel);
+	clReleaseKernel(m_sap2Kernel);
+	clReleaseKernel(m_prepareSumVarianceKernel);
+	
+
+}
+
+/// conservative test for overlap between two aabbs
+static bool TestAabbAgainstAabb2(const b3Vector3 &aabbMin1, const b3Vector3 &aabbMax1,
+								const b3Vector3 &aabbMin2, const b3Vector3 &aabbMax2)
+{
+	bool overlap = true;
+	overlap = (aabbMin1.getX() > aabbMax2.getX() || aabbMax1.getX() < aabbMin2.getX()) ? false : overlap;
+	overlap = (aabbMin1.getZ() > aabbMax2.getZ() || aabbMax1.getZ() < aabbMin2.getZ()) ? false : overlap;
+	overlap = (aabbMin1.getY() > aabbMax2.getY() || aabbMax1.getY() < aabbMin2.getY()) ? false : overlap;
+	return overlap;
+}
+
+
+
+//http://stereopsis.com/radix.html
+static unsigned int FloatFlip(float fl)
+{
+	unsigned int f = *(unsigned int*)&fl;
+	unsigned int mask = -(int)(f >> 31) | 0x80000000;
+	return f ^ mask;
+};
+
+void  b3GpuSapBroadphase::init3dSap()
+{
+	if (m_currentBuffer<0)
+	{
+		m_allAabbsGPU.copyToHost(m_allAabbsCPU);
+
+		m_currentBuffer = 0;
+		for (int axis=0;axis<3;axis++)
+		{
+			for (int buf=0;buf<2;buf++)
+			{
+				int totalNumAabbs = m_allAabbsCPU.size();
+				int numEndPoints = 2*totalNumAabbs;
+				m_sortedAxisCPU[axis][buf].resize(numEndPoints);
+
+				if (buf==m_currentBuffer)
+				{
+					for (int i=0;i<totalNumAabbs;i++)
+					{
+						m_sortedAxisCPU[axis][buf][i*2].m_key = FloatFlip(m_allAabbsCPU[i].m_min[axis])-1;
+						m_sortedAxisCPU[axis][buf][i*2].m_value = i*2;
+						m_sortedAxisCPU[axis][buf][i*2+1].m_key = FloatFlip(m_allAabbsCPU[i].m_max[axis])+1;
+						m_sortedAxisCPU[axis][buf][i*2+1].m_value = i*2+1;
+					}
+				}
+			}
+		}
+
+		for (int axis=0;axis<3;axis++)
+		{
+			m_sorter->executeHost(m_sortedAxisCPU[axis][m_currentBuffer]);
+		}
+
+		for (int axis=0;axis<3;axis++)
+		{
+			//int totalNumAabbs = m_allAabbsCPU.size();
+			int numEndPoints = m_sortedAxisCPU[axis][m_currentBuffer].size();
+			m_objectMinMaxIndexCPU[axis][m_currentBuffer].resize(numEndPoints);
+			for (int i=0;i<numEndPoints;i++)
+			{
+				int destIndex = m_sortedAxisCPU[axis][m_currentBuffer][i].m_value;
+				int newDest = destIndex/2;
+				if (destIndex&1)
+				{
+					m_objectMinMaxIndexCPU[axis][m_currentBuffer][newDest].y=i;
+				} else
+				{
+					m_objectMinMaxIndexCPU[axis][m_currentBuffer][newDest].x=i;
+				}
+			}
+		}
+
+	}
+}
+
+
+static bool b3PairCmp(const b3Int4& p, const b3Int4& q)
+{
+	return ((p.x<q.x) || ((p.x==q.x) && (p.y<q.y)));
+}
+
+
+static bool operator==(const b3Int4& a,const b3Int4& b)
+{
+	return a.x == b.x && a.y == b.y;
+};
+
+static bool operator<(const b3Int4& a,const b3Int4& b)
+{
+	return a.x < b.x || (a.x == b.x && a.y < b.y);
+};
+
+static bool operator>(const b3Int4& a,const b3Int4& b)
+{
+	return a.x > b.x || (a.x == b.x && a.y > b.y);
+};
+
+b3AlignedObjectArray<b3Int4> addedHostPairs;
+b3AlignedObjectArray<b3Int4> removedHostPairs;
+
+b3AlignedObjectArray<b3SapAabb>	preAabbs;
+
+void  b3GpuSapBroadphase::calculateOverlappingPairsHostIncremental3Sap()
+{
+	//static int framepje = 0;
+	//printf("framepje=%d\n",framepje++);
+	
+
+	B3_PROFILE("calculateOverlappingPairsHostIncremental3Sap");
+
+	addedHostPairs.resize(0);
+	removedHostPairs.resize(0);
+
+	b3Assert(m_currentBuffer>=0);
+	
+	{
+		preAabbs.resize(m_allAabbsCPU.size());
+		for (int i=0;i<preAabbs.size();i++)
+		{
+			preAabbs[i]=m_allAabbsCPU[i];
+		}
+	}
+	
+
+	if (m_currentBuffer<0)
+		return;
+	{
+		B3_PROFILE("m_allAabbsGPU.copyToHost");
+		m_allAabbsGPU.copyToHost(m_allAabbsCPU);
+	}
+
+	b3AlignedObjectArray<b3Int4> allPairs;
+	{
+		B3_PROFILE("m_overlappingPairs.copyToHost");
+		m_overlappingPairs.copyToHost(allPairs);
+	}
+	if (0)
+	{
+	{
+		printf("ab[40].min=%f,%f,%f,ab[40].max=%f,%f,%f\n",
+		m_allAabbsCPU[40].m_min[0],	m_allAabbsCPU[40].m_min[1],m_allAabbsCPU[40].m_min[2],
+		m_allAabbsCPU[40].m_max[0],	m_allAabbsCPU[40].m_max[1],m_allAabbsCPU[40].m_max[2]);
+	}
+
+	{
+		printf("ab[53].min=%f,%f,%f,ab[53].max=%f,%f,%f\n",
+		m_allAabbsCPU[53].m_min[0],	m_allAabbsCPU[53].m_min[1],m_allAabbsCPU[53].m_min[2],
+		m_allAabbsCPU[53].m_max[0],	m_allAabbsCPU[53].m_max[1],m_allAabbsCPU[53].m_max[2]);
+	}
+	
+
+	{
+	b3Int4 newPair;
+	newPair.x = 40;
+	newPair.y = 53;
+		int index = allPairs.findBinarySearch(newPair);
+		printf("hasPair(40,53)=%d out of %d\n",index, allPairs.size());
+
+		{
+			int overlap = TestAabbAgainstAabb2((const b3Vector3&)m_allAabbsCPU[40].m_min, (const b3Vector3&)m_allAabbsCPU[40].m_max,(const b3Vector3&)m_allAabbsCPU[53].m_min,(const b3Vector3&)m_allAabbsCPU[53].m_max);
+			printf("overlap=%d\n",overlap);
+		}
+
+		if (preAabbs.size())
+		{
+			int prevOverlap = TestAabbAgainstAabb2((const b3Vector3&)preAabbs[40].m_min, (const b3Vector3&)preAabbs[40].m_max,(const b3Vector3&)preAabbs[53].m_min,(const b3Vector3&)preAabbs[53].m_max);
+			printf("prevoverlap=%d\n",prevOverlap);
+		} else
+		{
+			printf("unknown prevoverlap\n");
+		}
+
+	}
+	}
+
+
+	if (0)
+	{
+		for (int i=0;i<m_allAabbsCPU.size();i++)
+		{
+			//printf("aabb[%d] min=%f,%f,%f max=%f,%f,%f\n",i,m_allAabbsCPU[i].m_min[0],m_allAabbsCPU[i].m_min[1],m_allAabbsCPU[i].m_min[2],			m_allAabbsCPU[i].m_max[0],m_allAabbsCPU[i].m_max[1],m_allAabbsCPU[i].m_max[2]);
+
+
+		}
+
+		for (int axis=0;axis<3;axis++)
+		{
+			for (int buf=0;buf<2;buf++)
+			{
+				b3Assert(m_sortedAxisCPU[axis][buf].size() == m_allAabbsCPU.size()*2);
+			}
+		}
+	}
+
+
+
+	m_currentBuffer = 1-m_currentBuffer;
+	
+
+
+	int totalNumAabbs = m_allAabbsCPU.size();
+
+	{
+		B3_PROFILE("assign m_sortedAxisCPU(FloatFlip)");
+		for (int i=0;i<totalNumAabbs;i++)
+		{
+		
+
+			unsigned int keyMin[3];
+			unsigned int keyMax[3];
+			for (int axis=0;axis<3;axis++)
+			{
+				float vmin=m_allAabbsCPU[i].m_min[axis];
+				float vmax = m_allAabbsCPU[i].m_max[axis];
+				keyMin[axis] = FloatFlip(vmin);
+				keyMax[axis] = FloatFlip(vmax);
+			
+				m_sortedAxisCPU[axis][m_currentBuffer][i*2].m_key = keyMin[axis]-1;
+				m_sortedAxisCPU[axis][m_currentBuffer][i*2].m_value = i*2;
+				m_sortedAxisCPU[axis][m_currentBuffer][i*2+1].m_key = keyMax[axis]+1;
+				m_sortedAxisCPU[axis][m_currentBuffer][i*2+1].m_value = i*2+1;
+			}
+			//printf("aabb[%d] min=%u,%u,%u max %u,%u,%u\n", i,keyMin[0],keyMin[1],keyMin[2],keyMax[0],keyMax[1],keyMax[2]);
+
+		}
+	}
+	
+	
+
+	{
+		B3_PROFILE("sort m_sortedAxisCPU");
+		for (int axis=0;axis<3;axis++)
+			m_sorter->executeHost(m_sortedAxisCPU[axis][m_currentBuffer]);
+	}
+
+#if 0
+	if (0)
+	{
+		for (int axis=0;axis<3;axis++)
+		{
+			//printf("axis %d\n",axis);
+			for (int i=0;i<m_sortedAxisCPU[axis][m_currentBuffer].size();i++)
+			{
+				//int key = m_sortedAxisCPU[axis][m_currentBuffer][i].m_key;
+				//int value = m_sortedAxisCPU[axis][m_currentBuffer][i].m_value;
+				//printf("[%d]=%d\n",i,value);
+			}
+
+		}
+	}
+#endif
+
+	{
+		B3_PROFILE("assign m_objectMinMaxIndexCPU");
+		for (int axis=0;axis<3;axis++)
+		{
+			int totalNumAabbs = m_allAabbsCPU.size();
+			int numEndPoints = m_sortedAxisCPU[axis][m_currentBuffer].size();
+			m_objectMinMaxIndexCPU[axis][m_currentBuffer].resize(totalNumAabbs);
+			for (int i=0;i<numEndPoints;i++)
+			{
+				int destIndex = m_sortedAxisCPU[axis][m_currentBuffer][i].m_value;
+				int newDest = destIndex/2;
+				if (destIndex&1)
+				{
+					m_objectMinMaxIndexCPU[axis][m_currentBuffer][newDest].y=i;
+				} else
+				{
+					m_objectMinMaxIndexCPU[axis][m_currentBuffer][newDest].x=i;
+				}
+			}
+		}
+	}
+
+#if 0
+	if (0)
+	{	
+		printf("==========================\n");
+		for (int axis=0;axis<3;axis++)
+		{
+			unsigned int curMinIndex40 = m_objectMinMaxIndexCPU[axis][m_currentBuffer][40].x;
+			unsigned int curMaxIndex40 = m_objectMinMaxIndexCPU[axis][m_currentBuffer][40].y;
+			unsigned int prevMaxIndex40 = m_objectMinMaxIndexCPU[axis][1-m_currentBuffer][40].y;
+			unsigned int prevMinIndex40 = m_objectMinMaxIndexCPU[axis][1-m_currentBuffer][40].x;
+
+			int dmin40 = curMinIndex40 - prevMinIndex40;
+			int dmax40 = curMinIndex40 - prevMinIndex40;
+			printf("axis %d curMinIndex40=%d prevMinIndex40=%d\n",axis,curMinIndex40, prevMinIndex40);
+			printf("axis %d curMaxIndex40=%d prevMaxIndex40=%d\n",axis,curMaxIndex40, prevMaxIndex40);
+		}
+		printf(".........................\n");
+		for (int axis=0;axis<3;axis++)
+		{
+			unsigned int curMinIndex53 = m_objectMinMaxIndexCPU[axis][m_currentBuffer][53].x;
+			unsigned int curMaxIndex53 = m_objectMinMaxIndexCPU[axis][m_currentBuffer][53].y;
+			unsigned int prevMaxIndex53 = m_objectMinMaxIndexCPU[axis][1-m_currentBuffer][53].y;
+			unsigned int prevMinIndex53 = m_objectMinMaxIndexCPU[axis][1-m_currentBuffer][53].x;
+
+			int dmin40 = curMinIndex53 - prevMinIndex53;
+			int dmax40 = curMinIndex53 - prevMinIndex53;
+			printf("axis %d curMinIndex53=%d prevMinIndex53=%d\n",axis,curMinIndex53, prevMinIndex53);
+			printf("axis %d curMaxIndex53=%d prevMaxIndex53=%d\n",axis,curMaxIndex53, prevMaxIndex53);
+		}
+
+	}
+#endif
+
+
+	int a = m_objectMinMaxIndexCPU[0][m_currentBuffer].size();
+	int b = m_objectMinMaxIndexCPU[1][m_currentBuffer].size();
+	int c = m_objectMinMaxIndexCPU[2][m_currentBuffer].size();
+	b3Assert(a==b);
+	b3Assert(b==c);
+	/*
+	if (searchIncremental3dSapOnGpu)
+	{
+		B3_PROFILE("computePairsIncremental3dSapKernelGPU");
+		int numObjects = m_objectMinMaxIndexCPU[0][m_currentBuffer].size();
+		int maxCapacity = 1024*1024;
+		{
+			B3_PROFILE("copy from host");
+			m_objectMinMaxIndexGPUaxis0.copyFromHost(m_objectMinMaxIndexCPU[0][m_currentBuffer]);
+			m_objectMinMaxIndexGPUaxis1.copyFromHost(m_objectMinMaxIndexCPU[1][m_currentBuffer]);
+			m_objectMinMaxIndexGPUaxis2.copyFromHost(m_objectMinMaxIndexCPU[2][m_currentBuffer]);
+			m_objectMinMaxIndexGPUaxis0prev.copyFromHost(m_objectMinMaxIndexCPU[0][1-m_currentBuffer]);
+			m_objectMinMaxIndexGPUaxis1prev.copyFromHost(m_objectMinMaxIndexCPU[1][1-m_currentBuffer]);
+			m_objectMinMaxIndexGPUaxis2prev.copyFromHost(m_objectMinMaxIndexCPU[2][1-m_currentBuffer]);
+
+			m_sortedAxisGPU0.copyFromHost(m_sortedAxisCPU[0][m_currentBuffer]);
+			m_sortedAxisGPU1.copyFromHost(m_sortedAxisCPU[1][m_currentBuffer]);
+			m_sortedAxisGPU2.copyFromHost(m_sortedAxisCPU[2][m_currentBuffer]);
+			m_sortedAxisGPU0prev.copyFromHost(m_sortedAxisCPU[0][1-m_currentBuffer]);
+			m_sortedAxisGPU1prev.copyFromHost(m_sortedAxisCPU[1][1-m_currentBuffer]);
+			m_sortedAxisGPU2prev.copyFromHost(m_sortedAxisCPU[2][1-m_currentBuffer]);
+
+		
+			m_addedHostPairsGPU.resize(maxCapacity);
+			m_removedHostPairsGPU.resize(maxCapacity);
+
+			m_addedCountGPU.resize(0);
+			m_addedCountGPU.push_back(0);
+			m_removedCountGPU.resize(0);
+			m_removedCountGPU.push_back(0);
+		}
+
+		{
+			B3_PROFILE("launch1D");
+			b3LauncherCL launcher(m_queue,  m_computePairsIncremental3dSapKernel,"m_computePairsIncremental3dSapKernel");
+			launcher.setBuffer(m_objectMinMaxIndexGPUaxis0.getBufferCL());
+			launcher.setBuffer(m_objectMinMaxIndexGPUaxis1.getBufferCL());
+			launcher.setBuffer(m_objectMinMaxIndexGPUaxis2.getBufferCL());
+			launcher.setBuffer(m_objectMinMaxIndexGPUaxis0prev.getBufferCL());
+			launcher.setBuffer(m_objectMinMaxIndexGPUaxis1prev.getBufferCL());
+			launcher.setBuffer(m_objectMinMaxIndexGPUaxis2prev.getBufferCL());
+
+			launcher.setBuffer(m_sortedAxisGPU0.getBufferCL());
+			launcher.setBuffer(m_sortedAxisGPU1.getBufferCL());
+			launcher.setBuffer(m_sortedAxisGPU2.getBufferCL());
+			launcher.setBuffer(m_sortedAxisGPU0prev.getBufferCL());
+			launcher.setBuffer(m_sortedAxisGPU1prev.getBufferCL());
+			launcher.setBuffer(m_sortedAxisGPU2prev.getBufferCL());
+
+		
+			launcher.setBuffer(m_addedHostPairsGPU.getBufferCL());
+			launcher.setBuffer(m_removedHostPairsGPU.getBufferCL());
+			launcher.setBuffer(m_addedCountGPU.getBufferCL());
+			launcher.setBuffer(m_removedCountGPU.getBufferCL());
+			launcher.setConst(maxCapacity);
+			launcher.setConst( numObjects);
+			launcher.launch1D( numObjects);
+			clFinish(m_queue);
+		}
+
+		{
+			B3_PROFILE("copy to host");
+			int addedCountGPU = m_addedCountGPU.at(0);
+			m_addedHostPairsGPU.resize(addedCountGPU);
+			m_addedHostPairsGPU.copyToHost(addedHostPairs);
+
+			//printf("addedCountGPU=%d\n",addedCountGPU);
+			int removedCountGPU = m_removedCountGPU.at(0);
+			m_removedHostPairsGPU.resize(removedCountGPU);
+			m_removedHostPairsGPU.copyToHost(removedHostPairs);
+			//printf("removedCountGPU=%d\n",removedCountGPU);
+
+		}
+
+
+
+	} 
+	else
+	*/
+	{
+		int numObjects = m_objectMinMaxIndexCPU[0][m_currentBuffer].size();
+
+		B3_PROFILE("actual search");
+		for (int i=0;i<numObjects;i++)
+		{
+			//int numObjects = m_objectMinMaxIndexCPU[axis][m_currentBuffer].size();
+			//int checkObjects[]={40,53};
+			//int numCheckObjects = sizeof(checkObjects)/sizeof(int);
+			
+			//for (int a=0;a<numCheckObjects ;a++)
+			
+			for (int axis=0;axis<3;axis++)
+			{
+				//int i = checkObjects[a];
+
+				unsigned int curMinIndex = m_objectMinMaxIndexCPU[axis][m_currentBuffer][i].x;
+				unsigned int curMaxIndex = m_objectMinMaxIndexCPU[axis][m_currentBuffer][i].y;
+				unsigned int prevMinIndex = m_objectMinMaxIndexCPU[axis][1-m_currentBuffer][i].x;
+				int dmin = curMinIndex - prevMinIndex;
+				
+				unsigned int prevMaxIndex = m_objectMinMaxIndexCPU[axis][1-m_currentBuffer][i].y;
+
+				
+
+				int dmax = curMaxIndex - prevMaxIndex;
+				if (dmin!=0)
+				{
+					//printf("for object %d, dmin=%d\n",i,dmin);
+				}
+				if (dmax!=0)
+				{
+					//printf("for object %d, dmax=%d\n",i,dmax);
+				}
+				for (int otherbuffer = 0;otherbuffer<2;otherbuffer++)
+				{
+					if (dmin!=0)
+					{
+						int stepMin = dmin<0 ? -1 : 1;
+						for (int j=prevMinIndex;j!=curMinIndex;j+=stepMin)
+						{
+							int otherIndex2 = m_sortedAxisCPU[axis][otherbuffer][j].y;
+							int otherIndex = otherIndex2/2;
+							if (otherIndex!=i)
+							{
+								bool otherIsMax = ((otherIndex2&1)!=0);
+
+								if (otherIsMax)
+								{
+									//bool overlap = TestAabbAgainstAabb2((const b3Vector3&)m_allAabbsCPU[i].m_min, (const b3Vector3&)m_allAabbsCPU[i].m_max,(const b3Vector3&)m_allAabbsCPU[otherIndex].m_min,(const b3Vector3&)m_allAabbsCPU[otherIndex].m_max);
+									//bool prevOverlap = TestAabbAgainstAabb2((const b3Vector3&)preAabbs[i].m_min, (const b3Vector3&)preAabbs[i].m_max,(const b3Vector3&)preAabbs[otherIndex].m_min,(const b3Vector3&)preAabbs[otherIndex].m_max);
+									
+									bool overlap = true;
+
+									for (int ax=0;ax<3;ax++)
+									{
+										if ((m_objectMinMaxIndexCPU[ax][m_currentBuffer][i].x > m_objectMinMaxIndexCPU[ax][m_currentBuffer][otherIndex].y) ||
+											(m_objectMinMaxIndexCPU[ax][m_currentBuffer][i].y < m_objectMinMaxIndexCPU[ax][m_currentBuffer][otherIndex].x))
+											overlap=false;
+									}
+
+								//	b3Assert(overlap2==overlap);
+
+									bool prevOverlap = true;
+
+									for (int ax=0;ax<3;ax++)
+									{
+										if ((m_objectMinMaxIndexCPU[ax][1-m_currentBuffer][i].x > m_objectMinMaxIndexCPU[ax][1-m_currentBuffer][otherIndex].y) ||
+											(m_objectMinMaxIndexCPU[ax][1-m_currentBuffer][i].y < m_objectMinMaxIndexCPU[ax][1-m_currentBuffer][otherIndex].x))
+											prevOverlap=false;
+									}
+									
+
+									//b3Assert(overlap==overlap2);
+								
+
+
+									if (dmin<0)
+									{
+										if (overlap && !prevOverlap)
+										{
+											//add a pair
+											b3Int4 newPair;
+											if (i<=otherIndex)
+											{
+												newPair.x = i;
+												newPair.y = otherIndex;
+											} else
+											{
+												newPair.x = otherIndex;
+												newPair.y = i;
+											}
+											addedHostPairs.push_back(newPair);
+										}
+									} 
+									else
+									{
+										if (!overlap && prevOverlap)
+										{
+									
+											//remove a pair
+											b3Int4 removedPair;
+											if (i<=otherIndex)
+											{
+												removedPair.x = i;
+												removedPair.y = otherIndex;
+											} else
+											{
+												removedPair.x = otherIndex;
+												removedPair.y = i;
+											}
+											removedHostPairs.push_back(removedPair);
+										}
+									}//otherisMax
+								}//if (dmin<0)
+							}//if (otherIndex!=i)
+						}//for (int j=
+					}
+				
+					if (dmax!=0)
+					{
+						int stepMax = dmax<0 ? -1 : 1;
+						for (int j=prevMaxIndex;j!=curMaxIndex;j+=stepMax)
+						{
+							int otherIndex2 = m_sortedAxisCPU[axis][otherbuffer][j].y;
+							int otherIndex = otherIndex2/2;
+							if (otherIndex!=i)
+							{
+								//bool otherIsMin = ((otherIndex2&1)==0);
+								//if (otherIsMin)
+								{
+									//bool overlap = TestAabbAgainstAabb2((const b3Vector3&)m_allAabbsCPU[i].m_min, (const b3Vector3&)m_allAabbsCPU[i].m_max,(const b3Vector3&)m_allAabbsCPU[otherIndex].m_min,(const b3Vector3&)m_allAabbsCPU[otherIndex].m_max);
+									//bool prevOverlap = TestAabbAgainstAabb2((const b3Vector3&)preAabbs[i].m_min, (const b3Vector3&)preAabbs[i].m_max,(const b3Vector3&)preAabbs[otherIndex].m_min,(const b3Vector3&)preAabbs[otherIndex].m_max);
+									
+									bool overlap = true;
+
+									for (int ax=0;ax<3;ax++)
+									{
+										if ((m_objectMinMaxIndexCPU[ax][m_currentBuffer][i].x > m_objectMinMaxIndexCPU[ax][m_currentBuffer][otherIndex].y) ||
+											(m_objectMinMaxIndexCPU[ax][m_currentBuffer][i].y < m_objectMinMaxIndexCPU[ax][m_currentBuffer][otherIndex].x))
+											overlap=false;
+									}
+									//b3Assert(overlap2==overlap);
+
+									bool prevOverlap = true;
+
+									for (int ax=0;ax<3;ax++)
+									{
+										if ((m_objectMinMaxIndexCPU[ax][1-m_currentBuffer][i].x > m_objectMinMaxIndexCPU[ax][1-m_currentBuffer][otherIndex].y) ||
+											(m_objectMinMaxIndexCPU[ax][1-m_currentBuffer][i].y < m_objectMinMaxIndexCPU[ax][1-m_currentBuffer][otherIndex].x))
+											prevOverlap=false;
+									}
+									
+
+									if (dmax>0)
+									{
+										if (overlap && !prevOverlap)
+										{
+											//add a pair
+											b3Int4 newPair;
+											if (i<=otherIndex)
+											{
+												newPair.x = i;
+												newPair.y = otherIndex;
+											} else
+											{
+												newPair.x = otherIndex;
+												newPair.y = i;
+											}
+											addedHostPairs.push_back(newPair);
+							
+										}
+									} 
+									else
+									{
+										if (!overlap && prevOverlap)
+										{
+											//if (otherIndex2&1==0) -> min?
+											//remove a pair
+											b3Int4 removedPair;
+											if (i<=otherIndex)
+											{
+												removedPair.x = i;
+												removedPair.y = otherIndex;
+											} else
+											{
+												removedPair.x = otherIndex;
+												removedPair.y = i;
+											}
+											removedHostPairs.push_back(removedPair);
+								
+										}
+									}
+							
+								}//if (dmin<0)
+							}//if (otherIndex!=i)
+						}//for (int j=
+					}
+				}//for (int otherbuffer
+			}//for (int axis=0;
+		}//for (int i=0;i<numObjects
+	}
+
+	//remove duplicates and add/remove then to existing m_overlappingPairs
+	
+	
+	
+	{
+		{
+			B3_PROFILE("sort allPairs");
+			allPairs.quickSort(b3PairCmp);
+		}
+		{
+			B3_PROFILE("sort addedHostPairs");
+			addedHostPairs.quickSort(b3PairCmp);
+		}
+		{
+			B3_PROFILE("sort removedHostPairs");
+			removedHostPairs.quickSort(b3PairCmp);
+		}
+	}
+
+	b3Int4 prevPair;
+	prevPair.x = -1;
+	prevPair.y = -1;
+	
+	int uniqueRemovedPairs = 0;
+
+	b3AlignedObjectArray<int> removedPositions;
+
+	{
+		B3_PROFILE("actual removing");
+		for (int i=0;i<removedHostPairs.size();i++)
+		{
+			b3Int4 removedPair = removedHostPairs[i];
+			if ((removedPair.x != prevPair.x) || (removedPair.y != prevPair.y))
+			{
+
+			int index1 = allPairs.findBinarySearch(removedPair);
+
+	//#ifdef _DEBUG
+				
+				
+	
+				int index2 = allPairs.findLinearSearch(removedPair);
+				b3Assert(index1==index2);
+	
+				//b3Assert(index1!=allPairs.size());
+				if (index1<allPairs.size())
+	//#endif//_DEBUG
+				{
+					uniqueRemovedPairs++;
+					removedPositions.push_back(index1);
+					{
+						//printf("framepje(%d) remove pair(%d):%d,%d\n",framepje,i,removedPair.x,removedPair.y);
+					}
+				}
+			}
+			prevPair = removedPair;
+		}
+
+		if (uniqueRemovedPairs)
+		{
+			for (int i=0;i<removedPositions.size();i++)
+			{
+				allPairs[removedPositions[i]].x = INT_MAX ;
+				allPairs[removedPositions[i]].y = INT_MAX ;
+			}
+			allPairs.quickSort(b3PairCmp);
+			allPairs.resize(allPairs.size()-uniqueRemovedPairs);
+		}
+	}
+	//if (uniqueRemovedPairs)
+	//	printf("uniqueRemovedPairs=%d\n",uniqueRemovedPairs);
+	//printf("removedHostPairs.size = %d\n",removedHostPairs.size());
+
+	prevPair.x = -1;
+	prevPair.y = -1;
+	
+	int uniqueAddedPairs=0;
+	b3AlignedObjectArray<b3Int4> actualAddedPairs;
+
+	{
+		B3_PROFILE("actual adding");
+		for (int i=0;i<addedHostPairs.size();i++)
+		{
+			b3Int4 newPair = addedHostPairs[i];
+			if ((newPair.x != prevPair.x) || (newPair.y != prevPair.y))
+			{
+//#ifdef _DEBUG		
+				int index1 = allPairs.findBinarySearch(newPair);
+				
+	
+				int index2 = allPairs.findLinearSearch(newPair);
+				b3Assert(index1==index2);
+	
+
+				b3Assert(index1==allPairs.size());
+				if (index1!=allPairs.size())
+				{
+					printf("??\n");
+				}
+
+				if (index1==allPairs.size())
+//#endif //_DEBUG
+				{
+					uniqueAddedPairs++;
+					actualAddedPairs.push_back(newPair);
+				}
+			}
+			prevPair = newPair;
+		}
+		for (int i=0;i<actualAddedPairs.size();i++)
+		{
+			//printf("framepje (%d), new pair(%d):%d,%d\n",framepje,i,actualAddedPairs[i].x,actualAddedPairs[i].y);
+			allPairs.push_back(actualAddedPairs[i]);
+		}
+	}
+	
+	//if (uniqueAddedPairs)
+	//	printf("uniqueAddedPairs=%d\n", uniqueAddedPairs);
+
+
+	{
+		B3_PROFILE("m_overlappingPairs.copyFromHost");
+		m_overlappingPairs.copyFromHost(allPairs);
+	}
+
+	
+}
+
+
+
+
+void  b3GpuSapBroadphase::calculateOverlappingPairsHost(int maxPairs)
+{
+	//test
+//	if (m_currentBuffer>=0)
+	//	return calculateOverlappingPairsHostIncremental3Sap();
+
+	b3Assert(m_allAabbsCPU.size() == m_allAabbsGPU.size());
+	m_allAabbsGPU.copyToHost(m_allAabbsCPU);
+
+
+	
+	int axis=0;
+	{
+		B3_PROFILE("CPU compute best variance axis");
+		b3Vector3 s=b3MakeVector3(0,0,0),s2=b3MakeVector3(0,0,0);
+		int numRigidBodies = m_smallAabbsMappingCPU.size();
+
+		for(int i=0;i<numRigidBodies;i++) 
+		{
+			b3SapAabb aabb = this->m_allAabbsCPU[m_smallAabbsMappingCPU[i]];
+
+			b3Vector3 maxAabb=b3MakeVector3(aabb.m_max[0],aabb.m_max[1],aabb.m_max[2]);
+			b3Vector3 minAabb=b3MakeVector3(aabb.m_min[0],aabb.m_min[1],aabb.m_min[2]);
+			b3Vector3 centerAabb=(maxAabb+minAabb)*0.5f;
+						
+			s += centerAabb;
+			s2 += centerAabb*centerAabb;
+		}
+		b3Vector3 v = s2 - (s*s) / (float)numRigidBodies;
+		
+		if(v[1] > v[0]) 
+			axis = 1;
+		if(v[2] > v[axis]) 
+			axis = 2;
+	}
+
+
+	
+
+	b3AlignedObjectArray<b3Int4> hostPairs;
+
+	{
+		int numSmallAabbs = m_smallAabbsMappingCPU.size();
+		for (int i=0;i<numSmallAabbs;i++)
+		{
+			b3SapAabb smallAabbi = m_allAabbsCPU[m_smallAabbsMappingCPU[i]];
+			//float reference = smallAabbi.m_max[axis];
+
+			for (int j=i+1;j<numSmallAabbs;j++)
+			{
+
+				b3SapAabb smallAabbj = m_allAabbsCPU[m_smallAabbsMappingCPU[j]];
+
+				if (TestAabbAgainstAabb2((b3Vector3&)smallAabbi.m_min, (b3Vector3&)smallAabbi.m_max,
+					(b3Vector3&)smallAabbj.m_min,(b3Vector3&)smallAabbj.m_max))
+				{
+					b3Int4 pair;
+					int a = smallAabbi.m_minIndices[3];
+					int b = smallAabbj.m_minIndices[3];
+					if (a<=b)
+					{
+						pair.x = a;//store the original index in the unsorted aabb array
+						pair.y = b;
+					} else
+					{
+						pair.x = b;//store the original index in the unsorted aabb array
+						pair.y = a;
+					}
+					hostPairs.push_back(pair);
+				}
+			}
+		}
+	}
+
+	
+	{
+		int numSmallAabbs = m_smallAabbsMappingCPU.size();
+		for (int i=0;i<numSmallAabbs;i++)
+		{
+			b3SapAabb smallAabbi = m_allAabbsCPU[m_smallAabbsMappingCPU[i]];
+
+			//float reference = smallAabbi.m_max[axis];
+			int numLargeAabbs = m_largeAabbsMappingCPU.size();
+
+			for (int j=0;j<numLargeAabbs;j++)
+			{
+				b3SapAabb largeAabbj = m_allAabbsCPU[m_largeAabbsMappingCPU[j]];
+				if (TestAabbAgainstAabb2((b3Vector3&)smallAabbi.m_min, (b3Vector3&)smallAabbi.m_max,
+					(b3Vector3&)largeAabbj.m_min,(b3Vector3&)largeAabbj.m_max))
+				{
+					b3Int4 pair;
+					int a = largeAabbj.m_minIndices[3];
+					int b = smallAabbi.m_minIndices[3];
+					if (a<=b)
+					{
+						pair.x = a; 
+						pair.y = b;//store the original index in the unsorted aabb array
+					} else
+					{
+						pair.x = b;
+						pair.y = a;//store the original index in the unsorted aabb array
+					}
+					
+					hostPairs.push_back(pair);
+				}
+			}
+		}
+	}
+
+	if (hostPairs.size() > maxPairs)
+	{
+		hostPairs.resize(maxPairs);
+	}
+
+	if (hostPairs.size())
+	{
+		m_overlappingPairs.copyFromHost(hostPairs);
+	} else
+	{
+		m_overlappingPairs.resize(0);
+	}
+
+	//init3dSap();
+
+}
+
+void  b3GpuSapBroadphase::reset()
+{
+	m_allAabbsGPU.resize(0);
+	m_allAabbsCPU.resize(0);
+
+
+	m_smallAabbsMappingGPU.resize(0);
+	m_smallAabbsMappingCPU.resize(0);
+
+	m_pairCount.resize(0);
+
+	m_largeAabbsMappingGPU.resize(0);
+	m_largeAabbsMappingCPU.resize(0);
+
+}
+
+
+void  b3GpuSapBroadphase::calculateOverlappingPairs(int maxPairs)
+{
+	if (m_sapKernel==0)
+	{
+		calculateOverlappingPairsHost(maxPairs);
+		return;
+	}
+
+	//if (m_currentBuffer>=0)
+	//	return calculateOverlappingPairsHostIncremental3Sap();
+
+	//calculateOverlappingPairsHost(maxPairs);
+
+	B3_PROFILE("GPU 1-axis SAP calculateOverlappingPairs");
+
+	int axis = 0;
+
+	{
+
+	//bool syncOnHost = false;
+
+	int numSmallAabbs = m_smallAabbsMappingCPU.size();
+	if (m_prefixScanFloat4 && numSmallAabbs)
+	{
+		B3_PROFILE("GPU compute best variance axis");
+		
+		if (m_dst.size()!=(numSmallAabbs+1))
+		{
+			m_dst.resize(numSmallAabbs+128);
+			m_sum.resize(numSmallAabbs+128);
+			m_sum2.resize(numSmallAabbs+128);
+			m_sum.at(numSmallAabbs)=b3MakeVector3(0,0,0); //slow?
+			m_sum2.at(numSmallAabbs)=b3MakeVector3(0,0,0); //slow?
+		}
+
+		b3LauncherCL launcher(m_queue, m_prepareSumVarianceKernel ,"m_prepareSumVarianceKernel");
+		launcher.setBuffer(m_allAabbsGPU.getBufferCL());
+		
+		launcher.setBuffer(m_smallAabbsMappingGPU.getBufferCL());
+		launcher.setBuffer(m_sum.getBufferCL());
+		launcher.setBuffer(m_sum2.getBufferCL());
+		launcher.setConst( numSmallAabbs  );
+		int num = numSmallAabbs;
+		launcher.launch1D( num);
+		
+
+		b3Vector3 s;
+		b3Vector3 s2;
+		m_prefixScanFloat4->execute(m_sum,m_dst,numSmallAabbs+1,&s);
+		m_prefixScanFloat4->execute(m_sum2,m_dst,numSmallAabbs+1,&s2);
+
+		b3Vector3 v = s2 - (s*s) / (float)numSmallAabbs;
+		
+		if(v[1] > v[0]) 
+			axis = 1;
+		if(v[2] > v[axis]) 
+			axis = 2;
+	}
+
+
+		
+		m_gpuSmallSortData.resize(numSmallAabbs);
+		
+
+#if 1
+		if (m_smallAabbsMappingGPU.size())
+		{
+			
+			B3_PROFILE("flipFloatKernel");
+			b3BufferInfoCL bInfo[] = { 
+				b3BufferInfoCL( m_allAabbsGPU.getBufferCL(), true ), 
+				b3BufferInfoCL( m_smallAabbsMappingGPU.getBufferCL(), true),
+				b3BufferInfoCL( m_gpuSmallSortData.getBufferCL())};
+			b3LauncherCL launcher(m_queue, m_flipFloatKernel ,"m_flipFloatKernel");
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst( numSmallAabbs  );
+			launcher.setConst( axis  );
+			
+			int num = numSmallAabbs;
+			launcher.launch1D( num);
+			clFinish(m_queue);
+		}
+
+		if (m_gpuSmallSortData.size())
+		{
+			B3_PROFILE("gpu radix sort");
+			m_sorter->execute(m_gpuSmallSortData);
+			clFinish(m_queue);
+		}
+
+		m_gpuSmallSortedAabbs.resize(numSmallAabbs);
+		if (numSmallAabbs)
+		{
+			B3_PROFILE("scatterKernel");
+			
+			b3BufferInfoCL bInfo[] = { 
+				b3BufferInfoCL( m_allAabbsGPU.getBufferCL(), true ),
+				b3BufferInfoCL( m_smallAabbsMappingGPU.getBufferCL(), true),
+				b3BufferInfoCL( m_gpuSmallSortData.getBufferCL(),true),
+				b3BufferInfoCL(m_gpuSmallSortedAabbs.getBufferCL())};
+			b3LauncherCL launcher(m_queue, m_scatterKernel ,"m_scatterKernel ");
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst( numSmallAabbs);
+			int num = numSmallAabbs;
+			launcher.launch1D( num);
+			clFinish(m_queue);
+			
+		}
+        
+
+			m_overlappingPairs.resize(maxPairs);
+
+			m_pairCount.resize(0);
+			m_pairCount.push_back(0);
+            int numPairs=0;
+
+			{
+				int numLargeAabbs = m_largeAabbsMappingGPU.size();
+				if (numLargeAabbs && numSmallAabbs)
+				{
+					//@todo
+					B3_PROFILE("sap2Kernel");
+					b3BufferInfoCL bInfo[] = { 
+						b3BufferInfoCL( m_allAabbsGPU.getBufferCL() ),
+						b3BufferInfoCL( m_largeAabbsMappingGPU.getBufferCL() ),
+						b3BufferInfoCL( m_smallAabbsMappingGPU.getBufferCL() ), 
+						b3BufferInfoCL( m_overlappingPairs.getBufferCL() ), 
+						b3BufferInfoCL(m_pairCount.getBufferCL())};
+					b3LauncherCL launcher(m_queue, m_sap2Kernel,"m_sap2Kernel");
+					launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+					launcher.setConst(   numLargeAabbs  );
+					launcher.setConst( numSmallAabbs);
+					launcher.setConst( axis  );
+					launcher.setConst( maxPairs  );
+//@todo: use actual maximum work item sizes of the device instead of hardcoded values
+					launcher.launch2D( numLargeAabbs, numSmallAabbs,4,64);
+                
+					numPairs = m_pairCount.at(0);
+					if (numPairs >maxPairs)
+					{
+						b3Error("Error running out of pairs: numPairs = %d, maxPairs = %d.\n", numPairs, maxPairs);
+						numPairs =maxPairs;
+					}
+				}
+			}
+			if (m_gpuSmallSortedAabbs.size())
+			{
+				B3_PROFILE("sapKernel");
+				b3BufferInfoCL bInfo[] = { b3BufferInfoCL( m_gpuSmallSortedAabbs.getBufferCL() ), b3BufferInfoCL( m_overlappingPairs.getBufferCL() ), b3BufferInfoCL(m_pairCount.getBufferCL())};
+				b3LauncherCL launcher(m_queue, m_sapKernel,"m_sapKernel");
+				launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+				launcher.setConst( numSmallAabbs  );
+				launcher.setConst( axis  );
+				launcher.setConst( maxPairs  );
+
+			
+				int num = numSmallAabbs;
+#if 0                
+                int buffSize = launcher.getSerializationBufferSize();
+                unsigned char* buf = new unsigned char[buffSize+sizeof(int)];
+                for (int i=0;i<buffSize+1;i++)
+                {
+                    unsigned char* ptr = (unsigned char*)&buf[i];
+                    *ptr = 0xff;
+                }
+                int actualWrite = launcher.serializeArguments(buf,buffSize);
+                
+                unsigned char* cptr = (unsigned char*)&buf[buffSize];
+    //            printf("buf[buffSize] = %d\n",*cptr);
+                
+                assert(buf[buffSize]==0xff);//check for buffer overrun
+                int* ptr = (int*)&buf[buffSize];
+                
+                *ptr = num;
+                
+                FILE* f = fopen("m_sapKernelArgs.bin","wb");
+                fwrite(buf,buffSize+sizeof(int),1,f);
+                fclose(f);
+#endif//
+
+                launcher.launch1D( num);
+				clFinish(m_queue);
+                
+                numPairs = m_pairCount.at(0);
+                if (numPairs>maxPairs)
+				{
+					b3Error("Error running out of pairs: numPairs = %d, maxPairs = %d.\n", numPairs, maxPairs);
+					numPairs = maxPairs;
+					m_pairCount.resize(0);
+					m_pairCount.push_back(maxPairs);
+				}
+			}
+			
+#else
+        int numPairs = 0;
+        
+        
+        b3LauncherCL launcher(m_queue, m_sapKernel);
+
+        const char* fileName = "m_sapKernelArgs.bin";
+        FILE* f = fopen(fileName,"rb");
+        if (f)
+        {
+            int sizeInBytes=0;
+            if (fseek(f, 0, SEEK_END) || (sizeInBytes = ftell(f)) == EOF || fseek(f, 0, SEEK_SET)) 
+            {
+                printf("error, cannot get file size\n");
+                exit(0);
+            }
+            
+            unsigned char* buf = (unsigned char*) malloc(sizeInBytes);
+            fread(buf,sizeInBytes,1,f);
+            int serializedBytes = launcher.deserializeArgs(buf, sizeInBytes,m_context);
+            int num = *(int*)&buf[serializedBytes];
+            launcher.launch1D( num);
+            
+            b3OpenCLArray<int> pairCount(m_context, m_queue);
+            int numElements = launcher.m_arrays[2]->size()/sizeof(int);
+            pairCount.setFromOpenCLBuffer(launcher.m_arrays[2]->getBufferCL(),numElements);
+            numPairs = pairCount.at(0);
+            //printf("overlapping pairs = %d\n",numPairs);
+            b3AlignedObjectArray<b3Int4>		hostOoverlappingPairs;
+            b3OpenCLArray<b3Int4> tmpGpuPairs(m_context,m_queue);
+            tmpGpuPairs.setFromOpenCLBuffer(launcher.m_arrays[1]->getBufferCL(),numPairs );
+   
+            tmpGpuPairs.copyToHost(hostOoverlappingPairs);
+            m_overlappingPairs.copyFromHost(hostOoverlappingPairs);
+            //printf("hello %d\n", m_overlappingPairs.size());
+            free(buf);
+            fclose(f);
+            
+        } else {
+            printf("error: cannot find file %s\n",fileName);
+        }
+        
+        clFinish(m_queue);
+
+        
+#endif
+
+			
+        m_overlappingPairs.resize(numPairs);
+		
+	}//B3_PROFILE("GPU_RADIX SORT");
+	//init3dSap();
+}
+
+void b3GpuSapBroadphase::writeAabbsToGpu()
+{
+	m_smallAabbsMappingGPU.copyFromHost(m_smallAabbsMappingCPU);
+	m_largeAabbsMappingGPU.copyFromHost(m_largeAabbsMappingCPU);
+
+	m_allAabbsGPU.copyFromHost(m_allAabbsCPU);//might not be necessary, the 'setupGpuAabbsFull' already takes care of this
+	
+	
+
+}
+
+void b3GpuSapBroadphase::createLargeProxy(const b3Vector3& aabbMin,  const b3Vector3& aabbMax, int userPtr , int collisionFilterGroup, int collisionFilterMask)
+{
+	int index = userPtr;
+	b3SapAabb aabb;
+	for (int i=0;i<4;i++)
+	{
+		aabb.m_min[i] = aabbMin[i];
+		aabb.m_max[i] = aabbMax[i];
+	}
+	aabb.m_minIndices[3] = index;
+	aabb.m_signedMaxIndices[3] = m_allAabbsCPU.size();
+	m_largeAabbsMappingCPU.push_back(m_allAabbsCPU.size());
+	
+	m_allAabbsCPU.push_back(aabb);
+}
+
+void b3GpuSapBroadphase::createProxy(const b3Vector3& aabbMin,  const b3Vector3& aabbMax, int userPtr , int collisionFilterGroup, int collisionFilterMask)
+{
+	int index = userPtr;
+	b3SapAabb aabb;
+	for (int i=0;i<4;i++)
+	{
+		aabb.m_min[i] = aabbMin[i];
+		aabb.m_max[i] = aabbMax[i];
+	}
+	aabb.m_minIndices[3] = index;
+	aabb.m_signedMaxIndices[3] = m_allAabbsCPU.size();
+	m_smallAabbsMappingCPU.push_back(m_allAabbsCPU.size());
+
+
+	m_allAabbsCPU.push_back(aabb);
+}
+
+cl_mem	b3GpuSapBroadphase::getAabbBufferWS()
+{
+	return m_allAabbsGPU.getBufferCL();
+}
+
+int	b3GpuSapBroadphase::getNumOverlap()
+{
+	return m_overlappingPairs.size();
+}
+cl_mem	b3GpuSapBroadphase::getOverlappingPairBuffer()
+{
+	return m_overlappingPairs.getBufferCL();
+}
+
+b3OpenCLArray<b3Int4>& b3GpuSapBroadphase::getOverlappingPairsGPU()
+{
+	return m_overlappingPairs;
+}
+b3OpenCLArray<int>& b3GpuSapBroadphase::getSmallAabbIndicesGPU()
+{
+	return m_smallAabbsMappingGPU;
+}
+b3OpenCLArray<int>& b3GpuSapBroadphase::getLargeAabbIndicesGPU()
+{
+	return m_largeAabbsMappingGPU;
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuSapBroadphase.h b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuSapBroadphase.h
new file mode 100644
index 0000000000..8d36ac78f2
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3GpuSapBroadphase.h
@@ -0,0 +1,151 @@
+#ifndef B3_GPU_SAP_BROADPHASE_H
+#define B3_GPU_SAP_BROADPHASE_H
+
+#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3FillCL.h" //b3Int2
+class b3Vector3;
+#include "Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h"
+
+#include "b3SapAabb.h"
+#include "Bullet3Common/shared/b3Int2.h"
+
+#include "b3GpuBroadphaseInterface.h"
+
+
+class b3GpuSapBroadphase : public b3GpuBroadphaseInterface
+{
+	
+	cl_context				m_context;
+	cl_device_id			m_device;
+	cl_command_queue		m_queue;
+	cl_kernel				m_flipFloatKernel;
+	cl_kernel				m_scatterKernel ;
+	cl_kernel				m_copyAabbsKernel;
+	cl_kernel				m_sapKernel;
+	cl_kernel				m_sap2Kernel;
+	cl_kernel				m_prepareSumVarianceKernel;
+	
+
+	class b3RadixSort32CL* m_sorter;
+
+	///test for 3d SAP
+	b3AlignedObjectArray<b3SortData>		m_sortedAxisCPU[3][2];
+	b3AlignedObjectArray<b3UnsignedInt2>	m_objectMinMaxIndexCPU[3][2];
+	b3OpenCLArray<b3UnsignedInt2>			m_objectMinMaxIndexGPUaxis0;
+	b3OpenCLArray<b3UnsignedInt2>			m_objectMinMaxIndexGPUaxis1;
+	b3OpenCLArray<b3UnsignedInt2>			m_objectMinMaxIndexGPUaxis2;
+	b3OpenCLArray<b3UnsignedInt2>			m_objectMinMaxIndexGPUaxis0prev;
+	b3OpenCLArray<b3UnsignedInt2>			m_objectMinMaxIndexGPUaxis1prev;
+	b3OpenCLArray<b3UnsignedInt2>			m_objectMinMaxIndexGPUaxis2prev;
+
+	b3OpenCLArray<b3SortData>				m_sortedAxisGPU0;
+	b3OpenCLArray<b3SortData>				m_sortedAxisGPU1;
+	b3OpenCLArray<b3SortData>				m_sortedAxisGPU2;
+	b3OpenCLArray<b3SortData>				m_sortedAxisGPU0prev;
+	b3OpenCLArray<b3SortData>				m_sortedAxisGPU1prev;
+	b3OpenCLArray<b3SortData>				m_sortedAxisGPU2prev;
+
+
+	b3OpenCLArray<b3Int4>					m_addedHostPairsGPU;
+	b3OpenCLArray<b3Int4>					m_removedHostPairsGPU;
+	b3OpenCLArray<int>						m_addedCountGPU;
+	b3OpenCLArray<int>						m_removedCountGPU;
+	
+	int	m_currentBuffer;
+
+public:
+
+	b3OpenCLArray<int> m_pairCount;
+
+
+	b3OpenCLArray<b3SapAabb>	m_allAabbsGPU;
+	b3AlignedObjectArray<b3SapAabb>	m_allAabbsCPU;
+
+	virtual b3OpenCLArray<b3SapAabb>&	getAllAabbsGPU()
+	{
+		return m_allAabbsGPU;
+	}
+	virtual b3AlignedObjectArray<b3SapAabb>&	getAllAabbsCPU()
+	{
+		return m_allAabbsCPU;
+	}
+
+	b3OpenCLArray<b3Vector3>	m_sum;
+	b3OpenCLArray<b3Vector3>	m_sum2;
+	b3OpenCLArray<b3Vector3>	m_dst;
+
+	b3OpenCLArray<int>	m_smallAabbsMappingGPU;
+	b3AlignedObjectArray<int> m_smallAabbsMappingCPU;
+
+	b3OpenCLArray<int>	m_largeAabbsMappingGPU;
+	b3AlignedObjectArray<int> m_largeAabbsMappingCPU;
+
+	
+	b3OpenCLArray<b3Int4>		m_overlappingPairs;
+
+	//temporary gpu work memory
+	b3OpenCLArray<b3SortData>	m_gpuSmallSortData;
+	b3OpenCLArray<b3SapAabb>	m_gpuSmallSortedAabbs;
+
+	class b3PrefixScanFloat4CL*		m_prefixScanFloat4;
+
+	enum b3GpuSapKernelType
+	{
+		B3_GPU_SAP_KERNEL_BRUTE_FORCE_CPU=1,
+		B3_GPU_SAP_KERNEL_BRUTE_FORCE_GPU,
+		B3_GPU_SAP_KERNEL_ORIGINAL,
+		B3_GPU_SAP_KERNEL_BARRIER,
+		B3_GPU_SAP_KERNEL_LOCAL_SHARED_MEMORY
+	};
+
+	b3GpuSapBroadphase(cl_context ctx,cl_device_id device, cl_command_queue  q , b3GpuSapKernelType kernelType=B3_GPU_SAP_KERNEL_LOCAL_SHARED_MEMORY);
+	virtual ~b3GpuSapBroadphase();
+	
+	static b3GpuBroadphaseInterface* CreateFuncBruteForceCpu(cl_context ctx,cl_device_id device, cl_command_queue  q)
+	{
+		return new b3GpuSapBroadphase(ctx,device,q,B3_GPU_SAP_KERNEL_BRUTE_FORCE_CPU);
+	}
+
+	static b3GpuBroadphaseInterface* CreateFuncBruteForceGpu(cl_context ctx,cl_device_id device, cl_command_queue  q)
+	{
+		return new b3GpuSapBroadphase(ctx,device,q,B3_GPU_SAP_KERNEL_BRUTE_FORCE_GPU);
+	}
+
+	static b3GpuBroadphaseInterface* CreateFuncOriginal(cl_context ctx,cl_device_id device, cl_command_queue  q)
+	{
+		return new b3GpuSapBroadphase(ctx,device,q,B3_GPU_SAP_KERNEL_ORIGINAL);
+	}
+	static b3GpuBroadphaseInterface* CreateFuncBarrier(cl_context ctx,cl_device_id device, cl_command_queue  q)
+	{
+		return new b3GpuSapBroadphase(ctx,device,q,B3_GPU_SAP_KERNEL_BARRIER);
+	}
+	static b3GpuBroadphaseInterface* CreateFuncLocalMemory(cl_context ctx,cl_device_id device, cl_command_queue  q)
+	{
+		return new b3GpuSapBroadphase(ctx,device,q,B3_GPU_SAP_KERNEL_LOCAL_SHARED_MEMORY);
+	}
+	
+
+	virtual void  calculateOverlappingPairs(int maxPairs);
+	virtual void  calculateOverlappingPairsHost(int maxPairs);
+	
+	void  reset();
+
+	void init3dSap();
+	virtual void calculateOverlappingPairsHostIncremental3Sap();
+
+	virtual void createProxy(const b3Vector3& aabbMin,  const b3Vector3& aabbMax, int userPtr , int collisionFilterGroup, int collisionFilterMask);
+	virtual void createLargeProxy(const b3Vector3& aabbMin,  const b3Vector3& aabbMax, int userPtr , int collisionFilterGroup, int collisionFilterMask);
+
+	//call writeAabbsToGpu after done making all changes (createProxy etc)
+	virtual void writeAabbsToGpu();
+
+	virtual cl_mem	getAabbBufferWS();
+	virtual int	getNumOverlap();
+	virtual cl_mem	getOverlappingPairBuffer();
+	
+	virtual b3OpenCLArray<b3Int4>& getOverlappingPairsGPU();
+	virtual b3OpenCLArray<int>& getSmallAabbIndicesGPU();
+	virtual b3OpenCLArray<int>& getLargeAabbIndicesGPU();
+};
+
+#endif //B3_GPU_SAP_BROADPHASE_H
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3SapAabb.h b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3SapAabb.h
new file mode 100644
index 0000000000..ea6550fede
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/b3SapAabb.h
@@ -0,0 +1,14 @@
+#ifndef B3_SAP_AABB_H
+#define B3_SAP_AABB_H
+
+#include "Bullet3Common/b3Scalar.h"
+#include "Bullet3Collision/BroadPhaseCollision/shared/b3Aabb.h"
+
+///just make sure that the b3Aabb is 16-byte aligned
+B3_ATTRIBUTE_ALIGNED16(struct) b3SapAabb : public b3Aabb
+{
+
+};
+
+
+#endif //B3_SAP_AABB_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/gridBroadphase.cl b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/gridBroadphase.cl
new file mode 100644
index 0000000000..ded4796d33
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/gridBroadphase.cl
@@ -0,0 +1,216 @@
+
+
+int getPosHash(int4 gridPos, __global float4* pParams)
+{
+	int4 gridDim = *((__global int4*)(pParams + 1));
+	gridPos.x &= gridDim.x - 1;
+	gridPos.y &= gridDim.y - 1;
+	gridPos.z &= gridDim.z - 1;
+	int hash = gridPos.z * gridDim.y * gridDim.x + gridPos.y * gridDim.x + gridPos.x;
+	return hash;
+} 
+
+int4 getGridPos(float4 worldPos, __global float4* pParams)
+{
+    int4 gridPos;
+	int4 gridDim = *((__global int4*)(pParams + 1));
+    gridPos.x = (int)floor(worldPos.x * pParams[0].x) & (gridDim.x - 1);
+    gridPos.y = (int)floor(worldPos.y * pParams[0].y) & (gridDim.y - 1);
+    gridPos.z = (int)floor(worldPos.z * pParams[0].z) & (gridDim.z - 1);
+    return gridPos;
+}
+
+
+// calculate grid hash value for each body using its AABB
+__kernel void kCalcHashAABB(int numObjects, __global float4* allpAABB, __global const int* smallAabbMapping, __global int2* pHash, __global float4* pParams )
+{
+    int index = get_global_id(0);
+    if(index >= numObjects)
+	{
+		return;
+	}
+	float4 bbMin = allpAABB[smallAabbMapping[index]*2];
+	float4 bbMax = allpAABB[smallAabbMapping[index]*2 + 1];
+	float4 pos;
+	pos.x = (bbMin.x + bbMax.x) * 0.5f;
+	pos.y = (bbMin.y + bbMax.y) * 0.5f;
+	pos.z = (bbMin.z + bbMax.z) * 0.5f;
+	pos.w = 0.f;
+    // get address in grid
+    int4 gridPos = getGridPos(pos, pParams);
+    int gridHash = getPosHash(gridPos, pParams);
+    // store grid hash and body index
+    int2 hashVal;
+    hashVal.x = gridHash;
+    hashVal.y = index;
+    pHash[index] = hashVal;
+}
+
+__kernel void kClearCellStart(	int numCells, 
+								__global int* pCellStart )
+{
+    int index = get_global_id(0);
+    if(index >= numCells)
+	{
+		return;
+	}
+	pCellStart[index] = -1;
+}
+
+__kernel void kFindCellStart(int numObjects, __global int2* pHash, __global int* cellStart )
+{
+	__local int sharedHash[513];
+    int index = get_global_id(0);
+	int2 sortedData;
+
+    if(index < numObjects)
+	{
+		sortedData = pHash[index];
+		// Load hash data into shared memory so that we can look 
+		// at neighboring body's hash value without loading
+		// two hash values per thread
+		sharedHash[get_local_id(0) + 1] = sortedData.x;
+		if((index > 0) && (get_local_id(0) == 0))
+		{
+			// first thread in block must load neighbor body hash
+			sharedHash[0] = pHash[index-1].x;
+		}
+	}
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if(index < numObjects)
+	{
+		if((index == 0) || (sortedData.x != sharedHash[get_local_id(0)]))
+		{
+			cellStart[sortedData.x] = index;
+		}
+	}
+}
+
+int testAABBOverlap(float4 min0, float4 max0, float4 min1, float4 max1)
+{
+	return	(min0.x <= max1.x)&& (min1.x <= max0.x) && 
+			(min0.y <= max1.y)&& (min1.y <= max0.y) && 
+			(min0.z <= max1.z)&& (min1.z <= max0.z); 
+}
+
+
+
+
+//search for AABB 'index' against other AABBs' in this cell
+void findPairsInCell(	int numObjects,
+						int4	gridPos,
+						int    index,
+						__global int2*  pHash,
+						__global int*   pCellStart,
+						__global float4* allpAABB, 
+						__global const int* smallAabbMapping,
+						__global float4* pParams,
+							volatile  __global int* pairCount,
+						__global int4*   pPairBuff2,
+						int maxPairs
+						)
+{
+	int4 pGridDim = *((__global int4*)(pParams + 1));
+	int maxBodiesPerCell = pGridDim.w;
+    int gridHash = getPosHash(gridPos, pParams);
+    // get start of bucket for this cell
+    int bucketStart = pCellStart[gridHash];
+    if (bucketStart == -1)
+	{
+        return;   // cell empty
+	}
+	// iterate over bodies in this cell
+    int2 sortedData = pHash[index];
+	int unsorted_indx = sortedData.y;
+    float4 min0 = allpAABB[smallAabbMapping[unsorted_indx]*2 + 0]; 
+	float4 max0 = allpAABB[smallAabbMapping[unsorted_indx]*2 + 1];
+	int handleIndex =  as_int(min0.w);
+	
+	int bucketEnd = bucketStart + maxBodiesPerCell;
+	bucketEnd = (bucketEnd > numObjects) ? numObjects : bucketEnd;
+	for(int index2 = bucketStart; index2 < bucketEnd; index2++) 
+	{
+        int2 cellData = pHash[index2];
+        if (cellData.x != gridHash)
+        {
+			break;   // no longer in same bucket
+		}
+		int unsorted_indx2 = cellData.y;
+        //if (unsorted_indx2 < unsorted_indx) // check not colliding with self
+		if (unsorted_indx2 != unsorted_indx) // check not colliding with self
+        {   
+			float4 min1 = allpAABB[smallAabbMapping[unsorted_indx2]*2 + 0];
+			float4 max1 = allpAABB[smallAabbMapping[unsorted_indx2]*2 + 1];
+			if(testAABBOverlap(min0, max0, min1, max1))
+			{
+				if (pairCount)
+				{
+					int handleIndex2 = as_int(min1.w);
+					if (handleIndex<handleIndex2)
+					{
+						int curPair = atomic_add(pairCount,1);
+						if (curPair<maxPairs)
+						{
+							int4 newpair;
+							newpair.x = handleIndex;
+							newpair.y = handleIndex2;
+							newpair.z = -1;
+							newpair.w = -1;
+							pPairBuff2[curPair] = newpair;
+						}
+					}
+				
+				}
+			}
+		}
+	}
+}
+
+__kernel void kFindOverlappingPairs(	int numObjects,
+										__global float4* allpAABB, 
+										__global const int* smallAabbMapping,
+										__global int2* pHash, 
+										__global int* pCellStart, 
+										__global float4* pParams ,
+										volatile  __global int* pairCount,
+										__global int4*   pPairBuff2,
+										int maxPairs
+										)
+
+{
+    int index = get_global_id(0);
+    if(index >= numObjects)
+	{
+		return;
+	}
+    int2 sortedData = pHash[index];
+	int unsorted_indx = sortedData.y;
+	float4 bbMin = allpAABB[smallAabbMapping[unsorted_indx]*2 + 0];
+	float4 bbMax = allpAABB[smallAabbMapping[unsorted_indx]*2 + 1];
+	float4 pos;
+	pos.x = (bbMin.x + bbMax.x) * 0.5f;
+	pos.y = (bbMin.y + bbMax.y) * 0.5f;
+	pos.z = (bbMin.z + bbMax.z) * 0.5f;
+    // get address in grid
+    int4 gridPosA = getGridPos(pos, pParams);
+    int4 gridPosB; 
+    // examine only neighbouring cells
+    for(int z=-1; z<=1; z++) 
+    {
+		gridPosB.z = gridPosA.z + z;
+        for(int y=-1; y<=1; y++) 
+        {
+			gridPosB.y = gridPosA.y + y;
+            for(int x=-1; x<=1; x++) 
+            {
+				gridPosB.x = gridPosA.x + x;
+                findPairsInCell(numObjects, gridPosB, index, pHash, pCellStart, allpAABB,smallAabbMapping, pParams, pairCount,pPairBuff2, maxPairs);
+            }
+        }
+    }
+}
+
+
+
+
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/gridBroadphaseKernels.h b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/gridBroadphaseKernels.h
new file mode 100644
index 0000000000..dad42477c3
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/gridBroadphaseKernels.h
@@ -0,0 +1,199 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* gridBroadphaseCL= \
+"int getPosHash(int4 gridPos, __global float4* pParams)\n"
+"{\n"
+"	int4 gridDim = *((__global int4*)(pParams + 1));\n"
+"	gridPos.x &= gridDim.x - 1;\n"
+"	gridPos.y &= gridDim.y - 1;\n"
+"	gridPos.z &= gridDim.z - 1;\n"
+"	int hash = gridPos.z * gridDim.y * gridDim.x + gridPos.y * gridDim.x + gridPos.x;\n"
+"	return hash;\n"
+"} \n"
+"int4 getGridPos(float4 worldPos, __global float4* pParams)\n"
+"{\n"
+"    int4 gridPos;\n"
+"	int4 gridDim = *((__global int4*)(pParams + 1));\n"
+"    gridPos.x = (int)floor(worldPos.x * pParams[0].x) & (gridDim.x - 1);\n"
+"    gridPos.y = (int)floor(worldPos.y * pParams[0].y) & (gridDim.y - 1);\n"
+"    gridPos.z = (int)floor(worldPos.z * pParams[0].z) & (gridDim.z - 1);\n"
+"    return gridPos;\n"
+"}\n"
+"// calculate grid hash value for each body using its AABB\n"
+"__kernel void kCalcHashAABB(int numObjects, __global float4* allpAABB, __global const int* smallAabbMapping, __global int2* pHash, __global float4* pParams )\n"
+"{\n"
+"    int index = get_global_id(0);\n"
+"    if(index >= numObjects)\n"
+"	{\n"
+"		return;\n"
+"	}\n"
+"	float4 bbMin = allpAABB[smallAabbMapping[index]*2];\n"
+"	float4 bbMax = allpAABB[smallAabbMapping[index]*2 + 1];\n"
+"	float4 pos;\n"
+"	pos.x = (bbMin.x + bbMax.x) * 0.5f;\n"
+"	pos.y = (bbMin.y + bbMax.y) * 0.5f;\n"
+"	pos.z = (bbMin.z + bbMax.z) * 0.5f;\n"
+"	pos.w = 0.f;\n"
+"    // get address in grid\n"
+"    int4 gridPos = getGridPos(pos, pParams);\n"
+"    int gridHash = getPosHash(gridPos, pParams);\n"
+"    // store grid hash and body index\n"
+"    int2 hashVal;\n"
+"    hashVal.x = gridHash;\n"
+"    hashVal.y = index;\n"
+"    pHash[index] = hashVal;\n"
+"}\n"
+"__kernel void kClearCellStart(	int numCells, \n"
+"								__global int* pCellStart )\n"
+"{\n"
+"    int index = get_global_id(0);\n"
+"    if(index >= numCells)\n"
+"	{\n"
+"		return;\n"
+"	}\n"
+"	pCellStart[index] = -1;\n"
+"}\n"
+"__kernel void kFindCellStart(int numObjects, __global int2* pHash, __global int* cellStart )\n"
+"{\n"
+"	__local int sharedHash[513];\n"
+"    int index = get_global_id(0);\n"
+"	int2 sortedData;\n"
+"    if(index < numObjects)\n"
+"	{\n"
+"		sortedData = pHash[index];\n"
+"		// Load hash data into shared memory so that we can look \n"
+"		// at neighboring body's hash value without loading\n"
+"		// two hash values per thread\n"
+"		sharedHash[get_local_id(0) + 1] = sortedData.x;\n"
+"		if((index > 0) && (get_local_id(0) == 0))\n"
+"		{\n"
+"			// first thread in block must load neighbor body hash\n"
+"			sharedHash[0] = pHash[index-1].x;\n"
+"		}\n"
+"	}\n"
+"    barrier(CLK_LOCAL_MEM_FENCE);\n"
+"    if(index < numObjects)\n"
+"	{\n"
+"		if((index == 0) || (sortedData.x != sharedHash[get_local_id(0)]))\n"
+"		{\n"
+"			cellStart[sortedData.x] = index;\n"
+"		}\n"
+"	}\n"
+"}\n"
+"int testAABBOverlap(float4 min0, float4 max0, float4 min1, float4 max1)\n"
+"{\n"
+"	return	(min0.x <= max1.x)&& (min1.x <= max0.x) && \n"
+"			(min0.y <= max1.y)&& (min1.y <= max0.y) && \n"
+"			(min0.z <= max1.z)&& (min1.z <= max0.z); \n"
+"}\n"
+"//search for AABB 'index' against other AABBs' in this cell\n"
+"void findPairsInCell(	int numObjects,\n"
+"						int4	gridPos,\n"
+"						int    index,\n"
+"						__global int2*  pHash,\n"
+"						__global int*   pCellStart,\n"
+"						__global float4* allpAABB, \n"
+"						__global const int* smallAabbMapping,\n"
+"						__global float4* pParams,\n"
+"							volatile  __global int* pairCount,\n"
+"						__global int4*   pPairBuff2,\n"
+"						int maxPairs\n"
+"						)\n"
+"{\n"
+"	int4 pGridDim = *((__global int4*)(pParams + 1));\n"
+"	int maxBodiesPerCell = pGridDim.w;\n"
+"    int gridHash = getPosHash(gridPos, pParams);\n"
+"    // get start of bucket for this cell\n"
+"    int bucketStart = pCellStart[gridHash];\n"
+"    if (bucketStart == -1)\n"
+"	{\n"
+"        return;   // cell empty\n"
+"	}\n"
+"	// iterate over bodies in this cell\n"
+"    int2 sortedData = pHash[index];\n"
+"	int unsorted_indx = sortedData.y;\n"
+"    float4 min0 = allpAABB[smallAabbMapping[unsorted_indx]*2 + 0]; \n"
+"	float4 max0 = allpAABB[smallAabbMapping[unsorted_indx]*2 + 1];\n"
+"	int handleIndex =  as_int(min0.w);\n"
+"	\n"
+"	int bucketEnd = bucketStart + maxBodiesPerCell;\n"
+"	bucketEnd = (bucketEnd > numObjects) ? numObjects : bucketEnd;\n"
+"	for(int index2 = bucketStart; index2 < bucketEnd; index2++) \n"
+"	{\n"
+"        int2 cellData = pHash[index2];\n"
+"        if (cellData.x != gridHash)\n"
+"        {\n"
+"			break;   // no longer in same bucket\n"
+"		}\n"
+"		int unsorted_indx2 = cellData.y;\n"
+"        //if (unsorted_indx2 < unsorted_indx) // check not colliding with self\n"
+"		if (unsorted_indx2 != unsorted_indx) // check not colliding with self\n"
+"        {   \n"
+"			float4 min1 = allpAABB[smallAabbMapping[unsorted_indx2]*2 + 0];\n"
+"			float4 max1 = allpAABB[smallAabbMapping[unsorted_indx2]*2 + 1];\n"
+"			if(testAABBOverlap(min0, max0, min1, max1))\n"
+"			{\n"
+"				if (pairCount)\n"
+"				{\n"
+"					int handleIndex2 = as_int(min1.w);\n"
+"					if (handleIndex<handleIndex2)\n"
+"					{\n"
+"						int curPair = atomic_add(pairCount,1);\n"
+"						if (curPair<maxPairs)\n"
+"						{\n"
+"							int4 newpair;\n"
+"							newpair.x = handleIndex;\n"
+"							newpair.y = handleIndex2;\n"
+"							newpair.z = -1;\n"
+"							newpair.w = -1;\n"
+"							pPairBuff2[curPair] = newpair;\n"
+"						}\n"
+"					}\n"
+"				\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"}\n"
+"__kernel void kFindOverlappingPairs(	int numObjects,\n"
+"										__global float4* allpAABB, \n"
+"										__global const int* smallAabbMapping,\n"
+"										__global int2* pHash, \n"
+"										__global int* pCellStart, \n"
+"										__global float4* pParams ,\n"
+"										volatile  __global int* pairCount,\n"
+"										__global int4*   pPairBuff2,\n"
+"										int maxPairs\n"
+"										)\n"
+"{\n"
+"    int index = get_global_id(0);\n"
+"    if(index >= numObjects)\n"
+"	{\n"
+"		return;\n"
+"	}\n"
+"    int2 sortedData = pHash[index];\n"
+"	int unsorted_indx = sortedData.y;\n"
+"	float4 bbMin = allpAABB[smallAabbMapping[unsorted_indx]*2 + 0];\n"
+"	float4 bbMax = allpAABB[smallAabbMapping[unsorted_indx]*2 + 1];\n"
+"	float4 pos;\n"
+"	pos.x = (bbMin.x + bbMax.x) * 0.5f;\n"
+"	pos.y = (bbMin.y + bbMax.y) * 0.5f;\n"
+"	pos.z = (bbMin.z + bbMax.z) * 0.5f;\n"
+"    // get address in grid\n"
+"    int4 gridPosA = getGridPos(pos, pParams);\n"
+"    int4 gridPosB; \n"
+"    // examine only neighbouring cells\n"
+"    for(int z=-1; z<=1; z++) \n"
+"    {\n"
+"		gridPosB.z = gridPosA.z + z;\n"
+"        for(int y=-1; y<=1; y++) \n"
+"        {\n"
+"			gridPosB.y = gridPosA.y + y;\n"
+"            for(int x=-1; x<=1; x++) \n"
+"            {\n"
+"				gridPosB.x = gridPosA.x + x;\n"
+"                findPairsInCell(numObjects, gridPosB, index, pHash, pCellStart, allpAABB,smallAabbMapping, pParams, pairCount,pPairBuff2, maxPairs);\n"
+"            }\n"
+"        }\n"
+"    }\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/parallelLinearBvh.cl b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/parallelLinearBvh.cl
new file mode 100644
index 0000000000..c375b9bf37
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/parallelLinearBvh.cl
@@ -0,0 +1,767 @@
+/*
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose,
+including commercial applications, and to alter it and redistribute it freely,
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Initial Author Jackson Lee, 2014
+
+typedef float b3Scalar;
+typedef float4 b3Vector3;
+#define b3Max max
+#define b3Min min
+#define b3Sqrt sqrt
+
+typedef struct
+{
+	unsigned int m_key;
+	unsigned int m_value;
+} SortDataCL;
+
+typedef struct 
+{
+	union
+	{
+		float4	m_min;
+		float   m_minElems[4];
+		int			m_minIndices[4];
+	};
+	union
+	{
+		float4	m_max;
+		float   m_maxElems[4];
+		int			m_maxIndices[4];
+	};
+} b3AabbCL;
+
+
+unsigned int interleaveBits(unsigned int x)
+{
+	//........ ........ ......12 3456789A	//x
+	//....1..2 ..3..4.. 5..6..7. .8..9..A	//x after interleaving bits
+	
+	//......12 3456789A ......12 3456789A	//x ^ (x << 16)
+	//11111111 ........ ........ 11111111	//0x FF 00 00 FF
+	//......12 ........ ........ 3456789A	//x = (x ^ (x << 16)) & 0xFF0000FF;
+	
+	//......12 ........ 3456789A 3456789A	//x ^ (x <<  8)
+	//......11 ........ 1111.... ....1111	//0x 03 00 F0 0F
+	//......12 ........ 3456.... ....789A	//x = (x ^ (x <<  8)) & 0x0300F00F;
+	
+	//..12..12 ....3456 3456.... 789A789A	//x ^ (x <<  4)
+	//......11 ....11.. ..11.... 11....11	//0x 03 0C 30 C3
+	//......12 ....34.. ..56.... 78....9A	//x = (x ^ (x <<  4)) & 0x030C30C3;
+	
+	//....1212 ..3434.. 5656..78 78..9A9A	//x ^ (x <<  2)
+	//....1..1 ..1..1.. 1..1..1. .1..1..1	//0x 09 24 92 49
+	//....1..2 ..3..4.. 5..6..7. .8..9..A	//x = (x ^ (x <<  2)) & 0x09249249;
+	
+	//........ ........ ......11 11111111	//0x000003FF
+	x &= 0x000003FF;		//Clear all bits above bit 10
+	
+	x = (x ^ (x << 16)) & 0xFF0000FF;
+	x = (x ^ (x <<  8)) & 0x0300F00F;
+	x = (x ^ (x <<  4)) & 0x030C30C3;
+	x = (x ^ (x <<  2)) & 0x09249249;
+	
+	return x;
+}
+unsigned int getMortonCode(unsigned int x, unsigned int y, unsigned int z)
+{
+	return interleaveBits(x) << 0 | interleaveBits(y) << 1 | interleaveBits(z) << 2;
+}
+
+__kernel void separateAabbs(__global b3AabbCL* unseparatedAabbs, __global int* aabbIndices, __global b3AabbCL* out_aabbs, int numAabbsToSeparate)
+{
+	int separatedAabbIndex = get_global_id(0);
+	if(separatedAabbIndex >= numAabbsToSeparate) return;
+
+	int unseparatedAabbIndex = aabbIndices[separatedAabbIndex];
+	out_aabbs[separatedAabbIndex] = unseparatedAabbs[unseparatedAabbIndex];
+}
+
+//Should replace with an optimized parallel reduction
+__kernel void findAllNodesMergedAabb(__global b3AabbCL* out_mergedAabb, int numAabbsNeedingMerge)
+{
+	//Each time this kernel is added to the command queue, 
+	//the number of AABBs needing to be merged is halved
+	//
+	//Example with 159 AABBs:
+	//	numRemainingAabbs == 159 / 2 + 159 % 2 == 80
+	//	numMergedAabbs == 159 - 80 == 79
+	//So, indices [0, 78] are merged with [0 + 80, 78 + 80]
+	
+	int numRemainingAabbs = numAabbsNeedingMerge / 2 + numAabbsNeedingMerge % 2;
+	int numMergedAabbs = numAabbsNeedingMerge - numRemainingAabbs;
+	
+	int aabbIndex = get_global_id(0);
+	if(aabbIndex >= numMergedAabbs) return;
+	
+	int otherAabbIndex = aabbIndex + numRemainingAabbs;
+	
+	b3AabbCL aabb = out_mergedAabb[aabbIndex];
+	b3AabbCL otherAabb = out_mergedAabb[otherAabbIndex];
+		
+	b3AabbCL mergedAabb;
+	mergedAabb.m_min = b3Min(aabb.m_min, otherAabb.m_min);
+	mergedAabb.m_max = b3Max(aabb.m_max, otherAabb.m_max);
+	out_mergedAabb[aabbIndex] = mergedAabb;
+}
+
+__kernel void assignMortonCodesAndAabbIndicies(__global b3AabbCL* worldSpaceAabbs, __global b3AabbCL* mergedAabbOfAllNodes, 
+												__global SortDataCL* out_mortonCodesAndAabbIndices, int numAabbs)
+{
+	int leafNodeIndex = get_global_id(0);	//Leaf node index == AABB index
+	if(leafNodeIndex >= numAabbs) return;
+	
+	b3AabbCL mergedAabb = mergedAabbOfAllNodes[0];
+	b3Vector3 gridCenter = (mergedAabb.m_min + mergedAabb.m_max) * 0.5f;
+	b3Vector3 gridCellSize = (mergedAabb.m_max - mergedAabb.m_min) / (float)1024;
+	
+	b3AabbCL aabb = worldSpaceAabbs[leafNodeIndex];
+	b3Vector3 aabbCenter = (aabb.m_min + aabb.m_max) * 0.5f;
+	b3Vector3 aabbCenterRelativeToGrid = aabbCenter - gridCenter;
+	
+	//Quantize into integer coordinates
+	//floor() is needed to prevent the center cell, at (0,0,0) from being twice the size
+	b3Vector3 gridPosition = aabbCenterRelativeToGrid / gridCellSize;
+	
+	int4 discretePosition;
+	discretePosition.x = (int)( (gridPosition.x >= 0.0f) ? gridPosition.x : floor(gridPosition.x) );
+	discretePosition.y = (int)( (gridPosition.y >= 0.0f) ? gridPosition.y : floor(gridPosition.y) );
+	discretePosition.z = (int)( (gridPosition.z >= 0.0f) ? gridPosition.z : floor(gridPosition.z) );
+	
+	//Clamp coordinates into [-512, 511], then convert range from [-512, 511] to [0, 1023]
+	discretePosition = b3Max( -512, b3Min(discretePosition, 511) );
+	discretePosition += 512;
+	
+	//Interleave bits(assign a morton code, also known as a z-curve)
+	unsigned int mortonCode = getMortonCode(discretePosition.x, discretePosition.y, discretePosition.z);
+	
+	//
+	SortDataCL mortonCodeIndexPair;
+	mortonCodeIndexPair.m_key = mortonCode;
+	mortonCodeIndexPair.m_value = leafNodeIndex;
+	
+	out_mortonCodesAndAabbIndices[leafNodeIndex] = mortonCodeIndexPair;
+}
+
+#define B3_PLVBH_TRAVERSE_MAX_STACK_SIZE 128
+
+//The most significant bit(0x80000000) of a int32 is used to distinguish between leaf and internal nodes.
+//If it is set, then the index is for an internal node; otherwise, it is a leaf node. 
+//In both cases, the bit should be cleared to access the actual node index.
+int isLeafNode(int index) { return (index >> 31 == 0); }
+int getIndexWithInternalNodeMarkerRemoved(int index) { return index & (~0x80000000); }
+int getIndexWithInternalNodeMarkerSet(int isLeaf, int index) { return (isLeaf) ? index : (index | 0x80000000); }
+
+//From sap.cl
+#define NEW_PAIR_MARKER -1
+
+bool TestAabbAgainstAabb2(const b3AabbCL* aabb1, const b3AabbCL* aabb2)
+{
+	bool overlap = true;
+	overlap = (aabb1->m_min.x > aabb2->m_max.x || aabb1->m_max.x < aabb2->m_min.x) ? false : overlap;
+	overlap = (aabb1->m_min.z > aabb2->m_max.z || aabb1->m_max.z < aabb2->m_min.z) ? false : overlap;
+	overlap = (aabb1->m_min.y > aabb2->m_max.y || aabb1->m_max.y < aabb2->m_min.y) ? false : overlap;
+	return overlap;
+}
+//From sap.cl
+
+__kernel void plbvhCalculateOverlappingPairs(__global b3AabbCL* rigidAabbs, 
+
+											__global int* rootNodeIndex, 
+											__global int2* internalNodeChildIndices, 
+											__global b3AabbCL* internalNodeAabbs,
+											__global int2* internalNodeLeafIndexRanges,
+											
+											__global SortDataCL* mortonCodesAndAabbIndices,
+											__global int* out_numPairs, __global int4* out_overlappingPairs, 
+											int maxPairs, int numQueryAabbs)
+{
+	//Using get_group_id()/get_local_id() is Faster than get_global_id(0) since
+	//mortonCodesAndAabbIndices[] contains rigid body indices sorted along the z-curve (more spatially coherent)
+	int queryBvhNodeIndex = get_group_id(0) * get_local_size(0) + get_local_id(0);
+	if(queryBvhNodeIndex >= numQueryAabbs) return;
+	
+	int queryRigidIndex = mortonCodesAndAabbIndices[queryBvhNodeIndex].m_value;
+	b3AabbCL queryAabb = rigidAabbs[queryRigidIndex];
+	
+	int stack[B3_PLVBH_TRAVERSE_MAX_STACK_SIZE];
+	
+	int stackSize = 1;
+	stack[0] = *rootNodeIndex;
+	
+	while(stackSize)
+	{
+		int internalOrLeafNodeIndex = stack[ stackSize - 1 ];
+		--stackSize;
+		
+		int isLeaf = isLeafNode(internalOrLeafNodeIndex);	//Internal node if false
+		int bvhNodeIndex = getIndexWithInternalNodeMarkerRemoved(internalOrLeafNodeIndex);
+		
+		//Optimization - if the BVH is structured as a binary radix tree, then
+		//each internal node corresponds to a contiguous range of leaf nodes(internalNodeLeafIndexRanges[]).
+		//This can be used to avoid testing each AABB-AABB pair twice, including preventing each node from colliding with itself.
+		{
+			int highestLeafIndex = (isLeaf) ? bvhNodeIndex : internalNodeLeafIndexRanges[bvhNodeIndex].y;
+			if(highestLeafIndex <= queryBvhNodeIndex) continue;
+		}
+		
+		//bvhRigidIndex is not used if internal node
+		int bvhRigidIndex = (isLeaf) ? mortonCodesAndAabbIndices[bvhNodeIndex].m_value : -1;
+	
+		b3AabbCL bvhNodeAabb = (isLeaf) ? rigidAabbs[bvhRigidIndex] : internalNodeAabbs[bvhNodeIndex];
+		if( TestAabbAgainstAabb2(&queryAabb, &bvhNodeAabb) )
+		{
+			if(isLeaf)
+			{
+				int4 pair;
+				pair.x = rigidAabbs[queryRigidIndex].m_minIndices[3];
+				pair.y = rigidAabbs[bvhRigidIndex].m_minIndices[3];
+				pair.z = NEW_PAIR_MARKER;
+				pair.w = NEW_PAIR_MARKER;
+				
+				int pairIndex = atomic_inc(out_numPairs);
+				if(pairIndex < maxPairs) out_overlappingPairs[pairIndex] = pair;
+			}
+			
+			if(!isLeaf)	//Internal node
+			{
+				if(stackSize + 2 > B3_PLVBH_TRAVERSE_MAX_STACK_SIZE)
+				{
+					//Error
+				}
+				else
+				{
+					stack[ stackSize++ ] = internalNodeChildIndices[bvhNodeIndex].x;
+					stack[ stackSize++ ] = internalNodeChildIndices[bvhNodeIndex].y;
+				}
+			}
+		}
+		
+	}
+}
+
+
+//From rayCastKernels.cl
+typedef struct
+{
+	float4 m_from;
+	float4 m_to;
+} b3RayInfo;
+//From rayCastKernels.cl
+
+b3Vector3 b3Vector3_normalize(b3Vector3 v)
+{
+	b3Vector3 normal = (b3Vector3){v.x, v.y, v.z, 0.f};
+	return normalize(normal);	//OpenCL normalize == vector4 normalize
+}
+b3Scalar b3Vector3_length2(b3Vector3 v) { return v.x*v.x + v.y*v.y + v.z*v.z; }
+b3Scalar b3Vector3_dot(b3Vector3 a, b3Vector3 b) { return a.x*b.x + a.y*b.y + a.z*b.z; }
+
+int rayIntersectsAabb(b3Vector3 rayOrigin, b3Scalar rayLength, b3Vector3 rayNormalizedDirection, b3AabbCL aabb)
+{
+	//AABB is considered as 3 pairs of 2 planes( {x_min, x_max}, {y_min, y_max}, {z_min, z_max} ).
+	//t_min is the point of intersection with the closer plane, t_max is the point of intersection with the farther plane.
+	//
+	//if (rayNormalizedDirection.x < 0.0f), then max.x will be the near plane 
+	//and min.x will be the far plane; otherwise, it is reversed.
+	//
+	//In order for there to be a collision, the t_min and t_max of each pair must overlap.
+	//This can be tested for by selecting the highest t_min and lowest t_max and comparing them.
+	
+	int4 isNegative = isless( rayNormalizedDirection, ((b3Vector3){0.0f, 0.0f, 0.0f, 0.0f}) );	//isless(x,y) returns (x < y)
+	
+	//When using vector types, the select() function checks the most signficant bit, 
+	//but isless() sets the least significant bit.
+	isNegative <<= 31;
+
+	//select(b, a, condition) == condition ? a : b
+	//When using select() with vector types, (condition[i]) is true if its most significant bit is 1
+	b3Vector3 t_min = ( select(aabb.m_min, aabb.m_max, isNegative) - rayOrigin ) / rayNormalizedDirection;
+	b3Vector3 t_max = ( select(aabb.m_max, aabb.m_min, isNegative) - rayOrigin ) / rayNormalizedDirection;
+	
+	b3Scalar t_min_final = 0.0f;
+	b3Scalar t_max_final = rayLength;
+	
+	//Must use fmin()/fmax(); if one of the parameters is NaN, then the parameter that is not NaN is returned. 
+	//Behavior of min()/max() with NaNs is undefined. (See OpenCL Specification 1.2 [6.12.2] and [6.12.4])
+	//Since the innermost fmin()/fmax() is always not NaN, this should never return NaN.
+	t_min_final = fmax( t_min.z, fmax(t_min.y, fmax(t_min.x, t_min_final)) );
+	t_max_final = fmin( t_max.z, fmin(t_max.y, fmin(t_max.x, t_max_final)) );
+	
+	return (t_min_final <= t_max_final);
+}
+
+__kernel void plbvhRayTraverse(__global b3AabbCL* rigidAabbs,
+
+								__global int* rootNodeIndex, 
+								__global int2* internalNodeChildIndices, 
+								__global b3AabbCL* internalNodeAabbs,
+								__global int2* internalNodeLeafIndexRanges,
+								__global SortDataCL* mortonCodesAndAabbIndices,
+								
+								__global b3RayInfo* rays,
+								
+								__global int* out_numRayRigidPairs, 
+								__global int2* out_rayRigidPairs,
+								int maxRayRigidPairs, int numRays)
+{
+	int rayIndex = get_global_id(0);
+	if(rayIndex >= numRays) return;
+	
+	//
+	b3Vector3 rayFrom = rays[rayIndex].m_from;
+	b3Vector3 rayTo = rays[rayIndex].m_to;
+	b3Vector3 rayNormalizedDirection = b3Vector3_normalize(rayTo - rayFrom);
+	b3Scalar rayLength = b3Sqrt( b3Vector3_length2(rayTo - rayFrom) );
+	
+	//
+	int stack[B3_PLVBH_TRAVERSE_MAX_STACK_SIZE];
+	
+	int stackSize = 1;
+	stack[0] = *rootNodeIndex;
+	
+	while(stackSize)
+	{
+		int internalOrLeafNodeIndex = stack[ stackSize - 1 ];
+		--stackSize;
+		
+		int isLeaf = isLeafNode(internalOrLeafNodeIndex);	//Internal node if false
+		int bvhNodeIndex = getIndexWithInternalNodeMarkerRemoved(internalOrLeafNodeIndex);
+		
+		//bvhRigidIndex is not used if internal node
+		int bvhRigidIndex = (isLeaf) ? mortonCodesAndAabbIndices[bvhNodeIndex].m_value : -1;
+	
+		b3AabbCL bvhNodeAabb = (isLeaf) ? rigidAabbs[bvhRigidIndex] : internalNodeAabbs[bvhNodeIndex];
+		if( rayIntersectsAabb(rayFrom, rayLength, rayNormalizedDirection, bvhNodeAabb)  )
+		{
+			if(isLeaf)
+			{
+				int2 rayRigidPair;
+				rayRigidPair.x = rayIndex;
+				rayRigidPair.y = rigidAabbs[bvhRigidIndex].m_minIndices[3];
+				
+				int pairIndex = atomic_inc(out_numRayRigidPairs);
+				if(pairIndex < maxRayRigidPairs) out_rayRigidPairs[pairIndex] = rayRigidPair;
+			}
+			
+			if(!isLeaf)	//Internal node
+			{
+				if(stackSize + 2 > B3_PLVBH_TRAVERSE_MAX_STACK_SIZE)
+				{
+					//Error
+				}
+				else
+				{
+					stack[ stackSize++ ] = internalNodeChildIndices[bvhNodeIndex].x;
+					stack[ stackSize++ ] = internalNodeChildIndices[bvhNodeIndex].y;
+				}
+			}
+		}
+	}
+}
+
+__kernel void plbvhLargeAabbAabbTest(__global b3AabbCL* smallAabbs, __global b3AabbCL* largeAabbs, 
+									__global int* out_numPairs, __global int4* out_overlappingPairs, 
+									int maxPairs, int numLargeAabbRigids, int numSmallAabbRigids)
+{
+	int smallAabbIndex = get_global_id(0);
+	if(smallAabbIndex >= numSmallAabbRigids) return;
+	
+	b3AabbCL smallAabb = smallAabbs[smallAabbIndex];
+	for(int i = 0; i < numLargeAabbRigids; ++i)
+	{
+		b3AabbCL largeAabb = largeAabbs[i];
+		if( TestAabbAgainstAabb2(&smallAabb, &largeAabb) )
+		{
+			int4 pair;
+			pair.x = largeAabb.m_minIndices[3];
+			pair.y = smallAabb.m_minIndices[3];
+			pair.z = NEW_PAIR_MARKER;
+			pair.w = NEW_PAIR_MARKER;
+			
+			int pairIndex = atomic_inc(out_numPairs);
+			if(pairIndex < maxPairs) out_overlappingPairs[pairIndex] = pair;
+		}
+	}
+}
+__kernel void plbvhLargeAabbRayTest(__global b3AabbCL* largeRigidAabbs, __global b3RayInfo* rays,
+									__global int* out_numRayRigidPairs,  __global int2* out_rayRigidPairs,
+									int numLargeAabbRigids, int maxRayRigidPairs, int numRays)
+{
+	int rayIndex = get_global_id(0);
+	if(rayIndex >= numRays) return;
+	
+	b3Vector3 rayFrom = rays[rayIndex].m_from;
+	b3Vector3 rayTo = rays[rayIndex].m_to;
+	b3Vector3 rayNormalizedDirection = b3Vector3_normalize(rayTo - rayFrom);
+	b3Scalar rayLength = b3Sqrt( b3Vector3_length2(rayTo - rayFrom) );
+	
+	for(int i = 0; i < numLargeAabbRigids; ++i)
+	{
+		b3AabbCL rigidAabb = largeRigidAabbs[i];
+		if( rayIntersectsAabb(rayFrom, rayLength, rayNormalizedDirection, rigidAabb) )
+		{
+			int2 rayRigidPair;
+			rayRigidPair.x = rayIndex;
+			rayRigidPair.y = rigidAabb.m_minIndices[3];
+			
+			int pairIndex = atomic_inc(out_numRayRigidPairs);
+			if(pairIndex < maxRayRigidPairs) out_rayRigidPairs[pairIndex] = rayRigidPair;
+		}
+	}
+}
+
+
+//Set so that it is always greater than the actual common prefixes, and never selected as a parent node.
+//If there are no duplicates, then the highest common prefix is 32 or 64, depending on the number of bits used for the z-curve.
+//Duplicate common prefixes increase the highest common prefix at most by the number of bits used to index the leaf node.
+//Since 32 bit ints are used to index leaf nodes, the max prefix is 64(32 + 32 bit z-curve) or 96(32 + 64 bit z-curve).
+#define B3_PLBVH_INVALID_COMMON_PREFIX 128
+
+#define B3_PLBVH_ROOT_NODE_MARKER -1
+
+#define b3Int64 long
+
+int computeCommonPrefixLength(b3Int64 i, b3Int64 j) { return (int)clz(i ^ j); }
+b3Int64 computeCommonPrefix(b3Int64 i, b3Int64 j) 
+{
+	//This function only needs to return (i & j) in order for the algorithm to work,
+	//but it may help with debugging to mask out the lower bits.
+
+	b3Int64 commonPrefixLength = (b3Int64)computeCommonPrefixLength(i, j);
+
+	b3Int64 sharedBits = i & j;
+	b3Int64 bitmask = ((b3Int64)(~0)) << (64 - commonPrefixLength);	//Set all bits after the common prefix to 0
+	
+	return sharedBits & bitmask;
+}
+
+//Same as computeCommonPrefixLength(), but allows for prefixes with different lengths
+int getSharedPrefixLength(b3Int64 prefixA, int prefixLengthA, b3Int64 prefixB, int prefixLengthB)
+{
+	return b3Min( computeCommonPrefixLength(prefixA, prefixB), b3Min(prefixLengthA, prefixLengthB) );
+}
+
+__kernel void computeAdjacentPairCommonPrefix(__global SortDataCL* mortonCodesAndAabbIndices,
+											__global b3Int64* out_commonPrefixes,
+											__global int* out_commonPrefixLengths,
+											int numInternalNodes)
+{
+	int internalNodeIndex = get_global_id(0);
+	if (internalNodeIndex >= numInternalNodes) return;
+	
+	//Here, (internalNodeIndex + 1) is never out of bounds since it is a leaf node index,
+	//and the number of internal nodes is always numLeafNodes - 1
+	int leftLeafIndex = internalNodeIndex;
+	int rightLeafIndex = internalNodeIndex + 1;
+	
+	int leftLeafMortonCode = mortonCodesAndAabbIndices[leftLeafIndex].m_key;
+	int rightLeafMortonCode = mortonCodesAndAabbIndices[rightLeafIndex].m_key;
+	
+	//Binary radix tree construction algorithm does not work if there are duplicate morton codes.
+	//Append the index of each leaf node to each morton code so that there are no duplicates.
+	//The algorithm also requires that the morton codes are sorted in ascending order; this requirement
+	//is also satisfied with this method, as (leftLeafIndex < rightLeafIndex) is always true.
+	//
+	//upsample(a, b) == ( ((b3Int64)a) << 32) | b
+	b3Int64 nonduplicateLeftMortonCode = upsample(leftLeafMortonCode, leftLeafIndex);
+	b3Int64 nonduplicateRightMortonCode = upsample(rightLeafMortonCode, rightLeafIndex);
+	
+	out_commonPrefixes[internalNodeIndex] = computeCommonPrefix(nonduplicateLeftMortonCode, nonduplicateRightMortonCode);
+	out_commonPrefixLengths[internalNodeIndex] = computeCommonPrefixLength(nonduplicateLeftMortonCode, nonduplicateRightMortonCode);
+}
+
+
+__kernel void buildBinaryRadixTreeLeafNodes(__global int* commonPrefixLengths, __global int* out_leafNodeParentNodes,
+											__global int2* out_childNodes, int numLeafNodes)
+{
+	int leafNodeIndex = get_global_id(0);
+	if (leafNodeIndex >= numLeafNodes) return;
+	
+	int numInternalNodes = numLeafNodes - 1;
+	
+	int leftSplitIndex = leafNodeIndex - 1;
+	int rightSplitIndex = leafNodeIndex;
+	
+	int leftCommonPrefix = (leftSplitIndex >= 0) ? commonPrefixLengths[leftSplitIndex] : B3_PLBVH_INVALID_COMMON_PREFIX;
+	int rightCommonPrefix = (rightSplitIndex < numInternalNodes) ? commonPrefixLengths[rightSplitIndex] : B3_PLBVH_INVALID_COMMON_PREFIX;
+	
+	//Parent node is the highest adjacent common prefix that is lower than the node's common prefix
+	//Leaf nodes are considered as having the highest common prefix
+	int isLeftHigherCommonPrefix = (leftCommonPrefix > rightCommonPrefix);
+	
+	//Handle cases for the edge nodes; the first and last node
+	//For leaf nodes, leftCommonPrefix and rightCommonPrefix should never both be B3_PLBVH_INVALID_COMMON_PREFIX
+	if(leftCommonPrefix == B3_PLBVH_INVALID_COMMON_PREFIX) isLeftHigherCommonPrefix = false;
+	if(rightCommonPrefix == B3_PLBVH_INVALID_COMMON_PREFIX) isLeftHigherCommonPrefix = true;
+	
+	int parentNodeIndex = (isLeftHigherCommonPrefix) ? leftSplitIndex : rightSplitIndex;
+	out_leafNodeParentNodes[leafNodeIndex] = parentNodeIndex;
+	
+	int isRightChild = (isLeftHigherCommonPrefix);	//If the left node is the parent, then this node is its right child and vice versa
+	
+	//out_childNodesAsInt[0] == int2.x == left child
+	//out_childNodesAsInt[1] == int2.y == right child
+	int isLeaf = 1;
+	__global int* out_childNodesAsInt = (__global int*)(&out_childNodes[parentNodeIndex]);
+	out_childNodesAsInt[isRightChild] = getIndexWithInternalNodeMarkerSet(isLeaf, leafNodeIndex);
+}
+
+__kernel void buildBinaryRadixTreeInternalNodes(__global b3Int64* commonPrefixes, __global int* commonPrefixLengths,
+												__global int2* out_childNodes,
+												__global int* out_internalNodeParentNodes, __global int* out_rootNodeIndex,
+												int numInternalNodes)
+{
+	int internalNodeIndex = get_group_id(0) * get_local_size(0) + get_local_id(0);
+	if(internalNodeIndex >= numInternalNodes) return;
+	
+	b3Int64 nodePrefix = commonPrefixes[internalNodeIndex];
+	int nodePrefixLength = commonPrefixLengths[internalNodeIndex];
+	
+//#define USE_LINEAR_SEARCH
+#ifdef USE_LINEAR_SEARCH
+	int leftIndex = -1;
+	int rightIndex = -1;
+	
+	//Find nearest element to left with a lower common prefix
+	for(int i = internalNodeIndex - 1; i >= 0; --i)
+	{
+		int nodeLeftSharedPrefixLength = getSharedPrefixLength(nodePrefix, nodePrefixLength, commonPrefixes[i], commonPrefixLengths[i]);
+		if(nodeLeftSharedPrefixLength < nodePrefixLength)
+		{
+			leftIndex = i;
+			break;
+		}
+	}
+	
+	//Find nearest element to right with a lower common prefix
+	for(int i = internalNodeIndex + 1; i < numInternalNodes; ++i)
+	{
+		int nodeRightSharedPrefixLength = getSharedPrefixLength(nodePrefix, nodePrefixLength, commonPrefixes[i], commonPrefixLengths[i]);
+		if(nodeRightSharedPrefixLength < nodePrefixLength)
+		{
+			rightIndex = i;
+			break;
+		}
+	}
+	
+#else //Use binary search
+
+	//Find nearest element to left with a lower common prefix
+	int leftIndex = -1;
+	{
+		int lower = 0;
+		int upper = internalNodeIndex - 1;
+		
+		while(lower <= upper)
+		{
+			int mid = (lower + upper) / 2;
+			b3Int64 midPrefix = commonPrefixes[mid];
+			int midPrefixLength = commonPrefixLengths[mid];
+			
+			int nodeMidSharedPrefixLength = getSharedPrefixLength(nodePrefix, nodePrefixLength, midPrefix, midPrefixLength);
+			if(nodeMidSharedPrefixLength < nodePrefixLength) 
+			{
+				int right = mid + 1;
+				if(right < internalNodeIndex)
+				{
+					b3Int64 rightPrefix = commonPrefixes[right];
+					int rightPrefixLength = commonPrefixLengths[right];
+					
+					int nodeRightSharedPrefixLength = getSharedPrefixLength(nodePrefix, nodePrefixLength, rightPrefix, rightPrefixLength);
+					if(nodeRightSharedPrefixLength < nodePrefixLength) 
+					{
+						lower = right;
+						leftIndex = right;
+					}
+					else 
+					{
+						leftIndex = mid;
+						break;
+					}
+				}
+				else 
+				{
+					leftIndex = mid;
+					break;
+				}
+			}
+			else upper = mid - 1;
+		}
+	}
+	
+	//Find nearest element to right with a lower common prefix
+	int rightIndex = -1;
+	{
+		int lower = internalNodeIndex + 1;
+		int upper = numInternalNodes - 1;
+		
+		while(lower <= upper)
+		{
+			int mid = (lower + upper) / 2;
+			b3Int64 midPrefix = commonPrefixes[mid];
+			int midPrefixLength = commonPrefixLengths[mid];
+			
+			int nodeMidSharedPrefixLength = getSharedPrefixLength(nodePrefix, nodePrefixLength, midPrefix, midPrefixLength);
+			if(nodeMidSharedPrefixLength < nodePrefixLength) 
+			{
+				int left = mid - 1;
+				if(left > internalNodeIndex)
+				{
+					b3Int64 leftPrefix = commonPrefixes[left];
+					int leftPrefixLength = commonPrefixLengths[left];
+				
+					int nodeLeftSharedPrefixLength = getSharedPrefixLength(nodePrefix, nodePrefixLength, leftPrefix, leftPrefixLength);
+					if(nodeLeftSharedPrefixLength < nodePrefixLength) 
+					{
+						upper = left;
+						rightIndex = left;
+					}
+					else 
+					{
+						rightIndex = mid;
+						break;
+					}
+				}
+				else 
+				{
+					rightIndex = mid;
+					break;
+				}
+			}
+			else lower = mid + 1;
+		}
+	}
+#endif
+	
+	//Select parent
+	{
+		int leftPrefixLength = (leftIndex != -1) ? commonPrefixLengths[leftIndex] : B3_PLBVH_INVALID_COMMON_PREFIX;
+		int rightPrefixLength =  (rightIndex != -1) ? commonPrefixLengths[rightIndex] : B3_PLBVH_INVALID_COMMON_PREFIX;
+		
+		int isLeftHigherPrefixLength = (leftPrefixLength > rightPrefixLength);
+		
+		if(leftPrefixLength == B3_PLBVH_INVALID_COMMON_PREFIX) isLeftHigherPrefixLength = false;
+		else if(rightPrefixLength == B3_PLBVH_INVALID_COMMON_PREFIX) isLeftHigherPrefixLength = true;
+		
+		int parentNodeIndex = (isLeftHigherPrefixLength) ? leftIndex : rightIndex;
+		
+		int isRootNode = (leftIndex == -1 && rightIndex == -1);
+		out_internalNodeParentNodes[internalNodeIndex] = (!isRootNode) ? parentNodeIndex : B3_PLBVH_ROOT_NODE_MARKER;
+		
+		int isLeaf = 0;
+		if(!isRootNode)
+		{
+			int isRightChild = (isLeftHigherPrefixLength);	//If the left node is the parent, then this node is its right child and vice versa
+			
+			//out_childNodesAsInt[0] == int2.x == left child
+			//out_childNodesAsInt[1] == int2.y == right child
+			__global int* out_childNodesAsInt = (__global int*)(&out_childNodes[parentNodeIndex]);
+			out_childNodesAsInt[isRightChild] = getIndexWithInternalNodeMarkerSet(isLeaf, internalNodeIndex);
+		}
+		else *out_rootNodeIndex = getIndexWithInternalNodeMarkerSet(isLeaf, internalNodeIndex);
+	}
+}
+
+__kernel void findDistanceFromRoot(__global int* rootNodeIndex, __global int* internalNodeParentNodes,
+									__global int* out_maxDistanceFromRoot, __global int* out_distanceFromRoot, int numInternalNodes)
+{
+	if( get_global_id(0) == 0 ) atomic_xchg(out_maxDistanceFromRoot, 0);
+
+	int internalNodeIndex = get_global_id(0);
+	if(internalNodeIndex >= numInternalNodes) return;
+	
+	//
+	int distanceFromRoot = 0;
+	{
+		int parentIndex = internalNodeParentNodes[internalNodeIndex];
+		while(parentIndex != B3_PLBVH_ROOT_NODE_MARKER)
+		{
+			parentIndex = internalNodeParentNodes[parentIndex];
+			++distanceFromRoot;
+		}
+	}
+	out_distanceFromRoot[internalNodeIndex] = distanceFromRoot;
+	
+	//
+	__local int localMaxDistanceFromRoot;
+	if( get_local_id(0) == 0 ) localMaxDistanceFromRoot = 0;
+	barrier(CLK_LOCAL_MEM_FENCE);
+	
+	atomic_max(&localMaxDistanceFromRoot, distanceFromRoot);
+	barrier(CLK_LOCAL_MEM_FENCE);
+	
+	if( get_local_id(0) == 0 ) atomic_max(out_maxDistanceFromRoot, localMaxDistanceFromRoot);
+}
+
+__kernel void buildBinaryRadixTreeAabbsRecursive(__global int* distanceFromRoot, __global SortDataCL* mortonCodesAndAabbIndices,
+												__global int2* childNodes,
+												__global b3AabbCL* leafNodeAabbs, __global b3AabbCL* internalNodeAabbs,
+												int maxDistanceFromRoot, int processedDistance, int numInternalNodes)
+{
+	int internalNodeIndex = get_global_id(0);
+	if(internalNodeIndex >= numInternalNodes) return;
+	
+	int distance = distanceFromRoot[internalNodeIndex];
+	
+	if(distance == processedDistance)
+	{
+		int leftChildIndex = childNodes[internalNodeIndex].x;
+		int rightChildIndex = childNodes[internalNodeIndex].y;
+		
+		int isLeftChildLeaf = isLeafNode(leftChildIndex);
+		int isRightChildLeaf = isLeafNode(rightChildIndex);
+		
+		leftChildIndex = getIndexWithInternalNodeMarkerRemoved(leftChildIndex);
+		rightChildIndex = getIndexWithInternalNodeMarkerRemoved(rightChildIndex);
+		
+		//leftRigidIndex/rightRigidIndex is not used if internal node
+		int leftRigidIndex = (isLeftChildLeaf) ? mortonCodesAndAabbIndices[leftChildIndex].m_value : -1;
+		int rightRigidIndex = (isRightChildLeaf) ? mortonCodesAndAabbIndices[rightChildIndex].m_value : -1;
+		
+		b3AabbCL leftChildAabb = (isLeftChildLeaf) ? leafNodeAabbs[leftRigidIndex] : internalNodeAabbs[leftChildIndex];
+		b3AabbCL rightChildAabb = (isRightChildLeaf) ? leafNodeAabbs[rightRigidIndex] : internalNodeAabbs[rightChildIndex];
+		
+		b3AabbCL mergedAabb;
+		mergedAabb.m_min = b3Min(leftChildAabb.m_min, rightChildAabb.m_min);
+		mergedAabb.m_max = b3Max(leftChildAabb.m_max, rightChildAabb.m_max);
+		internalNodeAabbs[internalNodeIndex] = mergedAabb;
+	}
+}
+
+__kernel void findLeafIndexRanges(__global int2* internalNodeChildNodes, __global int2* out_leafIndexRanges, int numInternalNodes)
+{
+	int internalNodeIndex = get_global_id(0);
+	if(internalNodeIndex >= numInternalNodes) return;
+	
+	int numLeafNodes = numInternalNodes + 1;
+	
+	int2 childNodes = internalNodeChildNodes[internalNodeIndex];
+	
+	int2 leafIndexRange;	//x == min leaf index, y == max leaf index
+	
+	//Find lowest leaf index covered by this internal node
+	{
+		int lowestIndex = childNodes.x;		//childNodes.x == Left child
+		while( !isLeafNode(lowestIndex) ) lowestIndex = internalNodeChildNodes[ getIndexWithInternalNodeMarkerRemoved(lowestIndex) ].x;
+		leafIndexRange.x = lowestIndex;
+	}
+	
+	//Find highest leaf index covered by this internal node
+	{
+		int highestIndex = childNodes.y;	//childNodes.y == Right child
+		while( !isLeafNode(highestIndex) ) highestIndex = internalNodeChildNodes[ getIndexWithInternalNodeMarkerRemoved(highestIndex) ].y;
+		leafIndexRange.y = highestIndex;
+	}
+	
+	//
+	out_leafIndexRanges[internalNodeIndex] = leafIndexRange;
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/parallelLinearBvhKernels.h b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/parallelLinearBvhKernels.h
new file mode 100644
index 0000000000..5eb8f45b16
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/parallelLinearBvhKernels.h
@@ -0,0 +1,729 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* parallelLinearBvhCL= \
+"/*\n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose,\n"
+"including commercial applications, and to alter it and redistribute it freely,\n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Initial Author Jackson Lee, 2014\n"
+"typedef float b3Scalar;\n"
+"typedef float4 b3Vector3;\n"
+"#define b3Max max\n"
+"#define b3Min min\n"
+"#define b3Sqrt sqrt\n"
+"typedef struct\n"
+"{\n"
+"	unsigned int m_key;\n"
+"	unsigned int m_value;\n"
+"} SortDataCL;\n"
+"typedef struct \n"
+"{\n"
+"	union\n"
+"	{\n"
+"		float4	m_min;\n"
+"		float   m_minElems[4];\n"
+"		int			m_minIndices[4];\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float4	m_max;\n"
+"		float   m_maxElems[4];\n"
+"		int			m_maxIndices[4];\n"
+"	};\n"
+"} b3AabbCL;\n"
+"unsigned int interleaveBits(unsigned int x)\n"
+"{\n"
+"	//........ ........ ......12 3456789A	//x\n"
+"	//....1..2 ..3..4.. 5..6..7. .8..9..A	//x after interleaving bits\n"
+"	\n"
+"	//......12 3456789A ......12 3456789A	//x ^ (x << 16)\n"
+"	//11111111 ........ ........ 11111111	//0x FF 00 00 FF\n"
+"	//......12 ........ ........ 3456789A	//x = (x ^ (x << 16)) & 0xFF0000FF;\n"
+"	\n"
+"	//......12 ........ 3456789A 3456789A	//x ^ (x <<  8)\n"
+"	//......11 ........ 1111.... ....1111	//0x 03 00 F0 0F\n"
+"	//......12 ........ 3456.... ....789A	//x = (x ^ (x <<  8)) & 0x0300F00F;\n"
+"	\n"
+"	//..12..12 ....3456 3456.... 789A789A	//x ^ (x <<  4)\n"
+"	//......11 ....11.. ..11.... 11....11	//0x 03 0C 30 C3\n"
+"	//......12 ....34.. ..56.... 78....9A	//x = (x ^ (x <<  4)) & 0x030C30C3;\n"
+"	\n"
+"	//....1212 ..3434.. 5656..78 78..9A9A	//x ^ (x <<  2)\n"
+"	//....1..1 ..1..1.. 1..1..1. .1..1..1	//0x 09 24 92 49\n"
+"	//....1..2 ..3..4.. 5..6..7. .8..9..A	//x = (x ^ (x <<  2)) & 0x09249249;\n"
+"	\n"
+"	//........ ........ ......11 11111111	//0x000003FF\n"
+"	x &= 0x000003FF;		//Clear all bits above bit 10\n"
+"	\n"
+"	x = (x ^ (x << 16)) & 0xFF0000FF;\n"
+"	x = (x ^ (x <<  8)) & 0x0300F00F;\n"
+"	x = (x ^ (x <<  4)) & 0x030C30C3;\n"
+"	x = (x ^ (x <<  2)) & 0x09249249;\n"
+"	\n"
+"	return x;\n"
+"}\n"
+"unsigned int getMortonCode(unsigned int x, unsigned int y, unsigned int z)\n"
+"{\n"
+"	return interleaveBits(x) << 0 | interleaveBits(y) << 1 | interleaveBits(z) << 2;\n"
+"}\n"
+"__kernel void separateAabbs(__global b3AabbCL* unseparatedAabbs, __global int* aabbIndices, __global b3AabbCL* out_aabbs, int numAabbsToSeparate)\n"
+"{\n"
+"	int separatedAabbIndex = get_global_id(0);\n"
+"	if(separatedAabbIndex >= numAabbsToSeparate) return;\n"
+"	int unseparatedAabbIndex = aabbIndices[separatedAabbIndex];\n"
+"	out_aabbs[separatedAabbIndex] = unseparatedAabbs[unseparatedAabbIndex];\n"
+"}\n"
+"//Should replace with an optimized parallel reduction\n"
+"__kernel void findAllNodesMergedAabb(__global b3AabbCL* out_mergedAabb, int numAabbsNeedingMerge)\n"
+"{\n"
+"	//Each time this kernel is added to the command queue, \n"
+"	//the number of AABBs needing to be merged is halved\n"
+"	//\n"
+"	//Example with 159 AABBs:\n"
+"	//	numRemainingAabbs == 159 / 2 + 159 % 2 == 80\n"
+"	//	numMergedAabbs == 159 - 80 == 79\n"
+"	//So, indices [0, 78] are merged with [0 + 80, 78 + 80]\n"
+"	\n"
+"	int numRemainingAabbs = numAabbsNeedingMerge / 2 + numAabbsNeedingMerge % 2;\n"
+"	int numMergedAabbs = numAabbsNeedingMerge - numRemainingAabbs;\n"
+"	\n"
+"	int aabbIndex = get_global_id(0);\n"
+"	if(aabbIndex >= numMergedAabbs) return;\n"
+"	\n"
+"	int otherAabbIndex = aabbIndex + numRemainingAabbs;\n"
+"	\n"
+"	b3AabbCL aabb = out_mergedAabb[aabbIndex];\n"
+"	b3AabbCL otherAabb = out_mergedAabb[otherAabbIndex];\n"
+"		\n"
+"	b3AabbCL mergedAabb;\n"
+"	mergedAabb.m_min = b3Min(aabb.m_min, otherAabb.m_min);\n"
+"	mergedAabb.m_max = b3Max(aabb.m_max, otherAabb.m_max);\n"
+"	out_mergedAabb[aabbIndex] = mergedAabb;\n"
+"}\n"
+"__kernel void assignMortonCodesAndAabbIndicies(__global b3AabbCL* worldSpaceAabbs, __global b3AabbCL* mergedAabbOfAllNodes, \n"
+"												__global SortDataCL* out_mortonCodesAndAabbIndices, int numAabbs)\n"
+"{\n"
+"	int leafNodeIndex = get_global_id(0);	//Leaf node index == AABB index\n"
+"	if(leafNodeIndex >= numAabbs) return;\n"
+"	\n"
+"	b3AabbCL mergedAabb = mergedAabbOfAllNodes[0];\n"
+"	b3Vector3 gridCenter = (mergedAabb.m_min + mergedAabb.m_max) * 0.5f;\n"
+"	b3Vector3 gridCellSize = (mergedAabb.m_max - mergedAabb.m_min) / (float)1024;\n"
+"	\n"
+"	b3AabbCL aabb = worldSpaceAabbs[leafNodeIndex];\n"
+"	b3Vector3 aabbCenter = (aabb.m_min + aabb.m_max) * 0.5f;\n"
+"	b3Vector3 aabbCenterRelativeToGrid = aabbCenter - gridCenter;\n"
+"	\n"
+"	//Quantize into integer coordinates\n"
+"	//floor() is needed to prevent the center cell, at (0,0,0) from being twice the size\n"
+"	b3Vector3 gridPosition = aabbCenterRelativeToGrid / gridCellSize;\n"
+"	\n"
+"	int4 discretePosition;\n"
+"	discretePosition.x = (int)( (gridPosition.x >= 0.0f) ? gridPosition.x : floor(gridPosition.x) );\n"
+"	discretePosition.y = (int)( (gridPosition.y >= 0.0f) ? gridPosition.y : floor(gridPosition.y) );\n"
+"	discretePosition.z = (int)( (gridPosition.z >= 0.0f) ? gridPosition.z : floor(gridPosition.z) );\n"
+"	\n"
+"	//Clamp coordinates into [-512, 511], then convert range from [-512, 511] to [0, 1023]\n"
+"	discretePosition = b3Max( -512, b3Min(discretePosition, 511) );\n"
+"	discretePosition += 512;\n"
+"	\n"
+"	//Interleave bits(assign a morton code, also known as a z-curve)\n"
+"	unsigned int mortonCode = getMortonCode(discretePosition.x, discretePosition.y, discretePosition.z);\n"
+"	\n"
+"	//\n"
+"	SortDataCL mortonCodeIndexPair;\n"
+"	mortonCodeIndexPair.m_key = mortonCode;\n"
+"	mortonCodeIndexPair.m_value = leafNodeIndex;\n"
+"	\n"
+"	out_mortonCodesAndAabbIndices[leafNodeIndex] = mortonCodeIndexPair;\n"
+"}\n"
+"#define B3_PLVBH_TRAVERSE_MAX_STACK_SIZE 128\n"
+"//The most significant bit(0x80000000) of a int32 is used to distinguish between leaf and internal nodes.\n"
+"//If it is set, then the index is for an internal node; otherwise, it is a leaf node. \n"
+"//In both cases, the bit should be cleared to access the actual node index.\n"
+"int isLeafNode(int index) { return (index >> 31 == 0); }\n"
+"int getIndexWithInternalNodeMarkerRemoved(int index) { return index & (~0x80000000); }\n"
+"int getIndexWithInternalNodeMarkerSet(int isLeaf, int index) { return (isLeaf) ? index : (index | 0x80000000); }\n"
+"//From sap.cl\n"
+"#define NEW_PAIR_MARKER -1\n"
+"bool TestAabbAgainstAabb2(const b3AabbCL* aabb1, const b3AabbCL* aabb2)\n"
+"{\n"
+"	bool overlap = true;\n"
+"	overlap = (aabb1->m_min.x > aabb2->m_max.x || aabb1->m_max.x < aabb2->m_min.x) ? false : overlap;\n"
+"	overlap = (aabb1->m_min.z > aabb2->m_max.z || aabb1->m_max.z < aabb2->m_min.z) ? false : overlap;\n"
+"	overlap = (aabb1->m_min.y > aabb2->m_max.y || aabb1->m_max.y < aabb2->m_min.y) ? false : overlap;\n"
+"	return overlap;\n"
+"}\n"
+"//From sap.cl\n"
+"__kernel void plbvhCalculateOverlappingPairs(__global b3AabbCL* rigidAabbs, \n"
+"											__global int* rootNodeIndex, \n"
+"											__global int2* internalNodeChildIndices, \n"
+"											__global b3AabbCL* internalNodeAabbs,\n"
+"											__global int2* internalNodeLeafIndexRanges,\n"
+"											\n"
+"											__global SortDataCL* mortonCodesAndAabbIndices,\n"
+"											__global int* out_numPairs, __global int4* out_overlappingPairs, \n"
+"											int maxPairs, int numQueryAabbs)\n"
+"{\n"
+"	//Using get_group_id()/get_local_id() is Faster than get_global_id(0) since\n"
+"	//mortonCodesAndAabbIndices[] contains rigid body indices sorted along the z-curve (more spatially coherent)\n"
+"	int queryBvhNodeIndex = get_group_id(0) * get_local_size(0) + get_local_id(0);\n"
+"	if(queryBvhNodeIndex >= numQueryAabbs) return;\n"
+"	\n"
+"	int queryRigidIndex = mortonCodesAndAabbIndices[queryBvhNodeIndex].m_value;\n"
+"	b3AabbCL queryAabb = rigidAabbs[queryRigidIndex];\n"
+"	\n"
+"	int stack[B3_PLVBH_TRAVERSE_MAX_STACK_SIZE];\n"
+"	\n"
+"	int stackSize = 1;\n"
+"	stack[0] = *rootNodeIndex;\n"
+"	\n"
+"	while(stackSize)\n"
+"	{\n"
+"		int internalOrLeafNodeIndex = stack[ stackSize - 1 ];\n"
+"		--stackSize;\n"
+"		\n"
+"		int isLeaf = isLeafNode(internalOrLeafNodeIndex);	//Internal node if false\n"
+"		int bvhNodeIndex = getIndexWithInternalNodeMarkerRemoved(internalOrLeafNodeIndex);\n"
+"		\n"
+"		//Optimization - if the BVH is structured as a binary radix tree, then\n"
+"		//each internal node corresponds to a contiguous range of leaf nodes(internalNodeLeafIndexRanges[]).\n"
+"		//This can be used to avoid testing each AABB-AABB pair twice, including preventing each node from colliding with itself.\n"
+"		{\n"
+"			int highestLeafIndex = (isLeaf) ? bvhNodeIndex : internalNodeLeafIndexRanges[bvhNodeIndex].y;\n"
+"			if(highestLeafIndex <= queryBvhNodeIndex) continue;\n"
+"		}\n"
+"		\n"
+"		//bvhRigidIndex is not used if internal node\n"
+"		int bvhRigidIndex = (isLeaf) ? mortonCodesAndAabbIndices[bvhNodeIndex].m_value : -1;\n"
+"	\n"
+"		b3AabbCL bvhNodeAabb = (isLeaf) ? rigidAabbs[bvhRigidIndex] : internalNodeAabbs[bvhNodeIndex];\n"
+"		if( TestAabbAgainstAabb2(&queryAabb, &bvhNodeAabb) )\n"
+"		{\n"
+"			if(isLeaf)\n"
+"			{\n"
+"				int4 pair;\n"
+"				pair.x = rigidAabbs[queryRigidIndex].m_minIndices[3];\n"
+"				pair.y = rigidAabbs[bvhRigidIndex].m_minIndices[3];\n"
+"				pair.z = NEW_PAIR_MARKER;\n"
+"				pair.w = NEW_PAIR_MARKER;\n"
+"				\n"
+"				int pairIndex = atomic_inc(out_numPairs);\n"
+"				if(pairIndex < maxPairs) out_overlappingPairs[pairIndex] = pair;\n"
+"			}\n"
+"			\n"
+"			if(!isLeaf)	//Internal node\n"
+"			{\n"
+"				if(stackSize + 2 > B3_PLVBH_TRAVERSE_MAX_STACK_SIZE)\n"
+"				{\n"
+"					//Error\n"
+"				}\n"
+"				else\n"
+"				{\n"
+"					stack[ stackSize++ ] = internalNodeChildIndices[bvhNodeIndex].x;\n"
+"					stack[ stackSize++ ] = internalNodeChildIndices[bvhNodeIndex].y;\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"		\n"
+"	}\n"
+"}\n"
+"//From rayCastKernels.cl\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_from;\n"
+"	float4 m_to;\n"
+"} b3RayInfo;\n"
+"//From rayCastKernels.cl\n"
+"b3Vector3 b3Vector3_normalize(b3Vector3 v)\n"
+"{\n"
+"	b3Vector3 normal = (b3Vector3){v.x, v.y, v.z, 0.f};\n"
+"	return normalize(normal);	//OpenCL normalize == vector4 normalize\n"
+"}\n"
+"b3Scalar b3Vector3_length2(b3Vector3 v) { return v.x*v.x + v.y*v.y + v.z*v.z; }\n"
+"b3Scalar b3Vector3_dot(b3Vector3 a, b3Vector3 b) { return a.x*b.x + a.y*b.y + a.z*b.z; }\n"
+"int rayIntersectsAabb(b3Vector3 rayOrigin, b3Scalar rayLength, b3Vector3 rayNormalizedDirection, b3AabbCL aabb)\n"
+"{\n"
+"	//AABB is considered as 3 pairs of 2 planes( {x_min, x_max}, {y_min, y_max}, {z_min, z_max} ).\n"
+"	//t_min is the point of intersection with the closer plane, t_max is the point of intersection with the farther plane.\n"
+"	//\n"
+"	//if (rayNormalizedDirection.x < 0.0f), then max.x will be the near plane \n"
+"	//and min.x will be the far plane; otherwise, it is reversed.\n"
+"	//\n"
+"	//In order for there to be a collision, the t_min and t_max of each pair must overlap.\n"
+"	//This can be tested for by selecting the highest t_min and lowest t_max and comparing them.\n"
+"	\n"
+"	int4 isNegative = isless( rayNormalizedDirection, ((b3Vector3){0.0f, 0.0f, 0.0f, 0.0f}) );	//isless(x,y) returns (x < y)\n"
+"	\n"
+"	//When using vector types, the select() function checks the most signficant bit, \n"
+"	//but isless() sets the least significant bit.\n"
+"	isNegative <<= 31;\n"
+"	//select(b, a, condition) == condition ? a : b\n"
+"	//When using select() with vector types, (condition[i]) is true if its most significant bit is 1\n"
+"	b3Vector3 t_min = ( select(aabb.m_min, aabb.m_max, isNegative) - rayOrigin ) / rayNormalizedDirection;\n"
+"	b3Vector3 t_max = ( select(aabb.m_max, aabb.m_min, isNegative) - rayOrigin ) / rayNormalizedDirection;\n"
+"	\n"
+"	b3Scalar t_min_final = 0.0f;\n"
+"	b3Scalar t_max_final = rayLength;\n"
+"	\n"
+"	//Must use fmin()/fmax(); if one of the parameters is NaN, then the parameter that is not NaN is returned. \n"
+"	//Behavior of min()/max() with NaNs is undefined. (See OpenCL Specification 1.2 [6.12.2] and [6.12.4])\n"
+"	//Since the innermost fmin()/fmax() is always not NaN, this should never return NaN.\n"
+"	t_min_final = fmax( t_min.z, fmax(t_min.y, fmax(t_min.x, t_min_final)) );\n"
+"	t_max_final = fmin( t_max.z, fmin(t_max.y, fmin(t_max.x, t_max_final)) );\n"
+"	\n"
+"	return (t_min_final <= t_max_final);\n"
+"}\n"
+"__kernel void plbvhRayTraverse(__global b3AabbCL* rigidAabbs,\n"
+"								__global int* rootNodeIndex, \n"
+"								__global int2* internalNodeChildIndices, \n"
+"								__global b3AabbCL* internalNodeAabbs,\n"
+"								__global int2* internalNodeLeafIndexRanges,\n"
+"								__global SortDataCL* mortonCodesAndAabbIndices,\n"
+"								\n"
+"								__global b3RayInfo* rays,\n"
+"								\n"
+"								__global int* out_numRayRigidPairs, \n"
+"								__global int2* out_rayRigidPairs,\n"
+"								int maxRayRigidPairs, int numRays)\n"
+"{\n"
+"	int rayIndex = get_global_id(0);\n"
+"	if(rayIndex >= numRays) return;\n"
+"	\n"
+"	//\n"
+"	b3Vector3 rayFrom = rays[rayIndex].m_from;\n"
+"	b3Vector3 rayTo = rays[rayIndex].m_to;\n"
+"	b3Vector3 rayNormalizedDirection = b3Vector3_normalize(rayTo - rayFrom);\n"
+"	b3Scalar rayLength = b3Sqrt( b3Vector3_length2(rayTo - rayFrom) );\n"
+"	\n"
+"	//\n"
+"	int stack[B3_PLVBH_TRAVERSE_MAX_STACK_SIZE];\n"
+"	\n"
+"	int stackSize = 1;\n"
+"	stack[0] = *rootNodeIndex;\n"
+"	\n"
+"	while(stackSize)\n"
+"	{\n"
+"		int internalOrLeafNodeIndex = stack[ stackSize - 1 ];\n"
+"		--stackSize;\n"
+"		\n"
+"		int isLeaf = isLeafNode(internalOrLeafNodeIndex);	//Internal node if false\n"
+"		int bvhNodeIndex = getIndexWithInternalNodeMarkerRemoved(internalOrLeafNodeIndex);\n"
+"		\n"
+"		//bvhRigidIndex is not used if internal node\n"
+"		int bvhRigidIndex = (isLeaf) ? mortonCodesAndAabbIndices[bvhNodeIndex].m_value : -1;\n"
+"	\n"
+"		b3AabbCL bvhNodeAabb = (isLeaf) ? rigidAabbs[bvhRigidIndex] : internalNodeAabbs[bvhNodeIndex];\n"
+"		if( rayIntersectsAabb(rayFrom, rayLength, rayNormalizedDirection, bvhNodeAabb)  )\n"
+"		{\n"
+"			if(isLeaf)\n"
+"			{\n"
+"				int2 rayRigidPair;\n"
+"				rayRigidPair.x = rayIndex;\n"
+"				rayRigidPair.y = rigidAabbs[bvhRigidIndex].m_minIndices[3];\n"
+"				\n"
+"				int pairIndex = atomic_inc(out_numRayRigidPairs);\n"
+"				if(pairIndex < maxRayRigidPairs) out_rayRigidPairs[pairIndex] = rayRigidPair;\n"
+"			}\n"
+"			\n"
+"			if(!isLeaf)	//Internal node\n"
+"			{\n"
+"				if(stackSize + 2 > B3_PLVBH_TRAVERSE_MAX_STACK_SIZE)\n"
+"				{\n"
+"					//Error\n"
+"				}\n"
+"				else\n"
+"				{\n"
+"					stack[ stackSize++ ] = internalNodeChildIndices[bvhNodeIndex].x;\n"
+"					stack[ stackSize++ ] = internalNodeChildIndices[bvhNodeIndex].y;\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"}\n"
+"__kernel void plbvhLargeAabbAabbTest(__global b3AabbCL* smallAabbs, __global b3AabbCL* largeAabbs, \n"
+"									__global int* out_numPairs, __global int4* out_overlappingPairs, \n"
+"									int maxPairs, int numLargeAabbRigids, int numSmallAabbRigids)\n"
+"{\n"
+"	int smallAabbIndex = get_global_id(0);\n"
+"	if(smallAabbIndex >= numSmallAabbRigids) return;\n"
+"	\n"
+"	b3AabbCL smallAabb = smallAabbs[smallAabbIndex];\n"
+"	for(int i = 0; i < numLargeAabbRigids; ++i)\n"
+"	{\n"
+"		b3AabbCL largeAabb = largeAabbs[i];\n"
+"		if( TestAabbAgainstAabb2(&smallAabb, &largeAabb) )\n"
+"		{\n"
+"			int4 pair;\n"
+"			pair.x = largeAabb.m_minIndices[3];\n"
+"			pair.y = smallAabb.m_minIndices[3];\n"
+"			pair.z = NEW_PAIR_MARKER;\n"
+"			pair.w = NEW_PAIR_MARKER;\n"
+"			\n"
+"			int pairIndex = atomic_inc(out_numPairs);\n"
+"			if(pairIndex < maxPairs) out_overlappingPairs[pairIndex] = pair;\n"
+"		}\n"
+"	}\n"
+"}\n"
+"__kernel void plbvhLargeAabbRayTest(__global b3AabbCL* largeRigidAabbs, __global b3RayInfo* rays,\n"
+"									__global int* out_numRayRigidPairs,  __global int2* out_rayRigidPairs,\n"
+"									int numLargeAabbRigids, int maxRayRigidPairs, int numRays)\n"
+"{\n"
+"	int rayIndex = get_global_id(0);\n"
+"	if(rayIndex >= numRays) return;\n"
+"	\n"
+"	b3Vector3 rayFrom = rays[rayIndex].m_from;\n"
+"	b3Vector3 rayTo = rays[rayIndex].m_to;\n"
+"	b3Vector3 rayNormalizedDirection = b3Vector3_normalize(rayTo - rayFrom);\n"
+"	b3Scalar rayLength = b3Sqrt( b3Vector3_length2(rayTo - rayFrom) );\n"
+"	\n"
+"	for(int i = 0; i < numLargeAabbRigids; ++i)\n"
+"	{\n"
+"		b3AabbCL rigidAabb = largeRigidAabbs[i];\n"
+"		if( rayIntersectsAabb(rayFrom, rayLength, rayNormalizedDirection, rigidAabb) )\n"
+"		{\n"
+"			int2 rayRigidPair;\n"
+"			rayRigidPair.x = rayIndex;\n"
+"			rayRigidPair.y = rigidAabb.m_minIndices[3];\n"
+"			\n"
+"			int pairIndex = atomic_inc(out_numRayRigidPairs);\n"
+"			if(pairIndex < maxRayRigidPairs) out_rayRigidPairs[pairIndex] = rayRigidPair;\n"
+"		}\n"
+"	}\n"
+"}\n"
+"//Set so that it is always greater than the actual common prefixes, and never selected as a parent node.\n"
+"//If there are no duplicates, then the highest common prefix is 32 or 64, depending on the number of bits used for the z-curve.\n"
+"//Duplicate common prefixes increase the highest common prefix at most by the number of bits used to index the leaf node.\n"
+"//Since 32 bit ints are used to index leaf nodes, the max prefix is 64(32 + 32 bit z-curve) or 96(32 + 64 bit z-curve).\n"
+"#define B3_PLBVH_INVALID_COMMON_PREFIX 128\n"
+"#define B3_PLBVH_ROOT_NODE_MARKER -1\n"
+"#define b3Int64 long\n"
+"int computeCommonPrefixLength(b3Int64 i, b3Int64 j) { return (int)clz(i ^ j); }\n"
+"b3Int64 computeCommonPrefix(b3Int64 i, b3Int64 j) \n"
+"{\n"
+"	//This function only needs to return (i & j) in order for the algorithm to work,\n"
+"	//but it may help with debugging to mask out the lower bits.\n"
+"	b3Int64 commonPrefixLength = (b3Int64)computeCommonPrefixLength(i, j);\n"
+"	b3Int64 sharedBits = i & j;\n"
+"	b3Int64 bitmask = ((b3Int64)(~0)) << (64 - commonPrefixLength);	//Set all bits after the common prefix to 0\n"
+"	\n"
+"	return sharedBits & bitmask;\n"
+"}\n"
+"//Same as computeCommonPrefixLength(), but allows for prefixes with different lengths\n"
+"int getSharedPrefixLength(b3Int64 prefixA, int prefixLengthA, b3Int64 prefixB, int prefixLengthB)\n"
+"{\n"
+"	return b3Min( computeCommonPrefixLength(prefixA, prefixB), b3Min(prefixLengthA, prefixLengthB) );\n"
+"}\n"
+"__kernel void computeAdjacentPairCommonPrefix(__global SortDataCL* mortonCodesAndAabbIndices,\n"
+"											__global b3Int64* out_commonPrefixes,\n"
+"											__global int* out_commonPrefixLengths,\n"
+"											int numInternalNodes)\n"
+"{\n"
+"	int internalNodeIndex = get_global_id(0);\n"
+"	if (internalNodeIndex >= numInternalNodes) return;\n"
+"	\n"
+"	//Here, (internalNodeIndex + 1) is never out of bounds since it is a leaf node index,\n"
+"	//and the number of internal nodes is always numLeafNodes - 1\n"
+"	int leftLeafIndex = internalNodeIndex;\n"
+"	int rightLeafIndex = internalNodeIndex + 1;\n"
+"	\n"
+"	int leftLeafMortonCode = mortonCodesAndAabbIndices[leftLeafIndex].m_key;\n"
+"	int rightLeafMortonCode = mortonCodesAndAabbIndices[rightLeafIndex].m_key;\n"
+"	\n"
+"	//Binary radix tree construction algorithm does not work if there are duplicate morton codes.\n"
+"	//Append the index of each leaf node to each morton code so that there are no duplicates.\n"
+"	//The algorithm also requires that the morton codes are sorted in ascending order; this requirement\n"
+"	//is also satisfied with this method, as (leftLeafIndex < rightLeafIndex) is always true.\n"
+"	//\n"
+"	//upsample(a, b) == ( ((b3Int64)a) << 32) | b\n"
+"	b3Int64 nonduplicateLeftMortonCode = upsample(leftLeafMortonCode, leftLeafIndex);\n"
+"	b3Int64 nonduplicateRightMortonCode = upsample(rightLeafMortonCode, rightLeafIndex);\n"
+"	\n"
+"	out_commonPrefixes[internalNodeIndex] = computeCommonPrefix(nonduplicateLeftMortonCode, nonduplicateRightMortonCode);\n"
+"	out_commonPrefixLengths[internalNodeIndex] = computeCommonPrefixLength(nonduplicateLeftMortonCode, nonduplicateRightMortonCode);\n"
+"}\n"
+"__kernel void buildBinaryRadixTreeLeafNodes(__global int* commonPrefixLengths, __global int* out_leafNodeParentNodes,\n"
+"											__global int2* out_childNodes, int numLeafNodes)\n"
+"{\n"
+"	int leafNodeIndex = get_global_id(0);\n"
+"	if (leafNodeIndex >= numLeafNodes) return;\n"
+"	\n"
+"	int numInternalNodes = numLeafNodes - 1;\n"
+"	\n"
+"	int leftSplitIndex = leafNodeIndex - 1;\n"
+"	int rightSplitIndex = leafNodeIndex;\n"
+"	\n"
+"	int leftCommonPrefix = (leftSplitIndex >= 0) ? commonPrefixLengths[leftSplitIndex] : B3_PLBVH_INVALID_COMMON_PREFIX;\n"
+"	int rightCommonPrefix = (rightSplitIndex < numInternalNodes) ? commonPrefixLengths[rightSplitIndex] : B3_PLBVH_INVALID_COMMON_PREFIX;\n"
+"	\n"
+"	//Parent node is the highest adjacent common prefix that is lower than the node's common prefix\n"
+"	//Leaf nodes are considered as having the highest common prefix\n"
+"	int isLeftHigherCommonPrefix = (leftCommonPrefix > rightCommonPrefix);\n"
+"	\n"
+"	//Handle cases for the edge nodes; the first and last node\n"
+"	//For leaf nodes, leftCommonPrefix and rightCommonPrefix should never both be B3_PLBVH_INVALID_COMMON_PREFIX\n"
+"	if(leftCommonPrefix == B3_PLBVH_INVALID_COMMON_PREFIX) isLeftHigherCommonPrefix = false;\n"
+"	if(rightCommonPrefix == B3_PLBVH_INVALID_COMMON_PREFIX) isLeftHigherCommonPrefix = true;\n"
+"	\n"
+"	int parentNodeIndex = (isLeftHigherCommonPrefix) ? leftSplitIndex : rightSplitIndex;\n"
+"	out_leafNodeParentNodes[leafNodeIndex] = parentNodeIndex;\n"
+"	\n"
+"	int isRightChild = (isLeftHigherCommonPrefix);	//If the left node is the parent, then this node is its right child and vice versa\n"
+"	\n"
+"	//out_childNodesAsInt[0] == int2.x == left child\n"
+"	//out_childNodesAsInt[1] == int2.y == right child\n"
+"	int isLeaf = 1;\n"
+"	__global int* out_childNodesAsInt = (__global int*)(&out_childNodes[parentNodeIndex]);\n"
+"	out_childNodesAsInt[isRightChild] = getIndexWithInternalNodeMarkerSet(isLeaf, leafNodeIndex);\n"
+"}\n"
+"__kernel void buildBinaryRadixTreeInternalNodes(__global b3Int64* commonPrefixes, __global int* commonPrefixLengths,\n"
+"												__global int2* out_childNodes,\n"
+"												__global int* out_internalNodeParentNodes, __global int* out_rootNodeIndex,\n"
+"												int numInternalNodes)\n"
+"{\n"
+"	int internalNodeIndex = get_group_id(0) * get_local_size(0) + get_local_id(0);\n"
+"	if(internalNodeIndex >= numInternalNodes) return;\n"
+"	\n"
+"	b3Int64 nodePrefix = commonPrefixes[internalNodeIndex];\n"
+"	int nodePrefixLength = commonPrefixLengths[internalNodeIndex];\n"
+"	\n"
+"//#define USE_LINEAR_SEARCH\n"
+"#ifdef USE_LINEAR_SEARCH\n"
+"	int leftIndex = -1;\n"
+"	int rightIndex = -1;\n"
+"	\n"
+"	//Find nearest element to left with a lower common prefix\n"
+"	for(int i = internalNodeIndex - 1; i >= 0; --i)\n"
+"	{\n"
+"		int nodeLeftSharedPrefixLength = getSharedPrefixLength(nodePrefix, nodePrefixLength, commonPrefixes[i], commonPrefixLengths[i]);\n"
+"		if(nodeLeftSharedPrefixLength < nodePrefixLength)\n"
+"		{\n"
+"			leftIndex = i;\n"
+"			break;\n"
+"		}\n"
+"	}\n"
+"	\n"
+"	//Find nearest element to right with a lower common prefix\n"
+"	for(int i = internalNodeIndex + 1; i < numInternalNodes; ++i)\n"
+"	{\n"
+"		int nodeRightSharedPrefixLength = getSharedPrefixLength(nodePrefix, nodePrefixLength, commonPrefixes[i], commonPrefixLengths[i]);\n"
+"		if(nodeRightSharedPrefixLength < nodePrefixLength)\n"
+"		{\n"
+"			rightIndex = i;\n"
+"			break;\n"
+"		}\n"
+"	}\n"
+"	\n"
+"#else //Use binary search\n"
+"	//Find nearest element to left with a lower common prefix\n"
+"	int leftIndex = -1;\n"
+"	{\n"
+"		int lower = 0;\n"
+"		int upper = internalNodeIndex - 1;\n"
+"		\n"
+"		while(lower <= upper)\n"
+"		{\n"
+"			int mid = (lower + upper) / 2;\n"
+"			b3Int64 midPrefix = commonPrefixes[mid];\n"
+"			int midPrefixLength = commonPrefixLengths[mid];\n"
+"			\n"
+"			int nodeMidSharedPrefixLength = getSharedPrefixLength(nodePrefix, nodePrefixLength, midPrefix, midPrefixLength);\n"
+"			if(nodeMidSharedPrefixLength < nodePrefixLength) \n"
+"			{\n"
+"				int right = mid + 1;\n"
+"				if(right < internalNodeIndex)\n"
+"				{\n"
+"					b3Int64 rightPrefix = commonPrefixes[right];\n"
+"					int rightPrefixLength = commonPrefixLengths[right];\n"
+"					\n"
+"					int nodeRightSharedPrefixLength = getSharedPrefixLength(nodePrefix, nodePrefixLength, rightPrefix, rightPrefixLength);\n"
+"					if(nodeRightSharedPrefixLength < nodePrefixLength) \n"
+"					{\n"
+"						lower = right;\n"
+"						leftIndex = right;\n"
+"					}\n"
+"					else \n"
+"					{\n"
+"						leftIndex = mid;\n"
+"						break;\n"
+"					}\n"
+"				}\n"
+"				else \n"
+"				{\n"
+"					leftIndex = mid;\n"
+"					break;\n"
+"				}\n"
+"			}\n"
+"			else upper = mid - 1;\n"
+"		}\n"
+"	}\n"
+"	\n"
+"	//Find nearest element to right with a lower common prefix\n"
+"	int rightIndex = -1;\n"
+"	{\n"
+"		int lower = internalNodeIndex + 1;\n"
+"		int upper = numInternalNodes - 1;\n"
+"		\n"
+"		while(lower <= upper)\n"
+"		{\n"
+"			int mid = (lower + upper) / 2;\n"
+"			b3Int64 midPrefix = commonPrefixes[mid];\n"
+"			int midPrefixLength = commonPrefixLengths[mid];\n"
+"			\n"
+"			int nodeMidSharedPrefixLength = getSharedPrefixLength(nodePrefix, nodePrefixLength, midPrefix, midPrefixLength);\n"
+"			if(nodeMidSharedPrefixLength < nodePrefixLength) \n"
+"			{\n"
+"				int left = mid - 1;\n"
+"				if(left > internalNodeIndex)\n"
+"				{\n"
+"					b3Int64 leftPrefix = commonPrefixes[left];\n"
+"					int leftPrefixLength = commonPrefixLengths[left];\n"
+"				\n"
+"					int nodeLeftSharedPrefixLength = getSharedPrefixLength(nodePrefix, nodePrefixLength, leftPrefix, leftPrefixLength);\n"
+"					if(nodeLeftSharedPrefixLength < nodePrefixLength) \n"
+"					{\n"
+"						upper = left;\n"
+"						rightIndex = left;\n"
+"					}\n"
+"					else \n"
+"					{\n"
+"						rightIndex = mid;\n"
+"						break;\n"
+"					}\n"
+"				}\n"
+"				else \n"
+"				{\n"
+"					rightIndex = mid;\n"
+"					break;\n"
+"				}\n"
+"			}\n"
+"			else lower = mid + 1;\n"
+"		}\n"
+"	}\n"
+"#endif\n"
+"	\n"
+"	//Select parent\n"
+"	{\n"
+"		int leftPrefixLength = (leftIndex != -1) ? commonPrefixLengths[leftIndex] : B3_PLBVH_INVALID_COMMON_PREFIX;\n"
+"		int rightPrefixLength =  (rightIndex != -1) ? commonPrefixLengths[rightIndex] : B3_PLBVH_INVALID_COMMON_PREFIX;\n"
+"		\n"
+"		int isLeftHigherPrefixLength = (leftPrefixLength > rightPrefixLength);\n"
+"		\n"
+"		if(leftPrefixLength == B3_PLBVH_INVALID_COMMON_PREFIX) isLeftHigherPrefixLength = false;\n"
+"		else if(rightPrefixLength == B3_PLBVH_INVALID_COMMON_PREFIX) isLeftHigherPrefixLength = true;\n"
+"		\n"
+"		int parentNodeIndex = (isLeftHigherPrefixLength) ? leftIndex : rightIndex;\n"
+"		\n"
+"		int isRootNode = (leftIndex == -1 && rightIndex == -1);\n"
+"		out_internalNodeParentNodes[internalNodeIndex] = (!isRootNode) ? parentNodeIndex : B3_PLBVH_ROOT_NODE_MARKER;\n"
+"		\n"
+"		int isLeaf = 0;\n"
+"		if(!isRootNode)\n"
+"		{\n"
+"			int isRightChild = (isLeftHigherPrefixLength);	//If the left node is the parent, then this node is its right child and vice versa\n"
+"			\n"
+"			//out_childNodesAsInt[0] == int2.x == left child\n"
+"			//out_childNodesAsInt[1] == int2.y == right child\n"
+"			__global int* out_childNodesAsInt = (__global int*)(&out_childNodes[parentNodeIndex]);\n"
+"			out_childNodesAsInt[isRightChild] = getIndexWithInternalNodeMarkerSet(isLeaf, internalNodeIndex);\n"
+"		}\n"
+"		else *out_rootNodeIndex = getIndexWithInternalNodeMarkerSet(isLeaf, internalNodeIndex);\n"
+"	}\n"
+"}\n"
+"__kernel void findDistanceFromRoot(__global int* rootNodeIndex, __global int* internalNodeParentNodes,\n"
+"									__global int* out_maxDistanceFromRoot, __global int* out_distanceFromRoot, int numInternalNodes)\n"
+"{\n"
+"	if( get_global_id(0) == 0 ) atomic_xchg(out_maxDistanceFromRoot, 0);\n"
+"	int internalNodeIndex = get_global_id(0);\n"
+"	if(internalNodeIndex >= numInternalNodes) return;\n"
+"	\n"
+"	//\n"
+"	int distanceFromRoot = 0;\n"
+"	{\n"
+"		int parentIndex = internalNodeParentNodes[internalNodeIndex];\n"
+"		while(parentIndex != B3_PLBVH_ROOT_NODE_MARKER)\n"
+"		{\n"
+"			parentIndex = internalNodeParentNodes[parentIndex];\n"
+"			++distanceFromRoot;\n"
+"		}\n"
+"	}\n"
+"	out_distanceFromRoot[internalNodeIndex] = distanceFromRoot;\n"
+"	\n"
+"	//\n"
+"	__local int localMaxDistanceFromRoot;\n"
+"	if( get_local_id(0) == 0 ) localMaxDistanceFromRoot = 0;\n"
+"	barrier(CLK_LOCAL_MEM_FENCE);\n"
+"	\n"
+"	atomic_max(&localMaxDistanceFromRoot, distanceFromRoot);\n"
+"	barrier(CLK_LOCAL_MEM_FENCE);\n"
+"	\n"
+"	if( get_local_id(0) == 0 ) atomic_max(out_maxDistanceFromRoot, localMaxDistanceFromRoot);\n"
+"}\n"
+"__kernel void buildBinaryRadixTreeAabbsRecursive(__global int* distanceFromRoot, __global SortDataCL* mortonCodesAndAabbIndices,\n"
+"												__global int2* childNodes,\n"
+"												__global b3AabbCL* leafNodeAabbs, __global b3AabbCL* internalNodeAabbs,\n"
+"												int maxDistanceFromRoot, int processedDistance, int numInternalNodes)\n"
+"{\n"
+"	int internalNodeIndex = get_global_id(0);\n"
+"	if(internalNodeIndex >= numInternalNodes) return;\n"
+"	\n"
+"	int distance = distanceFromRoot[internalNodeIndex];\n"
+"	\n"
+"	if(distance == processedDistance)\n"
+"	{\n"
+"		int leftChildIndex = childNodes[internalNodeIndex].x;\n"
+"		int rightChildIndex = childNodes[internalNodeIndex].y;\n"
+"		\n"
+"		int isLeftChildLeaf = isLeafNode(leftChildIndex);\n"
+"		int isRightChildLeaf = isLeafNode(rightChildIndex);\n"
+"		\n"
+"		leftChildIndex = getIndexWithInternalNodeMarkerRemoved(leftChildIndex);\n"
+"		rightChildIndex = getIndexWithInternalNodeMarkerRemoved(rightChildIndex);\n"
+"		\n"
+"		//leftRigidIndex/rightRigidIndex is not used if internal node\n"
+"		int leftRigidIndex = (isLeftChildLeaf) ? mortonCodesAndAabbIndices[leftChildIndex].m_value : -1;\n"
+"		int rightRigidIndex = (isRightChildLeaf) ? mortonCodesAndAabbIndices[rightChildIndex].m_value : -1;\n"
+"		\n"
+"		b3AabbCL leftChildAabb = (isLeftChildLeaf) ? leafNodeAabbs[leftRigidIndex] : internalNodeAabbs[leftChildIndex];\n"
+"		b3AabbCL rightChildAabb = (isRightChildLeaf) ? leafNodeAabbs[rightRigidIndex] : internalNodeAabbs[rightChildIndex];\n"
+"		\n"
+"		b3AabbCL mergedAabb;\n"
+"		mergedAabb.m_min = b3Min(leftChildAabb.m_min, rightChildAabb.m_min);\n"
+"		mergedAabb.m_max = b3Max(leftChildAabb.m_max, rightChildAabb.m_max);\n"
+"		internalNodeAabbs[internalNodeIndex] = mergedAabb;\n"
+"	}\n"
+"}\n"
+"__kernel void findLeafIndexRanges(__global int2* internalNodeChildNodes, __global int2* out_leafIndexRanges, int numInternalNodes)\n"
+"{\n"
+"	int internalNodeIndex = get_global_id(0);\n"
+"	if(internalNodeIndex >= numInternalNodes) return;\n"
+"	\n"
+"	int numLeafNodes = numInternalNodes + 1;\n"
+"	\n"
+"	int2 childNodes = internalNodeChildNodes[internalNodeIndex];\n"
+"	\n"
+"	int2 leafIndexRange;	//x == min leaf index, y == max leaf index\n"
+"	\n"
+"	//Find lowest leaf index covered by this internal node\n"
+"	{\n"
+"		int lowestIndex = childNodes.x;		//childNodes.x == Left child\n"
+"		while( !isLeafNode(lowestIndex) ) lowestIndex = internalNodeChildNodes[ getIndexWithInternalNodeMarkerRemoved(lowestIndex) ].x;\n"
+"		leafIndexRange.x = lowestIndex;\n"
+"	}\n"
+"	\n"
+"	//Find highest leaf index covered by this internal node\n"
+"	{\n"
+"		int highestIndex = childNodes.y;	//childNodes.y == Right child\n"
+"		while( !isLeafNode(highestIndex) ) highestIndex = internalNodeChildNodes[ getIndexWithInternalNodeMarkerRemoved(highestIndex) ].y;\n"
+"		leafIndexRange.y = highestIndex;\n"
+"	}\n"
+"	\n"
+"	//\n"
+"	out_leafIndexRanges[internalNodeIndex] = leafIndexRange;\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/sap.cl b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/sap.cl
new file mode 100644
index 0000000000..93f77a6433
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/sap.cl
@@ -0,0 +1,389 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+#define NEW_PAIR_MARKER -1
+
+typedef struct 
+{
+	union
+	{
+		float4	m_min;
+		float   m_minElems[4];
+		int			m_minIndices[4];
+	};
+	union
+	{
+		float4	m_max;
+		float   m_maxElems[4];
+		int			m_maxIndices[4];
+	};
+} btAabbCL;
+
+
+/// conservative test for overlap between two aabbs
+bool TestAabbAgainstAabb2(const btAabbCL* aabb1, __local const btAabbCL* aabb2);
+bool TestAabbAgainstAabb2(const btAabbCL* aabb1, __local const btAabbCL* aabb2)
+{
+	bool overlap = true;
+	overlap = (aabb1->m_min.x > aabb2->m_max.x || aabb1->m_max.x < aabb2->m_min.x) ? false : overlap;
+	overlap = (aabb1->m_min.z > aabb2->m_max.z || aabb1->m_max.z < aabb2->m_min.z) ? false : overlap;
+	overlap = (aabb1->m_min.y > aabb2->m_max.y || aabb1->m_max.y < aabb2->m_min.y) ? false : overlap;
+	return overlap;
+}
+bool TestAabbAgainstAabb2GlobalGlobal(__global const btAabbCL* aabb1, __global const btAabbCL* aabb2);
+bool TestAabbAgainstAabb2GlobalGlobal(__global const btAabbCL* aabb1, __global const btAabbCL* aabb2)
+{
+	bool overlap = true;
+	overlap = (aabb1->m_min.x > aabb2->m_max.x || aabb1->m_max.x < aabb2->m_min.x) ? false : overlap;
+	overlap = (aabb1->m_min.z > aabb2->m_max.z || aabb1->m_max.z < aabb2->m_min.z) ? false : overlap;
+	overlap = (aabb1->m_min.y > aabb2->m_max.y || aabb1->m_max.y < aabb2->m_min.y) ? false : overlap;
+	return overlap;
+}
+
+bool TestAabbAgainstAabb2Global(const btAabbCL* aabb1, __global const btAabbCL* aabb2);
+bool TestAabbAgainstAabb2Global(const btAabbCL* aabb1, __global const btAabbCL* aabb2)
+{
+	bool overlap = true;
+	overlap = (aabb1->m_min.x > aabb2->m_max.x || aabb1->m_max.x < aabb2->m_min.x) ? false : overlap;
+	overlap = (aabb1->m_min.z > aabb2->m_max.z || aabb1->m_max.z < aabb2->m_min.z) ? false : overlap;
+	overlap = (aabb1->m_min.y > aabb2->m_max.y || aabb1->m_max.y < aabb2->m_min.y) ? false : overlap;
+	return overlap;
+}
+
+
+__kernel void   computePairsKernelTwoArrays( __global const btAabbCL* unsortedAabbs, __global const int* unsortedAabbMapping,  __global const int* unsortedAabbMapping2, volatile __global int4* pairsOut,volatile  __global int* pairCount, int numUnsortedAabbs, int numUnSortedAabbs2, int axis, int maxPairs)
+{
+	int i = get_global_id(0);
+	if (i>=numUnsortedAabbs)
+		return;
+
+	int j = get_global_id(1);
+	if (j>=numUnSortedAabbs2)
+		return;
+
+
+	__global const btAabbCL* unsortedAabbPtr = &unsortedAabbs[unsortedAabbMapping[i]];
+	__global const btAabbCL* unsortedAabbPtr2 = &unsortedAabbs[unsortedAabbMapping2[j]];
+
+	if (TestAabbAgainstAabb2GlobalGlobal(unsortedAabbPtr,unsortedAabbPtr2))
+	{
+		int4 myPair;
+		
+		int xIndex = unsortedAabbPtr[0].m_minIndices[3];
+		int yIndex = unsortedAabbPtr2[0].m_minIndices[3];
+		if (xIndex>yIndex)
+		{
+			int tmp = xIndex;
+			xIndex=yIndex;
+			yIndex=tmp;
+		}
+		
+		myPair.x = xIndex;
+		myPair.y = yIndex;
+		myPair.z = NEW_PAIR_MARKER;
+		myPair.w = NEW_PAIR_MARKER;
+
+
+		int curPair = atomic_inc (pairCount);
+		if (curPair<maxPairs)
+		{
+				pairsOut[curPair] = myPair; //flush to main memory
+		}
+	}
+}
+
+
+
+__kernel void   computePairsKernelBruteForce( __global const btAabbCL* aabbs, volatile __global int4* pairsOut,volatile  __global int* pairCount, int numObjects, int axis, int maxPairs)
+{
+	int i = get_global_id(0);
+	if (i>=numObjects)
+		return;
+	for (int j=i+1;j<numObjects;j++)
+	{
+		if (TestAabbAgainstAabb2GlobalGlobal(&aabbs[i],&aabbs[j]))
+		{
+			int4 myPair;
+			myPair.x = aabbs[i].m_minIndices[3];
+			myPair.y = aabbs[j].m_minIndices[3];
+			myPair.z = NEW_PAIR_MARKER;
+			myPair.w = NEW_PAIR_MARKER;
+
+			int curPair = atomic_inc (pairCount);
+			if (curPair<maxPairs)
+			{
+					pairsOut[curPair] = myPair; //flush to main memory
+			}
+		}
+	}
+}
+
+__kernel void   computePairsKernelOriginal( __global const btAabbCL* aabbs, volatile __global int4* pairsOut,volatile  __global int* pairCount, int numObjects, int axis, int maxPairs)
+{
+	int i = get_global_id(0);
+	if (i>=numObjects)
+		return;
+	for (int j=i+1;j<numObjects;j++)
+	{
+  	if(aabbs[i].m_maxElems[axis] < (aabbs[j].m_minElems[axis])) 
+		{
+			break;
+		}
+		if (TestAabbAgainstAabb2GlobalGlobal(&aabbs[i],&aabbs[j]))
+		{
+			int4 myPair;
+			myPair.x = aabbs[i].m_minIndices[3];
+			myPair.y = aabbs[j].m_minIndices[3];
+			myPair.z = NEW_PAIR_MARKER;
+			myPair.w = NEW_PAIR_MARKER;
+
+			int curPair = atomic_inc (pairCount);
+			if (curPair<maxPairs)
+			{
+					pairsOut[curPair] = myPair; //flush to main memory
+			}
+		}
+	}
+}
+
+
+
+
+__kernel void   computePairsKernelBarrier( __global const btAabbCL* aabbs, volatile __global int4* pairsOut,volatile  __global int* pairCount, int numObjects, int axis, int maxPairs)
+{
+	int i = get_global_id(0);
+	int localId = get_local_id(0);
+
+	__local int numActiveWgItems[1];
+	__local int breakRequest[1];
+
+	if (localId==0)
+	{
+		numActiveWgItems[0] = 0;
+		breakRequest[0] = 0;
+	}
+	barrier(CLK_LOCAL_MEM_FENCE);
+	atomic_inc(numActiveWgItems);
+	barrier(CLK_LOCAL_MEM_FENCE);
+	int localBreak = 0;
+
+	int j=i+1;
+	do
+	{
+		barrier(CLK_LOCAL_MEM_FENCE);
+	
+		if (j<numObjects)
+		{
+	  	if(aabbs[i].m_maxElems[axis] < (aabbs[j].m_minElems[axis])) 
+			{
+				if (!localBreak)
+				{
+					atomic_inc(breakRequest);
+					localBreak = 1;
+				}
+			}
+		}
+		
+		barrier(CLK_LOCAL_MEM_FENCE);
+		
+		if (j>=numObjects && !localBreak)
+		{
+			atomic_inc(breakRequest);
+			localBreak = 1;
+		}
+		barrier(CLK_LOCAL_MEM_FENCE);
+		
+		if (!localBreak)
+		{
+			if (TestAabbAgainstAabb2GlobalGlobal(&aabbs[i],&aabbs[j]))
+			{
+				int4 myPair;
+				myPair.x = aabbs[i].m_minIndices[3];
+				myPair.y = aabbs[j].m_minIndices[3];
+				myPair.z = NEW_PAIR_MARKER;
+				myPair.w = NEW_PAIR_MARKER;
+
+				int curPair = atomic_inc (pairCount);
+				if (curPair<maxPairs)
+				{
+						pairsOut[curPair] = myPair; //flush to main memory
+				}
+			}
+		}
+		j++;
+
+	} while (breakRequest[0]<numActiveWgItems[0]);
+}
+
+
+__kernel void   computePairsKernelLocalSharedMemory( __global const btAabbCL* aabbs, volatile __global int4* pairsOut,volatile  __global int* pairCount, int numObjects, int axis, int maxPairs)
+{
+	int i = get_global_id(0);
+	int localId = get_local_id(0);
+
+	__local int numActiveWgItems[1];
+	__local int breakRequest[1];
+	__local btAabbCL localAabbs[128];// = aabbs[i];
+	
+	btAabbCL myAabb;
+	
+	myAabb = (i<numObjects)? aabbs[i]:aabbs[0];
+	float testValue = 	myAabb.m_maxElems[axis];
+	
+	if (localId==0)
+	{
+		numActiveWgItems[0] = 0;
+		breakRequest[0] = 0;
+	}
+	int localCount=0;
+	int block=0;
+	localAabbs[localId] = (i+block)<numObjects? aabbs[i+block] : aabbs[0];
+	localAabbs[localId+64] = (i+block+64)<numObjects? aabbs[i+block+64]: aabbs[0];
+	
+	barrier(CLK_LOCAL_MEM_FENCE);
+	atomic_inc(numActiveWgItems);
+	barrier(CLK_LOCAL_MEM_FENCE);
+	int localBreak = 0;
+	
+	int j=i+1;
+	do
+	{
+		barrier(CLK_LOCAL_MEM_FENCE);
+	
+		if (j<numObjects)
+		{
+	  	if(testValue < (localAabbs[localCount+localId+1].m_minElems[axis])) 
+			{
+				if (!localBreak)
+				{
+					atomic_inc(breakRequest);
+					localBreak = 1;
+				}
+			}
+		}
+		
+		barrier(CLK_LOCAL_MEM_FENCE);
+		
+		if (j>=numObjects && !localBreak)
+		{
+			atomic_inc(breakRequest);
+			localBreak = 1;
+		}
+		barrier(CLK_LOCAL_MEM_FENCE);
+		
+		if (!localBreak)
+		{
+			if (TestAabbAgainstAabb2(&myAabb,&localAabbs[localCount+localId+1]))
+			{
+				int4 myPair;
+				myPair.x = myAabb.m_minIndices[3];
+				myPair.y = localAabbs[localCount+localId+1].m_minIndices[3];
+				myPair.z = NEW_PAIR_MARKER;
+				myPair.w = NEW_PAIR_MARKER;
+
+				int curPair = atomic_inc (pairCount);
+				if (curPair<maxPairs)
+				{
+						pairsOut[curPair] = myPair; //flush to main memory
+				}
+			}
+		}
+		
+		barrier(CLK_LOCAL_MEM_FENCE);
+
+		localCount++;
+		if (localCount==64)
+		{
+			localCount = 0;
+			block+=64;			
+			localAabbs[localId] = ((i+block)<numObjects) ? aabbs[i+block] : aabbs[0];
+			localAabbs[localId+64] = ((i+64+block)<numObjects) ? aabbs[i+block+64] : aabbs[0];
+		}
+		j++;
+		
+	} while (breakRequest[0]<numActiveWgItems[0]);
+	
+}
+
+
+
+
+//http://stereopsis.com/radix.html
+unsigned int FloatFlip(float fl);
+unsigned int FloatFlip(float fl)
+{
+	unsigned int f = *(unsigned int*)&fl;
+	unsigned int mask = -(int)(f >> 31) | 0x80000000;
+	return f ^ mask;
+}
+float IFloatFlip(unsigned int f);
+float IFloatFlip(unsigned int f)
+{
+	unsigned int mask = ((f >> 31) - 1) | 0x80000000;
+	unsigned int fl = f ^ mask;
+	return *(float*)&fl;
+}
+
+
+
+
+__kernel void   copyAabbsKernel( __global const btAabbCL* allAabbs, __global btAabbCL* destAabbs, int numObjects)
+{
+	int i = get_global_id(0);
+	if (i>=numObjects)
+		return;
+	int src = destAabbs[i].m_maxIndices[3];
+	destAabbs[i] = allAabbs[src];
+	destAabbs[i].m_maxIndices[3] = src;
+}
+
+
+__kernel void   flipFloatKernel( __global const btAabbCL* allAabbs, __global const int* smallAabbMapping, __global int2* sortData, int numObjects, int axis)
+{
+	int i = get_global_id(0);
+	if (i>=numObjects)
+		return;
+	
+	
+	sortData[i].x = FloatFlip(allAabbs[smallAabbMapping[i]].m_minElems[axis]);
+	sortData[i].y = i;
+		
+}
+
+
+__kernel void   scatterKernel( __global const btAabbCL* allAabbs, __global const int* smallAabbMapping, volatile __global const int2* sortData, __global btAabbCL* sortedAabbs, int numObjects)
+{
+	int i = get_global_id(0);
+	if (i>=numObjects)
+		return;
+	
+	sortedAabbs[i] = allAabbs[smallAabbMapping[sortData[i].y]];
+}
+
+
+
+__kernel void   prepareSumVarianceKernel( __global const btAabbCL* allAabbs, __global const int* smallAabbMapping, __global float4* sum, __global float4* sum2,int numAabbs)
+{
+	int i = get_global_id(0);
+	if (i>=numAabbs)
+		return;
+	
+	btAabbCL smallAabb = allAabbs[smallAabbMapping[i]];
+	
+	float4 s;
+	s = (smallAabb.m_max+smallAabb.m_min)*0.5f;
+	sum[i]=s;
+	sum2[i]=s*s;	
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/sapKernels.h b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/sapKernels.h
new file mode 100644
index 0000000000..04d40fcf26
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/BroadphaseCollision/kernels/sapKernels.h
@@ -0,0 +1,342 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* sapCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Erwin Coumans\n"
+"#define NEW_PAIR_MARKER -1\n"
+"typedef struct \n"
+"{\n"
+"	union\n"
+"	{\n"
+"		float4	m_min;\n"
+"		float   m_minElems[4];\n"
+"		int			m_minIndices[4];\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float4	m_max;\n"
+"		float   m_maxElems[4];\n"
+"		int			m_maxIndices[4];\n"
+"	};\n"
+"} btAabbCL;\n"
+"/// conservative test for overlap between two aabbs\n"
+"bool TestAabbAgainstAabb2(const btAabbCL* aabb1, __local const btAabbCL* aabb2);\n"
+"bool TestAabbAgainstAabb2(const btAabbCL* aabb1, __local const btAabbCL* aabb2)\n"
+"{\n"
+"	bool overlap = true;\n"
+"	overlap = (aabb1->m_min.x > aabb2->m_max.x || aabb1->m_max.x < aabb2->m_min.x) ? false : overlap;\n"
+"	overlap = (aabb1->m_min.z > aabb2->m_max.z || aabb1->m_max.z < aabb2->m_min.z) ? false : overlap;\n"
+"	overlap = (aabb1->m_min.y > aabb2->m_max.y || aabb1->m_max.y < aabb2->m_min.y) ? false : overlap;\n"
+"	return overlap;\n"
+"}\n"
+"bool TestAabbAgainstAabb2GlobalGlobal(__global const btAabbCL* aabb1, __global const btAabbCL* aabb2);\n"
+"bool TestAabbAgainstAabb2GlobalGlobal(__global const btAabbCL* aabb1, __global const btAabbCL* aabb2)\n"
+"{\n"
+"	bool overlap = true;\n"
+"	overlap = (aabb1->m_min.x > aabb2->m_max.x || aabb1->m_max.x < aabb2->m_min.x) ? false : overlap;\n"
+"	overlap = (aabb1->m_min.z > aabb2->m_max.z || aabb1->m_max.z < aabb2->m_min.z) ? false : overlap;\n"
+"	overlap = (aabb1->m_min.y > aabb2->m_max.y || aabb1->m_max.y < aabb2->m_min.y) ? false : overlap;\n"
+"	return overlap;\n"
+"}\n"
+"bool TestAabbAgainstAabb2Global(const btAabbCL* aabb1, __global const btAabbCL* aabb2);\n"
+"bool TestAabbAgainstAabb2Global(const btAabbCL* aabb1, __global const btAabbCL* aabb2)\n"
+"{\n"
+"	bool overlap = true;\n"
+"	overlap = (aabb1->m_min.x > aabb2->m_max.x || aabb1->m_max.x < aabb2->m_min.x) ? false : overlap;\n"
+"	overlap = (aabb1->m_min.z > aabb2->m_max.z || aabb1->m_max.z < aabb2->m_min.z) ? false : overlap;\n"
+"	overlap = (aabb1->m_min.y > aabb2->m_max.y || aabb1->m_max.y < aabb2->m_min.y) ? false : overlap;\n"
+"	return overlap;\n"
+"}\n"
+"__kernel void   computePairsKernelTwoArrays( __global const btAabbCL* unsortedAabbs, __global const int* unsortedAabbMapping,  __global const int* unsortedAabbMapping2, volatile __global int4* pairsOut,volatile  __global int* pairCount, int numUnsortedAabbs, int numUnSortedAabbs2, int axis, int maxPairs)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numUnsortedAabbs)\n"
+"		return;\n"
+"	int j = get_global_id(1);\n"
+"	if (j>=numUnSortedAabbs2)\n"
+"		return;\n"
+"	__global const btAabbCL* unsortedAabbPtr = &unsortedAabbs[unsortedAabbMapping[i]];\n"
+"	__global const btAabbCL* unsortedAabbPtr2 = &unsortedAabbs[unsortedAabbMapping2[j]];\n"
+"	if (TestAabbAgainstAabb2GlobalGlobal(unsortedAabbPtr,unsortedAabbPtr2))\n"
+"	{\n"
+"		int4 myPair;\n"
+"		\n"
+"		int xIndex = unsortedAabbPtr[0].m_minIndices[3];\n"
+"		int yIndex = unsortedAabbPtr2[0].m_minIndices[3];\n"
+"		if (xIndex>yIndex)\n"
+"		{\n"
+"			int tmp = xIndex;\n"
+"			xIndex=yIndex;\n"
+"			yIndex=tmp;\n"
+"		}\n"
+"		\n"
+"		myPair.x = xIndex;\n"
+"		myPair.y = yIndex;\n"
+"		myPair.z = NEW_PAIR_MARKER;\n"
+"		myPair.w = NEW_PAIR_MARKER;\n"
+"		int curPair = atomic_inc (pairCount);\n"
+"		if (curPair<maxPairs)\n"
+"		{\n"
+"				pairsOut[curPair] = myPair; //flush to main memory\n"
+"		}\n"
+"	}\n"
+"}\n"
+"__kernel void   computePairsKernelBruteForce( __global const btAabbCL* aabbs, volatile __global int4* pairsOut,volatile  __global int* pairCount, int numObjects, int axis, int maxPairs)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numObjects)\n"
+"		return;\n"
+"	for (int j=i+1;j<numObjects;j++)\n"
+"	{\n"
+"		if (TestAabbAgainstAabb2GlobalGlobal(&aabbs[i],&aabbs[j]))\n"
+"		{\n"
+"			int4 myPair;\n"
+"			myPair.x = aabbs[i].m_minIndices[3];\n"
+"			myPair.y = aabbs[j].m_minIndices[3];\n"
+"			myPair.z = NEW_PAIR_MARKER;\n"
+"			myPair.w = NEW_PAIR_MARKER;\n"
+"			int curPair = atomic_inc (pairCount);\n"
+"			if (curPair<maxPairs)\n"
+"			{\n"
+"					pairsOut[curPair] = myPair; //flush to main memory\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"}\n"
+"__kernel void   computePairsKernelOriginal( __global const btAabbCL* aabbs, volatile __global int4* pairsOut,volatile  __global int* pairCount, int numObjects, int axis, int maxPairs)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numObjects)\n"
+"		return;\n"
+"	for (int j=i+1;j<numObjects;j++)\n"
+"	{\n"
+"  	if(aabbs[i].m_maxElems[axis] < (aabbs[j].m_minElems[axis])) \n"
+"		{\n"
+"			break;\n"
+"		}\n"
+"		if (TestAabbAgainstAabb2GlobalGlobal(&aabbs[i],&aabbs[j]))\n"
+"		{\n"
+"			int4 myPair;\n"
+"			myPair.x = aabbs[i].m_minIndices[3];\n"
+"			myPair.y = aabbs[j].m_minIndices[3];\n"
+"			myPair.z = NEW_PAIR_MARKER;\n"
+"			myPair.w = NEW_PAIR_MARKER;\n"
+"			int curPair = atomic_inc (pairCount);\n"
+"			if (curPair<maxPairs)\n"
+"			{\n"
+"					pairsOut[curPair] = myPair; //flush to main memory\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"}\n"
+"__kernel void   computePairsKernelBarrier( __global const btAabbCL* aabbs, volatile __global int4* pairsOut,volatile  __global int* pairCount, int numObjects, int axis, int maxPairs)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	int localId = get_local_id(0);\n"
+"	__local int numActiveWgItems[1];\n"
+"	__local int breakRequest[1];\n"
+"	if (localId==0)\n"
+"	{\n"
+"		numActiveWgItems[0] = 0;\n"
+"		breakRequest[0] = 0;\n"
+"	}\n"
+"	barrier(CLK_LOCAL_MEM_FENCE);\n"
+"	atomic_inc(numActiveWgItems);\n"
+"	barrier(CLK_LOCAL_MEM_FENCE);\n"
+"	int localBreak = 0;\n"
+"	int j=i+1;\n"
+"	do\n"
+"	{\n"
+"		barrier(CLK_LOCAL_MEM_FENCE);\n"
+"	\n"
+"		if (j<numObjects)\n"
+"		{\n"
+"	  	if(aabbs[i].m_maxElems[axis] < (aabbs[j].m_minElems[axis])) \n"
+"			{\n"
+"				if (!localBreak)\n"
+"				{\n"
+"					atomic_inc(breakRequest);\n"
+"					localBreak = 1;\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"		\n"
+"		barrier(CLK_LOCAL_MEM_FENCE);\n"
+"		\n"
+"		if (j>=numObjects && !localBreak)\n"
+"		{\n"
+"			atomic_inc(breakRequest);\n"
+"			localBreak = 1;\n"
+"		}\n"
+"		barrier(CLK_LOCAL_MEM_FENCE);\n"
+"		\n"
+"		if (!localBreak)\n"
+"		{\n"
+"			if (TestAabbAgainstAabb2GlobalGlobal(&aabbs[i],&aabbs[j]))\n"
+"			{\n"
+"				int4 myPair;\n"
+"				myPair.x = aabbs[i].m_minIndices[3];\n"
+"				myPair.y = aabbs[j].m_minIndices[3];\n"
+"				myPair.z = NEW_PAIR_MARKER;\n"
+"				myPair.w = NEW_PAIR_MARKER;\n"
+"				int curPair = atomic_inc (pairCount);\n"
+"				if (curPair<maxPairs)\n"
+"				{\n"
+"						pairsOut[curPair] = myPair; //flush to main memory\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"		j++;\n"
+"	} while (breakRequest[0]<numActiveWgItems[0]);\n"
+"}\n"
+"__kernel void   computePairsKernelLocalSharedMemory( __global const btAabbCL* aabbs, volatile __global int4* pairsOut,volatile  __global int* pairCount, int numObjects, int axis, int maxPairs)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	int localId = get_local_id(0);\n"
+"	__local int numActiveWgItems[1];\n"
+"	__local int breakRequest[1];\n"
+"	__local btAabbCL localAabbs[128];// = aabbs[i];\n"
+"	\n"
+"	btAabbCL myAabb;\n"
+"	\n"
+"	myAabb = (i<numObjects)? aabbs[i]:aabbs[0];\n"
+"	float testValue = 	myAabb.m_maxElems[axis];\n"
+"	\n"
+"	if (localId==0)\n"
+"	{\n"
+"		numActiveWgItems[0] = 0;\n"
+"		breakRequest[0] = 0;\n"
+"	}\n"
+"	int localCount=0;\n"
+"	int block=0;\n"
+"	localAabbs[localId] = (i+block)<numObjects? aabbs[i+block] : aabbs[0];\n"
+"	localAabbs[localId+64] = (i+block+64)<numObjects? aabbs[i+block+64]: aabbs[0];\n"
+"	\n"
+"	barrier(CLK_LOCAL_MEM_FENCE);\n"
+"	atomic_inc(numActiveWgItems);\n"
+"	barrier(CLK_LOCAL_MEM_FENCE);\n"
+"	int localBreak = 0;\n"
+"	\n"
+"	int j=i+1;\n"
+"	do\n"
+"	{\n"
+"		barrier(CLK_LOCAL_MEM_FENCE);\n"
+"	\n"
+"		if (j<numObjects)\n"
+"		{\n"
+"	  	if(testValue < (localAabbs[localCount+localId+1].m_minElems[axis])) \n"
+"			{\n"
+"				if (!localBreak)\n"
+"				{\n"
+"					atomic_inc(breakRequest);\n"
+"					localBreak = 1;\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"		\n"
+"		barrier(CLK_LOCAL_MEM_FENCE);\n"
+"		\n"
+"		if (j>=numObjects && !localBreak)\n"
+"		{\n"
+"			atomic_inc(breakRequest);\n"
+"			localBreak = 1;\n"
+"		}\n"
+"		barrier(CLK_LOCAL_MEM_FENCE);\n"
+"		\n"
+"		if (!localBreak)\n"
+"		{\n"
+"			if (TestAabbAgainstAabb2(&myAabb,&localAabbs[localCount+localId+1]))\n"
+"			{\n"
+"				int4 myPair;\n"
+"				myPair.x = myAabb.m_minIndices[3];\n"
+"				myPair.y = localAabbs[localCount+localId+1].m_minIndices[3];\n"
+"				myPair.z = NEW_PAIR_MARKER;\n"
+"				myPair.w = NEW_PAIR_MARKER;\n"
+"				int curPair = atomic_inc (pairCount);\n"
+"				if (curPair<maxPairs)\n"
+"				{\n"
+"						pairsOut[curPair] = myPair; //flush to main memory\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"		\n"
+"		barrier(CLK_LOCAL_MEM_FENCE);\n"
+"		localCount++;\n"
+"		if (localCount==64)\n"
+"		{\n"
+"			localCount = 0;\n"
+"			block+=64;			\n"
+"			localAabbs[localId] = ((i+block)<numObjects) ? aabbs[i+block] : aabbs[0];\n"
+"			localAabbs[localId+64] = ((i+64+block)<numObjects) ? aabbs[i+block+64] : aabbs[0];\n"
+"		}\n"
+"		j++;\n"
+"		\n"
+"	} while (breakRequest[0]<numActiveWgItems[0]);\n"
+"	\n"
+"}\n"
+"//http://stereopsis.com/radix.html\n"
+"unsigned int FloatFlip(float fl);\n"
+"unsigned int FloatFlip(float fl)\n"
+"{\n"
+"	unsigned int f = *(unsigned int*)&fl;\n"
+"	unsigned int mask = -(int)(f >> 31) | 0x80000000;\n"
+"	return f ^ mask;\n"
+"}\n"
+"float IFloatFlip(unsigned int f);\n"
+"float IFloatFlip(unsigned int f)\n"
+"{\n"
+"	unsigned int mask = ((f >> 31) - 1) | 0x80000000;\n"
+"	unsigned int fl = f ^ mask;\n"
+"	return *(float*)&fl;\n"
+"}\n"
+"__kernel void   copyAabbsKernel( __global const btAabbCL* allAabbs, __global btAabbCL* destAabbs, int numObjects)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numObjects)\n"
+"		return;\n"
+"	int src = destAabbs[i].m_maxIndices[3];\n"
+"	destAabbs[i] = allAabbs[src];\n"
+"	destAabbs[i].m_maxIndices[3] = src;\n"
+"}\n"
+"__kernel void   flipFloatKernel( __global const btAabbCL* allAabbs, __global const int* smallAabbMapping, __global int2* sortData, int numObjects, int axis)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numObjects)\n"
+"		return;\n"
+"	\n"
+"	\n"
+"	sortData[i].x = FloatFlip(allAabbs[smallAabbMapping[i]].m_minElems[axis]);\n"
+"	sortData[i].y = i;\n"
+"		\n"
+"}\n"
+"__kernel void   scatterKernel( __global const btAabbCL* allAabbs, __global const int* smallAabbMapping, volatile __global const int2* sortData, __global btAabbCL* sortedAabbs, int numObjects)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numObjects)\n"
+"		return;\n"
+"	\n"
+"	sortedAabbs[i] = allAabbs[smallAabbMapping[sortData[i].y]];\n"
+"}\n"
+"__kernel void   prepareSumVarianceKernel( __global const btAabbCL* allAabbs, __global const int* smallAabbMapping, __global float4* sum, __global float4* sum2,int numAabbs)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numAabbs)\n"
+"		return;\n"
+"	\n"
+"	btAabbCL smallAabb = allAabbs[smallAabbMapping[i]];\n"
+"	\n"
+"	float4 s;\n"
+"	s = (smallAabb.m_max+smallAabb.m_min)*0.5f;\n"
+"	sum[i]=s;\n"
+"	sum2[i]=s*s;	\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/Initialize/b3OpenCLInclude.h b/thirdparty/bullet/Bullet3OpenCL/Initialize/b3OpenCLInclude.h
new file mode 100644
index 0000000000..e79182d7cb
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/Initialize/b3OpenCLInclude.h
@@ -0,0 +1,48 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2011 Advanced Micro Devices, Inc.  http://bulletphysics.org
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+#ifndef B3_OPENCL_INCLUDE_H
+#define B3_OPENCL_INCLUDE_H
+
+#ifdef B3_USE_CLEW
+	#include "clew/clew.h"
+#else
+
+#ifdef __APPLE__
+#ifdef USE_MINICL
+#include <MiniCL/cl.h>
+#else
+#include <OpenCL/cl.h>
+#include <OpenCL/cl_ext.h> //clLogMessagesToStderrAPPLE
+#endif
+#else
+#ifdef USE_MINICL
+#include <MiniCL/cl.h>
+#else
+#include <CL/cl.h>
+#ifdef _WIN32
+#include "CL/cl_gl.h"
+#endif //_WIN32
+#endif
+#endif //__APPLE__
+#endif //B3_USE_CLEW
+
+#include <assert.h>
+#include <stdio.h>
+#define oclCHECKERROR(a, b) if((a)!=(b)) { printf("OCL Error : %d\n", (a)); assert((a) == (b)); }
+
+
+#endif //B3_OPENCL_INCLUDE_H
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/Initialize/b3OpenCLUtils.cpp b/thirdparty/bullet/Bullet3OpenCL/Initialize/b3OpenCLUtils.cpp
new file mode 100644
index 0000000000..dd194fc7ba
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/Initialize/b3OpenCLUtils.cpp
@@ -0,0 +1,1011 @@
+/*
+Bullet Continuous Collision Detection and Physics Library, http://bulletphysics.org
+Copyright (C) 2006 - 2011 Sony Computer Entertainment Inc.
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose,
+including commercial applications, and to alter it and redistribute it freely,
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+//Original author: Roman Ponomarev
+//Mostly Reimplemented by Erwin Coumans
+
+
+bool gDebugForceLoadingFromSource = false;
+bool gDebugSkipLoadingBinary = false;
+
+#include "Bullet3Common/b3Logging.h"
+
+#include <string.h>
+
+#ifdef _WIN32
+#pragma warning (disable:4996)
+#endif
+#include "b3OpenCLUtils.h"
+//#include "b3OpenCLInclude.h"
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#define B3_MAX_CL_DEVICES 16 //who needs 16 devices?
+
+#ifdef _WIN32
+#include <windows.h>
+#endif
+
+#include <assert.h>
+#define b3Assert assert
+#ifndef _WIN32
+#include <sys/stat.h>
+
+#endif
+
+static const char* sCachedBinaryPath="cache";
+
+
+//Set the preferred platform vendor using the OpenCL SDK
+static const char* spPlatformVendor =
+#if defined(CL_PLATFORM_MINI_CL)
+"MiniCL, SCEA";
+#elif defined(CL_PLATFORM_AMD)
+"Advanced Micro Devices, Inc.";
+#elif defined(CL_PLATFORM_NVIDIA)
+"NVIDIA Corporation";
+#elif defined(CL_PLATFORM_INTEL)
+"Intel(R) Corporation";
+#elif defined(B3_USE_CLEW)
+"clew (OpenCL Extension Wrangler library)";
+#else
+"Unknown Vendor";
+#endif
+
+#ifndef CL_PLATFORM_MINI_CL
+#ifdef _WIN32
+#ifndef B3_USE_CLEW
+#include "CL/cl_gl.h"
+#endif //B3_USE_CLEW
+#endif //_WIN32
+#endif
+
+
+void MyFatalBreakAPPLE(   const char *  errstr ,
+                       const void *  private_info ,
+                       size_t        cb ,
+                       void *        user_data  )
+{
+   
+
+    const char* patloc = strstr(errstr, "Warning");
+    //find out if it is a warning or error, exit if error
+
+    if (patloc)
+    {
+		b3Warning("Warning: %s\n", errstr);
+    } else
+    {
+		b3Error("Error: %s\n", errstr);
+        b3Assert(0);
+    }
+
+}
+
+#ifdef B3_USE_CLEW
+
+int b3OpenCLUtils_clewInit()
+{
+	int result = -1;
+
+#ifdef _WIN32
+        const char* cl = "OpenCL.dll";
+#elif defined __APPLE__
+        const char* cl = "/System/Library/Frameworks/OpenCL.framework/Versions/Current/OpenCL";
+#else//presumable Linux?
+        //linux (tested on Ubuntu 12.10 with Catalyst 13.4 beta drivers, not that there is no symbolic link from libOpenCL.so
+        const char* cl = "libOpenCL.so.1";
+        result = clewInit(cl);
+        if (result != CLEW_SUCCESS)
+        {
+                cl = "libOpenCL.so";
+        } else
+        {
+                clewExit();
+        }
+#endif
+        result = clewInit(cl);
+        if (result!=CLEW_SUCCESS)
+		{
+                b3Error("clewInit failed with error code %d\n",result);
+		}
+        else
+        {
+                b3Printf("clewInit succesfull using %s\n",cl);
+        }
+	return result;
+}
+#endif
+
+int b3OpenCLUtils_getNumPlatforms(cl_int* pErrNum)
+{
+#ifdef B3_USE_CLEW
+	b3OpenCLUtils_clewInit();
+#endif
+
+	cl_platform_id pPlatforms[10] = { 0 };
+
+    cl_uint numPlatforms = 0;
+    cl_int ciErrNum = clGetPlatformIDs(10, pPlatforms, &numPlatforms);
+	//cl_int ciErrNum = clGetPlatformIDs(0, NULL, &numPlatforms);
+
+	if(ciErrNum != CL_SUCCESS)
+	{
+		if(pErrNum != NULL)
+			*pErrNum = ciErrNum;
+	}
+	return numPlatforms;
+
+}
+
+const char* b3OpenCLUtils_getSdkVendorName()
+{
+	return spPlatformVendor;
+}
+
+void b3OpenCLUtils_setCachePath(const char* path)
+{
+	sCachedBinaryPath = path;
+}
+
+cl_platform_id b3OpenCLUtils_getPlatform(int platformIndex0, cl_int* pErrNum)
+{
+#ifdef B3_USE_CLEW
+        b3OpenCLUtils_clewInit();
+#endif
+
+	cl_platform_id platform = 0;
+	unsigned int platformIndex = (unsigned int )platformIndex0;
+	cl_uint numPlatforms;
+	cl_int ciErrNum = clGetPlatformIDs(0, NULL, &numPlatforms);
+
+	if (platformIndex<numPlatforms)
+	{
+		cl_platform_id* platforms = (cl_platform_id*) malloc (sizeof(cl_platform_id)*numPlatforms);
+		ciErrNum = clGetPlatformIDs(numPlatforms, platforms, NULL);
+		if(ciErrNum != CL_SUCCESS)
+		{
+			if(pErrNum != NULL)
+				*pErrNum = ciErrNum;
+			return platform;
+		}
+
+		platform = platforms[platformIndex];
+
+		free (platforms);
+	}
+
+	return platform;
+}
+
+void b3OpenCLUtils::getPlatformInfo(cl_platform_id platform, b3OpenCLPlatformInfo* platformInfo)
+{
+	b3Assert(platform);
+	cl_int ciErrNum;
+	ciErrNum = clGetPlatformInfo(	platform,CL_PLATFORM_VENDOR,B3_MAX_STRING_LENGTH,platformInfo->m_platformVendor,NULL);
+	oclCHECKERROR(ciErrNum,CL_SUCCESS);
+	ciErrNum = clGetPlatformInfo(	platform,CL_PLATFORM_NAME,B3_MAX_STRING_LENGTH,platformInfo->m_platformName,NULL);
+	oclCHECKERROR(ciErrNum,CL_SUCCESS);
+	ciErrNum = clGetPlatformInfo(	platform,CL_PLATFORM_VERSION,B3_MAX_STRING_LENGTH,platformInfo->m_platformVersion,NULL);
+	oclCHECKERROR(ciErrNum,CL_SUCCESS);
+}
+
+void b3OpenCLUtils_printPlatformInfo( cl_platform_id platform)
+{
+	b3OpenCLPlatformInfo platformInfo;
+	b3OpenCLUtils::getPlatformInfo (platform, &platformInfo);
+	b3Printf("Platform info:\n");
+	b3Printf("  CL_PLATFORM_VENDOR: \t\t\t%s\n",platformInfo.m_platformVendor);
+	b3Printf("  CL_PLATFORM_NAME: \t\t\t%s\n",platformInfo.m_platformName);
+	b3Printf("  CL_PLATFORM_VERSION: \t\t\t%s\n",platformInfo.m_platformVersion);
+}
+
+
+
+cl_context b3OpenCLUtils_createContextFromPlatform(cl_platform_id platform, cl_device_type deviceType, cl_int* pErrNum, void* pGLContext, void* pGLDC, int preferredDeviceIndex, int preferredPlatformIndex)
+{
+	cl_context retContext = 0;
+	cl_int ciErrNum=0;
+	cl_uint num_entries;
+	cl_device_id devices[B3_MAX_CL_DEVICES];
+	cl_uint num_devices;
+	cl_context_properties* cprops;
+
+	/*
+	* If we could find our platform, use it. Otherwise pass a NULL and get whatever the
+	* implementation thinks we should be using.
+	*/
+	cl_context_properties cps[7] = {0,0,0,0,0,0,0};
+	cps[0] = CL_CONTEXT_PLATFORM;
+	cps[1] = (cl_context_properties)platform;
+#ifdef _WIN32
+#ifndef B3_USE_CLEW
+	if (pGLContext && pGLDC)
+	{
+		cps[2] = CL_GL_CONTEXT_KHR;
+		cps[3] = (cl_context_properties)pGLContext;
+		cps[4] = CL_WGL_HDC_KHR;
+		cps[5] = (cl_context_properties)pGLDC;
+	}
+#endif //B3_USE_CLEW
+#endif //_WIN32
+	num_entries = B3_MAX_CL_DEVICES;
+
+
+	num_devices=-1;
+
+	ciErrNum = clGetDeviceIDs(
+		platform,
+		deviceType,
+ 		num_entries,
+ 		devices,
+ 		&num_devices);
+
+    if (ciErrNum<0)
+    {
+        b3Printf("clGetDeviceIDs returned %d\n",ciErrNum);
+        return 0;
+    }
+	cprops = (NULL == platform) ? NULL : cps;
+
+	if (!num_devices)
+		return 0;
+
+	if (pGLContext)
+	{
+		//search for the GPU that relates to the OpenCL context
+		unsigned int i;
+		for (i=0;i<num_devices;i++)
+		{
+			retContext = clCreateContext(cprops,1,&devices[i],NULL,NULL,&ciErrNum);
+			if (ciErrNum==CL_SUCCESS)
+				break;
+		}
+	}
+	else
+	{
+		if (preferredDeviceIndex>=0 && (unsigned int)preferredDeviceIndex<num_devices)
+		{
+			//create a context of the preferred device index
+			retContext = clCreateContext(cprops,1,&devices[preferredDeviceIndex],NULL,NULL,&ciErrNum);
+		} else
+		{
+			//create a context of all devices
+#if defined (__APPLE__)
+			retContext = clCreateContext(cprops,num_devices,devices,MyFatalBreakAPPLE,NULL,&ciErrNum);
+#else
+        b3Printf("numDevices=%d\n",num_devices);
+
+			retContext = clCreateContext(cprops,num_devices,devices,NULL,NULL,&ciErrNum);
+#endif
+		}
+	}
+	if(pErrNum != NULL)
+	{
+		*pErrNum = ciErrNum;
+	};
+
+	return retContext;
+}
+
+cl_context b3OpenCLUtils_createContextFromType(cl_device_type deviceType, cl_int* pErrNum, void* pGLContext, void* pGLDC , int preferredDeviceIndex, int preferredPlatformIndex, cl_platform_id* retPlatformId)
+{
+#ifdef B3_USE_CLEW
+        b3OpenCLUtils_clewInit();
+#endif
+
+
+	cl_uint numPlatforms;
+	cl_context retContext = 0;
+	unsigned int i;
+
+	cl_int ciErrNum = clGetPlatformIDs(0, NULL, &numPlatforms);
+	if(ciErrNum != CL_SUCCESS)
+	{
+		if(pErrNum != NULL) *pErrNum = ciErrNum;
+		return NULL;
+	}
+	if(numPlatforms > 0)
+	{
+		cl_platform_id* platforms = (cl_platform_id*) malloc (sizeof(cl_platform_id)*numPlatforms);
+		ciErrNum = clGetPlatformIDs(numPlatforms, platforms, NULL);
+		if(ciErrNum != CL_SUCCESS)
+		{
+			if(pErrNum != NULL)
+				*pErrNum = ciErrNum;
+			free(platforms);
+			return NULL;
+		}
+
+
+
+		for ( i = 0; i < numPlatforms; ++i)
+		{
+			char pbuf[128];
+			ciErrNum = clGetPlatformInfo(	platforms[i],
+				CL_PLATFORM_VENDOR,
+				sizeof(pbuf),
+				pbuf,
+				NULL);
+			if(ciErrNum != CL_SUCCESS)
+			{
+				if(pErrNum != NULL) *pErrNum = ciErrNum;
+				return NULL;
+			}
+
+			if (preferredPlatformIndex>=0 && i==preferredPlatformIndex)
+			{
+				cl_platform_id tmpPlatform = platforms[0];
+				platforms[0] = platforms[i];
+				platforms[i] = tmpPlatform;
+				break;
+			} else
+			{
+				if(!strcmp(pbuf, spPlatformVendor))
+				{
+					cl_platform_id tmpPlatform = platforms[0];
+					platforms[0] = platforms[i];
+					platforms[i] = tmpPlatform;
+				}
+			}
+		}
+
+		for (i = 0; i < numPlatforms; ++i)
+		{
+			cl_platform_id platform = platforms[i];
+			assert(platform);
+
+			retContext = b3OpenCLUtils_createContextFromPlatform(platform,deviceType,pErrNum,pGLContext,pGLDC,preferredDeviceIndex,preferredPlatformIndex);
+
+			if (retContext)
+			{
+//				printf("OpenCL platform details:\n");
+				b3OpenCLPlatformInfo platformInfo;
+
+				b3OpenCLUtils::getPlatformInfo(platform, &platformInfo);
+
+				if (retPlatformId)
+					*retPlatformId = platform;
+
+				break;
+			}
+		}
+
+		free (platforms);
+	}
+	return retContext;
+}
+
+
+//////////////////////////////////////////////////////////////////////////////
+//! Gets the id of the nth device from the context
+//!
+//! @return the id or -1 when out of range
+//! @param cxMainContext         OpenCL context
+//! @param device_idx            index of the device of interest
+//////////////////////////////////////////////////////////////////////////////
+cl_device_id b3OpenCLUtils_getDevice(cl_context cxMainContext, int deviceIndex)
+{
+	assert(cxMainContext);
+
+	size_t szParmDataBytes;
+	cl_device_id* cdDevices;
+	cl_device_id device ;
+
+	// get the list of devices associated with context
+	clGetContextInfo(cxMainContext, CL_CONTEXT_DEVICES, 0, NULL, &szParmDataBytes);
+
+	if( szParmDataBytes / sizeof(cl_device_id) < (unsigned int)deviceIndex ) {
+		return (cl_device_id)-1;
+	}
+
+	cdDevices = (cl_device_id*) malloc(szParmDataBytes);
+
+	clGetContextInfo(cxMainContext, CL_CONTEXT_DEVICES, szParmDataBytes, cdDevices, NULL);
+
+	device = cdDevices[deviceIndex];
+	free(cdDevices);
+
+	return device;
+}
+
+int b3OpenCLUtils_getNumDevices(cl_context cxMainContext)
+{
+	size_t szParamDataBytes;
+	int device_count;
+	clGetContextInfo(cxMainContext, CL_CONTEXT_DEVICES, 0, NULL, &szParamDataBytes);
+	device_count = (int) szParamDataBytes/ sizeof(cl_device_id);
+	return device_count;
+}
+
+
+
+void b3OpenCLUtils::getDeviceInfo(cl_device_id device, b3OpenCLDeviceInfo* info)
+{
+	// CL_DEVICE_NAME
+	clGetDeviceInfo(device, CL_DEVICE_NAME, B3_MAX_STRING_LENGTH, &info->m_deviceName, NULL);
+
+	// CL_DEVICE_VENDOR
+	clGetDeviceInfo(device, CL_DEVICE_VENDOR, B3_MAX_STRING_LENGTH, &info->m_deviceVendor, NULL);
+
+	// CL_DRIVER_VERSION
+	clGetDeviceInfo(device, CL_DRIVER_VERSION, B3_MAX_STRING_LENGTH, &info->m_driverVersion, NULL);
+
+	// CL_DEVICE_INFO
+	clGetDeviceInfo(device, CL_DEVICE_TYPE, sizeof(cl_device_type), &info->m_deviceType, NULL);
+
+	// CL_DEVICE_MAX_COMPUTE_UNITS
+	clGetDeviceInfo(device, CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(info->m_computeUnits), &info->m_computeUnits, NULL);
+
+	// CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS
+	clGetDeviceInfo(device, CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS, sizeof(info->m_workitemDims), &info->m_workitemDims, NULL);
+
+	// CL_DEVICE_MAX_WORK_ITEM_SIZES
+	clGetDeviceInfo(device, CL_DEVICE_MAX_WORK_ITEM_SIZES, sizeof(info->m_workItemSize), &info->m_workItemSize, NULL);
+
+	// CL_DEVICE_MAX_WORK_GROUP_SIZE
+	clGetDeviceInfo(device, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(info->m_workgroupSize), &info->m_workgroupSize, NULL);
+
+	// CL_DEVICE_MAX_CLOCK_FREQUENCY
+	clGetDeviceInfo(device, CL_DEVICE_MAX_CLOCK_FREQUENCY, sizeof(info->m_clockFrequency), &info->m_clockFrequency, NULL);
+
+	// CL_DEVICE_ADDRESS_BITS
+	clGetDeviceInfo(device, CL_DEVICE_ADDRESS_BITS, sizeof(info->m_addressBits), &info->m_addressBits, NULL);
+
+	// CL_DEVICE_MAX_MEM_ALLOC_SIZE
+	clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(info->m_maxMemAllocSize), &info->m_maxMemAllocSize, NULL);
+
+	// CL_DEVICE_GLOBAL_MEM_SIZE
+	clGetDeviceInfo(device, CL_DEVICE_GLOBAL_MEM_SIZE, sizeof(info->m_globalMemSize), &info->m_globalMemSize, NULL);
+
+	// CL_DEVICE_ERROR_CORRECTION_SUPPORT
+	clGetDeviceInfo(device, CL_DEVICE_ERROR_CORRECTION_SUPPORT, sizeof(info->m_errorCorrectionSupport), &info->m_errorCorrectionSupport, NULL);
+
+	// CL_DEVICE_LOCAL_MEM_TYPE
+	clGetDeviceInfo(device, CL_DEVICE_LOCAL_MEM_TYPE, sizeof(info->m_localMemType), &info->m_localMemType, NULL);
+
+	// CL_DEVICE_LOCAL_MEM_SIZE
+	clGetDeviceInfo(device, CL_DEVICE_LOCAL_MEM_SIZE, sizeof(info->m_localMemSize), &info->m_localMemSize, NULL);
+
+	// CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE
+	clGetDeviceInfo(device, CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE, sizeof(info->m_constantBufferSize), &info->m_constantBufferSize, NULL);
+
+	// CL_DEVICE_QUEUE_PROPERTIES
+	clGetDeviceInfo(device, CL_DEVICE_QUEUE_PROPERTIES, sizeof(info->m_queueProperties), &info->m_queueProperties, NULL);
+
+	// CL_DEVICE_IMAGE_SUPPORT
+	clGetDeviceInfo(device, CL_DEVICE_IMAGE_SUPPORT, sizeof(info->m_imageSupport), &info->m_imageSupport, NULL);
+
+	// CL_DEVICE_MAX_READ_IMAGE_ARGS
+	clGetDeviceInfo(device, CL_DEVICE_MAX_READ_IMAGE_ARGS, sizeof(info->m_maxReadImageArgs), &info->m_maxReadImageArgs, NULL);
+
+	// CL_DEVICE_MAX_WRITE_IMAGE_ARGS
+	clGetDeviceInfo(device, CL_DEVICE_MAX_WRITE_IMAGE_ARGS, sizeof(info->m_maxWriteImageArgs), &info->m_maxWriteImageArgs, NULL);
+
+	// CL_DEVICE_IMAGE2D_MAX_WIDTH, CL_DEVICE_IMAGE2D_MAX_HEIGHT, CL_DEVICE_IMAGE3D_MAX_WIDTH, CL_DEVICE_IMAGE3D_MAX_HEIGHT, CL_DEVICE_IMAGE3D_MAX_DEPTH
+	clGetDeviceInfo(device, CL_DEVICE_IMAGE2D_MAX_WIDTH, sizeof(size_t), &info->m_image2dMaxWidth, NULL);
+	clGetDeviceInfo(device, CL_DEVICE_IMAGE2D_MAX_HEIGHT, sizeof(size_t), &info->m_image2dMaxHeight, NULL);
+	clGetDeviceInfo(device, CL_DEVICE_IMAGE3D_MAX_WIDTH, sizeof(size_t), &info->m_image3dMaxWidth, NULL);
+	clGetDeviceInfo(device, CL_DEVICE_IMAGE3D_MAX_HEIGHT, sizeof(size_t), &info->m_image3dMaxHeight, NULL);
+	clGetDeviceInfo(device, CL_DEVICE_IMAGE3D_MAX_DEPTH, sizeof(size_t), &info->m_image3dMaxDepth, NULL);
+
+	// CL_DEVICE_EXTENSIONS: get device extensions, and if any then parse & log the string onto separate lines
+	clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, B3_MAX_STRING_LENGTH, &info->m_deviceExtensions, NULL);
+
+	// CL_DEVICE_PREFERRED_VECTOR_WIDTH_<type>
+	clGetDeviceInfo(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR, sizeof(cl_uint), &info->m_vecWidthChar, NULL);
+	clGetDeviceInfo(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT, sizeof(cl_uint), &info->m_vecWidthShort, NULL);
+	clGetDeviceInfo(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT, sizeof(cl_uint), &info->m_vecWidthInt, NULL);
+	clGetDeviceInfo(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG, sizeof(cl_uint), &info->m_vecWidthLong, NULL);
+	clGetDeviceInfo(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT, sizeof(cl_uint), &info->m_vecWidthFloat, NULL);
+	clGetDeviceInfo(device, CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE, sizeof(cl_uint), &info->m_vecWidthDouble, NULL);
+}
+
+
+void b3OpenCLUtils_printDeviceInfo(cl_device_id device)
+{
+	b3OpenCLDeviceInfo info;
+	b3OpenCLUtils::getDeviceInfo(device,&info);
+	b3Printf("Device Info:\n");
+	b3Printf("  CL_DEVICE_NAME: \t\t\t%s\n", info.m_deviceName);
+	b3Printf("  CL_DEVICE_VENDOR: \t\t\t%s\n", info.m_deviceVendor);
+	b3Printf("  CL_DRIVER_VERSION: \t\t\t%s\n", info.m_driverVersion);
+
+	if( info.m_deviceType & CL_DEVICE_TYPE_CPU )
+		b3Printf("  CL_DEVICE_TYPE:\t\t\t%s\n", "CL_DEVICE_TYPE_CPU");
+	if( info.m_deviceType & CL_DEVICE_TYPE_GPU )
+		b3Printf("  CL_DEVICE_TYPE:\t\t\t%s\n", "CL_DEVICE_TYPE_GPU");
+	if( info.m_deviceType & CL_DEVICE_TYPE_ACCELERATOR )
+		b3Printf("  CL_DEVICE_TYPE:\t\t\t%s\n", "CL_DEVICE_TYPE_ACCELERATOR");
+	if( info.m_deviceType & CL_DEVICE_TYPE_DEFAULT )
+		b3Printf("  CL_DEVICE_TYPE:\t\t\t%s\n", "CL_DEVICE_TYPE_DEFAULT");
+
+	b3Printf("  CL_DEVICE_MAX_COMPUTE_UNITS:\t\t%u\n", info.m_computeUnits);
+	b3Printf("  CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS:\t%u\n", info.m_workitemDims);
+	b3Printf("  CL_DEVICE_MAX_WORK_ITEM_SIZES:\t%u / %u / %u \n", info.m_workItemSize[0], info.m_workItemSize[1], info.m_workItemSize[2]);
+	b3Printf("  CL_DEVICE_MAX_WORK_GROUP_SIZE:\t%u\n", info.m_workgroupSize);
+	b3Printf("  CL_DEVICE_MAX_CLOCK_FREQUENCY:\t%u MHz\n", info.m_clockFrequency);
+	b3Printf("  CL_DEVICE_ADDRESS_BITS:\t\t%u\n", info.m_addressBits);
+	b3Printf("  CL_DEVICE_MAX_MEM_ALLOC_SIZE:\t\t%u MByte\n", (unsigned int)(info.m_maxMemAllocSize/ (1024 * 1024)));
+	b3Printf("  CL_DEVICE_GLOBAL_MEM_SIZE:\t\t%u MByte\n", (unsigned int)(info.m_globalMemSize/ (1024 * 1024)));
+	b3Printf("  CL_DEVICE_ERROR_CORRECTION_SUPPORT:\t%s\n", info.m_errorCorrectionSupport== CL_TRUE ? "yes" : "no");
+	b3Printf("  CL_DEVICE_LOCAL_MEM_TYPE:\t\t%s\n", info.m_localMemType == 1 ? "local" : "global");
+	b3Printf("  CL_DEVICE_LOCAL_MEM_SIZE:\t\t%u KByte\n", (unsigned int)(info.m_localMemSize / 1024));
+	b3Printf("  CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE:\t%u KByte\n", (unsigned int)(info.m_constantBufferSize / 1024));
+	if( info.m_queueProperties  & CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE )
+		b3Printf("  CL_DEVICE_QUEUE_PROPERTIES:\t\t%s\n", "CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE");
+	if( info.m_queueProperties & CL_QUEUE_PROFILING_ENABLE )
+		b3Printf("  CL_DEVICE_QUEUE_PROPERTIES:\t\t%s\n", "CL_QUEUE_PROFILING_ENABLE");
+
+	b3Printf("  CL_DEVICE_IMAGE_SUPPORT:\t\t%u\n", info.m_imageSupport);
+
+	b3Printf("  CL_DEVICE_MAX_READ_IMAGE_ARGS:\t%u\n", info.m_maxReadImageArgs);
+	b3Printf("  CL_DEVICE_MAX_WRITE_IMAGE_ARGS:\t%u\n", info.m_maxWriteImageArgs);
+	b3Printf("\n  CL_DEVICE_IMAGE <dim>");
+	b3Printf("\t\t\t2D_MAX_WIDTH\t %u\n", info.m_image2dMaxWidth);
+	b3Printf("\t\t\t\t\t2D_MAX_HEIGHT\t %u\n", info.m_image2dMaxHeight);
+	b3Printf("\t\t\t\t\t3D_MAX_WIDTH\t %u\n", info.m_image3dMaxWidth);
+	b3Printf("\t\t\t\t\t3D_MAX_HEIGHT\t %u\n", info.m_image3dMaxHeight);
+	b3Printf("\t\t\t\t\t3D_MAX_DEPTH\t %u\n", info.m_image3dMaxDepth);
+	if (*info.m_deviceExtensions != 0)
+	{
+		b3Printf("\n  CL_DEVICE_EXTENSIONS:%s\n",info.m_deviceExtensions);
+	}
+	else
+	{
+		b3Printf("  CL_DEVICE_EXTENSIONS: None\n");
+	}
+	b3Printf("  CL_DEVICE_PREFERRED_VECTOR_WIDTH_<t>\t");
+	b3Printf("CHAR %u, SHORT %u, INT %u,LONG %u, FLOAT %u, DOUBLE %u\n\n\n",
+		info.m_vecWidthChar, info.m_vecWidthShort, info.m_vecWidthInt, info.m_vecWidthLong,info.m_vecWidthFloat, info.m_vecWidthDouble);
+
+
+}
+
+
+static const char* strip2(const char* name, const char* pattern)
+{
+	  size_t const patlen = strlen(pattern);
+  	size_t patcnt = 0;
+	  const char * oriptr;
+	  const char * patloc;
+		// find how many times the pattern occurs in the original string
+	  for (oriptr = name; (patloc = strstr(oriptr, pattern)); oriptr = patloc + patlen)
+	  {
+		patcnt++;
+	  }
+	  return oriptr;
+}
+
+cl_program b3OpenCLUtils_compileCLProgramFromString(cl_context clContext, cl_device_id device, const char* kernelSourceOrg, cl_int* pErrNum, const char* additionalMacrosArg , const char* clFileNameForCaching, bool disableBinaryCaching)
+{
+	const char* additionalMacros = additionalMacrosArg?additionalMacrosArg:"";
+
+	if (disableBinaryCaching)
+	{
+		//kernelSourceOrg = 0;
+	}
+
+	cl_program m_cpProgram=0;
+	cl_int status;
+
+	char binaryFileName[B3_MAX_STRING_LENGTH];
+
+	char deviceName[256];
+	char driverVersion[256];
+	const char* strippedName;
+	int fileUpToDate = 0;
+#ifdef _WIN32
+	int binaryFileValid=0;
+#endif	
+	if (!disableBinaryCaching && clFileNameForCaching)
+	{
+		clGetDeviceInfo(device, CL_DEVICE_NAME, 256, &deviceName, NULL);
+		clGetDeviceInfo(device, CL_DRIVER_VERSION, 256, &driverVersion, NULL);
+	
+		strippedName = strip2(clFileNameForCaching,"\\");
+		strippedName = strip2(strippedName,"/");
+	
+#ifdef _MSVC_VER
+		sprintf_s(binaryFileName,B3_MAX_STRING_LENGTH,"%s/%s.%s.%s.bin",sCachedBinaryPath,strippedName, deviceName,driverVersion );
+#else
+		sprintf(binaryFileName,"%s/%s.%s.%s.bin",sCachedBinaryPath,strippedName, deviceName,driverVersion );
+#endif
+	}
+	if (clFileNameForCaching && !(disableBinaryCaching || gDebugSkipLoadingBinary||gDebugForceLoadingFromSource) )
+	{
+		
+#ifdef _WIN32
+	char* bla=0;
+
+	
+
+		//printf("searching for %s\n", binaryFileName);
+
+
+		FILETIME modtimeBinary;
+		CreateDirectoryA(sCachedBinaryPath,0);
+		{
+
+			HANDLE binaryFileHandle = CreateFileA(binaryFileName,GENERIC_READ,0,0,OPEN_EXISTING,FILE_ATTRIBUTE_NORMAL,0);
+			if (binaryFileHandle ==INVALID_HANDLE_VALUE)
+			{
+				DWORD errorCode;
+				errorCode = GetLastError();
+				switch (errorCode)
+				{
+				case ERROR_FILE_NOT_FOUND:
+					{
+						b3Warning("\nCached file not found %s\n", binaryFileName);
+						break;
+					}
+				case ERROR_PATH_NOT_FOUND:
+					{
+						b3Warning("\nCached file path not found %s\n", binaryFileName);
+						break;
+					}
+				default:
+					{
+						b3Warning("\nFailed reading cached file with errorCode = %d\n", errorCode);
+					}
+				}
+			} else
+			{
+				if (GetFileTime(binaryFileHandle, NULL, NULL, &modtimeBinary)==0)
+				{
+					DWORD errorCode;
+					errorCode = GetLastError();
+					b3Warning("\nGetFileTime errorCode = %d\n", errorCode);
+				} else
+				{
+					binaryFileValid = 1;
+				}
+				CloseHandle(binaryFileHandle);
+			}
+
+			if (binaryFileValid)
+			{
+				HANDLE srcFileHandle = CreateFileA(clFileNameForCaching,GENERIC_READ,0,0,OPEN_EXISTING,FILE_ATTRIBUTE_NORMAL,0);
+
+				if (srcFileHandle==INVALID_HANDLE_VALUE)
+				{
+					const char* prefix[]={"./","../","../../","../../../","../../../../"};
+					for (int i=0;(srcFileHandle==INVALID_HANDLE_VALUE) && i<5;i++)
+					{
+						char relativeFileName[1024];
+						sprintf(relativeFileName,"%s%s",prefix[i],clFileNameForCaching);
+						srcFileHandle = CreateFileA(relativeFileName,GENERIC_READ,0,0,OPEN_EXISTING,FILE_ATTRIBUTE_NORMAL,0);
+					}
+
+				}
+
+
+				if (srcFileHandle!=INVALID_HANDLE_VALUE)
+				{
+					FILETIME modtimeSrc;
+					if (GetFileTime(srcFileHandle, NULL, NULL, &modtimeSrc)==0)
+					{
+						DWORD errorCode;
+						errorCode = GetLastError();
+						b3Warning("\nGetFileTime errorCode = %d\n", errorCode);
+					}
+					if (  ( modtimeSrc.dwHighDateTime < modtimeBinary.dwHighDateTime)
+						||(( modtimeSrc.dwHighDateTime == modtimeBinary.dwHighDateTime)&&(modtimeSrc.dwLowDateTime <= modtimeBinary.dwLowDateTime)))
+					{
+						fileUpToDate=1;
+					} else
+					{
+						b3Warning("\nCached binary file out-of-date (%s)\n",binaryFileName);
+					}
+					CloseHandle(srcFileHandle);
+				}
+				else
+				{
+#ifdef _DEBUG
+					DWORD errorCode;
+					errorCode = GetLastError();
+					switch (errorCode)
+					{
+					case ERROR_FILE_NOT_FOUND:
+						{
+							b3Warning("\nSrc file not found %s\n", clFileNameForCaching);
+							break;
+						}
+					case ERROR_PATH_NOT_FOUND:
+						{
+							b3Warning("\nSrc path not found %s\n", clFileNameForCaching);
+							break;
+						}
+					default:
+						{
+							b3Warning("\nnSrc file reading errorCode = %d\n", errorCode);
+						}
+					}
+
+					//we should make sure the src file exists so we can verify the timestamp with binary
+//					assert(0);
+					b3Warning("Warning: cannot find OpenCL kernel %s to verify timestamp of binary cached kernel %s\n",clFileNameForCaching, binaryFileName);
+					fileUpToDate = true;
+#else
+					//if we cannot find the source, assume it is OK in release builds
+					fileUpToDate = true;
+#endif
+				}
+			}
+
+
+		}
+
+	
+
+#else
+	fileUpToDate = true;
+	if (mkdir(sCachedBinaryPath,0777) == -1)
+	{
+	}
+	else
+	{
+		b3Printf("Succesfully created cache directory: %s\n", sCachedBinaryPath);
+	}
+#endif //_WIN32
+	}
+	
+	
+	if( fileUpToDate)
+	{
+#ifdef _MSC_VER
+		FILE* file;
+		if (fopen_s(&file,binaryFileName, "rb")!=0)
+			file=0;
+#else
+		FILE* file = fopen(binaryFileName, "rb");
+#endif
+		
+		if (file)
+		{
+			size_t binarySize=0;
+			char* binary =0;
+			
+			fseek( file, 0L, SEEK_END );
+			binarySize = ftell( file );
+			rewind( file );
+			binary = (char*)malloc(sizeof(char)*binarySize);
+			int bytesRead;
+			bytesRead = fread( binary, sizeof(char), binarySize, file );
+			fclose( file );
+			
+			m_cpProgram = clCreateProgramWithBinary( clContext, 1,&device, &binarySize, (const unsigned char**)&binary, 0, &status );
+			b3Assert( status == CL_SUCCESS );
+			status = clBuildProgram( m_cpProgram, 1, &device, additionalMacros, 0, 0 );
+			b3Assert( status == CL_SUCCESS );
+			
+			if( status != CL_SUCCESS )
+			{
+				char *build_log;
+				size_t ret_val_size;
+				clGetProgramBuildInfo(m_cpProgram, device, CL_PROGRAM_BUILD_LOG, 0, NULL, &ret_val_size);
+				build_log = (char*)malloc(sizeof(char)*(ret_val_size+1));
+				clGetProgramBuildInfo(m_cpProgram, device, CL_PROGRAM_BUILD_LOG, ret_val_size, build_log, NULL);
+				build_log[ret_val_size] = '\0';
+				b3Error("%s\n", build_log);
+				free (build_log);
+				b3Assert(0);
+				m_cpProgram = 0;
+
+				b3Warning("clBuildProgram reported failure on cached binary: %s\n",binaryFileName);
+
+			} else
+			{
+				b3Printf("clBuildProgram successfully compiled cached binary: %s\n",binaryFileName);	
+			}
+			free (binary);
+			
+		} else
+		{
+			b3Warning("Cannot open cached binary: %s\n",binaryFileName);
+		}
+	}
+	
+	
+	
+	
+	
+	
+	
+	
+	
+	if (!m_cpProgram)
+	{
+
+		cl_int localErrNum;
+		char* compileFlags;
+		int flagsize;
+
+
+
+		const char* kernelSource = kernelSourceOrg;
+
+		if (!kernelSourceOrg || gDebugForceLoadingFromSource)
+		{
+			if (clFileNameForCaching)
+			{
+
+				FILE* file = fopen(clFileNameForCaching, "rb");
+				//in many cases the relative path is a few levels up the directory hierarchy, so try it
+				if (!file)
+				{
+					const char* prefix[]={"../","../../","../../../","../../../../"};
+					for (int i=0;!file && i<3;i++)
+					{
+						char relativeFileName[1024];
+						sprintf(relativeFileName,"%s%s",prefix[i],clFileNameForCaching);
+						file = fopen(relativeFileName, "rb");
+					}
+				}
+
+				if (file)
+				{
+					char* kernelSrc=0;
+					fseek( file, 0L, SEEK_END );
+					int kernelSize = ftell( file );
+					rewind( file );
+					kernelSrc = (char*)malloc(kernelSize+1);
+					int readBytes;
+					readBytes = fread((void*)kernelSrc,1,kernelSize, file);
+					kernelSrc[kernelSize] = 0;
+					fclose(file);
+					kernelSource = kernelSrc;
+				}
+			}
+		}
+
+		size_t program_length = kernelSource ? strlen(kernelSource) : 0;
+#ifdef MAC //or __APPLE__?
+		char* flags = "-cl-mad-enable -DMAC ";
+#else
+		const char* flags = "";
+#endif
+
+
+		m_cpProgram = clCreateProgramWithSource(clContext, 1, (const char**)&kernelSource, &program_length, &localErrNum);
+		if (localErrNum!= CL_SUCCESS)
+		{
+			if (pErrNum)
+				*pErrNum = localErrNum;
+			return 0;
+		}
+
+		// Build the program with 'mad' Optimization option
+
+
+
+        		flagsize = sizeof(char)*(strlen(additionalMacros) + strlen(flags) + 5);
+		compileFlags = (char*) malloc(flagsize);
+#ifdef _MSC_VER
+		sprintf_s(compileFlags,flagsize, "%s %s", flags, additionalMacros);
+#else
+		sprintf(compileFlags, "%s %s", flags, additionalMacros);
+#endif
+		localErrNum = clBuildProgram(m_cpProgram, 1, &device, compileFlags, NULL, NULL);
+		if (localErrNum!= CL_SUCCESS)
+		{
+			char *build_log;
+			size_t ret_val_size;
+			clGetProgramBuildInfo(m_cpProgram, device, CL_PROGRAM_BUILD_LOG, 0, NULL, &ret_val_size);
+			build_log = (char*) malloc(sizeof(char)*(ret_val_size+1));
+			clGetProgramBuildInfo(m_cpProgram, device, CL_PROGRAM_BUILD_LOG, ret_val_size, build_log, NULL);
+
+			// to be carefully, terminate with \0
+			// there's no information in the reference whether the string is 0 terminated or not
+			build_log[ret_val_size] = '\0';
+
+
+			b3Error("Error in clBuildProgram, Line %u in file %s, Log: \n%s\n !!!\n\n", __LINE__, __FILE__, build_log);
+			free (build_log);
+			if (pErrNum)
+				*pErrNum = localErrNum;
+			return 0;
+		}
+
+
+		if( !disableBinaryCaching && clFileNameForCaching )
+		{	//	write to binary
+
+			cl_uint numAssociatedDevices;
+			status = clGetProgramInfo( m_cpProgram, CL_PROGRAM_NUM_DEVICES, sizeof(cl_uint), &numAssociatedDevices, 0 );
+			b3Assert( status == CL_SUCCESS );
+			if (numAssociatedDevices==1)
+			{
+
+				size_t binarySize;
+				char* binary ;
+
+				status = clGetProgramInfo( m_cpProgram, CL_PROGRAM_BINARY_SIZES, sizeof(size_t), &binarySize, 0 );
+				b3Assert( status == CL_SUCCESS );
+
+				binary = (char*)malloc(sizeof(char)*binarySize);
+
+				status = clGetProgramInfo( m_cpProgram, CL_PROGRAM_BINARIES, sizeof(char*), &binary, 0 );
+				b3Assert( status == CL_SUCCESS );
+
+				{
+					FILE* file=0;
+#ifdef _MSC_VER
+					if (fopen_s(&file,binaryFileName, "wb")!=0)
+						file=0;
+#else
+					file = fopen(binaryFileName, "wb");
+#endif
+					if (file)
+					{
+						fwrite( binary, sizeof(char), binarySize, file );
+						fclose( file );
+					} else
+					{
+						b3Warning("cannot write file %s\n", binaryFileName);
+					}
+				}
+
+				free (binary);
+			}
+		}
+
+		free(compileFlags);
+
+	}
+	return m_cpProgram;
+}
+
+
+cl_kernel b3OpenCLUtils_compileCLKernelFromString(cl_context clContext, cl_device_id device, const char* kernelSource, const char* kernelName, cl_int* pErrNum, cl_program prog, const char* additionalMacros )
+{
+
+	cl_kernel kernel;
+	cl_int localErrNum;
+
+	cl_program m_cpProgram = prog;
+
+	b3Printf("compiling kernel %s ",kernelName);
+
+	if (!m_cpProgram)
+	{
+		m_cpProgram = b3OpenCLUtils_compileCLProgramFromString(clContext,device,kernelSource,pErrNum, additionalMacros,0, false);
+	}
+
+
+	// Create the kernel
+	kernel = clCreateKernel(m_cpProgram, kernelName, &localErrNum);
+	if (localErrNum != CL_SUCCESS)
+	{
+		b3Error("Error in clCreateKernel, Line %u in file %s, cannot find kernel function %s !!!\n\n", __LINE__, __FILE__, kernelName);
+        assert(0);
+		if (pErrNum)
+			*pErrNum = localErrNum;
+		return 0;
+	}
+
+	if (!prog && m_cpProgram)
+	{
+		clReleaseProgram(m_cpProgram);
+	}
+	b3Printf("ready. \n");
+
+
+	if (pErrNum)
+			*pErrNum = CL_SUCCESS;
+	return kernel;
+
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/Initialize/b3OpenCLUtils.h b/thirdparty/bullet/Bullet3OpenCL/Initialize/b3OpenCLUtils.h
new file mode 100644
index 0000000000..db6466e76b
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/Initialize/b3OpenCLUtils.h
@@ -0,0 +1,194 @@
+/*
+Bullet Continuous Collision Detection and Physics Library, http://bulletphysics.org
+Copyright (C) 2006 - 2011 Sony Computer Entertainment Inc. 
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+//original author: Roman Ponomarev
+//cleanup by Erwin Coumans
+
+#ifndef B3_OPENCL_UTILS_H
+#define B3_OPENCL_UTILS_H
+
+#include "b3OpenCLInclude.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+
+///C API for OpenCL utilities: convenience functions, see below for C++ API
+
+/// CL Context optionally takes a GL context. This is a generic type because we don't really want this code
+/// to have to understand GL types. It is a HGLRC in _WIN32 or a GLXContext otherwise.
+cl_context 	b3OpenCLUtils_createContextFromType(cl_device_type deviceType, cl_int* pErrNum, void* pGLCtx , void* pGLDC , int preferredDeviceIndex , int preferredPlatformIndex, cl_platform_id* platformId);
+	
+int b3OpenCLUtils_getNumDevices(cl_context cxMainContext);
+
+cl_device_id b3OpenCLUtils_getDevice(cl_context cxMainContext, int nr);
+
+void b3OpenCLUtils_printDeviceInfo(cl_device_id device);
+
+cl_kernel b3OpenCLUtils_compileCLKernelFromString( cl_context clContext,cl_device_id device, const char* kernelSource, const char* kernelName, cl_int* pErrNum, cl_program prog,const char* additionalMacros);
+
+//optional
+cl_program b3OpenCLUtils_compileCLProgramFromString( cl_context clContext,cl_device_id device, const char* kernelSource, cl_int* pErrNum,const char* additionalMacros  , const char* srcFileNameForCaching, bool disableBinaryCaching);
+
+//the following optional APIs provide access using specific platform information
+int b3OpenCLUtils_getNumPlatforms(cl_int* pErrNum);
+
+///get the nr'th platform, where nr is in the range [0..getNumPlatforms)
+cl_platform_id b3OpenCLUtils_getPlatform(int nr, cl_int* pErrNum);
+
+
+void b3OpenCLUtils_printPlatformInfo(cl_platform_id platform);
+
+const char* b3OpenCLUtils_getSdkVendorName();
+	
+///set the path (directory/folder) where the compiled OpenCL kernel are stored
+void b3OpenCLUtils_setCachePath(const char* path);
+	
+cl_context 	b3OpenCLUtils_createContextFromPlatform(cl_platform_id platform, cl_device_type deviceType, cl_int* pErrNum, void* pGLCtx , void* pGLDC ,int preferredDeviceIndex , int preferredPlatformIndex);
+
+#ifdef __cplusplus
+}
+
+#define B3_MAX_STRING_LENGTH 1024
+
+typedef struct
+{
+	char m_deviceName[B3_MAX_STRING_LENGTH];
+	char m_deviceVendor[B3_MAX_STRING_LENGTH];
+	char m_driverVersion[B3_MAX_STRING_LENGTH];
+	char m_deviceExtensions[B3_MAX_STRING_LENGTH];
+
+	cl_device_type		m_deviceType;
+	cl_uint 				m_computeUnits;
+	size_t 					m_workitemDims;
+	size_t 					m_workItemSize[3];
+	size_t 					m_image2dMaxWidth;
+	size_t 					m_image2dMaxHeight;
+	size_t 					m_image3dMaxWidth;
+	size_t 					m_image3dMaxHeight;
+	size_t 					m_image3dMaxDepth;
+	size_t 					m_workgroupSize;
+	cl_uint 				m_clockFrequency;
+	cl_ulong				m_constantBufferSize;
+	cl_ulong				m_localMemSize;
+	cl_ulong				m_globalMemSize;
+    cl_bool					m_errorCorrectionSupport;
+	cl_device_local_mem_type m_localMemType;
+	cl_uint					m_maxReadImageArgs;
+	cl_uint					m_maxWriteImageArgs;
+
+
+
+	cl_uint 				m_addressBits;
+	cl_ulong				m_maxMemAllocSize;
+	cl_command_queue_properties m_queueProperties;
+	cl_bool					m_imageSupport;
+	cl_uint					m_vecWidthChar;
+	cl_uint					m_vecWidthShort;
+	cl_uint					m_vecWidthInt;
+	cl_uint					m_vecWidthLong;
+	cl_uint					m_vecWidthFloat;
+	cl_uint					m_vecWidthDouble;
+
+} b3OpenCLDeviceInfo;
+
+struct b3OpenCLPlatformInfo
+{
+	char m_platformVendor[B3_MAX_STRING_LENGTH];
+	char m_platformName[B3_MAX_STRING_LENGTH];
+	char m_platformVersion[B3_MAX_STRING_LENGTH];
+	
+	b3OpenCLPlatformInfo()
+	{
+		m_platformVendor[0]=0;
+		m_platformName[0]=0;
+		m_platformVersion[0]=0;
+	}
+};
+
+
+///C++ API for OpenCL utilities: convenience functions
+struct b3OpenCLUtils
+{
+	/// CL Context optionally takes a GL context. This is a generic type because we don't really want this code
+	/// to have to understand GL types. It is a HGLRC in _WIN32 or a GLXContext otherwise.
+	static inline cl_context 	createContextFromType(cl_device_type deviceType, cl_int* pErrNum, void* pGLCtx = 0, void* pGLDC = 0, int preferredDeviceIndex = -1, int preferredPlatformIndex= - 1, cl_platform_id* platformId=0)
+	{
+		return b3OpenCLUtils_createContextFromType(deviceType, pErrNum, pGLCtx , pGLDC , preferredDeviceIndex, preferredPlatformIndex, platformId);
+	}
+	
+	static inline int getNumDevices(cl_context cxMainContext)
+	{
+		return b3OpenCLUtils_getNumDevices(cxMainContext);
+	}
+	static inline cl_device_id getDevice(cl_context cxMainContext, int nr)
+	{
+		return b3OpenCLUtils_getDevice(cxMainContext,nr);
+	}
+
+	static void getDeviceInfo(cl_device_id device, b3OpenCLDeviceInfo* info);
+
+	static inline void printDeviceInfo(cl_device_id device)
+	{
+		b3OpenCLUtils_printDeviceInfo(device);
+	}
+
+	static inline cl_kernel compileCLKernelFromString( cl_context clContext,cl_device_id device, const char* kernelSource, const char* kernelName, cl_int* pErrNum=0, cl_program prog=0,const char* additionalMacros = "" )
+	{
+		return b3OpenCLUtils_compileCLKernelFromString(clContext,device, kernelSource,  kernelName, pErrNum, prog,additionalMacros);
+	}
+
+	//optional
+	static inline cl_program compileCLProgramFromString( cl_context clContext,cl_device_id device, const char* kernelSource, cl_int* pErrNum=0,const char* additionalMacros = "" , const char* srcFileNameForCaching=0, bool disableBinaryCaching=false)
+	{
+		return b3OpenCLUtils_compileCLProgramFromString(clContext,device, kernelSource, pErrNum,additionalMacros, srcFileNameForCaching, disableBinaryCaching);
+	}
+
+	//the following optional APIs provide access using specific platform information
+	static inline int getNumPlatforms(cl_int* pErrNum=0)
+	{
+		return b3OpenCLUtils_getNumPlatforms(pErrNum);
+	}
+	///get the nr'th platform, where nr is in the range [0..getNumPlatforms)
+	static inline cl_platform_id getPlatform(int nr, cl_int* pErrNum=0)
+	{
+		return b3OpenCLUtils_getPlatform(nr,pErrNum);
+	}
+	
+	static void getPlatformInfo(cl_platform_id platform, b3OpenCLPlatformInfo* platformInfo);
+
+	static inline void printPlatformInfo(cl_platform_id platform)
+	{
+		b3OpenCLUtils_printPlatformInfo(platform);
+	}
+
+	static inline const char* getSdkVendorName()
+	{
+		return b3OpenCLUtils_getSdkVendorName();
+	}
+	static inline cl_context 	createContextFromPlatform(cl_platform_id platform, cl_device_type deviceType, cl_int* pErrNum, void* pGLCtx = 0, void* pGLDC = 0,int preferredDeviceIndex = -1, int preferredPlatformIndex= -1)
+	{
+		return b3OpenCLUtils_createContextFromPlatform(platform, deviceType, pErrNum, pGLCtx,pGLDC,preferredDeviceIndex, preferredPlatformIndex);
+	}
+	static void setCachePath(const char* path)
+	{
+		b3OpenCLUtils_setCachePath(path);
+	}
+};
+
+#endif //__cplusplus
+
+#endif // B3_OPENCL_UTILS_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3BvhInfo.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3BvhInfo.h
new file mode 100644
index 0000000000..872f039506
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3BvhInfo.h
@@ -0,0 +1,18 @@
+#ifndef B3_BVH_INFO_H
+#define B3_BVH_INFO_H
+
+#include "Bullet3Common/b3Vector3.h"
+
+struct b3BvhInfo
+{
+	b3Vector3	m_aabbMin;
+	b3Vector3	m_aabbMax;
+	b3Vector3	m_quantization;
+	int			m_numNodes;
+	int			m_numSubTrees;
+	int			m_nodeOffset;
+	int			m_subTreeOffset;
+
+};
+
+#endif //B3_BVH_INFO_H
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ContactCache.cpp b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ContactCache.cpp
new file mode 100644
index 0000000000..cb30ee939b
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ContactCache.cpp
@@ -0,0 +1,258 @@
+
+#if 0
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2006 Erwin Coumans  http://continuousphysics.com/Bullet/
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+
+#include "b3ContactCache.h"
+#include "Bullet3Common/b3Transform.h"
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Contact4Data.h"
+
+b3Scalar					gContactBreakingThreshold = b3Scalar(0.02);
+
+///gContactCalcArea3Points will approximate the convex hull area using 3 points
+///when setting it to false, it will use 4 points to compute the area: it is more accurate but slower
+bool						gContactCalcArea3Points = true;
+
+
+
+
+static inline b3Scalar calcArea4Points(const b3Vector3 &p0,const b3Vector3 &p1,const b3Vector3 &p2,const b3Vector3 &p3)
+{
+	// It calculates possible 3 area constructed from random 4 points and returns the biggest one.
+
+	b3Vector3 a[3],b[3];
+	a[0] = p0 - p1;
+	a[1] = p0 - p2;
+	a[2] = p0 - p3;
+	b[0] = p2 - p3;
+	b[1] = p1 - p3;
+	b[2] = p1 - p2;
+
+	//todo: Following 3 cross production can be easily optimized by SIMD.
+	b3Vector3 tmp0 = a[0].cross(b[0]);
+	b3Vector3 tmp1 = a[1].cross(b[1]);
+	b3Vector3 tmp2 = a[2].cross(b[2]);
+
+	return b3Max(b3Max(tmp0.length2(),tmp1.length2()),tmp2.length2());
+}
+#if 0
+
+//using localPointA for all points
+int b3ContactCache::sortCachedPoints(const b3Vector3& pt) 
+{
+		//calculate 4 possible cases areas, and take biggest area
+		//also need to keep 'deepest'
+		
+		int maxPenetrationIndex = -1;
+#define KEEP_DEEPEST_POINT 1
+#ifdef KEEP_DEEPEST_POINT
+		b3Scalar maxPenetration = pt.getDistance();
+		for (int i=0;i<4;i++)
+		{
+			if (m_pointCache[i].getDistance() < maxPenetration)
+			{
+				maxPenetrationIndex = i;
+				maxPenetration = m_pointCache[i].getDistance();
+			}
+		}
+#endif //KEEP_DEEPEST_POINT
+		
+		b3Scalar res0(b3Scalar(0.)),res1(b3Scalar(0.)),res2(b3Scalar(0.)),res3(b3Scalar(0.));
+
+	if (gContactCalcArea3Points)
+	{
+		if (maxPenetrationIndex != 0)
+		{
+			b3Vector3 a0 = pt.m_localPointA-m_pointCache[1].m_localPointA;
+			b3Vector3 b0 = m_pointCache[3].m_localPointA-m_pointCache[2].m_localPointA;
+			b3Vector3 cross = a0.cross(b0);
+			res0 = cross.length2();
+		}
+		if (maxPenetrationIndex != 1)
+		{
+			b3Vector3 a1 = pt.m_localPointA-m_pointCache[0].m_localPointA;
+			b3Vector3 b1 = m_pointCache[3].m_localPointA-m_pointCache[2].m_localPointA;
+			b3Vector3 cross = a1.cross(b1);
+			res1 = cross.length2();
+		}
+
+		if (maxPenetrationIndex != 2)
+		{
+			b3Vector3 a2 = pt.m_localPointA-m_pointCache[0].m_localPointA;
+			b3Vector3 b2 = m_pointCache[3].m_localPointA-m_pointCache[1].m_localPointA;
+			b3Vector3 cross = a2.cross(b2);
+			res2 = cross.length2();
+		}
+
+		if (maxPenetrationIndex != 3)
+		{
+			b3Vector3 a3 = pt.m_localPointA-m_pointCache[0].m_localPointA;
+			b3Vector3 b3 = m_pointCache[2].m_localPointA-m_pointCache[1].m_localPointA;
+			b3Vector3 cross = a3.cross(b3);
+			res3 = cross.length2();
+		}
+	} 
+	else
+	{
+		if(maxPenetrationIndex != 0) {
+			res0 = calcArea4Points(pt.m_localPointA,m_pointCache[1].m_localPointA,m_pointCache[2].m_localPointA,m_pointCache[3].m_localPointA);
+		}
+
+		if(maxPenetrationIndex != 1) {
+			res1 = calcArea4Points(pt.m_localPointA,m_pointCache[0].m_localPointA,m_pointCache[2].m_localPointA,m_pointCache[3].m_localPointA);
+		}
+
+		if(maxPenetrationIndex != 2) {
+			res2 = calcArea4Points(pt.m_localPointA,m_pointCache[0].m_localPointA,m_pointCache[1].m_localPointA,m_pointCache[3].m_localPointA);
+		}
+
+		if(maxPenetrationIndex != 3) {
+			res3 = calcArea4Points(pt.m_localPointA,m_pointCache[0].m_localPointA,m_pointCache[1].m_localPointA,m_pointCache[2].m_localPointA);
+		}
+	}
+	b3Vector4 maxvec(res0,res1,res2,res3);
+	int biggestarea = maxvec.closestAxis4();
+	return biggestarea;
+	
+}
+
+
+int b3ContactCache::getCacheEntry(const b3Vector3& newPoint) const
+{
+	b3Scalar shortestDist =  getContactBreakingThreshold() * getContactBreakingThreshold();
+	int size = getNumContacts();
+	int nearestPoint = -1;
+	for( int i = 0; i < size; i++ )
+	{
+		const b3Vector3 &mp = m_pointCache[i];
+
+		b3Vector3 diffA =  mp.m_localPointA- newPoint.m_localPointA;
+		const b3Scalar distToManiPoint = diffA.dot(diffA);
+		if( distToManiPoint < shortestDist )
+		{
+			shortestDist = distToManiPoint;
+			nearestPoint = i;
+		}
+	}
+	return nearestPoint;
+}
+
+int b3ContactCache::addManifoldPoint(const b3Vector3& newPoint)
+{
+	b3Assert(validContactDistance(newPoint));
+	
+	int insertIndex = getNumContacts();
+	if (insertIndex == MANIFOLD_CACHE_SIZE)
+	{
+#if MANIFOLD_CACHE_SIZE >= 4
+		//sort cache so best points come first, based on area
+		insertIndex = sortCachedPoints(newPoint);
+#else
+		insertIndex = 0;
+#endif
+		clearUserCache(m_pointCache[insertIndex]);
+		
+	} else
+	{
+		m_cachedPoints++;
+
+		
+	}
+	if (insertIndex<0)
+		insertIndex=0;
+
+	//b3Assert(m_pointCache[insertIndex].m_userPersistentData==0);
+	m_pointCache[insertIndex] = newPoint;
+	return insertIndex;
+}
+
+#endif
+
+bool b3ContactCache::validContactDistance(const b3Vector3& pt)
+{
+	return pt.w <= gContactBreakingThreshold;
+}
+
+void b3ContactCache::removeContactPoint(struct b3Contact4Data& newContactCache,int i)
+{
+	int numContacts = b3Contact4Data_getNumPoints(&newContactCache);
+	if (i!=(numContacts-1))
+	{
+		b3Swap(newContactCache.m_localPosA[i],newContactCache.m_localPosA[numContacts-1]);
+		b3Swap(newContactCache.m_localPosB[i],newContactCache.m_localPosB[numContacts-1]);
+		b3Swap(newContactCache.m_worldPosB[i],newContactCache.m_worldPosB[numContacts-1]);
+	}
+	b3Contact4Data_setNumPoints(&newContactCache,numContacts-1);
+
+}
+
+
+void b3ContactCache::refreshContactPoints(const b3Transform& trA,const b3Transform& trB, struct b3Contact4Data& contacts)
+{
+
+	int numContacts = b3Contact4Data_getNumPoints(&contacts);
+	
+
+	int i;
+	/// first refresh worldspace positions and distance
+	for (i=numContacts-1;i>=0;i--)
+	{
+		b3Vector3 worldPosA = trA( contacts.m_localPosA[i]);
+		b3Vector3 worldPosB = trB( contacts.m_localPosB[i]);
+		contacts.m_worldPosB[i] = worldPosB;
+		float distance = (worldPosA -  worldPosB).dot(contacts.m_worldNormalOnB);
+		contacts.m_worldPosB[i].w = distance;
+	}
+
+	/// then 
+	b3Scalar distance2d;
+	b3Vector3 projectedDifference,projectedPoint;
+	for (i=numContacts-1;i>=0;i--)
+	{
+		b3Vector3 worldPosA = trA( contacts.m_localPosA[i]);
+		b3Vector3 worldPosB = trB( contacts.m_localPosB[i]);
+		b3Vector3&pt = contacts.m_worldPosB[i];
+		//contact becomes invalid when signed distance exceeds margin (projected on contactnormal direction)
+		if (!validContactDistance(pt))
+		{
+			removeContactPoint(contacts,i);
+		} else
+		{
+			//contact also becomes invalid when relative movement orthogonal to normal exceeds margin
+			projectedPoint = worldPosA - contacts.m_worldNormalOnB * contacts.m_worldPosB[i].w;
+			projectedDifference = contacts.m_worldPosB[i] - projectedPoint;
+			distance2d = projectedDifference.dot(projectedDifference);
+			if (distance2d  > gContactBreakingThreshold*gContactBreakingThreshold )
+			{
+				removeContactPoint(contacts,i);
+			} else
+			{
+				////contact point processed callback
+				//if (gContactProcessedCallback)
+				//	(*gContactProcessedCallback)(manifoldPoint,(void*)m_body0,(void*)m_body1);
+			}
+		}
+	}
+	
+
+}
+
+
+
+
+
+#endif
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ContactCache.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ContactCache.h
new file mode 100644
index 0000000000..d6c9b0a07e
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ContactCache.h
@@ -0,0 +1,80 @@
+
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2013 Erwin Coumans  http://bulletphysics.org
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+#ifndef B3_CONTACT_CACHE_H
+#define B3_CONTACT_CACHE_H
+
+
+#include "Bullet3Common/b3Vector3.h"
+#include "Bullet3Common/b3Transform.h"
+#include "Bullet3Common/b3AlignedAllocator.h"
+
+
+///maximum contact breaking and merging threshold
+extern b3Scalar gContactBreakingThreshold;
+
+
+
+#define MANIFOLD_CACHE_SIZE 4
+
+///b3ContactCache is a contact point cache, it stays persistent as long as objects are overlapping in the broadphase.
+///Those contact points are created by the collision narrow phase.
+///The cache can be empty, or hold 1,2,3 or 4 points. Some collision algorithms (GJK) might only add one point at a time.
+///updates/refreshes old contact points, and throw them away if necessary (distance becomes too large)
+///reduces the cache to 4 points, when more then 4 points are added, using following rules:
+///the contact point with deepest penetration is always kept, and it tries to maximuze the area covered by the points
+///note that some pairs of objects might have more then one contact manifold.
+B3_ATTRIBUTE_ALIGNED16( class) b3ContactCache
+{
+
+	
+
+	
+	/// sort cached points so most isolated points come first
+	int	sortCachedPoints(const b3Vector3& pt);
+
+	
+
+public:
+
+	B3_DECLARE_ALIGNED_ALLOCATOR();
+
+	
+	
+	int addManifoldPoint( const b3Vector3& newPoint);
+
+	/*void replaceContactPoint(const b3Vector3& newPoint,int insertIndex)
+	{
+		b3Assert(validContactDistance(newPoint));
+		m_pointCache[insertIndex] = newPoint;
+	}
+	*/
+
+
+	
+	static bool validContactDistance(const b3Vector3& pt);
+	
+	/// calculated new worldspace coordinates and depth, and reject points that exceed the collision margin
+	static void	refreshContactPoints(  const b3Transform& trA,const b3Transform& trB, struct b3Contact4Data& newContactCache);
+
+	static void removeContactPoint(struct b3Contact4Data& newContactCache,int i);
+	
+
+};
+
+
+
+#endif //B3_CONTACT_CACHE_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ConvexHullContact.cpp b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ConvexHullContact.cpp
new file mode 100644
index 0000000000..fb435aa7fd
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ConvexHullContact.cpp
@@ -0,0 +1,4733 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2011 Advanced Micro Devices, Inc.  http://bulletphysics.org
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+bool findSeparatingAxisOnGpu = true;
+bool splitSearchSepAxisConcave = false;
+bool splitSearchSepAxisConvex = true;
+bool useMprGpu = true;//use mpr for edge-edge  (+contact point) or sat. Needs testing on main OpenCL platforms, before enabling...
+bool bvhTraversalKernelGPU = true;
+bool findConcaveSeparatingAxisKernelGPU = true;
+bool clipConcaveFacesAndFindContactsCPU = false;//false;//true;
+bool clipConvexFacesAndFindContactsCPU = false;//false;//true;
+bool reduceConcaveContactsOnGPU = true;//false;
+bool reduceConvexContactsOnGPU = true;//false;
+bool findConvexClippingFacesGPU = true;
+bool useGjk = false;///option for CPU/host testing, when findSeparatingAxisOnGpu = false
+bool useGjkContacts = false;//////option for CPU/host testing when findSeparatingAxisOnGpu = false
+
+
+static int myframecount=0;///for testing
+
+///This file was written by Erwin Coumans
+///Separating axis rest based on work from Pierre Terdiman, see
+///And contact clipping based on work from Simon Hobbs
+
+//#define B3_DEBUG_SAT_FACE
+
+//#define CHECK_ON_HOST
+
+#ifdef CHECK_ON_HOST
+//#define PERSISTENT_CONTACTS_HOST
+#endif
+
+int b3g_actualSATPairTests=0;
+
+#include "b3ConvexHullContact.h"
+#include <string.h>//memcpy
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3ConvexPolyhedronData.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3MprPenetration.h"
+
+#include "Bullet3OpenCL/NarrowphaseCollision/b3ContactCache.h"
+#include "Bullet3Geometry/b3AabbUtil.h"
+
+typedef b3AlignedObjectArray<b3Vector3> b3VertexArray;
+
+
+#include <float.h> //for FLT_MAX
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h"
+//#include "AdlQuaternion.h"
+
+#include "kernels/satKernels.h"
+#include "kernels/mprKernels.h"
+
+#include "kernels/satConcaveKernels.h"
+
+#include "kernels/satClipHullContacts.h"
+#include "kernels/bvhTraversal.h"
+#include "kernels/primitiveContacts.h"
+
+
+#include "Bullet3Geometry/b3AabbUtil.h"
+
+#define BT_NARROWPHASE_SAT_PATH "src/Bullet3OpenCL/NarrowphaseCollision/kernels/sat.cl"
+#define BT_NARROWPHASE_SAT_CONCAVE_PATH "src/Bullet3OpenCL/NarrowphaseCollision/kernels/satConcave.cl"
+
+#define BT_NARROWPHASE_MPR_PATH "src/Bullet3OpenCL/NarrowphaseCollision/kernels/mpr.cl"
+
+
+#define BT_NARROWPHASE_CLIPHULL_PATH "src/Bullet3OpenCL/NarrowphaseCollision/kernels/satClipHullContacts.cl"
+#define BT_NARROWPHASE_BVH_TRAVERSAL_PATH "src/Bullet3OpenCL/NarrowphaseCollision/kernels/bvhTraversal.cl"
+#define BT_NARROWPHASE_PRIMITIVE_CONTACT_PATH "src/Bullet3OpenCL/NarrowphaseCollision/kernels/primitiveContacts.cl"
+
+
+#ifndef __global
+#define __global
+#endif
+
+#ifndef __kernel
+#define __kernel
+#endif
+
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3BvhTraversal.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3FindConcaveSatAxis.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3ClipFaces.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3NewContactReduction.h"
+
+
+
+#define dot3F4 b3Dot
+
+GpuSatCollision::GpuSatCollision(cl_context ctx,cl_device_id device, cl_command_queue  q )
+:m_context(ctx),
+m_device(device),
+m_queue(q),
+
+m_findSeparatingAxisKernel(0),
+m_findSeparatingAxisVertexFaceKernel(0),
+m_findSeparatingAxisEdgeEdgeKernel(0),
+m_unitSphereDirections(m_context,m_queue),
+
+m_totalContactsOut(m_context, m_queue),
+m_sepNormals(m_context, m_queue),
+m_dmins(m_context,m_queue),
+
+m_hasSeparatingNormals(m_context, m_queue),
+m_concaveSepNormals(m_context, m_queue),
+m_concaveHasSeparatingNormals(m_context,m_queue),
+m_numConcavePairsOut(m_context, m_queue),
+
+
+m_gpuCompoundPairs(m_context, m_queue),
+
+
+m_gpuCompoundSepNormals(m_context, m_queue),
+m_gpuHasCompoundSepNormals(m_context, m_queue),
+
+m_numCompoundPairsOut(m_context, m_queue)
+{
+	m_totalContactsOut.push_back(0);
+	
+	cl_int errNum=0;
+
+	if (1)
+	{
+		const char* mprSrc = mprKernelsCL;
+		
+		const char* srcConcave = satConcaveKernelsCL;
+		char flags[1024]={0};
+//#ifdef CL_PLATFORM_INTEL
+//		sprintf(flags,"-g -s \"%s\"","C:/develop/bullet3_experiments2/opencl/gpu_narrowphase/kernels/sat.cl");
+//#endif
+		m_mprPenetrationKernel  = 0;
+		m_findSeparatingAxisUnitSphereKernel = 0;
+
+		if (useMprGpu)
+		{
+			cl_program mprProg = b3OpenCLUtils::compileCLProgramFromString(m_context,m_device,mprSrc,&errNum,flags,BT_NARROWPHASE_MPR_PATH);
+			b3Assert(errNum==CL_SUCCESS);
+		
+			m_mprPenetrationKernel  = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,mprSrc, "mprPenetrationKernel",&errNum,mprProg );
+			b3Assert(m_mprPenetrationKernel);
+			b3Assert(errNum==CL_SUCCESS);
+
+			m_findSeparatingAxisUnitSphereKernel =  b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,mprSrc, "findSeparatingAxisUnitSphereKernel",&errNum,mprProg );
+			b3Assert(m_findSeparatingAxisUnitSphereKernel);
+            b3Assert(errNum==CL_SUCCESS);
+
+
+			int numDirections = sizeof(unitSphere162)/sizeof(b3Vector3);
+			m_unitSphereDirections.resize(numDirections);
+			m_unitSphereDirections.copyFromHostPointer(unitSphere162,numDirections,0,true);
+
+
+		}
+
+
+		cl_program satProg = b3OpenCLUtils::compileCLProgramFromString(m_context,m_device,satKernelsCL,&errNum,flags,BT_NARROWPHASE_SAT_PATH);
+		b3Assert(errNum==CL_SUCCESS);
+
+		cl_program satConcaveProg = b3OpenCLUtils::compileCLProgramFromString(m_context,m_device,srcConcave,&errNum,flags,BT_NARROWPHASE_SAT_CONCAVE_PATH);
+		b3Assert(errNum==CL_SUCCESS);
+
+		m_findSeparatingAxisKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,satKernelsCL, "findSeparatingAxisKernel",&errNum,satProg );
+		b3Assert(m_findSeparatingAxisKernel);
+		b3Assert(errNum==CL_SUCCESS);
+
+
+		m_findSeparatingAxisVertexFaceKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,satKernelsCL, "findSeparatingAxisVertexFaceKernel",&errNum,satProg );
+		b3Assert(m_findSeparatingAxisVertexFaceKernel);
+
+		m_findSeparatingAxisEdgeEdgeKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,satKernelsCL, "findSeparatingAxisEdgeEdgeKernel",&errNum,satProg );
+		b3Assert(m_findSeparatingAxisVertexFaceKernel);
+
+
+		m_findConcaveSeparatingAxisKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,satKernelsCL, "findConcaveSeparatingAxisKernel",&errNum,satProg );
+		b3Assert(m_findConcaveSeparatingAxisKernel);
+		b3Assert(errNum==CL_SUCCESS);
+        
+        m_findConcaveSeparatingAxisVertexFaceKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,srcConcave, "findConcaveSeparatingAxisVertexFaceKernel",&errNum,satConcaveProg );
+		b3Assert(m_findConcaveSeparatingAxisVertexFaceKernel);
+		b3Assert(errNum==CL_SUCCESS);
+        
+        m_findConcaveSeparatingAxisEdgeEdgeKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,srcConcave, "findConcaveSeparatingAxisEdgeEdgeKernel",&errNum,satConcaveProg );
+		b3Assert(m_findConcaveSeparatingAxisEdgeEdgeKernel);
+		b3Assert(errNum==CL_SUCCESS);
+        
+     
+        
+		
+		m_findCompoundPairsKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,satKernelsCL, "findCompoundPairsKernel",&errNum,satProg );
+		b3Assert(m_findCompoundPairsKernel);
+		b3Assert(errNum==CL_SUCCESS);
+		m_processCompoundPairsKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,satKernelsCL, "processCompoundPairsKernel",&errNum,satProg );
+		b3Assert(m_processCompoundPairsKernel);
+		b3Assert(errNum==CL_SUCCESS);
+	}
+
+	if (1)
+	{
+		const char* srcClip = satClipKernelsCL;
+
+		char flags[1024]={0};
+//#ifdef CL_PLATFORM_INTEL
+//		sprintf(flags,"-g -s \"%s\"","C:/develop/bullet3_experiments2/opencl/gpu_narrowphase/kernels/satClipHullContacts.cl");
+//#endif
+
+		cl_program satClipContactsProg = b3OpenCLUtils::compileCLProgramFromString(m_context,m_device,srcClip,&errNum,flags,BT_NARROWPHASE_CLIPHULL_PATH);
+		b3Assert(errNum==CL_SUCCESS);
+
+		m_clipHullHullKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,srcClip, "clipHullHullKernel",&errNum,satClipContactsProg);
+		b3Assert(errNum==CL_SUCCESS);
+
+		m_clipCompoundsHullHullKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,srcClip, "clipCompoundsHullHullKernel",&errNum,satClipContactsProg);
+		b3Assert(errNum==CL_SUCCESS);
+		
+
+        m_findClippingFacesKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,srcClip, "findClippingFacesKernel",&errNum,satClipContactsProg);
+		b3Assert(errNum==CL_SUCCESS);
+
+        m_clipFacesAndFindContacts = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,srcClip, "clipFacesAndFindContactsKernel",&errNum,satClipContactsProg);
+		b3Assert(errNum==CL_SUCCESS);        
+
+		m_clipHullHullConcaveConvexKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,srcClip, "clipHullHullConcaveConvexKernel",&errNum,satClipContactsProg);
+		b3Assert(errNum==CL_SUCCESS);
+
+//		m_extractManifoldAndAddContactKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,srcClip, "extractManifoldAndAddContactKernel",&errNum,satClipContactsProg);
+	//	b3Assert(errNum==CL_SUCCESS);
+
+        m_newContactReductionKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,srcClip,
+                            "newContactReductionKernel",&errNum,satClipContactsProg);
+		b3Assert(errNum==CL_SUCCESS);
+	}
+   else
+	{
+		m_clipHullHullKernel=0;
+		m_clipCompoundsHullHullKernel = 0;
+        m_findClippingFacesKernel = 0;
+        m_newContactReductionKernel=0;
+        m_clipFacesAndFindContacts = 0;
+		m_clipHullHullConcaveConvexKernel = 0;
+//		m_extractManifoldAndAddContactKernel = 0;
+	}
+
+	 if (1)
+	{
+		const char* srcBvh = bvhTraversalKernelCL;
+		cl_program bvhTraversalProg = b3OpenCLUtils::compileCLProgramFromString(m_context,m_device,srcBvh,&errNum,"",BT_NARROWPHASE_BVH_TRAVERSAL_PATH);
+		b3Assert(errNum==CL_SUCCESS);
+
+		m_bvhTraversalKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,srcBvh, "bvhTraversalKernel",&errNum,bvhTraversalProg,"");
+		b3Assert(errNum==CL_SUCCESS);
+
+	}
+        
+	 {
+		 const char* primitiveContactsSrc = primitiveContactsKernelsCL;
+		cl_program primitiveContactsProg = b3OpenCLUtils::compileCLProgramFromString(m_context,m_device,primitiveContactsSrc,&errNum,"",BT_NARROWPHASE_PRIMITIVE_CONTACT_PATH);
+		b3Assert(errNum==CL_SUCCESS);
+
+		m_primitiveContactsKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,primitiveContactsSrc, "primitiveContactsKernel",&errNum,primitiveContactsProg,"");
+		b3Assert(errNum==CL_SUCCESS);
+
+		m_findConcaveSphereContactsKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,primitiveContactsSrc, "findConcaveSphereContactsKernel",&errNum,primitiveContactsProg );
+		b3Assert(errNum==CL_SUCCESS);
+		b3Assert(m_findConcaveSphereContactsKernel);
+
+		m_processCompoundPairsPrimitivesKernel = b3OpenCLUtils::compileCLKernelFromString(m_context, m_device,primitiveContactsSrc, "processCompoundPairsPrimitivesKernel",&errNum,primitiveContactsProg,"");
+		b3Assert(errNum==CL_SUCCESS);
+		b3Assert(m_processCompoundPairsPrimitivesKernel);
+		 
+	 }
+	
+
+}
+
+GpuSatCollision::~GpuSatCollision()
+{
+	
+	if (m_findSeparatingAxisVertexFaceKernel)
+		clReleaseKernel(m_findSeparatingAxisVertexFaceKernel);
+
+	if (m_findSeparatingAxisEdgeEdgeKernel)
+		clReleaseKernel(m_findSeparatingAxisEdgeEdgeKernel);
+
+	if (m_findSeparatingAxisUnitSphereKernel)
+		clReleaseKernel(m_findSeparatingAxisUnitSphereKernel);
+
+	if (m_mprPenetrationKernel)
+		clReleaseKernel(m_mprPenetrationKernel);
+
+
+	if (m_findSeparatingAxisKernel)
+		clReleaseKernel(m_findSeparatingAxisKernel);
+
+    if (m_findConcaveSeparatingAxisVertexFaceKernel)
+        clReleaseKernel(m_findConcaveSeparatingAxisVertexFaceKernel);
+
+    
+    if (m_findConcaveSeparatingAxisEdgeEdgeKernel)
+        clReleaseKernel(m_findConcaveSeparatingAxisEdgeEdgeKernel);
+    
+	if (m_findConcaveSeparatingAxisKernel)
+		clReleaseKernel(m_findConcaveSeparatingAxisKernel);
+
+	if (m_findCompoundPairsKernel)
+		clReleaseKernel(m_findCompoundPairsKernel);
+
+	if (m_processCompoundPairsKernel)
+		clReleaseKernel(m_processCompoundPairsKernel);
+    
+    if (m_findClippingFacesKernel)
+        clReleaseKernel(m_findClippingFacesKernel);
+   
+    if (m_clipFacesAndFindContacts)
+        clReleaseKernel(m_clipFacesAndFindContacts);
+    if (m_newContactReductionKernel)
+        clReleaseKernel(m_newContactReductionKernel);
+	if (m_primitiveContactsKernel)
+		clReleaseKernel(m_primitiveContactsKernel);
+    
+	if (m_findConcaveSphereContactsKernel)
+		clReleaseKernel(m_findConcaveSphereContactsKernel);
+
+	if (m_processCompoundPairsPrimitivesKernel)
+		clReleaseKernel(m_processCompoundPairsPrimitivesKernel);
+
+	if (m_clipHullHullKernel)
+		clReleaseKernel(m_clipHullHullKernel);
+	if (m_clipCompoundsHullHullKernel)
+		clReleaseKernel(m_clipCompoundsHullHullKernel);
+
+	if (m_clipHullHullConcaveConvexKernel)
+		clReleaseKernel(m_clipHullHullConcaveConvexKernel);
+//	if (m_extractManifoldAndAddContactKernel)
+	//	clReleaseKernel(m_extractManifoldAndAddContactKernel);
+
+	if (m_bvhTraversalKernel)
+		clReleaseKernel(m_bvhTraversalKernel);
+
+}
+
+struct MyTriangleCallback : public b3NodeOverlapCallback
+{
+	int m_bodyIndexA;
+	int m_bodyIndexB;
+
+	virtual void processNode(int subPart, int triangleIndex)
+	{
+		printf("bodyIndexA %d, bodyIndexB %d\n",m_bodyIndexA,m_bodyIndexB);
+		printf("triangleIndex %d\n", triangleIndex);
+	}
+};
+
+
+#define float4 b3Vector3
+#define make_float4(x,y,z,w) b3MakeVector3(x,y,z,w)
+
+float signedDistanceFromPointToPlane(const float4& point, const float4& planeEqn, float4* closestPointOnFace)
+{
+	float4 n = planeEqn;
+	n[3] = 0.f;
+	float dist = dot3F4(n, point) + planeEqn[3];
+	*closestPointOnFace = point - dist * n;
+	return dist;
+}
+
+
+
+#define cross3(a,b) (a.cross(b))
+b3Vector3 transform(const b3Vector3* v, const b3Vector3* pos, const b3Quaternion* orn)
+{
+	b3Transform tr;
+	tr.setIdentity();
+	tr.setOrigin(*pos);
+	tr.setRotation(*orn);
+	b3Vector3 res = tr(*v);
+	return res;
+}
+
+
+inline bool IsPointInPolygon(const float4& p, 
+							const b3GpuFace* face,
+							 const float4* baseVertex,
+							const  int* convexIndices,
+							float4* out)
+{
+    float4 a;
+    float4 b;
+    float4 ab;
+    float4 ap;
+    float4 v;
+
+	float4 plane = b3MakeVector3(face->m_plane.x,face->m_plane.y,face->m_plane.z,0.f);
+	
+	if (face->m_numIndices<2)
+		return false;
+
+	
+	float4 v0 = baseVertex[convexIndices[face->m_indexOffset + face->m_numIndices-1]];
+	b = v0;
+
+    for(unsigned i=0; i != face->m_numIndices; ++i)
+    {
+		a = b;
+		float4 vi = baseVertex[convexIndices[face->m_indexOffset + i]];
+		b = vi;
+        ab = b-a;
+        ap = p-a;
+        v = cross3(ab,plane);
+
+        if (b3Dot(ap, v) > 0.f)
+        {
+            float ab_m2 = b3Dot(ab, ab);
+            float rt = ab_m2 != 0.f ? b3Dot(ab, ap) / ab_m2 : 0.f;
+            if (rt <= 0.f)
+            {
+                *out = a;
+            }
+            else if (rt >= 1.f) 
+            {
+                *out = b;
+            }
+            else
+            {
+            	float s = 1.f - rt;
+				out[0].x = s * a.x + rt * b.x;
+				out[0].y = s * a.y + rt * b.y;
+				out[0].z = s * a.z + rt * b.z;
+            }
+            return false;
+        }
+    }
+    return true;
+}
+
+#define normalize3(a) (a.normalize())
+
+
+int extractManifoldSequentialGlobal( const float4* p, int nPoints, const float4& nearNormal, b3Int4* contactIdx)
+{
+	if( nPoints == 0 )
+        return 0;
+    
+    if (nPoints <=4)
+        return nPoints;
+    
+    
+    if (nPoints >64)
+        nPoints = 64;
+    
+	float4 center = b3MakeVector3(0,0,0,0);
+	{
+		
+		for (int i=0;i<nPoints;i++)
+			center += p[i];
+		center /= (float)nPoints;
+	}
+    
+	
+    
+	//	sample 4 directions
+    
+    float4 aVector = p[0] - center;
+    float4 u = cross3( nearNormal, aVector );
+    float4 v = cross3( nearNormal, u );
+    u = normalize3( u );
+    v = normalize3( v );
+    
+    
+    //keep point with deepest penetration
+    float minW= FLT_MAX;
+    
+    int minIndex=-1;
+    
+    float4 maxDots;
+    maxDots.x = FLT_MIN;
+    maxDots.y = FLT_MIN;
+    maxDots.z = FLT_MIN;
+    maxDots.w = FLT_MIN;
+    
+    //	idx, distance
+    for(int ie = 0; ie<nPoints; ie++ )
+    {
+        if (p[ie].w<minW)
+        {
+            minW = p[ie].w;
+            minIndex=ie;
+        }
+        float f;
+        float4 r = p[ie]-center;
+        f = dot3F4( u, r );
+        if (f<maxDots.x)
+        {
+            maxDots.x = f;
+            contactIdx[0].x = ie;
+        }
+        
+        f = dot3F4( -u, r );
+        if (f<maxDots.y)
+        {
+            maxDots.y = f;
+            contactIdx[0].y = ie;
+        }
+        
+        
+        f = dot3F4( v, r );
+        if (f<maxDots.z)
+        {
+            maxDots.z = f;
+            contactIdx[0].z = ie;
+        }
+        
+        f = dot3F4( -v, r );
+        if (f<maxDots.w)
+        {
+            maxDots.w = f;
+            contactIdx[0].w = ie;
+        }
+        
+    }
+    
+    if (contactIdx[0].x != minIndex && contactIdx[0].y != minIndex && contactIdx[0].z != minIndex && contactIdx[0].w != minIndex)
+    {
+        //replace the first contact with minimum (todo: replace contact with least penetration)
+        contactIdx[0].x = minIndex;
+    }
+    
+    return 4;
+    
+}
+
+
+
+#define MAX_VERTS 1024
+
+
+inline void project(const b3ConvexPolyhedronData& hull,  const float4& pos, const b3Quaternion& orn, const float4& dir, const b3AlignedObjectArray<b3Vector3>& vertices, b3Scalar& min, b3Scalar& max)
+{
+	min = FLT_MAX;
+	max = -FLT_MAX;
+	int numVerts = hull.m_numVertices;
+
+	const float4 localDir = b3QuatRotate(orn.inverse(),dir);
+
+	b3Scalar offset = dot3F4(pos,dir);
+
+	for(int i=0;i<numVerts;i++)
+	{
+		//b3Vector3 pt = trans * vertices[m_vertexOffset+i];
+		//b3Scalar dp = pt.dot(dir);
+		//b3Vector3 vertex = vertices[hull.m_vertexOffset+i];
+		b3Scalar dp = dot3F4((float4&)vertices[hull.m_vertexOffset+i],localDir);
+		//b3Assert(dp==dpL);
+		if(dp < min)	min = dp;
+		if(dp > max)	max = dp;
+	}
+	if(min>max)
+	{
+		b3Scalar tmp = min;
+		min = max;
+		max = tmp;
+	}
+	min += offset;
+	max += offset;
+}
+
+
+static bool TestSepAxis(const b3ConvexPolyhedronData& hullA, const b3ConvexPolyhedronData& hullB, 
+	const float4& posA,const b3Quaternion& ornA,
+	const float4& posB,const b3Quaternion& ornB,
+	const float4& sep_axis, const b3AlignedObjectArray<b3Vector3>& verticesA,const b3AlignedObjectArray<b3Vector3>& verticesB,b3Scalar& depth)
+{
+	b3Scalar Min0,Max0;
+	b3Scalar Min1,Max1;
+	project(hullA,posA,ornA,sep_axis,verticesA, Min0, Max0);
+	project(hullB,posB,ornB, sep_axis,verticesB, Min1, Max1);
+
+	if(Max0<Min1 || Max1<Min0)
+		return false;
+
+	b3Scalar d0 = Max0 - Min1;
+	assert(d0>=0.0f);
+	b3Scalar d1 = Max1 - Min0;
+	assert(d1>=0.0f);
+	depth = d0<d1 ? d0:d1;
+	return true;
+}
+
+inline bool IsAlmostZero(const b3Vector3& v)
+{
+	if(fabsf(v.x)>1e-6 || fabsf(v.y)>1e-6 || fabsf(v.z)>1e-6)	return false;
+	return true;
+}
+
+
+static bool findSeparatingAxis(	const b3ConvexPolyhedronData& hullA, const b3ConvexPolyhedronData& hullB, 
+	const float4& posA1,
+	const b3Quaternion& ornA,
+	const float4& posB1,
+	const b3Quaternion& ornB,
+	const b3AlignedObjectArray<b3Vector3>& verticesA,
+	const b3AlignedObjectArray<b3Vector3>& uniqueEdgesA, 
+	const b3AlignedObjectArray<b3GpuFace>& facesA,
+	const b3AlignedObjectArray<int>& indicesA,
+	const b3AlignedObjectArray<b3Vector3>& verticesB, 
+	const b3AlignedObjectArray<b3Vector3>& uniqueEdgesB, 
+	const b3AlignedObjectArray<b3GpuFace>& facesB,
+	const b3AlignedObjectArray<int>& indicesB,
+
+	b3Vector3& sep)
+{
+	B3_PROFILE("findSeparatingAxis");
+
+	b3g_actualSATPairTests++;
+	float4 posA = posA1;
+	posA.w = 0.f;
+	float4 posB = posB1;
+	posB.w = 0.f;
+//#ifdef TEST_INTERNAL_OBJECTS
+	float4 c0local = (float4&)hullA.m_localCenter;
+	float4 c0 = transform(&c0local, &posA, &ornA);
+	float4 c1local = (float4&)hullB.m_localCenter;
+	float4 c1 = transform(&c1local,&posB,&ornB);
+	const float4 deltaC2 = c0 - c1;
+//#endif
+
+	b3Scalar dmin = FLT_MAX;
+	int curPlaneTests=0;
+
+	int numFacesA = hullA.m_numFaces;
+	// Test normals from hullA
+	for(int i=0;i<numFacesA;i++)
+	{
+		const float4& normal = (float4&)facesA[hullA.m_faceOffset+i].m_plane;
+		float4 faceANormalWS = b3QuatRotate(ornA,normal);
+
+		if (dot3F4(deltaC2,faceANormalWS)<0)
+			faceANormalWS*=-1.f;
+
+		curPlaneTests++;
+#ifdef TEST_INTERNAL_OBJECTS
+		gExpectedNbTests++;
+		if(gUseInternalObject && !TestInternalObjects(transA,transB, DeltaC2, faceANormalWS, hullA, hullB, dmin))
+			continue;
+		gActualNbTests++;
+#endif
+
+		
+		b3Scalar d;
+		if(!TestSepAxis( hullA, hullB, posA,ornA,posB,ornB,faceANormalWS, verticesA, verticesB,d))
+			return false;
+
+		if(d<dmin)
+		{
+			dmin = d;
+			sep = (b3Vector3&)faceANormalWS;
+		}
+	}
+
+	int numFacesB = hullB.m_numFaces;
+	// Test normals from hullB
+	for(int i=0;i<numFacesB;i++)
+	{
+		float4 normal = (float4&)facesB[hullB.m_faceOffset+i].m_plane;
+		float4 WorldNormal = b3QuatRotate(ornB, normal);
+
+		if (dot3F4(deltaC2,WorldNormal)<0)
+		{
+			WorldNormal*=-1.f;
+		}
+		curPlaneTests++;
+#ifdef TEST_INTERNAL_OBJECTS
+		gExpectedNbTests++;
+		if(gUseInternalObject && !TestInternalObjects(transA,transB,DeltaC2, WorldNormal, hullA, hullB, dmin))
+			continue;
+		gActualNbTests++;
+#endif
+
+		b3Scalar d;
+		if(!TestSepAxis(hullA, hullB,posA,ornA,posB,ornB,WorldNormal,verticesA,verticesB,d))
+			return false;
+
+		if(d<dmin)
+		{
+			dmin = d;
+			sep = (b3Vector3&)WorldNormal;
+		}
+	}
+
+	int curEdgeEdge = 0;
+	// Test edges
+	for(int e0=0;e0<hullA.m_numUniqueEdges;e0++)
+	{
+		const float4& edge0 = (float4&) uniqueEdgesA[hullA.m_uniqueEdgesOffset+e0];
+		float4 edge0World = b3QuatRotate(ornA,(float4&)edge0);
+
+		for(int e1=0;e1<hullB.m_numUniqueEdges;e1++)
+		{
+			const b3Vector3 edge1 = uniqueEdgesB[hullB.m_uniqueEdgesOffset+e1];
+			float4 edge1World = b3QuatRotate(ornB,(float4&)edge1);
+
+
+			float4 crossje = cross3(edge0World,edge1World);
+
+			curEdgeEdge++;
+			if(!IsAlmostZero((b3Vector3&)crossje))
+			{
+				crossje = normalize3(crossje);
+				if (dot3F4(deltaC2,crossje)<0)
+					crossje*=-1.f;
+
+
+#ifdef TEST_INTERNAL_OBJECTS
+				gExpectedNbTests++;
+				if(gUseInternalObject && !TestInternalObjects(transA,transB,DeltaC2, Cross, hullA, hullB, dmin))
+					continue;
+				gActualNbTests++;
+#endif
+
+				b3Scalar dist;
+				if(!TestSepAxis( hullA, hullB, posA,ornA,posB,ornB,crossje, verticesA,verticesB,dist))
+					return false;
+
+				if(dist<dmin)
+				{
+					dmin = dist;
+					sep = (b3Vector3&)crossje;
+				}
+			}
+		}
+
+	}
+
+	
+	if((dot3F4(-deltaC2,(float4&)sep))>0.0f)
+		sep = -sep;
+
+	return true;
+}
+
+
+bool findSeparatingAxisEdgeEdge(	__global const b3ConvexPolyhedronData* hullA, __global const b3ConvexPolyhedronData* hullB, 
+	const b3Float4& posA1,
+	const b3Quat& ornA,
+	const b3Float4& posB1,
+	const b3Quat& ornB,
+	const b3Float4& DeltaC2,
+	__global const b3AlignedObjectArray<float4>& vertices, 
+	__global const b3AlignedObjectArray<float4>& uniqueEdges, 
+	__global const b3AlignedObjectArray<b3GpuFace>& faces,
+	__global const b3AlignedObjectArray<int>&  indices,
+	float4* sep,
+	float* dmin)
+{
+//	int i = get_global_id(0);
+
+	float4 posA = posA1;
+	posA.w = 0.f;
+	float4 posB = posB1;
+	posB.w = 0.f;
+
+	//int curPlaneTests=0;
+
+	int curEdgeEdge = 0;
+	// Test edges
+	for(int e0=0;e0<hullA->m_numUniqueEdges;e0++)
+	{
+		const float4 edge0 = uniqueEdges[hullA->m_uniqueEdgesOffset+e0];
+		float4 edge0World = b3QuatRotate(ornA,edge0);
+
+		for(int e1=0;e1<hullB->m_numUniqueEdges;e1++)
+		{
+			const float4 edge1 = uniqueEdges[hullB->m_uniqueEdgesOffset+e1];
+			float4 edge1World = b3QuatRotate(ornB,edge1);
+
+
+			float4 crossje = cross3(edge0World,edge1World);
+
+			curEdgeEdge++;
+			if(!IsAlmostZero(crossje))
+			{
+				crossje = normalize3(crossje);
+				if (dot3F4(DeltaC2,crossje)<0)
+					crossje*=-1.f;
+					
+				float dist;
+				bool result = true;
+				{
+					float Min0,Max0;
+					float Min1,Max1;
+					project(*hullA,posA,ornA,crossje,vertices, Min0, Max0);
+					project(*hullB,posB,ornB,crossje,vertices, Min1, Max1);
+				
+					if(Max0<Min1 || Max1<Min0)
+						result = false;
+				
+					float d0 = Max0 - Min1;
+					float d1 = Max1 - Min0;
+					dist = d0<d1 ? d0:d1;
+					result = true;
+
+				}
+				
+
+				if(dist<*dmin)
+				{
+					*dmin = dist;
+					*sep = crossje;
+				}
+			}
+		}
+
+	}
+
+	
+	if((dot3F4(-DeltaC2,*sep))>0.0f)
+	{
+		*sep = -(*sep);
+	}
+	return true;
+}
+
+
+__inline float4 lerp3(const float4& a,const float4& b, float  t)
+{
+	return b3MakeVector3(	a.x + (b.x - a.x) * t,
+						a.y + (b.y - a.y) * t,
+						a.z + (b.z - a.z) * t,
+						0.f);
+}
+
+
+// Clips a face to the back of a plane, return the number of vertices out, stored in ppVtxOut
+int clipFace(const float4* pVtxIn, int numVertsIn, float4& planeNormalWS,float planeEqWS, float4* ppVtxOut)
+{
+	
+	int ve;
+	float ds, de;
+	int numVertsOut = 0;
+	if (numVertsIn < 2)
+		return 0;
+
+	float4 firstVertex=pVtxIn[numVertsIn-1];
+	float4 endVertex = pVtxIn[0];
+	
+	ds = dot3F4(planeNormalWS,firstVertex)+planeEqWS;
+
+	for (ve = 0; ve < numVertsIn; ve++)
+	{
+		endVertex=pVtxIn[ve];
+
+		de = dot3F4(planeNormalWS,endVertex)+planeEqWS;
+
+		if (ds<0)
+		{
+			if (de<0)
+			{
+				// Start < 0, end < 0, so output endVertex
+				ppVtxOut[numVertsOut++] = endVertex;
+			}
+			else
+			{
+				// Start < 0, end >= 0, so output intersection
+				ppVtxOut[numVertsOut++] = lerp3(firstVertex, endVertex,(ds * 1.f/(ds - de)) );
+			}
+		}
+		else
+		{
+			if (de<0)
+			{
+				// Start >= 0, end < 0 so output intersection and end
+				ppVtxOut[numVertsOut++] = lerp3(firstVertex, endVertex,(ds * 1.f/(ds - de)) );
+				ppVtxOut[numVertsOut++] = endVertex;
+			}
+		}
+		firstVertex = endVertex;
+		ds = de;
+	}
+	return numVertsOut;
+}
+
+
+int clipFaceAgainstHull(const float4& separatingNormal, const b3ConvexPolyhedronData* hullA,  
+	const float4& posA, const b3Quaternion& ornA, float4* worldVertsB1, int numWorldVertsB1,
+	float4* worldVertsB2, int capacityWorldVertsB2,
+	const float minDist, float maxDist,
+	const b3AlignedObjectArray<float4>& verticesA,	const b3AlignedObjectArray<b3GpuFace>& facesA,	const b3AlignedObjectArray<int>& indicesA,
+	//const float4* verticesB,	const b3GpuFace* facesB,	const int* indicesB,
+	float4* contactsOut,
+	int contactCapacity)
+{
+	int numContactsOut = 0;
+
+	float4* pVtxIn = worldVertsB1;
+	float4* pVtxOut = worldVertsB2;
+	
+	int numVertsIn = numWorldVertsB1;
+	int numVertsOut = 0;
+
+	int closestFaceA=-1;
+	{
+		float dmin = FLT_MAX;
+		for(int face=0;face<hullA->m_numFaces;face++)
+		{
+			const float4 Normal = b3MakeVector3(
+				facesA[hullA->m_faceOffset+face].m_plane.x, 
+				facesA[hullA->m_faceOffset+face].m_plane.y, 
+				facesA[hullA->m_faceOffset+face].m_plane.z,0.f);
+			const float4 faceANormalWS = b3QuatRotate(ornA,Normal);
+		
+			float d = dot3F4(faceANormalWS,separatingNormal);
+			if (d < dmin)
+			{
+				dmin = d;
+				closestFaceA = face;
+			}
+		}
+	}
+	if (closestFaceA<0)
+		return numContactsOut;
+
+	b3GpuFace polyA = facesA[hullA->m_faceOffset+closestFaceA];
+
+	// clip polygon to back of planes of all faces of hull A that are adjacent to witness face
+//	int numContacts = numWorldVertsB1;
+	int numVerticesA = polyA.m_numIndices;
+	for(int e0=0;e0<numVerticesA;e0++)
+	{
+		const float4 a = verticesA[hullA->m_vertexOffset+indicesA[polyA.m_indexOffset+e0]];
+		const float4 b = verticesA[hullA->m_vertexOffset+indicesA[polyA.m_indexOffset+((e0+1)%numVerticesA)]];
+		const float4 edge0 = a - b;
+		const float4 WorldEdge0 = b3QuatRotate(ornA,edge0);
+		float4 planeNormalA = make_float4(polyA.m_plane.x,polyA.m_plane.y,polyA.m_plane.z,0.f);
+		float4 worldPlaneAnormal1 = b3QuatRotate(ornA,planeNormalA);
+
+		float4 planeNormalWS1 = -cross3(WorldEdge0,worldPlaneAnormal1);
+		float4 worldA1 = transform(&a,&posA,&ornA);
+		float planeEqWS1 = -dot3F4(worldA1,planeNormalWS1);
+		
+		float4 planeNormalWS = planeNormalWS1;
+		float planeEqWS=planeEqWS1;
+		
+		//clip face
+		//clipFace(*pVtxIn, *pVtxOut,planeNormalWS,planeEqWS);
+		numVertsOut = clipFace(pVtxIn, numVertsIn, planeNormalWS,planeEqWS, pVtxOut);
+
+		//btSwap(pVtxIn,pVtxOut);
+		float4* tmp = pVtxOut;
+		pVtxOut = pVtxIn;
+		pVtxIn = tmp;
+		numVertsIn = numVertsOut;
+		numVertsOut = 0;
+	}
+
+	
+	// only keep points that are behind the witness face
+	{
+		float4 localPlaneNormal  = make_float4(polyA.m_plane.x,polyA.m_plane.y,polyA.m_plane.z,0.f);
+		float localPlaneEq = polyA.m_plane.w;
+		float4 planeNormalWS = b3QuatRotate(ornA,localPlaneNormal);
+		float planeEqWS=localPlaneEq-dot3F4(planeNormalWS,posA);
+		for (int i=0;i<numVertsIn;i++)
+		{
+			float depth = dot3F4(planeNormalWS,pVtxIn[i])+planeEqWS;
+			if (depth <=minDist)
+			{
+				depth = minDist;
+			}
+			if (numContactsOut<contactCapacity)
+			{
+				if (depth <=maxDist)
+				{
+					float4 pointInWorld = pVtxIn[i];
+					//resultOut.addContactPoint(separatingNormal,point,depth);
+					contactsOut[numContactsOut++] = b3MakeVector3(pointInWorld.x,pointInWorld.y,pointInWorld.z,depth);
+					//printf("depth=%f\n",depth);
+				}
+			} else
+			{
+				b3Error("exceeding contact capacity (%d,%df)\n", numContactsOut,contactCapacity);
+			}
+		}
+	}
+
+	return numContactsOut;
+}
+
+
+
+static int	clipHullAgainstHull(const float4& separatingNormal, 
+	const b3ConvexPolyhedronData& hullA, const b3ConvexPolyhedronData& hullB, 
+	const float4& posA, const b3Quaternion& ornA,const float4& posB, const b3Quaternion& ornB, 
+	float4* worldVertsB1, float4* worldVertsB2, int capacityWorldVerts,
+	const float minDist, float maxDist,
+	const b3AlignedObjectArray<float4>& verticesA,	const b3AlignedObjectArray<b3GpuFace>& facesA,	const b3AlignedObjectArray<int>& indicesA,
+	const b3AlignedObjectArray<float4>& verticesB,	const b3AlignedObjectArray<b3GpuFace>& facesB,	const b3AlignedObjectArray<int>& indicesB,
+
+	float4*	contactsOut,
+	int contactCapacity)
+{
+	int numContactsOut = 0;
+	int numWorldVertsB1= 0;
+	
+	B3_PROFILE("clipHullAgainstHull");
+
+//	float curMaxDist=maxDist;
+	int closestFaceB=-1;
+	float dmax = -FLT_MAX;
+
+	{
+		//B3_PROFILE("closestFaceB");
+		if (hullB.m_numFaces!=1)
+		{
+			//printf("wtf\n");
+		}
+		static bool once = true;
+		//printf("separatingNormal=%f,%f,%f\n",separatingNormal.x,separatingNormal.y,separatingNormal.z);
+		
+		for(int face=0;face<hullB.m_numFaces;face++)
+		{
+#ifdef BT_DEBUG_SAT_FACE
+			if (once)
+				printf("face %d\n",face);
+			const b3GpuFace* faceB = &facesB[hullB.m_faceOffset+face];
+			if (once)
+			{
+				for (int i=0;i<faceB->m_numIndices;i++)
+				{
+					float4 vert = verticesB[hullB.m_vertexOffset+indicesB[faceB->m_indexOffset+i]];
+					printf("vert[%d] = %f,%f,%f\n",i,vert.x,vert.y,vert.z);
+				}
+			}
+#endif //BT_DEBUG_SAT_FACE
+			//if (facesB[hullB.m_faceOffset+face].m_numIndices>2)
+			{
+				const float4 Normal = b3MakeVector3(facesB[hullB.m_faceOffset+face].m_plane.x, 
+					facesB[hullB.m_faceOffset+face].m_plane.y, facesB[hullB.m_faceOffset+face].m_plane.z,0.f);
+				const float4 WorldNormal = b3QuatRotate(ornB, Normal);
+#ifdef BT_DEBUG_SAT_FACE
+				if (once)
+					printf("faceNormal = %f,%f,%f\n",Normal.x,Normal.y,Normal.z);
+#endif
+				float d = dot3F4(WorldNormal,separatingNormal);
+				if (d > dmax)
+				{
+					dmax = d;
+					closestFaceB = face;
+				}
+			}
+		}
+		once = false;
+	}
+
+	
+	b3Assert(closestFaceB>=0);
+	{
+		//B3_PROFILE("worldVertsB1");
+		const b3GpuFace& polyB = facesB[hullB.m_faceOffset+closestFaceB];
+		const int numVertices = polyB.m_numIndices;
+		for(int e0=0;e0<numVertices;e0++)
+		{
+			const float4& b = verticesB[hullB.m_vertexOffset+indicesB[polyB.m_indexOffset+e0]];
+			worldVertsB1[numWorldVertsB1++] = transform(&b,&posB,&ornB);
+		}
+	}
+
+	if (closestFaceB>=0)
+	{
+		//B3_PROFILE("clipFaceAgainstHull");
+		numContactsOut = clipFaceAgainstHull((float4&)separatingNormal, &hullA, 
+				posA,ornA,
+				worldVertsB1,numWorldVertsB1,worldVertsB2,capacityWorldVerts, minDist, maxDist,
+				verticesA,				facesA,				indicesA,
+				contactsOut,contactCapacity);
+	}
+
+	return numContactsOut;
+}
+
+
+
+
+
+
+#define PARALLEL_SUM(v, n) for(int j=1; j<n; j++) v[0] += v[j];
+#define PARALLEL_DO(execution, n) for(int ie=0; ie<n; ie++){execution;}
+#define REDUCE_MAX(v, n) {int i=0;\
+for(int offset=0; offset<n; offset++) v[i] = (v[i].y > v[i+offset].y)? v[i]: v[i+offset]; }
+#define REDUCE_MIN(v, n) {int i=0;\
+for(int offset=0; offset<n; offset++) v[i] = (v[i].y < v[i+offset].y)? v[i]: v[i+offset]; }
+
+int extractManifold(const float4* p, int nPoints, const float4& nearNormal, b3Int4* contactIdx)
+{
+	if( nPoints == 0 )
+        return 0;
+    
+    if (nPoints <=4)
+        return nPoints;
+    
+    
+    if (nPoints >64)
+        nPoints = 64;
+    
+	float4 center = make_float4(0,0,0,0);
+	{
+		
+		for (int i=0;i<nPoints;i++)
+			center += p[i];
+		center /= (float)nPoints;
+	}
+    
+	
+    
+	//	sample 4 directions
+    
+    float4 aVector = p[0] - center;
+    float4 u = cross3( nearNormal, aVector );
+    float4 v = cross3( nearNormal, u );
+    u = normalize3( u );
+    v = normalize3( v );
+    
+    
+    //keep point with deepest penetration
+    float minW= FLT_MAX;
+    
+    int minIndex=-1;
+    
+    float4 maxDots;
+    maxDots.x = FLT_MIN;
+    maxDots.y = FLT_MIN;
+    maxDots.z = FLT_MIN;
+    maxDots.w = FLT_MIN;
+    
+    //	idx, distance
+    for(int ie = 0; ie<nPoints; ie++ )
+    {
+        if (p[ie].w<minW)
+        {
+            minW = p[ie].w;
+            minIndex=ie;
+        }
+        float f;
+        float4 r = p[ie]-center;
+        f = dot3F4( u, r );
+        if (f<maxDots.x)
+        {
+            maxDots.x = f;
+            contactIdx[0].x = ie;
+        }
+        
+        f = dot3F4( -u, r );
+        if (f<maxDots.y)
+        {
+            maxDots.y = f;
+            contactIdx[0].y = ie;
+        }
+        
+        
+        f = dot3F4( v, r );
+        if (f<maxDots.z)
+        {
+            maxDots.z = f;
+            contactIdx[0].z = ie;
+        }
+        
+        f = dot3F4( -v, r );
+        if (f<maxDots.w)
+        {
+            maxDots.w = f;
+            contactIdx[0].w = ie;
+        }
+        
+    }
+    
+    if (contactIdx[0].x != minIndex && contactIdx[0].y != minIndex && contactIdx[0].z != minIndex && contactIdx[0].w != minIndex)
+    {
+        //replace the first contact with minimum (todo: replace contact with least penetration)
+        contactIdx[0].x = minIndex;
+    }
+    
+    return 4;
+    
+}
+
+
+
+
+int clipHullHullSingle(
+			int bodyIndexA, int bodyIndexB,
+										 const float4& posA,
+										 const b3Quaternion& ornA,
+										 const float4& posB,
+										 const b3Quaternion& ornB,
+
+			int collidableIndexA, int collidableIndexB,
+
+			const b3AlignedObjectArray<b3RigidBodyData>* bodyBuf, 
+			b3AlignedObjectArray<b3Contact4>* globalContactOut, 
+			int& nContacts,
+			
+			const b3AlignedObjectArray<b3ConvexPolyhedronData>& hostConvexDataA,
+			const b3AlignedObjectArray<b3ConvexPolyhedronData>& hostConvexDataB,
+	
+			const b3AlignedObjectArray<b3Vector3>& verticesA, 
+			const b3AlignedObjectArray<b3Vector3>& uniqueEdgesA, 
+			const b3AlignedObjectArray<b3GpuFace>& facesA,
+			const b3AlignedObjectArray<int>& indicesA,
+	
+			const b3AlignedObjectArray<b3Vector3>& verticesB,
+			const b3AlignedObjectArray<b3Vector3>& uniqueEdgesB,
+			const b3AlignedObjectArray<b3GpuFace>& facesB,
+			const b3AlignedObjectArray<int>& indicesB,
+
+			const b3AlignedObjectArray<b3Collidable>& hostCollidablesA,
+			const b3AlignedObjectArray<b3Collidable>& hostCollidablesB,
+			const b3Vector3& sepNormalWorldSpace,
+			int maxContactCapacity			)
+{
+	int contactIndex = -1;
+	b3ConvexPolyhedronData hullA, hullB;
+    
+    b3Collidable colA = hostCollidablesA[collidableIndexA];
+    hullA = hostConvexDataA[colA.m_shapeIndex];
+    //printf("numvertsA = %d\n",hullA.m_numVertices);
+    
+    
+    b3Collidable colB = hostCollidablesB[collidableIndexB];
+    hullB = hostConvexDataB[colB.m_shapeIndex];
+    //printf("numvertsB = %d\n",hullB.m_numVertices);
+    
+	
+	float4 contactsOut[MAX_VERTS];
+	int localContactCapacity = MAX_VERTS;
+
+#ifdef _WIN32
+	b3Assert(_finite(bodyBuf->at(bodyIndexA).m_pos.x));
+	b3Assert(_finite(bodyBuf->at(bodyIndexB).m_pos.x));
+#endif
+	
+	
+	{
+		
+		float4 worldVertsB1[MAX_VERTS];
+		float4 worldVertsB2[MAX_VERTS];
+		int capacityWorldVerts = MAX_VERTS;
+
+		float4 hostNormal = make_float4(sepNormalWorldSpace.x,sepNormalWorldSpace.y,sepNormalWorldSpace.z,0.f);
+		int shapeA = hostCollidablesA[collidableIndexA].m_shapeIndex;
+		int shapeB = hostCollidablesB[collidableIndexB].m_shapeIndex;
+
+		b3Scalar minDist = -1;
+		b3Scalar maxDist = 0.;
+
+		        
+
+		b3Transform trA,trB;
+		{
+		//B3_PROFILE("transform computation");
+		//trA.setIdentity();
+		trA.setOrigin(b3MakeVector3(posA.x,posA.y,posA.z));
+		trA.setRotation(b3Quaternion(ornA.x,ornA.y,ornA.z,ornA.w));
+				
+		//trB.setIdentity();
+		trB.setOrigin(b3MakeVector3(posB.x,posB.y,posB.z));
+		trB.setRotation(b3Quaternion(ornB.x,ornB.y,ornB.z,ornB.w));
+		}
+
+		b3Quaternion trAorn = trA.getRotation();
+        b3Quaternion trBorn = trB.getRotation();
+        
+		int numContactsOut = clipHullAgainstHull(hostNormal, 
+						hostConvexDataA.at(shapeA), 
+						hostConvexDataB.at(shapeB),
+								(float4&)trA.getOrigin(), (b3Quaternion&)trAorn,
+								(float4&)trB.getOrigin(), (b3Quaternion&)trBorn,
+								worldVertsB1,worldVertsB2,capacityWorldVerts,
+								minDist, maxDist,
+								verticesA,	facesA,indicesA,
+								verticesB,	facesB,indicesB,
+								
+								contactsOut,localContactCapacity);
+
+		if (numContactsOut>0)
+		{
+			B3_PROFILE("overlap");
+
+			float4 normalOnSurfaceB = (float4&)hostNormal;
+			
+			b3Int4 contactIdx;
+			contactIdx.x = 0;
+			contactIdx.y = 1;
+			contactIdx.z = 2;
+			contactIdx.w = 3;
+			
+			int numPoints = 0;
+					
+			{
+			//	B3_PROFILE("extractManifold");
+				numPoints = extractManifold(contactsOut, numContactsOut, normalOnSurfaceB,  &contactIdx);
+			}
+					
+			b3Assert(numPoints);
+					
+			if (nContacts<maxContactCapacity)
+			{
+				contactIndex = nContacts;
+				globalContactOut->expand();
+				b3Contact4& contact = globalContactOut->at(nContacts);
+				contact.m_batchIdx = 0;//i;
+				contact.m_bodyAPtrAndSignBit = (bodyBuf->at(bodyIndexA).m_invMass==0)? -bodyIndexA:bodyIndexA;
+				contact.m_bodyBPtrAndSignBit = (bodyBuf->at(bodyIndexB).m_invMass==0)? -bodyIndexB:bodyIndexB;
+
+				contact.m_frictionCoeffCmp = 45874;
+				contact.m_restituitionCoeffCmp = 0;
+					
+	//			float distance = 0.f;
+				for (int p=0;p<numPoints;p++)
+				{
+					contact.m_worldPosB[p] = contactsOut[contactIdx.s[p]];//check if it is actually on B
+					contact.m_worldNormalOnB = normalOnSurfaceB; 
+				}
+				//printf("bodyIndexA %d,bodyIndexB %d,normal=%f,%f,%f numPoints %d\n",bodyIndexA,bodyIndexB,normalOnSurfaceB.x,normalOnSurfaceB.y,normalOnSurfaceB.z,numPoints);
+				contact.m_worldNormalOnB.w = (b3Scalar)numPoints;
+				nContacts++;
+			} else
+			{
+				b3Error("Error: exceeding contact capacity (%d/%d)\n", nContacts,maxContactCapacity);
+			}
+		}
+	}
+	return contactIndex;
+}
+
+	
+
+
+
+void computeContactPlaneConvex(int pairIndex,
+																int bodyIndexA, int bodyIndexB, 
+																int collidableIndexA, int collidableIndexB, 
+																const b3RigidBodyData* rigidBodies, 
+																const b3Collidable* collidables,
+																const b3ConvexPolyhedronData* convexShapes,
+																const b3Vector3* convexVertices,
+																const int* convexIndices,
+																const b3GpuFace* faces,
+																b3Contact4* globalContactsOut,
+																int& nGlobalContactsOut,
+																int maxContactCapacity)
+{
+
+		int shapeIndex = collidables[collidableIndexB].m_shapeIndex;
+	const b3ConvexPolyhedronData* hullB = &convexShapes[shapeIndex];
+	
+	b3Vector3 posB = rigidBodies[bodyIndexB].m_pos;
+	b3Quaternion ornB = rigidBodies[bodyIndexB].m_quat;
+	b3Vector3 posA = rigidBodies[bodyIndexA].m_pos;
+	b3Quaternion ornA = rigidBodies[bodyIndexA].m_quat;
+
+//	int numContactsOut = 0;
+//	int numWorldVertsB1= 0;
+
+	b3Vector3 planeEq = faces[collidables[collidableIndexA].m_shapeIndex].m_plane;
+	b3Vector3 planeNormal=b3MakeVector3(planeEq.x,planeEq.y,planeEq.z);
+	b3Vector3 planeNormalWorld = b3QuatRotate(ornA,planeNormal);
+	float planeConstant = planeEq.w;
+	b3Transform convexWorldTransform;
+	convexWorldTransform.setIdentity();
+	convexWorldTransform.setOrigin(posB);
+	convexWorldTransform.setRotation(ornB);
+	b3Transform planeTransform;
+	planeTransform.setIdentity();
+	planeTransform.setOrigin(posA);
+	planeTransform.setRotation(ornA);
+
+	b3Transform planeInConvex;
+	planeInConvex= convexWorldTransform.inverse() * planeTransform;
+	b3Transform convexInPlane;
+	convexInPlane = planeTransform.inverse() * convexWorldTransform;
+	
+	b3Vector3 planeNormalInConvex = planeInConvex.getBasis()*-planeNormal;
+	float maxDot = -1e30;
+	int hitVertex=-1;
+	b3Vector3 hitVtx;
+
+#define MAX_PLANE_CONVEX_POINTS 64
+
+	b3Vector3 contactPoints[MAX_PLANE_CONVEX_POINTS];
+	int numPoints = 0;
+
+	b3Int4 contactIdx;
+	contactIdx.s[0] = 0;
+	contactIdx.s[1] = 1;
+	contactIdx.s[2] = 2;
+	contactIdx.s[3] = 3;
+	
+	for (int i=0;i<hullB->m_numVertices;i++)
+	{
+		b3Vector3 vtx = convexVertices[hullB->m_vertexOffset+i];
+		float curDot = vtx.dot(planeNormalInConvex);
+
+
+		if (curDot>maxDot)
+		{
+			hitVertex=i;
+			maxDot=curDot;
+			hitVtx = vtx;
+			//make sure the deepest points is always included
+			if (numPoints==MAX_PLANE_CONVEX_POINTS)
+				numPoints--;
+		}
+
+		if (numPoints<MAX_PLANE_CONVEX_POINTS)
+		{
+			b3Vector3 vtxWorld = convexWorldTransform*vtx;
+			b3Vector3 vtxInPlane = planeTransform.inverse()*vtxWorld;
+			float dist = planeNormal.dot(vtxInPlane)-planeConstant;
+			if (dist<0.f)
+			{
+				vtxWorld.w = dist;
+				contactPoints[numPoints] = vtxWorld;
+				numPoints++;
+			}
+		}
+
+	}
+
+	int numReducedPoints  = 0;
+
+	numReducedPoints = numPoints;
+	
+	if (numPoints>4)
+	{
+		numReducedPoints = extractManifoldSequentialGlobal( contactPoints, numPoints, planeNormalInConvex, &contactIdx);
+	}
+	int dstIdx;
+//    dstIdx = nGlobalContactsOut++;//AppendInc( nGlobalContactsOut, dstIdx );
+		
+	if (numReducedPoints>0)
+	{
+		if (nGlobalContactsOut < maxContactCapacity)
+		{
+			dstIdx=nGlobalContactsOut;
+			nGlobalContactsOut++;
+
+			b3Contact4* c = &globalContactsOut[dstIdx];
+			c->m_worldNormalOnB = -planeNormalWorld;
+			c->setFrictionCoeff(0.7);
+			c->setRestituitionCoeff(0.f);
+
+			c->m_batchIdx = pairIndex;
+			c->m_bodyAPtrAndSignBit = rigidBodies[bodyIndexA].m_invMass==0?-bodyIndexA:bodyIndexA;
+			c->m_bodyBPtrAndSignBit = rigidBodies[bodyIndexB].m_invMass==0?-bodyIndexB:bodyIndexB;
+			for (int i=0;i<numReducedPoints;i++)
+			{
+				b3Vector3 pOnB1 = contactPoints[contactIdx.s[i]];
+				c->m_worldPosB[i] = pOnB1;
+			}
+			c->m_worldNormalOnB.w = (b3Scalar)numReducedPoints;
+		}//if (dstIdx < numPairs)
+	}	
+		
+
+
+//	printf("computeContactPlaneConvex\n");
+}
+
+
+
+B3_FORCE_INLINE b3Vector3	MyUnQuantize(const unsigned short* vecIn, const b3Vector3& quantization, const b3Vector3& bvhAabbMin)
+	{
+			b3Vector3	vecOut;
+			vecOut.setValue(
+			(b3Scalar)(vecIn[0]) / (quantization.x),
+			(b3Scalar)(vecIn[1]) / (quantization.y),
+			(b3Scalar)(vecIn[2]) / (quantization.z));
+			vecOut += bvhAabbMin;
+			return vecOut;
+	}
+
+void traverseTreeTree()
+{
+
+}
+
+#include "Bullet3Common/shared/b3Mat3x3.h"
+
+int numAabbChecks = 0;
+int maxNumAabbChecks = 0;
+int maxDepth = 0;
+
+// work-in-progress
+__kernel void   findCompoundPairsKernel( 
+	int pairIndex,
+	int bodyIndexA,
+	int bodyIndexB,
+	int collidableIndexA,
+	int collidableIndexB,
+	__global const b3RigidBodyData* rigidBodies, 
+	__global const b3Collidable* collidables,
+	__global const b3ConvexPolyhedronData* convexShapes, 
+	__global const b3AlignedObjectArray<b3Float4>& vertices,
+	__global const b3AlignedObjectArray<b3Aabb>& aabbsWorldSpace,
+	__global const b3AlignedObjectArray<b3Aabb>& aabbsLocalSpace,
+	__global const b3GpuChildShape* gpuChildShapes,
+	__global b3Int4* gpuCompoundPairsOut,
+	__global  int* numCompoundPairsOut,
+	int maxNumCompoundPairsCapacity,
+	b3AlignedObjectArray<b3QuantizedBvhNode>&	treeNodesCPU,
+	b3AlignedObjectArray<b3BvhSubtreeInfo>&	subTreesCPU,
+	b3AlignedObjectArray<b3BvhInfo>&	bvhInfoCPU
+	)
+{
+	numAabbChecks=0;
+	maxNumAabbChecks=0;
+//	int i = pairIndex;
+	{
+		
+
+		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+
+
+		//once the broadphase avoids static-static pairs, we can remove this test
+		if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))
+		{
+			return;
+		}
+
+		if ((collidables[collidableIndexA].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS) &&(collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS))
+		{
+			int bvhA = collidables[collidableIndexA].m_compoundBvhIndex;
+			int bvhB = collidables[collidableIndexB].m_compoundBvhIndex;
+			int numSubTreesA = bvhInfoCPU[bvhA].m_numSubTrees;
+			int subTreesOffsetA = bvhInfoCPU[bvhA].m_subTreeOffset;
+			int subTreesOffsetB = bvhInfoCPU[bvhB].m_subTreeOffset;
+
+
+			int numSubTreesB = bvhInfoCPU[bvhB].m_numSubTrees;
+			
+			float4 posA = rigidBodies[bodyIndexA].m_pos;
+			b3Quat ornA = rigidBodies[bodyIndexA].m_quat;
+
+			b3Transform transA;
+			transA.setIdentity();
+			transA.setOrigin(posA);
+			transA.setRotation(ornA);
+
+			b3Quat ornB = rigidBodies[bodyIndexB].m_quat;
+			float4 posB = rigidBodies[bodyIndexB].m_pos;
+
+			b3Transform transB;
+			transB.setIdentity();
+			transB.setOrigin(posB);
+			transB.setRotation(ornB);
+
+
+
+			for (int p=0;p<numSubTreesA;p++)
+			{
+				b3BvhSubtreeInfo subtreeA = subTreesCPU[subTreesOffsetA+p];
+				//bvhInfoCPU[bvhA].m_quantization
+				b3Vector3 treeAminLocal = MyUnQuantize(subtreeA.m_quantizedAabbMin,bvhInfoCPU[bvhA].m_quantization,bvhInfoCPU[bvhA].m_aabbMin);
+				b3Vector3 treeAmaxLocal = MyUnQuantize(subtreeA.m_quantizedAabbMax,bvhInfoCPU[bvhA].m_quantization,bvhInfoCPU[bvhA].m_aabbMin);
+
+				b3Vector3 aabbAMinOut,aabbAMaxOut;
+				float margin=0.f;
+				b3TransformAabb2(treeAminLocal,treeAmaxLocal, margin,transA.getOrigin(),transA.getRotation(),&aabbAMinOut,&aabbAMaxOut);
+
+				for (int q=0;q<numSubTreesB;q++)
+				{
+					b3BvhSubtreeInfo subtreeB = subTreesCPU[subTreesOffsetB+q];
+
+					b3Vector3 treeBminLocal = MyUnQuantize(subtreeB.m_quantizedAabbMin,bvhInfoCPU[bvhB].m_quantization,bvhInfoCPU[bvhB].m_aabbMin);
+					b3Vector3 treeBmaxLocal = MyUnQuantize(subtreeB.m_quantizedAabbMax,bvhInfoCPU[bvhB].m_quantization,bvhInfoCPU[bvhB].m_aabbMin);
+
+					b3Vector3 aabbBMinOut,aabbBMaxOut;
+					float margin=0.f;
+					b3TransformAabb2(treeBminLocal,treeBmaxLocal, margin,transB.getOrigin(),transB.getRotation(),&aabbBMinOut,&aabbBMaxOut);
+
+					
+					numAabbChecks=0;
+					bool aabbOverlap = b3TestAabbAgainstAabb(aabbAMinOut,aabbAMaxOut,aabbBMinOut,aabbBMaxOut);
+					if (aabbOverlap)
+					{
+						
+						int startNodeIndexA = subtreeA.m_rootNodeIndex+bvhInfoCPU[bvhA].m_nodeOffset;
+		//				int endNodeIndexA = startNodeIndexA+subtreeA.m_subtreeSize;
+
+						int startNodeIndexB = subtreeB.m_rootNodeIndex+bvhInfoCPU[bvhB].m_nodeOffset;
+		//				int endNodeIndexB = startNodeIndexB+subtreeB.m_subtreeSize;
+
+						b3AlignedObjectArray<b3Int2> nodeStack;
+						b3Int2 node0;
+						node0.x = startNodeIndexA;
+						node0.y = startNodeIndexB;
+
+						int maxStackDepth = 1024;
+						nodeStack.resize(maxStackDepth);
+						int depth=0;
+						nodeStack[depth++]=node0;
+
+						do
+						{
+							if (depth > maxDepth)
+							{
+								maxDepth=depth;
+								printf("maxDepth=%d\n",maxDepth);
+							}
+							b3Int2 node = nodeStack[--depth];
+							
+							b3Vector3 aMinLocal = MyUnQuantize(treeNodesCPU[node.x].m_quantizedAabbMin,bvhInfoCPU[bvhA].m_quantization,bvhInfoCPU[bvhA].m_aabbMin);
+							b3Vector3 aMaxLocal = MyUnQuantize(treeNodesCPU[node.x].m_quantizedAabbMax,bvhInfoCPU[bvhA].m_quantization,bvhInfoCPU[bvhA].m_aabbMin);
+
+							b3Vector3 bMinLocal = MyUnQuantize(treeNodesCPU[node.y].m_quantizedAabbMin,bvhInfoCPU[bvhB].m_quantization,bvhInfoCPU[bvhB].m_aabbMin);
+							b3Vector3 bMaxLocal = MyUnQuantize(treeNodesCPU[node.y].m_quantizedAabbMax,bvhInfoCPU[bvhB].m_quantization,bvhInfoCPU[bvhB].m_aabbMin);
+
+							float margin=0.f;
+							b3Vector3 aabbAMinOut,aabbAMaxOut;
+							b3TransformAabb2(aMinLocal,aMaxLocal, margin,transA.getOrigin(),transA.getRotation(),&aabbAMinOut,&aabbAMaxOut);
+
+							b3Vector3 aabbBMinOut,aabbBMaxOut;
+							b3TransformAabb2(bMinLocal,bMaxLocal, margin,transB.getOrigin(),transB.getRotation(),&aabbBMinOut,&aabbBMaxOut);
+
+							numAabbChecks++;
+							bool nodeOverlap = b3TestAabbAgainstAabb(aabbAMinOut,aabbAMaxOut,aabbBMinOut,aabbBMaxOut);
+							if (nodeOverlap)
+							{
+								bool isLeafA = treeNodesCPU[node.x].isLeafNode();
+								bool isLeafB = treeNodesCPU[node.y].isLeafNode();
+								bool isInternalA = !isLeafA;
+								bool isInternalB = !isLeafB;
+
+								//fail, even though it might hit two leaf nodes
+								if (depth+4>maxStackDepth && !(isLeafA && isLeafB))
+								{
+									b3Error("Error: traversal exceeded maxStackDepth\n");
+									continue;
+								}
+
+								if(isInternalA)
+								{
+									int nodeAleftChild = node.x+1;
+									bool isNodeALeftChildLeaf = treeNodesCPU[node.x+1].isLeafNode();
+									int nodeArightChild = isNodeALeftChildLeaf? node.x+2 : node.x+1 + treeNodesCPU[node.x+1].getEscapeIndex();
+
+									if(isInternalB)
+									{					
+										int nodeBleftChild = node.y+1;
+										bool isNodeBLeftChildLeaf = treeNodesCPU[node.y+1].isLeafNode();
+										int nodeBrightChild = isNodeBLeftChildLeaf? node.y+2 : node.y+1 + treeNodesCPU[node.y+1].getEscapeIndex();
+
+										nodeStack[depth++] = b3MakeInt2(nodeAleftChild, nodeBleftChild);
+										nodeStack[depth++] = b3MakeInt2(nodeArightChild, nodeBleftChild);
+										nodeStack[depth++] = b3MakeInt2(nodeAleftChild, nodeBrightChild);
+										nodeStack[depth++] = b3MakeInt2(nodeArightChild, nodeBrightChild);
+									}
+									else
+									{
+										nodeStack[depth++] = b3MakeInt2(nodeAleftChild,node.y);
+										nodeStack[depth++] = b3MakeInt2(nodeArightChild,node.y);
+									}
+								}
+								else
+								{
+									if(isInternalB)
+									{
+										int nodeBleftChild = node.y+1;
+										bool isNodeBLeftChildLeaf = treeNodesCPU[node.y+1].isLeafNode();
+										int nodeBrightChild = isNodeBLeftChildLeaf? node.y+2 : node.y+1 + treeNodesCPU[node.y+1].getEscapeIndex();
+										nodeStack[depth++] = b3MakeInt2(node.x,nodeBleftChild);
+										nodeStack[depth++] = b3MakeInt2(node.x,nodeBrightChild);
+									}
+									else
+									{
+										int compoundPairIdx = b3AtomicInc(numCompoundPairsOut);
+										if (compoundPairIdx<maxNumCompoundPairsCapacity)
+										{
+											int childShapeIndexA = treeNodesCPU[node.x].getTriangleIndex();
+											int childShapeIndexB = treeNodesCPU[node.y].getTriangleIndex();
+											gpuCompoundPairsOut[compoundPairIdx]  = b3MakeInt4(bodyIndexA,bodyIndexB,childShapeIndexA,childShapeIndexB);
+										}
+									}
+								}
+							}
+						} while (depth);
+						maxNumAabbChecks = b3Max(numAabbChecks,maxNumAabbChecks);
+					}
+				}
+			}
+			
+			return;
+		}
+
+		if ((collidables[collidableIndexA].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS) ||(collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS))
+		{
+
+			if (collidables[collidableIndexA].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS) 
+			{
+
+				int numChildrenA = collidables[collidableIndexA].m_numChildShapes;
+				for (int c=0;c<numChildrenA;c++)
+				{
+					int childShapeIndexA = collidables[collidableIndexA].m_shapeIndex+c;
+					int childColIndexA = gpuChildShapes[childShapeIndexA].m_shapeIndex;
+
+					float4 posA = rigidBodies[bodyIndexA].m_pos;
+					b3Quat ornA = rigidBodies[bodyIndexA].m_quat;
+					float4 childPosA = gpuChildShapes[childShapeIndexA].m_childPosition;
+					b3Quat childOrnA = gpuChildShapes[childShapeIndexA].m_childOrientation;
+					float4 newPosA = b3QuatRotate(ornA,childPosA)+posA;
+					b3Quat newOrnA = b3QuatMul(ornA,childOrnA);
+					
+
+					
+					b3Aabb aabbA = aabbsLocalSpace[childColIndexA];
+
+					
+					b3Transform transA;
+					transA.setIdentity();
+					transA.setOrigin(newPosA);
+					transA.setRotation(newOrnA);
+					b3Scalar margin=0.0f;
+
+					b3Vector3 aabbAMinOut,aabbAMaxOut;
+
+					b3TransformAabb2((const b3Float4&)aabbA.m_min,(const b3Float4&)aabbA.m_max, margin,transA.getOrigin(),transA.getRotation(),&aabbAMinOut,&aabbAMaxOut);
+
+					if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+					{
+						int numChildrenB = collidables[collidableIndexB].m_numChildShapes;
+						for (int b=0;b<numChildrenB;b++)
+						{
+							int childShapeIndexB = collidables[collidableIndexB].m_shapeIndex+b;
+							int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+							b3Quat ornB = rigidBodies[bodyIndexB].m_quat;
+							float4 posB = rigidBodies[bodyIndexB].m_pos;
+							float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+							b3Quat childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+							float4 newPosB = transform(&childPosB,&posB,&ornB);
+							b3Quat newOrnB = b3QuatMul(ornB,childOrnB);
+
+							
+
+							b3Aabb aabbB = aabbsLocalSpace[childColIndexB];
+
+							b3Transform transB;
+							transB.setIdentity();
+							transB.setOrigin(newPosB);
+							transB.setRotation(newOrnB);
+
+							b3Vector3 aabbBMinOut,aabbBMaxOut;
+							b3TransformAabb2((const b3Float4&)aabbB.m_min,(const b3Float4&)aabbB.m_max, margin,transB.getOrigin(),transB.getRotation(),&aabbBMinOut,&aabbBMaxOut);
+
+							numAabbChecks++;
+							bool aabbOverlap = b3TestAabbAgainstAabb(aabbAMinOut,aabbAMaxOut,aabbBMinOut,aabbBMaxOut);
+							if (aabbOverlap)
+							{
+								/*
+								int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+								float dmin = FLT_MAX;
+								float4 posA = newPosA;
+								posA.w = 0.f;
+								float4 posB = newPosB;
+								posB.w = 0.f;
+								float4 c0local = convexShapes[shapeIndexA].m_localCenter;
+								b3Quat ornA = newOrnA;
+								float4 c0 = transform(&c0local, &posA, &ornA);
+								float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+								b3Quat ornB =newOrnB;
+								float4 c1 = transform(&c1local,&posB,&ornB);
+								const float4 DeltaC2 = c0 - c1;
+								*/
+								{//
+									int compoundPairIdx = b3AtomicInc(numCompoundPairsOut);
+									if (compoundPairIdx<maxNumCompoundPairsCapacity)
+									{
+										gpuCompoundPairsOut[compoundPairIdx]  = b3MakeInt4(bodyIndexA,bodyIndexB,childShapeIndexA,childShapeIndexB);
+									}
+								}//
+							}//fi(1)
+						} //for (int b=0
+					}//if (collidables[collidableIndexB].
+					else//if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+					{
+						if (1)
+						{
+						//	int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+						//	float dmin = FLT_MAX;
+							float4 posA = newPosA;
+							posA.w = 0.f;
+							float4 posB = rigidBodies[bodyIndexB].m_pos;
+							posB.w = 0.f;
+							float4 c0local = convexShapes[shapeIndexA].m_localCenter;
+							b3Quat ornA = newOrnA;
+							float4 c0;
+							c0 = transform(&c0local, &posA, &ornA);
+							float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+							b3Quat ornB = rigidBodies[bodyIndexB].m_quat;
+							float4 c1;
+							c1 = transform(&c1local,&posB,&ornB);
+						//	const float4 DeltaC2 = c0 - c1;
+
+							{
+								int compoundPairIdx = b3AtomicInc(numCompoundPairsOut);
+								if (compoundPairIdx<maxNumCompoundPairsCapacity)
+								{
+									gpuCompoundPairsOut[compoundPairIdx] = b3MakeInt4(bodyIndexA,bodyIndexB,childShapeIndexA,-1);
+								}//if (compoundPairIdx<maxNumCompoundPairsCapacity)
+							}//
+						}//fi (1)
+					}//if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+				}//for (int b=0;b<numChildrenB;b++)	
+				return;
+			}//if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+			if ((collidables[collidableIndexA].m_shapeType!=SHAPE_CONCAVE_TRIMESH) 
+				&& (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS))
+			{
+				int numChildrenB = collidables[collidableIndexB].m_numChildShapes;
+				for (int b=0;b<numChildrenB;b++)
+				{
+					int childShapeIndexB = collidables[collidableIndexB].m_shapeIndex+b;
+					int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+					b3Quat ornB = rigidBodies[bodyIndexB].m_quat;
+					float4 posB = rigidBodies[bodyIndexB].m_pos;
+					float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+					b3Quat childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+					float4 newPosB = b3QuatRotate(ornB,childPosB)+posB;
+					b3Quat newOrnB = b3QuatMul(ornB,childOrnB);
+
+					int shapeIndexB = collidables[childColIndexB].m_shapeIndex;
+
+
+					//////////////////////////////////////
+
+					if (1)
+					{
+					//	int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+					//	float dmin = FLT_MAX;
+						float4 posA = rigidBodies[bodyIndexA].m_pos;
+						posA.w = 0.f;
+						float4 posB = newPosB;
+						posB.w = 0.f;
+						float4 c0local = convexShapes[shapeIndexA].m_localCenter;
+						b3Quat ornA = rigidBodies[bodyIndexA].m_quat;
+						float4 c0;
+						c0 = transform(&c0local, &posA, &ornA);
+						float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+						b3Quat ornB =newOrnB;
+						float4 c1;
+						c1 = transform(&c1local,&posB,&ornB);
+					//	const float4 DeltaC2 = c0 - c1;
+						{//
+							int compoundPairIdx = b3AtomicInc(numCompoundPairsOut);
+							if (compoundPairIdx<maxNumCompoundPairsCapacity)
+							{
+								gpuCompoundPairsOut[compoundPairIdx] = b3MakeInt4(bodyIndexA,bodyIndexB,-1,childShapeIndexB);
+							}//fi (compoundPairIdx<maxNumCompoundPairsCapacity)
+						}//
+					}//fi (1)	
+				}//for (int b=0;b<numChildrenB;b++)
+				return;
+			}//if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+			return;
+		}//fi ((collidables[collidableIndexA].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS) ||(collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS))
+	}//i<numPairs
+}
+
+
+
+__kernel void   processCompoundPairsKernel( __global const b3Int4* gpuCompoundPairs,
+										__global const b3RigidBodyData* rigidBodies, 
+										__global const b3Collidable* collidables,
+										__global const b3ConvexPolyhedronData* convexShapes, 
+										__global const b3AlignedObjectArray<b3Float4>& vertices,
+										__global const b3AlignedObjectArray<b3Float4>& uniqueEdges,
+										__global const b3AlignedObjectArray<b3GpuFace>& faces,
+										__global const b3AlignedObjectArray<int>& indices,
+										__global b3Aabb* aabbs,
+										__global const b3GpuChildShape* gpuChildShapes,
+										__global b3AlignedObjectArray<b3Float4>& gpuCompoundSepNormalsOut,
+										__global b3AlignedObjectArray<int>& gpuHasCompoundSepNormalsOut,
+										int numCompoundPairs,
+										int i
+										)
+{
+
+//	int i = get_global_id(0);
+	if (i<numCompoundPairs)
+	{
+		int bodyIndexA = gpuCompoundPairs[i].x;
+		int bodyIndexB = gpuCompoundPairs[i].y;
+
+		int childShapeIndexA = gpuCompoundPairs[i].z;
+		int childShapeIndexB = gpuCompoundPairs[i].w;
+		
+		int collidableIndexA = -1;
+		int collidableIndexB = -1;
+		
+		b3Quat ornA = rigidBodies[bodyIndexA].m_quat;
+		float4 posA = rigidBodies[bodyIndexA].m_pos;
+		
+		b3Quat ornB = rigidBodies[bodyIndexB].m_quat;
+		float4 posB = rigidBodies[bodyIndexB].m_pos;
+							
+		if (childShapeIndexA >= 0)
+		{
+			collidableIndexA = gpuChildShapes[childShapeIndexA].m_shapeIndex;
+			float4 childPosA = gpuChildShapes[childShapeIndexA].m_childPosition;
+			b3Quat	childOrnA = gpuChildShapes[childShapeIndexA].m_childOrientation;
+			float4 newPosA = b3QuatRotate(ornA,childPosA)+posA;
+			b3Quat newOrnA = b3QuatMul(ornA,childOrnA);
+			posA = newPosA;
+			ornA = newOrnA;
+		} else
+		{
+			collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+		}
+		
+		if (childShapeIndexB>=0)
+		{
+			collidableIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+			b3Quat childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+			float4 newPosB = b3QuatRotate(ornB,childPosB)+posB;
+			b3Quat newOrnB = b3QuatMul(ornB,childOrnB);
+			posB = newPosB;
+			ornB = newOrnB;
+		} else
+		{
+			collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;	
+		}
+	
+		gpuHasCompoundSepNormalsOut[i] = 0;
+	
+		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+	
+		int shapeTypeA = collidables[collidableIndexA].m_shapeType;
+		int shapeTypeB = collidables[collidableIndexB].m_shapeType;
+	
+
+		if ((shapeTypeA != SHAPE_CONVEX_HULL) || (shapeTypeB != SHAPE_CONVEX_HULL))
+		{
+			return;
+		}
+
+		int hasSeparatingAxis = 5;
+							
+	//	int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+		float dmin = FLT_MAX;
+		posA.w = 0.f;
+		posB.w = 0.f;
+		float4 c0local = convexShapes[shapeIndexA].m_localCenter;
+		float4 c0 = transform(&c0local, &posA, &ornA);
+		float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+		float4 c1 = transform(&c1local,&posB,&ornB);
+		const float4 DeltaC2 = c0 - c1;
+		float4 sepNormal = make_float4(1,0,0,0);
+//		bool sepA = findSeparatingAxis(	convexShapes[shapeIndexA], convexShapes[shapeIndexB],posA,ornA,posB,ornB,DeltaC2,vertices,uniqueEdges,faces,indices,&sepNormal,&dmin);
+		bool sepA = findSeparatingAxis(	convexShapes[shapeIndexA], convexShapes[shapeIndexB],posA,ornA,posB,ornB,vertices,uniqueEdges,faces,indices,vertices,uniqueEdges,faces,indices,sepNormal);//,&dmin);
+	
+		hasSeparatingAxis = 4;
+		if (!sepA)
+		{
+			hasSeparatingAxis = 0;
+		} else
+		{
+			bool sepB = findSeparatingAxis(	convexShapes[shapeIndexB],convexShapes[shapeIndexA],posB,ornB,posA,ornA,vertices,uniqueEdges,faces,indices,vertices,uniqueEdges,faces,indices,sepNormal);//,&dmin);
+
+			if (!sepB)
+			{
+				hasSeparatingAxis = 0;
+			} else//(!sepB)
+			{
+				bool sepEE = findSeparatingAxisEdgeEdge(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,posB,ornB,DeltaC2,vertices,uniqueEdges,faces,indices,&sepNormal,&dmin);
+				if (sepEE)
+				{
+						gpuCompoundSepNormalsOut[i] = sepNormal;//fastNormalize4(sepNormal);
+						gpuHasCompoundSepNormalsOut[i] = 1;
+				}//sepEE
+			}//(!sepB)
+		}//(!sepA)
+		
+		
+	}
+		
+}
+
+
+__kernel void   clipCompoundsHullHullKernel( __global const b3Int4* gpuCompoundPairs, 
+																					__global const b3RigidBodyData* rigidBodies, 
+																					__global const b3Collidable* collidables,
+																					__global const b3ConvexPolyhedronData* convexShapes, 
+																					__global const b3AlignedObjectArray<b3Float4>& vertices,
+																					__global const b3AlignedObjectArray<b3Float4>& uniqueEdges,
+																					__global const b3AlignedObjectArray<b3GpuFace>& faces,
+																					__global const b3AlignedObjectArray<int>& indices,
+																					__global const b3GpuChildShape* gpuChildShapes,
+																					__global const b3AlignedObjectArray<b3Float4>& gpuCompoundSepNormalsOut,
+																					__global const b3AlignedObjectArray<int>& gpuHasCompoundSepNormalsOut,
+																					__global struct b3Contact4Data* globalContactsOut,
+																					int* nGlobalContactsOut,
+																					int numCompoundPairs, int maxContactCapacity, int i)
+{
+
+//	int i = get_global_id(0);
+	int pairIndex = i;
+	
+	float4 worldVertsB1[64];
+	float4 worldVertsB2[64];
+	int capacityWorldVerts = 64;	
+
+	float4 localContactsOut[64];
+	int localContactCapacity=64;
+	
+	float minDist = -1e30f;
+	float maxDist = 0.0f;
+
+	if (i<numCompoundPairs)
+	{
+
+		if (gpuHasCompoundSepNormalsOut[i])
+		{
+
+			int bodyIndexA = gpuCompoundPairs[i].x;
+			int bodyIndexB = gpuCompoundPairs[i].y;
+			
+			int childShapeIndexA = gpuCompoundPairs[i].z;
+			int childShapeIndexB = gpuCompoundPairs[i].w;
+			
+			int collidableIndexA = -1;
+			int collidableIndexB = -1;
+			
+			b3Quat ornA = rigidBodies[bodyIndexA].m_quat;
+			float4 posA = rigidBodies[bodyIndexA].m_pos;
+			
+			b3Quat ornB = rigidBodies[bodyIndexB].m_quat;
+			float4 posB = rigidBodies[bodyIndexB].m_pos;
+								
+			if (childShapeIndexA >= 0)
+			{
+				collidableIndexA = gpuChildShapes[childShapeIndexA].m_shapeIndex;
+				float4 childPosA = gpuChildShapes[childShapeIndexA].m_childPosition;
+				b3Quat childOrnA = gpuChildShapes[childShapeIndexA].m_childOrientation;
+				float4 newPosA = b3QuatRotate(ornA,childPosA)+posA;
+				b3Quat newOrnA = b3QuatMul(ornA,childOrnA);
+				posA = newPosA;
+				ornA = newOrnA;
+			} else
+			{
+				collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+			}
+			
+			if (childShapeIndexB>=0)
+			{
+				collidableIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+				float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+				b3Quat  childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+				float4 newPosB = b3QuatRotate(ornB,childPosB)+posB;
+				b3Quat  newOrnB = b3QuatMul(ornB,childOrnB);
+				posB = newPosB;
+				ornB = newOrnB;
+			} else
+			{
+				collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;	
+			}
+			
+			int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+			int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+		
+			int numLocalContactsOut = clipHullAgainstHull(gpuCompoundSepNormalsOut[i],
+														convexShapes[shapeIndexA], convexShapes[shapeIndexB],
+														posA,ornA,
+													  posB,ornB,
+													  worldVertsB1,worldVertsB2,capacityWorldVerts,
+														minDist, maxDist,
+														vertices,faces,indices,
+														vertices,faces,indices,
+														localContactsOut,localContactCapacity);
+												
+		if (numLocalContactsOut>0)
+		{
+				float4 normal = -gpuCompoundSepNormalsOut[i];
+				int nPoints = numLocalContactsOut;
+				float4* pointsIn = localContactsOut;
+				b3Int4 contactIdx;// = {-1,-1,-1,-1};
+
+				contactIdx.s[0] = 0;
+				contactIdx.s[1] = 1;
+				contactIdx.s[2] = 2;
+				contactIdx.s[3] = 3;
+
+				int nReducedContacts = extractManifoldSequentialGlobal(pointsIn, nPoints, normal, &contactIdx);
+		
+				int dstIdx;
+				dstIdx = b3AtomicInc( nGlobalContactsOut);
+				if ((dstIdx+nReducedContacts) < maxContactCapacity)
+				{
+					__global struct b3Contact4Data* c = globalContactsOut+ dstIdx;
+					c->m_worldNormalOnB = -normal;
+					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+					c->m_batchIdx = pairIndex;
+					int bodyA = gpuCompoundPairs[pairIndex].x;
+					int bodyB = gpuCompoundPairs[pairIndex].y;
+					c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;
+					c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;
+					c->m_childIndexA = childShapeIndexA;
+					c->m_childIndexB = childShapeIndexB;
+					for (int i=0;i<nReducedContacts;i++)
+					{
+						c->m_worldPosB[i] = pointsIn[contactIdx.s[i]];
+					}
+					b3Contact4Data_setNumPoints(c,nReducedContacts);
+				}
+				
+			}//		if (numContactsOut>0)
+		}//		if (gpuHasCompoundSepNormalsOut[i])
+	}//	if (i<numCompoundPairs)
+
+}
+
+
+void computeContactCompoundCompound(int pairIndex,
+																int bodyIndexA, int bodyIndexB, 
+																int collidableIndexA, int collidableIndexB, 
+																const b3RigidBodyData* rigidBodies, 
+																const b3Collidable* collidables,
+																const b3ConvexPolyhedronData* convexShapes,
+																const b3GpuChildShape* cpuChildShapes,
+																const b3AlignedObjectArray<b3Aabb>& hostAabbsWorldSpace,
+																const b3AlignedObjectArray<b3Aabb>& hostAabbsLocalSpace,
+
+																const b3AlignedObjectArray<b3Vector3>& convexVertices,
+																const b3AlignedObjectArray<b3Vector3>& hostUniqueEdges,
+																const b3AlignedObjectArray<int>& convexIndices,
+																const b3AlignedObjectArray<b3GpuFace>& faces,
+																
+																b3Contact4* globalContactsOut,
+																int& nGlobalContactsOut,
+																int maxContactCapacity,
+																b3AlignedObjectArray<b3QuantizedBvhNode>&	treeNodesCPU,
+																b3AlignedObjectArray<b3BvhSubtreeInfo>&	subTreesCPU,
+																b3AlignedObjectArray<b3BvhInfo>&	bvhInfoCPU
+																)
+{
+
+	int shapeTypeB = collidables[collidableIndexB].m_shapeType;
+	b3Assert(shapeTypeB == SHAPE_COMPOUND_OF_CONVEX_HULLS);
+
+	b3AlignedObjectArray<b3Int4> cpuCompoundPairsOut;
+	int numCompoundPairsOut=0;
+	int maxNumCompoundPairsCapacity = 8192;//1024;
+	cpuCompoundPairsOut.resize(maxNumCompoundPairsCapacity);
+
+	// work-in-progress
+	findCompoundPairsKernel( 
+							pairIndex,
+							bodyIndexA,bodyIndexB,
+							collidableIndexA,collidableIndexB,
+							rigidBodies, 
+							collidables,
+							convexShapes, 
+							convexVertices,
+							hostAabbsWorldSpace,
+							hostAabbsLocalSpace,
+							cpuChildShapes,
+							&cpuCompoundPairsOut[0],
+							&numCompoundPairsOut,
+							maxNumCompoundPairsCapacity	,
+							treeNodesCPU,
+							subTreesCPU,
+							bvhInfoCPU
+							);
+
+	printf("maxNumAabbChecks=%d\n",maxNumAabbChecks);
+	if (numCompoundPairsOut>maxNumCompoundPairsCapacity)
+	{
+		b3Error("numCompoundPairsOut exceeded maxNumCompoundPairsCapacity (%d)\n",maxNumCompoundPairsCapacity);
+		numCompoundPairsOut=maxNumCompoundPairsCapacity;
+	}
+	b3AlignedObjectArray<b3Float4> cpuCompoundSepNormalsOut;
+	b3AlignedObjectArray<int> cpuHasCompoundSepNormalsOut;
+	cpuCompoundSepNormalsOut.resize(numCompoundPairsOut);
+	cpuHasCompoundSepNormalsOut.resize(numCompoundPairsOut);
+
+	for (int i=0;i<numCompoundPairsOut;i++)
+	{
+
+		processCompoundPairsKernel(&cpuCompoundPairsOut[0],rigidBodies,collidables,convexShapes,convexVertices,hostUniqueEdges,faces,convexIndices,0,cpuChildShapes,
+			cpuCompoundSepNormalsOut,cpuHasCompoundSepNormalsOut,numCompoundPairsOut,i);
+	}
+
+	for (int i=0;i<numCompoundPairsOut;i++)
+	{
+		clipCompoundsHullHullKernel(&cpuCompoundPairsOut[0],rigidBodies,collidables,convexShapes,convexVertices,hostUniqueEdges,faces,convexIndices,cpuChildShapes,
+			cpuCompoundSepNormalsOut,cpuHasCompoundSepNormalsOut,globalContactsOut,&nGlobalContactsOut,numCompoundPairsOut,maxContactCapacity,i);
+	}
+		/*
+		int childColIndexA = gpuChildShapes[childShapeIndexA].m_shapeIndex;
+
+					float4 posA = rigidBodies[bodyIndexA].m_pos;
+					b3Quat ornA = rigidBodies[bodyIndexA].m_quat;
+					float4 childPosA = gpuChildShapes[childShapeIndexA].m_childPosition;
+					b3Quat childOrnA = gpuChildShapes[childShapeIndexA].m_childOrientation;
+					float4 newPosA = b3QuatRotate(ornA,childPosA)+posA;
+					b3Quat newOrnA = b3QuatMul(ornA,childOrnA);
+
+					int shapeIndexA = collidables[childColIndexA].m_shapeIndex;
+
+
+			bool foundSepAxis = findSeparatingAxis(hullA,hullB,
+							posA,
+							ornA,
+							posB,
+							ornB,
+
+							convexVertices,uniqueEdges,faces,convexIndices,
+							convexVertices,uniqueEdges,faces,convexIndices,
+							
+							sepNormalWorldSpace
+							);
+							*/
+
+	
+	/*
+	if (foundSepAxis)
+	{
+		
+		
+		contactIndex = clipHullHullSingle(
+			bodyIndexA, bodyIndexB,
+						   posA,ornA,
+						   posB,ornB,
+			collidableIndexA, collidableIndexB,
+			&rigidBodies, 
+			&globalContactsOut,
+			nGlobalContactsOut,
+			
+			convexShapes,
+			convexShapes,
+	
+			convexVertices, 
+			uniqueEdges, 
+			faces,
+			convexIndices,
+	
+			convexVertices,
+			uniqueEdges,
+			faces,
+			convexIndices,
+
+			collidables,
+			collidables,
+			sepNormalWorldSpace,
+			maxContactCapacity);
+			
+	}
+	*/
+
+//	return contactIndex;
+	
+	/*
+
+	int numChildrenB = collidables[collidableIndexB].m_numChildShapes;
+	for (int c=0;c<numChildrenB;c++)
+	{
+		int childShapeIndexB = collidables[collidableIndexB].m_shapeIndex+c;
+		int childColIndexB = cpuChildShapes[childShapeIndexB].m_shapeIndex;
+
+		float4 rootPosB = rigidBodies[bodyIndexB].m_pos;
+		b3Quaternion rootOrnB = rigidBodies[bodyIndexB].m_quat;
+		b3Vector3 childPosB = cpuChildShapes[childShapeIndexB].m_childPosition;
+		b3Quaternion childOrnB = cpuChildShapes[childShapeIndexB].m_childOrientation;
+		float4  posB = b3QuatRotate(rootOrnB,childPosB)+rootPosB;
+		b3Quaternion ornB = b3QuatMul(rootOrnB,childOrnB);//b3QuatMul(ornB,childOrnB);
+
+		int shapeIndexB = collidables[childColIndexB].m_shapeIndex;
+
+		const b3ConvexPolyhedronData* hullB = &convexShapes[shapeIndexB];
+
+	}
+	*/
+
+}
+
+void computeContactPlaneCompound(int pairIndex,
+																int bodyIndexA, int bodyIndexB, 
+																int collidableIndexA, int collidableIndexB, 
+																const b3RigidBodyData* rigidBodies, 
+																const b3Collidable* collidables,
+																const b3ConvexPolyhedronData* convexShapes,
+																const b3GpuChildShape* cpuChildShapes,
+																const b3Vector3* convexVertices,
+																const int* convexIndices,
+																const b3GpuFace* faces,
+																
+																b3Contact4* globalContactsOut,
+																int& nGlobalContactsOut,
+																int maxContactCapacity)
+{
+
+	int shapeTypeB = collidables[collidableIndexB].m_shapeType;
+	b3Assert(shapeTypeB == SHAPE_COMPOUND_OF_CONVEX_HULLS);
+
+
+	int numChildrenB = collidables[collidableIndexB].m_numChildShapes;
+	for (int c=0;c<numChildrenB;c++)
+	{
+		int childShapeIndexB = collidables[collidableIndexB].m_shapeIndex+c;
+		int childColIndexB = cpuChildShapes[childShapeIndexB].m_shapeIndex;
+
+		float4 rootPosB = rigidBodies[bodyIndexB].m_pos;
+		b3Quaternion rootOrnB = rigidBodies[bodyIndexB].m_quat;
+		b3Vector3 childPosB = cpuChildShapes[childShapeIndexB].m_childPosition;
+		b3Quaternion childOrnB = cpuChildShapes[childShapeIndexB].m_childOrientation;
+		float4  posB = b3QuatRotate(rootOrnB,childPosB)+rootPosB;
+		b3Quaternion ornB = rootOrnB*childOrnB;//b3QuatMul(ornB,childOrnB);
+
+		int shapeIndexB = collidables[childColIndexB].m_shapeIndex;
+
+		const b3ConvexPolyhedronData* hullB = &convexShapes[shapeIndexB];
+	
+		
+		b3Vector3 posA = rigidBodies[bodyIndexA].m_pos;
+		b3Quaternion ornA = rigidBodies[bodyIndexA].m_quat;
+
+	//	int numContactsOut = 0;
+	//	int numWorldVertsB1= 0;
+
+		b3Vector3 planeEq = faces[collidables[collidableIndexA].m_shapeIndex].m_plane;
+		b3Vector3 planeNormal=b3MakeVector3(planeEq.x,planeEq.y,planeEq.z);
+		b3Vector3 planeNormalWorld = b3QuatRotate(ornA,planeNormal);
+		float planeConstant = planeEq.w;
+		b3Transform convexWorldTransform;
+		convexWorldTransform.setIdentity();
+		convexWorldTransform.setOrigin(posB);
+		convexWorldTransform.setRotation(ornB);
+		b3Transform planeTransform;
+		planeTransform.setIdentity();
+		planeTransform.setOrigin(posA);
+		planeTransform.setRotation(ornA);
+
+		b3Transform planeInConvex;
+		planeInConvex= convexWorldTransform.inverse() * planeTransform;
+		b3Transform convexInPlane;
+		convexInPlane = planeTransform.inverse() * convexWorldTransform;
+	
+		b3Vector3 planeNormalInConvex = planeInConvex.getBasis()*-planeNormal;
+		float maxDot = -1e30;
+		int hitVertex=-1;
+		b3Vector3 hitVtx;
+
+	#define MAX_PLANE_CONVEX_POINTS 64
+
+		b3Vector3 contactPoints[MAX_PLANE_CONVEX_POINTS];
+		int numPoints = 0;
+
+		b3Int4 contactIdx;
+		contactIdx.s[0] = 0;
+		contactIdx.s[1] = 1;
+		contactIdx.s[2] = 2;
+		contactIdx.s[3] = 3;
+	
+		for (int i=0;i<hullB->m_numVertices;i++)
+		{
+			b3Vector3 vtx = convexVertices[hullB->m_vertexOffset+i];
+			float curDot = vtx.dot(planeNormalInConvex);
+
+
+			if (curDot>maxDot)
+			{
+				hitVertex=i;
+				maxDot=curDot;
+				hitVtx = vtx;
+				//make sure the deepest points is always included
+				if (numPoints==MAX_PLANE_CONVEX_POINTS)
+					numPoints--;
+			}
+
+			if (numPoints<MAX_PLANE_CONVEX_POINTS)
+			{
+				b3Vector3 vtxWorld = convexWorldTransform*vtx;
+				b3Vector3 vtxInPlane = planeTransform.inverse()*vtxWorld;
+				float dist = planeNormal.dot(vtxInPlane)-planeConstant;
+				if (dist<0.f)
+				{
+					vtxWorld.w = dist;
+					contactPoints[numPoints] = vtxWorld;
+					numPoints++;
+				}
+			}
+
+		}
+
+		int numReducedPoints  = 0;
+
+		numReducedPoints = numPoints;
+	
+		if (numPoints>4)
+		{
+			numReducedPoints = extractManifoldSequentialGlobal( contactPoints, numPoints, planeNormalInConvex, &contactIdx);
+		}
+		int dstIdx;
+	//    dstIdx = nGlobalContactsOut++;//AppendInc( nGlobalContactsOut, dstIdx );
+		
+		if (numReducedPoints>0)
+		{
+			if (nGlobalContactsOut < maxContactCapacity)
+			{
+				dstIdx=nGlobalContactsOut;
+				nGlobalContactsOut++;
+
+				b3Contact4* c = &globalContactsOut[dstIdx];
+				c->m_worldNormalOnB = -planeNormalWorld;
+				c->setFrictionCoeff(0.7);
+				c->setRestituitionCoeff(0.f);
+
+				c->m_batchIdx = pairIndex;
+				c->m_bodyAPtrAndSignBit = rigidBodies[bodyIndexA].m_invMass==0?-bodyIndexA:bodyIndexA;
+				c->m_bodyBPtrAndSignBit = rigidBodies[bodyIndexB].m_invMass==0?-bodyIndexB:bodyIndexB;
+				for (int i=0;i<numReducedPoints;i++)
+				{
+					b3Vector3 pOnB1 = contactPoints[contactIdx.s[i]];
+					c->m_worldPosB[i] = pOnB1;
+				}
+				c->m_worldNormalOnB.w = (b3Scalar)numReducedPoints;
+			}//if (dstIdx < numPairs)
+		}	
+		
+	}
+
+	
+}
+
+
+
+
+
+void	computeContactSphereConvex(int pairIndex,
+																int bodyIndexA, int bodyIndexB, 
+																int collidableIndexA, int collidableIndexB, 
+																const b3RigidBodyData* rigidBodies, 
+																const b3Collidable* collidables,
+																const b3ConvexPolyhedronData* convexShapes,
+																const b3Vector3* convexVertices,
+																const int* convexIndices,
+																const b3GpuFace* faces,
+																b3Contact4* globalContactsOut,
+																int& nGlobalContactsOut,
+																int maxContactCapacity)
+{
+
+	float radius = collidables[collidableIndexA].m_radius;
+	float4 spherePos1 = rigidBodies[bodyIndexA].m_pos;
+	b3Quaternion sphereOrn = rigidBodies[bodyIndexA].m_quat;
+
+
+
+	float4 pos = rigidBodies[bodyIndexB].m_pos;
+	
+
+	b3Quaternion quat = rigidBodies[bodyIndexB].m_quat;
+
+	b3Transform tr;
+	tr.setIdentity();
+	tr.setOrigin(pos);
+	tr.setRotation(quat);
+	b3Transform trInv = tr.inverse();
+
+	float4 spherePos = trInv(spherePos1);
+
+	int collidableIndex = rigidBodies[bodyIndexB].m_collidableIdx;
+	int shapeIndex = collidables[collidableIndex].m_shapeIndex;
+	int numFaces = convexShapes[shapeIndex].m_numFaces;
+	float4 closestPnt = b3MakeVector3(0, 0, 0, 0);
+//	float4 hitNormalWorld = b3MakeVector3(0, 0, 0, 0);
+	float minDist = -1000000.f; // TODO: What is the largest/smallest float?
+	bool bCollide = true;
+	int region = -1;
+	float4 localHitNormal;
+	for ( int f = 0; f < numFaces; f++ )
+	{
+		b3GpuFace face = faces[convexShapes[shapeIndex].m_faceOffset+f];
+		float4 planeEqn;
+		float4 localPlaneNormal = b3MakeVector3(face.m_plane.x,face.m_plane.y,face.m_plane.z,0.f);
+		float4 n1 = localPlaneNormal;//quatRotate(quat,localPlaneNormal);
+		planeEqn = n1;
+		planeEqn[3] = face.m_plane.w;
+
+		float4 pntReturn;
+		float dist = signedDistanceFromPointToPlane(spherePos, planeEqn, &pntReturn);
+
+		if ( dist > radius)
+		{
+			bCollide = false;
+			break;
+		}
+
+		if ( dist > 0 )
+		{
+			//might hit an edge or vertex
+			b3Vector3 out;
+
+			bool isInPoly = IsPointInPolygon(spherePos,
+					&face,
+					&convexVertices[convexShapes[shapeIndex].m_vertexOffset],
+					convexIndices,
+                    &out);
+			if (isInPoly)
+			{
+				if (dist>minDist)
+				{
+					minDist = dist;
+					closestPnt = pntReturn;
+					localHitNormal = planeEqn;
+					region=1;
+				}
+			} else
+			{
+				b3Vector3 tmp = spherePos-out;
+				b3Scalar l2 = tmp.length2();
+				if (l2<radius*radius)
+				{
+					dist  = b3Sqrt(l2);
+					if (dist>minDist)
+					{
+						minDist = dist;
+						closestPnt = out;
+						localHitNormal = tmp/dist;
+						region=2;
+					}
+					
+				} else
+				{
+					bCollide = false;
+					break;
+				}
+			}
+		}
+		else
+		{
+			if ( dist > minDist )
+			{
+				minDist = dist;
+				closestPnt = pntReturn;
+				localHitNormal = planeEqn;
+				region=3;
+			}
+		}
+	}
+	static int numChecks = 0;
+	numChecks++;
+
+	if (bCollide && minDist > -10000)
+	{
+		
+		float4 normalOnSurfaceB1 = tr.getBasis()*localHitNormal;//-hitNormalWorld;
+		float4 pOnB1 = tr(closestPnt);
+		//printf("dist ,%f,",minDist);
+		float actualDepth = minDist-radius;
+		if (actualDepth<0)
+		{
+		//printf("actualDepth = ,%f,", actualDepth);
+		//printf("normalOnSurfaceB1 = ,%f,%f,%f,", normalOnSurfaceB1.x,normalOnSurfaceB1.y,normalOnSurfaceB1.z);
+		//printf("region=,%d,\n", region);
+		pOnB1[3] = actualDepth;
+
+		int dstIdx;
+//    dstIdx = nGlobalContactsOut++;//AppendInc( nGlobalContactsOut, dstIdx );
+		
+		if (nGlobalContactsOut < maxContactCapacity)
+		{
+			dstIdx=nGlobalContactsOut;
+			nGlobalContactsOut++;
+
+			b3Contact4* c = &globalContactsOut[dstIdx];
+			c->m_worldNormalOnB = normalOnSurfaceB1;
+			c->setFrictionCoeff(0.7);
+			c->setRestituitionCoeff(0.f);
+
+			c->m_batchIdx = pairIndex;
+			c->m_bodyAPtrAndSignBit = rigidBodies[bodyIndexA].m_invMass==0?-bodyIndexA:bodyIndexA;
+			c->m_bodyBPtrAndSignBit = rigidBodies[bodyIndexB].m_invMass==0?-bodyIndexB:bodyIndexB;
+			c->m_worldPosB[0] = pOnB1;
+			int numPoints = 1;
+			c->m_worldNormalOnB.w = (b3Scalar)numPoints;
+		}//if (dstIdx < numPairs)
+		}
+	}//if (hasCollision)
+	
+}
+
+
+
+
+int computeContactConvexConvex2(
+																int pairIndex,
+																int bodyIndexA, int bodyIndexB, 
+																int collidableIndexA, int collidableIndexB, 
+																const b3AlignedObjectArray<b3RigidBodyData>& rigidBodies, 
+																const b3AlignedObjectArray<b3Collidable>& collidables,
+																const b3AlignedObjectArray<b3ConvexPolyhedronData>& convexShapes,
+																const b3AlignedObjectArray<b3Vector3>& convexVertices,
+																const b3AlignedObjectArray<b3Vector3>& uniqueEdges,
+																const b3AlignedObjectArray<int>& convexIndices,
+																const b3AlignedObjectArray<b3GpuFace>& faces,
+																b3AlignedObjectArray<b3Contact4>& globalContactsOut,
+																int& nGlobalContactsOut,
+																int maxContactCapacity,
+																const b3AlignedObjectArray<b3Contact4>& oldContacts
+																)
+{
+	int contactIndex = -1;
+	b3Vector3 posA = rigidBodies[bodyIndexA].m_pos;
+	b3Quaternion ornA = rigidBodies[bodyIndexA].m_quat;
+	b3Vector3 posB = rigidBodies[bodyIndexB].m_pos;
+	b3Quaternion ornB = rigidBodies[bodyIndexB].m_quat;
+	
+
+	b3ConvexPolyhedronData hullA, hullB;
+    
+	b3Vector3 sepNormalWorldSpace;
+
+	
+
+    b3Collidable colA = collidables[collidableIndexA];
+    hullA = convexShapes[colA.m_shapeIndex];
+    //printf("numvertsA = %d\n",hullA.m_numVertices);
+    
+    
+    b3Collidable colB = collidables[collidableIndexB];
+    hullB = convexShapes[colB.m_shapeIndex];
+    //printf("numvertsB = %d\n",hullB.m_numVertices);
+
+//	int contactCapacity = MAX_VERTS;
+	//int numContactsOut=0;
+
+
+#ifdef _WIN32
+	b3Assert(_finite(rigidBodies[bodyIndexA].m_pos.x));
+	b3Assert(_finite(rigidBodies[bodyIndexB].m_pos.x));
+#endif
+	
+		bool foundSepAxis = findSeparatingAxis(hullA,hullB,
+							posA,
+							ornA,
+							posB,
+							ornB,
+
+							convexVertices,uniqueEdges,faces,convexIndices,
+							convexVertices,uniqueEdges,faces,convexIndices,
+							
+							sepNormalWorldSpace
+							);
+
+	
+	if (foundSepAxis)
+	{
+		
+		
+		contactIndex = clipHullHullSingle(
+			bodyIndexA, bodyIndexB,
+						   posA,ornA,
+						   posB,ornB,
+			collidableIndexA, collidableIndexB,
+			&rigidBodies, 
+			&globalContactsOut,
+			nGlobalContactsOut,
+			
+			convexShapes,
+			convexShapes,
+	
+			convexVertices, 
+			uniqueEdges, 
+			faces,
+			convexIndices,
+	
+			convexVertices,
+			uniqueEdges,
+			faces,
+			convexIndices,
+
+			collidables,
+			collidables,
+			sepNormalWorldSpace,
+			maxContactCapacity);
+			
+	}
+
+	return contactIndex;
+}
+
+
+
+
+
+																
+																
+void GpuSatCollision::computeConvexConvexContactsGPUSAT( b3OpenCLArray<b3Int4>* pairs, int nPairs,
+			const b3OpenCLArray<b3RigidBodyData>* bodyBuf,
+			b3OpenCLArray<b3Contact4>* contactOut, int& nContacts,
+			const b3OpenCLArray<b3Contact4>* oldContacts,
+			int maxContactCapacity,
+			int compoundPairCapacity,
+			const b3OpenCLArray<b3ConvexPolyhedronData>& convexData,
+			const b3OpenCLArray<b3Vector3>& gpuVertices,
+			const b3OpenCLArray<b3Vector3>& gpuUniqueEdges,
+			const b3OpenCLArray<b3GpuFace>& gpuFaces,
+			const b3OpenCLArray<int>& gpuIndices,
+			const b3OpenCLArray<b3Collidable>& gpuCollidables,
+			const b3OpenCLArray<b3GpuChildShape>& gpuChildShapes,
+
+			const b3OpenCLArray<b3Aabb>& clAabbsWorldSpace,
+			const b3OpenCLArray<b3Aabb>& clAabbsLocalSpace,
+
+            b3OpenCLArray<b3Vector3>& worldVertsB1GPU,
+            b3OpenCLArray<b3Int4>& clippingFacesOutGPU,
+            b3OpenCLArray<b3Vector3>& worldNormalsAGPU,
+            b3OpenCLArray<b3Vector3>& worldVertsA1GPU,
+            b3OpenCLArray<b3Vector3>& worldVertsB2GPU,    
+			b3AlignedObjectArray<class b3OptimizedBvh*>& bvhDataUnused,
+			b3OpenCLArray<b3QuantizedBvhNode>*	treeNodesGPU,
+			b3OpenCLArray<b3BvhSubtreeInfo>*	subTreesGPU,
+			b3OpenCLArray<b3BvhInfo>*	bvhInfo,
+
+			int numObjects,
+			int maxTriConvexPairCapacity,
+			b3OpenCLArray<b3Int4>& triangleConvexPairsOut,
+			int& numTriConvexPairsOut
+			)
+{
+	myframecount++;
+
+	if (!nPairs)
+		return;
+
+#ifdef CHECK_ON_HOST
+
+
+	b3AlignedObjectArray<b3QuantizedBvhNode>	treeNodesCPU;
+	treeNodesGPU->copyToHost(treeNodesCPU);
+
+	b3AlignedObjectArray<b3BvhSubtreeInfo>	subTreesCPU;
+	subTreesGPU->copyToHost(subTreesCPU);
+
+	b3AlignedObjectArray<b3BvhInfo>	bvhInfoCPU;
+	bvhInfo->copyToHost(bvhInfoCPU);
+
+	b3AlignedObjectArray<b3Aabb> hostAabbsWorldSpace;
+	clAabbsWorldSpace.copyToHost(hostAabbsWorldSpace);
+
+	b3AlignedObjectArray<b3Aabb> hostAabbsLocalSpace;
+	clAabbsLocalSpace.copyToHost(hostAabbsLocalSpace);
+
+	b3AlignedObjectArray<b3Int4> hostPairs;
+	pairs->copyToHost(hostPairs);
+
+	b3AlignedObjectArray<b3RigidBodyData> hostBodyBuf;
+	bodyBuf->copyToHost(hostBodyBuf);
+
+	
+
+	b3AlignedObjectArray<b3ConvexPolyhedronData> hostConvexData;
+	convexData.copyToHost(hostConvexData);
+
+	b3AlignedObjectArray<b3Vector3> hostVertices;
+	gpuVertices.copyToHost(hostVertices);
+
+	b3AlignedObjectArray<b3Vector3> hostUniqueEdges;
+	gpuUniqueEdges.copyToHost(hostUniqueEdges);
+	b3AlignedObjectArray<b3GpuFace> hostFaces;
+	gpuFaces.copyToHost(hostFaces);
+	b3AlignedObjectArray<int> hostIndices;
+	gpuIndices.copyToHost(hostIndices);
+	b3AlignedObjectArray<b3Collidable> hostCollidables;
+	gpuCollidables.copyToHost(hostCollidables);
+	
+	b3AlignedObjectArray<b3GpuChildShape> cpuChildShapes;
+	gpuChildShapes.copyToHost(cpuChildShapes);
+	
+
+	b3AlignedObjectArray<b3Int4> hostTriangleConvexPairs;
+
+	b3AlignedObjectArray<b3Contact4> hostContacts;
+	if (nContacts)
+	{
+		contactOut->copyToHost(hostContacts);
+	}
+
+	b3AlignedObjectArray<b3Contact4> oldHostContacts;
+	
+	if (oldContacts->size())
+	{
+		oldContacts->copyToHost(oldHostContacts);
+	}
+
+
+	hostContacts.resize(maxContactCapacity);
+
+	for (int i=0;i<nPairs;i++)
+	{
+		int bodyIndexA = hostPairs[i].x;
+		int bodyIndexB = hostPairs[i].y;
+		int collidableIndexA = hostBodyBuf[bodyIndexA].m_collidableIdx;
+		int collidableIndexB = hostBodyBuf[bodyIndexB].m_collidableIdx;
+
+		if (hostCollidables[collidableIndexA].m_shapeType == SHAPE_SPHERE &&
+			hostCollidables[collidableIndexB].m_shapeType == SHAPE_CONVEX_HULL)
+		{
+			computeContactSphereConvex(i,bodyIndexA,bodyIndexB,collidableIndexA,collidableIndexB,&hostBodyBuf[0],
+				&hostCollidables[0],&hostConvexData[0],&hostVertices[0],&hostIndices[0],&hostFaces[0],&hostContacts[0],nContacts,maxContactCapacity);
+		}
+
+		if (hostCollidables[collidableIndexA].m_shapeType == SHAPE_CONVEX_HULL &&
+			hostCollidables[collidableIndexB].m_shapeType == SHAPE_SPHERE)
+		{
+			computeContactSphereConvex(i,bodyIndexB,bodyIndexA,collidableIndexB,collidableIndexA,&hostBodyBuf[0],
+				&hostCollidables[0],&hostConvexData[0],&hostVertices[0],&hostIndices[0],&hostFaces[0],&hostContacts[0],nContacts,maxContactCapacity);
+			//printf("convex-sphere\n");
+			
+		}
+
+		if (hostCollidables[collidableIndexA].m_shapeType == SHAPE_CONVEX_HULL &&
+			hostCollidables[collidableIndexB].m_shapeType == SHAPE_PLANE)
+		{
+			computeContactPlaneConvex(i,bodyIndexB,bodyIndexA,collidableIndexB,collidableIndexA,&hostBodyBuf[0],
+			&hostCollidables[0],&hostConvexData[0],&hostVertices[0],&hostIndices[0],&hostFaces[0],&hostContacts[0],nContacts,maxContactCapacity);
+//			printf("convex-plane\n");
+			
+		}
+
+		if (hostCollidables[collidableIndexA].m_shapeType == SHAPE_PLANE &&
+			hostCollidables[collidableIndexB].m_shapeType == SHAPE_CONVEX_HULL)
+		{
+			computeContactPlaneConvex(i,bodyIndexA,bodyIndexB,collidableIndexA,collidableIndexB,&hostBodyBuf[0],
+			&hostCollidables[0],&hostConvexData[0],&hostVertices[0],&hostIndices[0],&hostFaces[0],&hostContacts[0],nContacts,maxContactCapacity);
+//			printf("plane-convex\n");
+			
+		}
+
+			if (hostCollidables[collidableIndexA].m_shapeType == SHAPE_COMPOUND_OF_CONVEX_HULLS &&
+			hostCollidables[collidableIndexB].m_shapeType == SHAPE_COMPOUND_OF_CONVEX_HULLS)
+		{
+			computeContactCompoundCompound(i,bodyIndexB,bodyIndexA,collidableIndexB,collidableIndexA,&hostBodyBuf[0],
+			&hostCollidables[0],&hostConvexData[0],&cpuChildShapes[0], hostAabbsWorldSpace,hostAabbsLocalSpace,hostVertices,hostUniqueEdges,hostIndices,hostFaces,&hostContacts[0],
+			nContacts,maxContactCapacity,treeNodesCPU,subTreesCPU,bvhInfoCPU);	
+//			printf("convex-plane\n");
+			
+		}
+
+
+				if (hostCollidables[collidableIndexA].m_shapeType == SHAPE_COMPOUND_OF_CONVEX_HULLS &&
+			hostCollidables[collidableIndexB].m_shapeType == SHAPE_PLANE)
+		{
+			computeContactPlaneCompound(i,bodyIndexB,bodyIndexA,collidableIndexB,collidableIndexA,&hostBodyBuf[0],
+			&hostCollidables[0],&hostConvexData[0],&cpuChildShapes[0], &hostVertices[0],&hostIndices[0],&hostFaces[0],&hostContacts[0],nContacts,maxContactCapacity);
+//			printf("convex-plane\n");
+			
+		}
+
+		if (hostCollidables[collidableIndexA].m_shapeType == SHAPE_PLANE &&
+			hostCollidables[collidableIndexB].m_shapeType == SHAPE_COMPOUND_OF_CONVEX_HULLS)
+		{
+			computeContactPlaneCompound(i,bodyIndexA,bodyIndexB,collidableIndexA,collidableIndexB,&hostBodyBuf[0],
+			&hostCollidables[0],&hostConvexData[0],&cpuChildShapes[0],&hostVertices[0],&hostIndices[0],&hostFaces[0],&hostContacts[0],nContacts,maxContactCapacity);
+//			printf("plane-convex\n");
+			
+		}
+
+		if (hostCollidables[collidableIndexA].m_shapeType == SHAPE_CONVEX_HULL &&
+			hostCollidables[collidableIndexB].m_shapeType == SHAPE_CONVEX_HULL)
+		{
+			//printf("hostPairs[i].z=%d\n",hostPairs[i].z);
+			int contactIndex = computeContactConvexConvex2(           i,bodyIndexA,bodyIndexB,collidableIndexA,collidableIndexB,hostBodyBuf, hostCollidables,hostConvexData,hostVertices,hostUniqueEdges,hostIndices,hostFaces,hostContacts,nContacts,maxContactCapacity,oldHostContacts);
+			//int contactIndex = computeContactConvexConvex(hostPairs,i,bodyIndexA,bodyIndexB,collidableIndexA,collidableIndexB,hostBodyBuf,hostCollidables,hostConvexData,hostVertices,hostUniqueEdges,hostIndices,hostFaces,hostContacts,nContacts,maxContactCapacity,oldHostContacts);
+
+
+			if (contactIndex>=0)
+			{
+//				printf("convex convex contactIndex = %d\n",contactIndex);
+				hostPairs[i].z = contactIndex;
+			}
+//			printf("plane-convex\n");
+			
+		}
+
+
+	}
+
+	if (hostPairs.size())
+	{
+		pairs->copyFromHost(hostPairs);
+	}
+
+	hostContacts.resize(nContacts);
+	if (nContacts)
+		{
+			
+			contactOut->copyFromHost(hostContacts);
+		} else
+	{
+		contactOut->resize(0);
+		}
+
+		m_totalContactsOut.copyFromHostPointer(&nContacts,1,0,true);
+		//printf("(HOST) nContacts = %d\n",nContacts);
+
+#else
+
+	{
+		if (nPairs)
+		{
+			m_totalContactsOut.copyFromHostPointer(&nContacts,1,0,true);
+
+			B3_PROFILE("primitiveContactsKernel");
+			b3BufferInfoCL bInfo[] = {
+				b3BufferInfoCL( pairs->getBufferCL(), true ), 
+				b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+				b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+				b3BufferInfoCL( convexData.getBufferCL(),true),
+				b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+				b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+				b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+				b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+				b3BufferInfoCL( contactOut->getBufferCL()),
+				b3BufferInfoCL( m_totalContactsOut.getBufferCL())	
+			};
+			
+			b3LauncherCL launcher(m_queue, m_primitiveContactsKernel,"m_primitiveContactsKernel");
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst( nPairs  );
+			launcher.setConst(maxContactCapacity);
+			int num = nPairs;
+			launcher.launch1D( num);
+			clFinish(m_queue);
+		
+			nContacts = m_totalContactsOut.at(0);
+			contactOut->resize(nContacts);
+		}
+	}
+
+	
+#endif//CHECK_ON_HOST
+	
+	B3_PROFILE("computeConvexConvexContactsGPUSAT");
+   // printf("nContacts = %d\n",nContacts);
+    
+	
+	m_sepNormals.resize(nPairs);
+	m_hasSeparatingNormals.resize(nPairs);
+	
+	int concaveCapacity=maxTriConvexPairCapacity;
+	m_concaveSepNormals.resize(concaveCapacity);
+	m_concaveHasSeparatingNormals.resize(concaveCapacity);
+	m_numConcavePairsOut.resize(0);
+	m_numConcavePairsOut.push_back(0);
+
+	
+	m_gpuCompoundPairs.resize(compoundPairCapacity);
+
+	m_gpuCompoundSepNormals.resize(compoundPairCapacity);
+	
+	
+	m_gpuHasCompoundSepNormals.resize(compoundPairCapacity);
+	
+	m_numCompoundPairsOut.resize(0);
+	m_numCompoundPairsOut.push_back(0);
+
+	int numCompoundPairs = 0;
+
+	int numConcavePairs =0;
+
+	{
+		clFinish(m_queue);
+		if (findSeparatingAxisOnGpu)
+		{
+			m_dmins.resize(nPairs);
+			if (splitSearchSepAxisConvex)
+			{
+					
+
+				if (useMprGpu)
+				{
+					nContacts = m_totalContactsOut.at(0);
+					{
+						B3_PROFILE("mprPenetrationKernel");
+						b3BufferInfoCL bInfo[] = { 
+							b3BufferInfoCL( pairs->getBufferCL(), true ), 
+							b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+							b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+							b3BufferInfoCL( convexData.getBufferCL(),true),
+							b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+							b3BufferInfoCL( m_sepNormals.getBufferCL()),
+							b3BufferInfoCL( m_hasSeparatingNormals.getBufferCL()),
+							b3BufferInfoCL( contactOut->getBufferCL()),
+							b3BufferInfoCL( m_totalContactsOut.getBufferCL())
+						};
+
+						b3LauncherCL launcher(m_queue, m_mprPenetrationKernel,"mprPenetrationKernel");
+						launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+
+						launcher.setConst(maxContactCapacity);
+						launcher.setConst( nPairs  );
+
+						int num = nPairs;
+						launcher.launch1D( num);
+						clFinish(m_queue);
+						/*
+						b3AlignedObjectArray<int>hostHasSepAxis;
+						m_hasSeparatingNormals.copyToHost(hostHasSepAxis);
+						b3AlignedObjectArray<b3Vector3>hostSepAxis;
+						m_sepNormals.copyToHost(hostSepAxis);
+						*/
+						nContacts = m_totalContactsOut.at(0);
+						contactOut->resize(nContacts);
+					//	printf("nContacts (after mprPenetrationKernel) = %d\n",nContacts);
+						if (nContacts>maxContactCapacity)
+						{
+                
+							b3Error("Error: contacts exceeds capacity (%d/%d)\n", nContacts, maxContactCapacity);
+							nContacts = maxContactCapacity;
+						}
+
+					}
+				}
+				
+				if (1)
+				{
+
+					if (1)
+					{
+					{
+						B3_PROFILE("findSeparatingAxisVertexFaceKernel");
+						b3BufferInfoCL bInfo[] = { 
+							b3BufferInfoCL( pairs->getBufferCL(), true ), 
+							b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+							b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+							b3BufferInfoCL( convexData.getBufferCL(),true),
+							b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+							b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+							b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+							b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+							b3BufferInfoCL( clAabbsWorldSpace.getBufferCL(),true),
+							b3BufferInfoCL( m_sepNormals.getBufferCL()),
+							b3BufferInfoCL( m_hasSeparatingNormals.getBufferCL()),
+							b3BufferInfoCL( m_dmins.getBufferCL())
+						};
+
+						b3LauncherCL launcher(m_queue, m_findSeparatingAxisVertexFaceKernel,"findSeparatingAxisVertexFaceKernel");
+						launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+						launcher.setConst( nPairs  );
+
+						int num = nPairs;
+						launcher.launch1D( num);
+						clFinish(m_queue);
+					}
+
+
+					int numDirections = sizeof(unitSphere162)/sizeof(b3Vector3);
+					
+					{
+						B3_PROFILE("findSeparatingAxisEdgeEdgeKernel");
+						b3BufferInfoCL bInfo[] = { 
+							b3BufferInfoCL( pairs->getBufferCL(), true ), 
+							b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+							b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+							b3BufferInfoCL( convexData.getBufferCL(),true),
+							b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+							b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+							b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+							b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+							b3BufferInfoCL( clAabbsWorldSpace.getBufferCL(),true),
+							b3BufferInfoCL( m_sepNormals.getBufferCL()),
+							b3BufferInfoCL( m_hasSeparatingNormals.getBufferCL()),
+							b3BufferInfoCL( m_dmins.getBufferCL()),
+							b3BufferInfoCL( m_unitSphereDirections.getBufferCL(),true)
+
+						};
+
+						b3LauncherCL launcher(m_queue, m_findSeparatingAxisEdgeEdgeKernel,"findSeparatingAxisEdgeEdgeKernel");
+						launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+						launcher.setConst( numDirections);
+						launcher.setConst( nPairs  );
+						int num = nPairs;
+						launcher.launch1D( num);
+						clFinish(m_queue);
+
+					}
+					}
+					if (useMprGpu)
+					{
+						B3_PROFILE("findSeparatingAxisUnitSphereKernel");
+						b3BufferInfoCL bInfo[] = { 
+								b3BufferInfoCL( pairs->getBufferCL(), true ), 
+								b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+								b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+								b3BufferInfoCL( convexData.getBufferCL(),true),
+								b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+								b3BufferInfoCL( m_unitSphereDirections.getBufferCL(),true),
+								b3BufferInfoCL( m_sepNormals.getBufferCL()),
+								b3BufferInfoCL( m_hasSeparatingNormals.getBufferCL()),
+								b3BufferInfoCL( m_dmins.getBufferCL())
+						};
+
+						b3LauncherCL launcher(m_queue, m_findSeparatingAxisUnitSphereKernel,"findSeparatingAxisUnitSphereKernel");
+						launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+						int numDirections = sizeof(unitSphere162)/sizeof(b3Vector3);
+						launcher.setConst( numDirections);
+
+						launcher.setConst( nPairs  );
+                                                
+						int num = nPairs;
+						launcher.launch1D( num);
+						clFinish(m_queue);
+					}
+			}
+				
+
+			} else
+			{
+				B3_PROFILE("findSeparatingAxisKernel");
+				b3BufferInfoCL bInfo[] = { 
+					b3BufferInfoCL( pairs->getBufferCL(), true ), 
+					b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+					b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+					b3BufferInfoCL( convexData.getBufferCL(),true),
+					b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+					b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+					b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+					b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+					b3BufferInfoCL( clAabbsWorldSpace.getBufferCL(),true),
+					b3BufferInfoCL( m_sepNormals.getBufferCL()),
+					b3BufferInfoCL( m_hasSeparatingNormals.getBufferCL())
+				};
+
+				b3LauncherCL launcher(m_queue, m_findSeparatingAxisKernel,"m_findSeparatingAxisKernel");
+				launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+				launcher.setConst( nPairs  );
+
+				int num = nPairs;
+				launcher.launch1D( num);
+				clFinish(m_queue);
+			}
+			
+			
+		}
+        else
+        {
+            
+			B3_PROFILE("findSeparatingAxisKernel CPU");
+           
+            
+            b3AlignedObjectArray<b3Int4> hostPairs;
+            pairs->copyToHost(hostPairs);
+            b3AlignedObjectArray<b3RigidBodyData> hostBodyBuf;
+            bodyBuf->copyToHost(hostBodyBuf);
+
+            b3AlignedObjectArray<b3Collidable> hostCollidables;
+            gpuCollidables.copyToHost(hostCollidables);
+            
+            b3AlignedObjectArray<b3GpuChildShape> cpuChildShapes;
+            gpuChildShapes.copyToHost(cpuChildShapes);
+            
+            b3AlignedObjectArray<b3ConvexPolyhedronData> hostConvexShapeData;
+            convexData.copyToHost(hostConvexShapeData);
+            
+            b3AlignedObjectArray<b3Vector3> hostVertices;
+            gpuVertices.copyToHost(hostVertices);
+            
+            b3AlignedObjectArray<int> hostHasSepAxis;
+            hostHasSepAxis.resize(nPairs);
+            b3AlignedObjectArray<b3Vector3> hostSepAxis;
+            hostSepAxis.resize(nPairs);
+            
+            b3AlignedObjectArray<b3Vector3> hostUniqueEdges;
+            gpuUniqueEdges.copyToHost(hostUniqueEdges);
+            b3AlignedObjectArray<b3GpuFace> hostFaces;
+            gpuFaces.copyToHost(hostFaces);
+            
+            b3AlignedObjectArray<int> hostIndices;
+            gpuIndices.copyToHost(hostIndices);
+			
+			b3AlignedObjectArray<b3Contact4> hostContacts;
+			if (nContacts)
+			{
+				contactOut->copyToHost(hostContacts);
+			}
+			hostContacts.resize(maxContactCapacity);
+			int nGlobalContactsOut = nContacts;
+			
+            
+            for (int i=0;i<nPairs;i++)
+            {
+                
+                int bodyIndexA = hostPairs[i].x;
+                int bodyIndexB = hostPairs[i].y;
+                int collidableIndexA = hostBodyBuf[bodyIndexA].m_collidableIdx;
+                int collidableIndexB = hostBodyBuf[bodyIndexB].m_collidableIdx;
+                
+                int shapeIndexA = hostCollidables[collidableIndexA].m_shapeIndex;
+                int shapeIndexB = hostCollidables[collidableIndexB].m_shapeIndex;
+                
+                hostHasSepAxis[i] = 0;
+                
+                //once the broadphase avoids static-static pairs, we can remove this test
+                if ((hostBodyBuf[bodyIndexA].m_invMass==0) &&(hostBodyBuf[bodyIndexB].m_invMass==0))
+                {
+                    continue;
+                }
+                
+                
+                if ((hostCollidables[collidableIndexA].m_shapeType!=SHAPE_CONVEX_HULL) ||(hostCollidables[collidableIndexB].m_shapeType!=SHAPE_CONVEX_HULL))
+                {
+                    continue;
+                }
+                
+                float dmin = FLT_MAX;
+                
+                b3ConvexPolyhedronData* convexShapeA = &hostConvexShapeData[shapeIndexA];
+                b3ConvexPolyhedronData* convexShapeB = &hostConvexShapeData[shapeIndexB];
+                b3Vector3 posA = hostBodyBuf[bodyIndexA].m_pos;
+                b3Vector3 posB = hostBodyBuf[bodyIndexB].m_pos;
+                b3Quaternion ornA =hostBodyBuf[bodyIndexA].m_quat;
+                b3Quaternion ornB =hostBodyBuf[bodyIndexB].m_quat;
+				
+				
+				if (useGjk)
+				{
+
+					//first approximate the separating axis, to 'fail-proof' GJK+EPA or MPR
+					{
+						b3Vector3 c0local = hostConvexShapeData[shapeIndexA].m_localCenter;
+						b3Vector3 c0 = b3TransformPoint(c0local, posA, ornA);
+						b3Vector3 c1local = hostConvexShapeData[shapeIndexB].m_localCenter;
+						b3Vector3 c1 = b3TransformPoint(c1local,posB,ornB);
+						b3Vector3 DeltaC2 = c0 - c1;
+                
+						b3Vector3 sepAxis;
+                
+						bool hasSepAxisA = b3FindSeparatingAxis(convexShapeA, convexShapeB, posA, ornA, posB, ornB, DeltaC2,
+							&hostVertices.at(0), &hostUniqueEdges.at(0), &hostFaces.at(0), &hostIndices.at(0),
+							&hostVertices.at(0), &hostUniqueEdges.at(0), &hostFaces.at(0), &hostIndices.at(0),
+											 &sepAxis, &dmin);
+                
+						if (hasSepAxisA)
+						{
+							bool hasSepAxisB = b3FindSeparatingAxis(convexShapeB, convexShapeA, posB, ornB, posA, ornA, DeltaC2,
+																	&hostVertices.at(0), &hostUniqueEdges.at(0), &hostFaces.at(0), &hostIndices.at(0),
+																	&hostVertices.at(0), &hostUniqueEdges.at(0), &hostFaces.at(0), &hostIndices.at(0),
+																	&sepAxis, &dmin);
+							if (hasSepAxisB)
+							{
+								bool hasEdgeEdge =b3FindSeparatingAxisEdgeEdge(convexShapeA, convexShapeB, posA, ornA, posB, ornB, DeltaC2,
+															 &hostVertices.at(0), &hostUniqueEdges.at(0), &hostFaces.at(0), &hostIndices.at(0),
+															 &hostVertices.at(0), &hostUniqueEdges.at(0), &hostFaces.at(0), &hostIndices.at(0),
+															 &sepAxis, &dmin,false);
+													 
+								if (hasEdgeEdge)
+								{
+									hostHasSepAxis[i] = 1;
+									hostSepAxis[i] = sepAxis;
+									hostSepAxis[i].w = dmin;
+								}
+							}
+						}
+					}
+
+					if (hostHasSepAxis[i])
+					{
+						int pairIndex = i;
+				
+						bool useMpr = true;
+						if (useMpr)
+						{
+							int res=0;
+							float depth = 0.f;
+							b3Vector3 sepAxis2 = b3MakeVector3(1,0,0);
+							b3Vector3 resultPointOnBWorld = b3MakeVector3(0,0,0);
+
+						float depthOut;
+						b3Vector3 dirOut;
+						b3Vector3 posOut;
+						
+
+						//res = b3MprPenetration(bodyIndexA,bodyIndexB,hostBodyBuf,hostConvexShapeData,hostCollidables,hostVertices,&mprConfig,&depthOut,&dirOut,&posOut);
+						res = b3MprPenetration(pairIndex,bodyIndexA,bodyIndexB,&hostBodyBuf[0],&hostConvexShapeData[0],&hostCollidables[0],&hostVertices[0],&hostSepAxis[0],&hostHasSepAxis[0],&depthOut,&dirOut,&posOut);
+						depth = depthOut;
+						sepAxis2 =  b3MakeVector3(-dirOut.x,-dirOut.y,-dirOut.z);
+						resultPointOnBWorld = posOut;
+						//hostHasSepAxis[i] = 0;
+
+
+						if (res==0)
+						{
+							//add point?
+							//printf("depth = %f\n",depth);
+							//printf("normal = %f,%f,%f\n",dir.v[0],dir.v[1],dir.v[2]);
+							//qprintf("pos = %f,%f,%f\n",pos.v[0],pos.v[1],pos.v[2]);
+						
+							
+
+							float dist=0.f;
+
+							const b3ConvexPolyhedronData& hullA = hostConvexShapeData[hostCollidables[hostBodyBuf[bodyIndexA].m_collidableIdx].m_shapeIndex];
+							const b3ConvexPolyhedronData& hullB = hostConvexShapeData[hostCollidables[hostBodyBuf[bodyIndexB].m_collidableIdx].m_shapeIndex];
+
+							if(b3TestSepAxis( &hullA, &hullB, posA,ornA,posB,ornB,&sepAxis2, &hostVertices[0], &hostVertices[0],&dist))
+							{
+								if (depth > dist)
+								{
+									float diff = depth - dist;
+									
+									static float maxdiff = 0.f;
+									if (maxdiff < diff)
+									{
+										maxdiff = diff;
+										printf("maxdiff = %20.10f\n",maxdiff);
+									}
+								}
+							}
+							if (depth > dmin)
+							{
+								b3Vector3 oldAxis = hostSepAxis[i];
+								depth = dmin;
+								sepAxis2 = oldAxis;
+							}
+
+							
+
+							if(b3TestSepAxis( &hullA, &hullB, posA,ornA,posB,ornB,&sepAxis2, &hostVertices[0], &hostVertices[0],&dist))
+							{
+								if (depth > dist)
+								{
+									float diff = depth - dist;
+									//printf("?diff  = %f\n",diff );
+									static float maxdiff = 0.f;
+									if (maxdiff < diff)
+									{
+										maxdiff = diff;
+										printf("maxdiff = %20.10f\n",maxdiff);
+									}
+								}
+								//this is used for SAT
+								//hostHasSepAxis[i] = 1;
+								//hostSepAxis[i] = sepAxis2;
+
+								//add contact point
+
+								//int contactIndex = nGlobalContactsOut;
+								b3Contact4& newContact = hostContacts.at(nGlobalContactsOut);
+								nGlobalContactsOut++;
+								newContact.m_batchIdx = 0;//i;
+								newContact.m_bodyAPtrAndSignBit = (hostBodyBuf.at(bodyIndexA).m_invMass==0)? -bodyIndexA:bodyIndexA;
+								newContact.m_bodyBPtrAndSignBit = (hostBodyBuf.at(bodyIndexB).m_invMass==0)? -bodyIndexB:bodyIndexB;
+
+								newContact.m_frictionCoeffCmp = 45874;
+								newContact.m_restituitionCoeffCmp = 0;
+						
+								
+								static float maxDepth = 0.f;
+							
+								if (depth > maxDepth)
+								{
+									maxDepth  = depth;
+									printf("MPR maxdepth = %f\n",maxDepth );
+							
+								}
+							
+
+								resultPointOnBWorld.w = -depth;
+								newContact.m_worldPosB[0] = resultPointOnBWorld;
+								//b3Vector3 resultPointOnAWorld = resultPointOnBWorld+depth*sepAxis2;
+								newContact.m_worldNormalOnB = sepAxis2;
+								newContact.m_worldNormalOnB.w = (b3Scalar)1;
+							} else
+							{
+								printf("rejected\n");
+							}
+
+			
+						}
+						} else
+						{
+
+			
+				
+			//int contactIndex = computeContactConvexConvex2(           i,bodyIndexA,bodyIndexB,collidableIndexA,collidableIndexB,hostBodyBuf, hostCollidables,hostConvexData,hostVertices,hostUniqueEdges,hostIndices,hostFaces,hostContacts,nContacts,maxContactCapacity,oldHostContacts);
+							b3AlignedObjectArray<b3Contact4> oldHostContacts;	
+							int result;
+							result = computeContactConvexConvex2( //hostPairs,
+													   pairIndex,
+													bodyIndexA, bodyIndexB,
+													   collidableIndexA, collidableIndexB,
+													   hostBodyBuf,
+													   hostCollidables,
+													   hostConvexShapeData,
+													   hostVertices,
+													   hostUniqueEdges,
+													   hostIndices,
+													   hostFaces,
+													   hostContacts,
+													   nGlobalContactsOut,
+														maxContactCapacity,
+														oldHostContacts
+														//hostHasSepAxis,
+														//hostSepAxis
+														
+																);
+						}//mpr
+					}//hostHasSepAxis[i] = 1;
+					
+				} else
+				{
+				
+					b3Vector3 c0local = hostConvexShapeData[shapeIndexA].m_localCenter;
+					b3Vector3 c0 = b3TransformPoint(c0local, posA, ornA);
+					b3Vector3 c1local = hostConvexShapeData[shapeIndexB].m_localCenter;
+					b3Vector3 c1 = b3TransformPoint(c1local,posB,ornB);
+					b3Vector3 DeltaC2 = c0 - c1;
+                
+					b3Vector3 sepAxis;
+                
+					bool hasSepAxisA = b3FindSeparatingAxis(convexShapeA, convexShapeB, posA, ornA, posB, ornB, DeltaC2,
+						&hostVertices.at(0), &hostUniqueEdges.at(0), &hostFaces.at(0), &hostIndices.at(0),
+						&hostVertices.at(0), &hostUniqueEdges.at(0), &hostFaces.at(0), &hostIndices.at(0),
+										 &sepAxis, &dmin);
+                
+					if (hasSepAxisA)
+					{
+						bool hasSepAxisB = b3FindSeparatingAxis(convexShapeB, convexShapeA, posB, ornB, posA, ornA, DeltaC2,
+																&hostVertices.at(0), &hostUniqueEdges.at(0), &hostFaces.at(0), &hostIndices.at(0),
+																&hostVertices.at(0), &hostUniqueEdges.at(0), &hostFaces.at(0), &hostIndices.at(0),
+																&sepAxis, &dmin);
+						if (hasSepAxisB)
+						{
+							bool hasEdgeEdge =b3FindSeparatingAxisEdgeEdge(convexShapeA, convexShapeB, posA, ornA, posB, ornB, DeltaC2,
+														 &hostVertices.at(0), &hostUniqueEdges.at(0), &hostFaces.at(0), &hostIndices.at(0),
+														 &hostVertices.at(0), &hostUniqueEdges.at(0), &hostFaces.at(0), &hostIndices.at(0),
+														 &sepAxis, &dmin,true);
+													 
+							if (hasEdgeEdge)
+							{
+								hostHasSepAxis[i] = 1;
+								hostSepAxis[i] = sepAxis;
+							}
+						}
+					}
+				}
+            }
+            
+			if (useGjkContacts)//nGlobalContactsOut>0)
+			{
+				//printf("nGlobalContactsOut=%d\n",nGlobalContactsOut);
+				nContacts = nGlobalContactsOut;
+				contactOut->copyFromHost(hostContacts);
+	
+				m_totalContactsOut.copyFromHostPointer(&nContacts,1,0,true);
+			}
+            
+            m_hasSeparatingNormals.copyFromHost(hostHasSepAxis);
+            m_sepNormals.copyFromHost(hostSepAxis);
+        
+            /*
+             //double-check results from GPU (comment-out the 'else' so both paths are executed
+            b3AlignedObjectArray<int> checkHasSepAxis;
+            m_hasSeparatingNormals.copyToHost(checkHasSepAxis);
+            static int frameCount = 0;
+            frameCount++;
+            for (int i=0;i<nPairs;i++)
+            {
+                if (hostHasSepAxis[i] != checkHasSepAxis[i])
+                {
+                    printf("at frameCount %d hostHasSepAxis[%d] = %d but checkHasSepAxis[i] = %d\n",
+                           frameCount,i,hostHasSepAxis[i],checkHasSepAxis[i]);
+                }
+            }
+            //m_hasSeparatingNormals.copyFromHost(hostHasSepAxis);
+            //    m_sepNormals.copyFromHost(hostSepAxis);
+            */
+        }
+        
+        
+        numCompoundPairs = m_numCompoundPairsOut.at(0);
+        bool useGpuFindCompoundPairs=true;
+        if (useGpuFindCompoundPairs)
+        {
+            B3_PROFILE("findCompoundPairsKernel");
+            b3BufferInfoCL bInfo[] = 
+            { 
+                b3BufferInfoCL( pairs->getBufferCL(), true ), 
+                b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+                b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+                b3BufferInfoCL( convexData.getBufferCL(),true),
+                b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+                b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+                b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+                b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+                b3BufferInfoCL( clAabbsLocalSpace.getBufferCL(),true),
+                b3BufferInfoCL( gpuChildShapes.getBufferCL(),true),
+                b3BufferInfoCL( m_gpuCompoundPairs.getBufferCL()),
+                b3BufferInfoCL( m_numCompoundPairsOut.getBufferCL()),
+                b3BufferInfoCL(subTreesGPU->getBufferCL()),
+                b3BufferInfoCL(treeNodesGPU->getBufferCL()),
+                b3BufferInfoCL(bvhInfo->getBufferCL())
+            };
+
+            b3LauncherCL launcher(m_queue, m_findCompoundPairsKernel,"m_findCompoundPairsKernel");
+            launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+            launcher.setConst( nPairs  );
+            launcher.setConst( compoundPairCapacity);
+
+            int num = nPairs;
+            launcher.launch1D( num);
+            clFinish(m_queue);
+
+            numCompoundPairs = m_numCompoundPairsOut.at(0);
+            //printf("numCompoundPairs =%d\n",numCompoundPairs );
+            if (numCompoundPairs)
+            {
+                //printf("numCompoundPairs=%d\n",numCompoundPairs);
+            }
+            
+
+        } else
+        {
+
+
+            b3AlignedObjectArray<b3QuantizedBvhNode>	treeNodesCPU;
+            treeNodesGPU->copyToHost(treeNodesCPU);
+
+            b3AlignedObjectArray<b3BvhSubtreeInfo>	subTreesCPU;
+            subTreesGPU->copyToHost(subTreesCPU);
+
+            b3AlignedObjectArray<b3BvhInfo>	bvhInfoCPU;
+            bvhInfo->copyToHost(bvhInfoCPU);
+
+            b3AlignedObjectArray<b3Aabb> hostAabbsWorldSpace;
+            clAabbsWorldSpace.copyToHost(hostAabbsWorldSpace);
+
+            b3AlignedObjectArray<b3Aabb> hostAabbsLocalSpace;
+            clAabbsLocalSpace.copyToHost(hostAabbsLocalSpace);
+
+            b3AlignedObjectArray<b3Int4> hostPairs;
+            pairs->copyToHost(hostPairs);
+
+            b3AlignedObjectArray<b3RigidBodyData> hostBodyBuf;
+            bodyBuf->copyToHost(hostBodyBuf);
+
+
+            b3AlignedObjectArray<b3Int4> cpuCompoundPairsOut;
+            cpuCompoundPairsOut.resize(compoundPairCapacity);
+
+            b3AlignedObjectArray<b3Collidable> hostCollidables;
+            gpuCollidables.copyToHost(hostCollidables);
+
+            b3AlignedObjectArray<b3GpuChildShape> cpuChildShapes;
+            gpuChildShapes.copyToHost(cpuChildShapes);
+
+            b3AlignedObjectArray<b3ConvexPolyhedronData> hostConvexData;
+            convexData.copyToHost(hostConvexData);
+
+            b3AlignedObjectArray<b3Vector3> hostVertices;
+            gpuVertices.copyToHost(hostVertices);
+
+
+
+
+            for (int pairIndex=0;pairIndex<nPairs;pairIndex++)
+            {
+                int bodyIndexA = hostPairs[pairIndex].x;
+                int bodyIndexB = hostPairs[pairIndex].y;
+                int collidableIndexA = hostBodyBuf[bodyIndexA].m_collidableIdx;
+                int collidableIndexB = hostBodyBuf[bodyIndexB].m_collidableIdx;
+				if (cpuChildShapes.size())
+				{
+                findCompoundPairsKernel( 
+                            pairIndex,
+                            bodyIndexA,
+                            bodyIndexB,
+                            collidableIndexA,
+                            collidableIndexB,
+                            &hostBodyBuf[0],
+                            &hostCollidables[0],
+                            &hostConvexData[0],
+                            hostVertices,
+                            hostAabbsWorldSpace,
+                            hostAabbsLocalSpace,
+                            &cpuChildShapes[0],
+                            &cpuCompoundPairsOut[0],
+                            &numCompoundPairs,
+                            compoundPairCapacity,
+                            treeNodesCPU,
+                            subTreesCPU,
+                            bvhInfoCPU
+                            );
+				}
+            }
+            
+
+			m_numCompoundPairsOut.copyFromHostPointer(&numCompoundPairs,1,0,true);
+			if (numCompoundPairs)
+			{
+				b3CompoundOverlappingPair* ptr = (b3CompoundOverlappingPair*)&cpuCompoundPairsOut[0];
+				m_gpuCompoundPairs.copyFromHostPointer(ptr,numCompoundPairs,0,true);
+			}
+			//cpuCompoundPairsOut
+            
+        }
+		if (numCompoundPairs)
+		{
+			printf("numCompoundPairs=%d\n",numCompoundPairs);
+		}
+
+        if (numCompoundPairs > compoundPairCapacity)
+        {
+            b3Error("Exceeded compound pair capacity (%d/%d)\n", numCompoundPairs,  compoundPairCapacity);
+            numCompoundPairs = compoundPairCapacity;
+        }
+
+        
+
+        m_gpuCompoundPairs.resize(numCompoundPairs);
+        m_gpuHasCompoundSepNormals.resize(numCompoundPairs);
+        m_gpuCompoundSepNormals.resize(numCompoundPairs);
+        
+
+        if (numCompoundPairs)
+        {
+            B3_PROFILE("processCompoundPairsPrimitivesKernel");
+            b3BufferInfoCL bInfo[] = 
+            { 
+                b3BufferInfoCL( m_gpuCompoundPairs.getBufferCL(), true ), 
+                b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+                b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+                b3BufferInfoCL( convexData.getBufferCL(),true),
+                b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+                b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+                b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+                b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+                b3BufferInfoCL( clAabbsWorldSpace.getBufferCL(),true),
+                b3BufferInfoCL( gpuChildShapes.getBufferCL(),true),
+                b3BufferInfoCL( contactOut->getBufferCL()),
+                b3BufferInfoCL( m_totalContactsOut.getBufferCL())	
+            };
+
+            b3LauncherCL launcher(m_queue, m_processCompoundPairsPrimitivesKernel,"m_processCompoundPairsPrimitivesKernel");
+            launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+            launcher.setConst( numCompoundPairs  );
+            launcher.setConst(maxContactCapacity);
+
+            int num = numCompoundPairs;
+            launcher.launch1D( num);
+            clFinish(m_queue);
+            nContacts = m_totalContactsOut.at(0);
+            //printf("nContacts (after processCompoundPairsPrimitivesKernel) = %d\n",nContacts);
+            if (nContacts>maxContactCapacity)
+            {
+                
+                b3Error("Error: contacts exceeds capacity (%d/%d)\n", nContacts, maxContactCapacity);
+                nContacts = maxContactCapacity;
+            }
+        }
+        
+
+        if (numCompoundPairs)
+        {
+            B3_PROFILE("processCompoundPairsKernel");
+            b3BufferInfoCL bInfo[] = 
+            { 
+                b3BufferInfoCL( m_gpuCompoundPairs.getBufferCL(), true ), 
+                b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+                b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+                b3BufferInfoCL( convexData.getBufferCL(),true),
+                b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+                b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+                b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+                b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+                b3BufferInfoCL( clAabbsWorldSpace.getBufferCL(),true),
+                b3BufferInfoCL( gpuChildShapes.getBufferCL(),true),
+                b3BufferInfoCL( m_gpuCompoundSepNormals.getBufferCL()),
+                b3BufferInfoCL( m_gpuHasCompoundSepNormals.getBufferCL())
+            };
+
+            b3LauncherCL launcher(m_queue, m_processCompoundPairsKernel,"m_processCompoundPairsKernel");
+            launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+            launcher.setConst( numCompoundPairs  );
+
+            int num = numCompoundPairs;
+            launcher.launch1D( num);
+            clFinish(m_queue);
+        
+        }
+
+
+        //printf("numConcave  = %d\n",numConcave);
+
+    
+
+//		printf("hostNormals.size()=%d\n",hostNormals.size());
+		//int numPairs = pairCount.at(0);
+		
+		
+		
+	}
+	int vertexFaceCapacity = 64;
+
+
+		
+	{
+		//now perform the tree query on GPU
+			
+					
+				
+				
+		if (treeNodesGPU->size() && treeNodesGPU->size())
+		{
+			if (bvhTraversalKernelGPU)
+			{
+						
+				B3_PROFILE("m_bvhTraversalKernel");
+						
+						
+				numConcavePairs = m_numConcavePairsOut.at(0);
+						
+				b3LauncherCL launcher(m_queue, m_bvhTraversalKernel,"m_bvhTraversalKernel");
+				launcher.setBuffer( pairs->getBufferCL());
+				launcher.setBuffer(  bodyBuf->getBufferCL());
+				launcher.setBuffer( gpuCollidables.getBufferCL());
+				launcher.setBuffer( clAabbsWorldSpace.getBufferCL());
+				launcher.setBuffer( triangleConvexPairsOut.getBufferCL());
+				launcher.setBuffer( m_numConcavePairsOut.getBufferCL());
+				launcher.setBuffer( subTreesGPU->getBufferCL());
+				launcher.setBuffer( treeNodesGPU->getBufferCL());
+				launcher.setBuffer( bvhInfo->getBufferCL());
+						
+				launcher.setConst( nPairs  );
+				launcher.setConst( maxTriConvexPairCapacity);
+				int num = nPairs;
+				launcher.launch1D( num);
+				clFinish(m_queue);
+				numConcavePairs = m_numConcavePairsOut.at(0);
+			} else
+			{
+					b3AlignedObjectArray<b3Int4> hostPairs;
+					pairs->copyToHost(hostPairs);
+					b3AlignedObjectArray<b3RigidBodyData> hostBodyBuf;
+					bodyBuf->copyToHost(hostBodyBuf);
+					b3AlignedObjectArray<b3Collidable> hostCollidables;
+					gpuCollidables.copyToHost(hostCollidables);
+					b3AlignedObjectArray<b3Aabb> hostAabbsWorldSpace;
+					clAabbsWorldSpace.copyToHost(hostAabbsWorldSpace);
+
+					//int maxTriConvexPairCapacity,
+					b3AlignedObjectArray<b3Int4> triangleConvexPairsOutHost;
+					triangleConvexPairsOutHost.resize(maxTriConvexPairCapacity);
+
+					//int numTriConvexPairsOutHost=0;
+					numConcavePairs = 0;
+					//m_numConcavePairsOut
+
+					b3AlignedObjectArray<b3QuantizedBvhNode>	treeNodesCPU;
+					treeNodesGPU->copyToHost(treeNodesCPU);
+					b3AlignedObjectArray<b3BvhSubtreeInfo>	subTreesCPU;
+					subTreesGPU->copyToHost(subTreesCPU);
+					b3AlignedObjectArray<b3BvhInfo>	bvhInfoCPU;
+					bvhInfo->copyToHost(bvhInfoCPU);
+					//compute it...
+
+					volatile int hostNumConcavePairsOut=0;
+
+					//
+					for (int i=0;i<nPairs;i++)
+					{
+						b3BvhTraversal( &hostPairs.at(0), 
+						&hostBodyBuf.at(0),
+						&hostCollidables.at(0),
+						&hostAabbsWorldSpace.at(0),
+						&triangleConvexPairsOutHost.at(0),
+						&hostNumConcavePairsOut,
+						&subTreesCPU.at(0),
+						&treeNodesCPU.at(0),
+						&bvhInfoCPU.at(0),
+						nPairs,
+						maxTriConvexPairCapacity,
+						i);
+					}
+					numConcavePairs = hostNumConcavePairsOut;
+
+					if (hostNumConcavePairsOut)
+					{
+						triangleConvexPairsOutHost.resize(hostNumConcavePairsOut);
+						triangleConvexPairsOut.copyFromHost(triangleConvexPairsOutHost);
+					}
+					//
+
+					m_numConcavePairsOut.resize(0);
+					m_numConcavePairsOut.push_back(numConcavePairs);
+			}
+
+				//printf("numConcavePairs=%d (max = %d\n",numConcavePairs,maxTriConvexPairCapacity);
+						
+			if (numConcavePairs > maxTriConvexPairCapacity)
+			{
+				static int exceeded_maxTriConvexPairCapacity_count = 0;
+				b3Error("Exceeded the maxTriConvexPairCapacity (found %d but max is %d, it happened %d times)\n",
+					numConcavePairs,maxTriConvexPairCapacity,exceeded_maxTriConvexPairCapacity_count++);
+				numConcavePairs = maxTriConvexPairCapacity;
+			}
+			triangleConvexPairsOut.resize(numConcavePairs);
+	
+			if (numConcavePairs)
+			{
+
+				
+						
+	
+				clippingFacesOutGPU.resize(numConcavePairs);
+				worldNormalsAGPU.resize(numConcavePairs);
+				worldVertsA1GPU.resize(vertexFaceCapacity*(numConcavePairs));
+				worldVertsB1GPU.resize(vertexFaceCapacity*(numConcavePairs));
+
+
+				if (findConcaveSeparatingAxisKernelGPU)
+				{
+
+					/*
+					m_concaveHasSeparatingNormals.copyFromHost(concaveHasSeparatingNormalsCPU);
+						clippingFacesOutGPU.copyFromHost(clippingFacesOutCPU);
+						worldVertsA1GPU.copyFromHost(worldVertsA1CPU);
+						worldNormalsAGPU.copyFromHost(worldNormalsACPU);
+						worldVertsB1GPU.copyFromHost(worldVertsB1CPU);
+					*/
+
+					//now perform a SAT test for each triangle-convex element (stored in triangleConvexPairsOut)
+                    if (splitSearchSepAxisConcave)
+                    {
+                        //printf("numConcavePairs = %d\n",numConcavePairs);
+                        m_dmins.resize(numConcavePairs);
+                        {
+                            B3_PROFILE("findConcaveSeparatingAxisVertexFaceKernel");
+                            b3BufferInfoCL bInfo[] = {
+                                b3BufferInfoCL( triangleConvexPairsOut.getBufferCL() ),
+                                b3BufferInfoCL( bodyBuf->getBufferCL(),true),
+                                b3BufferInfoCL( gpuCollidables.getBufferCL(),true),
+                                b3BufferInfoCL( convexData.getBufferCL(),true),
+                                b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+                                b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+                                b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+                                b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+                                b3BufferInfoCL( gpuChildShapes.getBufferCL(),true),
+                                b3BufferInfoCL( clAabbsWorldSpace.getBufferCL(),true),
+                                b3BufferInfoCL( m_concaveSepNormals.getBufferCL()),
+                                b3BufferInfoCL( m_concaveHasSeparatingNormals.getBufferCL()),
+                                b3BufferInfoCL( clippingFacesOutGPU.getBufferCL()),
+                                b3BufferInfoCL( worldVertsA1GPU.getBufferCL()),
+                                b3BufferInfoCL(worldNormalsAGPU.getBufferCL()),
+                                b3BufferInfoCL(worldVertsB1GPU.getBufferCL()),
+                                b3BufferInfoCL(m_dmins.getBufferCL())
+                            };
+                            
+                            b3LauncherCL launcher(m_queue, m_findConcaveSeparatingAxisVertexFaceKernel,"m_findConcaveSeparatingAxisVertexFaceKernel");
+                            launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+                            launcher.setConst(vertexFaceCapacity);
+                            launcher.setConst( numConcavePairs  );
+                            
+                            int num = numConcavePairs;
+                            launcher.launch1D( num);
+                            clFinish(m_queue);
+
+                            
+                        }
+//                        numConcavePairs = 0;
+                        if (1)
+                        {
+                            B3_PROFILE("findConcaveSeparatingAxisEdgeEdgeKernel");
+                            b3BufferInfoCL bInfo[] = {
+                                b3BufferInfoCL( triangleConvexPairsOut.getBufferCL() ),
+                                b3BufferInfoCL( bodyBuf->getBufferCL(),true),
+                                b3BufferInfoCL( gpuCollidables.getBufferCL(),true),
+                                b3BufferInfoCL( convexData.getBufferCL(),true),
+                                b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+                                b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+                                b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+                                b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+                                b3BufferInfoCL( gpuChildShapes.getBufferCL(),true),
+                                b3BufferInfoCL( clAabbsWorldSpace.getBufferCL(),true),
+                                b3BufferInfoCL( m_concaveSepNormals.getBufferCL()),
+                                b3BufferInfoCL( m_concaveHasSeparatingNormals.getBufferCL()),
+                                b3BufferInfoCL( clippingFacesOutGPU.getBufferCL()),
+                                b3BufferInfoCL( worldVertsA1GPU.getBufferCL()),
+                                b3BufferInfoCL(worldNormalsAGPU.getBufferCL()),
+                                b3BufferInfoCL(worldVertsB1GPU.getBufferCL()),
+                                b3BufferInfoCL(m_dmins.getBufferCL())
+                            };
+                            
+                            b3LauncherCL launcher(m_queue, m_findConcaveSeparatingAxisEdgeEdgeKernel,"m_findConcaveSeparatingAxisEdgeEdgeKernel");
+                            launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+                            launcher.setConst(vertexFaceCapacity);
+                            launcher.setConst( numConcavePairs  );
+                            
+                            int num = numConcavePairs;
+                            launcher.launch1D( num);
+                            clFinish(m_queue);
+                        }
+                      
+                        
+                        // numConcavePairs = 0;
+                        
+                        
+                        
+                        
+                        
+                        
+                    } else
+                    {
+                        B3_PROFILE("findConcaveSeparatingAxisKernel");
+                        b3BufferInfoCL bInfo[] = { 
+                            b3BufferInfoCL( triangleConvexPairsOut.getBufferCL() ), 
+                            b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+                            b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+                            b3BufferInfoCL( convexData.getBufferCL(),true),
+                            b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+                            b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+                            b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+                            b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+                            b3BufferInfoCL( gpuChildShapes.getBufferCL(),true),
+                            b3BufferInfoCL( clAabbsWorldSpace.getBufferCL(),true),
+                            b3BufferInfoCL( m_concaveSepNormals.getBufferCL()),
+                            b3BufferInfoCL( m_concaveHasSeparatingNormals.getBufferCL()),
+                            b3BufferInfoCL( clippingFacesOutGPU.getBufferCL()),
+                            b3BufferInfoCL( worldVertsA1GPU.getBufferCL()),
+                            b3BufferInfoCL(worldNormalsAGPU.getBufferCL()),
+                            b3BufferInfoCL(worldVertsB1GPU.getBufferCL())
+                        };
+
+                        b3LauncherCL launcher(m_queue, m_findConcaveSeparatingAxisKernel,"m_findConcaveSeparatingAxisKernel");
+                        launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+                        launcher.setConst(vertexFaceCapacity);
+                        launcher.setConst( numConcavePairs  );
+
+                        int num = numConcavePairs;
+                        launcher.launch1D( num);
+                        clFinish(m_queue);
+                    }
+                    
+                    
+				} else
+				{
+
+						b3AlignedObjectArray<b3Int4> clippingFacesOutCPU;
+						b3AlignedObjectArray<b3Vector3> worldVertsA1CPU;
+						b3AlignedObjectArray<b3Vector3> worldNormalsACPU;
+						b3AlignedObjectArray<b3Vector3> worldVertsB1CPU;
+						b3AlignedObjectArray<int>concaveHasSeparatingNormalsCPU;
+
+						b3AlignedObjectArray<b3Int4> triangleConvexPairsOutHost;
+						triangleConvexPairsOut.copyToHost(triangleConvexPairsOutHost);
+						//triangleConvexPairsOutHost.resize(maxTriConvexPairCapacity);
+						b3AlignedObjectArray<b3RigidBodyData> hostBodyBuf;
+						bodyBuf->copyToHost(hostBodyBuf);
+						b3AlignedObjectArray<b3Collidable> hostCollidables;
+						gpuCollidables.copyToHost(hostCollidables);
+						b3AlignedObjectArray<b3Aabb> hostAabbsWorldSpace;
+						clAabbsWorldSpace.copyToHost(hostAabbsWorldSpace);
+
+						b3AlignedObjectArray<b3ConvexPolyhedronData> hostConvexData;
+						convexData.copyToHost(hostConvexData);
+
+						b3AlignedObjectArray<b3Vector3> hostVertices;
+						gpuVertices.copyToHost(hostVertices);
+
+						b3AlignedObjectArray<b3Vector3> hostUniqueEdges;
+						gpuUniqueEdges.copyToHost(hostUniqueEdges);
+						b3AlignedObjectArray<b3GpuFace> hostFaces;
+						gpuFaces.copyToHost(hostFaces);
+						b3AlignedObjectArray<int> hostIndices;
+						gpuIndices.copyToHost(hostIndices);
+						b3AlignedObjectArray<b3GpuChildShape> cpuChildShapes;
+						gpuChildShapes.copyToHost(cpuChildShapes);
+
+
+								
+						b3AlignedObjectArray<b3Vector3> concaveSepNormalsHost;
+						m_concaveSepNormals.copyToHost(concaveSepNormalsHost);
+						concaveHasSeparatingNormalsCPU.resize(concaveSepNormalsHost.size());
+
+						b3GpuChildShape* childShapePointerCPU = 0;
+						if (cpuChildShapes.size())
+							childShapePointerCPU  = &cpuChildShapes.at(0);
+
+						clippingFacesOutCPU.resize(clippingFacesOutGPU.size());
+						worldVertsA1CPU.resize(worldVertsA1GPU.size());
+    					worldNormalsACPU.resize(worldNormalsAGPU.size());
+						worldVertsB1CPU.resize(worldVertsB1GPU.size());
+
+						for (int i=0;i<numConcavePairs;i++)
+						{
+							b3FindConcaveSeparatingAxisKernel(&triangleConvexPairsOutHost.at(0),
+								&hostBodyBuf.at(0),
+								&hostCollidables.at(0),
+								&hostConvexData.at(0), &hostVertices.at(0),&hostUniqueEdges.at(0),
+								&hostFaces.at(0),&hostIndices.at(0),childShapePointerCPU,
+								&hostAabbsWorldSpace.at(0),
+								&concaveSepNormalsHost.at(0),
+								&clippingFacesOutCPU.at(0),
+								&worldVertsA1CPU.at(0),
+								&worldNormalsACPU.at(0),
+								&worldVertsB1CPU.at(0),
+								&concaveHasSeparatingNormalsCPU.at(0),
+								vertexFaceCapacity,
+								numConcavePairs,i);
+						};
+
+						m_concaveSepNormals.copyFromHost(concaveSepNormalsHost);
+						m_concaveHasSeparatingNormals.copyFromHost(concaveHasSeparatingNormalsCPU);
+						clippingFacesOutGPU.copyFromHost(clippingFacesOutCPU);
+						worldVertsA1GPU.copyFromHost(worldVertsA1CPU);
+						worldNormalsAGPU.copyFromHost(worldNormalsACPU);
+						worldVertsB1GPU.copyFromHost(worldVertsB1CPU);
+
+
+
+				}
+//							b3AlignedObjectArray<b3Vector3> cpuCompoundSepNormals;
+//						m_concaveSepNormals.copyToHost(cpuCompoundSepNormals);
+//					b3AlignedObjectArray<b3Int4> cpuConcavePairs;
+//				triangleConvexPairsOut.copyToHost(cpuConcavePairs);
+
+
+			}
+		}
+		
+		
+	}
+
+	if (numConcavePairs)
+	{
+			if (numConcavePairs)
+		{
+			B3_PROFILE("findConcaveSphereContactsKernel");
+				nContacts = m_totalContactsOut.at(0);
+//				printf("nContacts1 = %d\n",nContacts);
+			b3BufferInfoCL bInfo[] = { 
+				b3BufferInfoCL( triangleConvexPairsOut.getBufferCL() ), 
+				b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+				b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+				b3BufferInfoCL( convexData.getBufferCL(),true),
+				b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+				b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+				b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+				b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+				b3BufferInfoCL( clAabbsWorldSpace.getBufferCL(),true),
+				b3BufferInfoCL( contactOut->getBufferCL()),
+				b3BufferInfoCL( m_totalContactsOut.getBufferCL())
+			};
+
+			b3LauncherCL launcher(m_queue, m_findConcaveSphereContactsKernel,"m_findConcaveSphereContactsKernel");
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+
+			launcher.setConst( numConcavePairs  );
+			launcher.setConst(maxContactCapacity);
+
+			int num = numConcavePairs;
+			launcher.launch1D( num);
+			clFinish(m_queue);
+			nContacts = m_totalContactsOut.at(0);
+			//printf("nContacts (after findConcaveSphereContactsKernel) = %d\n",nContacts);
+
+			//printf("nContacts2 = %d\n",nContacts);
+
+			if (nContacts >= maxContactCapacity)
+			{
+				b3Error("Error: contacts exceeds capacity (%d/%d)\n", nContacts, maxContactCapacity);
+				nContacts = maxContactCapacity;
+			}
+		}
+		
+	}
+
+
+
+#ifdef __APPLE__
+	bool contactClippingOnGpu = true;
+#else
+	bool contactClippingOnGpu = true;
+#endif
+
+	if (contactClippingOnGpu)
+	{
+		m_totalContactsOut.copyFromHostPointer(&nContacts,1,0,true);
+//		printf("nContacts3 = %d\n",nContacts);
+
+
+		//B3_PROFILE("clipHullHullKernel");
+
+		bool breakupConcaveConvexKernel = true;
+
+#ifdef __APPLE__
+		//actually, some Apple OpenCL platform/device combinations work fine...
+		breakupConcaveConvexKernel = true;
+#endif
+		//concave-convex contact clipping
+		if (numConcavePairs)
+		{
+			//			printf("numConcavePairs = %d\n", numConcavePairs);
+			//		nContacts = m_totalContactsOut.at(0);
+			//	printf("nContacts before = %d\n", nContacts);
+
+			if (breakupConcaveConvexKernel)
+			{
+
+				worldVertsB2GPU.resize(vertexFaceCapacity*numConcavePairs);
+
+
+				//clipFacesAndFindContacts
+
+				if (clipConcaveFacesAndFindContactsCPU)
+				{
+
+					b3AlignedObjectArray<b3Int4> clippingFacesOutCPU;
+					b3AlignedObjectArray<b3Vector3> worldVertsA1CPU;
+					b3AlignedObjectArray<b3Vector3> worldNormalsACPU;
+					b3AlignedObjectArray<b3Vector3> worldVertsB1CPU;
+
+					clippingFacesOutGPU.copyToHost(clippingFacesOutCPU);
+					worldVertsA1GPU.copyToHost(worldVertsA1CPU);
+					worldNormalsAGPU.copyToHost(worldNormalsACPU);
+					worldVertsB1GPU.copyToHost(worldVertsB1CPU);
+
+
+
+					b3AlignedObjectArray<int>concaveHasSeparatingNormalsCPU;
+					m_concaveHasSeparatingNormals.copyToHost(concaveHasSeparatingNormalsCPU);
+
+					b3AlignedObjectArray<b3Vector3> concaveSepNormalsHost;
+					m_concaveSepNormals.copyToHost(concaveSepNormalsHost);
+
+					b3AlignedObjectArray<b3Vector3> worldVertsB2CPU;  
+					worldVertsB2CPU.resize(worldVertsB2GPU.size());
+
+
+					for (int i=0;i<numConcavePairs;i++)
+					{
+
+						clipFacesAndFindContactsKernel(   &concaveSepNormalsHost.at(0),
+							&concaveHasSeparatingNormalsCPU.at(0),
+							&clippingFacesOutCPU.at(0),
+							&worldVertsA1CPU.at(0),
+							&worldNormalsACPU.at(0),
+							&worldVertsB1CPU.at(0),
+							&worldVertsB2CPU.at(0),
+							vertexFaceCapacity,
+							i);
+					}
+
+					clippingFacesOutGPU.copyFromHost(clippingFacesOutCPU);
+					worldVertsB2GPU.copyFromHost(worldVertsB2CPU);
+
+
+				} else
+				{
+
+					if (1)
+					{
+
+
+
+						B3_PROFILE("clipFacesAndFindContacts");
+						//nContacts = m_totalContactsOut.at(0);
+						//int h = m_hasSeparatingNormals.at(0);
+						//int4 p = clippingFacesOutGPU.at(0);
+						b3BufferInfoCL bInfo[] = {
+							b3BufferInfoCL( m_concaveSepNormals.getBufferCL()),
+							b3BufferInfoCL( m_concaveHasSeparatingNormals.getBufferCL()),
+							b3BufferInfoCL( clippingFacesOutGPU.getBufferCL()),
+							b3BufferInfoCL( worldVertsA1GPU.getBufferCL()),
+							b3BufferInfoCL( worldNormalsAGPU.getBufferCL()),
+							b3BufferInfoCL( worldVertsB1GPU.getBufferCL()),
+							b3BufferInfoCL( worldVertsB2GPU.getBufferCL())
+						};
+						b3LauncherCL launcher(m_queue, m_clipFacesAndFindContacts,"m_clipFacesAndFindContacts");
+						launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+						launcher.setConst(vertexFaceCapacity);
+
+						launcher.setConst( numConcavePairs  );
+						int debugMode = 0;
+						launcher.setConst( debugMode);
+						int num = numConcavePairs;
+						launcher.launch1D( num);
+						clFinish(m_queue);
+						//int bla = m_totalContactsOut.at(0);
+					}
+				}
+				//contactReduction
+				{
+					int newContactCapacity=nContacts+numConcavePairs; 
+					contactOut->reserve(newContactCapacity);
+					if (reduceConcaveContactsOnGPU)
+					{
+//						printf("newReservation = %d\n",newReservation);
+						{
+							B3_PROFILE("newContactReductionKernel");
+							b3BufferInfoCL bInfo[] =
+							{
+								b3BufferInfoCL( triangleConvexPairsOut.getBufferCL(), true ),
+								b3BufferInfoCL( bodyBuf->getBufferCL(),true),
+								b3BufferInfoCL( m_concaveSepNormals.getBufferCL()),
+								b3BufferInfoCL( m_concaveHasSeparatingNormals.getBufferCL()),
+								b3BufferInfoCL( contactOut->getBufferCL()),
+								b3BufferInfoCL( clippingFacesOutGPU.getBufferCL()),
+								b3BufferInfoCL( worldVertsB2GPU.getBufferCL()),
+								b3BufferInfoCL( m_totalContactsOut.getBufferCL())
+							};
+
+							b3LauncherCL launcher(m_queue, m_newContactReductionKernel,"m_newContactReductionKernel");
+							launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+							launcher.setConst(vertexFaceCapacity);
+							launcher.setConst(newContactCapacity);
+							launcher.setConst( numConcavePairs  );
+							int num = numConcavePairs;
+
+							launcher.launch1D( num);
+						}
+						nContacts = m_totalContactsOut.at(0);
+						contactOut->resize(nContacts);
+
+						//printf("contactOut4 (after newContactReductionKernel) = %d\n",nContacts);
+					}else
+					{
+						
+						volatile int nGlobalContactsOut = nContacts;
+						b3AlignedObjectArray<b3Int4> triangleConvexPairsOutHost;
+						triangleConvexPairsOut.copyToHost(triangleConvexPairsOutHost);
+						b3AlignedObjectArray<b3RigidBodyData> hostBodyBuf;
+						bodyBuf->copyToHost(hostBodyBuf);
+
+						b3AlignedObjectArray<int>concaveHasSeparatingNormalsCPU;
+						m_concaveHasSeparatingNormals.copyToHost(concaveHasSeparatingNormalsCPU);
+
+						b3AlignedObjectArray<b3Vector3> concaveSepNormalsHost;
+						m_concaveSepNormals.copyToHost(concaveSepNormalsHost);
+
+
+						b3AlignedObjectArray<b3Contact4> hostContacts;
+						if (nContacts)
+						{
+							contactOut->copyToHost(hostContacts);
+						}
+						hostContacts.resize(newContactCapacity);
+
+						b3AlignedObjectArray<b3Int4> clippingFacesOutCPU;
+						b3AlignedObjectArray<b3Vector3> worldVertsB2CPU;
+
+						clippingFacesOutGPU.copyToHost(clippingFacesOutCPU);
+						worldVertsB2GPU.copyToHost(worldVertsB2CPU);
+
+
+
+						for (int i=0;i<numConcavePairs;i++)
+						{
+							b3NewContactReductionKernel( &triangleConvexPairsOutHost.at(0),
+                                                   &hostBodyBuf.at(0),
+												   &concaveSepNormalsHost.at(0),
+												   &concaveHasSeparatingNormalsCPU.at(0),
+												   &hostContacts.at(0),
+                                                   &clippingFacesOutCPU.at(0),
+                                                   &worldVertsB2CPU.at(0),
+                                                   &nGlobalContactsOut,
+                                                   vertexFaceCapacity,
+												   newContactCapacity,
+                                                   numConcavePairs,
+												   i
+                                                   );
+
+						}
+
+
+						nContacts = nGlobalContactsOut;
+						m_totalContactsOut.copyFromHostPointer(&nContacts,1,0,true);
+//						nContacts = m_totalContactsOut.at(0);
+						//contactOut->resize(nContacts);
+						hostContacts.resize(nContacts);
+						//printf("contactOut4 (after newContactReductionKernel) = %d\n",nContacts);
+						contactOut->copyFromHost(hostContacts);
+					}
+
+				}
+				//re-use?
+
+
+			} else
+			{
+				B3_PROFILE("clipHullHullConcaveConvexKernel");
+				nContacts = m_totalContactsOut.at(0);
+				int newContactCapacity = contactOut->capacity();
+
+				//printf("contactOut5 = %d\n",nContacts);
+				b3BufferInfoCL bInfo[] = { 
+					b3BufferInfoCL( triangleConvexPairsOut.getBufferCL(), true ), 
+					b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+					b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+					b3BufferInfoCL( convexData.getBufferCL(),true),
+					b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+					b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+					b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+					b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+					b3BufferInfoCL( gpuChildShapes.getBufferCL(),true),
+					b3BufferInfoCL( m_concaveSepNormals.getBufferCL()),
+					b3BufferInfoCL( contactOut->getBufferCL()),
+					b3BufferInfoCL( m_totalContactsOut.getBufferCL())	
+				};
+				b3LauncherCL launcher(m_queue, m_clipHullHullConcaveConvexKernel,"m_clipHullHullConcaveConvexKernel");
+				launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+				launcher.setConst(newContactCapacity);
+				launcher.setConst( numConcavePairs  );
+				int num = numConcavePairs;
+				launcher.launch1D( num);
+				clFinish(m_queue);
+				nContacts = m_totalContactsOut.at(0);
+				contactOut->resize(nContacts);
+				//printf("contactOut6 = %d\n",nContacts);
+				b3AlignedObjectArray<b3Contact4> cpuContacts;
+				contactOut->copyToHost(cpuContacts);
+			}
+			//			printf("nContacts after = %d\n", nContacts);
+		}//numConcavePairs
+
+
+
+		//convex-convex contact clipping
+		
+		bool breakupKernel = false;
+
+#ifdef __APPLE__
+		breakupKernel = true;
+#endif
+
+#ifdef CHECK_ON_HOST
+	bool computeConvexConvex = false;
+#else
+	bool computeConvexConvex = true;
+#endif//CHECK_ON_HOST
+		if (computeConvexConvex)
+		{
+			B3_PROFILE("clipHullHullKernel");
+		if (breakupKernel)
+		{
+
+
+
+
+			worldVertsB1GPU.resize(vertexFaceCapacity*nPairs);
+			clippingFacesOutGPU.resize(nPairs);
+			worldNormalsAGPU.resize(nPairs);
+			worldVertsA1GPU.resize(vertexFaceCapacity*nPairs);
+			worldVertsB2GPU.resize(vertexFaceCapacity*nPairs);
+
+			if (findConvexClippingFacesGPU)
+			{
+				B3_PROFILE("findClippingFacesKernel");
+				b3BufferInfoCL bInfo[] = {
+					b3BufferInfoCL( pairs->getBufferCL(), true ),
+					b3BufferInfoCL( bodyBuf->getBufferCL(),true),
+					b3BufferInfoCL( gpuCollidables.getBufferCL(),true),
+					b3BufferInfoCL( convexData.getBufferCL(),true),
+					b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+					b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+					b3BufferInfoCL( gpuFaces.getBufferCL(),true), 
+					b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+					b3BufferInfoCL( m_sepNormals.getBufferCL()),
+					b3BufferInfoCL( m_hasSeparatingNormals.getBufferCL()),
+					b3BufferInfoCL( clippingFacesOutGPU.getBufferCL()),
+					b3BufferInfoCL( worldVertsA1GPU.getBufferCL()),
+					b3BufferInfoCL( worldNormalsAGPU.getBufferCL()),
+					b3BufferInfoCL( worldVertsB1GPU.getBufferCL())
+				};
+
+				b3LauncherCL launcher(m_queue, m_findClippingFacesKernel,"m_findClippingFacesKernel");
+				launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+				launcher.setConst( vertexFaceCapacity);
+				launcher.setConst( nPairs  );
+				int num = nPairs;
+				launcher.launch1D( num);
+				clFinish(m_queue);
+
+			} else
+			{
+				
+				float minDist = -1e30f;
+				float maxDist = 0.02f;
+
+				b3AlignedObjectArray<b3ConvexPolyhedronData> hostConvexData;
+				convexData.copyToHost(hostConvexData);
+				b3AlignedObjectArray<b3Collidable> hostCollidables;
+				gpuCollidables.copyToHost(hostCollidables);
+
+				b3AlignedObjectArray<int> hostHasSepNormals;
+				m_hasSeparatingNormals.copyToHost(hostHasSepNormals);
+				b3AlignedObjectArray<b3Vector3> cpuSepNormals;
+				m_sepNormals.copyToHost(cpuSepNormals);
+
+				b3AlignedObjectArray<b3Int4> hostPairs;
+				pairs->copyToHost(hostPairs);
+				b3AlignedObjectArray<b3RigidBodyData> hostBodyBuf;
+				bodyBuf->copyToHost(hostBodyBuf);
+
+
+				//worldVertsB1GPU.resize(vertexFaceCapacity*nPairs);
+				b3AlignedObjectArray<b3Vector3> worldVertsB1CPU;
+				worldVertsB1GPU.copyToHost(worldVertsB1CPU);
+
+				b3AlignedObjectArray<b3Int4> clippingFacesOutCPU;
+				clippingFacesOutGPU.copyToHost(clippingFacesOutCPU);
+
+				b3AlignedObjectArray<b3Vector3> worldNormalsACPU;
+				worldNormalsACPU.resize(nPairs);
+
+				b3AlignedObjectArray<b3Vector3> worldVertsA1CPU;
+				worldVertsA1CPU.resize(worldVertsA1GPU.size());
+			
+			
+				b3AlignedObjectArray<b3Vector3> hostVertices;
+				gpuVertices.copyToHost(hostVertices);
+				b3AlignedObjectArray<b3GpuFace> hostFaces;
+				gpuFaces.copyToHost(hostFaces);
+				b3AlignedObjectArray<int> hostIndices;
+				gpuIndices.copyToHost(hostIndices);
+				
+
+				for (int i=0;i<nPairs;i++)
+				{
+
+					int bodyIndexA = hostPairs[i].x;
+					int bodyIndexB = hostPairs[i].y;
+			
+					int collidableIndexA = hostBodyBuf[bodyIndexA].m_collidableIdx;
+					int collidableIndexB = hostBodyBuf[bodyIndexB].m_collidableIdx;
+			
+					int shapeIndexA = hostCollidables[collidableIndexA].m_shapeIndex;
+					int shapeIndexB = hostCollidables[collidableIndexB].m_shapeIndex;
+			
+
+					if (hostHasSepNormals[i])
+					{
+						b3FindClippingFaces(cpuSepNormals[i],
+							&hostConvexData[shapeIndexA],
+							&hostConvexData[shapeIndexB],
+							hostBodyBuf[bodyIndexA].m_pos,hostBodyBuf[bodyIndexA].m_quat,
+							hostBodyBuf[bodyIndexB].m_pos,hostBodyBuf[bodyIndexB].m_quat,
+							&worldVertsA1CPU.at(0),&worldNormalsACPU.at(0),
+							&worldVertsB1CPU.at(0),
+							vertexFaceCapacity,minDist,maxDist,
+							&hostVertices.at(0),&hostFaces.at(0),
+							&hostIndices.at(0),
+							&hostVertices.at(0),&hostFaces.at(0),
+							&hostIndices.at(0),&clippingFacesOutCPU.at(0),i);
+					}
+				}
+
+				clippingFacesOutGPU.copyFromHost(clippingFacesOutCPU);
+				worldVertsA1GPU.copyFromHost(worldVertsA1CPU);
+				worldNormalsAGPU.copyFromHost(worldNormalsACPU);
+				worldVertsB1GPU.copyFromHost(worldVertsB1CPU);
+
+			}
+
+
+
+
+
+			///clip face B against face A, reduce contacts and append them to a global contact array
+			if (1)
+			{
+				if (clipConvexFacesAndFindContactsCPU)
+				{
+
+					//b3AlignedObjectArray<b3Int4> hostPairs;
+					//pairs->copyToHost(hostPairs);
+
+					b3AlignedObjectArray<b3Vector3> hostSepNormals;
+					m_sepNormals.copyToHost(hostSepNormals);
+					b3AlignedObjectArray<int> hostHasSepAxis;
+					m_hasSeparatingNormals.copyToHost(hostHasSepAxis);
+
+					b3AlignedObjectArray<b3Int4> hostClippingFaces;
+					clippingFacesOutGPU.copyToHost(hostClippingFaces);
+					b3AlignedObjectArray<b3Vector3> worldVertsB2CPU;
+					worldVertsB2CPU.resize(vertexFaceCapacity*nPairs);
+					
+					b3AlignedObjectArray<b3Vector3>worldVertsA1CPU;
+					worldVertsA1GPU.copyToHost(worldVertsA1CPU);
+					b3AlignedObjectArray<b3Vector3> worldNormalsACPU;
+					worldNormalsAGPU.copyToHost(worldNormalsACPU);
+
+					b3AlignedObjectArray<b3Vector3>  worldVertsB1CPU;
+					worldVertsB1GPU.copyToHost(worldVertsB1CPU);
+
+					/*
+					  __global const b3Float4* separatingNormals,
+                                                   __global const int* hasSeparatingAxis,
+                                                   __global b3Int4* clippingFacesOut,
+                                                   __global b3Float4* worldVertsA1,
+                                                   __global b3Float4* worldNormalsA1,
+                                                   __global b3Float4* worldVertsB1,
+                                                   __global b3Float4* worldVertsB2,
+                                                    int vertexFaceCapacity,
+															int pairIndex
+					*/
+					for (int i=0;i<nPairs;i++)
+					{
+						clipFacesAndFindContactsKernel(
+							&hostSepNormals.at(0),
+							&hostHasSepAxis.at(0),
+							&hostClippingFaces.at(0),
+							&worldVertsA1CPU.at(0),
+							&worldNormalsACPU.at(0),
+							&worldVertsB1CPU.at(0),
+							&worldVertsB2CPU.at(0),
+
+						vertexFaceCapacity,
+							i);
+					}
+					
+					clippingFacesOutGPU.copyFromHost(hostClippingFaces);
+					worldVertsB2GPU.copyFromHost(worldVertsB2CPU);
+
+				} else
+				{
+					B3_PROFILE("clipFacesAndFindContacts");
+					//nContacts = m_totalContactsOut.at(0);
+					//int h = m_hasSeparatingNormals.at(0);
+					//int4 p = clippingFacesOutGPU.at(0);
+					b3BufferInfoCL bInfo[] = {
+						b3BufferInfoCL( m_sepNormals.getBufferCL()),
+						b3BufferInfoCL( m_hasSeparatingNormals.getBufferCL()),
+						b3BufferInfoCL( clippingFacesOutGPU.getBufferCL()),
+						b3BufferInfoCL( worldVertsA1GPU.getBufferCL()),
+						b3BufferInfoCL( worldNormalsAGPU.getBufferCL()),
+						b3BufferInfoCL( worldVertsB1GPU.getBufferCL()),
+						b3BufferInfoCL( worldVertsB2GPU.getBufferCL())
+					};
+
+					b3LauncherCL launcher(m_queue, m_clipFacesAndFindContacts,"m_clipFacesAndFindContacts");
+					launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+					launcher.setConst(vertexFaceCapacity);
+
+					launcher.setConst( nPairs  );
+					int debugMode = 0;
+					launcher.setConst( debugMode);
+					int num = nPairs;
+					launcher.launch1D( num);
+					clFinish(m_queue);
+				} 
+
+				{
+					nContacts = m_totalContactsOut.at(0);
+					//printf("nContacts = %d\n",nContacts);
+
+					int newContactCapacity = nContacts+nPairs;
+					contactOut->reserve(newContactCapacity);
+
+					if (reduceConvexContactsOnGPU)
+					{
+						{
+							B3_PROFILE("newContactReductionKernel");
+							b3BufferInfoCL bInfo[] =
+							{
+								b3BufferInfoCL( pairs->getBufferCL(), true ),
+								b3BufferInfoCL( bodyBuf->getBufferCL(),true),
+								b3BufferInfoCL( m_sepNormals.getBufferCL()),
+								b3BufferInfoCL( m_hasSeparatingNormals.getBufferCL()),
+								b3BufferInfoCL( contactOut->getBufferCL()),
+								b3BufferInfoCL( clippingFacesOutGPU.getBufferCL()),
+								b3BufferInfoCL( worldVertsB2GPU.getBufferCL()),
+								b3BufferInfoCL( m_totalContactsOut.getBufferCL())
+							};
+
+							b3LauncherCL launcher(m_queue, m_newContactReductionKernel,"m_newContactReductionKernel");
+							launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+							launcher.setConst(vertexFaceCapacity);
+							launcher.setConst(newContactCapacity);
+							launcher.setConst( nPairs  );
+							int num = nPairs;
+
+							launcher.launch1D( num);
+						}
+						nContacts = m_totalContactsOut.at(0);
+						contactOut->resize(nContacts);
+					} else
+					{
+
+						volatile int nGlobalContactsOut = nContacts;
+						b3AlignedObjectArray<b3Int4> hostPairs;
+						pairs->copyToHost(hostPairs);
+						b3AlignedObjectArray<b3RigidBodyData> hostBodyBuf;
+						bodyBuf->copyToHost(hostBodyBuf);
+						b3AlignedObjectArray<b3Vector3> hostSepNormals;
+						m_sepNormals.copyToHost(hostSepNormals);
+						b3AlignedObjectArray<int> hostHasSepAxis;
+						m_hasSeparatingNormals.copyToHost(hostHasSepAxis);
+						b3AlignedObjectArray<b3Contact4> hostContactsOut;
+						contactOut->copyToHost(hostContactsOut);
+						hostContactsOut.resize(newContactCapacity);
+
+						b3AlignedObjectArray<b3Int4> hostClippingFaces;
+						clippingFacesOutGPU.copyToHost(hostClippingFaces);
+						b3AlignedObjectArray<b3Vector3> worldVertsB2CPU;
+						worldVertsB2GPU.copyToHost(worldVertsB2CPU);
+
+						for (int i=0;i<nPairs;i++)
+						{
+							b3NewContactReductionKernel(&hostPairs.at(0),
+								&hostBodyBuf.at(0),
+								&hostSepNormals.at(0),
+								&hostHasSepAxis.at(0),
+								&hostContactsOut.at(0),
+								&hostClippingFaces.at(0),
+								&worldVertsB2CPU.at(0),
+								&nGlobalContactsOut,
+								vertexFaceCapacity,
+								newContactCapacity,
+								nPairs,
+								i);
+						}
+
+						nContacts = nGlobalContactsOut;
+						m_totalContactsOut.copyFromHostPointer(&nContacts,1,0,true);
+						hostContactsOut.resize(nContacts);
+						//printf("contactOut4 (after newContactReductionKernel) = %d\n",nContacts);
+						contactOut->copyFromHost(hostContactsOut);
+					}
+					//                    b3Contact4 pt = contactOut->at(0);
+					//                  printf("nContacts = %d\n",nContacts);
+				}
+			}
+		}            
+		else//breakupKernel
+		{
+
+			if (nPairs)
+			{
+				b3BufferInfoCL bInfo[] = {
+					b3BufferInfoCL( pairs->getBufferCL(), true ), 
+					b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+					b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+					b3BufferInfoCL( convexData.getBufferCL(),true),
+					b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+					b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+					b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+					b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+					b3BufferInfoCL( m_sepNormals.getBufferCL()),
+					b3BufferInfoCL( m_hasSeparatingNormals.getBufferCL()),
+					b3BufferInfoCL( contactOut->getBufferCL()),
+					b3BufferInfoCL( m_totalContactsOut.getBufferCL())	
+				};
+				b3LauncherCL launcher(m_queue, m_clipHullHullKernel,"m_clipHullHullKernel");
+				launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+				launcher.setConst( nPairs  );
+				launcher.setConst(maxContactCapacity);
+
+				int num = nPairs;
+				launcher.launch1D( num);
+				clFinish(m_queue);
+
+				nContacts = m_totalContactsOut.at(0);
+				if (nContacts >= maxContactCapacity)
+				{
+					b3Error("Exceeded contact capacity (%d/%d)\n",nContacts,maxContactCapacity);
+					nContacts = maxContactCapacity;
+				}
+				contactOut->resize(nContacts);
+			}
+		}
+
+
+		int nCompoundsPairs = m_gpuCompoundPairs.size();
+
+		if (nCompoundsPairs)
+		{
+			b3BufferInfoCL bInfo[] = {
+				b3BufferInfoCL( m_gpuCompoundPairs.getBufferCL(), true ), 
+				b3BufferInfoCL( bodyBuf->getBufferCL(),true), 
+				b3BufferInfoCL( gpuCollidables.getBufferCL(),true), 
+				b3BufferInfoCL( convexData.getBufferCL(),true),
+				b3BufferInfoCL( gpuVertices.getBufferCL(),true),
+				b3BufferInfoCL( gpuUniqueEdges.getBufferCL(),true),
+				b3BufferInfoCL( gpuFaces.getBufferCL(),true),
+				b3BufferInfoCL( gpuIndices.getBufferCL(),true),
+				b3BufferInfoCL( gpuChildShapes.getBufferCL(),true),
+				b3BufferInfoCL( m_gpuCompoundSepNormals.getBufferCL(),true),
+				b3BufferInfoCL( m_gpuHasCompoundSepNormals.getBufferCL(),true),
+				b3BufferInfoCL( contactOut->getBufferCL()),
+				b3BufferInfoCL( m_totalContactsOut.getBufferCL())	
+			};
+			b3LauncherCL launcher(m_queue, m_clipCompoundsHullHullKernel,"m_clipCompoundsHullHullKernel");
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst( nCompoundsPairs  );
+			launcher.setConst(maxContactCapacity);
+
+			int num = nCompoundsPairs;
+			launcher.launch1D( num);
+			clFinish(m_queue);
+
+			nContacts = m_totalContactsOut.at(0);
+			if (nContacts>maxContactCapacity)
+			{
+
+				b3Error("Error: contacts exceeds capacity (%d/%d)\n", nContacts, maxContactCapacity);
+				nContacts = maxContactCapacity;
+			}
+			contactOut->resize(nContacts);
+		}//if nCompoundsPairs
+		}
+	}//contactClippingOnGpu
+
+	//printf("nContacts end = %d\n",nContacts);
+	
+	//printf("frameCount = %d\n",frameCount++);
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ConvexHullContact.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ConvexHullContact.h
new file mode 100644
index 0000000000..e24c1579c6
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ConvexHullContact.h
@@ -0,0 +1,118 @@
+
+#ifndef _CONVEX_HULL_CONTACT_H
+#define _CONVEX_HULL_CONTACT_H
+
+#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3RigidBodyData.h"
+#include "Bullet3Common/b3AlignedObjectArray.h"
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3ConvexPolyhedronData.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Collidable.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3Contact4.h"
+#include "Bullet3Common/shared/b3Int2.h"
+#include "Bullet3Common/shared/b3Int4.h"
+#include "b3OptimizedBvh.h"
+#include "b3BvhInfo.h"
+#include "Bullet3Collision/BroadPhaseCollision/shared/b3Aabb.h"
+
+//#include "../../dynamics/basic_demo/Stubs/ChNarrowPhase.h"
+
+
+
+
+struct GpuSatCollision
+{
+	cl_context				m_context;
+	cl_device_id			m_device;
+	cl_command_queue		m_queue;
+	cl_kernel				m_findSeparatingAxisKernel;
+	cl_kernel				m_mprPenetrationKernel;
+	cl_kernel				m_findSeparatingAxisUnitSphereKernel;
+
+
+	cl_kernel m_findSeparatingAxisVertexFaceKernel;
+	cl_kernel m_findSeparatingAxisEdgeEdgeKernel;
+	
+	cl_kernel				m_findConcaveSeparatingAxisKernel;
+    cl_kernel				m_findConcaveSeparatingAxisVertexFaceKernel;
+    cl_kernel				m_findConcaveSeparatingAxisEdgeEdgeKernel;
+ 
+    
+    
+    
+	cl_kernel				m_findCompoundPairsKernel;
+	cl_kernel				m_processCompoundPairsKernel;
+
+	cl_kernel				m_clipHullHullKernel;
+	cl_kernel				m_clipCompoundsHullHullKernel;
+    
+    cl_kernel               m_clipFacesAndFindContacts;
+    cl_kernel               m_findClippingFacesKernel;
+    
+	cl_kernel				m_clipHullHullConcaveConvexKernel;
+//	cl_kernel				m_extractManifoldAndAddContactKernel;
+    cl_kernel               m_newContactReductionKernel;
+
+	cl_kernel				m_bvhTraversalKernel;
+	cl_kernel				m_primitiveContactsKernel;
+	cl_kernel				m_findConcaveSphereContactsKernel;
+
+	cl_kernel				m_processCompoundPairsPrimitivesKernel;
+    
+	b3OpenCLArray<b3Vector3> m_unitSphereDirections;
+
+	b3OpenCLArray<int>		m_totalContactsOut;
+
+	b3OpenCLArray<b3Vector3> m_sepNormals;
+	b3OpenCLArray<float> m_dmins;
+
+	b3OpenCLArray<int>		m_hasSeparatingNormals;
+	b3OpenCLArray<b3Vector3> m_concaveSepNormals;
+	b3OpenCLArray<int>		m_concaveHasSeparatingNormals;
+	b3OpenCLArray<int>		m_numConcavePairsOut;
+	b3OpenCLArray<b3CompoundOverlappingPair> m_gpuCompoundPairs;
+	b3OpenCLArray<b3Vector3> m_gpuCompoundSepNormals;
+	b3OpenCLArray<int>		m_gpuHasCompoundSepNormals;
+	b3OpenCLArray<int>		m_numCompoundPairsOut;
+	
+
+	GpuSatCollision(cl_context ctx,cl_device_id device, cl_command_queue  q );
+	virtual ~GpuSatCollision();
+	
+
+	void computeConvexConvexContactsGPUSAT( b3OpenCLArray<b3Int4>* pairs, int nPairs, 
+			const b3OpenCLArray<b3RigidBodyData>* bodyBuf,
+			b3OpenCLArray<b3Contact4>* contactOut, int& nContacts,
+			const b3OpenCLArray<b3Contact4>* oldContacts,
+			int maxContactCapacity,
+			int compoundPairCapacity,
+			const b3OpenCLArray<b3ConvexPolyhedronData>& hostConvexData,
+			const b3OpenCLArray<b3Vector3>& vertices,
+			const b3OpenCLArray<b3Vector3>& uniqueEdges,
+			const b3OpenCLArray<b3GpuFace>& faces,
+			const b3OpenCLArray<int>& indices,
+			const b3OpenCLArray<b3Collidable>& gpuCollidables,
+			const b3OpenCLArray<b3GpuChildShape>& gpuChildShapes,
+
+			const b3OpenCLArray<b3Aabb>& clAabbsWorldSpace,
+			const b3OpenCLArray<b3Aabb>& clAabbsLocalSpace,
+
+           b3OpenCLArray<b3Vector3>& worldVertsB1GPU,
+           b3OpenCLArray<b3Int4>& clippingFacesOutGPU,
+           b3OpenCLArray<b3Vector3>& worldNormalsAGPU,
+           b3OpenCLArray<b3Vector3>& worldVertsA1GPU,
+           b3OpenCLArray<b3Vector3>& worldVertsB2GPU,
+		   b3AlignedObjectArray<class b3OptimizedBvh*>& bvhData,
+		   b3OpenCLArray<b3QuantizedBvhNode>*	treeNodesGPU,
+			b3OpenCLArray<b3BvhSubtreeInfo>*	subTreesGPU,
+			b3OpenCLArray<b3BvhInfo>*	bvhInfo,
+			int numObjects,
+			int maxTriConvexPairCapacity,
+			b3OpenCLArray<b3Int4>& triangleConvexPairs,
+			int& numTriConvexPairsOut
+			);
+
+
+};
+
+#endif //_CONVEX_HULL_CONTACT_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ConvexPolyhedronCL.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ConvexPolyhedronCL.h
new file mode 100644
index 0000000000..337100fb1a
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3ConvexPolyhedronCL.h
@@ -0,0 +1,9 @@
+#ifndef CONVEX_POLYHEDRON_CL
+#define CONVEX_POLYHEDRON_CL
+
+#include "Bullet3Common/b3Transform.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3ConvexPolyhedronData.h"
+
+
+
+#endif //CONVEX_POLYHEDRON_CL
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3GjkEpa.cpp b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3GjkEpa.cpp
new file mode 100644
index 0000000000..d636f983c6
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3GjkEpa.cpp
@@ -0,0 +1,1014 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2008 Erwin Coumans  http://continuousphysics.com/Bullet/
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the
+use of this software.
+Permission is granted to anyone to use this software for any purpose,
+including commercial applications, and to alter it and redistribute it
+freely,
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not
+claim that you wrote the original software. If you use this software in a
+product, an acknowledgment in the product documentation would be appreciated
+but is not required.
+2. Altered source versions must be plainly marked as such, and must not be
+misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+/*
+GJK-EPA collision solver by Nathanael Presson, 2008
+*/
+
+#include "b3GjkEpa.h"
+
+#include "b3SupportMappings.h"
+
+namespace gjkepa2_impl2
+{
+
+	// Config
+
+	/* GJK	*/ 
+#define GJK_MAX_ITERATIONS	128
+#define GJK_ACCURACY		((b3Scalar)0.0001)
+#define GJK_MIN_DISTANCE	((b3Scalar)0.0001)
+#define GJK_DUPLICATED_EPS	((b3Scalar)0.0001)
+#define GJK_SIMPLEX2_EPS	((b3Scalar)0.0)
+#define GJK_SIMPLEX3_EPS	((b3Scalar)0.0)
+#define GJK_SIMPLEX4_EPS	((b3Scalar)0.0)
+
+	/* EPA	*/ 
+#define EPA_MAX_VERTICES	64
+#define EPA_MAX_FACES		(EPA_MAX_VERTICES*2)
+#define EPA_MAX_ITERATIONS	255
+#define EPA_ACCURACY		((b3Scalar)0.0001)
+#define EPA_FALLBACK		(10*EPA_ACCURACY)
+#define EPA_PLANE_EPS		((b3Scalar)0.00001)
+#define EPA_INSIDE_EPS		((b3Scalar)0.01)
+
+
+	// Shorthands
+	
+
+	// MinkowskiDiff
+	struct	b3MinkowskiDiff
+	{
+		
+
+		const b3ConvexPolyhedronData*	m_shapes[2];
+		
+
+		b3Matrix3x3				m_toshape1;
+		b3Transform				m_toshape0;
+
+		bool					m_enableMargin;
+		
+
+			void					EnableMargin(bool enable)
+		{
+			m_enableMargin = enable;
+		}	
+		inline b3Vector3		Support0(const b3Vector3& d, const b3AlignedObjectArray<b3Vector3>& verticesA) const
+		{
+			if (m_enableMargin)
+			{
+				return localGetSupportVertexWithMargin(d,m_shapes[0],verticesA,0.f);
+			} else
+			{
+				return localGetSupportVertexWithoutMargin(d,m_shapes[0],verticesA);
+			}
+		}
+		inline b3Vector3		Support1(const b3Vector3& d, const b3AlignedObjectArray<b3Vector3>& verticesB) const
+		{
+			if (m_enableMargin)
+			{
+				return m_toshape0*(localGetSupportVertexWithMargin(m_toshape1*d,m_shapes[1],verticesB,0.f));
+			} else
+			{
+				return m_toshape0*(localGetSupportVertexWithoutMargin(m_toshape1*d,m_shapes[1],verticesB));
+			}
+		}
+
+		inline b3Vector3		Support(const b3Vector3& d,  const b3AlignedObjectArray<b3Vector3>& verticesA,  const b3AlignedObjectArray<b3Vector3>& verticesB) const
+		{
+			return(Support0(d,verticesA)-Support1(-d,verticesB));
+		}
+		b3Vector3				Support(const b3Vector3& d,unsigned int index,const b3AlignedObjectArray<b3Vector3>& verticesA,  const b3AlignedObjectArray<b3Vector3>& verticesB) const
+		{
+			if(index)
+				return(Support1(d,verticesA));
+			else
+				return(Support0(d,verticesB));
+		}
+	};
+
+	typedef	b3MinkowskiDiff	tShape;
+
+
+	// GJK
+	struct	b3GJK
+	{
+		/* Types		*/ 
+		struct	sSV
+		{
+			b3Vector3	d,w;
+		};
+		struct	sSimplex
+		{
+			sSV*		c[4];
+			b3Scalar	p[4];
+			unsigned int			rank;
+		};
+		struct	eStatus	{ enum _ {
+			Valid,
+			Inside,
+			Failed		};};
+			/* Fields		*/ 
+			tShape			m_shape;
+			const b3AlignedObjectArray<b3Vector3>& m_verticesA;
+			const b3AlignedObjectArray<b3Vector3>& m_verticesB;
+			b3Vector3		m_ray;
+			b3Scalar		m_distance;
+			sSimplex		m_simplices[2];
+			sSV				m_store[4];
+			sSV*			m_free[4];
+			unsigned int				m_nfree;
+			unsigned int				m_current;
+			sSimplex*		m_simplex;
+			eStatus::_		m_status;
+			/* Methods		*/ 
+			b3GJK(const b3AlignedObjectArray<b3Vector3>& verticesA,const b3AlignedObjectArray<b3Vector3>& verticesB)
+				:m_verticesA(verticesA),m_verticesB(verticesB)
+			{
+				Initialize();
+			}
+			void				Initialize()
+			{
+				m_ray		=	b3MakeVector3(0,0,0);
+				m_nfree		=	0;
+				m_status	=	eStatus::Failed;
+				m_current	=	0;
+				m_distance	=	0;
+			}
+			eStatus::_			Evaluate(const tShape& shapearg,const b3Vector3& guess)
+			{
+				unsigned int			iterations=0;
+				b3Scalar	sqdist=0;
+				b3Scalar	alpha=0;
+				b3Vector3	lastw[4];
+				unsigned int			clastw=0;
+				/* Initialize solver		*/ 
+				m_free[0]			=	&m_store[0];
+				m_free[1]			=	&m_store[1];
+				m_free[2]			=	&m_store[2];
+				m_free[3]			=	&m_store[3];
+				m_nfree				=	4;
+				m_current			=	0;
+				m_status			=	eStatus::Valid;
+				m_shape				=	shapearg;
+				m_distance			=	0;
+				/* Initialize simplex		*/ 
+				m_simplices[0].rank	=	0;
+				m_ray				=	guess;
+				const b3Scalar	sqrl=	m_ray.length2();
+				appendvertice(m_simplices[0],sqrl>0?-m_ray:b3MakeVector3(1,0,0));
+				m_simplices[0].p[0]	=	1;
+				m_ray				=	m_simplices[0].c[0]->w;	
+				sqdist				=	sqrl;
+				lastw[0]			=
+					lastw[1]			=
+					lastw[2]			=
+					lastw[3]			=	m_ray;
+				/* Loop						*/ 
+				do	{
+					const unsigned int		next=1-m_current;
+					sSimplex&	cs=m_simplices[m_current];
+					sSimplex&	ns=m_simplices[next];
+					/* Check zero							*/ 
+					const b3Scalar	rl=m_ray.length();
+					if(rl<GJK_MIN_DISTANCE)
+					{/* Touching or inside				*/ 
+						m_status=eStatus::Inside;
+						break;
+					}
+					/* Append new vertice in -'v' direction	*/ 
+					appendvertice(cs,-m_ray);
+					const b3Vector3&	w=cs.c[cs.rank-1]->w;
+					bool				found=false;
+					for(unsigned int i=0;i<4;++i)
+					{
+						if((w-lastw[i]).length2()<GJK_DUPLICATED_EPS)
+						{ found=true;break; }
+					}
+					if(found)
+					{/* Return old simplex				*/ 
+						removevertice(m_simplices[m_current]);
+						break;
+					}
+					else
+					{/* Update lastw					*/ 
+						lastw[clastw=(clastw+1)&3]=w;
+					}
+					/* Check for termination				*/ 
+					const b3Scalar	omega=b3Dot(m_ray,w)/rl;
+					alpha=b3Max(omega,alpha);
+					if(((rl-alpha)-(GJK_ACCURACY*rl))<=0)
+					{/* Return old simplex				*/ 
+						removevertice(m_simplices[m_current]);
+						break;
+					}		
+					/* Reduce simplex						*/ 
+					b3Scalar	weights[4];
+					unsigned int			mask=0;
+					switch(cs.rank)
+					{
+					case	2:	sqdist=projectorigin(	cs.c[0]->w,
+									cs.c[1]->w,
+									weights,mask);break;
+					case	3:	sqdist=projectorigin(	cs.c[0]->w,
+									cs.c[1]->w,
+									cs.c[2]->w,
+									weights,mask);break;
+					case	4:	sqdist=projectorigin(	cs.c[0]->w,
+									cs.c[1]->w,
+									cs.c[2]->w,
+									cs.c[3]->w,
+									weights,mask);break;
+					}
+					if(sqdist>=0)
+					{/* Valid	*/ 
+						ns.rank		=	0;
+						m_ray		=	b3MakeVector3(0,0,0);
+						m_current	=	next;
+						for(unsigned int i=0,ni=cs.rank;i<ni;++i)
+						{
+							if(mask&(1<<i))
+							{
+								ns.c[ns.rank]		=	cs.c[i];
+								ns.p[ns.rank++]		=	weights[i];
+								m_ray				+=	cs.c[i]->w*weights[i];
+							}
+							else
+							{
+								m_free[m_nfree++]	=	cs.c[i];
+							}
+						}
+						if(mask==15) m_status=eStatus::Inside;
+					}
+					else
+					{/* Return old simplex				*/ 
+						removevertice(m_simplices[m_current]);
+						break;
+					}
+					m_status=((++iterations)<GJK_MAX_ITERATIONS)?m_status:eStatus::Failed;
+				} while(m_status==eStatus::Valid);
+				m_simplex=&m_simplices[m_current];
+				switch(m_status)
+				{
+				case	eStatus::Valid:		m_distance=m_ray.length();break;
+				case	eStatus::Inside:	m_distance=0;break;
+				default:
+					{
+					}
+				}	
+				return(m_status);
+			}
+			bool					EncloseOrigin()
+			{
+				switch(m_simplex->rank)
+				{
+				case	1:
+					{
+						for(unsigned int i=0;i<3;++i)
+						{
+							b3Vector3		axis=b3MakeVector3(0,0,0);
+							axis[i]=1;
+							appendvertice(*m_simplex, axis);
+							if(EncloseOrigin())	return(true);
+							removevertice(*m_simplex);
+							appendvertice(*m_simplex,-axis);
+							if(EncloseOrigin())	return(true);
+							removevertice(*m_simplex);
+						}
+					}
+					break;
+				case	2:
+					{
+						const b3Vector3	d=m_simplex->c[1]->w-m_simplex->c[0]->w;
+						for(unsigned int i=0;i<3;++i)
+						{
+							b3Vector3		axis=b3MakeVector3(0,0,0);
+							axis[i]=1;
+							const b3Vector3	p=b3Cross(d,axis);
+							if(p.length2()>0)
+							{
+								appendvertice(*m_simplex, p);
+								if(EncloseOrigin())	return(true);
+								removevertice(*m_simplex);
+								appendvertice(*m_simplex,-p);
+								if(EncloseOrigin())	return(true);
+								removevertice(*m_simplex);
+							}
+						}
+					}
+					break;
+				case	3:
+					{
+						const b3Vector3	n=b3Cross(m_simplex->c[1]->w-m_simplex->c[0]->w,
+							m_simplex->c[2]->w-m_simplex->c[0]->w);
+						if(n.length2()>0)
+						{
+							appendvertice(*m_simplex,n);
+							if(EncloseOrigin())	return(true);
+							removevertice(*m_simplex);
+							appendvertice(*m_simplex,-n);
+							if(EncloseOrigin())	return(true);
+							removevertice(*m_simplex);
+						}
+					}
+					break;
+				case	4:
+					{
+						if(b3Fabs(det(	m_simplex->c[0]->w-m_simplex->c[3]->w,
+							m_simplex->c[1]->w-m_simplex->c[3]->w,
+							m_simplex->c[2]->w-m_simplex->c[3]->w))>0)
+							return(true);
+					}
+					break;
+				}
+				return(false);
+			}
+			/* Internals	*/ 
+			void				getsupport(const b3Vector3& d,sSV& sv) const
+			{
+				sv.d	=	d/d.length();
+				sv.w	=	m_shape.Support(sv.d,m_verticesA,m_verticesB);
+			}
+			void				removevertice(sSimplex& simplex)
+			{
+				m_free[m_nfree++]=simplex.c[--simplex.rank];
+			}
+			void				appendvertice(sSimplex& simplex,const b3Vector3& v)
+			{
+				simplex.p[simplex.rank]=0;
+				simplex.c[simplex.rank]=m_free[--m_nfree];
+				getsupport(v,*simplex.c[simplex.rank++]);
+			}
+			static b3Scalar		det(const b3Vector3& a,const b3Vector3& b,const b3Vector3& c)
+			{
+				return(	a.y*b.z*c.x+a.z*b.x*c.y-
+					a.x*b.z*c.y-a.y*b.x*c.z+
+					a.x*b.y*c.z-a.z*b.y*c.x);
+			}
+			static b3Scalar		projectorigin(	const b3Vector3& a,
+				const b3Vector3& b,
+				b3Scalar* w,unsigned int& m)
+			{
+				const b3Vector3	d=b-a;
+				const b3Scalar	l=d.length2();
+				if(l>GJK_SIMPLEX2_EPS)
+				{
+					const b3Scalar	t(l>0?-b3Dot(a,d)/l:0);
+					if(t>=1)		{ w[0]=0;w[1]=1;m=2;return(b.length2()); }
+					else if(t<=0)	{ w[0]=1;w[1]=0;m=1;return(a.length2()); }
+					else			{ w[0]=1-(w[1]=t);m=3;return((a+d*t).length2()); }
+				}
+				return(-1);
+			}
+			static b3Scalar		projectorigin(	const b3Vector3& a,
+				const b3Vector3& b,
+				const b3Vector3& c,
+				b3Scalar* w,unsigned int& m)
+			{
+				static const unsigned int		imd3[]={1,2,0};
+				const b3Vector3*	vt[]={&a,&b,&c};
+				const b3Vector3		dl[]={a-b,b-c,c-a};
+				const b3Vector3		n=b3Cross(dl[0],dl[1]);
+				const b3Scalar		l=n.length2();
+				if(l>GJK_SIMPLEX3_EPS)
+				{
+					b3Scalar	mindist=-1;
+					b3Scalar	subw[2]={0.f,0.f};
+					unsigned int			subm(0);
+					for(unsigned int i=0;i<3;++i)
+					{
+						if(b3Dot(*vt[i],b3Cross(dl[i],n))>0)
+						{
+							const unsigned int			j=imd3[i];
+							const b3Scalar	subd(projectorigin(*vt[i],*vt[j],subw,subm));
+							if((mindist<0)||(subd<mindist))
+							{
+								mindist		=	subd;
+								m			=	static_cast<unsigned int>(((subm&1)?1<<i:0)+((subm&2)?1<<j:0));
+								w[i]		=	subw[0];
+								w[j]		=	subw[1];
+								w[imd3[j]]	=	0;				
+							}
+						}
+					}
+					if(mindist<0)
+					{
+						const b3Scalar	d=b3Dot(a,n);	
+						const b3Scalar	s=b3Sqrt(l);
+						const b3Vector3	p=n*(d/l);
+						mindist	=	p.length2();
+						m		=	7;
+						w[0]	=	(b3Cross(dl[1],b-p)).length()/s;
+						w[1]	=	(b3Cross(dl[2],c-p)).length()/s;
+						w[2]	=	1-(w[0]+w[1]);
+					}
+					return(mindist);
+				}
+				return(-1);
+			}
+			static b3Scalar		projectorigin(	const b3Vector3& a,
+				const b3Vector3& b,
+				const b3Vector3& c,
+				const b3Vector3& d,
+				b3Scalar* w,unsigned int& m)
+			{
+				static const unsigned int		imd3[]={1,2,0};
+				const b3Vector3*	vt[]={&a,&b,&c,&d};
+				const b3Vector3		dl[]={a-d,b-d,c-d};
+				const b3Scalar		vl=det(dl[0],dl[1],dl[2]);
+				const bool			ng=(vl*b3Dot(a,b3Cross(b-c,a-b)))<=0;
+				if(ng&&(b3Fabs(vl)>GJK_SIMPLEX4_EPS))
+				{
+					b3Scalar	mindist=-1;
+					b3Scalar	subw[3]={0.f,0.f,0.f};
+					unsigned int			subm(0);
+					for(unsigned int i=0;i<3;++i)
+					{
+						const unsigned int			j=imd3[i];
+						const b3Scalar	s=vl*b3Dot(d,b3Cross(dl[i],dl[j]));
+						if(s>0)
+						{
+							const b3Scalar	subd=projectorigin(*vt[i],*vt[j],d,subw,subm);
+							if((mindist<0)||(subd<mindist))
+							{
+								mindist		=	subd;
+								m			=	static_cast<unsigned int>((subm&1?1<<i:0)+
+									(subm&2?1<<j:0)+
+									(subm&4?8:0));
+								w[i]		=	subw[0];
+								w[j]		=	subw[1];
+								w[imd3[j]]	=	0;
+								w[3]		=	subw[2];
+							}
+						}
+					}
+					if(mindist<0)
+					{
+						mindist	=	0;
+						m		=	15;
+						w[0]	=	det(c,b,d)/vl;
+						w[1]	=	det(a,c,d)/vl;
+						w[2]	=	det(b,a,d)/vl;
+						w[3]	=	1-(w[0]+w[1]+w[2]);
+					}
+					return(mindist);
+				}
+				return(-1);
+			}
+	};
+
+	// EPA
+	struct	b3EPA
+	{
+		/* Types		*/ 
+		typedef	b3GJK::sSV	sSV;
+		struct	sFace
+		{
+			b3Vector3	n;
+			b3Scalar	d;
+			sSV*		c[3];
+			sFace*		f[3];
+			sFace*		l[2];
+			unsigned char			e[3];
+			unsigned char			pass;
+		};
+		struct	sList
+		{
+			sFace*		root;
+			unsigned int			count;
+			sList() : root(0),count(0)	{}
+		};
+		struct	sHorizon
+		{
+			sFace*		cf;
+			sFace*		ff;
+			unsigned int			nf;
+			sHorizon() : cf(0),ff(0),nf(0)	{}
+		};
+		struct	eStatus { enum _ {
+			Valid,
+			Touching,
+			Degenerated,
+			NonConvex,
+			InvalidHull,		
+			OutOfFaces,
+			OutOfVertices,
+			AccuraryReached,
+			FallBack,
+			Failed		};};
+			/* Fields		*/ 
+			eStatus::_		m_status;
+			b3GJK::sSimplex	m_result;
+			b3Vector3		m_normal;
+			b3Scalar		m_depth;
+			sSV				m_sv_store[EPA_MAX_VERTICES];
+			sFace			m_fc_store[EPA_MAX_FACES];
+			unsigned int				m_nextsv;
+			sList			m_hull;
+			sList			m_stock;
+			/* Methods		*/ 
+			b3EPA()
+			{
+				Initialize();	
+			}
+
+
+			static inline void		bind(sFace* fa,unsigned int ea,sFace* fb,unsigned int eb)
+			{
+				fa->e[ea]=(unsigned char)eb;fa->f[ea]=fb;
+				fb->e[eb]=(unsigned char)ea;fb->f[eb]=fa;
+			}
+			static inline void		append(sList& list,sFace* face)
+			{
+				face->l[0]	=	0;
+				face->l[1]	=	list.root;
+				if(list.root) list.root->l[0]=face;
+				list.root	=	face;
+				++list.count;
+			}
+			static inline void		remove(sList& list,sFace* face)
+			{
+				if(face->l[1]) face->l[1]->l[0]=face->l[0];
+				if(face->l[0]) face->l[0]->l[1]=face->l[1];
+				if(face==list.root) list.root=face->l[1];
+				--list.count;
+			}
+
+
+			void				Initialize()
+			{
+				m_status	=	eStatus::Failed;
+				m_normal	=	b3MakeVector3(0,0,0);
+				m_depth		=	0;
+				m_nextsv	=	0;
+				for(unsigned int i=0;i<EPA_MAX_FACES;++i)
+				{
+					append(m_stock,&m_fc_store[EPA_MAX_FACES-i-1]);
+				}
+			}
+			eStatus::_			Evaluate(b3GJK& gjk,const b3Vector3& guess)
+			{
+				b3GJK::sSimplex&	simplex=*gjk.m_simplex;
+				if((simplex.rank>1)&&gjk.EncloseOrigin())
+				{
+
+					/* Clean up				*/ 
+					while(m_hull.root)
+					{
+						sFace*	f = m_hull.root;
+						remove(m_hull,f);
+						append(m_stock,f);
+					}
+					m_status	=	eStatus::Valid;
+					m_nextsv	=	0;
+					/* Orient simplex		*/ 
+					if(gjk.det(	simplex.c[0]->w-simplex.c[3]->w,
+						simplex.c[1]->w-simplex.c[3]->w,
+						simplex.c[2]->w-simplex.c[3]->w)<0)
+					{
+						b3Swap(simplex.c[0],simplex.c[1]);
+						b3Swap(simplex.p[0],simplex.p[1]);
+					}
+					/* Build initial hull	*/ 
+					sFace*	tetra[]={newface(simplex.c[0],simplex.c[1],simplex.c[2],true),
+						newface(simplex.c[1],simplex.c[0],simplex.c[3],true),
+						newface(simplex.c[2],simplex.c[1],simplex.c[3],true),
+						newface(simplex.c[0],simplex.c[2],simplex.c[3],true)};
+					if(m_hull.count==4)
+					{
+						sFace*		best=findbest();
+						sFace		outer=*best;
+						unsigned int			pass=0;
+						unsigned int			iterations=0;
+						bind(tetra[0],0,tetra[1],0);
+						bind(tetra[0],1,tetra[2],0);
+						bind(tetra[0],2,tetra[3],0);
+						bind(tetra[1],1,tetra[3],2);
+						bind(tetra[1],2,tetra[2],1);
+						bind(tetra[2],2,tetra[3],1);
+						m_status=eStatus::Valid;
+						for(;iterations<EPA_MAX_ITERATIONS;++iterations)
+						{
+							if(m_nextsv<EPA_MAX_VERTICES)
+							{	
+								sHorizon		horizon;
+								sSV*			w=&m_sv_store[m_nextsv++];
+								bool			valid=true;					
+								best->pass	=	(unsigned char)(++pass);
+								gjk.getsupport(best->n,*w);
+								const b3Scalar	wdist=b3Dot(best->n,w->w)-best->d;
+								if(wdist>EPA_ACCURACY)
+								{
+									for(unsigned int j=0;(j<3)&&valid;++j)
+									{
+										valid&=expand(	pass,w,
+											best->f[j],best->e[j],
+											horizon);
+									}
+									if(valid&&(horizon.nf>=3))
+									{
+										bind(horizon.cf,1,horizon.ff,2);
+										remove(m_hull,best);
+										append(m_stock,best);
+										best=findbest();
+										outer=*best;
+									} else { 
+										m_status=eStatus::Failed;
+										//m_status=eStatus::InvalidHull;
+									break; }
+								} else { m_status=eStatus::AccuraryReached;break; }
+							} else { m_status=eStatus::OutOfVertices;break; }
+						}
+						const b3Vector3	projection=outer.n*outer.d;
+						m_normal	=	outer.n;
+						m_depth		=	outer.d;
+						m_result.rank	=	3;
+						m_result.c[0]	=	outer.c[0];
+						m_result.c[1]	=	outer.c[1];
+						m_result.c[2]	=	outer.c[2];
+						m_result.p[0]	=	b3Cross(	outer.c[1]->w-projection,
+							outer.c[2]->w-projection).length();
+						m_result.p[1]	=	b3Cross(	outer.c[2]->w-projection,
+							outer.c[0]->w-projection).length();
+						m_result.p[2]	=	b3Cross(	outer.c[0]->w-projection,
+							outer.c[1]->w-projection).length();
+						const b3Scalar	sum=m_result.p[0]+m_result.p[1]+m_result.p[2];
+						m_result.p[0]	/=	sum;
+						m_result.p[1]	/=	sum;
+						m_result.p[2]	/=	sum;
+						return(m_status);
+					}
+				}
+				/* Fallback		*/ 
+				m_status	=	eStatus::FallBack;
+				m_normal	=	-guess;
+				const b3Scalar	nl=m_normal.length();
+				if(nl>0)
+					m_normal	=	m_normal/nl;
+				else
+					m_normal	=	b3MakeVector3(1,0,0);
+				m_depth	=	0;
+				m_result.rank=1;
+				m_result.c[0]=simplex.c[0];
+				m_result.p[0]=1;	
+				return(m_status);
+			}
+			bool getedgedist(sFace* face, sSV* a, sSV* b, b3Scalar& dist)
+			{
+				const b3Vector3 ba = b->w - a->w;
+				const b3Vector3 n_ab = b3Cross(ba, face->n); // Outward facing edge normal direction, on triangle plane
+				const b3Scalar a_dot_nab = b3Dot(a->w, n_ab); // Only care about the sign to determine inside/outside, so not normalization required
+
+				if(a_dot_nab < 0)
+				{
+					// Outside of edge a->b
+
+					const b3Scalar ba_l2 = ba.length2();
+					const b3Scalar a_dot_ba = b3Dot(a->w, ba);
+					const b3Scalar b_dot_ba = b3Dot(b->w, ba);
+
+					if(a_dot_ba > 0)
+					{
+						// Pick distance vertex a
+						dist = a->w.length();
+					}
+					else if(b_dot_ba < 0)
+					{
+						// Pick distance vertex b
+						dist = b->w.length();
+					}
+					else
+					{
+						// Pick distance to edge a->b
+						const b3Scalar a_dot_b = b3Dot(a->w, b->w);
+						dist = b3Sqrt(b3Max((a->w.length2() * b->w.length2() - a_dot_b * a_dot_b) / ba_l2, (b3Scalar)0));
+					}
+
+					return true;
+				}
+
+				return false;
+			}
+			sFace*				newface(sSV* a,sSV* b,sSV* c,bool forced)
+			{
+				if(m_stock.root)
+				{
+					sFace*	face=m_stock.root;
+					remove(m_stock,face);
+					append(m_hull,face);
+					face->pass	=	0;
+					face->c[0]	=	a;
+					face->c[1]	=	b;
+					face->c[2]	=	c;
+					face->n		=	b3Cross(b->w-a->w,c->w-a->w);
+					const b3Scalar	l=face->n.length();
+					const bool		v=l>EPA_ACCURACY;
+
+					if(v)
+					{
+						if(!(getedgedist(face, a, b, face->d) ||
+							 getedgedist(face, b, c, face->d) ||
+							 getedgedist(face, c, a, face->d)))
+						{
+							// Origin projects to the interior of the triangle
+							// Use distance to triangle plane
+							face->d = b3Dot(a->w, face->n) / l;
+						}
+
+						face->n /= l;
+						if(forced || (face->d >= -EPA_PLANE_EPS))
+						{
+							return face;
+						}
+						else
+							m_status=eStatus::NonConvex;
+					}
+					else
+						m_status=eStatus::Degenerated;
+
+					remove(m_hull, face);
+					append(m_stock, face);
+					return 0;
+
+				}
+				m_status = m_stock.root ? eStatus::OutOfVertices : eStatus::OutOfFaces;
+				return 0;
+			}
+			sFace*				findbest()
+			{
+				sFace*		minf=m_hull.root;
+				b3Scalar	mind=minf->d*minf->d;
+				for(sFace* f=minf->l[1];f;f=f->l[1])
+				{
+					const b3Scalar	sqd=f->d*f->d;
+					if(sqd<mind)
+					{
+						minf=f;
+						mind=sqd;
+					}
+				}
+				return(minf);
+			}
+			bool				expand(unsigned int pass,sSV* w,sFace* f,unsigned int e,sHorizon& horizon)
+			{
+				static const unsigned int	i1m3[]={1,2,0};
+				static const unsigned int	i2m3[]={2,0,1};
+				if(f->pass!=pass)
+				{
+					const unsigned int	e1=i1m3[e];
+					if((b3Dot(f->n,w->w)-f->d)<-EPA_PLANE_EPS)
+					{
+						sFace*	nf=newface(f->c[e1],f->c[e],w,false);
+						if(nf)
+						{
+							bind(nf,0,f,e);
+							if(horizon.cf) bind(horizon.cf,1,nf,2); else horizon.ff=nf;
+							horizon.cf=nf;
+							++horizon.nf;
+							return(true);
+						}
+					}
+					else
+					{
+						const unsigned int	e2=i2m3[e];
+						f->pass		=	(unsigned char)pass;
+						if(	expand(pass,w,f->f[e1],f->e[e1],horizon)&&
+							expand(pass,w,f->f[e2],f->e[e2],horizon))
+						{
+							remove(m_hull,f);
+							append(m_stock,f);
+							return(true);
+						}
+					}
+				}
+				return(false);
+			}
+
+	};
+
+	//
+	static void	Initialize(const b3Transform&	transA, const b3Transform&	transB,
+								const b3ConvexPolyhedronData* hullA, const b3ConvexPolyhedronData* hullB, 
+								const b3AlignedObjectArray<b3Vector3>& verticesA,
+								const b3AlignedObjectArray<b3Vector3>& verticesB,
+		b3GjkEpaSolver2::sResults& results,
+		tShape& shape,
+		bool withmargins)
+	{
+		/* Results		*/ 
+		results.witnesses[0]	=
+			results.witnesses[1]	=	b3MakeVector3(0,0,0);
+		results.status			=	b3GjkEpaSolver2::sResults::Separated;
+		/* Shape		*/ 
+		shape.m_shapes[0]		=	hullA;
+		shape.m_shapes[1]		=	hullB;
+		shape.m_toshape1		=	transB.getBasis().transposeTimes(transA.getBasis());
+		shape.m_toshape0		=	transA.inverseTimes(transB);
+		shape.EnableMargin(withmargins);
+	}
+
+}
+
+//
+// Api
+//
+
+using namespace	gjkepa2_impl2;
+
+//
+int			b3GjkEpaSolver2::StackSizeRequirement()
+{
+	return(sizeof(b3GJK)+sizeof(b3EPA));
+}
+
+//
+bool		b3GjkEpaSolver2::Distance(	const b3Transform&	transA, const b3Transform&	transB,
+										const b3ConvexPolyhedronData* hullA, const b3ConvexPolyhedronData* hullB, 
+										const b3AlignedObjectArray<b3Vector3>& verticesA,
+										const b3AlignedObjectArray<b3Vector3>& verticesB,
+									  const b3Vector3&		guess,
+									  sResults&				results)
+{
+	tShape			shape;
+	Initialize(transA,transB,hullA,hullB,verticesA,verticesB,results,shape,false);
+	b3GJK				gjk(verticesA,verticesB);
+	b3GJK::eStatus::_	gjk_status=gjk.Evaluate(shape,guess);
+	if(gjk_status==b3GJK::eStatus::Valid)
+	{
+		b3Vector3	w0=b3MakeVector3(0,0,0);
+		b3Vector3	w1=b3MakeVector3(0,0,0);
+		for(unsigned int i=0;i<gjk.m_simplex->rank;++i)
+		{
+			const b3Scalar	p=gjk.m_simplex->p[i];
+			w0+=shape.Support( gjk.m_simplex->c[i]->d,0,verticesA,verticesB)*p;
+			w1+=shape.Support(-gjk.m_simplex->c[i]->d,1,verticesA,verticesB)*p;
+		}
+		results.witnesses[0]	=	transA*w0;
+		results.witnesses[1]	=	transA*w1;
+		results.normal			=	w0-w1;
+		results.distance		=	results.normal.length();
+		results.normal			/=	results.distance>GJK_MIN_DISTANCE?results.distance:1;
+		return(true);
+	}
+	else
+	{
+		results.status	=	gjk_status==b3GJK::eStatus::Inside?
+			sResults::Penetrating	:
+		sResults::GJK_Failed	;
+		return(false);
+	}
+}
+
+//
+bool	b3GjkEpaSolver2::Penetration(	const b3Transform&	transA, const b3Transform&	transB,
+										const b3ConvexPolyhedronData* hullA, const b3ConvexPolyhedronData* hullB, 
+										const b3AlignedObjectArray<b3Vector3>& verticesA,
+										const b3AlignedObjectArray<b3Vector3>& verticesB,
+									 const b3Vector3&		guess,
+									 sResults&				results,
+									 bool					usemargins)
+{
+
+	tShape			shape;
+	Initialize(transA,transB,hullA,hullB,verticesA,verticesB,results,shape,usemargins);
+	b3GJK				gjk(verticesA,verticesB);
+	b3GJK::eStatus::_	gjk_status=gjk.Evaluate(shape,guess);
+	switch(gjk_status)
+	{
+	case	b3GJK::eStatus::Inside:
+		{
+			b3EPA				epa;
+			b3EPA::eStatus::_	epa_status=epa.Evaluate(gjk,-guess);
+			if(epa_status!=b3EPA::eStatus::Failed)
+			{
+				b3Vector3	w0=b3MakeVector3(0,0,0);
+				for(unsigned int i=0;i<epa.m_result.rank;++i)
+				{
+					w0+=shape.Support(epa.m_result.c[i]->d,0,verticesA,verticesB)*epa.m_result.p[i];
+				}
+				results.status			=	sResults::Penetrating;
+				results.witnesses[0]	=	transA*w0;
+				results.witnesses[1]	=	transA*(w0-epa.m_normal*epa.m_depth);
+				results.normal			=	-epa.m_normal;
+				results.distance		=	-epa.m_depth;
+				return(true);
+			} else results.status=sResults::EPA_Failed;
+		}
+		break;
+	case	b3GJK::eStatus::Failed:
+		results.status=sResults::GJK_Failed;
+		break;
+		default:
+					{
+					}
+	}
+	return(false);
+}
+
+
+#if 0
+//
+b3Scalar	b3GjkEpaSolver2::SignedDistance(const b3Vector3& position,
+											b3Scalar margin,
+											const b3Transform&	transA,
+											const b3ConvexPolyhedronData& hullA, 
+											const b3AlignedObjectArray<b3Vector3>& verticesA,
+											sResults& results)
+{
+	tShape			shape;
+	btSphereShape	shape1(margin);
+	b3Transform		wtrs1(b3Quaternion(0,0,0,1),position);
+	Initialize(shape0,wtrs0,&shape1,wtrs1,results,shape,false);
+	GJK				gjk;	
+	GJK::eStatus::_	gjk_status=gjk.Evaluate(shape,b3Vector3(1,1,1));
+	if(gjk_status==GJK::eStatus::Valid)
+	{
+		b3Vector3	w0=b3Vector3(0,0,0);
+		b3Vector3	w1=b3Vector3(0,0,0);
+		for(unsigned int i=0;i<gjk.m_simplex->rank;++i)
+		{
+			const b3Scalar	p=gjk.m_simplex->p[i];
+			w0+=shape.Support( gjk.m_simplex->c[i]->d,0)*p;
+			w1+=shape.Support(-gjk.m_simplex->c[i]->d,1)*p;
+		}
+		results.witnesses[0]	=	wtrs0*w0;
+		results.witnesses[1]	=	wtrs0*w1;
+		const b3Vector3	delta=	results.witnesses[1]-
+			results.witnesses[0];
+		const b3Scalar	margin=	shape0->getMarginNonVirtual()+
+			shape1.getMarginNonVirtual();
+		const b3Scalar	length=	delta.length();	
+		results.normal			=	delta/length;
+		results.witnesses[0]	+=	results.normal*margin;
+		return(length-margin);
+	}
+	else
+	{
+		if(gjk_status==GJK::eStatus::Inside)
+		{
+			if(Penetration(shape0,wtrs0,&shape1,wtrs1,gjk.m_ray,results))
+			{
+				const b3Vector3	delta=	results.witnesses[0]-
+					results.witnesses[1];
+				const b3Scalar	length=	delta.length();
+				if (length >= B3_EPSILON)
+					results.normal	=	delta/length;			
+				return(-length);
+			}
+		}	
+	}
+	return(B3_INFINITY);
+}
+
+//
+bool	b3GjkEpaSolver2::SignedDistance(const btConvexShape*	shape0,
+										const b3Transform&		wtrs0,
+										const btConvexShape*	shape1,
+										const b3Transform&		wtrs1,
+										const b3Vector3&		guess,
+										sResults&				results)
+{
+	if(!Distance(shape0,wtrs0,shape1,wtrs1,guess,results))
+		return(Penetration(shape0,wtrs0,shape1,wtrs1,guess,results,false));
+	else
+		return(true);
+}
+#endif
+
+
+/* Symbols cleanup		*/ 
+
+#undef GJK_MAX_ITERATIONS
+#undef GJK_ACCURACY
+#undef GJK_MIN_DISTANCE
+#undef GJK_DUPLICATED_EPS
+#undef GJK_SIMPLEX2_EPS
+#undef GJK_SIMPLEX3_EPS
+#undef GJK_SIMPLEX4_EPS
+
+#undef EPA_MAX_VERTICES
+#undef EPA_MAX_FACES
+#undef EPA_MAX_ITERATIONS
+#undef EPA_ACCURACY
+#undef EPA_FALLBACK
+#undef EPA_PLANE_EPS
+#undef EPA_INSIDE_EPS
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3GjkEpa.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3GjkEpa.h
new file mode 100644
index 0000000000..976238a04c
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3GjkEpa.h
@@ -0,0 +1,82 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2008 Erwin Coumans  http://continuousphysics.com/Bullet/
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the
+use of this software.
+Permission is granted to anyone to use this software for any purpose,
+including commercial applications, and to alter it and redistribute it
+freely,
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not
+claim that you wrote the original software. If you use this software in a
+product, an acknowledgment in the product documentation would be appreciated
+but is not required.
+2. Altered source versions must be plainly marked as such, and must not be
+misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+/*
+GJK-EPA collision solver by Nathanael Presson, 2008
+*/
+#ifndef B3_GJK_EPA2_H
+#define B3_GJK_EPA2_H
+
+#include "Bullet3Common/b3AlignedObjectArray.h"
+#include "Bullet3Common/b3Transform.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3ConvexPolyhedronData.h"
+
+
+///btGjkEpaSolver contributed under zlib by Nathanael Presson
+struct	b3GjkEpaSolver2
+{
+struct	sResults
+	{
+	enum eStatus
+		{
+		Separated,		/* Shapes doesnt penetrate												*/ 
+		Penetrating,	/* Shapes are penetrating												*/ 
+		GJK_Failed,		/* GJK phase fail, no big issue, shapes are probably just 'touching'	*/ 
+		EPA_Failed		/* EPA phase fail, bigger problem, need to save parameters, and debug	*/ 
+		}		status;
+	b3Vector3	witnesses[2];
+	b3Vector3	normal;
+	b3Scalar	distance;
+	};
+
+static int		StackSizeRequirement();
+
+static bool		Distance(	 const b3Transform&	transA, const b3Transform&	transB,
+							const b3ConvexPolyhedronData* hullA, const b3ConvexPolyhedronData* hullB, 
+							const b3AlignedObjectArray<b3Vector3>& verticesA,
+							const b3AlignedObjectArray<b3Vector3>& verticesB,
+							const b3Vector3& guess,
+							sResults& results);
+
+static bool		Penetration( const b3Transform&	transA, const b3Transform&	transB,
+							const b3ConvexPolyhedronData* hullA, const b3ConvexPolyhedronData* hullB, 
+							const b3AlignedObjectArray<b3Vector3>& verticesA,
+							const b3AlignedObjectArray<b3Vector3>& verticesB,
+							const b3Vector3& guess,
+							sResults& results,
+							bool usemargins=true);
+#if 0
+static b3Scalar	SignedDistance(	const b3Vector3& position,
+								b3Scalar margin,
+								const btConvexShape* shape,
+								const btTransform& wtrs,
+								sResults& results);
+							
+static bool		SignedDistance(	const btConvexShape* shape0,const btTransform& wtrs0,
+								const btConvexShape* shape1,const btTransform& wtrs1,
+								const b3Vector3& guess,
+								sResults& results);
+#endif 
+
+};
+
+#endif //B3_GJK_EPA2_H
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3OptimizedBvh.cpp b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3OptimizedBvh.cpp
new file mode 100644
index 0000000000..e9e51d5a36
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3OptimizedBvh.cpp
@@ -0,0 +1,390 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2009 Erwin Coumans  http://bulletphysics.org
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+
+#include "b3OptimizedBvh.h"
+#include "b3StridingMeshInterface.h"
+#include "Bullet3Geometry/b3AabbUtil.h"
+
+
+b3OptimizedBvh::b3OptimizedBvh()
+{ 
+}
+
+b3OptimizedBvh::~b3OptimizedBvh()
+{
+}
+
+
+void b3OptimizedBvh::build(b3StridingMeshInterface* triangles, bool useQuantizedAabbCompression, const b3Vector3& bvhAabbMin, const b3Vector3& bvhAabbMax)
+{
+	m_useQuantization = useQuantizedAabbCompression;
+
+
+	// NodeArray	triangleNodes;
+
+	struct	NodeTriangleCallback : public b3InternalTriangleIndexCallback
+	{
+
+		NodeArray&	m_triangleNodes;
+
+		NodeTriangleCallback& operator=(NodeTriangleCallback& other)
+		{
+			m_triangleNodes.copyFromArray(other.m_triangleNodes);
+			return *this;
+		}
+		
+		NodeTriangleCallback(NodeArray&	triangleNodes)
+			:m_triangleNodes(triangleNodes)
+		{
+		}
+
+		virtual void internalProcessTriangleIndex(b3Vector3* triangle,int partId,int  triangleIndex)
+		{
+			b3OptimizedBvhNode node;
+			b3Vector3	aabbMin,aabbMax;
+			aabbMin.setValue(b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT));
+			aabbMax.setValue(b3Scalar(-B3_LARGE_FLOAT),b3Scalar(-B3_LARGE_FLOAT),b3Scalar(-B3_LARGE_FLOAT)); 
+			aabbMin.setMin(triangle[0]);
+			aabbMax.setMax(triangle[0]);
+			aabbMin.setMin(triangle[1]);
+			aabbMax.setMax(triangle[1]);
+			aabbMin.setMin(triangle[2]);
+			aabbMax.setMax(triangle[2]);
+
+			//with quantization?
+			node.m_aabbMinOrg = aabbMin;
+			node.m_aabbMaxOrg = aabbMax;
+
+			node.m_escapeIndex = -1;
+	
+			//for child nodes
+			node.m_subPart = partId;
+			node.m_triangleIndex = triangleIndex;
+			m_triangleNodes.push_back(node);
+		}
+	};
+	struct	QuantizedNodeTriangleCallback : public b3InternalTriangleIndexCallback
+	{
+		QuantizedNodeArray&	m_triangleNodes;
+		const b3QuantizedBvh* m_optimizedTree; // for quantization
+
+		QuantizedNodeTriangleCallback& operator=(QuantizedNodeTriangleCallback& other)
+		{
+			m_triangleNodes.copyFromArray(other.m_triangleNodes);
+			m_optimizedTree = other.m_optimizedTree;
+			return *this;
+		}
+
+		QuantizedNodeTriangleCallback(QuantizedNodeArray&	triangleNodes,const b3QuantizedBvh* tree)
+			:m_triangleNodes(triangleNodes),m_optimizedTree(tree)
+		{
+		}
+
+		virtual void internalProcessTriangleIndex(b3Vector3* triangle,int partId,int  triangleIndex)
+		{
+			// The partId and triangle index must fit in the same (positive) integer
+			b3Assert(partId < (1<<MAX_NUM_PARTS_IN_BITS));
+			b3Assert(triangleIndex < (1<<(31-MAX_NUM_PARTS_IN_BITS)));
+			//negative indices are reserved for escapeIndex
+			b3Assert(triangleIndex>=0);
+
+			b3QuantizedBvhNode node;
+			b3Vector3	aabbMin,aabbMax;
+			aabbMin.setValue(b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT));
+			aabbMax.setValue(b3Scalar(-B3_LARGE_FLOAT),b3Scalar(-B3_LARGE_FLOAT),b3Scalar(-B3_LARGE_FLOAT)); 
+			aabbMin.setMin(triangle[0]);
+			aabbMax.setMax(triangle[0]);
+			aabbMin.setMin(triangle[1]);
+			aabbMax.setMax(triangle[1]);
+			aabbMin.setMin(triangle[2]);
+			aabbMax.setMax(triangle[2]);
+
+			//PCK: add these checks for zero dimensions of aabb
+			const b3Scalar MIN_AABB_DIMENSION = b3Scalar(0.002);
+			const b3Scalar MIN_AABB_HALF_DIMENSION = b3Scalar(0.001);
+			if (aabbMax.getX() - aabbMin.getX() < MIN_AABB_DIMENSION)
+			{
+				aabbMax.setX(aabbMax.getX() + MIN_AABB_HALF_DIMENSION);
+				aabbMin.setX(aabbMin.getX() - MIN_AABB_HALF_DIMENSION);
+			}
+			if (aabbMax.getY() - aabbMin.getY() < MIN_AABB_DIMENSION)
+			{
+				aabbMax.setY(aabbMax.getY() + MIN_AABB_HALF_DIMENSION);
+				aabbMin.setY(aabbMin.getY() - MIN_AABB_HALF_DIMENSION);
+			}
+			if (aabbMax.getZ() - aabbMin.getZ() < MIN_AABB_DIMENSION)
+			{
+				aabbMax.setZ(aabbMax.getZ() + MIN_AABB_HALF_DIMENSION);
+				aabbMin.setZ(aabbMin.getZ() - MIN_AABB_HALF_DIMENSION);
+			}
+
+			m_optimizedTree->quantize(&node.m_quantizedAabbMin[0],aabbMin,0);
+			m_optimizedTree->quantize(&node.m_quantizedAabbMax[0],aabbMax,1);
+
+			node.m_escapeIndexOrTriangleIndex = (partId<<(31-MAX_NUM_PARTS_IN_BITS)) | triangleIndex;
+
+			m_triangleNodes.push_back(node);
+		}
+	};
+	
+
+
+	int numLeafNodes = 0;
+
+	
+	if (m_useQuantization)
+	{
+
+		//initialize quantization values
+		setQuantizationValues(bvhAabbMin,bvhAabbMax);
+
+		QuantizedNodeTriangleCallback	callback(m_quantizedLeafNodes,this);
+
+	
+		triangles->InternalProcessAllTriangles(&callback,m_bvhAabbMin,m_bvhAabbMax);
+
+		//now we have an array of leafnodes in m_leafNodes
+		numLeafNodes = m_quantizedLeafNodes.size();
+
+
+		m_quantizedContiguousNodes.resize(2*numLeafNodes);
+
+
+	} else
+	{
+		NodeTriangleCallback	callback(m_leafNodes);
+
+		b3Vector3 aabbMin=b3MakeVector3(b3Scalar(-B3_LARGE_FLOAT),b3Scalar(-B3_LARGE_FLOAT),b3Scalar(-B3_LARGE_FLOAT));
+		b3Vector3 aabbMax=b3MakeVector3(b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT));
+
+		triangles->InternalProcessAllTriangles(&callback,aabbMin,aabbMax);
+
+		//now we have an array of leafnodes in m_leafNodes
+		numLeafNodes = m_leafNodes.size();
+
+		m_contiguousNodes.resize(2*numLeafNodes);
+	}
+
+	m_curNodeIndex = 0;
+
+	buildTree(0,numLeafNodes);
+
+	///if the entire tree is small then subtree size, we need to create a header info for the tree
+	if(m_useQuantization && !m_SubtreeHeaders.size())
+	{
+		b3BvhSubtreeInfo& subtree = m_SubtreeHeaders.expand();
+		subtree.setAabbFromQuantizeNode(m_quantizedContiguousNodes[0]);
+		subtree.m_rootNodeIndex = 0;
+		subtree.m_subtreeSize = m_quantizedContiguousNodes[0].isLeafNode() ? 1 : m_quantizedContiguousNodes[0].getEscapeIndex();
+	}
+
+	//PCK: update the copy of the size
+	m_subtreeHeaderCount = m_SubtreeHeaders.size();
+
+	//PCK: clear m_quantizedLeafNodes and m_leafNodes, they are temporary
+	m_quantizedLeafNodes.clear();
+	m_leafNodes.clear();
+}
+
+
+
+
+void	b3OptimizedBvh::refit(b3StridingMeshInterface* meshInterface,const b3Vector3& aabbMin,const b3Vector3& aabbMax)
+{
+	if (m_useQuantization)
+	{
+
+		setQuantizationValues(aabbMin,aabbMax);
+
+		updateBvhNodes(meshInterface,0,m_curNodeIndex,0);
+
+		///now update all subtree headers
+
+		int i;
+		for (i=0;i<m_SubtreeHeaders.size();i++)
+		{
+			b3BvhSubtreeInfo& subtree = m_SubtreeHeaders[i];
+			subtree.setAabbFromQuantizeNode(m_quantizedContiguousNodes[subtree.m_rootNodeIndex]);
+		}
+
+	} else
+	{
+
+	}
+}
+
+
+
+
+void	b3OptimizedBvh::refitPartial(b3StridingMeshInterface* meshInterface,const b3Vector3& aabbMin,const b3Vector3& aabbMax)
+{
+	//incrementally initialize quantization values
+	b3Assert(m_useQuantization);
+
+	b3Assert(aabbMin.getX() > m_bvhAabbMin.getX());
+	b3Assert(aabbMin.getY() > m_bvhAabbMin.getY());
+	b3Assert(aabbMin.getZ() > m_bvhAabbMin.getZ());
+
+	b3Assert(aabbMax.getX() < m_bvhAabbMax.getX());
+	b3Assert(aabbMax.getY() < m_bvhAabbMax.getY());
+	b3Assert(aabbMax.getZ() < m_bvhAabbMax.getZ());
+
+	///we should update all quantization values, using updateBvhNodes(meshInterface);
+	///but we only update chunks that overlap the given aabb
+	
+	unsigned short	quantizedQueryAabbMin[3];
+	unsigned short	quantizedQueryAabbMax[3];
+
+	quantize(&quantizedQueryAabbMin[0],aabbMin,0);
+	quantize(&quantizedQueryAabbMax[0],aabbMax,1);
+
+	int i;
+	for (i=0;i<this->m_SubtreeHeaders.size();i++)
+	{
+		b3BvhSubtreeInfo& subtree = m_SubtreeHeaders[i];
+
+		//PCK: unsigned instead of bool
+		unsigned overlap = b3TestQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,subtree.m_quantizedAabbMin,subtree.m_quantizedAabbMax);
+		if (overlap != 0)
+		{
+			updateBvhNodes(meshInterface,subtree.m_rootNodeIndex,subtree.m_rootNodeIndex+subtree.m_subtreeSize,i);
+
+			subtree.setAabbFromQuantizeNode(m_quantizedContiguousNodes[subtree.m_rootNodeIndex]);
+		}
+	}
+	
+}
+
+void	b3OptimizedBvh::updateBvhNodes(b3StridingMeshInterface* meshInterface,int firstNode,int endNode,int index)
+{
+	(void)index;
+
+	b3Assert(m_useQuantization);
+
+	int curNodeSubPart=-1;
+
+	//get access info to trianglemesh data
+		const unsigned char *vertexbase = 0;
+		int numverts = 0;
+		PHY_ScalarType type = PHY_INTEGER;
+		int stride = 0;
+		const unsigned char *indexbase = 0;
+		int indexstride = 0;
+		int numfaces = 0;
+		PHY_ScalarType indicestype = PHY_INTEGER;
+
+		b3Vector3	triangleVerts[3];
+		b3Vector3	aabbMin,aabbMax;
+		const b3Vector3& meshScaling = meshInterface->getScaling();
+		
+		int i;
+		for (i=endNode-1;i>=firstNode;i--)
+		{
+
+
+			b3QuantizedBvhNode& curNode = m_quantizedContiguousNodes[i];
+			if (curNode.isLeafNode())
+			{
+				//recalc aabb from triangle data
+				int nodeSubPart = curNode.getPartId();
+				int nodeTriangleIndex = curNode.getTriangleIndex();
+				if (nodeSubPart != curNodeSubPart)
+				{
+					if (curNodeSubPart >= 0)
+						meshInterface->unLockReadOnlyVertexBase(curNodeSubPart);
+					meshInterface->getLockedReadOnlyVertexIndexBase(&vertexbase,numverts,	type,stride,&indexbase,indexstride,numfaces,indicestype,nodeSubPart);
+
+					curNodeSubPart = nodeSubPart;
+					b3Assert(indicestype==PHY_INTEGER||indicestype==PHY_SHORT);
+				}
+				//triangles->getLockedReadOnlyVertexIndexBase(vertexBase,numVerts,
+
+				unsigned int* gfxbase = (unsigned int*)(indexbase+nodeTriangleIndex*indexstride);
+				
+				
+				for (int j=2;j>=0;j--)
+				{
+					
+					int graphicsindex = indicestype==PHY_SHORT?((unsigned short*)gfxbase)[j]:gfxbase[j];
+					if (type == PHY_FLOAT)
+					{
+						float* graphicsbase = (float*)(vertexbase+graphicsindex*stride);
+						triangleVerts[j] = b3MakeVector3(
+							graphicsbase[0]*meshScaling.getX(),
+							graphicsbase[1]*meshScaling.getY(),
+							graphicsbase[2]*meshScaling.getZ());
+					}
+					else
+					{
+						double* graphicsbase = (double*)(vertexbase+graphicsindex*stride);
+						triangleVerts[j] = b3MakeVector3( b3Scalar(graphicsbase[0]*meshScaling.getX()), b3Scalar(graphicsbase[1]*meshScaling.getY()), b3Scalar(graphicsbase[2]*meshScaling.getZ()));
+					}
+				}
+
+
+				
+				aabbMin.setValue(b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT));
+				aabbMax.setValue(b3Scalar(-B3_LARGE_FLOAT),b3Scalar(-B3_LARGE_FLOAT),b3Scalar(-B3_LARGE_FLOAT)); 
+				aabbMin.setMin(triangleVerts[0]);
+				aabbMax.setMax(triangleVerts[0]);
+				aabbMin.setMin(triangleVerts[1]);
+				aabbMax.setMax(triangleVerts[1]);
+				aabbMin.setMin(triangleVerts[2]);
+				aabbMax.setMax(triangleVerts[2]);
+
+				quantize(&curNode.m_quantizedAabbMin[0],aabbMin,0);
+				quantize(&curNode.m_quantizedAabbMax[0],aabbMax,1);
+				
+			} else
+			{
+				//combine aabb from both children
+
+				b3QuantizedBvhNode* leftChildNode = &m_quantizedContiguousNodes[i+1];
+				
+				b3QuantizedBvhNode* rightChildNode = leftChildNode->isLeafNode() ? &m_quantizedContiguousNodes[i+2] :
+					&m_quantizedContiguousNodes[i+1+leftChildNode->getEscapeIndex()];
+				
+
+				{
+					for (int i=0;i<3;i++)
+					{
+						curNode.m_quantizedAabbMin[i] = leftChildNode->m_quantizedAabbMin[i];
+						if (curNode.m_quantizedAabbMin[i]>rightChildNode->m_quantizedAabbMin[i])
+							curNode.m_quantizedAabbMin[i]=rightChildNode->m_quantizedAabbMin[i];
+
+						curNode.m_quantizedAabbMax[i] = leftChildNode->m_quantizedAabbMax[i];
+						if (curNode.m_quantizedAabbMax[i] < rightChildNode->m_quantizedAabbMax[i])
+							curNode.m_quantizedAabbMax[i] = rightChildNode->m_quantizedAabbMax[i];
+					}
+				}
+			}
+
+		}
+
+		if (curNodeSubPart >= 0)
+			meshInterface->unLockReadOnlyVertexBase(curNodeSubPart);
+
+		
+}
+
+///deSerializeInPlace loads and initializes a BVH from a buffer in memory 'in place'
+b3OptimizedBvh* b3OptimizedBvh::deSerializeInPlace(void *i_alignedDataBuffer, unsigned int i_dataBufferSize, bool i_swapEndian)
+{
+	b3QuantizedBvh* bvh = b3QuantizedBvh::deSerializeInPlace(i_alignedDataBuffer,i_dataBufferSize,i_swapEndian);
+	
+	//we don't add additional data so just do a static upcast
+	return static_cast<b3OptimizedBvh*>(bvh);
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3OptimizedBvh.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3OptimizedBvh.h
new file mode 100644
index 0000000000..0272ef83bf
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3OptimizedBvh.h
@@ -0,0 +1,65 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2009 Erwin Coumans  http://bulletphysics.org
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+///Contains contributions from Disney Studio's
+
+#ifndef B3_OPTIMIZED_BVH_H
+#define B3_OPTIMIZED_BVH_H
+
+#include "b3QuantizedBvh.h"
+
+class b3StridingMeshInterface;
+
+
+///The b3OptimizedBvh extends the b3QuantizedBvh to create AABB tree for triangle meshes, through the b3StridingMeshInterface.
+B3_ATTRIBUTE_ALIGNED16(class) b3OptimizedBvh : public b3QuantizedBvh
+{
+	
+public:
+	B3_DECLARE_ALIGNED_ALLOCATOR();
+
+protected:
+
+public:
+
+	b3OptimizedBvh();
+
+	virtual ~b3OptimizedBvh();
+
+	void	build(b3StridingMeshInterface* triangles,bool useQuantizedAabbCompression, const b3Vector3& bvhAabbMin, const b3Vector3& bvhAabbMax);
+
+	void	refit(b3StridingMeshInterface* triangles,const b3Vector3& aabbMin,const b3Vector3& aabbMax);
+
+	void	refitPartial(b3StridingMeshInterface* triangles,const b3Vector3& aabbMin, const b3Vector3& aabbMax);
+
+	void	updateBvhNodes(b3StridingMeshInterface* meshInterface,int firstNode,int endNode,int index);
+
+	/// Data buffer MUST be 16 byte aligned
+	virtual bool serializeInPlace(void *o_alignedDataBuffer, unsigned i_dataBufferSize, bool i_swapEndian) const
+	{
+		return b3QuantizedBvh::serialize(o_alignedDataBuffer,i_dataBufferSize,i_swapEndian);
+
+	}
+
+	///deSerializeInPlace loads and initializes a BVH from a buffer in memory 'in place'
+	static b3OptimizedBvh *deSerializeInPlace(void *i_alignedDataBuffer, unsigned int i_dataBufferSize, bool i_swapEndian);
+
+
+};
+
+
+#endif //B3_OPTIMIZED_BVH_H
+
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3QuantizedBvh.cpp b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3QuantizedBvh.cpp
new file mode 100644
index 0000000000..52027e1118
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3QuantizedBvh.cpp
@@ -0,0 +1,1301 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2006 Erwin Coumans  http://continuousphysics.com/Bullet/
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+#include "b3QuantizedBvh.h"
+
+#include "Bullet3Geometry/b3AabbUtil.h"
+
+
+#define RAYAABB2
+
+b3QuantizedBvh::b3QuantizedBvh() : 
+					m_bulletVersion(B3_BULLET_VERSION),
+					m_useQuantization(false), 
+					m_traversalMode(TRAVERSAL_STACKLESS_CACHE_FRIENDLY)
+					//m_traversalMode(TRAVERSAL_STACKLESS)
+					//m_traversalMode(TRAVERSAL_RECURSIVE)
+					,m_subtreeHeaderCount(0) //PCK: add this line
+{
+	m_bvhAabbMin.setValue(-B3_INFINITY,-B3_INFINITY,-B3_INFINITY);
+	m_bvhAabbMax.setValue(B3_INFINITY,B3_INFINITY,B3_INFINITY);
+}
+
+
+
+
+
+void b3QuantizedBvh::buildInternal()
+{
+	///assumes that caller filled in the m_quantizedLeafNodes
+	m_useQuantization = true;
+	int numLeafNodes = 0;
+	
+	if (m_useQuantization)
+	{
+		//now we have an array of leafnodes in m_leafNodes
+		numLeafNodes = m_quantizedLeafNodes.size();
+
+		m_quantizedContiguousNodes.resize(2*numLeafNodes);
+
+	}
+
+	m_curNodeIndex = 0;
+
+	buildTree(0,numLeafNodes);
+
+	///if the entire tree is small then subtree size, we need to create a header info for the tree
+	if(m_useQuantization && !m_SubtreeHeaders.size())
+	{
+		b3BvhSubtreeInfo& subtree = m_SubtreeHeaders.expand();
+		subtree.setAabbFromQuantizeNode(m_quantizedContiguousNodes[0]);
+		subtree.m_rootNodeIndex = 0;
+		subtree.m_subtreeSize = m_quantizedContiguousNodes[0].isLeafNode() ? 1 : m_quantizedContiguousNodes[0].getEscapeIndex();
+	}
+
+	//PCK: update the copy of the size
+	m_subtreeHeaderCount = m_SubtreeHeaders.size();
+
+	//PCK: clear m_quantizedLeafNodes and m_leafNodes, they are temporary
+	m_quantizedLeafNodes.clear();
+	m_leafNodes.clear();
+}
+
+
+
+///just for debugging, to visualize the individual patches/subtrees
+#ifdef DEBUG_PATCH_COLORS
+b3Vector3 color[4]=
+{
+	b3Vector3(1,0,0),
+	b3Vector3(0,1,0),
+	b3Vector3(0,0,1),
+	b3Vector3(0,1,1)
+};
+#endif //DEBUG_PATCH_COLORS
+
+
+
+void	b3QuantizedBvh::setQuantizationValues(const b3Vector3& bvhAabbMin,const b3Vector3& bvhAabbMax,b3Scalar quantizationMargin)
+{
+	//enlarge the AABB to avoid division by zero when initializing the quantization values
+	b3Vector3 clampValue =b3MakeVector3(quantizationMargin,quantizationMargin,quantizationMargin);
+	m_bvhAabbMin = bvhAabbMin - clampValue;
+	m_bvhAabbMax = bvhAabbMax + clampValue;
+	b3Vector3 aabbSize = m_bvhAabbMax - m_bvhAabbMin;
+	m_bvhQuantization = b3MakeVector3(b3Scalar(65533.0),b3Scalar(65533.0),b3Scalar(65533.0)) / aabbSize;
+	m_useQuantization = true;
+}
+
+
+
+
+b3QuantizedBvh::~b3QuantizedBvh()
+{
+}
+
+#ifdef DEBUG_TREE_BUILDING
+int gStackDepth = 0;
+int gMaxStackDepth = 0;
+#endif //DEBUG_TREE_BUILDING
+
+void	b3QuantizedBvh::buildTree	(int startIndex,int endIndex)
+{
+#ifdef DEBUG_TREE_BUILDING
+	gStackDepth++;
+	if (gStackDepth > gMaxStackDepth)
+		gMaxStackDepth = gStackDepth;
+#endif //DEBUG_TREE_BUILDING
+
+
+	int splitAxis, splitIndex, i;
+	int numIndices =endIndex-startIndex;
+	int curIndex = m_curNodeIndex;
+
+	b3Assert(numIndices>0);
+
+	if (numIndices==1)
+	{
+#ifdef DEBUG_TREE_BUILDING
+		gStackDepth--;
+#endif //DEBUG_TREE_BUILDING
+		
+		assignInternalNodeFromLeafNode(m_curNodeIndex,startIndex);
+
+		m_curNodeIndex++;
+		return;	
+	}
+	//calculate Best Splitting Axis and where to split it. Sort the incoming 'leafNodes' array within range 'startIndex/endIndex'.
+	
+	splitAxis = calcSplittingAxis(startIndex,endIndex);
+
+	splitIndex = sortAndCalcSplittingIndex(startIndex,endIndex,splitAxis);
+
+	int internalNodeIndex = m_curNodeIndex;
+	
+	//set the min aabb to 'inf' or a max value, and set the max aabb to a -inf/minimum value.
+	//the aabb will be expanded during buildTree/mergeInternalNodeAabb with actual node values
+	setInternalNodeAabbMin(m_curNodeIndex,m_bvhAabbMax);//can't use b3Vector3(B3_INFINITY,B3_INFINITY,B3_INFINITY)) because of quantization
+	setInternalNodeAabbMax(m_curNodeIndex,m_bvhAabbMin);//can't use b3Vector3(-B3_INFINITY,-B3_INFINITY,-B3_INFINITY)) because of quantization
+	
+	
+	for (i=startIndex;i<endIndex;i++)
+	{
+		mergeInternalNodeAabb(m_curNodeIndex,getAabbMin(i),getAabbMax(i));
+	}
+
+	m_curNodeIndex++;
+	
+
+	//internalNode->m_escapeIndex;
+	
+	int leftChildNodexIndex = m_curNodeIndex;
+
+	//build left child tree
+	buildTree(startIndex,splitIndex);
+
+	int rightChildNodexIndex = m_curNodeIndex;
+	//build right child tree
+	buildTree(splitIndex,endIndex);
+
+#ifdef DEBUG_TREE_BUILDING
+	gStackDepth--;
+#endif //DEBUG_TREE_BUILDING
+
+	int escapeIndex = m_curNodeIndex - curIndex;
+
+	if (m_useQuantization)
+	{
+		//escapeIndex is the number of nodes of this subtree
+		const int sizeQuantizedNode =sizeof(b3QuantizedBvhNode);
+		const int treeSizeInBytes = escapeIndex * sizeQuantizedNode;
+		if (treeSizeInBytes > MAX_SUBTREE_SIZE_IN_BYTES)
+		{
+			updateSubtreeHeaders(leftChildNodexIndex,rightChildNodexIndex);
+		}
+	} else
+	{
+
+	}
+
+	setInternalNodeEscapeIndex(internalNodeIndex,escapeIndex);
+
+}
+
+void	b3QuantizedBvh::updateSubtreeHeaders(int leftChildNodexIndex,int rightChildNodexIndex)
+{
+	b3Assert(m_useQuantization);
+
+	b3QuantizedBvhNode& leftChildNode = m_quantizedContiguousNodes[leftChildNodexIndex];
+	int leftSubTreeSize = leftChildNode.isLeafNode() ? 1 : leftChildNode.getEscapeIndex();
+	int leftSubTreeSizeInBytes =  leftSubTreeSize * static_cast<int>(sizeof(b3QuantizedBvhNode));
+	
+	b3QuantizedBvhNode& rightChildNode = m_quantizedContiguousNodes[rightChildNodexIndex];
+	int rightSubTreeSize = rightChildNode.isLeafNode() ? 1 : rightChildNode.getEscapeIndex();
+	int rightSubTreeSizeInBytes =  rightSubTreeSize *  static_cast<int>(sizeof(b3QuantizedBvhNode));
+
+	if(leftSubTreeSizeInBytes <= MAX_SUBTREE_SIZE_IN_BYTES)
+	{
+		b3BvhSubtreeInfo& subtree = m_SubtreeHeaders.expand();
+		subtree.setAabbFromQuantizeNode(leftChildNode);
+		subtree.m_rootNodeIndex = leftChildNodexIndex;
+		subtree.m_subtreeSize = leftSubTreeSize;
+	}
+
+	if(rightSubTreeSizeInBytes <= MAX_SUBTREE_SIZE_IN_BYTES)
+	{
+		b3BvhSubtreeInfo& subtree = m_SubtreeHeaders.expand();
+		subtree.setAabbFromQuantizeNode(rightChildNode);
+		subtree.m_rootNodeIndex = rightChildNodexIndex;
+		subtree.m_subtreeSize = rightSubTreeSize;
+	}
+
+	//PCK: update the copy of the size
+	m_subtreeHeaderCount = m_SubtreeHeaders.size();
+}
+
+
+int	b3QuantizedBvh::sortAndCalcSplittingIndex(int startIndex,int endIndex,int splitAxis)
+{
+	int i;
+	int splitIndex =startIndex;
+	int numIndices = endIndex - startIndex;
+	b3Scalar splitValue;
+
+	b3Vector3 means=b3MakeVector3(b3Scalar(0.),b3Scalar(0.),b3Scalar(0.));
+	for (i=startIndex;i<endIndex;i++)
+	{
+		b3Vector3 center = b3Scalar(0.5)*(getAabbMax(i)+getAabbMin(i));
+		means+=center;
+	}
+	means *= (b3Scalar(1.)/(b3Scalar)numIndices);
+	
+	splitValue = means[splitAxis];
+	
+	//sort leafNodes so all values larger then splitValue comes first, and smaller values start from 'splitIndex'.
+	for (i=startIndex;i<endIndex;i++)
+	{
+		b3Vector3 center = b3Scalar(0.5)*(getAabbMax(i)+getAabbMin(i));
+		if (center[splitAxis] > splitValue)
+		{
+			//swap
+			swapLeafNodes(i,splitIndex);
+			splitIndex++;
+		}
+	}
+
+	//if the splitIndex causes unbalanced trees, fix this by using the center in between startIndex and endIndex
+	//otherwise the tree-building might fail due to stack-overflows in certain cases.
+	//unbalanced1 is unsafe: it can cause stack overflows
+	//bool unbalanced1 = ((splitIndex==startIndex) || (splitIndex == (endIndex-1)));
+
+	//unbalanced2 should work too: always use center (perfect balanced trees)	
+	//bool unbalanced2 = true;
+
+	//this should be safe too:
+	int rangeBalancedIndices = numIndices/3;
+	bool unbalanced = ((splitIndex<=(startIndex+rangeBalancedIndices)) || (splitIndex >=(endIndex-1-rangeBalancedIndices)));
+	
+	if (unbalanced)
+	{
+		splitIndex = startIndex+ (numIndices>>1);
+	}
+
+	bool unbal = (splitIndex==startIndex) || (splitIndex == (endIndex));
+	(void)unbal;
+	b3Assert(!unbal);
+
+	return splitIndex;
+}
+
+
+int	b3QuantizedBvh::calcSplittingAxis(int startIndex,int endIndex)
+{
+	int i;
+
+	b3Vector3 means=b3MakeVector3(b3Scalar(0.),b3Scalar(0.),b3Scalar(0.));
+	b3Vector3 variance=b3MakeVector3(b3Scalar(0.),b3Scalar(0.),b3Scalar(0.));
+	int numIndices = endIndex-startIndex;
+
+	for (i=startIndex;i<endIndex;i++)
+	{
+		b3Vector3 center = b3Scalar(0.5)*(getAabbMax(i)+getAabbMin(i));
+		means+=center;
+	}
+	means *= (b3Scalar(1.)/(b3Scalar)numIndices);
+		
+	for (i=startIndex;i<endIndex;i++)
+	{
+		b3Vector3 center = b3Scalar(0.5)*(getAabbMax(i)+getAabbMin(i));
+		b3Vector3 diff2 = center-means;
+		diff2 = diff2 * diff2;
+		variance += diff2;
+	}
+	variance *= (b3Scalar(1.)/	((b3Scalar)numIndices-1)	);
+	
+	return variance.maxAxis();
+}
+
+
+
+void	b3QuantizedBvh::reportAabbOverlappingNodex(b3NodeOverlapCallback* nodeCallback,const b3Vector3& aabbMin,const b3Vector3& aabbMax) const
+{
+	//either choose recursive traversal (walkTree) or stackless (walkStacklessTree)
+
+	if (m_useQuantization)
+	{
+		///quantize query AABB
+		unsigned short int quantizedQueryAabbMin[3];
+		unsigned short int quantizedQueryAabbMax[3];
+		quantizeWithClamp(quantizedQueryAabbMin,aabbMin,0);
+		quantizeWithClamp(quantizedQueryAabbMax,aabbMax,1);
+
+		switch (m_traversalMode)
+		{
+		case TRAVERSAL_STACKLESS:
+				walkStacklessQuantizedTree(nodeCallback,quantizedQueryAabbMin,quantizedQueryAabbMax,0,m_curNodeIndex);
+			break;
+		case TRAVERSAL_STACKLESS_CACHE_FRIENDLY:
+				walkStacklessQuantizedTreeCacheFriendly(nodeCallback,quantizedQueryAabbMin,quantizedQueryAabbMax);
+			break;
+		case TRAVERSAL_RECURSIVE:
+			{
+				const b3QuantizedBvhNode* rootNode = &m_quantizedContiguousNodes[0];
+				walkRecursiveQuantizedTreeAgainstQueryAabb(rootNode,nodeCallback,quantizedQueryAabbMin,quantizedQueryAabbMax);
+			}
+			break;
+		default:
+			//unsupported
+			b3Assert(0);
+		}
+	} else
+	{
+		walkStacklessTree(nodeCallback,aabbMin,aabbMax);
+	}
+}
+
+
+static int b3s_maxIterations = 0;
+
+
+void	b3QuantizedBvh::walkStacklessTree(b3NodeOverlapCallback* nodeCallback,const b3Vector3& aabbMin,const b3Vector3& aabbMax) const
+{
+	b3Assert(!m_useQuantization);
+
+	const b3OptimizedBvhNode* rootNode = &m_contiguousNodes[0];
+	int escapeIndex, curIndex = 0;
+	int walkIterations = 0;
+	bool isLeafNode;
+	//PCK: unsigned instead of bool
+	unsigned aabbOverlap;
+
+	while (curIndex < m_curNodeIndex)
+	{
+		//catch bugs in tree data
+		b3Assert (walkIterations < m_curNodeIndex);
+
+		walkIterations++;
+		aabbOverlap = b3TestAabbAgainstAabb2(aabbMin,aabbMax,rootNode->m_aabbMinOrg,rootNode->m_aabbMaxOrg);
+		isLeafNode = rootNode->m_escapeIndex == -1;
+		
+		//PCK: unsigned instead of bool
+		if (isLeafNode && (aabbOverlap != 0))
+		{
+			nodeCallback->processNode(rootNode->m_subPart,rootNode->m_triangleIndex);
+		} 
+		
+		//PCK: unsigned instead of bool
+		if ((aabbOverlap != 0) || isLeafNode)
+		{
+			rootNode++;
+			curIndex++;
+		} else
+		{
+			escapeIndex = rootNode->m_escapeIndex;
+			rootNode += escapeIndex;
+			curIndex += escapeIndex;
+		}
+	}
+	if (b3s_maxIterations < walkIterations)
+		b3s_maxIterations = walkIterations;
+
+}
+
+/*
+///this was the original recursive traversal, before we optimized towards stackless traversal
+void	b3QuantizedBvh::walkTree(b3OptimizedBvhNode* rootNode,b3NodeOverlapCallback* nodeCallback,const b3Vector3& aabbMin,const b3Vector3& aabbMax) const
+{
+	bool isLeafNode, aabbOverlap = TestAabbAgainstAabb2(aabbMin,aabbMax,rootNode->m_aabbMin,rootNode->m_aabbMax);
+	if (aabbOverlap)
+	{
+		isLeafNode = (!rootNode->m_leftChild && !rootNode->m_rightChild);
+		if (isLeafNode)
+		{
+			nodeCallback->processNode(rootNode);
+		} else
+		{
+			walkTree(rootNode->m_leftChild,nodeCallback,aabbMin,aabbMax);
+			walkTree(rootNode->m_rightChild,nodeCallback,aabbMin,aabbMax);
+		}
+	}
+
+}
+*/
+
+void b3QuantizedBvh::walkRecursiveQuantizedTreeAgainstQueryAabb(const b3QuantizedBvhNode* currentNode,b3NodeOverlapCallback* nodeCallback,unsigned short int* quantizedQueryAabbMin,unsigned short int* quantizedQueryAabbMax) const
+{
+	b3Assert(m_useQuantization);
+	
+	bool isLeafNode;
+	//PCK: unsigned instead of bool
+	unsigned aabbOverlap;
+
+	//PCK: unsigned instead of bool
+	aabbOverlap = b3TestQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,currentNode->m_quantizedAabbMin,currentNode->m_quantizedAabbMax);
+	isLeafNode = currentNode->isLeafNode();
+		
+	//PCK: unsigned instead of bool
+	if (aabbOverlap != 0)
+	{
+		if (isLeafNode)
+		{
+			nodeCallback->processNode(currentNode->getPartId(),currentNode->getTriangleIndex());
+		} else
+		{
+			//process left and right children
+			const b3QuantizedBvhNode* leftChildNode = currentNode+1;
+			walkRecursiveQuantizedTreeAgainstQueryAabb(leftChildNode,nodeCallback,quantizedQueryAabbMin,quantizedQueryAabbMax);
+
+			const b3QuantizedBvhNode* rightChildNode = leftChildNode->isLeafNode() ? leftChildNode+1:leftChildNode+leftChildNode->getEscapeIndex();
+			walkRecursiveQuantizedTreeAgainstQueryAabb(rightChildNode,nodeCallback,quantizedQueryAabbMin,quantizedQueryAabbMax);
+		}
+	}		
+}
+
+
+
+void	b3QuantizedBvh::walkStacklessTreeAgainstRay(b3NodeOverlapCallback* nodeCallback, const b3Vector3& raySource, const b3Vector3& rayTarget, const b3Vector3& aabbMin, const b3Vector3& aabbMax, int startNodeIndex,int endNodeIndex) const
+{
+	b3Assert(!m_useQuantization);
+
+	const b3OptimizedBvhNode* rootNode = &m_contiguousNodes[0];
+	int escapeIndex, curIndex = 0;
+	int walkIterations = 0;
+	bool isLeafNode;
+	//PCK: unsigned instead of bool
+	unsigned aabbOverlap=0;
+	unsigned rayBoxOverlap=0;
+	b3Scalar lambda_max = 1.0;
+	
+		/* Quick pruning by quantized box */
+	b3Vector3 rayAabbMin = raySource;
+	b3Vector3 rayAabbMax = raySource;
+	rayAabbMin.setMin(rayTarget);
+	rayAabbMax.setMax(rayTarget);
+
+	/* Add box cast extents to bounding box */
+	rayAabbMin += aabbMin;
+	rayAabbMax += aabbMax;
+
+#ifdef RAYAABB2
+	b3Vector3 rayDir = (rayTarget-raySource);
+	rayDir.normalize ();
+	lambda_max = rayDir.dot(rayTarget-raySource);
+	///what about division by zero? --> just set rayDirection[i] to 1.0
+	b3Vector3 rayDirectionInverse;
+	rayDirectionInverse[0] = rayDir[0] == b3Scalar(0.0) ? b3Scalar(B3_LARGE_FLOAT) : b3Scalar(1.0) / rayDir[0];
+	rayDirectionInverse[1] = rayDir[1] == b3Scalar(0.0) ? b3Scalar(B3_LARGE_FLOAT) : b3Scalar(1.0) / rayDir[1];
+	rayDirectionInverse[2] = rayDir[2] == b3Scalar(0.0) ? b3Scalar(B3_LARGE_FLOAT) : b3Scalar(1.0) / rayDir[2];
+	unsigned int sign[3] = { rayDirectionInverse[0] < 0.0, rayDirectionInverse[1] < 0.0, rayDirectionInverse[2] < 0.0};
+#endif
+
+	b3Vector3 bounds[2];
+
+	while (curIndex < m_curNodeIndex)
+	{
+		b3Scalar param = 1.0;
+		//catch bugs in tree data
+		b3Assert (walkIterations < m_curNodeIndex);
+
+		walkIterations++;
+
+		bounds[0] = rootNode->m_aabbMinOrg;
+		bounds[1] = rootNode->m_aabbMaxOrg;
+		/* Add box cast extents */
+		bounds[0] -= aabbMax;
+		bounds[1] -= aabbMin;
+
+		aabbOverlap = b3TestAabbAgainstAabb2(rayAabbMin,rayAabbMax,rootNode->m_aabbMinOrg,rootNode->m_aabbMaxOrg);
+		//perhaps profile if it is worth doing the aabbOverlap test first
+
+#ifdef RAYAABB2
+			///careful with this check: need to check division by zero (above) and fix the unQuantize method
+			///thanks Joerg/hiker for the reproduction case!
+			///http://www.bulletphysics.com/Bullet/phpBB3/viewtopic.php?f=9&t=1858
+		rayBoxOverlap = aabbOverlap ? b3RayAabb2 (raySource, rayDirectionInverse, sign, bounds, param, 0.0f, lambda_max) : false;
+
+#else
+		b3Vector3 normal;
+		rayBoxOverlap = b3RayAabb(raySource, rayTarget,bounds[0],bounds[1],param, normal);
+#endif
+
+		isLeafNode = rootNode->m_escapeIndex == -1;
+		
+		//PCK: unsigned instead of bool
+		if (isLeafNode && (rayBoxOverlap != 0))
+		{
+			nodeCallback->processNode(rootNode->m_subPart,rootNode->m_triangleIndex);
+		} 
+		
+		//PCK: unsigned instead of bool
+		if ((rayBoxOverlap != 0) || isLeafNode)
+		{
+			rootNode++;
+			curIndex++;
+		} else
+		{
+			escapeIndex = rootNode->m_escapeIndex;
+			rootNode += escapeIndex;
+			curIndex += escapeIndex;
+		}
+	}
+	if (b3s_maxIterations < walkIterations)
+		b3s_maxIterations = walkIterations;
+
+}
+
+
+
+void	b3QuantizedBvh::walkStacklessQuantizedTreeAgainstRay(b3NodeOverlapCallback* nodeCallback, const b3Vector3& raySource, const b3Vector3& rayTarget, const b3Vector3& aabbMin, const b3Vector3& aabbMax, int startNodeIndex,int endNodeIndex) const
+{
+	b3Assert(m_useQuantization);
+	
+	int curIndex = startNodeIndex;
+	int walkIterations = 0;
+	int subTreeSize = endNodeIndex - startNodeIndex;
+	(void)subTreeSize;
+
+	const b3QuantizedBvhNode* rootNode = &m_quantizedContiguousNodes[startNodeIndex];
+	int escapeIndex;
+	
+	bool isLeafNode;
+	//PCK: unsigned instead of bool
+	unsigned boxBoxOverlap = 0;
+	unsigned rayBoxOverlap = 0;
+
+	b3Scalar lambda_max = 1.0;
+
+#ifdef RAYAABB2
+	b3Vector3 rayDirection = (rayTarget-raySource);
+	rayDirection.normalize ();
+	lambda_max = rayDirection.dot(rayTarget-raySource);
+	///what about division by zero? --> just set rayDirection[i] to 1.0
+	rayDirection[0] = rayDirection[0] == b3Scalar(0.0) ? b3Scalar(B3_LARGE_FLOAT) : b3Scalar(1.0) / rayDirection[0];
+	rayDirection[1] = rayDirection[1] == b3Scalar(0.0) ? b3Scalar(B3_LARGE_FLOAT) : b3Scalar(1.0) / rayDirection[1];
+	rayDirection[2] = rayDirection[2] == b3Scalar(0.0) ? b3Scalar(B3_LARGE_FLOAT) : b3Scalar(1.0) / rayDirection[2];
+	unsigned int sign[3] = { rayDirection[0] < 0.0, rayDirection[1] < 0.0, rayDirection[2] < 0.0};
+#endif
+
+	/* Quick pruning by quantized box */
+	b3Vector3 rayAabbMin = raySource;
+	b3Vector3 rayAabbMax = raySource;
+	rayAabbMin.setMin(rayTarget);
+	rayAabbMax.setMax(rayTarget);
+
+	/* Add box cast extents to bounding box */
+	rayAabbMin += aabbMin;
+	rayAabbMax += aabbMax;
+
+	unsigned short int quantizedQueryAabbMin[3];
+	unsigned short int quantizedQueryAabbMax[3];
+	quantizeWithClamp(quantizedQueryAabbMin,rayAabbMin,0);
+	quantizeWithClamp(quantizedQueryAabbMax,rayAabbMax,1);
+
+	while (curIndex < endNodeIndex)
+	{
+
+//#define VISUALLY_ANALYZE_BVH 1
+#ifdef VISUALLY_ANALYZE_BVH
+		//some code snippet to debugDraw aabb, to visually analyze bvh structure
+		static int drawPatch = 0;
+		//need some global access to a debugDrawer
+		extern b3IDebugDraw* debugDrawerPtr;
+		if (curIndex==drawPatch)
+		{
+			b3Vector3 aabbMin,aabbMax;
+			aabbMin = unQuantize(rootNode->m_quantizedAabbMin);
+			aabbMax = unQuantize(rootNode->m_quantizedAabbMax);
+			b3Vector3	color(1,0,0);
+			debugDrawerPtr->drawAabb(aabbMin,aabbMax,color);
+		}
+#endif//VISUALLY_ANALYZE_BVH
+
+		//catch bugs in tree data
+		b3Assert (walkIterations < subTreeSize);
+
+		walkIterations++;
+		//PCK: unsigned instead of bool
+		// only interested if this is closer than any previous hit
+		b3Scalar param = 1.0;
+		rayBoxOverlap = 0;
+		boxBoxOverlap = b3TestQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,rootNode->m_quantizedAabbMin,rootNode->m_quantizedAabbMax);
+		isLeafNode = rootNode->isLeafNode();
+		if (boxBoxOverlap)
+		{
+			b3Vector3 bounds[2];
+			bounds[0] = unQuantize(rootNode->m_quantizedAabbMin);
+			bounds[1] = unQuantize(rootNode->m_quantizedAabbMax);
+			/* Add box cast extents */
+			bounds[0] -= aabbMax;
+			bounds[1] -= aabbMin;
+#if 0
+			b3Vector3 normal;
+			bool ra2 = b3RayAabb2 (raySource, rayDirection, sign, bounds, param, 0.0, lambda_max);
+			bool ra = b3RayAabb (raySource, rayTarget, bounds[0], bounds[1], param, normal);
+			if (ra2 != ra)
+			{
+				printf("functions don't match\n");
+			}
+#endif
+#ifdef RAYAABB2
+			///careful with this check: need to check division by zero (above) and fix the unQuantize method
+			///thanks Joerg/hiker for the reproduction case!
+			///http://www.bulletphysics.com/Bullet/phpBB3/viewtopic.php?f=9&t=1858
+
+			//B3_PROFILE("b3RayAabb2");
+			rayBoxOverlap = b3RayAabb2 (raySource, rayDirection, sign, bounds, param, 0.0f, lambda_max);
+			
+#else
+			rayBoxOverlap = true;//b3RayAabb(raySource, rayTarget, bounds[0], bounds[1], param, normal);
+#endif
+		}
+		
+		if (isLeafNode && rayBoxOverlap)
+		{
+			nodeCallback->processNode(rootNode->getPartId(),rootNode->getTriangleIndex());
+		}
+		
+		//PCK: unsigned instead of bool
+		if ((rayBoxOverlap != 0) || isLeafNode)
+		{
+			rootNode++;
+			curIndex++;
+		} else
+		{
+			escapeIndex = rootNode->getEscapeIndex();
+			rootNode += escapeIndex;
+			curIndex += escapeIndex;
+		}
+	}
+	if (b3s_maxIterations < walkIterations)
+		b3s_maxIterations = walkIterations;
+
+}
+
+void	b3QuantizedBvh::walkStacklessQuantizedTree(b3NodeOverlapCallback* nodeCallback,unsigned short int* quantizedQueryAabbMin,unsigned short int* quantizedQueryAabbMax,int startNodeIndex,int endNodeIndex) const
+{
+	b3Assert(m_useQuantization);
+	
+	int curIndex = startNodeIndex;
+	int walkIterations = 0;
+	int subTreeSize = endNodeIndex - startNodeIndex;
+	(void)subTreeSize;
+
+	const b3QuantizedBvhNode* rootNode = &m_quantizedContiguousNodes[startNodeIndex];
+	int escapeIndex;
+	
+	bool isLeafNode;
+	//PCK: unsigned instead of bool
+	unsigned aabbOverlap;
+
+	while (curIndex < endNodeIndex)
+	{
+
+//#define VISUALLY_ANALYZE_BVH 1
+#ifdef VISUALLY_ANALYZE_BVH
+		//some code snippet to debugDraw aabb, to visually analyze bvh structure
+		static int drawPatch = 0;
+		//need some global access to a debugDrawer
+		extern b3IDebugDraw* debugDrawerPtr;
+		if (curIndex==drawPatch)
+		{
+			b3Vector3 aabbMin,aabbMax;
+			aabbMin = unQuantize(rootNode->m_quantizedAabbMin);
+			aabbMax = unQuantize(rootNode->m_quantizedAabbMax);
+			b3Vector3	color(1,0,0);
+			debugDrawerPtr->drawAabb(aabbMin,aabbMax,color);
+		}
+#endif//VISUALLY_ANALYZE_BVH
+
+		//catch bugs in tree data
+		b3Assert (walkIterations < subTreeSize);
+
+		walkIterations++;
+		//PCK: unsigned instead of bool
+		aabbOverlap = b3TestQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,rootNode->m_quantizedAabbMin,rootNode->m_quantizedAabbMax);
+		isLeafNode = rootNode->isLeafNode();
+		
+		if (isLeafNode && aabbOverlap)
+		{
+			nodeCallback->processNode(rootNode->getPartId(),rootNode->getTriangleIndex());
+		} 
+		
+		//PCK: unsigned instead of bool
+		if ((aabbOverlap != 0) || isLeafNode)
+		{
+			rootNode++;
+			curIndex++;
+		} else
+		{
+			escapeIndex = rootNode->getEscapeIndex();
+			rootNode += escapeIndex;
+			curIndex += escapeIndex;
+		}
+	}
+	if (b3s_maxIterations < walkIterations)
+		b3s_maxIterations = walkIterations;
+
+}
+
+//This traversal can be called from Playstation 3 SPU
+void	b3QuantizedBvh::walkStacklessQuantizedTreeCacheFriendly(b3NodeOverlapCallback* nodeCallback,unsigned short int* quantizedQueryAabbMin,unsigned short int* quantizedQueryAabbMax) const
+{
+	b3Assert(m_useQuantization);
+
+	int i;
+
+
+	for (i=0;i<this->m_SubtreeHeaders.size();i++)
+	{
+		const b3BvhSubtreeInfo& subtree = m_SubtreeHeaders[i];
+
+		//PCK: unsigned instead of bool
+		unsigned overlap = b3TestQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,subtree.m_quantizedAabbMin,subtree.m_quantizedAabbMax);
+		if (overlap != 0)
+		{
+			walkStacklessQuantizedTree(nodeCallback,quantizedQueryAabbMin,quantizedQueryAabbMax,
+				subtree.m_rootNodeIndex,
+				subtree.m_rootNodeIndex+subtree.m_subtreeSize);
+		}
+	}
+}
+
+
+void	b3QuantizedBvh::reportRayOverlappingNodex (b3NodeOverlapCallback* nodeCallback, const b3Vector3& raySource, const b3Vector3& rayTarget) const
+{
+	reportBoxCastOverlappingNodex(nodeCallback,raySource,rayTarget,b3MakeVector3(0,0,0),b3MakeVector3(0,0,0));
+}
+
+
+void	b3QuantizedBvh::reportBoxCastOverlappingNodex(b3NodeOverlapCallback* nodeCallback, const b3Vector3& raySource, const b3Vector3& rayTarget, const b3Vector3& aabbMin,const b3Vector3& aabbMax) const
+{
+	//always use stackless
+
+	if (m_useQuantization)
+	{
+		walkStacklessQuantizedTreeAgainstRay(nodeCallback, raySource, rayTarget, aabbMin, aabbMax, 0, m_curNodeIndex);
+	}
+	else
+	{
+		walkStacklessTreeAgainstRay(nodeCallback, raySource, rayTarget, aabbMin, aabbMax, 0, m_curNodeIndex);
+	}
+	/*
+	{
+		//recursive traversal
+		b3Vector3 qaabbMin = raySource;
+		b3Vector3 qaabbMax = raySource;
+		qaabbMin.setMin(rayTarget);
+		qaabbMax.setMax(rayTarget);
+		qaabbMin += aabbMin;
+		qaabbMax += aabbMax;
+		reportAabbOverlappingNodex(nodeCallback,qaabbMin,qaabbMax);
+	}
+	*/
+
+}
+
+
+void	b3QuantizedBvh::swapLeafNodes(int i,int splitIndex)
+{
+	if (m_useQuantization)
+	{
+			b3QuantizedBvhNode tmp = m_quantizedLeafNodes[i];
+			m_quantizedLeafNodes[i] = m_quantizedLeafNodes[splitIndex];
+			m_quantizedLeafNodes[splitIndex] = tmp;
+	} else
+	{
+			b3OptimizedBvhNode tmp = m_leafNodes[i];
+			m_leafNodes[i] = m_leafNodes[splitIndex];
+			m_leafNodes[splitIndex] = tmp;
+	}
+}
+
+void	b3QuantizedBvh::assignInternalNodeFromLeafNode(int internalNode,int leafNodeIndex)
+{
+	if (m_useQuantization)
+	{
+		m_quantizedContiguousNodes[internalNode] = m_quantizedLeafNodes[leafNodeIndex];
+	} else
+	{
+		m_contiguousNodes[internalNode] = m_leafNodes[leafNodeIndex];
+	}
+}
+
+//PCK: include
+#include <new>
+
+#if 0
+//PCK: consts
+static const unsigned BVH_ALIGNMENT = 16;
+static const unsigned BVH_ALIGNMENT_MASK = BVH_ALIGNMENT-1;
+
+static const unsigned BVH_ALIGNMENT_BLOCKS = 2;
+#endif
+
+
+unsigned int b3QuantizedBvh::getAlignmentSerializationPadding()
+{
+	// I changed this to 0 since the extra padding is not needed or used.
+	return 0;//BVH_ALIGNMENT_BLOCKS * BVH_ALIGNMENT;
+}
+
+unsigned b3QuantizedBvh::calculateSerializeBufferSize() const
+{
+	unsigned baseSize = sizeof(b3QuantizedBvh) + getAlignmentSerializationPadding();
+	baseSize += sizeof(b3BvhSubtreeInfo) * m_subtreeHeaderCount;
+	if (m_useQuantization)
+	{
+		return baseSize + m_curNodeIndex * sizeof(b3QuantizedBvhNode);
+	}
+	return baseSize + m_curNodeIndex * sizeof(b3OptimizedBvhNode);
+}
+
+bool b3QuantizedBvh::serialize(void *o_alignedDataBuffer, unsigned /*i_dataBufferSize */, bool i_swapEndian) const
+{
+	b3Assert(m_subtreeHeaderCount == m_SubtreeHeaders.size());
+	m_subtreeHeaderCount = m_SubtreeHeaders.size();
+
+/*	if (i_dataBufferSize < calculateSerializeBufferSize() || o_alignedDataBuffer == NULL || (((unsigned)o_alignedDataBuffer & BVH_ALIGNMENT_MASK) != 0))
+	{
+		///check alignedment for buffer?
+		b3Assert(0);
+		return false;
+	}
+*/
+
+	b3QuantizedBvh *targetBvh = (b3QuantizedBvh *)o_alignedDataBuffer;
+
+	// construct the class so the virtual function table, etc will be set up
+	// Also, m_leafNodes and m_quantizedLeafNodes will be initialized to default values by the constructor
+	new (targetBvh) b3QuantizedBvh;
+
+	if (i_swapEndian)
+	{
+		targetBvh->m_curNodeIndex = static_cast<int>(b3SwapEndian(m_curNodeIndex));
+
+
+		b3SwapVector3Endian(m_bvhAabbMin,targetBvh->m_bvhAabbMin);
+		b3SwapVector3Endian(m_bvhAabbMax,targetBvh->m_bvhAabbMax);
+		b3SwapVector3Endian(m_bvhQuantization,targetBvh->m_bvhQuantization);
+
+		targetBvh->m_traversalMode = (b3TraversalMode)b3SwapEndian(m_traversalMode);
+		targetBvh->m_subtreeHeaderCount = static_cast<int>(b3SwapEndian(m_subtreeHeaderCount));
+	}
+	else
+	{
+		targetBvh->m_curNodeIndex = m_curNodeIndex;
+		targetBvh->m_bvhAabbMin = m_bvhAabbMin;
+		targetBvh->m_bvhAabbMax = m_bvhAabbMax;
+		targetBvh->m_bvhQuantization = m_bvhQuantization;
+		targetBvh->m_traversalMode = m_traversalMode;
+		targetBvh->m_subtreeHeaderCount = m_subtreeHeaderCount;
+	}
+
+	targetBvh->m_useQuantization = m_useQuantization;
+
+	unsigned char *nodeData = (unsigned char *)targetBvh;
+	nodeData += sizeof(b3QuantizedBvh);
+	
+	unsigned sizeToAdd = 0;//(BVH_ALIGNMENT-((unsigned)nodeData & BVH_ALIGNMENT_MASK))&BVH_ALIGNMENT_MASK;
+	nodeData += sizeToAdd;
+	
+	int nodeCount = m_curNodeIndex;
+
+	if (m_useQuantization)
+	{
+		targetBvh->m_quantizedContiguousNodes.initializeFromBuffer(nodeData, nodeCount, nodeCount);
+
+		if (i_swapEndian)
+		{
+			for (int nodeIndex = 0; nodeIndex < nodeCount; nodeIndex++)
+			{
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[0] = b3SwapEndian(m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[0]);
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[1] = b3SwapEndian(m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[1]);
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[2] = b3SwapEndian(m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[2]);
+
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[0] = b3SwapEndian(m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[0]);
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[1] = b3SwapEndian(m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[1]);
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[2] = b3SwapEndian(m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[2]);
+
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_escapeIndexOrTriangleIndex = static_cast<int>(b3SwapEndian(m_quantizedContiguousNodes[nodeIndex].m_escapeIndexOrTriangleIndex));
+			}
+		}
+		else
+		{
+			for (int nodeIndex = 0; nodeIndex < nodeCount; nodeIndex++)
+			{
+	
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[0] = m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[0];
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[1] = m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[1];
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[2] = m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[2];
+
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[0] = m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[0];
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[1] = m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[1];
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[2] = m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[2];
+
+				targetBvh->m_quantizedContiguousNodes[nodeIndex].m_escapeIndexOrTriangleIndex = m_quantizedContiguousNodes[nodeIndex].m_escapeIndexOrTriangleIndex;
+
+
+			}
+		}
+		nodeData += sizeof(b3QuantizedBvhNode) * nodeCount;
+
+		// this clears the pointer in the member variable it doesn't really do anything to the data
+		// it does call the destructor on the contained objects, but they are all classes with no destructor defined
+		// so the memory (which is not freed) is left alone
+		targetBvh->m_quantizedContiguousNodes.initializeFromBuffer(NULL, 0, 0);
+	}
+	else
+	{
+		targetBvh->m_contiguousNodes.initializeFromBuffer(nodeData, nodeCount, nodeCount);
+
+		if (i_swapEndian)
+		{
+			for (int nodeIndex = 0; nodeIndex < nodeCount; nodeIndex++)
+			{
+				b3SwapVector3Endian(m_contiguousNodes[nodeIndex].m_aabbMinOrg, targetBvh->m_contiguousNodes[nodeIndex].m_aabbMinOrg);
+				b3SwapVector3Endian(m_contiguousNodes[nodeIndex].m_aabbMaxOrg, targetBvh->m_contiguousNodes[nodeIndex].m_aabbMaxOrg);
+
+				targetBvh->m_contiguousNodes[nodeIndex].m_escapeIndex = static_cast<int>(b3SwapEndian(m_contiguousNodes[nodeIndex].m_escapeIndex));
+				targetBvh->m_contiguousNodes[nodeIndex].m_subPart = static_cast<int>(b3SwapEndian(m_contiguousNodes[nodeIndex].m_subPart));
+				targetBvh->m_contiguousNodes[nodeIndex].m_triangleIndex = static_cast<int>(b3SwapEndian(m_contiguousNodes[nodeIndex].m_triangleIndex));
+			}
+		}
+		else
+		{
+			for (int nodeIndex = 0; nodeIndex < nodeCount; nodeIndex++)
+			{
+				targetBvh->m_contiguousNodes[nodeIndex].m_aabbMinOrg = m_contiguousNodes[nodeIndex].m_aabbMinOrg;
+				targetBvh->m_contiguousNodes[nodeIndex].m_aabbMaxOrg = m_contiguousNodes[nodeIndex].m_aabbMaxOrg;
+
+				targetBvh->m_contiguousNodes[nodeIndex].m_escapeIndex = m_contiguousNodes[nodeIndex].m_escapeIndex;
+				targetBvh->m_contiguousNodes[nodeIndex].m_subPart = m_contiguousNodes[nodeIndex].m_subPart;
+				targetBvh->m_contiguousNodes[nodeIndex].m_triangleIndex = m_contiguousNodes[nodeIndex].m_triangleIndex;
+			}
+		}
+		nodeData += sizeof(b3OptimizedBvhNode) * nodeCount;
+
+		// this clears the pointer in the member variable it doesn't really do anything to the data
+		// it does call the destructor on the contained objects, but they are all classes with no destructor defined
+		// so the memory (which is not freed) is left alone
+		targetBvh->m_contiguousNodes.initializeFromBuffer(NULL, 0, 0);
+	}
+
+	sizeToAdd = 0;//(BVH_ALIGNMENT-((unsigned)nodeData & BVH_ALIGNMENT_MASK))&BVH_ALIGNMENT_MASK;
+	nodeData += sizeToAdd;
+
+	// Now serialize the subtree headers
+	targetBvh->m_SubtreeHeaders.initializeFromBuffer(nodeData, m_subtreeHeaderCount, m_subtreeHeaderCount);
+	if (i_swapEndian)
+	{
+		for (int i = 0; i < m_subtreeHeaderCount; i++)
+		{
+			targetBvh->m_SubtreeHeaders[i].m_quantizedAabbMin[0] = b3SwapEndian(m_SubtreeHeaders[i].m_quantizedAabbMin[0]);
+			targetBvh->m_SubtreeHeaders[i].m_quantizedAabbMin[1] = b3SwapEndian(m_SubtreeHeaders[i].m_quantizedAabbMin[1]);
+			targetBvh->m_SubtreeHeaders[i].m_quantizedAabbMin[2] = b3SwapEndian(m_SubtreeHeaders[i].m_quantizedAabbMin[2]);
+
+			targetBvh->m_SubtreeHeaders[i].m_quantizedAabbMax[0] = b3SwapEndian(m_SubtreeHeaders[i].m_quantizedAabbMax[0]);
+			targetBvh->m_SubtreeHeaders[i].m_quantizedAabbMax[1] = b3SwapEndian(m_SubtreeHeaders[i].m_quantizedAabbMax[1]);
+			targetBvh->m_SubtreeHeaders[i].m_quantizedAabbMax[2] = b3SwapEndian(m_SubtreeHeaders[i].m_quantizedAabbMax[2]);
+
+			targetBvh->m_SubtreeHeaders[i].m_rootNodeIndex = static_cast<int>(b3SwapEndian(m_SubtreeHeaders[i].m_rootNodeIndex));
+			targetBvh->m_SubtreeHeaders[i].m_subtreeSize = static_cast<int>(b3SwapEndian(m_SubtreeHeaders[i].m_subtreeSize));
+		}
+	}
+	else
+	{
+		for (int i = 0; i < m_subtreeHeaderCount; i++)
+		{
+			targetBvh->m_SubtreeHeaders[i].m_quantizedAabbMin[0] = (m_SubtreeHeaders[i].m_quantizedAabbMin[0]);
+			targetBvh->m_SubtreeHeaders[i].m_quantizedAabbMin[1] = (m_SubtreeHeaders[i].m_quantizedAabbMin[1]);
+			targetBvh->m_SubtreeHeaders[i].m_quantizedAabbMin[2] = (m_SubtreeHeaders[i].m_quantizedAabbMin[2]);
+
+			targetBvh->m_SubtreeHeaders[i].m_quantizedAabbMax[0] = (m_SubtreeHeaders[i].m_quantizedAabbMax[0]);
+			targetBvh->m_SubtreeHeaders[i].m_quantizedAabbMax[1] = (m_SubtreeHeaders[i].m_quantizedAabbMax[1]);
+			targetBvh->m_SubtreeHeaders[i].m_quantizedAabbMax[2] = (m_SubtreeHeaders[i].m_quantizedAabbMax[2]);
+
+			targetBvh->m_SubtreeHeaders[i].m_rootNodeIndex = (m_SubtreeHeaders[i].m_rootNodeIndex);
+			targetBvh->m_SubtreeHeaders[i].m_subtreeSize = (m_SubtreeHeaders[i].m_subtreeSize);
+
+			// need to clear padding in destination buffer
+			targetBvh->m_SubtreeHeaders[i].m_padding[0] = 0;
+			targetBvh->m_SubtreeHeaders[i].m_padding[1] = 0;
+			targetBvh->m_SubtreeHeaders[i].m_padding[2] = 0;
+		}
+	}
+	nodeData += sizeof(b3BvhSubtreeInfo) * m_subtreeHeaderCount;
+
+	// this clears the pointer in the member variable it doesn't really do anything to the data
+	// it does call the destructor on the contained objects, but they are all classes with no destructor defined
+	// so the memory (which is not freed) is left alone
+	targetBvh->m_SubtreeHeaders.initializeFromBuffer(NULL, 0, 0);
+
+	// this wipes the virtual function table pointer at the start of the buffer for the class
+	*((void**)o_alignedDataBuffer) = NULL;
+
+	return true;
+}
+
+b3QuantizedBvh *b3QuantizedBvh::deSerializeInPlace(void *i_alignedDataBuffer, unsigned int i_dataBufferSize, bool i_swapEndian)
+{
+
+	if (i_alignedDataBuffer == NULL)// || (((unsigned)i_alignedDataBuffer & BVH_ALIGNMENT_MASK) != 0))
+	{
+		return NULL;
+	}
+	b3QuantizedBvh *bvh = (b3QuantizedBvh *)i_alignedDataBuffer;
+
+	if (i_swapEndian)
+	{
+		bvh->m_curNodeIndex = static_cast<int>(b3SwapEndian(bvh->m_curNodeIndex));
+
+		b3UnSwapVector3Endian(bvh->m_bvhAabbMin);
+		b3UnSwapVector3Endian(bvh->m_bvhAabbMax);
+		b3UnSwapVector3Endian(bvh->m_bvhQuantization);
+
+		bvh->m_traversalMode = (b3TraversalMode)b3SwapEndian(bvh->m_traversalMode);
+		bvh->m_subtreeHeaderCount = static_cast<int>(b3SwapEndian(bvh->m_subtreeHeaderCount));
+	}
+
+	unsigned int calculatedBufSize = bvh->calculateSerializeBufferSize();
+	b3Assert(calculatedBufSize <= i_dataBufferSize);
+
+	if (calculatedBufSize > i_dataBufferSize)
+	{
+		return NULL;
+	}
+
+	unsigned char *nodeData = (unsigned char *)bvh;
+	nodeData += sizeof(b3QuantizedBvh);
+	
+	unsigned sizeToAdd = 0;//(BVH_ALIGNMENT-((unsigned)nodeData & BVH_ALIGNMENT_MASK))&BVH_ALIGNMENT_MASK;
+	nodeData += sizeToAdd;
+	
+	int nodeCount = bvh->m_curNodeIndex;
+
+	// Must call placement new to fill in virtual function table, etc, but we don't want to overwrite most data, so call a special version of the constructor
+	// Also, m_leafNodes and m_quantizedLeafNodes will be initialized to default values by the constructor
+	new (bvh) b3QuantizedBvh(*bvh, false);
+
+	if (bvh->m_useQuantization)
+	{
+		bvh->m_quantizedContiguousNodes.initializeFromBuffer(nodeData, nodeCount, nodeCount);
+
+		if (i_swapEndian)
+		{
+			for (int nodeIndex = 0; nodeIndex < nodeCount; nodeIndex++)
+			{
+				bvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[0] = b3SwapEndian(bvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[0]);
+				bvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[1] = b3SwapEndian(bvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[1]);
+				bvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[2] = b3SwapEndian(bvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[2]);
+
+				bvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[0] = b3SwapEndian(bvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[0]);
+				bvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[1] = b3SwapEndian(bvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[1]);
+				bvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[2] = b3SwapEndian(bvh->m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[2]);
+
+				bvh->m_quantizedContiguousNodes[nodeIndex].m_escapeIndexOrTriangleIndex = static_cast<int>(b3SwapEndian(bvh->m_quantizedContiguousNodes[nodeIndex].m_escapeIndexOrTriangleIndex));
+			}
+		}
+		nodeData += sizeof(b3QuantizedBvhNode) * nodeCount;
+	}
+	else
+	{
+		bvh->m_contiguousNodes.initializeFromBuffer(nodeData, nodeCount, nodeCount);
+
+		if (i_swapEndian)
+		{
+			for (int nodeIndex = 0; nodeIndex < nodeCount; nodeIndex++)
+			{
+				b3UnSwapVector3Endian(bvh->m_contiguousNodes[nodeIndex].m_aabbMinOrg);
+				b3UnSwapVector3Endian(bvh->m_contiguousNodes[nodeIndex].m_aabbMaxOrg);
+				
+				bvh->m_contiguousNodes[nodeIndex].m_escapeIndex = static_cast<int>(b3SwapEndian(bvh->m_contiguousNodes[nodeIndex].m_escapeIndex));
+				bvh->m_contiguousNodes[nodeIndex].m_subPart = static_cast<int>(b3SwapEndian(bvh->m_contiguousNodes[nodeIndex].m_subPart));
+				bvh->m_contiguousNodes[nodeIndex].m_triangleIndex = static_cast<int>(b3SwapEndian(bvh->m_contiguousNodes[nodeIndex].m_triangleIndex));
+			}
+		}
+		nodeData += sizeof(b3OptimizedBvhNode) * nodeCount;
+	}
+
+	sizeToAdd = 0;//(BVH_ALIGNMENT-((unsigned)nodeData & BVH_ALIGNMENT_MASK))&BVH_ALIGNMENT_MASK;
+	nodeData += sizeToAdd;
+
+	// Now serialize the subtree headers
+	bvh->m_SubtreeHeaders.initializeFromBuffer(nodeData, bvh->m_subtreeHeaderCount, bvh->m_subtreeHeaderCount);
+	if (i_swapEndian)
+	{
+		for (int i = 0; i < bvh->m_subtreeHeaderCount; i++)
+		{
+			bvh->m_SubtreeHeaders[i].m_quantizedAabbMin[0] = b3SwapEndian(bvh->m_SubtreeHeaders[i].m_quantizedAabbMin[0]);
+			bvh->m_SubtreeHeaders[i].m_quantizedAabbMin[1] = b3SwapEndian(bvh->m_SubtreeHeaders[i].m_quantizedAabbMin[1]);
+			bvh->m_SubtreeHeaders[i].m_quantizedAabbMin[2] = b3SwapEndian(bvh->m_SubtreeHeaders[i].m_quantizedAabbMin[2]);
+
+			bvh->m_SubtreeHeaders[i].m_quantizedAabbMax[0] = b3SwapEndian(bvh->m_SubtreeHeaders[i].m_quantizedAabbMax[0]);
+			bvh->m_SubtreeHeaders[i].m_quantizedAabbMax[1] = b3SwapEndian(bvh->m_SubtreeHeaders[i].m_quantizedAabbMax[1]);
+			bvh->m_SubtreeHeaders[i].m_quantizedAabbMax[2] = b3SwapEndian(bvh->m_SubtreeHeaders[i].m_quantizedAabbMax[2]);
+
+			bvh->m_SubtreeHeaders[i].m_rootNodeIndex = static_cast<int>(b3SwapEndian(bvh->m_SubtreeHeaders[i].m_rootNodeIndex));
+			bvh->m_SubtreeHeaders[i].m_subtreeSize = static_cast<int>(b3SwapEndian(bvh->m_SubtreeHeaders[i].m_subtreeSize));
+		}
+	}
+
+	return bvh;
+}
+
+// Constructor that prevents b3Vector3's default constructor from being called
+b3QuantizedBvh::b3QuantizedBvh(b3QuantizedBvh &self, bool /* ownsMemory */) :
+m_bvhAabbMin(self.m_bvhAabbMin),
+m_bvhAabbMax(self.m_bvhAabbMax),
+m_bvhQuantization(self.m_bvhQuantization),
+m_bulletVersion(B3_BULLET_VERSION)
+{
+
+}
+
+void b3QuantizedBvh::deSerializeFloat(struct b3QuantizedBvhFloatData& quantizedBvhFloatData)
+{
+	m_bvhAabbMax.deSerializeFloat(quantizedBvhFloatData.m_bvhAabbMax);
+	m_bvhAabbMin.deSerializeFloat(quantizedBvhFloatData.m_bvhAabbMin);
+	m_bvhQuantization.deSerializeFloat(quantizedBvhFloatData.m_bvhQuantization);
+
+	m_curNodeIndex = quantizedBvhFloatData.m_curNodeIndex;
+	m_useQuantization = quantizedBvhFloatData.m_useQuantization!=0;
+	
+	{
+		int numElem = quantizedBvhFloatData.m_numContiguousLeafNodes;
+		m_contiguousNodes.resize(numElem);
+
+		if (numElem)
+		{
+			b3OptimizedBvhNodeFloatData* memPtr = quantizedBvhFloatData.m_contiguousNodesPtr;
+
+			for (int i=0;i<numElem;i++,memPtr++)
+			{
+				m_contiguousNodes[i].m_aabbMaxOrg.deSerializeFloat(memPtr->m_aabbMaxOrg);
+				m_contiguousNodes[i].m_aabbMinOrg.deSerializeFloat(memPtr->m_aabbMinOrg);
+				m_contiguousNodes[i].m_escapeIndex = memPtr->m_escapeIndex;
+				m_contiguousNodes[i].m_subPart = memPtr->m_subPart;
+				m_contiguousNodes[i].m_triangleIndex = memPtr->m_triangleIndex;
+			}
+		}
+	}
+
+	{
+		int numElem = quantizedBvhFloatData.m_numQuantizedContiguousNodes;
+		m_quantizedContiguousNodes.resize(numElem);
+		
+		if (numElem)
+		{
+			b3QuantizedBvhNodeData* memPtr = quantizedBvhFloatData.m_quantizedContiguousNodesPtr;
+			for (int i=0;i<numElem;i++,memPtr++)
+			{
+				m_quantizedContiguousNodes[i].m_escapeIndexOrTriangleIndex = memPtr->m_escapeIndexOrTriangleIndex;
+				m_quantizedContiguousNodes[i].m_quantizedAabbMax[0] = memPtr->m_quantizedAabbMax[0];
+				m_quantizedContiguousNodes[i].m_quantizedAabbMax[1] = memPtr->m_quantizedAabbMax[1];
+				m_quantizedContiguousNodes[i].m_quantizedAabbMax[2] = memPtr->m_quantizedAabbMax[2];
+				m_quantizedContiguousNodes[i].m_quantizedAabbMin[0] = memPtr->m_quantizedAabbMin[0];
+				m_quantizedContiguousNodes[i].m_quantizedAabbMin[1] = memPtr->m_quantizedAabbMin[1];
+				m_quantizedContiguousNodes[i].m_quantizedAabbMin[2] = memPtr->m_quantizedAabbMin[2];
+			}
+		}
+	}
+
+	m_traversalMode = b3TraversalMode(quantizedBvhFloatData.m_traversalMode);
+	
+	{
+		int numElem = quantizedBvhFloatData.m_numSubtreeHeaders;
+		m_SubtreeHeaders.resize(numElem);
+		if (numElem)
+		{
+			b3BvhSubtreeInfoData* memPtr = quantizedBvhFloatData.m_subTreeInfoPtr;
+			for (int i=0;i<numElem;i++,memPtr++)
+			{
+				m_SubtreeHeaders[i].m_quantizedAabbMax[0] = memPtr->m_quantizedAabbMax[0] ;
+				m_SubtreeHeaders[i].m_quantizedAabbMax[1] = memPtr->m_quantizedAabbMax[1];
+				m_SubtreeHeaders[i].m_quantizedAabbMax[2] = memPtr->m_quantizedAabbMax[2];
+				m_SubtreeHeaders[i].m_quantizedAabbMin[0] = memPtr->m_quantizedAabbMin[0];
+				m_SubtreeHeaders[i].m_quantizedAabbMin[1] = memPtr->m_quantizedAabbMin[1];
+				m_SubtreeHeaders[i].m_quantizedAabbMin[2] = memPtr->m_quantizedAabbMin[2];
+				m_SubtreeHeaders[i].m_rootNodeIndex = memPtr->m_rootNodeIndex;
+				m_SubtreeHeaders[i].m_subtreeSize = memPtr->m_subtreeSize;
+			}
+		}
+	}
+}
+
+void b3QuantizedBvh::deSerializeDouble(struct b3QuantizedBvhDoubleData& quantizedBvhDoubleData)
+{
+	m_bvhAabbMax.deSerializeDouble(quantizedBvhDoubleData.m_bvhAabbMax);
+	m_bvhAabbMin.deSerializeDouble(quantizedBvhDoubleData.m_bvhAabbMin);
+	m_bvhQuantization.deSerializeDouble(quantizedBvhDoubleData.m_bvhQuantization);
+
+	m_curNodeIndex = quantizedBvhDoubleData.m_curNodeIndex;
+	m_useQuantization = quantizedBvhDoubleData.m_useQuantization!=0;
+	
+	{
+		int numElem = quantizedBvhDoubleData.m_numContiguousLeafNodes;
+		m_contiguousNodes.resize(numElem);
+
+		if (numElem)
+		{
+			b3OptimizedBvhNodeDoubleData* memPtr = quantizedBvhDoubleData.m_contiguousNodesPtr;
+
+			for (int i=0;i<numElem;i++,memPtr++)
+			{
+				m_contiguousNodes[i].m_aabbMaxOrg.deSerializeDouble(memPtr->m_aabbMaxOrg);
+				m_contiguousNodes[i].m_aabbMinOrg.deSerializeDouble(memPtr->m_aabbMinOrg);
+				m_contiguousNodes[i].m_escapeIndex = memPtr->m_escapeIndex;
+				m_contiguousNodes[i].m_subPart = memPtr->m_subPart;
+				m_contiguousNodes[i].m_triangleIndex = memPtr->m_triangleIndex;
+			}
+		}
+	}
+
+	{
+		int numElem = quantizedBvhDoubleData.m_numQuantizedContiguousNodes;
+		m_quantizedContiguousNodes.resize(numElem);
+		
+		if (numElem)
+		{
+			b3QuantizedBvhNodeData* memPtr = quantizedBvhDoubleData.m_quantizedContiguousNodesPtr;
+			for (int i=0;i<numElem;i++,memPtr++)
+			{
+				m_quantizedContiguousNodes[i].m_escapeIndexOrTriangleIndex = memPtr->m_escapeIndexOrTriangleIndex;
+				m_quantizedContiguousNodes[i].m_quantizedAabbMax[0] = memPtr->m_quantizedAabbMax[0];
+				m_quantizedContiguousNodes[i].m_quantizedAabbMax[1] = memPtr->m_quantizedAabbMax[1];
+				m_quantizedContiguousNodes[i].m_quantizedAabbMax[2] = memPtr->m_quantizedAabbMax[2];
+				m_quantizedContiguousNodes[i].m_quantizedAabbMin[0] = memPtr->m_quantizedAabbMin[0];
+				m_quantizedContiguousNodes[i].m_quantizedAabbMin[1] = memPtr->m_quantizedAabbMin[1];
+				m_quantizedContiguousNodes[i].m_quantizedAabbMin[2] = memPtr->m_quantizedAabbMin[2];
+			}
+		}
+	}
+
+	m_traversalMode = b3TraversalMode(quantizedBvhDoubleData.m_traversalMode);
+	
+	{
+		int numElem = quantizedBvhDoubleData.m_numSubtreeHeaders;
+		m_SubtreeHeaders.resize(numElem);
+		if (numElem)
+		{
+			b3BvhSubtreeInfoData* memPtr = quantizedBvhDoubleData.m_subTreeInfoPtr;
+			for (int i=0;i<numElem;i++,memPtr++)
+			{
+				m_SubtreeHeaders[i].m_quantizedAabbMax[0] = memPtr->m_quantizedAabbMax[0] ;
+				m_SubtreeHeaders[i].m_quantizedAabbMax[1] = memPtr->m_quantizedAabbMax[1];
+				m_SubtreeHeaders[i].m_quantizedAabbMax[2] = memPtr->m_quantizedAabbMax[2];
+				m_SubtreeHeaders[i].m_quantizedAabbMin[0] = memPtr->m_quantizedAabbMin[0];
+				m_SubtreeHeaders[i].m_quantizedAabbMin[1] = memPtr->m_quantizedAabbMin[1];
+				m_SubtreeHeaders[i].m_quantizedAabbMin[2] = memPtr->m_quantizedAabbMin[2];
+				m_SubtreeHeaders[i].m_rootNodeIndex = memPtr->m_rootNodeIndex;
+				m_SubtreeHeaders[i].m_subtreeSize = memPtr->m_subtreeSize;
+			}
+		}
+	}
+
+}
+
+
+
+///fills the dataBuffer and returns the struct name (and 0 on failure)
+const char*	b3QuantizedBvh::serialize(void* dataBuffer, b3Serializer* serializer) const
+{
+	b3Assert(0);
+	return 0;
+}
+
+
+
+
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3QuantizedBvh.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3QuantizedBvh.h
new file mode 100644
index 0000000000..63c523c758
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3QuantizedBvh.h
@@ -0,0 +1,556 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2006 Erwin Coumans  http://continuousphysics.com/Bullet/
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+#ifndef B3_QUANTIZED_BVH_H
+#define B3_QUANTIZED_BVH_H
+
+class b3Serializer;
+
+//#define DEBUG_CHECK_DEQUANTIZATION 1
+#ifdef DEBUG_CHECK_DEQUANTIZATION
+#ifdef __SPU__
+#define printf spu_printf
+#endif //__SPU__
+
+#include <stdio.h>
+#include <stdlib.h>
+#endif //DEBUG_CHECK_DEQUANTIZATION
+
+#include "Bullet3Common/b3Vector3.h"
+#include "Bullet3Common/b3AlignedAllocator.h"
+
+#ifdef B3_USE_DOUBLE_PRECISION
+#define b3QuantizedBvhData b3QuantizedBvhDoubleData
+#define b3OptimizedBvhNodeData b3OptimizedBvhNodeDoubleData
+#define b3QuantizedBvhDataName "b3QuantizedBvhDoubleData"
+#else
+#define b3QuantizedBvhData b3QuantizedBvhFloatData
+#define b3OptimizedBvhNodeData b3OptimizedBvhNodeFloatData
+#define b3QuantizedBvhDataName "b3QuantizedBvhFloatData"
+#endif
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3QuantizedBvhNodeData.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3BvhSubtreeInfoData.h"
+
+
+
+//http://msdn.microsoft.com/library/default.asp?url=/library/en-us/vclang/html/vclrf__m128.asp
+
+
+//Note: currently we have 16 bytes per quantized node
+#define MAX_SUBTREE_SIZE_IN_BYTES  2048
+
+// 10 gives the potential for 1024 parts, with at most 2^21 (2097152) (minus one
+// actually) triangles each (since the sign bit is reserved
+#define MAX_NUM_PARTS_IN_BITS 10
+
+///b3QuantizedBvhNode is a compressed aabb node, 16 bytes.
+///Node can be used for leafnode or internal node. Leafnodes can point to 32-bit triangle index (non-negative range).
+B3_ATTRIBUTE_ALIGNED16	(struct) b3QuantizedBvhNode : public b3QuantizedBvhNodeData
+{
+	B3_DECLARE_ALIGNED_ALLOCATOR();
+
+	bool isLeafNode() const
+	{
+		//skipindex is negative (internal node), triangleindex >=0 (leafnode)
+		return (m_escapeIndexOrTriangleIndex >= 0);
+	}
+	int getEscapeIndex() const
+	{
+		b3Assert(!isLeafNode());
+		return -m_escapeIndexOrTriangleIndex;
+	}
+	int	getTriangleIndex() const
+	{
+		b3Assert(isLeafNode());
+		unsigned int x=0;
+		unsigned int y = (~(x&0))<<(31-MAX_NUM_PARTS_IN_BITS);
+		// Get only the lower bits where the triangle index is stored
+		return (m_escapeIndexOrTriangleIndex&~(y));
+	}
+	int	getPartId() const
+	{
+		b3Assert(isLeafNode());
+		// Get only the highest bits where the part index is stored
+		return (m_escapeIndexOrTriangleIndex>>(31-MAX_NUM_PARTS_IN_BITS));
+	}
+}
+;
+
+/// b3OptimizedBvhNode contains both internal and leaf node information.
+/// Total node size is 44 bytes / node. You can use the compressed version of 16 bytes.
+B3_ATTRIBUTE_ALIGNED16 (struct) b3OptimizedBvhNode
+{
+	B3_DECLARE_ALIGNED_ALLOCATOR();
+
+	//32 bytes
+	b3Vector3	m_aabbMinOrg;
+	b3Vector3	m_aabbMaxOrg;
+
+	//4
+	int	m_escapeIndex;
+
+	//8
+	//for child nodes
+	int	m_subPart;
+	int	m_triangleIndex;
+
+//pad the size to 64 bytes
+	char	m_padding[20];
+};
+
+
+///b3BvhSubtreeInfo provides info to gather a subtree of limited size
+B3_ATTRIBUTE_ALIGNED16(class) b3BvhSubtreeInfo : public b3BvhSubtreeInfoData
+{
+public:
+	B3_DECLARE_ALIGNED_ALLOCATOR();
+
+	b3BvhSubtreeInfo()
+	{
+		//memset(&m_padding[0], 0, sizeof(m_padding));
+	}
+
+
+	void	setAabbFromQuantizeNode(const b3QuantizedBvhNode& quantizedNode)
+	{
+		m_quantizedAabbMin[0] = quantizedNode.m_quantizedAabbMin[0];
+		m_quantizedAabbMin[1] = quantizedNode.m_quantizedAabbMin[1];
+		m_quantizedAabbMin[2] = quantizedNode.m_quantizedAabbMin[2];
+		m_quantizedAabbMax[0] = quantizedNode.m_quantizedAabbMax[0];
+		m_quantizedAabbMax[1] = quantizedNode.m_quantizedAabbMax[1];
+		m_quantizedAabbMax[2] = quantizedNode.m_quantizedAabbMax[2];
+	}
+}
+;
+
+
+class b3NodeOverlapCallback
+{
+public:
+	virtual ~b3NodeOverlapCallback() {};
+
+	virtual void processNode(int subPart, int triangleIndex) = 0;
+};
+
+#include "Bullet3Common/b3AlignedAllocator.h"
+#include "Bullet3Common/b3AlignedObjectArray.h"
+
+
+
+///for code readability:
+typedef b3AlignedObjectArray<b3OptimizedBvhNode>	NodeArray;
+typedef b3AlignedObjectArray<b3QuantizedBvhNode>	QuantizedNodeArray;
+typedef b3AlignedObjectArray<b3BvhSubtreeInfo>		BvhSubtreeInfoArray;
+
+
+///The b3QuantizedBvh class stores an AABB tree that can be quickly traversed on CPU and Cell SPU.
+///It is used by the b3BvhTriangleMeshShape as midphase
+///It is recommended to use quantization for better performance and lower memory requirements.
+B3_ATTRIBUTE_ALIGNED16(class) b3QuantizedBvh
+{
+public:
+	enum b3TraversalMode
+	{
+		TRAVERSAL_STACKLESS = 0,
+		TRAVERSAL_STACKLESS_CACHE_FRIENDLY,
+		TRAVERSAL_RECURSIVE
+	};
+
+
+
+
+	b3Vector3			m_bvhAabbMin;
+	b3Vector3			m_bvhAabbMax;
+	b3Vector3			m_bvhQuantization;
+
+protected:
+	int					m_bulletVersion;	//for serialization versioning. It could also be used to detect endianess.
+
+	int					m_curNodeIndex;
+	//quantization data
+	bool				m_useQuantization;
+
+
+
+	NodeArray			m_leafNodes;
+	NodeArray			m_contiguousNodes;
+	QuantizedNodeArray	m_quantizedLeafNodes;
+	QuantizedNodeArray	m_quantizedContiguousNodes;
+	
+	b3TraversalMode	m_traversalMode;
+	BvhSubtreeInfoArray		m_SubtreeHeaders;
+
+	//This is only used for serialization so we don't have to add serialization directly to b3AlignedObjectArray
+	mutable int m_subtreeHeaderCount;
+
+	
+
+
+
+	///two versions, one for quantized and normal nodes. This allows code-reuse while maintaining readability (no template/macro!)
+	///this might be refactored into a virtual, it is usually not calculated at run-time
+	void	setInternalNodeAabbMin(int nodeIndex, const b3Vector3& aabbMin)
+	{
+		if (m_useQuantization)
+		{
+			quantize(&m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[0] ,aabbMin,0);
+		} else
+		{
+			m_contiguousNodes[nodeIndex].m_aabbMinOrg = aabbMin;
+
+		}
+	}
+	void	setInternalNodeAabbMax(int nodeIndex,const b3Vector3& aabbMax)
+	{
+		if (m_useQuantization)
+		{
+			quantize(&m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[0],aabbMax,1);
+		} else
+		{
+			m_contiguousNodes[nodeIndex].m_aabbMaxOrg = aabbMax;
+		}
+	}
+
+	b3Vector3 getAabbMin(int nodeIndex) const
+	{
+		if (m_useQuantization)
+		{
+			return unQuantize(&m_quantizedLeafNodes[nodeIndex].m_quantizedAabbMin[0]);
+		}
+		//non-quantized
+		return m_leafNodes[nodeIndex].m_aabbMinOrg;
+
+	}
+	b3Vector3 getAabbMax(int nodeIndex) const
+	{
+		if (m_useQuantization)
+		{
+			return unQuantize(&m_quantizedLeafNodes[nodeIndex].m_quantizedAabbMax[0]);
+		} 
+		//non-quantized
+		return m_leafNodes[nodeIndex].m_aabbMaxOrg;
+		
+	}
+
+	
+	void	setInternalNodeEscapeIndex(int nodeIndex, int escapeIndex)
+	{
+		if (m_useQuantization)
+		{
+			m_quantizedContiguousNodes[nodeIndex].m_escapeIndexOrTriangleIndex = -escapeIndex;
+		} 
+		else
+		{
+			m_contiguousNodes[nodeIndex].m_escapeIndex = escapeIndex;
+		}
+
+	}
+
+	void mergeInternalNodeAabb(int nodeIndex,const b3Vector3& newAabbMin,const b3Vector3& newAabbMax) 
+	{
+		if (m_useQuantization)
+		{
+			unsigned short int quantizedAabbMin[3];
+			unsigned short int quantizedAabbMax[3];
+			quantize(quantizedAabbMin,newAabbMin,0);
+			quantize(quantizedAabbMax,newAabbMax,1);
+			for (int i=0;i<3;i++)
+			{
+				if (m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[i] > quantizedAabbMin[i])
+					m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMin[i] = quantizedAabbMin[i];
+
+				if (m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[i] < quantizedAabbMax[i])
+					m_quantizedContiguousNodes[nodeIndex].m_quantizedAabbMax[i] = quantizedAabbMax[i];
+
+			}
+		} else
+		{
+			//non-quantized
+			m_contiguousNodes[nodeIndex].m_aabbMinOrg.setMin(newAabbMin);
+			m_contiguousNodes[nodeIndex].m_aabbMaxOrg.setMax(newAabbMax);		
+		}
+	}
+
+	void	swapLeafNodes(int firstIndex,int secondIndex);
+
+	void	assignInternalNodeFromLeafNode(int internalNode,int leafNodeIndex);
+
+protected:
+
+	
+
+	void	buildTree	(int startIndex,int endIndex);
+
+	int	calcSplittingAxis(int startIndex,int endIndex);
+
+	int	sortAndCalcSplittingIndex(int startIndex,int endIndex,int splitAxis);
+	
+	void	walkStacklessTree(b3NodeOverlapCallback* nodeCallback,const b3Vector3& aabbMin,const b3Vector3& aabbMax) const;
+
+	void	walkStacklessQuantizedTreeAgainstRay(b3NodeOverlapCallback* nodeCallback, const b3Vector3& raySource, const b3Vector3& rayTarget, const b3Vector3& aabbMin, const b3Vector3& aabbMax, int startNodeIndex,int endNodeIndex) const;
+	void	walkStacklessQuantizedTree(b3NodeOverlapCallback* nodeCallback,unsigned short int* quantizedQueryAabbMin,unsigned short int* quantizedQueryAabbMax,int startNodeIndex,int endNodeIndex) const;
+	void	walkStacklessTreeAgainstRay(b3NodeOverlapCallback* nodeCallback, const b3Vector3& raySource, const b3Vector3& rayTarget, const b3Vector3& aabbMin, const b3Vector3& aabbMax, int startNodeIndex,int endNodeIndex) const;
+
+	///tree traversal designed for small-memory processors like PS3 SPU
+	void	walkStacklessQuantizedTreeCacheFriendly(b3NodeOverlapCallback* nodeCallback,unsigned short int* quantizedQueryAabbMin,unsigned short int* quantizedQueryAabbMax) const;
+
+	///use the 16-byte stackless 'skipindex' node tree to do a recursive traversal
+	void	walkRecursiveQuantizedTreeAgainstQueryAabb(const b3QuantizedBvhNode* currentNode,b3NodeOverlapCallback* nodeCallback,unsigned short int* quantizedQueryAabbMin,unsigned short int* quantizedQueryAabbMax) const;
+
+	///use the 16-byte stackless 'skipindex' node tree to do a recursive traversal
+	void	walkRecursiveQuantizedTreeAgainstQuantizedTree(const b3QuantizedBvhNode* treeNodeA,const b3QuantizedBvhNode* treeNodeB,b3NodeOverlapCallback* nodeCallback) const;
+	
+
+
+
+	void	updateSubtreeHeaders(int leftChildNodexIndex,int rightChildNodexIndex);
+
+public:
+	
+	B3_DECLARE_ALIGNED_ALLOCATOR();
+
+	b3QuantizedBvh();
+
+	virtual ~b3QuantizedBvh();
+
+	
+	///***************************************** expert/internal use only *************************
+	void	setQuantizationValues(const b3Vector3& bvhAabbMin,const b3Vector3& bvhAabbMax,b3Scalar quantizationMargin=b3Scalar(1.0));
+	QuantizedNodeArray&	getLeafNodeArray() {			return	m_quantizedLeafNodes;	}
+	///buildInternal is expert use only: assumes that setQuantizationValues and LeafNodeArray are initialized
+	void	buildInternal();
+	///***************************************** expert/internal use only *************************
+
+	void	reportAabbOverlappingNodex(b3NodeOverlapCallback* nodeCallback,const b3Vector3& aabbMin,const b3Vector3& aabbMax) const;
+	void	reportRayOverlappingNodex (b3NodeOverlapCallback* nodeCallback, const b3Vector3& raySource, const b3Vector3& rayTarget) const;
+	void	reportBoxCastOverlappingNodex(b3NodeOverlapCallback* nodeCallback, const b3Vector3& raySource, const b3Vector3& rayTarget, const b3Vector3& aabbMin,const b3Vector3& aabbMax) const;
+
+		B3_FORCE_INLINE void quantize(unsigned short* out, const b3Vector3& point,int isMax) const
+	{
+
+		b3Assert(m_useQuantization);
+
+		b3Assert(point.getX() <= m_bvhAabbMax.getX());
+		b3Assert(point.getY() <= m_bvhAabbMax.getY());
+		b3Assert(point.getZ() <= m_bvhAabbMax.getZ());
+
+		b3Assert(point.getX() >= m_bvhAabbMin.getX());
+		b3Assert(point.getY() >= m_bvhAabbMin.getY());
+		b3Assert(point.getZ() >= m_bvhAabbMin.getZ());
+
+		b3Vector3 v = (point - m_bvhAabbMin) * m_bvhQuantization;
+		///Make sure rounding is done in a way that unQuantize(quantizeWithClamp(...)) is conservative
+		///end-points always set the first bit, so that they are sorted properly (so that neighbouring AABBs overlap properly)
+		///@todo: double-check this
+		if (isMax)
+		{
+			out[0] = (unsigned short) (((unsigned short)(v.getX()+b3Scalar(1.)) | 1));
+			out[1] = (unsigned short) (((unsigned short)(v.getY()+b3Scalar(1.)) | 1));
+			out[2] = (unsigned short) (((unsigned short)(v.getZ()+b3Scalar(1.)) | 1));
+		} else
+		{
+			out[0] = (unsigned short) (((unsigned short)(v.getX()) & 0xfffe));
+			out[1] = (unsigned short) (((unsigned short)(v.getY()) & 0xfffe));
+			out[2] = (unsigned short) (((unsigned short)(v.getZ()) & 0xfffe));
+		}
+
+
+#ifdef DEBUG_CHECK_DEQUANTIZATION
+		b3Vector3 newPoint = unQuantize(out);
+		if (isMax)
+		{
+			if (newPoint.getX() < point.getX())
+			{
+				printf("unconservative X, diffX = %f, oldX=%f,newX=%f\n",newPoint.getX()-point.getX(), newPoint.getX(),point.getX());
+			}
+			if (newPoint.getY() < point.getY())
+			{
+				printf("unconservative Y, diffY = %f, oldY=%f,newY=%f\n",newPoint.getY()-point.getY(), newPoint.getY(),point.getY());
+			}
+			if (newPoint.getZ() < point.getZ())
+			{
+
+				printf("unconservative Z, diffZ = %f, oldZ=%f,newZ=%f\n",newPoint.getZ()-point.getZ(), newPoint.getZ(),point.getZ());
+			}
+		} else
+		{
+			if (newPoint.getX() > point.getX())
+			{
+				printf("unconservative X, diffX = %f, oldX=%f,newX=%f\n",newPoint.getX()-point.getX(), newPoint.getX(),point.getX());
+			}
+			if (newPoint.getY() > point.getY())
+			{
+				printf("unconservative Y, diffY = %f, oldY=%f,newY=%f\n",newPoint.getY()-point.getY(), newPoint.getY(),point.getY());
+			}
+			if (newPoint.getZ() > point.getZ())
+			{
+				printf("unconservative Z, diffZ = %f, oldZ=%f,newZ=%f\n",newPoint.getZ()-point.getZ(), newPoint.getZ(),point.getZ());
+			}
+		}
+#endif //DEBUG_CHECK_DEQUANTIZATION
+
+	}
+
+
+	B3_FORCE_INLINE void quantizeWithClamp(unsigned short* out, const b3Vector3& point2,int isMax) const
+	{
+
+		b3Assert(m_useQuantization);
+
+		b3Vector3 clampedPoint(point2);
+		clampedPoint.setMax(m_bvhAabbMin);
+		clampedPoint.setMin(m_bvhAabbMax);
+
+		quantize(out,clampedPoint,isMax);
+
+	}
+	
+	B3_FORCE_INLINE b3Vector3	unQuantize(const unsigned short* vecIn) const
+	{
+			b3Vector3	vecOut;
+			vecOut.setValue(
+			(b3Scalar)(vecIn[0]) / (m_bvhQuantization.getX()),
+			(b3Scalar)(vecIn[1]) / (m_bvhQuantization.getY()),
+			(b3Scalar)(vecIn[2]) / (m_bvhQuantization.getZ()));
+			vecOut += m_bvhAabbMin;
+			return vecOut;
+	}
+
+	///setTraversalMode let's you choose between stackless, recursive or stackless cache friendly tree traversal. Note this is only implemented for quantized trees.
+	void	setTraversalMode(b3TraversalMode	traversalMode)
+	{
+		m_traversalMode = traversalMode;
+	}
+
+
+	B3_FORCE_INLINE QuantizedNodeArray&	getQuantizedNodeArray()
+	{	
+		return	m_quantizedContiguousNodes;
+	}
+
+
+	B3_FORCE_INLINE BvhSubtreeInfoArray&	getSubtreeInfoArray()
+	{
+		return m_SubtreeHeaders;
+	}
+
+////////////////////////////////////////////////////////////////////
+
+	/////Calculate space needed to store BVH for serialization
+	unsigned calculateSerializeBufferSize() const;
+
+	/// Data buffer MUST be 16 byte aligned
+	virtual bool serialize(void *o_alignedDataBuffer, unsigned i_dataBufferSize, bool i_swapEndian) const;
+
+	///deSerializeInPlace loads and initializes a BVH from a buffer in memory 'in place'
+	static b3QuantizedBvh *deSerializeInPlace(void *i_alignedDataBuffer, unsigned int i_dataBufferSize, bool i_swapEndian);
+
+	static unsigned int getAlignmentSerializationPadding();
+//////////////////////////////////////////////////////////////////////
+
+	
+	virtual	int	calculateSerializeBufferSizeNew() const;
+
+	///fills the dataBuffer and returns the struct name (and 0 on failure)
+	virtual	const char*	serialize(void* dataBuffer, b3Serializer* serializer) const;
+
+	virtual	void deSerializeFloat(struct b3QuantizedBvhFloatData& quantizedBvhFloatData);
+
+	virtual	void deSerializeDouble(struct b3QuantizedBvhDoubleData& quantizedBvhDoubleData);
+
+
+////////////////////////////////////////////////////////////////////
+
+	B3_FORCE_INLINE bool isQuantized()
+	{
+		return m_useQuantization;
+	}
+
+private:
+	// Special "copy" constructor that allows for in-place deserialization
+	// Prevents b3Vector3's default constructor from being called, but doesn't inialize much else
+	// ownsMemory should most likely be false if deserializing, and if you are not, don't call this (it also changes the function signature, which we need)
+	b3QuantizedBvh(b3QuantizedBvh &other, bool ownsMemory);
+
+}
+;
+
+
+struct b3OptimizedBvhNodeFloatData
+{
+	b3Vector3FloatData	m_aabbMinOrg;
+	b3Vector3FloatData	m_aabbMaxOrg;
+	int	m_escapeIndex;
+	int	m_subPart;
+	int	m_triangleIndex;
+	char m_pad[4];
+};
+
+struct b3OptimizedBvhNodeDoubleData
+{
+	b3Vector3DoubleData	m_aabbMinOrg;
+	b3Vector3DoubleData	m_aabbMaxOrg;
+	int	m_escapeIndex;
+	int	m_subPart;
+	int	m_triangleIndex;
+	char	m_pad[4];
+};
+
+
+
+struct	b3QuantizedBvhFloatData
+{
+	b3Vector3FloatData			m_bvhAabbMin;
+	b3Vector3FloatData			m_bvhAabbMax;
+	b3Vector3FloatData			m_bvhQuantization;
+	int					m_curNodeIndex;
+	int					m_useQuantization;
+	int					m_numContiguousLeafNodes;
+	int					m_numQuantizedContiguousNodes;
+	b3OptimizedBvhNodeFloatData	*m_contiguousNodesPtr;
+	b3QuantizedBvhNodeData		*m_quantizedContiguousNodesPtr;
+	b3BvhSubtreeInfoData	*m_subTreeInfoPtr;
+	int					m_traversalMode;
+	int					m_numSubtreeHeaders;
+	
+};
+
+struct	b3QuantizedBvhDoubleData
+{
+	b3Vector3DoubleData			m_bvhAabbMin;
+	b3Vector3DoubleData			m_bvhAabbMax;
+	b3Vector3DoubleData			m_bvhQuantization;
+	int							m_curNodeIndex;
+	int							m_useQuantization;
+	int							m_numContiguousLeafNodes;
+	int							m_numQuantizedContiguousNodes;
+	b3OptimizedBvhNodeDoubleData	*m_contiguousNodesPtr;
+	b3QuantizedBvhNodeData			*m_quantizedContiguousNodesPtr;
+
+	int							m_traversalMode;
+	int							m_numSubtreeHeaders;
+	b3BvhSubtreeInfoData		*m_subTreeInfoPtr;
+};
+
+
+B3_FORCE_INLINE	int	b3QuantizedBvh::calculateSerializeBufferSizeNew() const
+{
+	return sizeof(b3QuantizedBvhData);
+}
+
+
+
+#endif //B3_QUANTIZED_BVH_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3StridingMeshInterface.cpp b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3StridingMeshInterface.cpp
new file mode 100644
index 0000000000..4d97f7f62b
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3StridingMeshInterface.cpp
@@ -0,0 +1,214 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2009 Erwin Coumans  http://bulletphysics.org
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+#include "b3StridingMeshInterface.h"
+
+
+b3StridingMeshInterface::~b3StridingMeshInterface()
+{
+
+}
+
+
+void	b3StridingMeshInterface::InternalProcessAllTriangles(b3InternalTriangleIndexCallback* callback,const b3Vector3& aabbMin,const b3Vector3& aabbMax) const
+{
+	(void)aabbMin;
+	(void)aabbMax;
+	int numtotalphysicsverts = 0;
+	int part,graphicssubparts = getNumSubParts();
+	const unsigned char * vertexbase;
+	const unsigned char * indexbase;
+	int indexstride;
+	PHY_ScalarType type;
+	PHY_ScalarType gfxindextype;
+	int stride,numverts,numtriangles;
+	int gfxindex;
+	b3Vector3 triangle[3];
+
+	b3Vector3 meshScaling = getScaling();
+
+	///if the number of parts is big, the performance might drop due to the innerloop switch on indextype
+	for (part=0;part<graphicssubparts ;part++)
+	{
+		getLockedReadOnlyVertexIndexBase(&vertexbase,numverts,type,stride,&indexbase,indexstride,numtriangles,gfxindextype,part);
+		numtotalphysicsverts+=numtriangles*3; //upper bound
+
+		///unlike that developers want to pass in double-precision meshes in single-precision Bullet build
+		///so disable this feature by default
+		///see patch http://code.google.com/p/bullet/issues/detail?id=213
+
+		switch (type)
+		{
+		case PHY_FLOAT:
+		 {
+
+			 float* graphicsbase;
+
+			 switch (gfxindextype)
+			 {
+			 case PHY_INTEGER:
+				 {
+					 for (gfxindex=0;gfxindex<numtriangles;gfxindex++)
+					 {
+						 unsigned int* tri_indices= (unsigned int*)(indexbase+gfxindex*indexstride);
+						 graphicsbase = (float*)(vertexbase+tri_indices[0]*stride);
+						 triangle[0].setValue(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),graphicsbase[2]*meshScaling.getZ());
+						 graphicsbase = (float*)(vertexbase+tri_indices[1]*stride);
+						 triangle[1].setValue(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),	graphicsbase[2]*meshScaling.getZ());
+						 graphicsbase = (float*)(vertexbase+tri_indices[2]*stride);
+						 triangle[2].setValue(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),	graphicsbase[2]*meshScaling.getZ());
+						 callback->internalProcessTriangleIndex(triangle,part,gfxindex);
+					 }
+					 break;
+				 }
+			 case PHY_SHORT:
+				 {
+					 for (gfxindex=0;gfxindex<numtriangles;gfxindex++)
+					 {
+						 unsigned short int* tri_indices= (unsigned short int*)(indexbase+gfxindex*indexstride);
+						 graphicsbase = (float*)(vertexbase+tri_indices[0]*stride);
+						 triangle[0].setValue(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),graphicsbase[2]*meshScaling.getZ());
+						 graphicsbase = (float*)(vertexbase+tri_indices[1]*stride);
+						 triangle[1].setValue(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),	graphicsbase[2]*meshScaling.getZ());
+						 graphicsbase = (float*)(vertexbase+tri_indices[2]*stride);
+						 triangle[2].setValue(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),	graphicsbase[2]*meshScaling.getZ());
+						 callback->internalProcessTriangleIndex(triangle,part,gfxindex);
+					 }
+					 break;
+				 }
+			case PHY_UCHAR:
+				 {
+					 for (gfxindex=0;gfxindex<numtriangles;gfxindex++)
+					 {
+						 unsigned char* tri_indices= (unsigned char*)(indexbase+gfxindex*indexstride);
+						 graphicsbase = (float*)(vertexbase+tri_indices[0]*stride);
+						 triangle[0].setValue(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),graphicsbase[2]*meshScaling.getZ());
+						 graphicsbase = (float*)(vertexbase+tri_indices[1]*stride);
+						 triangle[1].setValue(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),	graphicsbase[2]*meshScaling.getZ());
+						 graphicsbase = (float*)(vertexbase+tri_indices[2]*stride);
+						 triangle[2].setValue(graphicsbase[0]*meshScaling.getX(),graphicsbase[1]*meshScaling.getY(),	graphicsbase[2]*meshScaling.getZ());
+						 callback->internalProcessTriangleIndex(triangle,part,gfxindex);
+					 }
+					 break;
+				 }
+			 default:
+				 b3Assert((gfxindextype == PHY_INTEGER) || (gfxindextype == PHY_SHORT));
+			 }
+			 break;
+		 }
+
+		case PHY_DOUBLE:
+			{
+				double* graphicsbase;
+
+				switch (gfxindextype)
+				{
+				case PHY_INTEGER:
+					{
+						for (gfxindex=0;gfxindex<numtriangles;gfxindex++)
+						{
+							unsigned int* tri_indices= (unsigned int*)(indexbase+gfxindex*indexstride);
+							graphicsbase = (double*)(vertexbase+tri_indices[0]*stride);
+							triangle[0].setValue((b3Scalar)graphicsbase[0]*meshScaling.getX(),(b3Scalar)graphicsbase[1]*meshScaling.getY(),(b3Scalar)graphicsbase[2]*meshScaling.getZ());
+							graphicsbase = (double*)(vertexbase+tri_indices[1]*stride);
+							triangle[1].setValue((b3Scalar)graphicsbase[0]*meshScaling.getX(),(b3Scalar)graphicsbase[1]*meshScaling.getY(),  (b3Scalar)graphicsbase[2]*meshScaling.getZ());
+							graphicsbase = (double*)(vertexbase+tri_indices[2]*stride);
+							triangle[2].setValue((b3Scalar)graphicsbase[0]*meshScaling.getX(),(b3Scalar)graphicsbase[1]*meshScaling.getY(),  (b3Scalar)graphicsbase[2]*meshScaling.getZ());
+							callback->internalProcessTriangleIndex(triangle,part,gfxindex);
+						}
+						break;
+					}
+				case PHY_SHORT:
+					{
+						for (gfxindex=0;gfxindex<numtriangles;gfxindex++)
+						{
+							unsigned short int* tri_indices= (unsigned short int*)(indexbase+gfxindex*indexstride);
+							graphicsbase = (double*)(vertexbase+tri_indices[0]*stride);
+							triangle[0].setValue((b3Scalar)graphicsbase[0]*meshScaling.getX(),(b3Scalar)graphicsbase[1]*meshScaling.getY(),(b3Scalar)graphicsbase[2]*meshScaling.getZ());
+							graphicsbase = (double*)(vertexbase+tri_indices[1]*stride);
+							triangle[1].setValue((b3Scalar)graphicsbase[0]*meshScaling.getX(),(b3Scalar)graphicsbase[1]*meshScaling.getY(),  (b3Scalar)graphicsbase[2]*meshScaling.getZ());
+							graphicsbase = (double*)(vertexbase+tri_indices[2]*stride);
+							triangle[2].setValue((b3Scalar)graphicsbase[0]*meshScaling.getX(),(b3Scalar)graphicsbase[1]*meshScaling.getY(),  (b3Scalar)graphicsbase[2]*meshScaling.getZ());
+							callback->internalProcessTriangleIndex(triangle,part,gfxindex);
+						}
+						break;
+					}
+				case PHY_UCHAR:
+					{
+						for (gfxindex=0;gfxindex<numtriangles;gfxindex++)
+						{
+							unsigned char* tri_indices= (unsigned char*)(indexbase+gfxindex*indexstride);
+							graphicsbase = (double*)(vertexbase+tri_indices[0]*stride);
+							triangle[0].setValue((b3Scalar)graphicsbase[0]*meshScaling.getX(),(b3Scalar)graphicsbase[1]*meshScaling.getY(),(b3Scalar)graphicsbase[2]*meshScaling.getZ());
+							graphicsbase = (double*)(vertexbase+tri_indices[1]*stride);
+							triangle[1].setValue((b3Scalar)graphicsbase[0]*meshScaling.getX(),(b3Scalar)graphicsbase[1]*meshScaling.getY(),  (b3Scalar)graphicsbase[2]*meshScaling.getZ());
+							graphicsbase = (double*)(vertexbase+tri_indices[2]*stride);
+							triangle[2].setValue((b3Scalar)graphicsbase[0]*meshScaling.getX(),(b3Scalar)graphicsbase[1]*meshScaling.getY(),  (b3Scalar)graphicsbase[2]*meshScaling.getZ());
+							callback->internalProcessTriangleIndex(triangle,part,gfxindex);
+						}
+						break;
+					}
+				default:
+					b3Assert((gfxindextype == PHY_INTEGER) || (gfxindextype == PHY_SHORT));
+				}
+				break;
+			}
+		default:
+			b3Assert((type == PHY_FLOAT) || (type == PHY_DOUBLE));
+		}
+
+		unLockReadOnlyVertexBase(part);
+	}
+}
+
+void	b3StridingMeshInterface::calculateAabbBruteForce(b3Vector3& aabbMin,b3Vector3& aabbMax)
+{
+
+	struct	AabbCalculationCallback : public b3InternalTriangleIndexCallback
+	{
+		b3Vector3	m_aabbMin;
+		b3Vector3	m_aabbMax;
+
+		AabbCalculationCallback()
+		{
+			m_aabbMin.setValue(b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT));
+			m_aabbMax.setValue(b3Scalar(-B3_LARGE_FLOAT),b3Scalar(-B3_LARGE_FLOAT),b3Scalar(-B3_LARGE_FLOAT));
+		}
+
+		virtual void internalProcessTriangleIndex(b3Vector3* triangle,int partId,int  triangleIndex)
+		{
+			(void)partId;
+			(void)triangleIndex;
+
+			m_aabbMin.setMin(triangle[0]);
+			m_aabbMax.setMax(triangle[0]);
+			m_aabbMin.setMin(triangle[1]);
+			m_aabbMax.setMax(triangle[1]);
+			m_aabbMin.setMin(triangle[2]);
+			m_aabbMax.setMax(triangle[2]);
+		}
+	};
+
+	//first calculate the total aabb for all triangles
+	AabbCalculationCallback	aabbCallback;
+	aabbMin.setValue(b3Scalar(-B3_LARGE_FLOAT),b3Scalar(-B3_LARGE_FLOAT),b3Scalar(-B3_LARGE_FLOAT));
+	aabbMax.setValue(b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT));
+	InternalProcessAllTriangles(&aabbCallback,aabbMin,aabbMax);
+
+	aabbMin = aabbCallback.m_aabbMin;
+	aabbMax = aabbCallback.m_aabbMax;
+}
+
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3StridingMeshInterface.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3StridingMeshInterface.h
new file mode 100644
index 0000000000..9513f68f77
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3StridingMeshInterface.h
@@ -0,0 +1,167 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2009 Erwin Coumans  http://bulletphysics.org
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+#ifndef B3_STRIDING_MESHINTERFACE_H
+#define B3_STRIDING_MESHINTERFACE_H
+
+#include "Bullet3Common/b3Vector3.h"
+#include "b3TriangleCallback.h"
+//#include "b3ConcaveShape.h"
+
+
+enum  	PHY_ScalarType { 
+  PHY_FLOAT, PHY_DOUBLE, PHY_INTEGER, PHY_SHORT, 
+  PHY_FIXEDPOINT88, PHY_UCHAR 
+};
+
+
+///	The b3StridingMeshInterface is the interface class for high performance generic access to triangle meshes, used in combination with b3BvhTriangleMeshShape and some other collision shapes.
+/// Using index striding of 3*sizeof(integer) it can use triangle arrays, using index striding of 1*sizeof(integer) it can handle triangle strips.
+/// It allows for sharing graphics and collision meshes. Also it provides locking/unlocking of graphics meshes that are in gpu memory.
+B3_ATTRIBUTE_ALIGNED16(class ) b3StridingMeshInterface
+{
+	protected:
+	
+		b3Vector3 m_scaling;
+
+	public:
+		B3_DECLARE_ALIGNED_ALLOCATOR();
+		
+		b3StridingMeshInterface() :m_scaling(b3MakeVector3(b3Scalar(1.),b3Scalar(1.),b3Scalar(1.)))
+		{
+
+		}
+
+		virtual ~b3StridingMeshInterface();
+
+
+
+		virtual void	InternalProcessAllTriangles(b3InternalTriangleIndexCallback* callback,const b3Vector3& aabbMin,const b3Vector3& aabbMax) const;
+
+		///brute force method to calculate aabb
+		void	calculateAabbBruteForce(b3Vector3& aabbMin,b3Vector3& aabbMax);
+
+		/// get read and write access to a subpart of a triangle mesh
+		/// this subpart has a continuous array of vertices and indices
+		/// in this way the mesh can be handled as chunks of memory with striding
+		/// very similar to OpenGL vertexarray support
+		/// make a call to unLockVertexBase when the read and write access is finished	
+		virtual void	getLockedVertexIndexBase(unsigned char **vertexbase, int& numverts,PHY_ScalarType& type, int& stride,unsigned char **indexbase,int & indexstride,int& numfaces,PHY_ScalarType& indicestype,int subpart=0)=0;
+		
+		virtual void	getLockedReadOnlyVertexIndexBase(const unsigned char **vertexbase, int& numverts,PHY_ScalarType& type, int& stride,const unsigned char **indexbase,int & indexstride,int& numfaces,PHY_ScalarType& indicestype,int subpart=0) const=0;
+	
+		/// unLockVertexBase finishes the access to a subpart of the triangle mesh
+		/// make a call to unLockVertexBase when the read and write access (using getLockedVertexIndexBase) is finished
+		virtual void	unLockVertexBase(int subpart)=0;
+
+		virtual void	unLockReadOnlyVertexBase(int subpart) const=0;
+
+
+		/// getNumSubParts returns the number of seperate subparts
+		/// each subpart has a continuous array of vertices and indices
+		virtual int		getNumSubParts() const=0;
+
+		virtual void	preallocateVertices(int numverts)=0;
+		virtual void	preallocateIndices(int numindices)=0;
+
+		virtual bool	hasPremadeAabb() const { return false; }
+		virtual void	setPremadeAabb(const b3Vector3& aabbMin, const b3Vector3& aabbMax ) const
+                {
+                        (void) aabbMin;
+                        (void) aabbMax;
+                }
+		virtual void	getPremadeAabb(b3Vector3* aabbMin, b3Vector3* aabbMax ) const
+        {
+            (void) aabbMin;
+            (void) aabbMax;
+        }
+
+		const b3Vector3&	getScaling() const {
+			return m_scaling;
+		}
+		void	setScaling(const b3Vector3& scaling)
+		{
+			m_scaling = scaling;
+		}
+
+		virtual	int	calculateSerializeBufferSize() const;
+
+		///fills the dataBuffer and returns the struct name (and 0 on failure)
+		//virtual	const char*	serialize(void* dataBuffer, b3Serializer* serializer) const;
+
+
+};
+
+struct	b3IntIndexData
+{
+	int	m_value;
+};
+
+struct	b3ShortIntIndexData
+{
+	short m_value;
+	char m_pad[2];
+};
+
+struct	b3ShortIntIndexTripletData
+{
+	short	m_values[3];
+	char	m_pad[2];
+};
+
+struct	b3CharIndexTripletData
+{
+	unsigned char m_values[3];
+	char	m_pad;
+};
+
+
+///do not change those serialization structures, it requires an updated sBulletDNAstr/sBulletDNAstr64
+struct	b3MeshPartData
+{
+	b3Vector3FloatData			*m_vertices3f;
+	b3Vector3DoubleData			*m_vertices3d;
+
+	b3IntIndexData				*m_indices32;
+	b3ShortIntIndexTripletData	*m_3indices16;
+	b3CharIndexTripletData		*m_3indices8;
+
+	b3ShortIntIndexData			*m_indices16;//backwards compatibility
+
+	int                     m_numTriangles;//length of m_indices = m_numTriangles
+	int                     m_numVertices;
+};
+
+
+///do not change those serialization structures, it requires an updated sBulletDNAstr/sBulletDNAstr64
+struct	b3StridingMeshInterfaceData
+{
+	b3MeshPartData	*m_meshPartsPtr;
+	b3Vector3FloatData	m_scaling;
+	int	m_numMeshParts;
+	char m_padding[4];
+};
+
+
+
+
+B3_FORCE_INLINE	int	b3StridingMeshInterface::calculateSerializeBufferSize() const
+{
+	return sizeof(b3StridingMeshInterfaceData);
+}
+
+
+
+#endif //B3_STRIDING_MESHINTERFACE_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3SupportMappings.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3SupportMappings.h
new file mode 100644
index 0000000000..d073ee57c3
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3SupportMappings.h
@@ -0,0 +1,38 @@
+
+#ifndef B3_SUPPORT_MAPPINGS_H
+#define B3_SUPPORT_MAPPINGS_H
+
+#include "Bullet3Common/b3Transform.h"
+#include "Bullet3Common/b3AlignedObjectArray.h"
+#include "b3VectorFloat4.h"
+
+
+struct b3GjkPairDetector;
+
+
+
+inline b3Vector3 localGetSupportVertexWithMargin(const float4& supportVec,const struct b3ConvexPolyhedronData* hull, 
+	const b3AlignedObjectArray<b3Vector3>& verticesA, b3Scalar margin)
+{
+	b3Vector3 supVec = b3MakeVector3(b3Scalar(0.),b3Scalar(0.),b3Scalar(0.));
+	b3Scalar maxDot = b3Scalar(-B3_LARGE_FLOAT);
+
+    // Here we take advantage of dot(a, b*c) = dot(a*b, c).  Note: This is true mathematically, but not numerically. 
+    if( 0 < hull->m_numVertices )
+    {
+        const b3Vector3 scaled = supportVec;
+		int index = (int) scaled.maxDot( &verticesA[hull->m_vertexOffset], hull->m_numVertices, maxDot); 
+        return verticesA[hull->m_vertexOffset+index];
+    }
+
+    return supVec;
+
+}
+
+inline b3Vector3 localGetSupportVertexWithoutMargin(const float4& supportVec,const struct b3ConvexPolyhedronData* hull, 
+	const b3AlignedObjectArray<b3Vector3>& verticesA)
+{
+	return localGetSupportVertexWithMargin(supportVec,hull,verticesA,0.f);
+}
+
+#endif //B3_SUPPORT_MAPPINGS_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3TriangleCallback.cpp b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3TriangleCallback.cpp
new file mode 100644
index 0000000000..9066451884
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3TriangleCallback.cpp
@@ -0,0 +1,28 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2009 Erwin Coumans  http://bulletphysics.org
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+#include "b3TriangleCallback.h"
+
+b3TriangleCallback::~b3TriangleCallback()
+{
+	
+}
+
+
+b3InternalTriangleIndexCallback::~b3InternalTriangleIndexCallback()
+{
+
+}
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3TriangleCallback.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3TriangleCallback.h
new file mode 100644
index 0000000000..3059fa4f21
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3TriangleCallback.h
@@ -0,0 +1,42 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2009 Erwin Coumans  http://bulletphysics.org
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+#ifndef B3_TRIANGLE_CALLBACK_H
+#define B3_TRIANGLE_CALLBACK_H
+
+#include "Bullet3Common/b3Vector3.h"
+
+
+///The b3TriangleCallback provides a callback for each overlapping triangle when calling processAllTriangles.
+///This callback is called by processAllTriangles for all b3ConcaveShape derived class, such as  b3BvhTriangleMeshShape, b3StaticPlaneShape and b3HeightfieldTerrainShape.
+class b3TriangleCallback
+{
+public:
+
+	virtual ~b3TriangleCallback();
+	virtual void processTriangle(b3Vector3* triangle, int partId, int triangleIndex) = 0;
+};
+
+class b3InternalTriangleIndexCallback
+{
+public:
+
+	virtual ~b3InternalTriangleIndexCallback();
+	virtual void internalProcessTriangleIndex(b3Vector3* triangle,int partId,int  triangleIndex) = 0;
+};
+
+
+
+#endif //B3_TRIANGLE_CALLBACK_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3TriangleIndexVertexArray.cpp b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3TriangleIndexVertexArray.cpp
new file mode 100644
index 0000000000..a0f59babbe
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3TriangleIndexVertexArray.cpp
@@ -0,0 +1,95 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2009 Erwin Coumans  http://bulletphysics.org
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+#include "b3TriangleIndexVertexArray.h"
+
+b3TriangleIndexVertexArray::b3TriangleIndexVertexArray(int numTriangles,int* triangleIndexBase,int triangleIndexStride,int numVertices,b3Scalar* vertexBase,int vertexStride)
+: m_hasAabb(0)
+{
+	b3IndexedMesh mesh;
+
+	mesh.m_numTriangles = numTriangles;
+	mesh.m_triangleIndexBase = (const unsigned char *)triangleIndexBase;
+	mesh.m_triangleIndexStride = triangleIndexStride;
+	mesh.m_numVertices = numVertices;
+	mesh.m_vertexBase = (const unsigned char *)vertexBase;
+	mesh.m_vertexStride = vertexStride;
+
+	addIndexedMesh(mesh);
+
+}
+
+b3TriangleIndexVertexArray::~b3TriangleIndexVertexArray()
+{
+
+}
+
+void	b3TriangleIndexVertexArray::getLockedVertexIndexBase(unsigned char **vertexbase, int& numverts,PHY_ScalarType& type, int& vertexStride,unsigned char **indexbase,int & indexstride,int& numfaces,PHY_ScalarType& indicestype,int subpart)
+{
+	b3Assert(subpart< getNumSubParts() );
+
+	b3IndexedMesh& mesh = m_indexedMeshes[subpart];
+
+	numverts = mesh.m_numVertices;
+	(*vertexbase) = (unsigned char *) mesh.m_vertexBase;
+
+   type = mesh.m_vertexType;
+
+	vertexStride = mesh.m_vertexStride;
+
+	numfaces = mesh.m_numTriangles;
+
+	(*indexbase) = (unsigned char *)mesh.m_triangleIndexBase;
+	indexstride = mesh.m_triangleIndexStride;
+	indicestype = mesh.m_indexType;
+}
+
+void	b3TriangleIndexVertexArray::getLockedReadOnlyVertexIndexBase(const unsigned char **vertexbase, int& numverts,PHY_ScalarType& type, int& vertexStride,const unsigned char **indexbase,int & indexstride,int& numfaces,PHY_ScalarType& indicestype,int subpart) const
+{
+	const b3IndexedMesh& mesh = m_indexedMeshes[subpart];
+
+	numverts = mesh.m_numVertices;
+	(*vertexbase) = (const unsigned char *)mesh.m_vertexBase;
+
+   type = mesh.m_vertexType;
+   
+	vertexStride = mesh.m_vertexStride;
+
+	numfaces = mesh.m_numTriangles;
+	(*indexbase) = (const unsigned char *)mesh.m_triangleIndexBase;
+	indexstride = mesh.m_triangleIndexStride;
+	indicestype = mesh.m_indexType;
+}
+
+bool	b3TriangleIndexVertexArray::hasPremadeAabb() const
+{
+	return (m_hasAabb == 1);
+}
+
+
+void	b3TriangleIndexVertexArray::setPremadeAabb(const b3Vector3& aabbMin, const b3Vector3& aabbMax ) const
+{
+	m_aabbMin = aabbMin;
+	m_aabbMax = aabbMax;
+	m_hasAabb = 1; // this is intentionally an int see notes in header
+}
+
+void	b3TriangleIndexVertexArray::getPremadeAabb(b3Vector3* aabbMin, b3Vector3* aabbMax ) const
+{
+	*aabbMin = m_aabbMin;
+	*aabbMax = m_aabbMax;
+}
+
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3TriangleIndexVertexArray.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3TriangleIndexVertexArray.h
new file mode 100644
index 0000000000..d26b2893bc
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3TriangleIndexVertexArray.h
@@ -0,0 +1,133 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2009 Erwin Coumans  http://bulletphysics.org
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+#ifndef B3_TRIANGLE_INDEX_VERTEX_ARRAY_H
+#define B3_TRIANGLE_INDEX_VERTEX_ARRAY_H
+
+#include "b3StridingMeshInterface.h"
+#include "Bullet3Common/b3AlignedObjectArray.h"
+#include "Bullet3Common/b3Scalar.h"
+
+
+///The b3IndexedMesh indexes a single vertex and index array. Multiple b3IndexedMesh objects can be passed into a b3TriangleIndexVertexArray using addIndexedMesh.
+///Instead of the number of indices, we pass the number of triangles.
+B3_ATTRIBUTE_ALIGNED16( struct)	b3IndexedMesh
+{
+	B3_DECLARE_ALIGNED_ALLOCATOR();
+
+   int                     m_numTriangles;
+   const unsigned char *   m_triangleIndexBase;
+   // Size in byte of the indices for one triangle (3*sizeof(index_type) if the indices are tightly packed)
+   int                     m_triangleIndexStride;
+   int                     m_numVertices;
+   const unsigned char *   m_vertexBase;
+   // Size of a vertex, in bytes
+   int                     m_vertexStride;
+
+   // The index type is set when adding an indexed mesh to the
+   // b3TriangleIndexVertexArray, do not set it manually
+   PHY_ScalarType m_indexType;
+
+   // The vertex type has a default type similar to Bullet's precision mode (float or double)
+   // but can be set manually if you for example run Bullet with double precision but have
+   // mesh data in single precision..
+   PHY_ScalarType m_vertexType;
+
+
+   b3IndexedMesh()
+	   :m_indexType(PHY_INTEGER),
+#ifdef B3_USE_DOUBLE_PRECISION
+      m_vertexType(PHY_DOUBLE)
+#else // B3_USE_DOUBLE_PRECISION
+      m_vertexType(PHY_FLOAT)
+#endif // B3_USE_DOUBLE_PRECISION
+      {
+      }
+}
+;
+
+
+typedef b3AlignedObjectArray<b3IndexedMesh>	IndexedMeshArray;
+
+///The b3TriangleIndexVertexArray allows to access multiple triangle meshes, by indexing into existing triangle/index arrays.
+///Additional meshes can be added using addIndexedMesh
+///No duplcate is made of the vertex/index data, it only indexes into external vertex/index arrays.
+///So keep those arrays around during the lifetime of this b3TriangleIndexVertexArray.
+B3_ATTRIBUTE_ALIGNED16( class) b3TriangleIndexVertexArray : public b3StridingMeshInterface
+{
+protected:
+	IndexedMeshArray	m_indexedMeshes;
+	int m_pad[2];
+	mutable int m_hasAabb; // using int instead of bool to maintain alignment
+	mutable b3Vector3 m_aabbMin;
+	mutable b3Vector3 m_aabbMax;
+
+public:
+
+	B3_DECLARE_ALIGNED_ALLOCATOR();
+
+	b3TriangleIndexVertexArray() : m_hasAabb(0)
+	{
+	}
+
+	virtual ~b3TriangleIndexVertexArray();
+
+	//just to be backwards compatible
+	b3TriangleIndexVertexArray(int numTriangles,int* triangleIndexBase,int triangleIndexStride,int numVertices,b3Scalar* vertexBase,int vertexStride);
+	
+	void	addIndexedMesh(const b3IndexedMesh& mesh, PHY_ScalarType indexType = PHY_INTEGER)
+	{
+		m_indexedMeshes.push_back(mesh);
+		m_indexedMeshes[m_indexedMeshes.size()-1].m_indexType = indexType;
+	}
+	
+	
+	virtual void	getLockedVertexIndexBase(unsigned char **vertexbase, int& numverts,PHY_ScalarType& type, int& vertexStride,unsigned char **indexbase,int & indexstride,int& numfaces,PHY_ScalarType& indicestype,int subpart=0);
+
+	virtual void	getLockedReadOnlyVertexIndexBase(const unsigned char **vertexbase, int& numverts,PHY_ScalarType& type, int& vertexStride,const unsigned char **indexbase,int & indexstride,int& numfaces,PHY_ScalarType& indicestype,int subpart=0) const;
+
+	/// unLockVertexBase finishes the access to a subpart of the triangle mesh
+	/// make a call to unLockVertexBase when the read and write access (using getLockedVertexIndexBase) is finished
+	virtual void	unLockVertexBase(int subpart) {(void)subpart;}
+
+	virtual void	unLockReadOnlyVertexBase(int subpart) const {(void)subpart;}
+
+	/// getNumSubParts returns the number of seperate subparts
+	/// each subpart has a continuous array of vertices and indices
+	virtual int		getNumSubParts() const { 
+		return (int)m_indexedMeshes.size();
+	}
+
+	IndexedMeshArray&	getIndexedMeshArray()
+	{
+		return m_indexedMeshes;
+	}
+
+	const IndexedMeshArray&	getIndexedMeshArray() const
+	{
+		return m_indexedMeshes;
+	}
+
+	virtual void	preallocateVertices(int numverts){(void) numverts;}
+	virtual void	preallocateIndices(int numindices){(void) numindices;}
+
+	virtual bool	hasPremadeAabb() const;
+	virtual void	setPremadeAabb(const b3Vector3& aabbMin, const b3Vector3& aabbMax ) const;
+	virtual void	getPremadeAabb(b3Vector3* aabbMin, b3Vector3* aabbMax ) const;
+
+}
+;
+
+#endif //B3_TRIANGLE_INDEX_VERTEX_ARRAY_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3VectorFloat4.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3VectorFloat4.h
new file mode 100644
index 0000000000..f6f65f7719
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3VectorFloat4.h
@@ -0,0 +1,11 @@
+#ifndef B3_VECTOR_FLOAT4_H
+#define B3_VECTOR_FLOAT4_H
+
+#include "Bullet3Common/b3Transform.h"
+
+//#define cross3(a,b) (a.cross(b))
+#define float4 b3Vector3
+//#define make_float4(x,y,z,w) b3Vector4(x,y,z,w)
+
+
+#endif //B3_VECTOR_FLOAT4_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3VoronoiSimplexSolver.cpp b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3VoronoiSimplexSolver.cpp
new file mode 100644
index 0000000000..cf3d5ef49d
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3VoronoiSimplexSolver.cpp
@@ -0,0 +1,609 @@
+
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2006 Erwin Coumans  http://continuousphysics.com/Bullet/
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+	
+	Elsevier CDROM license agreements grants nonexclusive license to use the software
+	for any purpose, commercial or non-commercial as long as the following credit is included
+	identifying the original source of the software:
+
+	Parts of the source are "from the book Real-Time Collision Detection by
+	Christer Ericson, published by Morgan Kaufmann Publishers,
+	(c) 2005 Elsevier Inc."
+		
+*/
+
+
+#include "b3VoronoiSimplexSolver.h"
+
+#define VERTA  0
+#define VERTB  1
+#define VERTC  2
+#define VERTD  3
+
+#define B3_CATCH_DEGENERATE_TETRAHEDRON 1
+void	b3VoronoiSimplexSolver::removeVertex(int index)
+{
+	
+	b3Assert(m_numVertices>0);
+	m_numVertices--;
+	m_simplexVectorW[index] = m_simplexVectorW[m_numVertices];
+	m_simplexPointsP[index] = m_simplexPointsP[m_numVertices];
+	m_simplexPointsQ[index] = m_simplexPointsQ[m_numVertices];
+}
+
+void	b3VoronoiSimplexSolver::reduceVertices (const b3UsageBitfield& usedVerts)
+{
+	if ((numVertices() >= 4) && (!usedVerts.usedVertexD))
+		removeVertex(3);
+
+	if ((numVertices() >= 3) && (!usedVerts.usedVertexC))
+		removeVertex(2);
+
+	if ((numVertices() >= 2) && (!usedVerts.usedVertexB))
+		removeVertex(1);
+	
+	if ((numVertices() >= 1) && (!usedVerts.usedVertexA))
+		removeVertex(0);
+
+}
+
+
+
+
+
+//clear the simplex, remove all the vertices
+void b3VoronoiSimplexSolver::reset()
+{
+	m_cachedValidClosest = false;
+	m_numVertices = 0;
+	m_needsUpdate = true;
+	m_lastW = b3MakeVector3(b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT),b3Scalar(B3_LARGE_FLOAT));
+	m_cachedBC.reset();
+}
+
+
+
+	//add a vertex
+void b3VoronoiSimplexSolver::addVertex(const b3Vector3& w, const b3Vector3& p, const b3Vector3& q)
+{
+	m_lastW = w;
+	m_needsUpdate = true;
+
+	m_simplexVectorW[m_numVertices] = w;
+	m_simplexPointsP[m_numVertices] = p;
+	m_simplexPointsQ[m_numVertices] = q;
+
+	m_numVertices++;
+}
+
+bool	b3VoronoiSimplexSolver::updateClosestVectorAndPoints()
+{
+	
+	if (m_needsUpdate)
+	{
+		m_cachedBC.reset();
+
+		m_needsUpdate = false;
+
+		switch (numVertices())
+		{
+		case 0:
+				m_cachedValidClosest = false;
+				break;
+		case 1:
+			{
+				m_cachedP1 = m_simplexPointsP[0];
+				m_cachedP2 = m_simplexPointsQ[0];
+				m_cachedV = m_cachedP1-m_cachedP2; //== m_simplexVectorW[0]
+				m_cachedBC.reset();
+				m_cachedBC.setBarycentricCoordinates(b3Scalar(1.),b3Scalar(0.),b3Scalar(0.),b3Scalar(0.));
+				m_cachedValidClosest = m_cachedBC.isValid();
+				break;
+			};
+		case 2:
+			{
+			//closest point origin from line segment
+					const b3Vector3& from = m_simplexVectorW[0];
+					const b3Vector3& to = m_simplexVectorW[1];
+					b3Vector3 nearest;
+
+					b3Vector3 p =b3MakeVector3(b3Scalar(0.),b3Scalar(0.),b3Scalar(0.));
+					b3Vector3 diff = p - from;
+					b3Vector3 v = to - from;
+					b3Scalar t = v.dot(diff);
+					
+					if (t > 0) {
+						b3Scalar dotVV = v.dot(v);
+						if (t < dotVV) {
+							t /= dotVV;
+							diff -= t*v;
+							m_cachedBC.m_usedVertices.usedVertexA = true;
+							m_cachedBC.m_usedVertices.usedVertexB = true;
+						} else {
+							t = 1;
+							diff -= v;
+							//reduce to 1 point
+							m_cachedBC.m_usedVertices.usedVertexB = true;
+						}
+					} else
+					{
+						t = 0;
+						//reduce to 1 point
+						m_cachedBC.m_usedVertices.usedVertexA = true;
+					}
+					m_cachedBC.setBarycentricCoordinates(1-t,t);
+					nearest = from + t*v;
+
+					m_cachedP1 = m_simplexPointsP[0] + t * (m_simplexPointsP[1] - m_simplexPointsP[0]);
+					m_cachedP2 = m_simplexPointsQ[0] + t * (m_simplexPointsQ[1] - m_simplexPointsQ[0]);
+					m_cachedV = m_cachedP1 - m_cachedP2;
+					
+					reduceVertices(m_cachedBC.m_usedVertices);
+
+					m_cachedValidClosest = m_cachedBC.isValid();
+					break;
+			}
+		case 3: 
+			{ 
+				//closest point origin from triangle 
+				b3Vector3 p =b3MakeVector3(b3Scalar(0.),b3Scalar(0.),b3Scalar(0.)); 
+
+				const b3Vector3& a = m_simplexVectorW[0]; 
+				const b3Vector3& b = m_simplexVectorW[1]; 
+				const b3Vector3& c = m_simplexVectorW[2]; 
+
+				closestPtPointTriangle(p,a,b,c,m_cachedBC); 
+				m_cachedP1 = m_simplexPointsP[0] * m_cachedBC.m_barycentricCoords[0] + 
+				m_simplexPointsP[1] * m_cachedBC.m_barycentricCoords[1] + 
+				m_simplexPointsP[2] * m_cachedBC.m_barycentricCoords[2]; 
+
+				m_cachedP2 = m_simplexPointsQ[0] * m_cachedBC.m_barycentricCoords[0] + 
+				m_simplexPointsQ[1] * m_cachedBC.m_barycentricCoords[1] + 
+				m_simplexPointsQ[2] * m_cachedBC.m_barycentricCoords[2]; 
+
+				m_cachedV = m_cachedP1-m_cachedP2; 
+
+				reduceVertices (m_cachedBC.m_usedVertices); 
+				m_cachedValidClosest = m_cachedBC.isValid(); 
+
+				break; 
+			}
+		case 4:
+			{
+
+				
+				b3Vector3 p =b3MakeVector3(b3Scalar(0.),b3Scalar(0.),b3Scalar(0.));
+				
+				const b3Vector3& a = m_simplexVectorW[0];
+				const b3Vector3& b = m_simplexVectorW[1];
+				const b3Vector3& c = m_simplexVectorW[2];
+				const b3Vector3& d = m_simplexVectorW[3];
+
+				bool hasSeperation = closestPtPointTetrahedron(p,a,b,c,d,m_cachedBC);
+
+				if (hasSeperation)
+				{
+
+					m_cachedP1 = m_simplexPointsP[0] * m_cachedBC.m_barycentricCoords[0] +
+						m_simplexPointsP[1] * m_cachedBC.m_barycentricCoords[1] +
+						m_simplexPointsP[2] * m_cachedBC.m_barycentricCoords[2] +
+						m_simplexPointsP[3] * m_cachedBC.m_barycentricCoords[3];
+
+					m_cachedP2 = m_simplexPointsQ[0] * m_cachedBC.m_barycentricCoords[0] +
+						m_simplexPointsQ[1] * m_cachedBC.m_barycentricCoords[1] +
+						m_simplexPointsQ[2] * m_cachedBC.m_barycentricCoords[2] +
+						m_simplexPointsQ[3] * m_cachedBC.m_barycentricCoords[3];
+
+					m_cachedV = m_cachedP1-m_cachedP2;
+					reduceVertices (m_cachedBC.m_usedVertices);
+				} else
+				{
+//					printf("sub distance got penetration\n");
+
+					if (m_cachedBC.m_degenerate)
+					{
+						m_cachedValidClosest = false;
+					} else
+					{
+						m_cachedValidClosest = true;
+						//degenerate case == false, penetration = true + zero
+						m_cachedV.setValue(b3Scalar(0.),b3Scalar(0.),b3Scalar(0.));
+					}
+					break;
+				}
+
+				m_cachedValidClosest = m_cachedBC.isValid();
+
+				//closest point origin from tetrahedron
+				break;
+			}
+		default:
+			{
+				m_cachedValidClosest = false;
+			}
+		};
+	}
+
+	return m_cachedValidClosest;
+
+}
+
+//return/calculate the closest vertex
+bool b3VoronoiSimplexSolver::closest(b3Vector3& v)
+{
+	bool succes = updateClosestVectorAndPoints();
+	v = m_cachedV;
+	return succes;
+}
+
+
+
+b3Scalar b3VoronoiSimplexSolver::maxVertex()
+{
+	int i, numverts = numVertices();
+	b3Scalar maxV = b3Scalar(0.);
+	for (i=0;i<numverts;i++)
+	{
+		b3Scalar curLen2 = m_simplexVectorW[i].length2();
+		if (maxV < curLen2)
+			maxV = curLen2;
+	}
+	return maxV;
+}
+
+
+
+	//return the current simplex
+int b3VoronoiSimplexSolver::getSimplex(b3Vector3 *pBuf, b3Vector3 *qBuf, b3Vector3 *yBuf) const
+{
+	int i;
+	for (i=0;i<numVertices();i++)
+	{
+		yBuf[i] = m_simplexVectorW[i];
+		pBuf[i] = m_simplexPointsP[i];
+		qBuf[i] = m_simplexPointsQ[i];
+	}
+	return numVertices();
+}
+
+
+
+
+bool b3VoronoiSimplexSolver::inSimplex(const b3Vector3& w)
+{
+	bool found = false;
+	int i, numverts = numVertices();
+	//b3Scalar maxV = b3Scalar(0.);
+	
+	//w is in the current (reduced) simplex
+	for (i=0;i<numverts;i++)
+	{
+#ifdef BT_USE_EQUAL_VERTEX_THRESHOLD
+		if ( m_simplexVectorW[i].distance2(w) <= m_equalVertexThreshold)
+#else
+		if (m_simplexVectorW[i] == w)
+#endif
+			found = true;
+	}
+
+	//check in case lastW is already removed
+	if (w == m_lastW)
+		return true;
+    	
+	return found;
+}
+
+void b3VoronoiSimplexSolver::backup_closest(b3Vector3& v) 
+{
+	v = m_cachedV;
+}
+
+
+bool b3VoronoiSimplexSolver::emptySimplex() const 
+{
+	return (numVertices() == 0);
+
+}
+
+void b3VoronoiSimplexSolver::compute_points(b3Vector3& p1, b3Vector3& p2) 
+{
+	updateClosestVectorAndPoints();
+	p1 = m_cachedP1;
+	p2 = m_cachedP2;
+
+}
+
+
+
+
+bool	b3VoronoiSimplexSolver::closestPtPointTriangle(const b3Vector3& p, const b3Vector3& a, const b3Vector3& b, const b3Vector3& c,b3SubSimplexClosestResult& result)
+{
+	result.m_usedVertices.reset();
+
+    // Check if P in vertex region outside A
+    b3Vector3 ab = b - a;
+    b3Vector3 ac = c - a;
+    b3Vector3 ap = p - a;
+    b3Scalar d1 = ab.dot(ap);
+    b3Scalar d2 = ac.dot(ap);
+    if (d1 <= b3Scalar(0.0) && d2 <= b3Scalar(0.0)) 
+	{
+		result.m_closestPointOnSimplex = a;
+		result.m_usedVertices.usedVertexA = true;
+		result.setBarycentricCoordinates(1,0,0);
+		return true;// a; // barycentric coordinates (1,0,0)
+	}
+
+    // Check if P in vertex region outside B
+    b3Vector3 bp = p - b;
+    b3Scalar d3 = ab.dot(bp);
+    b3Scalar d4 = ac.dot(bp);
+    if (d3 >= b3Scalar(0.0) && d4 <= d3) 
+	{
+		result.m_closestPointOnSimplex = b;
+		result.m_usedVertices.usedVertexB = true;
+		result.setBarycentricCoordinates(0,1,0);
+
+		return true; // b; // barycentric coordinates (0,1,0)
+	}
+    // Check if P in edge region of AB, if so return projection of P onto AB
+    b3Scalar vc = d1*d4 - d3*d2;
+    if (vc <= b3Scalar(0.0) && d1 >= b3Scalar(0.0) && d3 <= b3Scalar(0.0)) {
+        b3Scalar v = d1 / (d1 - d3);
+		result.m_closestPointOnSimplex = a + v * ab;
+		result.m_usedVertices.usedVertexA = true;
+		result.m_usedVertices.usedVertexB = true;
+		result.setBarycentricCoordinates(1-v,v,0);
+		return true;
+        //return a + v * ab; // barycentric coordinates (1-v,v,0)
+    }
+
+    // Check if P in vertex region outside C
+    b3Vector3 cp = p - c;
+    b3Scalar d5 = ab.dot(cp);
+    b3Scalar d6 = ac.dot(cp);
+    if (d6 >= b3Scalar(0.0) && d5 <= d6) 
+	{
+		result.m_closestPointOnSimplex = c;
+		result.m_usedVertices.usedVertexC = true;
+		result.setBarycentricCoordinates(0,0,1);
+		return true;//c; // barycentric coordinates (0,0,1)
+	}
+
+    // Check if P in edge region of AC, if so return projection of P onto AC
+    b3Scalar vb = d5*d2 - d1*d6;
+    if (vb <= b3Scalar(0.0) && d2 >= b3Scalar(0.0) && d6 <= b3Scalar(0.0)) {
+        b3Scalar w = d2 / (d2 - d6);
+		result.m_closestPointOnSimplex = a + w * ac;
+		result.m_usedVertices.usedVertexA = true;
+		result.m_usedVertices.usedVertexC = true;
+		result.setBarycentricCoordinates(1-w,0,w);
+		return true;
+        //return a + w * ac; // barycentric coordinates (1-w,0,w)
+    }
+
+    // Check if P in edge region of BC, if so return projection of P onto BC
+    b3Scalar va = d3*d6 - d5*d4;
+    if (va <= b3Scalar(0.0) && (d4 - d3) >= b3Scalar(0.0) && (d5 - d6) >= b3Scalar(0.0)) {
+        b3Scalar w = (d4 - d3) / ((d4 - d3) + (d5 - d6));
+		
+		result.m_closestPointOnSimplex = b + w * (c - b);
+		result.m_usedVertices.usedVertexB = true;
+		result.m_usedVertices.usedVertexC = true;
+		result.setBarycentricCoordinates(0,1-w,w);
+		return true;		
+       // return b + w * (c - b); // barycentric coordinates (0,1-w,w)
+    }
+
+    // P inside face region. Compute Q through its barycentric coordinates (u,v,w)
+    b3Scalar denom = b3Scalar(1.0) / (va + vb + vc);
+    b3Scalar v = vb * denom;
+    b3Scalar w = vc * denom;
+    
+	result.m_closestPointOnSimplex = a + ab * v + ac * w;
+	result.m_usedVertices.usedVertexA = true;
+	result.m_usedVertices.usedVertexB = true;
+	result.m_usedVertices.usedVertexC = true;
+	result.setBarycentricCoordinates(1-v-w,v,w);
+	
+	return true;
+//	return a + ab * v + ac * w; // = u*a + v*b + w*c, u = va * denom = b3Scalar(1.0) - v - w
+
+}
+
+
+
+
+
+/// Test if point p and d lie on opposite sides of plane through abc
+int b3VoronoiSimplexSolver::pointOutsideOfPlane(const b3Vector3& p, const b3Vector3& a, const b3Vector3& b, const b3Vector3& c, const b3Vector3& d)
+{
+	b3Vector3 normal = (b-a).cross(c-a);
+
+    b3Scalar signp = (p - a).dot(normal); // [AP AB AC]
+    b3Scalar signd = (d - a).dot( normal); // [AD AB AC]
+
+#ifdef B3_CATCH_DEGENERATE_TETRAHEDRON
+#ifdef BT_USE_DOUBLE_PRECISION
+if (signd * signd < (b3Scalar(1e-8) * b3Scalar(1e-8)))
+	{
+		return -1;
+	}
+#else
+	if (signd * signd < (b3Scalar(1e-4) * b3Scalar(1e-4)))
+	{
+//		printf("affine dependent/degenerate\n");//
+		return -1;
+	}
+#endif
+
+#endif
+	// Points on opposite sides if expression signs are opposite
+    return signp * signd < b3Scalar(0.);
+}
+
+
+bool	b3VoronoiSimplexSolver::closestPtPointTetrahedron(const b3Vector3& p, const b3Vector3& a, const b3Vector3& b, const b3Vector3& c, const b3Vector3& d, b3SubSimplexClosestResult& finalResult)
+{
+	b3SubSimplexClosestResult tempResult;
+
+    // Start out assuming point inside all halfspaces, so closest to itself
+	finalResult.m_closestPointOnSimplex = p;
+	finalResult.m_usedVertices.reset();
+    finalResult.m_usedVertices.usedVertexA = true;
+	finalResult.m_usedVertices.usedVertexB = true;
+	finalResult.m_usedVertices.usedVertexC = true;
+	finalResult.m_usedVertices.usedVertexD = true;
+
+    int pointOutsideABC = pointOutsideOfPlane(p, a, b, c, d);
+	int pointOutsideACD = pointOutsideOfPlane(p, a, c, d, b);
+  	int	pointOutsideADB = pointOutsideOfPlane(p, a, d, b, c);
+	int	pointOutsideBDC = pointOutsideOfPlane(p, b, d, c, a);
+
+   if (pointOutsideABC < 0 || pointOutsideACD < 0 || pointOutsideADB < 0 || pointOutsideBDC < 0)
+   {
+	   finalResult.m_degenerate = true;
+	   return false;
+   }
+
+   if (!pointOutsideABC  && !pointOutsideACD && !pointOutsideADB && !pointOutsideBDC)
+	 {
+		 return false;
+	 }
+
+
+    b3Scalar bestSqDist = FLT_MAX;
+    // If point outside face abc then compute closest point on abc
+	if (pointOutsideABC) 
+	{
+        closestPtPointTriangle(p, a, b, c,tempResult);
+		b3Vector3 q = tempResult.m_closestPointOnSimplex;
+		
+        b3Scalar sqDist = (q - p).dot( q - p);
+        // Update best closest point if (squared) distance is less than current best
+        if (sqDist < bestSqDist) {
+			bestSqDist = sqDist;
+			finalResult.m_closestPointOnSimplex = q;
+			//convert result bitmask!
+			finalResult.m_usedVertices.reset();
+			finalResult.m_usedVertices.usedVertexA = tempResult.m_usedVertices.usedVertexA;
+			finalResult.m_usedVertices.usedVertexB = tempResult.m_usedVertices.usedVertexB;
+			finalResult.m_usedVertices.usedVertexC = tempResult.m_usedVertices.usedVertexC;
+			finalResult.setBarycentricCoordinates(
+					tempResult.m_barycentricCoords[VERTA],
+					tempResult.m_barycentricCoords[VERTB],
+					tempResult.m_barycentricCoords[VERTC],
+					0
+			);
+
+		}
+    }
+  
+
+	// Repeat test for face acd
+	if (pointOutsideACD) 
+	{
+        closestPtPointTriangle(p, a, c, d,tempResult);
+		b3Vector3 q = tempResult.m_closestPointOnSimplex;
+		//convert result bitmask!
+
+        b3Scalar sqDist = (q - p).dot( q - p);
+        if (sqDist < bestSqDist) 
+		{
+			bestSqDist = sqDist;
+			finalResult.m_closestPointOnSimplex = q;
+			finalResult.m_usedVertices.reset();
+			finalResult.m_usedVertices.usedVertexA = tempResult.m_usedVertices.usedVertexA;
+
+			finalResult.m_usedVertices.usedVertexC = tempResult.m_usedVertices.usedVertexB;
+			finalResult.m_usedVertices.usedVertexD = tempResult.m_usedVertices.usedVertexC;
+			finalResult.setBarycentricCoordinates(
+					tempResult.m_barycentricCoords[VERTA],
+					0,
+					tempResult.m_barycentricCoords[VERTB],
+					tempResult.m_barycentricCoords[VERTC]
+			);
+
+		}
+    }
+    // Repeat test for face adb
+
+	
+	if (pointOutsideADB)
+	{
+		closestPtPointTriangle(p, a, d, b,tempResult);
+		b3Vector3 q = tempResult.m_closestPointOnSimplex;
+		//convert result bitmask!
+
+        b3Scalar sqDist = (q - p).dot( q - p);
+        if (sqDist < bestSqDist) 
+		{
+			bestSqDist = sqDist;
+			finalResult.m_closestPointOnSimplex = q;
+			finalResult.m_usedVertices.reset();
+			finalResult.m_usedVertices.usedVertexA = tempResult.m_usedVertices.usedVertexA;
+			finalResult.m_usedVertices.usedVertexB = tempResult.m_usedVertices.usedVertexC;
+			
+			finalResult.m_usedVertices.usedVertexD = tempResult.m_usedVertices.usedVertexB;
+			finalResult.setBarycentricCoordinates(
+					tempResult.m_barycentricCoords[VERTA],
+					tempResult.m_barycentricCoords[VERTC],
+					0,
+					tempResult.m_barycentricCoords[VERTB]
+			);
+
+		}
+    }
+    // Repeat test for face bdc
+    
+
+	if (pointOutsideBDC)
+	{
+        closestPtPointTriangle(p, b, d, c,tempResult);
+		b3Vector3 q = tempResult.m_closestPointOnSimplex;
+		//convert result bitmask!
+        b3Scalar sqDist = (q - p).dot( q - p);
+        if (sqDist < bestSqDist) 
+		{
+			bestSqDist = sqDist;
+			finalResult.m_closestPointOnSimplex = q;
+			finalResult.m_usedVertices.reset();
+			//
+			finalResult.m_usedVertices.usedVertexB = tempResult.m_usedVertices.usedVertexA;
+			finalResult.m_usedVertices.usedVertexC = tempResult.m_usedVertices.usedVertexC;
+			finalResult.m_usedVertices.usedVertexD = tempResult.m_usedVertices.usedVertexB;
+
+			finalResult.setBarycentricCoordinates(
+					0,
+					tempResult.m_barycentricCoords[VERTA],
+					tempResult.m_barycentricCoords[VERTC],
+					tempResult.m_barycentricCoords[VERTB]
+			);
+
+		}
+    }
+
+	//help! we ended up full !
+	
+	if (finalResult.m_usedVertices.usedVertexA &&
+		finalResult.m_usedVertices.usedVertexB &&
+		finalResult.m_usedVertices.usedVertexC &&
+		finalResult.m_usedVertices.usedVertexD) 
+	{
+		return true;
+	}
+
+    return true;
+}
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3VoronoiSimplexSolver.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3VoronoiSimplexSolver.h
new file mode 100644
index 0000000000..a6e27667d8
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/b3VoronoiSimplexSolver.h
@@ -0,0 +1,177 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2003-2006 Erwin Coumans  http://continuousphysics.com/Bullet/
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+
+
+#ifndef B3_VORONOI_SIMPLEX_SOLVER_H
+#define B3_VORONOI_SIMPLEX_SOLVER_H
+
+#include "Bullet3Common/b3Vector3.h"
+
+
+#define VORONOI_SIMPLEX_MAX_VERTS 5
+
+///disable next define, or use defaultCollisionConfiguration->getSimplexSolver()->setEqualVertexThreshold(0.f) to disable/configure
+//#define BT_USE_EQUAL_VERTEX_THRESHOLD
+#define VORONOI_DEFAULT_EQUAL_VERTEX_THRESHOLD 0.0001f
+
+
+struct b3UsageBitfield{
+	b3UsageBitfield()
+	{
+		reset();
+	}
+
+	void reset()
+	{
+		usedVertexA = false;
+		usedVertexB = false;
+		usedVertexC = false;
+		usedVertexD = false;
+	}
+	unsigned short usedVertexA	: 1;
+	unsigned short usedVertexB	: 1;
+	unsigned short usedVertexC	: 1;
+	unsigned short usedVertexD	: 1;
+	unsigned short unused1		: 1;
+	unsigned short unused2		: 1;
+	unsigned short unused3		: 1;
+	unsigned short unused4		: 1;
+};
+
+
+struct	b3SubSimplexClosestResult
+{
+	b3Vector3	m_closestPointOnSimplex;
+	//MASK for m_usedVertices
+	//stores the simplex vertex-usage, using the MASK, 
+	// if m_usedVertices & MASK then the related vertex is used
+	b3UsageBitfield	m_usedVertices;
+	b3Scalar	m_barycentricCoords[4];
+	bool m_degenerate;
+
+	void	reset()
+	{
+		m_degenerate = false;
+		setBarycentricCoordinates();
+		m_usedVertices.reset();
+	}
+	bool	isValid()
+	{
+		bool valid = (m_barycentricCoords[0] >= b3Scalar(0.)) &&
+			(m_barycentricCoords[1] >= b3Scalar(0.)) &&
+			(m_barycentricCoords[2] >= b3Scalar(0.)) &&
+			(m_barycentricCoords[3] >= b3Scalar(0.));
+
+
+		return valid;
+	}
+	void	setBarycentricCoordinates(b3Scalar a=b3Scalar(0.),b3Scalar b=b3Scalar(0.),b3Scalar c=b3Scalar(0.),b3Scalar d=b3Scalar(0.))
+	{
+		m_barycentricCoords[0] = a;
+		m_barycentricCoords[1] = b;
+		m_barycentricCoords[2] = c;
+		m_barycentricCoords[3] = d;
+	}
+
+};
+
+/// b3VoronoiSimplexSolver is an implementation of the closest point distance algorithm from a 1-4 points simplex to the origin.
+/// Can be used with GJK, as an alternative to Johnson distance algorithm.
+
+B3_ATTRIBUTE_ALIGNED16(class) b3VoronoiSimplexSolver 
+{
+public:
+
+	B3_DECLARE_ALIGNED_ALLOCATOR();
+
+	int	m_numVertices;
+
+	b3Vector3	m_simplexVectorW[VORONOI_SIMPLEX_MAX_VERTS];
+	b3Vector3	m_simplexPointsP[VORONOI_SIMPLEX_MAX_VERTS];
+	b3Vector3	m_simplexPointsQ[VORONOI_SIMPLEX_MAX_VERTS];
+
+	
+
+	b3Vector3	m_cachedP1;
+	b3Vector3	m_cachedP2;
+	b3Vector3	m_cachedV;
+	b3Vector3	m_lastW;
+	
+	b3Scalar	m_equalVertexThreshold;
+	bool		m_cachedValidClosest;
+
+
+	b3SubSimplexClosestResult m_cachedBC;
+
+	bool	m_needsUpdate;
+	
+	void	removeVertex(int index);
+	void	reduceVertices (const b3UsageBitfield& usedVerts);
+	bool	updateClosestVectorAndPoints();
+
+	bool	closestPtPointTetrahedron(const b3Vector3& p, const b3Vector3& a, const b3Vector3& b, const b3Vector3& c, const b3Vector3& d, b3SubSimplexClosestResult& finalResult);
+	int		pointOutsideOfPlane(const b3Vector3& p, const b3Vector3& a, const b3Vector3& b, const b3Vector3& c, const b3Vector3& d);
+	bool	closestPtPointTriangle(const b3Vector3& p, const b3Vector3& a, const b3Vector3& b, const b3Vector3& c,b3SubSimplexClosestResult& result);
+
+public:
+
+	b3VoronoiSimplexSolver()
+		:  m_equalVertexThreshold(VORONOI_DEFAULT_EQUAL_VERTEX_THRESHOLD)
+	{
+	}
+	 void reset();
+
+	 void addVertex(const b3Vector3& w, const b3Vector3& p, const b3Vector3& q);
+
+	 void	setEqualVertexThreshold(b3Scalar threshold)
+	 {
+		 m_equalVertexThreshold = threshold;
+	 }
+
+	 b3Scalar	getEqualVertexThreshold() const
+	 {
+		 return m_equalVertexThreshold;
+	 }
+
+	 bool closest(b3Vector3& v);
+
+	 b3Scalar maxVertex();
+
+	 bool fullSimplex() const
+	 {
+		 return (m_numVertices == 4);
+	 }
+
+	 int getSimplex(b3Vector3 *pBuf, b3Vector3 *qBuf, b3Vector3 *yBuf) const;
+
+	 bool inSimplex(const b3Vector3& w);
+	
+	 void backup_closest(b3Vector3& v) ;
+
+	 bool emptySimplex() const ;
+
+	 void compute_points(b3Vector3& p1, b3Vector3& p2) ;
+
+	 int numVertices() const 
+	 {
+		 return m_numVertices;
+	 }
+
+
+};
+
+#endif //B3_VORONOI_SIMPLEX_SOLVER_H
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/bvhTraversal.cl b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/bvhTraversal.cl
new file mode 100644
index 0000000000..faa413441c
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/bvhTraversal.cl
@@ -0,0 +1,283 @@
+//keep this enum in sync with the CPU version (in btCollidable.h)
+//written by Erwin Coumans
+
+#define SHAPE_CONVEX_HULL 3
+#define SHAPE_CONCAVE_TRIMESH 5
+#define TRIANGLE_NUM_CONVEX_FACES 5
+#define SHAPE_COMPOUND_OF_CONVEX_HULLS 6
+#define SHAPE_SPHERE 7
+
+typedef unsigned int u32;
+
+#define MAX_NUM_PARTS_IN_BITS 10
+
+///btQuantizedBvhNode is a compressed aabb node, 16 bytes.
+///Node can be used for leafnode or internal node. Leafnodes can point to 32-bit triangle index (non-negative range).
+typedef struct
+{
+	//12 bytes
+	unsigned short int	m_quantizedAabbMin[3];
+	unsigned short int	m_quantizedAabbMax[3];
+	//4 bytes
+	int	m_escapeIndexOrTriangleIndex;
+} btQuantizedBvhNode;
+
+typedef struct
+{
+	float4		m_aabbMin;
+	float4		m_aabbMax;
+	float4		m_quantization;
+	int			m_numNodes;
+	int			m_numSubTrees;
+	int			m_nodeOffset;
+	int			m_subTreeOffset;
+
+} b3BvhInfo;
+
+int	getTriangleIndex(const btQuantizedBvhNode* rootNode)
+{
+	unsigned int x=0;
+	unsigned int y = (~(x&0))<<(31-MAX_NUM_PARTS_IN_BITS);
+	// Get only the lower bits where the triangle index is stored
+	return (rootNode->m_escapeIndexOrTriangleIndex&~(y));
+}
+
+int isLeaf(const btQuantizedBvhNode* rootNode)
+{
+	//skipindex is negative (internal node), triangleindex >=0 (leafnode)
+	return (rootNode->m_escapeIndexOrTriangleIndex >= 0)? 1 : 0;
+}
+	
+int getEscapeIndex(const btQuantizedBvhNode* rootNode)
+{
+	return -rootNode->m_escapeIndexOrTriangleIndex;
+}
+
+typedef struct
+{
+	//12 bytes
+	unsigned short int	m_quantizedAabbMin[3];
+	unsigned short int	m_quantizedAabbMax[3];
+	//4 bytes, points to the root of the subtree
+	int			m_rootNodeIndex;
+	//4 bytes
+	int			m_subtreeSize;
+	int			m_padding[3];
+} btBvhSubtreeInfo;
+
+///keep this in sync with btCollidable.h
+typedef struct
+{
+	int m_numChildShapes;
+	int blaat2;
+	int m_shapeType;
+	int m_shapeIndex;
+	
+} btCollidableGpu;
+
+typedef struct
+{
+	float4	m_childPosition;
+	float4	m_childOrientation;
+	int m_shapeIndex;
+	int m_unused0;
+	int m_unused1;
+	int m_unused2;
+} btGpuChildShape;
+
+
+typedef struct
+{
+	float4 m_pos;
+	float4 m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_collidableIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} BodyData;
+
+typedef struct 
+{
+	union
+	{
+		float4	m_min;
+		float   m_minElems[4];
+		int			m_minIndices[4];
+	};
+	union
+	{
+		float4	m_max;
+		float   m_maxElems[4];
+		int			m_maxIndices[4];
+	};
+} btAabbCL;
+
+
+int testQuantizedAabbAgainstQuantizedAabb(
+								const unsigned short int* aabbMin1,
+								const unsigned short int* aabbMax1,
+								const unsigned short int* aabbMin2,
+								const unsigned short int* aabbMax2)
+{
+	//int overlap = 1;
+	if (aabbMin1[0] > aabbMax2[0])
+		return 0;
+	if (aabbMax1[0] < aabbMin2[0])
+		return 0;
+	if (aabbMin1[1] > aabbMax2[1])
+		return 0;
+	if (aabbMax1[1] < aabbMin2[1])
+		return 0;
+	if (aabbMin1[2] > aabbMax2[2])
+		return 0;
+	if (aabbMax1[2] < aabbMin2[2])
+		return 0;
+	return 1;
+	//overlap = ((aabbMin1[0] > aabbMax2[0]) || (aabbMax1[0] < aabbMin2[0])) ? 0 : overlap;
+	//overlap = ((aabbMin1[2] > aabbMax2[2]) || (aabbMax1[2] < aabbMin2[2])) ? 0 : overlap;
+	//overlap = ((aabbMin1[1] > aabbMax2[1]) || (aabbMax1[1] < aabbMin2[1])) ? 0 : overlap;
+	//return overlap;
+}
+
+
+void quantizeWithClamp(unsigned short* out, float4 point2,int isMax, float4 bvhAabbMin, float4 bvhAabbMax, float4 bvhQuantization)
+{
+	float4 clampedPoint = max(point2,bvhAabbMin);
+	clampedPoint = min (clampedPoint, bvhAabbMax);
+
+	float4 v = (clampedPoint - bvhAabbMin) * bvhQuantization;
+	if (isMax)
+	{
+		out[0] = (unsigned short) (((unsigned short)(v.x+1.f) | 1));
+		out[1] = (unsigned short) (((unsigned short)(v.y+1.f) | 1));
+		out[2] = (unsigned short) (((unsigned short)(v.z+1.f) | 1));
+	} else
+	{
+		out[0] = (unsigned short) (((unsigned short)(v.x) & 0xfffe));
+		out[1] = (unsigned short) (((unsigned short)(v.y) & 0xfffe));
+		out[2] = (unsigned short) (((unsigned short)(v.z) & 0xfffe));
+	}
+
+}
+
+
+// work-in-progress
+__kernel void   bvhTraversalKernel( __global const int4* pairs, 
+									__global const BodyData* rigidBodies, 
+									__global const btCollidableGpu* collidables,
+									__global btAabbCL* aabbs,
+									__global int4* concavePairsOut,
+									__global volatile int* numConcavePairsOut,
+									__global const btBvhSubtreeInfo* subtreeHeadersRoot,
+									__global const btQuantizedBvhNode* quantizedNodesRoot,
+									__global const b3BvhInfo* bvhInfos,
+									int numPairs,
+									int maxNumConcavePairsCapacity)
+{
+	int id = get_global_id(0);
+	if (id>=numPairs)
+		return;
+	
+	int bodyIndexA = pairs[id].x;
+	int bodyIndexB = pairs[id].y;
+	int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+	int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+	
+	//once the broadphase avoids static-static pairs, we can remove this test
+	if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))
+	{
+		return;
+	}
+		
+	if (collidables[collidableIndexA].m_shapeType!=SHAPE_CONCAVE_TRIMESH)
+		return;
+
+	int shapeTypeB = collidables[collidableIndexB].m_shapeType;
+		
+	if (shapeTypeB!=SHAPE_CONVEX_HULL &&
+		shapeTypeB!=SHAPE_SPHERE	&&
+		shapeTypeB!=SHAPE_COMPOUND_OF_CONVEX_HULLS
+		)
+		return;
+
+	b3BvhInfo bvhInfo = bvhInfos[collidables[collidableIndexA].m_numChildShapes];
+
+	float4 bvhAabbMin = bvhInfo.m_aabbMin;
+	float4 bvhAabbMax = bvhInfo.m_aabbMax;
+	float4 bvhQuantization = bvhInfo.m_quantization;
+	int numSubtreeHeaders = bvhInfo.m_numSubTrees;
+	__global const btBvhSubtreeInfo* subtreeHeaders = &subtreeHeadersRoot[bvhInfo.m_subTreeOffset];
+	__global const btQuantizedBvhNode* quantizedNodes = &quantizedNodesRoot[bvhInfo.m_nodeOffset];
+	
+
+	unsigned short int quantizedQueryAabbMin[3];
+	unsigned short int quantizedQueryAabbMax[3];
+	quantizeWithClamp(quantizedQueryAabbMin,aabbs[bodyIndexB].m_min,false,bvhAabbMin, bvhAabbMax,bvhQuantization);
+	quantizeWithClamp(quantizedQueryAabbMax,aabbs[bodyIndexB].m_max,true ,bvhAabbMin, bvhAabbMax,bvhQuantization);
+	
+	for (int i=0;i<numSubtreeHeaders;i++)
+	{
+		btBvhSubtreeInfo subtree = subtreeHeaders[i];
+				
+		int overlap = testQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,subtree.m_quantizedAabbMin,subtree.m_quantizedAabbMax);
+		if (overlap != 0)
+		{
+			int startNodeIndex = subtree.m_rootNodeIndex;
+			int endNodeIndex = subtree.m_rootNodeIndex+subtree.m_subtreeSize;
+			int curIndex = startNodeIndex;
+			int escapeIndex;
+			int isLeafNode;
+			int aabbOverlap;
+			while (curIndex < endNodeIndex)
+			{
+				btQuantizedBvhNode rootNode = quantizedNodes[curIndex];
+				aabbOverlap = testQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,rootNode.m_quantizedAabbMin,rootNode.m_quantizedAabbMax);
+				isLeafNode = isLeaf(&rootNode);
+				if (aabbOverlap)
+				{
+					if (isLeafNode)
+					{
+						int triangleIndex = getTriangleIndex(&rootNode);
+						if (shapeTypeB==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+						{
+								int numChildrenB = collidables[collidableIndexB].m_numChildShapes;
+								int pairIdx = atomic_add(numConcavePairsOut,numChildrenB);
+								for (int b=0;b<numChildrenB;b++)
+								{
+									if ((pairIdx+b)<maxNumConcavePairsCapacity)
+									{
+										int childShapeIndexB = collidables[collidableIndexB].m_shapeIndex+b;
+										int4 newPair = (int4)(bodyIndexA,bodyIndexB,triangleIndex,childShapeIndexB);
+										concavePairsOut[pairIdx+b] = newPair;
+									}
+								}
+						} else
+						{
+							int pairIdx = atomic_inc(numConcavePairsOut);
+							if (pairIdx<maxNumConcavePairsCapacity)
+							{
+								int4 newPair = (int4)(bodyIndexA,bodyIndexB,triangleIndex,0);
+								concavePairsOut[pairIdx] = newPair;
+							}
+						}
+					} 
+					curIndex++;
+				} else
+				{
+					if (isLeafNode)
+					{
+						curIndex++;
+					} else
+					{
+						escapeIndex = getEscapeIndex(&rootNode);
+						curIndex += escapeIndex;
+					}
+				}
+			}
+		}
+	}
+
+}
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/bvhTraversal.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/bvhTraversal.h
new file mode 100644
index 0000000000..4b3b49eae8
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/bvhTraversal.h
@@ -0,0 +1,258 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* bvhTraversalKernelCL= \
+"//keep this enum in sync with the CPU version (in btCollidable.h)\n"
+"//written by Erwin Coumans\n"
+"#define SHAPE_CONVEX_HULL 3\n"
+"#define SHAPE_CONCAVE_TRIMESH 5\n"
+"#define TRIANGLE_NUM_CONVEX_FACES 5\n"
+"#define SHAPE_COMPOUND_OF_CONVEX_HULLS 6\n"
+"#define SHAPE_SPHERE 7\n"
+"typedef unsigned int u32;\n"
+"#define MAX_NUM_PARTS_IN_BITS 10\n"
+"///btQuantizedBvhNode is a compressed aabb node, 16 bytes.\n"
+"///Node can be used for leafnode or internal node. Leafnodes can point to 32-bit triangle index (non-negative range).\n"
+"typedef struct\n"
+"{\n"
+"	//12 bytes\n"
+"	unsigned short int	m_quantizedAabbMin[3];\n"
+"	unsigned short int	m_quantizedAabbMax[3];\n"
+"	//4 bytes\n"
+"	int	m_escapeIndexOrTriangleIndex;\n"
+"} btQuantizedBvhNode;\n"
+"typedef struct\n"
+"{\n"
+"	float4		m_aabbMin;\n"
+"	float4		m_aabbMax;\n"
+"	float4		m_quantization;\n"
+"	int			m_numNodes;\n"
+"	int			m_numSubTrees;\n"
+"	int			m_nodeOffset;\n"
+"	int			m_subTreeOffset;\n"
+"} b3BvhInfo;\n"
+"int	getTriangleIndex(const btQuantizedBvhNode* rootNode)\n"
+"{\n"
+"	unsigned int x=0;\n"
+"	unsigned int y = (~(x&0))<<(31-MAX_NUM_PARTS_IN_BITS);\n"
+"	// Get only the lower bits where the triangle index is stored\n"
+"	return (rootNode->m_escapeIndexOrTriangleIndex&~(y));\n"
+"}\n"
+"int isLeaf(const btQuantizedBvhNode* rootNode)\n"
+"{\n"
+"	//skipindex is negative (internal node), triangleindex >=0 (leafnode)\n"
+"	return (rootNode->m_escapeIndexOrTriangleIndex >= 0)? 1 : 0;\n"
+"}\n"
+"	\n"
+"int getEscapeIndex(const btQuantizedBvhNode* rootNode)\n"
+"{\n"
+"	return -rootNode->m_escapeIndexOrTriangleIndex;\n"
+"}\n"
+"typedef struct\n"
+"{\n"
+"	//12 bytes\n"
+"	unsigned short int	m_quantizedAabbMin[3];\n"
+"	unsigned short int	m_quantizedAabbMax[3];\n"
+"	//4 bytes, points to the root of the subtree\n"
+"	int			m_rootNodeIndex;\n"
+"	//4 bytes\n"
+"	int			m_subtreeSize;\n"
+"	int			m_padding[3];\n"
+"} btBvhSubtreeInfo;\n"
+"///keep this in sync with btCollidable.h\n"
+"typedef struct\n"
+"{\n"
+"	int m_numChildShapes;\n"
+"	int blaat2;\n"
+"	int m_shapeType;\n"
+"	int m_shapeIndex;\n"
+"	\n"
+"} btCollidableGpu;\n"
+"typedef struct\n"
+"{\n"
+"	float4	m_childPosition;\n"
+"	float4	m_childOrientation;\n"
+"	int m_shapeIndex;\n"
+"	int m_unused0;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"} btGpuChildShape;\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	float4 m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"	u32 m_collidableIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} BodyData;\n"
+"typedef struct \n"
+"{\n"
+"	union\n"
+"	{\n"
+"		float4	m_min;\n"
+"		float   m_minElems[4];\n"
+"		int			m_minIndices[4];\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float4	m_max;\n"
+"		float   m_maxElems[4];\n"
+"		int			m_maxIndices[4];\n"
+"	};\n"
+"} btAabbCL;\n"
+"int testQuantizedAabbAgainstQuantizedAabb(\n"
+"								const unsigned short int* aabbMin1,\n"
+"								const unsigned short int* aabbMax1,\n"
+"								const unsigned short int* aabbMin2,\n"
+"								const unsigned short int* aabbMax2)\n"
+"{\n"
+"	//int overlap = 1;\n"
+"	if (aabbMin1[0] > aabbMax2[0])\n"
+"		return 0;\n"
+"	if (aabbMax1[0] < aabbMin2[0])\n"
+"		return 0;\n"
+"	if (aabbMin1[1] > aabbMax2[1])\n"
+"		return 0;\n"
+"	if (aabbMax1[1] < aabbMin2[1])\n"
+"		return 0;\n"
+"	if (aabbMin1[2] > aabbMax2[2])\n"
+"		return 0;\n"
+"	if (aabbMax1[2] < aabbMin2[2])\n"
+"		return 0;\n"
+"	return 1;\n"
+"	//overlap = ((aabbMin1[0] > aabbMax2[0]) || (aabbMax1[0] < aabbMin2[0])) ? 0 : overlap;\n"
+"	//overlap = ((aabbMin1[2] > aabbMax2[2]) || (aabbMax1[2] < aabbMin2[2])) ? 0 : overlap;\n"
+"	//overlap = ((aabbMin1[1] > aabbMax2[1]) || (aabbMax1[1] < aabbMin2[1])) ? 0 : overlap;\n"
+"	//return overlap;\n"
+"}\n"
+"void quantizeWithClamp(unsigned short* out, float4 point2,int isMax, float4 bvhAabbMin, float4 bvhAabbMax, float4 bvhQuantization)\n"
+"{\n"
+"	float4 clampedPoint = max(point2,bvhAabbMin);\n"
+"	clampedPoint = min (clampedPoint, bvhAabbMax);\n"
+"	float4 v = (clampedPoint - bvhAabbMin) * bvhQuantization;\n"
+"	if (isMax)\n"
+"	{\n"
+"		out[0] = (unsigned short) (((unsigned short)(v.x+1.f) | 1));\n"
+"		out[1] = (unsigned short) (((unsigned short)(v.y+1.f) | 1));\n"
+"		out[2] = (unsigned short) (((unsigned short)(v.z+1.f) | 1));\n"
+"	} else\n"
+"	{\n"
+"		out[0] = (unsigned short) (((unsigned short)(v.x) & 0xfffe));\n"
+"		out[1] = (unsigned short) (((unsigned short)(v.y) & 0xfffe));\n"
+"		out[2] = (unsigned short) (((unsigned short)(v.z) & 0xfffe));\n"
+"	}\n"
+"}\n"
+"// work-in-progress\n"
+"__kernel void   bvhTraversalKernel( __global const int4* pairs, \n"
+"									__global const BodyData* rigidBodies, \n"
+"									__global const btCollidableGpu* collidables,\n"
+"									__global btAabbCL* aabbs,\n"
+"									__global int4* concavePairsOut,\n"
+"									__global volatile int* numConcavePairsOut,\n"
+"									__global const btBvhSubtreeInfo* subtreeHeadersRoot,\n"
+"									__global const btQuantizedBvhNode* quantizedNodesRoot,\n"
+"									__global const b3BvhInfo* bvhInfos,\n"
+"									int numPairs,\n"
+"									int maxNumConcavePairsCapacity)\n"
+"{\n"
+"	int id = get_global_id(0);\n"
+"	if (id>=numPairs)\n"
+"		return;\n"
+"	\n"
+"	int bodyIndexA = pairs[id].x;\n"
+"	int bodyIndexB = pairs[id].y;\n"
+"	int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"	int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"	\n"
+"	//once the broadphase avoids static-static pairs, we can remove this test\n"
+"	if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))\n"
+"	{\n"
+"		return;\n"
+"	}\n"
+"		\n"
+"	if (collidables[collidableIndexA].m_shapeType!=SHAPE_CONCAVE_TRIMESH)\n"
+"		return;\n"
+"	int shapeTypeB = collidables[collidableIndexB].m_shapeType;\n"
+"		\n"
+"	if (shapeTypeB!=SHAPE_CONVEX_HULL &&\n"
+"		shapeTypeB!=SHAPE_SPHERE	&&\n"
+"		shapeTypeB!=SHAPE_COMPOUND_OF_CONVEX_HULLS\n"
+"		)\n"
+"		return;\n"
+"	b3BvhInfo bvhInfo = bvhInfos[collidables[collidableIndexA].m_numChildShapes];\n"
+"	float4 bvhAabbMin = bvhInfo.m_aabbMin;\n"
+"	float4 bvhAabbMax = bvhInfo.m_aabbMax;\n"
+"	float4 bvhQuantization = bvhInfo.m_quantization;\n"
+"	int numSubtreeHeaders = bvhInfo.m_numSubTrees;\n"
+"	__global const btBvhSubtreeInfo* subtreeHeaders = &subtreeHeadersRoot[bvhInfo.m_subTreeOffset];\n"
+"	__global const btQuantizedBvhNode* quantizedNodes = &quantizedNodesRoot[bvhInfo.m_nodeOffset];\n"
+"	\n"
+"	unsigned short int quantizedQueryAabbMin[3];\n"
+"	unsigned short int quantizedQueryAabbMax[3];\n"
+"	quantizeWithClamp(quantizedQueryAabbMin,aabbs[bodyIndexB].m_min,false,bvhAabbMin, bvhAabbMax,bvhQuantization);\n"
+"	quantizeWithClamp(quantizedQueryAabbMax,aabbs[bodyIndexB].m_max,true ,bvhAabbMin, bvhAabbMax,bvhQuantization);\n"
+"	\n"
+"	for (int i=0;i<numSubtreeHeaders;i++)\n"
+"	{\n"
+"		btBvhSubtreeInfo subtree = subtreeHeaders[i];\n"
+"				\n"
+"		int overlap = testQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,subtree.m_quantizedAabbMin,subtree.m_quantizedAabbMax);\n"
+"		if (overlap != 0)\n"
+"		{\n"
+"			int startNodeIndex = subtree.m_rootNodeIndex;\n"
+"			int endNodeIndex = subtree.m_rootNodeIndex+subtree.m_subtreeSize;\n"
+"			int curIndex = startNodeIndex;\n"
+"			int escapeIndex;\n"
+"			int isLeafNode;\n"
+"			int aabbOverlap;\n"
+"			while (curIndex < endNodeIndex)\n"
+"			{\n"
+"				btQuantizedBvhNode rootNode = quantizedNodes[curIndex];\n"
+"				aabbOverlap = testQuantizedAabbAgainstQuantizedAabb(quantizedQueryAabbMin,quantizedQueryAabbMax,rootNode.m_quantizedAabbMin,rootNode.m_quantizedAabbMax);\n"
+"				isLeafNode = isLeaf(&rootNode);\n"
+"				if (aabbOverlap)\n"
+"				{\n"
+"					if (isLeafNode)\n"
+"					{\n"
+"						int triangleIndex = getTriangleIndex(&rootNode);\n"
+"						if (shapeTypeB==SHAPE_COMPOUND_OF_CONVEX_HULLS)\n"
+"						{\n"
+"								int numChildrenB = collidables[collidableIndexB].m_numChildShapes;\n"
+"								int pairIdx = atomic_add(numConcavePairsOut,numChildrenB);\n"
+"								for (int b=0;b<numChildrenB;b++)\n"
+"								{\n"
+"									if ((pairIdx+b)<maxNumConcavePairsCapacity)\n"
+"									{\n"
+"										int childShapeIndexB = collidables[collidableIndexB].m_shapeIndex+b;\n"
+"										int4 newPair = (int4)(bodyIndexA,bodyIndexB,triangleIndex,childShapeIndexB);\n"
+"										concavePairsOut[pairIdx+b] = newPair;\n"
+"									}\n"
+"								}\n"
+"						} else\n"
+"						{\n"
+"							int pairIdx = atomic_inc(numConcavePairsOut);\n"
+"							if (pairIdx<maxNumConcavePairsCapacity)\n"
+"							{\n"
+"								int4 newPair = (int4)(bodyIndexA,bodyIndexB,triangleIndex,0);\n"
+"								concavePairsOut[pairIdx] = newPair;\n"
+"							}\n"
+"						}\n"
+"					} \n"
+"					curIndex++;\n"
+"				} else\n"
+"				{\n"
+"					if (isLeafNode)\n"
+"					{\n"
+"						curIndex++;\n"
+"					} else\n"
+"					{\n"
+"						escapeIndex = getEscapeIndex(&rootNode);\n"
+"						curIndex += escapeIndex;\n"
+"					}\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/mpr.cl b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/mpr.cl
new file mode 100644
index 0000000000..e754f4e1da
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/mpr.cl
@@ -0,0 +1,311 @@
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3MprPenetration.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Contact4Data.h"
+
+#define AppendInc(x, out) out = atomic_inc(x)
+#define GET_NPOINTS(x) (x).m_worldNormalOnB.w
+#ifdef cl_ext_atomic_counters_32
+	#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+	#define counter32_t volatile __global int*
+#endif
+
+
+__kernel void   mprPenetrationKernel( __global int4* pairs,
+																					__global const b3RigidBodyData_t* rigidBodies, 
+																					__global const b3Collidable_t* collidables,
+																					__global const b3ConvexPolyhedronData_t* convexShapes, 
+																					__global const float4* vertices,
+																					__global float4* separatingNormals,
+																					__global int* hasSeparatingAxis,
+																					__global struct b3Contact4Data* restrict globalContactsOut,
+																					counter32_t nGlobalContactsOut,
+																					int contactCapacity,
+																					int numPairs)
+{
+	int i = get_global_id(0);
+	int pairIndex = i;
+	if (i<numPairs)
+	{
+		int bodyIndexA = pairs[i].x;
+		int bodyIndexB = pairs[i].y;
+
+		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+	
+		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+		
+		
+		//once the broadphase avoids static-static pairs, we can remove this test
+		if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))
+		{
+			return;
+		}
+		
+
+		if ((collidables[collidableIndexA].m_shapeType!=SHAPE_CONVEX_HULL) ||(collidables[collidableIndexB].m_shapeType!=SHAPE_CONVEX_HULL))
+		{
+			return;
+		}
+
+		float depthOut;
+		b3Float4 dirOut;
+		b3Float4 posOut;
+
+
+		int res = b3MprPenetration(pairIndex, bodyIndexA, bodyIndexB,rigidBodies,convexShapes,collidables,vertices,separatingNormals,hasSeparatingAxis,&depthOut, &dirOut, &posOut);
+		
+		
+		
+		
+
+		if (res==0)
+		{
+			//add a contact
+
+			int dstIdx;
+			AppendInc( nGlobalContactsOut, dstIdx );
+			if (dstIdx<contactCapacity)
+			{
+				pairs[pairIndex].z = dstIdx;
+				__global struct b3Contact4Data* c = globalContactsOut + dstIdx;
+				c->m_worldNormalOnB = -dirOut;//normal;
+				c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+				c->m_batchIdx = pairIndex;
+				int bodyA = pairs[pairIndex].x;
+				int bodyB = pairs[pairIndex].y;
+				c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0 ? -bodyA:bodyA;
+				c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0 ? -bodyB:bodyB;
+				c->m_childIndexA = -1;
+				c->m_childIndexB = -1;
+				//for (int i=0;i<nContacts;i++)
+				posOut.w = -depthOut;
+				c->m_worldPosB[0] = posOut;//localPoints[contactIdx[i]];
+				GET_NPOINTS(*c) = 1;//nContacts;
+			}
+		}
+
+	}
+}
+
+typedef float4 Quaternion;
+#define make_float4 (float4)
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = make_float4(a.xyz,0.f);
+	float4 b1 = make_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+
+
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross3( a, b );
+	ans += a.w*b+b.w*a;
+//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+__inline
+float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(qtMul(q,vcpy),qInv);
+	return out;
+}
+
+__inline
+float4 transform(const float4* p, const float4* translation, const Quaternion* orientation)
+{
+	return qtRotate( *orientation, *p ) + (*translation);
+}
+
+
+__inline
+float4 qtInvRotate(const Quaternion q, float4 vec)
+{
+	return qtRotate( qtInvert( q ), vec );
+}
+
+
+inline void project(__global const b3ConvexPolyhedronData_t* hull,  const float4 pos, const float4 orn, 
+const float4* dir, __global const float4* vertices, float* min, float* max)
+{
+	min[0] = FLT_MAX;
+	max[0] = -FLT_MAX;
+	int numVerts = hull->m_numVertices;
+
+	const float4 localDir = qtInvRotate(orn,*dir);
+	float offset = dot(pos,*dir);
+	for(int i=0;i<numVerts;i++)
+	{
+		float dp = dot(vertices[hull->m_vertexOffset+i],localDir);
+		if(dp < min[0])	
+			min[0] = dp;
+		if(dp > max[0])	
+			max[0] = dp;
+	}
+	if(min[0]>max[0])
+	{
+		float tmp = min[0];
+		min[0] = max[0];
+		max[0] = tmp;
+	}
+	min[0] += offset;
+	max[0] += offset;
+}
+
+
+bool findSeparatingAxisUnitSphere(	__global const b3ConvexPolyhedronData_t* hullA, __global const b3ConvexPolyhedronData_t* hullB, 
+	const float4 posA1,
+	const float4 ornA,
+	const float4 posB1,
+	const float4 ornB,
+	const float4 DeltaC2,
+	__global const float4* vertices,
+	__global const float4* unitSphereDirections,
+	int numUnitSphereDirections,
+	float4* sep,
+	float* dmin)
+{
+	
+	float4 posA = posA1;
+	posA.w = 0.f;
+	float4 posB = posB1;
+	posB.w = 0.f;
+
+	int curPlaneTests=0;
+
+	int curEdgeEdge = 0;
+	// Test unit sphere directions
+	for (int i=0;i<numUnitSphereDirections;i++)
+	{
+
+		float4 crossje;
+		crossje = unitSphereDirections[i];	
+
+		if (dot3F4(DeltaC2,crossje)>0)
+			crossje *= -1.f;
+		{
+			float dist;
+			bool result = true;
+			float Min0,Max0;
+			float Min1,Max1;
+			project(hullA,posA,ornA,&crossje,vertices, &Min0, &Max0);
+			project(hullB,posB,ornB,&crossje,vertices, &Min1, &Max1);
+		
+			if(Max0<Min1 || Max1<Min0)
+				return false;
+		
+			float d0 = Max0 - Min1;
+			float d1 = Max1 - Min0;
+			dist = d0<d1 ? d0:d1;
+			result = true;
+	
+			if(dist<*dmin)
+			{
+				*dmin = dist;
+				*sep = crossje;
+			}
+		}
+	}
+
+	
+	if((dot3F4(-DeltaC2,*sep))>0.0f)
+	{
+		*sep = -(*sep);
+	}
+	return true;
+}
+
+
+
+__kernel void   findSeparatingAxisUnitSphereKernel( __global const int4* pairs, 
+																					__global const b3RigidBodyData_t* rigidBodies, 
+																					__global const b3Collidable_t* collidables,
+																					__global const b3ConvexPolyhedronData_t* convexShapes, 
+																					__global const float4* vertices,
+																					__global const float4* unitSphereDirections,
+																					__global  float4* separatingNormals,
+																					__global  int* hasSeparatingAxis,
+																					__global  float* dmins,
+																					int numUnitSphereDirections,
+																					int numPairs
+																					)
+{
+
+	int i = get_global_id(0);
+	
+	if (i<numPairs)
+	{
+
+		if (hasSeparatingAxis[i])
+		{
+	
+			int bodyIndexA = pairs[i].x;
+			int bodyIndexB = pairs[i].y;
+	
+			int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+			int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+		
+			int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+			int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+			
+			
+			int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+	
+			float dmin = dmins[i];
+	
+			float4 posA = rigidBodies[bodyIndexA].m_pos;
+			posA.w = 0.f;
+			float4 posB = rigidBodies[bodyIndexB].m_pos;
+			posB.w = 0.f;
+			float4 c0local = convexShapes[shapeIndexA].m_localCenter;
+			float4 ornA = rigidBodies[bodyIndexA].m_quat;
+			float4 c0 = transform(&c0local, &posA, &ornA);
+			float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+			float4 ornB =rigidBodies[bodyIndexB].m_quat;
+			float4 c1 = transform(&c1local,&posB,&ornB);
+			const float4 DeltaC2 = c0 - c1;
+			float4 sepNormal = separatingNormals[i];
+			
+			int numEdgeEdgeDirections = convexShapes[shapeIndexA].m_numUniqueEdges*convexShapes[shapeIndexB].m_numUniqueEdges;
+			if (numEdgeEdgeDirections>numUnitSphereDirections)
+			{
+				bool sepEE = findSeparatingAxisUnitSphere(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,
+																										posB,ornB,
+																										DeltaC2,
+																										vertices,unitSphereDirections,numUnitSphereDirections,&sepNormal,&dmin);
+				if (!sepEE)
+				{
+					hasSeparatingAxis[i] = 0;
+				} else
+				{
+					hasSeparatingAxis[i] = 1;
+					separatingNormals[i] = sepNormal;
+				}
+			}
+		}		//if (hasSeparatingAxis[i])
+	}//(i<numPairs)
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/mprKernels.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/mprKernels.h
new file mode 100644
index 0000000000..7ed4b382c3
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/mprKernels.h
@@ -0,0 +1,1446 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* mprKernelsCL= \
+"/***\n"
+" * ---------------------------------\n"
+" * Copyright (c)2012 Daniel Fiser <danfis@danfis.cz>\n"
+" *\n"
+" *  This file was ported from mpr.c file, part of libccd.\n"
+" *  The Minkoski Portal Refinement implementation was ported \n"
+" *  to OpenCL by Erwin Coumans for the Bullet 3 Physics library.\n"
+" *  at http://github.com/erwincoumans/bullet3\n"
+" *\n"
+" *  Distributed under the OSI-approved BSD License (the \"License\");\n"
+" *  see <http://www.opensource.org/licenses/bsd-license.php>.\n"
+" *  This software is distributed WITHOUT ANY WARRANTY; without even the\n"
+" *  implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n"
+" *  See the License for more information.\n"
+" */\n"
+"#ifndef B3_MPR_PENETRATION_H\n"
+"#define B3_MPR_PENETRATION_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#define B3_PLATFORM_DEFINITIONS_H\n"
+"struct MyTest\n"
+"{\n"
+"	int bla;\n"
+"};\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"//keep B3_LARGE_FLOAT*B3_LARGE_FLOAT < FLT_MAX\n"
+"#define B3_LARGE_FLOAT 1e18f\n"
+"#define B3_INFINITY 1e18f\n"
+"#define b3Assert(a)\n"
+"#define b3ConstArray(a) __global const a*\n"
+"#define b3AtomicInc atomic_inc\n"
+"#define b3AtomicAdd atomic_add\n"
+"#define b3Fabs fabs\n"
+"#define b3Sqrt native_sqrt\n"
+"#define b3Sin native_sin\n"
+"#define b3Cos native_cos\n"
+"#define B3_STATIC\n"
+"#endif\n"
+"#endif\n"
+"#ifndef B3_FLOAT4_H\n"
+"#define B3_FLOAT4_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif\n"
+"#endif\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Float4;\n"
+"	#define b3Float4ConstArg const b3Float4\n"
+"	#define b3MakeFloat4 (float4)\n"
+"	float b3Dot3F4(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return dot(a1, b1);\n"
+"	}\n"
+"	b3Float4 b3Cross3(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return cross(a1, b1);\n"
+"	}\n"
+"	#define b3MinFloat4 min\n"
+"	#define b3MaxFloat4 max\n"
+"	#define b3Normalized(a) normalize(a)\n"
+"#endif \n"
+"		\n"
+"inline bool b3IsAlmostZero(b3Float4ConstArg v)\n"
+"{\n"
+"	if(b3Fabs(v.x)>1e-6 || b3Fabs(v.y)>1e-6 || b3Fabs(v.z)>1e-6)	\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"inline int    b3MaxDot( b3Float4ConstArg vec, __global const b3Float4* vecArray, int vecLen, float* dotOut )\n"
+"{\n"
+"    float maxDot = -B3_INFINITY;\n"
+"    int i = 0;\n"
+"    int ptIndex = -1;\n"
+"    for( i = 0; i < vecLen; i++ )\n"
+"    {\n"
+"        float dot = b3Dot3F4(vecArray[i],vec);\n"
+"            \n"
+"        if( dot > maxDot )\n"
+"        {\n"
+"            maxDot = dot;\n"
+"            ptIndex = i;\n"
+"        }\n"
+"    }\n"
+"	b3Assert(ptIndex>=0);\n"
+"    if (ptIndex<0)\n"
+"	{\n"
+"		ptIndex = 0;\n"
+"	}\n"
+"    *dotOut = maxDot;\n"
+"    return ptIndex;\n"
+"}\n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_RIGIDBODY_DATA_H\n"
+"#define B3_RIGIDBODY_DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_QUAT_H\n"
+"#define B3_QUAT_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif\n"
+"#endif\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Quat;\n"
+"	#define b3QuatConstArg const b3Quat\n"
+"	\n"
+"	\n"
+"inline float4 b3FastNormalize4(float4 v)\n"
+"{\n"
+"	v = (float4)(v.xyz,0.f);\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"	\n"
+"inline b3Quat b3QuatMul(b3Quat a, b3Quat b);\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in);\n"
+"inline b3Quat b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec);\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatMul(b3QuatConstArg a, b3QuatConstArg b)\n"
+"{\n"
+"	b3Quat ans;\n"
+"	ans = b3Cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - b3Dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in)\n"
+"{\n"
+"	b3Quat q;\n"
+"	q=in;\n"
+"	//return b3FastNormalize4(in);\n"
+"	float len = native_sqrt(dot(q, q));\n"
+"	if(len > 0.f)\n"
+"	{\n"
+"		q *= 1.f / len;\n"
+"	}\n"
+"	else\n"
+"	{\n"
+"		q.x = q.y = q.z = 0.f;\n"
+"		q.w = 1.f;\n"
+"	}\n"
+"	return q;\n"
+"}\n"
+"inline float4 b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	b3Quat qInv = b3QuatInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = b3QuatMul(b3QuatMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline float4 b3QuatInvRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	return b3QuatRotate( b3QuatInvert( q ), vec );\n"
+"}\n"
+"inline b3Float4 b3TransformPoint(b3Float4ConstArg point, b3Float4ConstArg translation, b3QuatConstArg  orientation)\n"
+"{\n"
+"	return b3QuatRotate( orientation, point ) + (translation);\n"
+"}\n"
+"	\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"#ifndef B3_MAT3x3_H\n"
+"#define B3_MAT3x3_H\n"
+"#ifndef B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"typedef struct\n"
+"{\n"
+"	b3Float4 m_row[3];\n"
+"}b3Mat3x3;\n"
+"#define b3Mat3x3ConstArg const b3Mat3x3\n"
+"#define b3GetRow(m,row) (m.m_row[row])\n"
+"inline b3Mat3x3 b3QuatGetRotationMatrix(b3Quat quat)\n"
+"{\n"
+"	b3Float4 quat2 = (b3Float4)(quat.x*quat.x, quat.y*quat.y, quat.z*quat.z, 0.f);\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0].x=1-2*quat2.y-2*quat2.z;\n"
+"	out.m_row[0].y=2*quat.x*quat.y-2*quat.w*quat.z;\n"
+"	out.m_row[0].z=2*quat.x*quat.z+2*quat.w*quat.y;\n"
+"	out.m_row[0].w = 0.f;\n"
+"	out.m_row[1].x=2*quat.x*quat.y+2*quat.w*quat.z;\n"
+"	out.m_row[1].y=1-2*quat2.x-2*quat2.z;\n"
+"	out.m_row[1].z=2*quat.y*quat.z-2*quat.w*quat.x;\n"
+"	out.m_row[1].w = 0.f;\n"
+"	out.m_row[2].x=2*quat.x*quat.z-2*quat.w*quat.y;\n"
+"	out.m_row[2].y=2*quat.y*quat.z+2*quat.w*quat.x;\n"
+"	out.m_row[2].z=1-2*quat2.x-2*quat2.y;\n"
+"	out.m_row[2].w = 0.f;\n"
+"	return out;\n"
+"}\n"
+"inline b3Mat3x3 b3AbsoluteMat3x3(b3Mat3x3ConstArg matIn)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = fabs(matIn.m_row[0]);\n"
+"	out.m_row[1] = fabs(matIn.m_row[1]);\n"
+"	out.m_row[2] = fabs(matIn.m_row[2]);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtZero();\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity();\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m);\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b);\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Mat3x3 mtZero()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(0.f);\n"
+"	m.m_row[1] = (b3Float4)(0.f);\n"
+"	m.m_row[2] = (b3Float4)(0.f);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(1,0,0,0);\n"
+"	m.m_row[1] = (b3Float4)(0,1,0,0);\n"
+"	m.m_row[2] = (b3Float4)(0,0,1,0);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = (b3Float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (b3Float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (b3Float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Mat3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	b3Mat3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = b3Dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = b3Dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = b3Dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b)\n"
+"{\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a.m_row[0], b );\n"
+"	ans.y = b3Dot3F4( a.m_row[1], b );\n"
+"	ans.z = b3Dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Float4 colx = b3MakeFloat4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	b3Float4 coly = b3MakeFloat4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	b3Float4 colz = b3MakeFloat4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a, colx );\n"
+"	ans.y = b3Dot3F4( a, coly );\n"
+"	ans.z = b3Dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"#endif\n"
+"#endif //B3_MAT3x3_H\n"
+"typedef struct b3RigidBodyData b3RigidBodyData_t;\n"
+"struct b3RigidBodyData\n"
+"{\n"
+"	b3Float4				m_pos;\n"
+"	b3Quat					m_quat;\n"
+"	b3Float4				m_linVel;\n"
+"	b3Float4				m_angVel;\n"
+"	int 					m_collidableIdx;\n"
+"	float 				m_invMass;\n"
+"	float 				m_restituitionCoeff;\n"
+"	float 				m_frictionCoeff;\n"
+"};\n"
+"typedef struct b3InertiaData b3InertiaData_t;\n"
+"struct b3InertiaData\n"
+"{\n"
+"	b3Mat3x3 m_invInertiaWorld;\n"
+"	b3Mat3x3 m_initInvInertia;\n"
+"};\n"
+"#endif //B3_RIGIDBODY_DATA_H\n"
+"	\n"
+"#ifndef B3_CONVEX_POLYHEDRON_DATA_H\n"
+"#define B3_CONVEX_POLYHEDRON_DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"typedef struct b3GpuFace b3GpuFace_t;\n"
+"struct b3GpuFace\n"
+"{\n"
+"	b3Float4 m_plane;\n"
+"	int m_indexOffset;\n"
+"	int m_numIndices;\n"
+"	int m_unusedPadding1;\n"
+"	int m_unusedPadding2;\n"
+"};\n"
+"typedef struct b3ConvexPolyhedronData b3ConvexPolyhedronData_t;\n"
+"struct b3ConvexPolyhedronData\n"
+"{\n"
+"	b3Float4		m_localCenter;\n"
+"	b3Float4		m_extents;\n"
+"	b3Float4		mC;\n"
+"	b3Float4		mE;\n"
+"	float			m_radius;\n"
+"	int	m_faceOffset;\n"
+"	int m_numFaces;\n"
+"	int	m_numVertices;\n"
+"	int m_vertexOffset;\n"
+"	int	m_uniqueEdgesOffset;\n"
+"	int	m_numUniqueEdges;\n"
+"	int m_unused;\n"
+"};\n"
+"#endif //B3_CONVEX_POLYHEDRON_DATA_H\n"
+"#ifndef B3_COLLIDABLE_H\n"
+"#define B3_COLLIDABLE_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"enum b3ShapeTypes\n"
+"{\n"
+"	SHAPE_HEIGHT_FIELD=1,\n"
+"	SHAPE_CONVEX_HULL=3,\n"
+"	SHAPE_PLANE=4,\n"
+"	SHAPE_CONCAVE_TRIMESH=5,\n"
+"	SHAPE_COMPOUND_OF_CONVEX_HULLS=6,\n"
+"	SHAPE_SPHERE=7,\n"
+"	MAX_NUM_SHAPE_TYPES,\n"
+"};\n"
+"typedef struct b3Collidable b3Collidable_t;\n"
+"struct b3Collidable\n"
+"{\n"
+"	union {\n"
+"		int m_numChildShapes;\n"
+"		int m_bvhIndex;\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float m_radius;\n"
+"		int	m_compoundBvhIndex;\n"
+"	};\n"
+"	int m_shapeType;\n"
+"	int m_shapeIndex;\n"
+"};\n"
+"typedef struct b3GpuChildShape b3GpuChildShape_t;\n"
+"struct b3GpuChildShape\n"
+"{\n"
+"	b3Float4	m_childPosition;\n"
+"	b3Quat		m_childOrientation;\n"
+"	int m_shapeIndex;\n"
+"	int m_unused0;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"};\n"
+"struct b3CompoundOverlappingPair\n"
+"{\n"
+"	int m_bodyIndexA;\n"
+"	int m_bodyIndexB;\n"
+"//	int	m_pairType;\n"
+"	int m_childShapeIndexA;\n"
+"	int m_childShapeIndexB;\n"
+"};\n"
+"#endif //B3_COLLIDABLE_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#define B3_MPR_SQRT sqrt\n"
+"#endif\n"
+"#define B3_MPR_FMIN(x, y) ((x) < (y) ? (x) : (y))\n"
+"#define B3_MPR_FABS fabs\n"
+"#define B3_MPR_TOLERANCE 1E-6f\n"
+"#define B3_MPR_MAX_ITERATIONS 1000\n"
+"struct _b3MprSupport_t \n"
+"{\n"
+"    b3Float4 v;  //!< Support point in minkowski sum\n"
+"    b3Float4 v1; //!< Support point in obj1\n"
+"    b3Float4 v2; //!< Support point in obj2\n"
+"};\n"
+"typedef struct _b3MprSupport_t b3MprSupport_t;\n"
+"struct _b3MprSimplex_t \n"
+"{\n"
+"    b3MprSupport_t ps[4];\n"
+"    int last; //!< index of last added point\n"
+"};\n"
+"typedef struct _b3MprSimplex_t b3MprSimplex_t;\n"
+"inline b3MprSupport_t* b3MprSimplexPointW(b3MprSimplex_t *s, int idx)\n"
+"{\n"
+"    return &s->ps[idx];\n"
+"}\n"
+"inline void b3MprSimplexSetSize(b3MprSimplex_t *s, int size)\n"
+"{\n"
+"    s->last = size - 1;\n"
+"}\n"
+"inline int b3MprSimplexSize(const b3MprSimplex_t *s)\n"
+"{\n"
+"    return s->last + 1;\n"
+"}\n"
+"inline const b3MprSupport_t* b3MprSimplexPoint(const b3MprSimplex_t* s, int idx)\n"
+"{\n"
+"    // here is no check on boundaries\n"
+"    return &s->ps[idx];\n"
+"}\n"
+"inline void b3MprSupportCopy(b3MprSupport_t *d, const b3MprSupport_t *s)\n"
+"{\n"
+"    *d = *s;\n"
+"}\n"
+"inline void b3MprSimplexSet(b3MprSimplex_t *s, size_t pos, const b3MprSupport_t *a)\n"
+"{\n"
+"    b3MprSupportCopy(s->ps + pos, a);\n"
+"}\n"
+"inline void b3MprSimplexSwap(b3MprSimplex_t *s, size_t pos1, size_t pos2)\n"
+"{\n"
+"    b3MprSupport_t supp;\n"
+"    b3MprSupportCopy(&supp, &s->ps[pos1]);\n"
+"    b3MprSupportCopy(&s->ps[pos1], &s->ps[pos2]);\n"
+"    b3MprSupportCopy(&s->ps[pos2], &supp);\n"
+"}\n"
+"inline int b3MprIsZero(float val)\n"
+"{\n"
+"    return B3_MPR_FABS(val) < FLT_EPSILON;\n"
+"}\n"
+"inline int b3MprEq(float _a, float _b)\n"
+"{\n"
+"    float ab;\n"
+"    float a, b;\n"
+"    ab = B3_MPR_FABS(_a - _b);\n"
+"    if (B3_MPR_FABS(ab) < FLT_EPSILON)\n"
+"        return 1;\n"
+"    a = B3_MPR_FABS(_a);\n"
+"    b = B3_MPR_FABS(_b);\n"
+"    if (b > a){\n"
+"        return ab < FLT_EPSILON * b;\n"
+"    }else{\n"
+"        return ab < FLT_EPSILON * a;\n"
+"    }\n"
+"}\n"
+"inline int b3MprVec3Eq(const b3Float4* a, const b3Float4 *b)\n"
+"{\n"
+"    return b3MprEq((*a).x, (*b).x)\n"
+"            && b3MprEq((*a).y, (*b).y)\n"
+"            && b3MprEq((*a).z, (*b).z);\n"
+"}\n"
+"inline b3Float4 b3LocalGetSupportVertex(b3Float4ConstArg supportVec,__global const b3ConvexPolyhedronData_t* hull, 	b3ConstArray(b3Float4) verticesA)\n"
+"{\n"
+"	b3Float4 supVec = b3MakeFloat4(0,0,0,0);\n"
+"	float maxDot = -B3_LARGE_FLOAT;\n"
+"    if( 0 < hull->m_numVertices )\n"
+"    {\n"
+"        const b3Float4 scaled = supportVec;\n"
+"		int index = b3MaxDot(scaled, &verticesA[hull->m_vertexOffset], hull->m_numVertices, &maxDot);\n"
+"        return verticesA[hull->m_vertexOffset+index];\n"
+"    }\n"
+"    return supVec;\n"
+"}\n"
+"B3_STATIC void b3MprConvexSupport(int pairIndex,int bodyIndex,  b3ConstArray(b3RigidBodyData_t) cpuBodyBuf, \n"
+"													b3ConstArray(b3ConvexPolyhedronData_t) cpuConvexData, \n"
+"													b3ConstArray(b3Collidable_t)				cpuCollidables,\n"
+"													b3ConstArray(b3Float4)					cpuVertices,\n"
+"													__global b3Float4* sepAxis,\n"
+"														const b3Float4* _dir, b3Float4* outp, int logme)\n"
+"{\n"
+"	//dir is in worldspace, move to local space\n"
+"	\n"
+"	b3Float4 pos = cpuBodyBuf[bodyIndex].m_pos;\n"
+"	b3Quat orn = cpuBodyBuf[bodyIndex].m_quat;\n"
+"	\n"
+"	b3Float4 dir = b3MakeFloat4((*_dir).x,(*_dir).y,(*_dir).z,0.f);\n"
+"	\n"
+"	const b3Float4 localDir = b3QuatRotate(b3QuatInverse(orn),dir);\n"
+"	\n"
+"	//find local support vertex\n"
+"	int colIndex = cpuBodyBuf[bodyIndex].m_collidableIdx;\n"
+"	\n"
+"	b3Assert(cpuCollidables[colIndex].m_shapeType==SHAPE_CONVEX_HULL);\n"
+"	__global const b3ConvexPolyhedronData_t* hull = &cpuConvexData[cpuCollidables[colIndex].m_shapeIndex];\n"
+"	\n"
+"	b3Float4 pInA;\n"
+"	if (logme)\n"
+"	{\n"
+"		b3Float4 supVec = b3MakeFloat4(0,0,0,0);\n"
+"		float maxDot = -B3_LARGE_FLOAT;\n"
+"		if( 0 < hull->m_numVertices )\n"
+"		{\n"
+"			const b3Float4 scaled = localDir;\n"
+"			int index = b3MaxDot(scaled, &cpuVertices[hull->m_vertexOffset], hull->m_numVertices, &maxDot);\n"
+"			pInA = cpuVertices[hull->m_vertexOffset+index];\n"
+"			\n"
+"		}\n"
+"	} else\n"
+"	{\n"
+"		pInA = b3LocalGetSupportVertex(localDir,hull,cpuVertices);\n"
+"	}\n"
+"	//move vertex to world space\n"
+"	*outp = b3TransformPoint(pInA,pos,orn);\n"
+"	\n"
+"}\n"
+"inline void b3MprSupport(int pairIndex,int bodyIndexA, int bodyIndexB,   b3ConstArray(b3RigidBodyData_t) cpuBodyBuf, \n"
+"													b3ConstArray(b3ConvexPolyhedronData_t) cpuConvexData, \n"
+"													b3ConstArray(b3Collidable_t)				cpuCollidables,\n"
+"													b3ConstArray(b3Float4)					cpuVertices,\n"
+"													__global b3Float4* sepAxis,\n"
+"													const b3Float4* _dir, b3MprSupport_t *supp)\n"
+"{\n"
+"    b3Float4 dir;\n"
+"	dir = *_dir;\n"
+"	b3MprConvexSupport(pairIndex,bodyIndexA,cpuBodyBuf,cpuConvexData,cpuCollidables,cpuVertices,sepAxis,&dir, &supp->v1,0);\n"
+"    dir = *_dir*-1.f;\n"
+"	b3MprConvexSupport(pairIndex,bodyIndexB,cpuBodyBuf,cpuConvexData,cpuCollidables,cpuVertices,sepAxis,&dir, &supp->v2,0);\n"
+"    supp->v = supp->v1 - supp->v2;\n"
+"}\n"
+"inline void b3FindOrigin(int bodyIndexA, int bodyIndexB, b3ConstArray(b3RigidBodyData_t) cpuBodyBuf, b3MprSupport_t *center)\n"
+"{\n"
+"    center->v1 = cpuBodyBuf[bodyIndexA].m_pos;\n"
+"	center->v2 = cpuBodyBuf[bodyIndexB].m_pos;\n"
+"    center->v = center->v1 - center->v2;\n"
+"}\n"
+"inline void b3MprVec3Set(b3Float4 *v, float x, float y, float z)\n"
+"{\n"
+"	(*v).x = x;\n"
+"	(*v).y = y;\n"
+"	(*v).z = z;\n"
+"	(*v).w = 0.f;\n"
+"}\n"
+"inline void b3MprVec3Add(b3Float4 *v, const b3Float4 *w)\n"
+"{\n"
+"    (*v).x += (*w).x;\n"
+"    (*v).y += (*w).y;\n"
+"    (*v).z += (*w).z;\n"
+"}\n"
+"inline void b3MprVec3Copy(b3Float4 *v, const b3Float4 *w)\n"
+"{\n"
+"    *v = *w;\n"
+"}\n"
+"inline void b3MprVec3Scale(b3Float4 *d, float k)\n"
+"{\n"
+"    *d *= k;\n"
+"}\n"
+"inline float b3MprVec3Dot(const b3Float4 *a, const b3Float4 *b)\n"
+"{\n"
+"    float dot;\n"
+"	dot = b3Dot3F4(*a,*b);\n"
+"    return dot;\n"
+"}\n"
+"inline float b3MprVec3Len2(const b3Float4 *v)\n"
+"{\n"
+"    return b3MprVec3Dot(v, v);\n"
+"}\n"
+"inline void b3MprVec3Normalize(b3Float4 *d)\n"
+"{\n"
+"    float k = 1.f / B3_MPR_SQRT(b3MprVec3Len2(d));\n"
+"    b3MprVec3Scale(d, k);\n"
+"}\n"
+"inline void b3MprVec3Cross(b3Float4 *d, const b3Float4 *a, const b3Float4 *b)\n"
+"{\n"
+"	*d = b3Cross3(*a,*b);\n"
+"	\n"
+"}\n"
+"inline void b3MprVec3Sub2(b3Float4 *d, const b3Float4 *v, const b3Float4 *w)\n"
+"{\n"
+"	*d = *v - *w;\n"
+"}\n"
+"inline void b3PortalDir(const b3MprSimplex_t *portal, b3Float4 *dir)\n"
+"{\n"
+"    b3Float4 v2v1, v3v1;\n"
+"    b3MprVec3Sub2(&v2v1, &b3MprSimplexPoint(portal, 2)->v,\n"
+"                       &b3MprSimplexPoint(portal, 1)->v);\n"
+"    b3MprVec3Sub2(&v3v1, &b3MprSimplexPoint(portal, 3)->v,\n"
+"                       &b3MprSimplexPoint(portal, 1)->v);\n"
+"    b3MprVec3Cross(dir, &v2v1, &v3v1);\n"
+"    b3MprVec3Normalize(dir);\n"
+"}\n"
+"inline int portalEncapsulesOrigin(const b3MprSimplex_t *portal,\n"
+"                                       const b3Float4 *dir)\n"
+"{\n"
+"    float dot;\n"
+"    dot = b3MprVec3Dot(dir, &b3MprSimplexPoint(portal, 1)->v);\n"
+"    return b3MprIsZero(dot) || dot > 0.f;\n"
+"}\n"
+"inline int portalReachTolerance(const b3MprSimplex_t *portal,\n"
+"                                     const b3MprSupport_t *v4,\n"
+"                                     const b3Float4 *dir)\n"
+"{\n"
+"    float dv1, dv2, dv3, dv4;\n"
+"    float dot1, dot2, dot3;\n"
+"    // find the smallest dot product of dir and {v1-v4, v2-v4, v3-v4}\n"
+"    dv1 = b3MprVec3Dot(&b3MprSimplexPoint(portal, 1)->v, dir);\n"
+"    dv2 = b3MprVec3Dot(&b3MprSimplexPoint(portal, 2)->v, dir);\n"
+"    dv3 = b3MprVec3Dot(&b3MprSimplexPoint(portal, 3)->v, dir);\n"
+"    dv4 = b3MprVec3Dot(&v4->v, dir);\n"
+"    dot1 = dv4 - dv1;\n"
+"    dot2 = dv4 - dv2;\n"
+"    dot3 = dv4 - dv3;\n"
+"    dot1 = B3_MPR_FMIN(dot1, dot2);\n"
+"    dot1 = B3_MPR_FMIN(dot1, dot3);\n"
+"    return b3MprEq(dot1, B3_MPR_TOLERANCE) || dot1 < B3_MPR_TOLERANCE;\n"
+"}\n"
+"inline int portalCanEncapsuleOrigin(const b3MprSimplex_t *portal,   \n"
+"                                         const b3MprSupport_t *v4,\n"
+"                                         const b3Float4 *dir)\n"
+"{\n"
+"    float dot;\n"
+"    dot = b3MprVec3Dot(&v4->v, dir);\n"
+"    return b3MprIsZero(dot) || dot > 0.f;\n"
+"}\n"
+"inline void b3ExpandPortal(b3MprSimplex_t *portal,\n"
+"                              const b3MprSupport_t *v4)\n"
+"{\n"
+"    float dot;\n"
+"    b3Float4 v4v0;\n"
+"    b3MprVec3Cross(&v4v0, &v4->v, &b3MprSimplexPoint(portal, 0)->v);\n"
+"    dot = b3MprVec3Dot(&b3MprSimplexPoint(portal, 1)->v, &v4v0);\n"
+"    if (dot > 0.f){\n"
+"        dot = b3MprVec3Dot(&b3MprSimplexPoint(portal, 2)->v, &v4v0);\n"
+"        if (dot > 0.f){\n"
+"            b3MprSimplexSet(portal, 1, v4);\n"
+"        }else{\n"
+"            b3MprSimplexSet(portal, 3, v4);\n"
+"        }\n"
+"    }else{\n"
+"        dot = b3MprVec3Dot(&b3MprSimplexPoint(portal, 3)->v, &v4v0);\n"
+"        if (dot > 0.f){\n"
+"            b3MprSimplexSet(portal, 2, v4);\n"
+"        }else{\n"
+"            b3MprSimplexSet(portal, 1, v4);\n"
+"        }\n"
+"    }\n"
+"}\n"
+"B3_STATIC int b3DiscoverPortal(int pairIndex, int bodyIndexA, int bodyIndexB,  b3ConstArray(b3RigidBodyData_t) cpuBodyBuf, \n"
+"													b3ConstArray(b3ConvexPolyhedronData_t) cpuConvexData, \n"
+"													b3ConstArray(b3Collidable_t)				cpuCollidables,\n"
+"													b3ConstArray(b3Float4)					cpuVertices,\n"
+"													__global b3Float4* sepAxis,\n"
+"													__global int*	hasSepAxis,\n"
+"													b3MprSimplex_t *portal)\n"
+"{\n"
+"    b3Float4 dir, va, vb;\n"
+"    float dot;\n"
+"    int cont;\n"
+"	\n"
+"	\n"
+"    // vertex 0 is center of portal\n"
+"    b3FindOrigin(bodyIndexA,bodyIndexB,cpuBodyBuf, b3MprSimplexPointW(portal, 0));\n"
+"    // vertex 0 is center of portal\n"
+"    b3MprSimplexSetSize(portal, 1);\n"
+"	\n"
+"	b3Float4 zero = b3MakeFloat4(0,0,0,0);\n"
+"	b3Float4* b3mpr_vec3_origin = &zero;\n"
+"    if (b3MprVec3Eq(&b3MprSimplexPoint(portal, 0)->v, b3mpr_vec3_origin)){\n"
+"        // Portal's center lies on origin (0,0,0) => we know that objects\n"
+"        // intersect but we would need to know penetration info.\n"
+"        // So move center little bit...\n"
+"        b3MprVec3Set(&va, FLT_EPSILON * 10.f, 0.f, 0.f);\n"
+"        b3MprVec3Add(&b3MprSimplexPointW(portal, 0)->v, &va);\n"
+"    }\n"
+"    // vertex 1 = support in direction of origin\n"
+"    b3MprVec3Copy(&dir, &b3MprSimplexPoint(portal, 0)->v);\n"
+"    b3MprVec3Scale(&dir, -1.f);\n"
+"    b3MprVec3Normalize(&dir);\n"
+"    b3MprSupport(pairIndex,bodyIndexA,bodyIndexB,cpuBodyBuf,cpuConvexData,cpuCollidables,cpuVertices, sepAxis,&dir, b3MprSimplexPointW(portal, 1));\n"
+"    b3MprSimplexSetSize(portal, 2);\n"
+"    // test if origin isn't outside of v1\n"
+"    dot = b3MprVec3Dot(&b3MprSimplexPoint(portal, 1)->v, &dir);\n"
+"	\n"
+"    if (b3MprIsZero(dot) || dot < 0.f)\n"
+"        return -1;\n"
+"    // vertex 2\n"
+"    b3MprVec3Cross(&dir, &b3MprSimplexPoint(portal, 0)->v,\n"
+"                       &b3MprSimplexPoint(portal, 1)->v);\n"
+"    if (b3MprIsZero(b3MprVec3Len2(&dir))){\n"
+"        if (b3MprVec3Eq(&b3MprSimplexPoint(portal, 1)->v, b3mpr_vec3_origin)){\n"
+"            // origin lies on v1\n"
+"            return 1;\n"
+"        }else{\n"
+"            // origin lies on v0-v1 segment\n"
+"            return 2;\n"
+"        }\n"
+"    }\n"
+"    b3MprVec3Normalize(&dir);\n"
+"	 b3MprSupport(pairIndex,bodyIndexA,bodyIndexB,cpuBodyBuf,cpuConvexData,cpuCollidables,cpuVertices, sepAxis,&dir, b3MprSimplexPointW(portal, 2));\n"
+"    \n"
+"    dot = b3MprVec3Dot(&b3MprSimplexPoint(portal, 2)->v, &dir);\n"
+"    if (b3MprIsZero(dot) || dot < 0.f)\n"
+"        return -1;\n"
+"    b3MprSimplexSetSize(portal, 3);\n"
+"    // vertex 3 direction\n"
+"    b3MprVec3Sub2(&va, &b3MprSimplexPoint(portal, 1)->v,\n"
+"                     &b3MprSimplexPoint(portal, 0)->v);\n"
+"    b3MprVec3Sub2(&vb, &b3MprSimplexPoint(portal, 2)->v,\n"
+"                     &b3MprSimplexPoint(portal, 0)->v);\n"
+"    b3MprVec3Cross(&dir, &va, &vb);\n"
+"    b3MprVec3Normalize(&dir);\n"
+"    // it is better to form portal faces to be oriented \"outside\" origin\n"
+"    dot = b3MprVec3Dot(&dir, &b3MprSimplexPoint(portal, 0)->v);\n"
+"    if (dot > 0.f){\n"
+"        b3MprSimplexSwap(portal, 1, 2);\n"
+"        b3MprVec3Scale(&dir, -1.f);\n"
+"    }\n"
+"    while (b3MprSimplexSize(portal) < 4){\n"
+"		 b3MprSupport(pairIndex,bodyIndexA,bodyIndexB,cpuBodyBuf,cpuConvexData,cpuCollidables,cpuVertices, sepAxis,&dir, b3MprSimplexPointW(portal, 3));\n"
+"        \n"
+"        dot = b3MprVec3Dot(&b3MprSimplexPoint(portal, 3)->v, &dir);\n"
+"        if (b3MprIsZero(dot) || dot < 0.f)\n"
+"            return -1;\n"
+"        cont = 0;\n"
+"        // test if origin is outside (v1, v0, v3) - set v2 as v3 and\n"
+"        // continue\n"
+"        b3MprVec3Cross(&va, &b3MprSimplexPoint(portal, 1)->v,\n"
+"                          &b3MprSimplexPoint(portal, 3)->v);\n"
+"        dot = b3MprVec3Dot(&va, &b3MprSimplexPoint(portal, 0)->v);\n"
+"        if (dot < 0.f && !b3MprIsZero(dot)){\n"
+"            b3MprSimplexSet(portal, 2, b3MprSimplexPoint(portal, 3));\n"
+"            cont = 1;\n"
+"        }\n"
+"        if (!cont){\n"
+"            // test if origin is outside (v3, v0, v2) - set v1 as v3 and\n"
+"            // continue\n"
+"            b3MprVec3Cross(&va, &b3MprSimplexPoint(portal, 3)->v,\n"
+"                              &b3MprSimplexPoint(portal, 2)->v);\n"
+"            dot = b3MprVec3Dot(&va, &b3MprSimplexPoint(portal, 0)->v);\n"
+"            if (dot < 0.f && !b3MprIsZero(dot)){\n"
+"                b3MprSimplexSet(portal, 1, b3MprSimplexPoint(portal, 3));\n"
+"                cont = 1;\n"
+"            }\n"
+"        }\n"
+"        if (cont){\n"
+"            b3MprVec3Sub2(&va, &b3MprSimplexPoint(portal, 1)->v,\n"
+"                             &b3MprSimplexPoint(portal, 0)->v);\n"
+"            b3MprVec3Sub2(&vb, &b3MprSimplexPoint(portal, 2)->v,\n"
+"                             &b3MprSimplexPoint(portal, 0)->v);\n"
+"            b3MprVec3Cross(&dir, &va, &vb);\n"
+"            b3MprVec3Normalize(&dir);\n"
+"        }else{\n"
+"            b3MprSimplexSetSize(portal, 4);\n"
+"        }\n"
+"    }\n"
+"    return 0;\n"
+"}\n"
+"B3_STATIC int b3RefinePortal(int pairIndex,int bodyIndexA, int bodyIndexB,  b3ConstArray(b3RigidBodyData_t) cpuBodyBuf, \n"
+"													b3ConstArray(b3ConvexPolyhedronData_t) cpuConvexData, \n"
+"													b3ConstArray(b3Collidable_t)				cpuCollidables,\n"
+"													b3ConstArray(b3Float4)					cpuVertices,\n"
+"													__global b3Float4* sepAxis,\n"
+"													b3MprSimplex_t *portal)\n"
+"{\n"
+"    b3Float4 dir;\n"
+"    b3MprSupport_t v4;\n"
+"	for (int i=0;i<B3_MPR_MAX_ITERATIONS;i++)\n"
+"    //while (1)\n"
+"	{\n"
+"        // compute direction outside the portal (from v0 throught v1,v2,v3\n"
+"        // face)\n"
+"        b3PortalDir(portal, &dir);\n"
+"        // test if origin is inside the portal\n"
+"        if (portalEncapsulesOrigin(portal, &dir))\n"
+"            return 0;\n"
+"        // get next support point\n"
+"        \n"
+"		 b3MprSupport(pairIndex,bodyIndexA,bodyIndexB,cpuBodyBuf,cpuConvexData,cpuCollidables,cpuVertices, sepAxis,&dir, &v4);\n"
+"        // test if v4 can expand portal to contain origin and if portal\n"
+"        // expanding doesn't reach given tolerance\n"
+"        if (!portalCanEncapsuleOrigin(portal, &v4, &dir)\n"
+"                || portalReachTolerance(portal, &v4, &dir))\n"
+"		{\n"
+"            return -1;\n"
+"        }\n"
+"        // v1-v2-v3 triangle must be rearranged to face outside Minkowski\n"
+"        // difference (direction from v0).\n"
+"        b3ExpandPortal(portal, &v4);\n"
+"    }\n"
+"    return -1;\n"
+"}\n"
+"B3_STATIC void b3FindPos(const b3MprSimplex_t *portal, b3Float4 *pos)\n"
+"{\n"
+"	b3Float4 zero = b3MakeFloat4(0,0,0,0);\n"
+"	b3Float4* b3mpr_vec3_origin = &zero;\n"
+"    b3Float4 dir;\n"
+"    size_t i;\n"
+"    float b[4], sum, inv;\n"
+"    b3Float4 vec, p1, p2;\n"
+"    b3PortalDir(portal, &dir);\n"
+"    // use barycentric coordinates of tetrahedron to find origin\n"
+"    b3MprVec3Cross(&vec, &b3MprSimplexPoint(portal, 1)->v,\n"
+"                       &b3MprSimplexPoint(portal, 2)->v);\n"
+"    b[0] = b3MprVec3Dot(&vec, &b3MprSimplexPoint(portal, 3)->v);\n"
+"    b3MprVec3Cross(&vec, &b3MprSimplexPoint(portal, 3)->v,\n"
+"                       &b3MprSimplexPoint(portal, 2)->v);\n"
+"    b[1] = b3MprVec3Dot(&vec, &b3MprSimplexPoint(portal, 0)->v);\n"
+"    b3MprVec3Cross(&vec, &b3MprSimplexPoint(portal, 0)->v,\n"
+"                       &b3MprSimplexPoint(portal, 1)->v);\n"
+"    b[2] = b3MprVec3Dot(&vec, &b3MprSimplexPoint(portal, 3)->v);\n"
+"    b3MprVec3Cross(&vec, &b3MprSimplexPoint(portal, 2)->v,\n"
+"                       &b3MprSimplexPoint(portal, 1)->v);\n"
+"    b[3] = b3MprVec3Dot(&vec, &b3MprSimplexPoint(portal, 0)->v);\n"
+"	sum = b[0] + b[1] + b[2] + b[3];\n"
+"    if (b3MprIsZero(sum) || sum < 0.f){\n"
+"		b[0] = 0.f;\n"
+"        b3MprVec3Cross(&vec, &b3MprSimplexPoint(portal, 2)->v,\n"
+"                           &b3MprSimplexPoint(portal, 3)->v);\n"
+"        b[1] = b3MprVec3Dot(&vec, &dir);\n"
+"        b3MprVec3Cross(&vec, &b3MprSimplexPoint(portal, 3)->v,\n"
+"                           &b3MprSimplexPoint(portal, 1)->v);\n"
+"        b[2] = b3MprVec3Dot(&vec, &dir);\n"
+"        b3MprVec3Cross(&vec, &b3MprSimplexPoint(portal, 1)->v,\n"
+"                           &b3MprSimplexPoint(portal, 2)->v);\n"
+"        b[3] = b3MprVec3Dot(&vec, &dir);\n"
+"		sum = b[1] + b[2] + b[3];\n"
+"	}\n"
+"	inv = 1.f / sum;\n"
+"    b3MprVec3Copy(&p1, b3mpr_vec3_origin);\n"
+"    b3MprVec3Copy(&p2, b3mpr_vec3_origin);\n"
+"    for (i = 0; i < 4; i++){\n"
+"        b3MprVec3Copy(&vec, &b3MprSimplexPoint(portal, i)->v1);\n"
+"        b3MprVec3Scale(&vec, b[i]);\n"
+"        b3MprVec3Add(&p1, &vec);\n"
+"        b3MprVec3Copy(&vec, &b3MprSimplexPoint(portal, i)->v2);\n"
+"        b3MprVec3Scale(&vec, b[i]);\n"
+"        b3MprVec3Add(&p2, &vec);\n"
+"    }\n"
+"    b3MprVec3Scale(&p1, inv);\n"
+"    b3MprVec3Scale(&p2, inv);\n"
+"    b3MprVec3Copy(pos, &p1);\n"
+"    b3MprVec3Add(pos, &p2);\n"
+"    b3MprVec3Scale(pos, 0.5);\n"
+"}\n"
+"inline float b3MprVec3Dist2(const b3Float4 *a, const b3Float4 *b)\n"
+"{\n"
+"    b3Float4 ab;\n"
+"    b3MprVec3Sub2(&ab, a, b);\n"
+"    return b3MprVec3Len2(&ab);\n"
+"}\n"
+"inline float _b3MprVec3PointSegmentDist2(const b3Float4 *P,\n"
+"                                                  const b3Float4 *x0,\n"
+"                                                  const b3Float4 *b,\n"
+"                                                  b3Float4 *witness)\n"
+"{\n"
+"    // The computation comes from solving equation of segment:\n"
+"    //      S(t) = x0 + t.d\n"
+"    //          where - x0 is initial point of segment\n"
+"    //                - d is direction of segment from x0 (|d| > 0)\n"
+"    //                - t belongs to <0, 1> interval\n"
+"    // \n"
+"    // Than, distance from a segment to some point P can be expressed:\n"
+"    //      D(t) = |x0 + t.d - P|^2\n"
+"    //          which is distance from any point on segment. Minimization\n"
+"    //          of this function brings distance from P to segment.\n"
+"    // Minimization of D(t) leads to simple quadratic equation that's\n"
+"    // solving is straightforward.\n"
+"    //\n"
+"    // Bonus of this method is witness point for free.\n"
+"    float dist, t;\n"
+"    b3Float4 d, a;\n"
+"    // direction of segment\n"
+"    b3MprVec3Sub2(&d, b, x0);\n"
+"    // precompute vector from P to x0\n"
+"    b3MprVec3Sub2(&a, x0, P);\n"
+"    t  = -1.f * b3MprVec3Dot(&a, &d);\n"
+"    t /= b3MprVec3Len2(&d);\n"
+"    if (t < 0.f || b3MprIsZero(t)){\n"
+"        dist = b3MprVec3Dist2(x0, P);\n"
+"        if (witness)\n"
+"            b3MprVec3Copy(witness, x0);\n"
+"    }else if (t > 1.f || b3MprEq(t, 1.f)){\n"
+"        dist = b3MprVec3Dist2(b, P);\n"
+"        if (witness)\n"
+"            b3MprVec3Copy(witness, b);\n"
+"    }else{\n"
+"        if (witness){\n"
+"            b3MprVec3Copy(witness, &d);\n"
+"            b3MprVec3Scale(witness, t);\n"
+"            b3MprVec3Add(witness, x0);\n"
+"            dist = b3MprVec3Dist2(witness, P);\n"
+"        }else{\n"
+"            // recycling variables\n"
+"            b3MprVec3Scale(&d, t);\n"
+"            b3MprVec3Add(&d, &a);\n"
+"            dist = b3MprVec3Len2(&d);\n"
+"        }\n"
+"    }\n"
+"    return dist;\n"
+"}\n"
+"inline float b3MprVec3PointTriDist2(const b3Float4 *P,\n"
+"                                const b3Float4 *x0, const b3Float4 *B,\n"
+"                                const b3Float4 *C,\n"
+"                                b3Float4 *witness)\n"
+"{\n"
+"    // Computation comes from analytic expression for triangle (x0, B, C)\n"
+"    //      T(s, t) = x0 + s.d1 + t.d2, where d1 = B - x0 and d2 = C - x0 and\n"
+"    // Then equation for distance is:\n"
+"    //      D(s, t) = | T(s, t) - P |^2\n"
+"    // This leads to minimization of quadratic function of two variables.\n"
+"    // The solution from is taken only if s is between 0 and 1, t is\n"
+"    // between 0 and 1 and t + s < 1, otherwise distance from segment is\n"
+"    // computed.\n"
+"    b3Float4 d1, d2, a;\n"
+"    float u, v, w, p, q, r;\n"
+"    float s, t, dist, dist2;\n"
+"    b3Float4 witness2;\n"
+"    b3MprVec3Sub2(&d1, B, x0);\n"
+"    b3MprVec3Sub2(&d2, C, x0);\n"
+"    b3MprVec3Sub2(&a, x0, P);\n"
+"    u = b3MprVec3Dot(&a, &a);\n"
+"    v = b3MprVec3Dot(&d1, &d1);\n"
+"    w = b3MprVec3Dot(&d2, &d2);\n"
+"    p = b3MprVec3Dot(&a, &d1);\n"
+"    q = b3MprVec3Dot(&a, &d2);\n"
+"    r = b3MprVec3Dot(&d1, &d2);\n"
+"    s = (q * r - w * p) / (w * v - r * r);\n"
+"    t = (-s * r - q) / w;\n"
+"    if ((b3MprIsZero(s) || s > 0.f)\n"
+"            && (b3MprEq(s, 1.f) || s < 1.f)\n"
+"            && (b3MprIsZero(t) || t > 0.f)\n"
+"            && (b3MprEq(t, 1.f) || t < 1.f)\n"
+"            && (b3MprEq(t + s, 1.f) || t + s < 1.f)){\n"
+"        if (witness){\n"
+"            b3MprVec3Scale(&d1, s);\n"
+"            b3MprVec3Scale(&d2, t);\n"
+"            b3MprVec3Copy(witness, x0);\n"
+"            b3MprVec3Add(witness, &d1);\n"
+"            b3MprVec3Add(witness, &d2);\n"
+"            dist = b3MprVec3Dist2(witness, P);\n"
+"        }else{\n"
+"            dist  = s * s * v;\n"
+"            dist += t * t * w;\n"
+"            dist += 2.f * s * t * r;\n"
+"            dist += 2.f * s * p;\n"
+"            dist += 2.f * t * q;\n"
+"            dist += u;\n"
+"        }\n"
+"    }else{\n"
+"        dist = _b3MprVec3PointSegmentDist2(P, x0, B, witness);\n"
+"        dist2 = _b3MprVec3PointSegmentDist2(P, x0, C, &witness2);\n"
+"        if (dist2 < dist){\n"
+"            dist = dist2;\n"
+"            if (witness)\n"
+"                b3MprVec3Copy(witness, &witness2);\n"
+"        }\n"
+"        dist2 = _b3MprVec3PointSegmentDist2(P, B, C, &witness2);\n"
+"        if (dist2 < dist){\n"
+"            dist = dist2;\n"
+"            if (witness)\n"
+"                b3MprVec3Copy(witness, &witness2);\n"
+"        }\n"
+"    }\n"
+"    return dist;\n"
+"}\n"
+"B3_STATIC void b3FindPenetr(int pairIndex,int bodyIndexA, int bodyIndexB,  b3ConstArray(b3RigidBodyData_t) cpuBodyBuf, \n"
+"													b3ConstArray(b3ConvexPolyhedronData_t) cpuConvexData, \n"
+"													b3ConstArray(b3Collidable_t)				cpuCollidables,\n"
+"													b3ConstArray(b3Float4)					cpuVertices,\n"
+"													__global b3Float4* sepAxis,\n"
+"                       b3MprSimplex_t *portal,\n"
+"                       float *depth, b3Float4 *pdir, b3Float4 *pos)\n"
+"{\n"
+"    b3Float4 dir;\n"
+"    b3MprSupport_t v4;\n"
+"    unsigned long iterations;\n"
+"	b3Float4 zero = b3MakeFloat4(0,0,0,0);\n"
+"	b3Float4* b3mpr_vec3_origin = &zero;\n"
+"    iterations = 1UL;\n"
+"	for (int i=0;i<B3_MPR_MAX_ITERATIONS;i++)\n"
+"    //while (1)\n"
+"	{\n"
+"        // compute portal direction and obtain next support point\n"
+"        b3PortalDir(portal, &dir);\n"
+"        \n"
+"		 b3MprSupport(pairIndex,bodyIndexA,bodyIndexB,cpuBodyBuf,cpuConvexData,cpuCollidables,cpuVertices, sepAxis,&dir, &v4);\n"
+"        // reached tolerance -> find penetration info\n"
+"        if (portalReachTolerance(portal, &v4, &dir)\n"
+"                || iterations ==B3_MPR_MAX_ITERATIONS)\n"
+"		{\n"
+"            *depth = b3MprVec3PointTriDist2(b3mpr_vec3_origin,&b3MprSimplexPoint(portal, 1)->v,&b3MprSimplexPoint(portal, 2)->v,&b3MprSimplexPoint(portal, 3)->v,pdir);\n"
+"            *depth = B3_MPR_SQRT(*depth);\n"
+"			\n"
+"			if (b3MprIsZero((*pdir).x) && b3MprIsZero((*pdir).y) && b3MprIsZero((*pdir).z))\n"
+"			{\n"
+"				\n"
+"				*pdir = dir;\n"
+"			} \n"
+"			b3MprVec3Normalize(pdir);\n"
+"			\n"
+"            // barycentric coordinates:\n"
+"            b3FindPos(portal, pos);\n"
+"            return;\n"
+"        }\n"
+"        b3ExpandPortal(portal, &v4);\n"
+"        iterations++;\n"
+"    }\n"
+"}\n"
+"B3_STATIC void b3FindPenetrTouch(b3MprSimplex_t *portal,float *depth, b3Float4 *dir, b3Float4 *pos)\n"
+"{\n"
+"    // Touching contact on portal's v1 - so depth is zero and direction\n"
+"    // is unimportant and pos can be guessed\n"
+"    *depth = 0.f;\n"
+"    b3Float4 zero = b3MakeFloat4(0,0,0,0);\n"
+"	b3Float4* b3mpr_vec3_origin = &zero;\n"
+"	b3MprVec3Copy(dir, b3mpr_vec3_origin);\n"
+"    b3MprVec3Copy(pos, &b3MprSimplexPoint(portal, 1)->v1);\n"
+"    b3MprVec3Add(pos, &b3MprSimplexPoint(portal, 1)->v2);\n"
+"    b3MprVec3Scale(pos, 0.5);\n"
+"}\n"
+"B3_STATIC void b3FindPenetrSegment(b3MprSimplex_t *portal,\n"
+"                              float *depth, b3Float4 *dir, b3Float4 *pos)\n"
+"{\n"
+"    \n"
+"    // Origin lies on v0-v1 segment.\n"
+"    // Depth is distance to v1, direction also and position must be\n"
+"    // computed\n"
+"    b3MprVec3Copy(pos, &b3MprSimplexPoint(portal, 1)->v1);\n"
+"    b3MprVec3Add(pos, &b3MprSimplexPoint(portal, 1)->v2);\n"
+"    b3MprVec3Scale(pos, 0.5f);\n"
+"    \n"
+"    b3MprVec3Copy(dir, &b3MprSimplexPoint(portal, 1)->v);\n"
+"    *depth = B3_MPR_SQRT(b3MprVec3Len2(dir));\n"
+"    b3MprVec3Normalize(dir);\n"
+"}\n"
+"inline int b3MprPenetration(int pairIndex, int bodyIndexA, int bodyIndexB,\n"
+"					b3ConstArray(b3RigidBodyData_t) cpuBodyBuf,\n"
+"					b3ConstArray(b3ConvexPolyhedronData_t) cpuConvexData, \n"
+"					b3ConstArray(b3Collidable_t)	cpuCollidables,\n"
+"					b3ConstArray(b3Float4)	cpuVertices,\n"
+"					__global b3Float4* sepAxis,\n"
+"					__global int*	hasSepAxis,\n"
+"					float *depthOut, b3Float4* dirOut, b3Float4* posOut)\n"
+"{\n"
+"	\n"
+"	 b3MprSimplex_t portal;\n"
+"	 \n"
+"//	if (!hasSepAxis[pairIndex])\n"
+"	//	return -1;\n"
+"	\n"
+"	hasSepAxis[pairIndex] = 0;\n"
+"	 int res;\n"
+"    // Phase 1: Portal discovery\n"
+"    res = b3DiscoverPortal(pairIndex,bodyIndexA,bodyIndexB,cpuBodyBuf,cpuConvexData,cpuCollidables,cpuVertices,sepAxis,hasSepAxis, &portal);\n"
+"	\n"
+"	  \n"
+"	//sepAxis[pairIndex] = *pdir;//or -dir?\n"
+"	switch (res)\n"
+"	{\n"
+"	case 0:\n"
+"		{\n"
+"			// Phase 2: Portal refinement\n"
+"		\n"
+"			res = b3RefinePortal(pairIndex,bodyIndexA,bodyIndexB,cpuBodyBuf,cpuConvexData,cpuCollidables,cpuVertices, sepAxis,&portal);\n"
+"			if (res < 0)\n"
+"				return -1;\n"
+"			// Phase 3. Penetration info\n"
+"			b3FindPenetr(pairIndex,bodyIndexA,bodyIndexB,cpuBodyBuf,cpuConvexData,cpuCollidables,cpuVertices, sepAxis,&portal, depthOut, dirOut, posOut);\n"
+"			hasSepAxis[pairIndex] = 1;\n"
+"			sepAxis[pairIndex] = -*dirOut;\n"
+"			break;\n"
+"		}\n"
+"	case 1:\n"
+"		{\n"
+"			 // Touching contact on portal's v1.\n"
+"			b3FindPenetrTouch(&portal, depthOut, dirOut, posOut);\n"
+"			break;\n"
+"		}\n"
+"	case 2:\n"
+"		{\n"
+"			\n"
+"			b3FindPenetrSegment( &portal, depthOut, dirOut, posOut);\n"
+"			break;\n"
+"		}\n"
+"	default:\n"
+"		{\n"
+"			hasSepAxis[pairIndex]=0;\n"
+"			//if (res < 0)\n"
+"			//{\n"
+"				// Origin isn't inside portal - no collision.\n"
+"				return -1;\n"
+"			//}\n"
+"		}\n"
+"	};\n"
+"	\n"
+"	return 0;\n"
+"};\n"
+"#endif //B3_MPR_PENETRATION_H\n"
+"#ifndef B3_CONTACT4DATA_H\n"
+"#define B3_CONTACT4DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"typedef  struct b3Contact4Data b3Contact4Data_t;\n"
+"struct b3Contact4Data\n"
+"{\n"
+"	b3Float4	m_worldPosB[4];\n"
+"//	b3Float4	m_localPosA[4];\n"
+"//	b3Float4	m_localPosB[4];\n"
+"	b3Float4	m_worldNormalOnB;	//	w: m_nPoints\n"
+"	unsigned short  m_restituitionCoeffCmp;\n"
+"	unsigned short  m_frictionCoeffCmp;\n"
+"	int m_batchIdx;\n"
+"	int m_bodyAPtrAndSignBit;//x:m_bodyAPtr, y:m_bodyBPtr\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"	int	m_childIndexA;\n"
+"	int	m_childIndexB;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"};\n"
+"inline int b3Contact4Data_getNumPoints(const struct b3Contact4Data* contact)\n"
+"{\n"
+"	return (int)contact->m_worldNormalOnB.w;\n"
+"};\n"
+"inline void b3Contact4Data_setNumPoints(struct b3Contact4Data* contact, int numPoints)\n"
+"{\n"
+"	contact->m_worldNormalOnB.w = (float)numPoints;\n"
+"};\n"
+"#endif //B3_CONTACT4DATA_H\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define GET_NPOINTS(x) (x).m_worldNormalOnB.w\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"	#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"	#define counter32_t volatile __global int*\n"
+"#endif\n"
+"__kernel void   mprPenetrationKernel( __global int4* pairs,\n"
+"																					__global const b3RigidBodyData_t* rigidBodies, \n"
+"																					__global const b3Collidable_t* collidables,\n"
+"																					__global const b3ConvexPolyhedronData_t* convexShapes, \n"
+"																					__global const float4* vertices,\n"
+"																					__global float4* separatingNormals,\n"
+"																					__global int* hasSeparatingAxis,\n"
+"																					__global struct b3Contact4Data* restrict globalContactsOut,\n"
+"																					counter32_t nGlobalContactsOut,\n"
+"																					int contactCapacity,\n"
+"																					int numPairs)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	int pairIndex = i;\n"
+"	if (i<numPairs)\n"
+"	{\n"
+"		int bodyIndexA = pairs[i].x;\n"
+"		int bodyIndexB = pairs[i].y;\n"
+"		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"	\n"
+"		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"		\n"
+"		\n"
+"		//once the broadphase avoids static-static pairs, we can remove this test\n"
+"		if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))\n"
+"		{\n"
+"			return;\n"
+"		}\n"
+"		\n"
+"		if ((collidables[collidableIndexA].m_shapeType!=SHAPE_CONVEX_HULL) ||(collidables[collidableIndexB].m_shapeType!=SHAPE_CONVEX_HULL))\n"
+"		{\n"
+"			return;\n"
+"		}\n"
+"		float depthOut;\n"
+"		b3Float4 dirOut;\n"
+"		b3Float4 posOut;\n"
+"		int res = b3MprPenetration(pairIndex, bodyIndexA, bodyIndexB,rigidBodies,convexShapes,collidables,vertices,separatingNormals,hasSeparatingAxis,&depthOut, &dirOut, &posOut);\n"
+"		\n"
+"		\n"
+"		\n"
+"		\n"
+"		if (res==0)\n"
+"		{\n"
+"			//add a contact\n"
+"			int dstIdx;\n"
+"			AppendInc( nGlobalContactsOut, dstIdx );\n"
+"			if (dstIdx<contactCapacity)\n"
+"			{\n"
+"				pairs[pairIndex].z = dstIdx;\n"
+"				__global struct b3Contact4Data* c = globalContactsOut + dstIdx;\n"
+"				c->m_worldNormalOnB = -dirOut;//normal;\n"
+"				c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);\n"
+"				c->m_batchIdx = pairIndex;\n"
+"				int bodyA = pairs[pairIndex].x;\n"
+"				int bodyB = pairs[pairIndex].y;\n"
+"				c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0 ? -bodyA:bodyA;\n"
+"				c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0 ? -bodyB:bodyB;\n"
+"				c->m_childIndexA = -1;\n"
+"				c->m_childIndexB = -1;\n"
+"				//for (int i=0;i<nContacts;i++)\n"
+"				posOut.w = -depthOut;\n"
+"				c->m_worldPosB[0] = posOut;//localPoints[contactIdx[i]];\n"
+"				GET_NPOINTS(*c) = 1;//nContacts;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"}\n"
+"typedef float4 Quaternion;\n"
+"#define make_float4 (float4)\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = make_float4(a.xyz,0.f);\n"
+"	float4 b1 = make_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(qtMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"float4 transform(const float4* p, const float4* translation, const Quaternion* orientation)\n"
+"{\n"
+"	return qtRotate( *orientation, *p ) + (*translation);\n"
+"}\n"
+"__inline\n"
+"float4 qtInvRotate(const Quaternion q, float4 vec)\n"
+"{\n"
+"	return qtRotate( qtInvert( q ), vec );\n"
+"}\n"
+"inline void project(__global const b3ConvexPolyhedronData_t* hull,  const float4 pos, const float4 orn, \n"
+"const float4* dir, __global const float4* vertices, float* min, float* max)\n"
+"{\n"
+"	min[0] = FLT_MAX;\n"
+"	max[0] = -FLT_MAX;\n"
+"	int numVerts = hull->m_numVertices;\n"
+"	const float4 localDir = qtInvRotate(orn,*dir);\n"
+"	float offset = dot(pos,*dir);\n"
+"	for(int i=0;i<numVerts;i++)\n"
+"	{\n"
+"		float dp = dot(vertices[hull->m_vertexOffset+i],localDir);\n"
+"		if(dp < min[0])	\n"
+"			min[0] = dp;\n"
+"		if(dp > max[0])	\n"
+"			max[0] = dp;\n"
+"	}\n"
+"	if(min[0]>max[0])\n"
+"	{\n"
+"		float tmp = min[0];\n"
+"		min[0] = max[0];\n"
+"		max[0] = tmp;\n"
+"	}\n"
+"	min[0] += offset;\n"
+"	max[0] += offset;\n"
+"}\n"
+"bool findSeparatingAxisUnitSphere(	__global const b3ConvexPolyhedronData_t* hullA, __global const b3ConvexPolyhedronData_t* hullB, \n"
+"	const float4 posA1,\n"
+"	const float4 ornA,\n"
+"	const float4 posB1,\n"
+"	const float4 ornB,\n"
+"	const float4 DeltaC2,\n"
+"	__global const float4* vertices,\n"
+"	__global const float4* unitSphereDirections,\n"
+"	int numUnitSphereDirections,\n"
+"	float4* sep,\n"
+"	float* dmin)\n"
+"{\n"
+"	\n"
+"	float4 posA = posA1;\n"
+"	posA.w = 0.f;\n"
+"	float4 posB = posB1;\n"
+"	posB.w = 0.f;\n"
+"	int curPlaneTests=0;\n"
+"	int curEdgeEdge = 0;\n"
+"	// Test unit sphere directions\n"
+"	for (int i=0;i<numUnitSphereDirections;i++)\n"
+"	{\n"
+"		float4 crossje;\n"
+"		crossje = unitSphereDirections[i];	\n"
+"		if (dot3F4(DeltaC2,crossje)>0)\n"
+"			crossje *= -1.f;\n"
+"		{\n"
+"			float dist;\n"
+"			bool result = true;\n"
+"			float Min0,Max0;\n"
+"			float Min1,Max1;\n"
+"			project(hullA,posA,ornA,&crossje,vertices, &Min0, &Max0);\n"
+"			project(hullB,posB,ornB,&crossje,vertices, &Min1, &Max1);\n"
+"		\n"
+"			if(Max0<Min1 || Max1<Min0)\n"
+"				return false;\n"
+"		\n"
+"			float d0 = Max0 - Min1;\n"
+"			float d1 = Max1 - Min0;\n"
+"			dist = d0<d1 ? d0:d1;\n"
+"			result = true;\n"
+"	\n"
+"			if(dist<*dmin)\n"
+"			{\n"
+"				*dmin = dist;\n"
+"				*sep = crossje;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	\n"
+"	if((dot3F4(-DeltaC2,*sep))>0.0f)\n"
+"	{\n"
+"		*sep = -(*sep);\n"
+"	}\n"
+"	return true;\n"
+"}\n"
+"__kernel void   findSeparatingAxisUnitSphereKernel( __global const int4* pairs, \n"
+"																					__global const b3RigidBodyData_t* rigidBodies, \n"
+"																					__global const b3Collidable_t* collidables,\n"
+"																					__global const b3ConvexPolyhedronData_t* convexShapes, \n"
+"																					__global const float4* vertices,\n"
+"																					__global const float4* unitSphereDirections,\n"
+"																					__global  float4* separatingNormals,\n"
+"																					__global  int* hasSeparatingAxis,\n"
+"																					__global  float* dmins,\n"
+"																					int numUnitSphereDirections,\n"
+"																					int numPairs\n"
+"																					)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	\n"
+"	if (i<numPairs)\n"
+"	{\n"
+"		if (hasSeparatingAxis[i])\n"
+"		{\n"
+"	\n"
+"			int bodyIndexA = pairs[i].x;\n"
+"			int bodyIndexB = pairs[i].y;\n"
+"	\n"
+"			int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"			int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"		\n"
+"			int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"			int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"			\n"
+"			\n"
+"			int numFacesA = convexShapes[shapeIndexA].m_numFaces;\n"
+"	\n"
+"			float dmin = dmins[i];\n"
+"	\n"
+"			float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"			posA.w = 0.f;\n"
+"			float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"			posB.w = 0.f;\n"
+"			float4 c0local = convexShapes[shapeIndexA].m_localCenter;\n"
+"			float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"			float4 c0 = transform(&c0local, &posA, &ornA);\n"
+"			float4 c1local = convexShapes[shapeIndexB].m_localCenter;\n"
+"			float4 ornB =rigidBodies[bodyIndexB].m_quat;\n"
+"			float4 c1 = transform(&c1local,&posB,&ornB);\n"
+"			const float4 DeltaC2 = c0 - c1;\n"
+"			float4 sepNormal = separatingNormals[i];\n"
+"			\n"
+"			int numEdgeEdgeDirections = convexShapes[shapeIndexA].m_numUniqueEdges*convexShapes[shapeIndexB].m_numUniqueEdges;\n"
+"			if (numEdgeEdgeDirections>numUnitSphereDirections)\n"
+"			{\n"
+"				bool sepEE = findSeparatingAxisUnitSphere(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,\n"
+"																										posB,ornB,\n"
+"																										DeltaC2,\n"
+"																										vertices,unitSphereDirections,numUnitSphereDirections,&sepNormal,&dmin);\n"
+"				if (!sepEE)\n"
+"				{\n"
+"					hasSeparatingAxis[i] = 0;\n"
+"				} else\n"
+"				{\n"
+"					hasSeparatingAxis[i] = 1;\n"
+"					separatingNormals[i] = sepNormal;\n"
+"				}\n"
+"			}\n"
+"		}		//if (hasSeparatingAxis[i])\n"
+"	}//(i<numPairs)\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/primitiveContacts.cl b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/primitiveContacts.cl
new file mode 100644
index 0000000000..9c9e920f13
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/primitiveContacts.cl
@@ -0,0 +1,1374 @@
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Contact4Data.h"
+
+#define SHAPE_CONVEX_HULL 3
+#define SHAPE_PLANE 4
+#define SHAPE_CONCAVE_TRIMESH 5
+#define SHAPE_COMPOUND_OF_CONVEX_HULLS 6
+#define SHAPE_SPHERE 7
+
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile __global int*
+#endif
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+#define max2 max
+#define min2 min
+
+typedef unsigned int u32;
+
+
+
+
+typedef struct 
+{
+	union
+	{
+		float4	m_min;
+		float   m_minElems[4];
+		int			m_minIndices[4];
+	};
+	union
+	{
+		float4	m_max;
+		float   m_maxElems[4];
+		int			m_maxIndices[4];
+	};
+} btAabbCL;
+
+///keep this in sync with btCollidable.h
+typedef struct
+{
+	int m_numChildShapes;
+	float m_radius;
+	int m_shapeType;
+	int m_shapeIndex;
+	
+} btCollidableGpu;
+
+typedef struct
+{
+	float4	m_childPosition;
+	float4	m_childOrientation;
+	int m_shapeIndex;
+	int m_unused0;
+	int m_unused1;
+	int m_unused2;
+} btGpuChildShape;
+
+#define GET_NPOINTS(x) (x).m_worldNormalOnB.w
+
+typedef struct
+{
+	float4 m_pos;
+	float4 m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_collidableIdx;	
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} BodyData;
+
+
+typedef struct  
+{
+	float4		m_localCenter;
+	float4		m_extents;
+	float4		mC;
+	float4		mE;
+	
+	float			m_radius;
+	int	m_faceOffset;
+	int m_numFaces;
+	int	m_numVertices;
+	
+	int m_vertexOffset;
+	int	m_uniqueEdgesOffset;
+	int	m_numUniqueEdges;
+	int m_unused;
+
+} ConvexPolyhedronCL;
+
+typedef struct
+{
+	float4 m_plane;
+	int m_indexOffset;
+	int m_numIndices;
+} btGpuFace;
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define make_float4 (float4)
+#define make_float2 (float2)
+#define make_uint4 (uint4)
+#define make_int4 (int4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+
+__inline
+float fastDiv(float numerator, float denominator)
+{
+	return native_divide(numerator, denominator);	
+//	return numerator/denominator;	
+}
+
+__inline
+float4 fastDiv4(float4 numerator, float4 denominator)
+{
+	return native_divide(numerator, denominator);	
+}
+
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+
+//#define dot3F4 dot
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = make_float4(a.xyz,0.f);
+	float4 b1 = make_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	return fast_normalize(v);
+}
+
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b);
+
+__inline
+Quaternion qtNormalize(Quaternion in);
+
+__inline
+float4 qtRotate(Quaternion q, float4 vec);
+
+__inline
+Quaternion qtInvert(Quaternion q);
+
+
+
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross3( a, b );
+	ans += a.w*b+b.w*a;
+//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+Quaternion qtNormalize(Quaternion in)
+{
+	return fastNormalize4(in);
+//	in /= length( in );
+//	return in;
+}
+__inline
+float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(qtMul(q,vcpy),qInv);
+	return out;
+}
+
+__inline
+Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+__inline
+float4 qtInvRotate(const Quaternion q, float4 vec)
+{
+	return qtRotate( qtInvert( q ), vec );
+}
+
+__inline
+float4 transform(const float4* p, const float4* translation, const Quaternion* orientation)
+{
+	return qtRotate( *orientation, *p ) + (*translation);
+}
+
+void	trInverse(float4 translationIn, Quaternion orientationIn,
+		float4* translationOut, Quaternion* orientationOut)
+{
+	*orientationOut = qtInvert(orientationIn);
+	*translationOut = qtRotate(*orientationOut, -translationIn);
+}
+
+void	trMul(float4 translationA, Quaternion orientationA,
+						float4 translationB, Quaternion orientationB,
+		float4* translationOut, Quaternion* orientationOut)
+{
+	*orientationOut = qtMul(orientationA,orientationB);
+	*translationOut = transform(&translationB,&translationA,&orientationA);
+}
+
+
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = make_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+}
+
+
+__inline float4 lerp3(const float4 a,const float4 b, float  t)
+{
+	return make_float4(	a.x + (b.x - a.x) * t,
+						a.y + (b.y - a.y) * t,
+						a.z + (b.z - a.z) * t,
+						0.f);
+}
+
+
+float signedDistanceFromPointToPlane(float4 point, float4 planeEqn, float4* closestPointOnFace)
+{
+	float4 n = (float4)(planeEqn.x, planeEqn.y, planeEqn.z, 0);
+	float dist = dot3F4(n, point) + planeEqn.w;
+	*closestPointOnFace = point - dist * n;
+	return dist;
+}
+
+
+
+inline bool IsPointInPolygon(float4 p, 
+							const btGpuFace* face,
+							__global const float4* baseVertex,
+							__global const  int* convexIndices,
+							float4* out)
+{
+    float4 a;
+    float4 b;
+    float4 ab;
+    float4 ap;
+    float4 v;
+
+	float4 plane = make_float4(face->m_plane.x,face->m_plane.y,face->m_plane.z,0.f);
+	
+	if (face->m_numIndices<2)
+		return false;
+
+	
+	float4 v0 = baseVertex[convexIndices[face->m_indexOffset + face->m_numIndices-1]];
+	
+	b = v0;
+
+    for(unsigned i=0; i != face->m_numIndices; ++i)
+    {
+		a = b;
+		float4 vi = baseVertex[convexIndices[face->m_indexOffset + i]];
+		b = vi;
+        ab = b-a;
+        ap = p-a;
+        v = cross3(ab,plane);
+
+        if (dot(ap, v) > 0.f)
+        {
+            float ab_m2 = dot(ab, ab);
+            float rt = ab_m2 != 0.f ? dot(ab, ap) / ab_m2 : 0.f;
+            if (rt <= 0.f)
+            {
+                *out = a;
+            }
+            else if (rt >= 1.f) 
+            {
+                *out = b;
+            }
+            else
+            {
+            	float s = 1.f - rt;
+				out[0].x = s * a.x + rt * b.x;
+				out[0].y = s * a.y + rt * b.y;
+				out[0].z = s * a.z + rt * b.z;
+            }
+            return false;
+        }
+    }
+    return true;
+}
+
+
+
+
+void	computeContactSphereConvex(int pairIndex,
+																int bodyIndexA, int bodyIndexB, 
+																int collidableIndexA, int collidableIndexB, 
+																__global const BodyData* rigidBodies, 
+																__global const btCollidableGpu* collidables,
+																__global const ConvexPolyhedronCL* convexShapes,
+																__global const float4* convexVertices,
+																__global const int* convexIndices,
+																__global const btGpuFace* faces,
+																__global struct b3Contact4Data* restrict globalContactsOut,
+																counter32_t nGlobalContactsOut,
+																int maxContactCapacity,
+																float4 spherePos2,
+																float radius,
+																float4 pos,
+																float4 quat
+																)
+{
+
+	float4 invPos;
+	float4 invOrn;
+
+	trInverse(pos,quat, &invPos,&invOrn);
+
+	float4 spherePos = transform(&spherePos2,&invPos,&invOrn);
+
+	int shapeIndex = collidables[collidableIndexB].m_shapeIndex;
+	int numFaces = convexShapes[shapeIndex].m_numFaces;
+	float4 closestPnt = (float4)(0, 0, 0, 0);
+	float4 hitNormalWorld = (float4)(0, 0, 0, 0);
+	float minDist = -1000000.f;
+	bool bCollide = true;
+
+	for ( int f = 0; f < numFaces; f++ )
+	{
+		btGpuFace face = faces[convexShapes[shapeIndex].m_faceOffset+f];
+
+		// set up a plane equation 
+		float4 planeEqn;
+		float4 n1 = face.m_plane;
+		n1.w = 0.f;
+		planeEqn = n1;
+		planeEqn.w = face.m_plane.w;
+		
+	
+		// compute a signed distance from the vertex in cloth to the face of rigidbody.
+		float4 pntReturn;
+		float dist = signedDistanceFromPointToPlane(spherePos, planeEqn, &pntReturn);
+
+		// If the distance is positive, the plane is a separating plane. 
+		if ( dist > radius )
+		{
+			bCollide = false;
+			break;
+		}
+
+
+		if (dist>0)
+		{
+			//might hit an edge or vertex
+			float4 out;
+			float4 zeroPos = make_float4(0,0,0,0);
+
+			bool isInPoly = IsPointInPolygon(spherePos,
+					&face,
+					&convexVertices[convexShapes[shapeIndex].m_vertexOffset],
+					convexIndices,
+           &out);
+			if (isInPoly)
+			{
+				if (dist>minDist)
+				{
+					minDist = dist;
+					closestPnt = pntReturn;
+					hitNormalWorld = planeEqn;
+					
+				}
+			} else
+			{
+				float4 tmp = spherePos-out;
+				float l2 = dot(tmp,tmp);
+				if (l2<radius*radius)
+				{
+					dist  = sqrt(l2);
+					if (dist>minDist)
+					{
+						minDist = dist;
+						closestPnt = out;
+						hitNormalWorld = tmp/dist;
+						
+					}
+					
+				} else
+				{
+					bCollide = false;
+					break;
+				}
+			}
+		} else
+		{
+			if ( dist > minDist )
+			{
+				minDist = dist;
+				closestPnt = pntReturn;
+				hitNormalWorld.xyz = planeEqn.xyz;
+			}
+		}
+		
+	}
+
+	
+
+	if (bCollide && minDist > -10000)
+	{
+		float4 normalOnSurfaceB1 = qtRotate(quat,-hitNormalWorld);
+		float4 pOnB1 = transform(&closestPnt,&pos,&quat);
+		
+		float actualDepth = minDist-radius;
+		if (actualDepth<=0.f)
+		{
+			
+
+			pOnB1.w = actualDepth;
+
+			int dstIdx;
+			AppendInc( nGlobalContactsOut, dstIdx );
+		
+			
+			if (1)//dstIdx < maxContactCapacity)
+			{
+				__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];
+				c->m_worldNormalOnB = -normalOnSurfaceB1;
+				c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+				c->m_batchIdx = pairIndex;
+				c->m_bodyAPtrAndSignBit = rigidBodies[bodyIndexA].m_invMass==0?-bodyIndexA:bodyIndexA;
+				c->m_bodyBPtrAndSignBit = rigidBodies[bodyIndexB].m_invMass==0?-bodyIndexB:bodyIndexB;
+				c->m_worldPosB[0] = pOnB1;
+				c->m_childIndexA = -1;
+				c->m_childIndexB = -1;
+
+				GET_NPOINTS(*c) = 1;
+			} 
+
+		}
+	}//if (hasCollision)
+
+}
+							
+
+
+int extractManifoldSequential(const float4* p, int nPoints, float4 nearNormal, int4* contactIdx)
+{
+	if( nPoints == 0 )
+        return 0;
+    
+    if (nPoints <=4)
+        return nPoints;
+    
+    
+    if (nPoints >64)
+        nPoints = 64;
+    
+	float4 center = make_float4(0.f);
+	{
+		
+		for (int i=0;i<nPoints;i++)
+			center += p[i];
+		center /= (float)nPoints;
+	}
+    
+	
+    
+	//	sample 4 directions
+    
+    float4 aVector = p[0] - center;
+    float4 u = cross3( nearNormal, aVector );
+    float4 v = cross3( nearNormal, u );
+    u = normalize3( u );
+    v = normalize3( v );
+    
+    
+    //keep point with deepest penetration
+    float minW= FLT_MAX;
+    
+    int minIndex=-1;
+    
+    float4 maxDots;
+    maxDots.x = FLT_MIN;
+    maxDots.y = FLT_MIN;
+    maxDots.z = FLT_MIN;
+    maxDots.w = FLT_MIN;
+    
+    //	idx, distance
+    for(int ie = 0; ie<nPoints; ie++ )
+    {
+        if (p[ie].w<minW)
+        {
+            minW = p[ie].w;
+            minIndex=ie;
+        }
+        float f;
+        float4 r = p[ie]-center;
+        f = dot3F4( u, r );
+        if (f<maxDots.x)
+        {
+            maxDots.x = f;
+            contactIdx[0].x = ie;
+        }
+        
+        f = dot3F4( -u, r );
+        if (f<maxDots.y)
+        {
+            maxDots.y = f;
+            contactIdx[0].y = ie;
+        }
+        
+        
+        f = dot3F4( v, r );
+        if (f<maxDots.z)
+        {
+            maxDots.z = f;
+            contactIdx[0].z = ie;
+        }
+        
+        f = dot3F4( -v, r );
+        if (f<maxDots.w)
+        {
+            maxDots.w = f;
+            contactIdx[0].w = ie;
+        }
+        
+    }
+    
+    if (contactIdx[0].x != minIndex && contactIdx[0].y != minIndex && contactIdx[0].z != minIndex && contactIdx[0].w != minIndex)
+    {
+        //replace the first contact with minimum (todo: replace contact with least penetration)
+        contactIdx[0].x = minIndex;
+    }
+    
+    return 4;
+    
+}
+
+#define MAX_PLANE_CONVEX_POINTS 64
+
+int computeContactPlaneConvex(int pairIndex,
+								int bodyIndexA, int bodyIndexB, 
+								int collidableIndexA, int collidableIndexB, 
+								__global const BodyData* rigidBodies, 
+								__global const btCollidableGpu*collidables,
+								__global const ConvexPolyhedronCL* convexShapes,
+								__global const float4* convexVertices,
+								__global const int* convexIndices,
+								__global const btGpuFace* faces,
+								__global struct b3Contact4Data* restrict globalContactsOut,
+								counter32_t nGlobalContactsOut,
+								int maxContactCapacity,
+								float4 posB,
+								Quaternion ornB
+								)
+{
+	int resultIndex=-1;
+
+		int shapeIndex = collidables[collidableIndexB].m_shapeIndex;
+	__global const ConvexPolyhedronCL* hullB = &convexShapes[shapeIndex];
+	
+	float4 posA;
+	posA = rigidBodies[bodyIndexA].m_pos;
+	Quaternion ornA;
+	ornA = rigidBodies[bodyIndexA].m_quat;
+
+	int numContactsOut = 0;
+	int numWorldVertsB1= 0;
+
+	float4 planeEq;
+	 planeEq = faces[collidables[collidableIndexA].m_shapeIndex].m_plane;
+	float4 planeNormal = make_float4(planeEq.x,planeEq.y,planeEq.z,0.f);
+	float4 planeNormalWorld;
+	planeNormalWorld = qtRotate(ornA,planeNormal);
+	float planeConstant = planeEq.w;
+	
+	float4 invPosA;Quaternion invOrnA;
+	float4 convexInPlaneTransPos1; Quaternion convexInPlaneTransOrn1;
+	{
+		
+		trInverse(posA,ornA,&invPosA,&invOrnA);
+		trMul(invPosA,invOrnA,posB,ornB,&convexInPlaneTransPos1,&convexInPlaneTransOrn1);
+	}
+	float4 invPosB;Quaternion invOrnB;
+	float4 planeInConvexPos1;	Quaternion planeInConvexOrn1;
+	{
+		
+		trInverse(posB,ornB,&invPosB,&invOrnB);
+		trMul(invPosB,invOrnB,posA,ornA,&planeInConvexPos1,&planeInConvexOrn1);	
+	}
+
+	
+	float4 planeNormalInConvex = qtRotate(planeInConvexOrn1,-planeNormal);
+	float maxDot = -1e30;
+	int hitVertex=-1;
+	float4 hitVtx;
+
+
+
+	float4 contactPoints[MAX_PLANE_CONVEX_POINTS];
+	int numPoints = 0;
+
+	int4 contactIdx;
+	contactIdx=make_int4(0,1,2,3);
+    
+	
+	for (int i=0;i<hullB->m_numVertices;i++)
+	{
+		float4 vtx = convexVertices[hullB->m_vertexOffset+i];
+		float curDot = dot(vtx,planeNormalInConvex);
+
+
+		if (curDot>maxDot)
+		{
+			hitVertex=i;
+			maxDot=curDot;
+			hitVtx = vtx;
+			//make sure the deepest points is always included
+			if (numPoints==MAX_PLANE_CONVEX_POINTS)
+				numPoints--;
+		}
+
+		if (numPoints<MAX_PLANE_CONVEX_POINTS)
+		{
+			float4 vtxWorld = transform(&vtx, &posB, &ornB);
+			float4 vtxInPlane = transform(&vtxWorld, &invPosA, &invOrnA);//oplaneTransform.inverse()*vtxWorld;
+			float dist = dot(planeNormal,vtxInPlane)-planeConstant;
+			if (dist<0.f)
+			{
+				vtxWorld.w = dist;
+				contactPoints[numPoints] = vtxWorld;
+				numPoints++;
+			}
+		}
+
+	}
+
+	int numReducedPoints  = numPoints;
+	if (numPoints>4)
+	{
+		numReducedPoints = extractManifoldSequential( contactPoints, numPoints, planeNormalInConvex, &contactIdx);
+	}
+
+	if (numReducedPoints>0)
+	{
+		int dstIdx;
+	    AppendInc( nGlobalContactsOut, dstIdx );
+
+		if (dstIdx < maxContactCapacity)
+		{
+			resultIndex = dstIdx;
+			__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];
+			c->m_worldNormalOnB = -planeNormalWorld;
+			//c->setFrictionCoeff(0.7);
+			//c->setRestituitionCoeff(0.f);
+			c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+			c->m_batchIdx = pairIndex;
+			c->m_bodyAPtrAndSignBit = rigidBodies[bodyIndexA].m_invMass==0?-bodyIndexA:bodyIndexA;
+			c->m_bodyBPtrAndSignBit = rigidBodies[bodyIndexB].m_invMass==0?-bodyIndexB:bodyIndexB;
+			c->m_childIndexA = -1;
+			c->m_childIndexB = -1;
+
+			switch (numReducedPoints)
+            {
+                case 4:
+                    c->m_worldPosB[3] = contactPoints[contactIdx.w];
+                case 3:
+                    c->m_worldPosB[2] = contactPoints[contactIdx.z];
+                case 2:
+                    c->m_worldPosB[1] = contactPoints[contactIdx.y];
+                case 1:
+                    c->m_worldPosB[0] = contactPoints[contactIdx.x];
+                default:
+                {
+                }
+            };
+			
+			GET_NPOINTS(*c) = numReducedPoints;
+		}//if (dstIdx < numPairs)
+	}	
+
+	return resultIndex;
+}
+
+
+void	computeContactPlaneSphere(int pairIndex,
+																int bodyIndexA, int bodyIndexB, 
+																int collidableIndexA, int collidableIndexB, 
+																__global const BodyData* rigidBodies, 
+																__global const btCollidableGpu* collidables,
+																__global const btGpuFace* faces,
+																__global struct b3Contact4Data* restrict globalContactsOut,
+																counter32_t nGlobalContactsOut,
+																int maxContactCapacity)
+{
+	float4 planeEq = faces[collidables[collidableIndexA].m_shapeIndex].m_plane;
+	float radius = collidables[collidableIndexB].m_radius;
+	float4 posA1 = rigidBodies[bodyIndexA].m_pos;
+	float4 ornA1 = rigidBodies[bodyIndexA].m_quat;
+	float4 posB1 = rigidBodies[bodyIndexB].m_pos;
+	float4 ornB1 = rigidBodies[bodyIndexB].m_quat;
+	
+	bool hasCollision = false;
+	float4 planeNormal1 = make_float4(planeEq.x,planeEq.y,planeEq.z,0.f);
+	float planeConstant = planeEq.w;
+	float4 convexInPlaneTransPos1; Quaternion convexInPlaneTransOrn1;
+	{
+		float4 invPosA;Quaternion invOrnA;
+		trInverse(posA1,ornA1,&invPosA,&invOrnA);
+		trMul(invPosA,invOrnA,posB1,ornB1,&convexInPlaneTransPos1,&convexInPlaneTransOrn1);
+	}
+	float4 planeInConvexPos1;	Quaternion planeInConvexOrn1;
+	{
+		float4 invPosB;Quaternion invOrnB;
+		trInverse(posB1,ornB1,&invPosB,&invOrnB);
+		trMul(invPosB,invOrnB,posA1,ornA1,&planeInConvexPos1,&planeInConvexOrn1);	
+	}
+	float4 vtx1 = qtRotate(planeInConvexOrn1,-planeNormal1)*radius;
+	float4 vtxInPlane1 = transform(&vtx1,&convexInPlaneTransPos1,&convexInPlaneTransOrn1);
+	float distance = dot3F4(planeNormal1,vtxInPlane1) - planeConstant;
+	hasCollision = distance < 0.f;//m_manifoldPtr->getContactBreakingThreshold();
+	if (hasCollision)
+	{
+		float4 vtxInPlaneProjected1 = vtxInPlane1 -   distance*planeNormal1;
+		float4 vtxInPlaneWorld1 = transform(&vtxInPlaneProjected1,&posA1,&ornA1);
+		float4 normalOnSurfaceB1 = qtRotate(ornA1,planeNormal1);
+		float4 pOnB1 = vtxInPlaneWorld1+normalOnSurfaceB1*distance;
+		pOnB1.w = distance;
+
+		int dstIdx;
+    AppendInc( nGlobalContactsOut, dstIdx );
+		
+		if (dstIdx < maxContactCapacity)
+		{
+			__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];
+			c->m_worldNormalOnB = -normalOnSurfaceB1;
+			c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+			c->m_batchIdx = pairIndex;
+			c->m_bodyAPtrAndSignBit = rigidBodies[bodyIndexA].m_invMass==0?-bodyIndexA:bodyIndexA;
+			c->m_bodyBPtrAndSignBit = rigidBodies[bodyIndexB].m_invMass==0?-bodyIndexB:bodyIndexB;
+			c->m_worldPosB[0] = pOnB1;
+			c->m_childIndexA = -1;
+			c->m_childIndexB = -1;
+			GET_NPOINTS(*c) = 1;
+		}//if (dstIdx < numPairs)
+	}//if (hasCollision)
+}
+
+
+__kernel void   primitiveContactsKernel( __global int4* pairs, 
+																					__global const BodyData* rigidBodies, 
+																					__global const btCollidableGpu* collidables,
+																					__global const ConvexPolyhedronCL* convexShapes, 
+																					__global const float4* vertices,
+																					__global const float4* uniqueEdges,
+																					__global const btGpuFace* faces,
+																					__global const int* indices,
+																					__global struct b3Contact4Data* restrict globalContactsOut,
+																					counter32_t nGlobalContactsOut,
+																					int numPairs, int maxContactCapacity)
+{
+
+	int i = get_global_id(0);
+	int pairIndex = i;
+	
+	float4 worldVertsB1[64];
+	float4 worldVertsB2[64];
+	int capacityWorldVerts = 64;	
+
+	float4 localContactsOut[64];
+	int localContactCapacity=64;
+	
+	float minDist = -1e30f;
+	float maxDist = 0.02f;
+
+	if (i<numPairs)
+	{
+
+		int bodyIndexA = pairs[i].x;
+		int bodyIndexB = pairs[i].y;
+			
+		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+	
+		if (collidables[collidableIndexA].m_shapeType == SHAPE_PLANE &&
+			collidables[collidableIndexB].m_shapeType == SHAPE_CONVEX_HULL)
+		{
+
+			float4 posB;
+			posB = rigidBodies[bodyIndexB].m_pos;
+			Quaternion ornB;
+			ornB = rigidBodies[bodyIndexB].m_quat;
+			int contactIndex = computeContactPlaneConvex(pairIndex, bodyIndexA, bodyIndexB, collidableIndexA, collidableIndexB, 
+																rigidBodies,collidables,convexShapes,vertices,indices,
+																faces,	globalContactsOut, nGlobalContactsOut,maxContactCapacity, posB,ornB);
+			if (contactIndex>=0)
+				pairs[pairIndex].z = contactIndex;
+
+			return;
+		}
+
+
+		if (collidables[collidableIndexA].m_shapeType == SHAPE_CONVEX_HULL &&
+			collidables[collidableIndexB].m_shapeType == SHAPE_PLANE)
+		{
+
+			float4 posA;
+			posA = rigidBodies[bodyIndexA].m_pos;
+			Quaternion ornA;
+			ornA = rigidBodies[bodyIndexA].m_quat;
+
+
+			int contactIndex = computeContactPlaneConvex( pairIndex, bodyIndexB,bodyIndexA,  collidableIndexB,collidableIndexA, 
+																rigidBodies,collidables,convexShapes,vertices,indices,
+																faces,	globalContactsOut, nGlobalContactsOut,maxContactCapacity,posA,ornA);
+
+			if (contactIndex>=0)
+				pairs[pairIndex].z = contactIndex;
+
+			return;
+		}
+
+		if (collidables[collidableIndexA].m_shapeType == SHAPE_PLANE &&
+			collidables[collidableIndexB].m_shapeType == SHAPE_SPHERE)
+		{
+			computeContactPlaneSphere(pairIndex, bodyIndexA, bodyIndexB, collidableIndexA, collidableIndexB, 
+																rigidBodies,collidables,faces,	globalContactsOut, nGlobalContactsOut,maxContactCapacity);
+			return;
+		}
+
+
+		if (collidables[collidableIndexA].m_shapeType == SHAPE_SPHERE &&
+			collidables[collidableIndexB].m_shapeType == SHAPE_PLANE)
+		{
+
+
+			computeContactPlaneSphere( pairIndex, bodyIndexB,bodyIndexA,  collidableIndexB,collidableIndexA, 
+																rigidBodies,collidables,
+																faces,	globalContactsOut, nGlobalContactsOut,maxContactCapacity);
+
+			return;
+		}
+
+		
+
+	
+		if (collidables[collidableIndexA].m_shapeType == SHAPE_SPHERE &&
+			collidables[collidableIndexB].m_shapeType == SHAPE_CONVEX_HULL)
+		{
+		
+			float4 spherePos = rigidBodies[bodyIndexA].m_pos;
+			float sphereRadius = collidables[collidableIndexA].m_radius;
+			float4 convexPos = rigidBodies[bodyIndexB].m_pos;
+			float4 convexOrn = rigidBodies[bodyIndexB].m_quat;
+
+			computeContactSphereConvex(pairIndex, bodyIndexA, bodyIndexB, collidableIndexA, collidableIndexB, 
+																rigidBodies,collidables,convexShapes,vertices,indices,faces, globalContactsOut, nGlobalContactsOut,maxContactCapacity,
+																spherePos,sphereRadius,convexPos,convexOrn);
+
+			return;
+		}
+
+		if (collidables[collidableIndexA].m_shapeType == SHAPE_CONVEX_HULL &&
+			collidables[collidableIndexB].m_shapeType == SHAPE_SPHERE)
+		{
+		
+			float4 spherePos = rigidBodies[bodyIndexB].m_pos;
+			float sphereRadius = collidables[collidableIndexB].m_radius;
+			float4 convexPos = rigidBodies[bodyIndexA].m_pos;
+			float4 convexOrn = rigidBodies[bodyIndexA].m_quat;
+
+			computeContactSphereConvex(pairIndex, bodyIndexB, bodyIndexA, collidableIndexB, collidableIndexA, 
+																rigidBodies,collidables,convexShapes,vertices,indices,faces, globalContactsOut, nGlobalContactsOut,maxContactCapacity,
+																spherePos,sphereRadius,convexPos,convexOrn);
+			return;
+		}
+	
+	
+	
+		
+	
+	
+		if (collidables[collidableIndexA].m_shapeType == SHAPE_SPHERE &&
+			collidables[collidableIndexB].m_shapeType == SHAPE_SPHERE)
+		{
+			//sphere-sphere
+			float radiusA = collidables[collidableIndexA].m_radius;
+			float radiusB = collidables[collidableIndexB].m_radius;
+			float4 posA = rigidBodies[bodyIndexA].m_pos;
+			float4 posB = rigidBodies[bodyIndexB].m_pos;
+
+			float4 diff = posA-posB;
+			float len = length(diff);
+			
+			///iff distance positive, don't generate a new contact
+			if ( len <= (radiusA+radiusB))
+			{
+				///distance (negative means penetration)
+				float dist = len - (radiusA+radiusB);
+				float4 normalOnSurfaceB = make_float4(1.f,0.f,0.f,0.f);
+				if (len > 0.00001)
+				{
+					normalOnSurfaceB = diff / len;
+				}
+				float4 contactPosB = posB + normalOnSurfaceB*radiusB;
+				contactPosB.w = dist;
+								
+				int dstIdx;
+				 AppendInc( nGlobalContactsOut, dstIdx );
+				
+				if (dstIdx < maxContactCapacity)
+				{
+					__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];
+					c->m_worldNormalOnB = normalOnSurfaceB;
+					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+					c->m_batchIdx = pairIndex;
+					int bodyA = pairs[pairIndex].x;
+					int bodyB = pairs[pairIndex].y;
+					c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;
+					c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;
+					c->m_worldPosB[0] = contactPosB;
+					c->m_childIndexA = -1;
+					c->m_childIndexB = -1;
+					GET_NPOINTS(*c) = 1;
+				}//if (dstIdx < numPairs)
+			}//if ( len <= (radiusA+radiusB))
+
+			return;
+		}//SHAPE_SPHERE SHAPE_SPHERE
+
+	}//	if (i<numPairs)
+
+}
+
+
+// work-in-progress
+__kernel void   processCompoundPairsPrimitivesKernel( __global const int4* gpuCompoundPairs,
+													__global const BodyData* rigidBodies, 
+													__global const btCollidableGpu* collidables,
+													__global const ConvexPolyhedronCL* convexShapes, 
+													__global const float4* vertices,
+													__global const float4* uniqueEdges,
+													__global const btGpuFace* faces,
+													__global const int* indices,
+													__global btAabbCL* aabbs,
+													__global const btGpuChildShape* gpuChildShapes,
+													__global struct b3Contact4Data* restrict globalContactsOut,
+													counter32_t nGlobalContactsOut,
+													int numCompoundPairs, int maxContactCapacity
+													)
+{
+
+	int i = get_global_id(0);
+	if (i<numCompoundPairs)
+	{
+		int bodyIndexA = gpuCompoundPairs[i].x;
+		int bodyIndexB = gpuCompoundPairs[i].y;
+
+		int childShapeIndexA = gpuCompoundPairs[i].z;
+		int childShapeIndexB = gpuCompoundPairs[i].w;
+		
+		int collidableIndexA = -1;
+		int collidableIndexB = -1;
+		
+		float4 ornA = rigidBodies[bodyIndexA].m_quat;
+		float4 posA = rigidBodies[bodyIndexA].m_pos;
+		
+		float4 ornB = rigidBodies[bodyIndexB].m_quat;
+		float4 posB = rigidBodies[bodyIndexB].m_pos;
+							
+		if (childShapeIndexA >= 0)
+		{
+			collidableIndexA = gpuChildShapes[childShapeIndexA].m_shapeIndex;
+			float4 childPosA = gpuChildShapes[childShapeIndexA].m_childPosition;
+			float4 childOrnA = gpuChildShapes[childShapeIndexA].m_childOrientation;
+			float4 newPosA = qtRotate(ornA,childPosA)+posA;
+			float4 newOrnA = qtMul(ornA,childOrnA);
+			posA = newPosA;
+			ornA = newOrnA;
+		} else
+		{
+			collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+		}
+		
+		if (childShapeIndexB>=0)
+		{
+			collidableIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+			float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+			float4 newPosB = transform(&childPosB,&posB,&ornB);
+			float4 newOrnB = qtMul(ornB,childOrnB);
+			posB = newPosB;
+			ornB = newOrnB;
+		} else
+		{
+			collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;	
+		}
+	
+		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+	
+		int shapeTypeA = collidables[collidableIndexA].m_shapeType;
+		int shapeTypeB = collidables[collidableIndexB].m_shapeType;
+
+		int pairIndex = i;
+		if ((shapeTypeA == SHAPE_PLANE) && (shapeTypeB==SHAPE_CONVEX_HULL))
+		{
+
+			computeContactPlaneConvex( pairIndex, bodyIndexA,bodyIndexB,  collidableIndexA,collidableIndexB, 
+																rigidBodies,collidables,convexShapes,vertices,indices,
+																faces,	globalContactsOut, nGlobalContactsOut,maxContactCapacity,posB,ornB);
+			return;
+		}
+
+		if ((shapeTypeA == SHAPE_CONVEX_HULL) && (shapeTypeB==SHAPE_PLANE))
+		{
+
+			computeContactPlaneConvex( pairIndex, bodyIndexB,bodyIndexA,  collidableIndexB,collidableIndexA, 
+																rigidBodies,collidables,convexShapes,vertices,indices,
+																faces,	globalContactsOut, nGlobalContactsOut,maxContactCapacity,posA,ornA);
+			return;
+		}
+
+		if ((shapeTypeA == SHAPE_CONVEX_HULL) && (shapeTypeB == SHAPE_SPHERE))
+		{
+			float4 spherePos = rigidBodies[bodyIndexB].m_pos;
+			float sphereRadius = collidables[collidableIndexB].m_radius;
+			float4 convexPos = posA;
+			float4 convexOrn = ornA;
+			
+			computeContactSphereConvex(pairIndex, bodyIndexB, bodyIndexA , collidableIndexB,collidableIndexA, 
+										rigidBodies,collidables,convexShapes,vertices,indices,faces, globalContactsOut, nGlobalContactsOut,maxContactCapacity,
+										spherePos,sphereRadius,convexPos,convexOrn);
+	
+			return;
+		}
+
+		if ((shapeTypeA == SHAPE_SPHERE) && (shapeTypeB == SHAPE_CONVEX_HULL))
+		{
+
+			float4 spherePos = rigidBodies[bodyIndexA].m_pos;
+			float sphereRadius = collidables[collidableIndexA].m_radius;
+			float4 convexPos = posB;
+			float4 convexOrn = ornB;
+
+			
+			computeContactSphereConvex(pairIndex, bodyIndexA, bodyIndexB, collidableIndexA, collidableIndexB, 
+										rigidBodies,collidables,convexShapes,vertices,indices,faces, globalContactsOut, nGlobalContactsOut,maxContactCapacity,
+										spherePos,sphereRadius,convexPos,convexOrn);
+	
+			return;
+		}
+	}//	if (i<numCompoundPairs)
+}
+
+
+bool pointInTriangle(const float4* vertices, const float4* normal, float4 *p )
+{
+
+	const float4* p1 = &vertices[0];
+	const float4* p2 = &vertices[1];
+	const float4* p3 = &vertices[2];
+
+	float4 edge1;	edge1 = (*p2 - *p1);
+	float4 edge2;	edge2 = ( *p3 - *p2 );
+	float4 edge3;	edge3 = ( *p1 - *p3 );
+
+	
+	float4 p1_to_p; p1_to_p = ( *p - *p1 );
+	float4 p2_to_p; p2_to_p = ( *p - *p2 );
+	float4 p3_to_p; p3_to_p = ( *p - *p3 );
+
+	float4 edge1_normal; edge1_normal = ( cross(edge1,*normal));
+	float4 edge2_normal; edge2_normal = ( cross(edge2,*normal));
+	float4 edge3_normal; edge3_normal = ( cross(edge3,*normal));
+
+	
+	
+	float r1, r2, r3;
+	r1 = dot(edge1_normal,p1_to_p );
+	r2 = dot(edge2_normal,p2_to_p );
+	r3 = dot(edge3_normal,p3_to_p );
+	
+	if ( r1 > 0 && r2 > 0 && r3 > 0 )
+		return true;
+    if ( r1 <= 0 && r2 <= 0 && r3 <= 0 ) 
+		return true;
+	return false;
+
+}
+
+
+float segmentSqrDistance(float4 from, float4 to,float4 p, float4* nearest) 
+{
+	float4 diff = p - from;
+	float4 v = to - from;
+	float t = dot(v,diff);
+	
+	if (t > 0) 
+	{
+		float dotVV = dot(v,v);
+		if (t < dotVV) 
+		{
+			t /= dotVV;
+			diff -= t*v;
+		} else 
+		{
+			t = 1;
+			diff -= v;
+		}
+	} else
+	{
+		t = 0;
+	}
+	*nearest = from + t*v;
+	return dot(diff,diff);	
+}
+
+
+void	computeContactSphereTriangle(int pairIndex,
+									int bodyIndexA, int bodyIndexB,
+									int collidableIndexA, int collidableIndexB, 
+									__global const BodyData* rigidBodies, 
+									__global const btCollidableGpu* collidables,
+									const float4* triangleVertices,
+									__global struct b3Contact4Data* restrict globalContactsOut,
+									counter32_t nGlobalContactsOut,
+									int maxContactCapacity,
+									float4 spherePos2,
+									float radius,
+									float4 pos,
+									float4 quat,
+									int faceIndex
+									)
+{
+
+	float4 invPos;
+	float4 invOrn;
+
+	trInverse(pos,quat, &invPos,&invOrn);
+	float4 spherePos = transform(&spherePos2,&invPos,&invOrn);
+	int numFaces = 3;
+	float4 closestPnt = (float4)(0, 0, 0, 0);
+	float4 hitNormalWorld = (float4)(0, 0, 0, 0);
+	float minDist = -1000000.f;
+	bool bCollide = false;
+
+	
+	//////////////////////////////////////
+
+	float4 sphereCenter;
+	sphereCenter = spherePos;
+
+	const float4* vertices = triangleVertices;
+	float contactBreakingThreshold = 0.f;//todo?
+	float radiusWithThreshold = radius + contactBreakingThreshold;
+	float4 edge10;
+	edge10 = vertices[1]-vertices[0];
+	edge10.w = 0.f;//is this needed?
+	float4 edge20;
+	edge20 = vertices[2]-vertices[0];
+	edge20.w = 0.f;//is this needed?
+	float4 normal = cross3(edge10,edge20);
+	normal = normalize(normal);
+	float4 p1ToCenter;
+	p1ToCenter = sphereCenter - vertices[0];
+	
+	float distanceFromPlane = dot(p1ToCenter,normal);
+
+	if (distanceFromPlane < 0.f)
+	{
+		//triangle facing the other way
+		distanceFromPlane *= -1.f;
+		normal *= -1.f;
+	}
+	hitNormalWorld = normal;
+
+	bool isInsideContactPlane = distanceFromPlane < radiusWithThreshold;
+	
+	// Check for contact / intersection
+	bool hasContact = false;
+	float4 contactPoint;
+	if (isInsideContactPlane) 
+	{
+	
+		if (pointInTriangle(vertices,&normal, &sphereCenter)) 
+		{
+			// Inside the contact wedge - touches a point on the shell plane
+			hasContact = true;
+			contactPoint = sphereCenter - normal*distanceFromPlane;
+			
+		} else {
+			// Could be inside one of the contact capsules
+			float contactCapsuleRadiusSqr = radiusWithThreshold*radiusWithThreshold;
+			float4 nearestOnEdge;
+			int numEdges = 3;
+			for (int i = 0; i < numEdges; i++) 
+			{
+				float4 pa =vertices[i];
+				float4 pb = vertices[(i+1)%3];
+
+				float distanceSqr = segmentSqrDistance(pa,pb,sphereCenter, &nearestOnEdge);
+				if (distanceSqr < contactCapsuleRadiusSqr) 
+				{
+					// Yep, we're inside a capsule
+					hasContact = true;
+					contactPoint = nearestOnEdge;
+					
+				}
+				
+			}
+		}
+	}
+
+	if (hasContact) 
+	{
+
+		closestPnt = contactPoint;
+		float4 contactToCenter = sphereCenter - contactPoint;
+		minDist = length(contactToCenter);
+		if (minDist>FLT_EPSILON)
+		{
+			hitNormalWorld = normalize(contactToCenter);//*(1./minDist);
+			bCollide  = true;
+		}
+		
+	}
+
+
+	/////////////////////////////////////
+
+	if (bCollide && minDist > -10000)
+	{
+		
+		float4 normalOnSurfaceB1 = qtRotate(quat,-hitNormalWorld);
+		float4 pOnB1 = transform(&closestPnt,&pos,&quat);
+		float actualDepth = minDist-radius;
+
+		
+		if (actualDepth<=0.f)
+		{
+			pOnB1.w = actualDepth;
+			int dstIdx;
+
+			
+			float lenSqr = dot3F4(normalOnSurfaceB1,normalOnSurfaceB1);
+			if (lenSqr>FLT_EPSILON)
+			{
+				AppendInc( nGlobalContactsOut, dstIdx );
+			
+				if (dstIdx < maxContactCapacity)
+				{
+					__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];
+					c->m_worldNormalOnB = -normalOnSurfaceB1;
+					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+					c->m_batchIdx = pairIndex;
+					c->m_bodyAPtrAndSignBit = rigidBodies[bodyIndexA].m_invMass==0?-bodyIndexA:bodyIndexA;
+					c->m_bodyBPtrAndSignBit = rigidBodies[bodyIndexB].m_invMass==0?-bodyIndexB:bodyIndexB;
+					c->m_worldPosB[0] = pOnB1;
+
+					c->m_childIndexA = -1;
+					c->m_childIndexB = faceIndex;
+
+					GET_NPOINTS(*c) = 1;
+				} 
+			}
+
+		}
+	}//if (hasCollision)
+
+}
+
+
+
+// work-in-progress
+__kernel void   findConcaveSphereContactsKernel( __global int4* concavePairs,
+												__global const BodyData* rigidBodies,
+												__global const btCollidableGpu* collidables,
+												__global const ConvexPolyhedronCL* convexShapes, 
+												__global const float4* vertices,
+												__global const float4* uniqueEdges,
+												__global const btGpuFace* faces,
+												__global const int* indices,
+												__global btAabbCL* aabbs,
+												__global struct b3Contact4Data* restrict globalContactsOut,
+												counter32_t nGlobalContactsOut,
+													int numConcavePairs, int maxContactCapacity
+												)
+{
+
+	int i = get_global_id(0);
+	if (i>=numConcavePairs)
+		return;
+	int pairIdx = i;
+
+	int bodyIndexA = concavePairs[i].x;
+	int bodyIndexB = concavePairs[i].y;
+
+	int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+	int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+
+	int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+	int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+
+	if (collidables[collidableIndexB].m_shapeType==SHAPE_SPHERE)
+	{
+		int f = concavePairs[i].z;
+		btGpuFace face = faces[convexShapes[shapeIndexA].m_faceOffset+f];
+		
+		float4 verticesA[3];
+		for (int i=0;i<3;i++)
+		{
+			int index = indices[face.m_indexOffset+i];
+			float4 vert = vertices[convexShapes[shapeIndexA].m_vertexOffset+index];
+			verticesA[i] = vert;
+		}
+
+		float4 spherePos = rigidBodies[bodyIndexB].m_pos;
+		float sphereRadius = collidables[collidableIndexB].m_radius;
+		float4 convexPos = rigidBodies[bodyIndexA].m_pos;
+		float4 convexOrn = rigidBodies[bodyIndexA].m_quat;
+
+		computeContactSphereTriangle(i, bodyIndexB, bodyIndexA, collidableIndexB, collidableIndexA, 
+																rigidBodies,collidables,
+																verticesA,
+																globalContactsOut, nGlobalContactsOut,maxContactCapacity,
+																spherePos,sphereRadius,convexPos,convexOrn, f);
+
+		return;
+	}
+}
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/primitiveContacts.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/primitiveContacts.h
new file mode 100644
index 0000000000..b0103fe674
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/primitiveContacts.h
@@ -0,0 +1,1289 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* primitiveContactsKernelsCL= \
+"#ifndef B3_CONTACT4DATA_H\n"
+"#define B3_CONTACT4DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#define B3_FLOAT4_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#define B3_PLATFORM_DEFINITIONS_H\n"
+"struct MyTest\n"
+"{\n"
+"	int bla;\n"
+"};\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"//keep B3_LARGE_FLOAT*B3_LARGE_FLOAT < FLT_MAX\n"
+"#define B3_LARGE_FLOAT 1e18f\n"
+"#define B3_INFINITY 1e18f\n"
+"#define b3Assert(a)\n"
+"#define b3ConstArray(a) __global const a*\n"
+"#define b3AtomicInc atomic_inc\n"
+"#define b3AtomicAdd atomic_add\n"
+"#define b3Fabs fabs\n"
+"#define b3Sqrt native_sqrt\n"
+"#define b3Sin native_sin\n"
+"#define b3Cos native_cos\n"
+"#define B3_STATIC\n"
+"#endif\n"
+"#endif\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Float4;\n"
+"	#define b3Float4ConstArg const b3Float4\n"
+"	#define b3MakeFloat4 (float4)\n"
+"	float b3Dot3F4(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return dot(a1, b1);\n"
+"	}\n"
+"	b3Float4 b3Cross3(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return cross(a1, b1);\n"
+"	}\n"
+"	#define b3MinFloat4 min\n"
+"	#define b3MaxFloat4 max\n"
+"	#define b3Normalized(a) normalize(a)\n"
+"#endif \n"
+"		\n"
+"inline bool b3IsAlmostZero(b3Float4ConstArg v)\n"
+"{\n"
+"	if(b3Fabs(v.x)>1e-6 || b3Fabs(v.y)>1e-6 || b3Fabs(v.z)>1e-6)	\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"inline int    b3MaxDot( b3Float4ConstArg vec, __global const b3Float4* vecArray, int vecLen, float* dotOut )\n"
+"{\n"
+"    float maxDot = -B3_INFINITY;\n"
+"    int i = 0;\n"
+"    int ptIndex = -1;\n"
+"    for( i = 0; i < vecLen; i++ )\n"
+"    {\n"
+"        float dot = b3Dot3F4(vecArray[i],vec);\n"
+"            \n"
+"        if( dot > maxDot )\n"
+"        {\n"
+"            maxDot = dot;\n"
+"            ptIndex = i;\n"
+"        }\n"
+"    }\n"
+"	b3Assert(ptIndex>=0);\n"
+"    if (ptIndex<0)\n"
+"	{\n"
+"		ptIndex = 0;\n"
+"	}\n"
+"    *dotOut = maxDot;\n"
+"    return ptIndex;\n"
+"}\n"
+"#endif //B3_FLOAT4_H\n"
+"typedef  struct b3Contact4Data b3Contact4Data_t;\n"
+"struct b3Contact4Data\n"
+"{\n"
+"	b3Float4	m_worldPosB[4];\n"
+"//	b3Float4	m_localPosA[4];\n"
+"//	b3Float4	m_localPosB[4];\n"
+"	b3Float4	m_worldNormalOnB;	//	w: m_nPoints\n"
+"	unsigned short  m_restituitionCoeffCmp;\n"
+"	unsigned short  m_frictionCoeffCmp;\n"
+"	int m_batchIdx;\n"
+"	int m_bodyAPtrAndSignBit;//x:m_bodyAPtr, y:m_bodyBPtr\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"	int	m_childIndexA;\n"
+"	int	m_childIndexB;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"};\n"
+"inline int b3Contact4Data_getNumPoints(const struct b3Contact4Data* contact)\n"
+"{\n"
+"	return (int)contact->m_worldNormalOnB.w;\n"
+"};\n"
+"inline void b3Contact4Data_setNumPoints(struct b3Contact4Data* contact, int numPoints)\n"
+"{\n"
+"	contact->m_worldNormalOnB.w = (float)numPoints;\n"
+"};\n"
+"#endif //B3_CONTACT4DATA_H\n"
+"#define SHAPE_CONVEX_HULL 3\n"
+"#define SHAPE_PLANE 4\n"
+"#define SHAPE_CONCAVE_TRIMESH 5\n"
+"#define SHAPE_COMPOUND_OF_CONVEX_HULLS 6\n"
+"#define SHAPE_SPHERE 7\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile __global int*\n"
+"#endif\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"typedef unsigned int u32;\n"
+"typedef struct \n"
+"{\n"
+"	union\n"
+"	{\n"
+"		float4	m_min;\n"
+"		float   m_minElems[4];\n"
+"		int			m_minIndices[4];\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float4	m_max;\n"
+"		float   m_maxElems[4];\n"
+"		int			m_maxIndices[4];\n"
+"	};\n"
+"} btAabbCL;\n"
+"///keep this in sync with btCollidable.h\n"
+"typedef struct\n"
+"{\n"
+"	int m_numChildShapes;\n"
+"	float m_radius;\n"
+"	int m_shapeType;\n"
+"	int m_shapeIndex;\n"
+"	\n"
+"} btCollidableGpu;\n"
+"typedef struct\n"
+"{\n"
+"	float4	m_childPosition;\n"
+"	float4	m_childOrientation;\n"
+"	int m_shapeIndex;\n"
+"	int m_unused0;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"} btGpuChildShape;\n"
+"#define GET_NPOINTS(x) (x).m_worldNormalOnB.w\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	float4 m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"	u32 m_collidableIdx;	\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} BodyData;\n"
+"typedef struct  \n"
+"{\n"
+"	float4		m_localCenter;\n"
+"	float4		m_extents;\n"
+"	float4		mC;\n"
+"	float4		mE;\n"
+"	\n"
+"	float			m_radius;\n"
+"	int	m_faceOffset;\n"
+"	int m_numFaces;\n"
+"	int	m_numVertices;\n"
+"	\n"
+"	int m_vertexOffset;\n"
+"	int	m_uniqueEdgesOffset;\n"
+"	int	m_numUniqueEdges;\n"
+"	int m_unused;\n"
+"} ConvexPolyhedronCL;\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_plane;\n"
+"	int m_indexOffset;\n"
+"	int m_numIndices;\n"
+"} btGpuFace;\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"#define make_float4 (float4)\n"
+"#define make_float2 (float2)\n"
+"#define make_uint4 (uint4)\n"
+"#define make_int4 (int4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"__inline\n"
+"float fastDiv(float numerator, float denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"//	return numerator/denominator;	\n"
+"}\n"
+"__inline\n"
+"float4 fastDiv4(float4 numerator, float4 denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"}\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"//#define dot3F4 dot\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = make_float4(a.xyz,0.f);\n"
+"	float4 b1 = make_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"typedef float4 Quaternion;\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b);\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in);\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec);\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q);\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in)\n"
+"{\n"
+"	return fastNormalize4(in);\n"
+"//	in /= length( in );\n"
+"//	return in;\n"
+"}\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(qtMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"__inline\n"
+"float4 qtInvRotate(const Quaternion q, float4 vec)\n"
+"{\n"
+"	return qtRotate( qtInvert( q ), vec );\n"
+"}\n"
+"__inline\n"
+"float4 transform(const float4* p, const float4* translation, const Quaternion* orientation)\n"
+"{\n"
+"	return qtRotate( *orientation, *p ) + (*translation);\n"
+"}\n"
+"void	trInverse(float4 translationIn, Quaternion orientationIn,\n"
+"		float4* translationOut, Quaternion* orientationOut)\n"
+"{\n"
+"	*orientationOut = qtInvert(orientationIn);\n"
+"	*translationOut = qtRotate(*orientationOut, -translationIn);\n"
+"}\n"
+"void	trMul(float4 translationA, Quaternion orientationA,\n"
+"						float4 translationB, Quaternion orientationB,\n"
+"		float4* translationOut, Quaternion* orientationOut)\n"
+"{\n"
+"	*orientationOut = qtMul(orientationA,orientationB);\n"
+"	*translationOut = transform(&translationB,&translationA,&orientationA);\n"
+"}\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = make_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"}\n"
+"__inline float4 lerp3(const float4 a,const float4 b, float  t)\n"
+"{\n"
+"	return make_float4(	a.x + (b.x - a.x) * t,\n"
+"						a.y + (b.y - a.y) * t,\n"
+"						a.z + (b.z - a.z) * t,\n"
+"						0.f);\n"
+"}\n"
+"float signedDistanceFromPointToPlane(float4 point, float4 planeEqn, float4* closestPointOnFace)\n"
+"{\n"
+"	float4 n = (float4)(planeEqn.x, planeEqn.y, planeEqn.z, 0);\n"
+"	float dist = dot3F4(n, point) + planeEqn.w;\n"
+"	*closestPointOnFace = point - dist * n;\n"
+"	return dist;\n"
+"}\n"
+"inline bool IsPointInPolygon(float4 p, \n"
+"							const btGpuFace* face,\n"
+"							__global const float4* baseVertex,\n"
+"							__global const  int* convexIndices,\n"
+"							float4* out)\n"
+"{\n"
+"    float4 a;\n"
+"    float4 b;\n"
+"    float4 ab;\n"
+"    float4 ap;\n"
+"    float4 v;\n"
+"	float4 plane = make_float4(face->m_plane.x,face->m_plane.y,face->m_plane.z,0.f);\n"
+"	\n"
+"	if (face->m_numIndices<2)\n"
+"		return false;\n"
+"	\n"
+"	float4 v0 = baseVertex[convexIndices[face->m_indexOffset + face->m_numIndices-1]];\n"
+"	\n"
+"	b = v0;\n"
+"    for(unsigned i=0; i != face->m_numIndices; ++i)\n"
+"    {\n"
+"		a = b;\n"
+"		float4 vi = baseVertex[convexIndices[face->m_indexOffset + i]];\n"
+"		b = vi;\n"
+"        ab = b-a;\n"
+"        ap = p-a;\n"
+"        v = cross3(ab,plane);\n"
+"        if (dot(ap, v) > 0.f)\n"
+"        {\n"
+"            float ab_m2 = dot(ab, ab);\n"
+"            float rt = ab_m2 != 0.f ? dot(ab, ap) / ab_m2 : 0.f;\n"
+"            if (rt <= 0.f)\n"
+"            {\n"
+"                *out = a;\n"
+"            }\n"
+"            else if (rt >= 1.f) \n"
+"            {\n"
+"                *out = b;\n"
+"            }\n"
+"            else\n"
+"            {\n"
+"            	float s = 1.f - rt;\n"
+"				out[0].x = s * a.x + rt * b.x;\n"
+"				out[0].y = s * a.y + rt * b.y;\n"
+"				out[0].z = s * a.z + rt * b.z;\n"
+"            }\n"
+"            return false;\n"
+"        }\n"
+"    }\n"
+"    return true;\n"
+"}\n"
+"void	computeContactSphereConvex(int pairIndex,\n"
+"																int bodyIndexA, int bodyIndexB, \n"
+"																int collidableIndexA, int collidableIndexB, \n"
+"																__global const BodyData* rigidBodies, \n"
+"																__global const btCollidableGpu* collidables,\n"
+"																__global const ConvexPolyhedronCL* convexShapes,\n"
+"																__global const float4* convexVertices,\n"
+"																__global const int* convexIndices,\n"
+"																__global const btGpuFace* faces,\n"
+"																__global struct b3Contact4Data* restrict globalContactsOut,\n"
+"																counter32_t nGlobalContactsOut,\n"
+"																int maxContactCapacity,\n"
+"																float4 spherePos2,\n"
+"																float radius,\n"
+"																float4 pos,\n"
+"																float4 quat\n"
+"																)\n"
+"{\n"
+"	float4 invPos;\n"
+"	float4 invOrn;\n"
+"	trInverse(pos,quat, &invPos,&invOrn);\n"
+"	float4 spherePos = transform(&spherePos2,&invPos,&invOrn);\n"
+"	int shapeIndex = collidables[collidableIndexB].m_shapeIndex;\n"
+"	int numFaces = convexShapes[shapeIndex].m_numFaces;\n"
+"	float4 closestPnt = (float4)(0, 0, 0, 0);\n"
+"	float4 hitNormalWorld = (float4)(0, 0, 0, 0);\n"
+"	float minDist = -1000000.f;\n"
+"	bool bCollide = true;\n"
+"	for ( int f = 0; f < numFaces; f++ )\n"
+"	{\n"
+"		btGpuFace face = faces[convexShapes[shapeIndex].m_faceOffset+f];\n"
+"		// set up a plane equation \n"
+"		float4 planeEqn;\n"
+"		float4 n1 = face.m_plane;\n"
+"		n1.w = 0.f;\n"
+"		planeEqn = n1;\n"
+"		planeEqn.w = face.m_plane.w;\n"
+"		\n"
+"	\n"
+"		// compute a signed distance from the vertex in cloth to the face of rigidbody.\n"
+"		float4 pntReturn;\n"
+"		float dist = signedDistanceFromPointToPlane(spherePos, planeEqn, &pntReturn);\n"
+"		// If the distance is positive, the plane is a separating plane. \n"
+"		if ( dist > radius )\n"
+"		{\n"
+"			bCollide = false;\n"
+"			break;\n"
+"		}\n"
+"		if (dist>0)\n"
+"		{\n"
+"			//might hit an edge or vertex\n"
+"			float4 out;\n"
+"			float4 zeroPos = make_float4(0,0,0,0);\n"
+"			bool isInPoly = IsPointInPolygon(spherePos,\n"
+"					&face,\n"
+"					&convexVertices[convexShapes[shapeIndex].m_vertexOffset],\n"
+"					convexIndices,\n"
+"           &out);\n"
+"			if (isInPoly)\n"
+"			{\n"
+"				if (dist>minDist)\n"
+"				{\n"
+"					minDist = dist;\n"
+"					closestPnt = pntReturn;\n"
+"					hitNormalWorld = planeEqn;\n"
+"					\n"
+"				}\n"
+"			} else\n"
+"			{\n"
+"				float4 tmp = spherePos-out;\n"
+"				float l2 = dot(tmp,tmp);\n"
+"				if (l2<radius*radius)\n"
+"				{\n"
+"					dist  = sqrt(l2);\n"
+"					if (dist>minDist)\n"
+"					{\n"
+"						minDist = dist;\n"
+"						closestPnt = out;\n"
+"						hitNormalWorld = tmp/dist;\n"
+"						\n"
+"					}\n"
+"					\n"
+"				} else\n"
+"				{\n"
+"					bCollide = false;\n"
+"					break;\n"
+"				}\n"
+"			}\n"
+"		} else\n"
+"		{\n"
+"			if ( dist > minDist )\n"
+"			{\n"
+"				minDist = dist;\n"
+"				closestPnt = pntReturn;\n"
+"				hitNormalWorld.xyz = planeEqn.xyz;\n"
+"			}\n"
+"		}\n"
+"		\n"
+"	}\n"
+"	\n"
+"	if (bCollide && minDist > -10000)\n"
+"	{\n"
+"		float4 normalOnSurfaceB1 = qtRotate(quat,-hitNormalWorld);\n"
+"		float4 pOnB1 = transform(&closestPnt,&pos,&quat);\n"
+"		\n"
+"		float actualDepth = minDist-radius;\n"
+"		if (actualDepth<=0.f)\n"
+"		{\n"
+"			\n"
+"			pOnB1.w = actualDepth;\n"
+"			int dstIdx;\n"
+"			AppendInc( nGlobalContactsOut, dstIdx );\n"
+"		\n"
+"			\n"
+"			if (1)//dstIdx < maxContactCapacity)\n"
+"			{\n"
+"				__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];\n"
+"				c->m_worldNormalOnB = -normalOnSurfaceB1;\n"
+"				c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);\n"
+"				c->m_batchIdx = pairIndex;\n"
+"				c->m_bodyAPtrAndSignBit = rigidBodies[bodyIndexA].m_invMass==0?-bodyIndexA:bodyIndexA;\n"
+"				c->m_bodyBPtrAndSignBit = rigidBodies[bodyIndexB].m_invMass==0?-bodyIndexB:bodyIndexB;\n"
+"				c->m_worldPosB[0] = pOnB1;\n"
+"				c->m_childIndexA = -1;\n"
+"				c->m_childIndexB = -1;\n"
+"				GET_NPOINTS(*c) = 1;\n"
+"			} \n"
+"		}\n"
+"	}//if (hasCollision)\n"
+"}\n"
+"							\n"
+"int extractManifoldSequential(const float4* p, int nPoints, float4 nearNormal, int4* contactIdx)\n"
+"{\n"
+"	if( nPoints == 0 )\n"
+"        return 0;\n"
+"    \n"
+"    if (nPoints <=4)\n"
+"        return nPoints;\n"
+"    \n"
+"    \n"
+"    if (nPoints >64)\n"
+"        nPoints = 64;\n"
+"    \n"
+"	float4 center = make_float4(0.f);\n"
+"	{\n"
+"		\n"
+"		for (int i=0;i<nPoints;i++)\n"
+"			center += p[i];\n"
+"		center /= (float)nPoints;\n"
+"	}\n"
+"    \n"
+"	\n"
+"    \n"
+"	//	sample 4 directions\n"
+"    \n"
+"    float4 aVector = p[0] - center;\n"
+"    float4 u = cross3( nearNormal, aVector );\n"
+"    float4 v = cross3( nearNormal, u );\n"
+"    u = normalize3( u );\n"
+"    v = normalize3( v );\n"
+"    \n"
+"    \n"
+"    //keep point with deepest penetration\n"
+"    float minW= FLT_MAX;\n"
+"    \n"
+"    int minIndex=-1;\n"
+"    \n"
+"    float4 maxDots;\n"
+"    maxDots.x = FLT_MIN;\n"
+"    maxDots.y = FLT_MIN;\n"
+"    maxDots.z = FLT_MIN;\n"
+"    maxDots.w = FLT_MIN;\n"
+"    \n"
+"    //	idx, distance\n"
+"    for(int ie = 0; ie<nPoints; ie++ )\n"
+"    {\n"
+"        if (p[ie].w<minW)\n"
+"        {\n"
+"            minW = p[ie].w;\n"
+"            minIndex=ie;\n"
+"        }\n"
+"        float f;\n"
+"        float4 r = p[ie]-center;\n"
+"        f = dot3F4( u, r );\n"
+"        if (f<maxDots.x)\n"
+"        {\n"
+"            maxDots.x = f;\n"
+"            contactIdx[0].x = ie;\n"
+"        }\n"
+"        \n"
+"        f = dot3F4( -u, r );\n"
+"        if (f<maxDots.y)\n"
+"        {\n"
+"            maxDots.y = f;\n"
+"            contactIdx[0].y = ie;\n"
+"        }\n"
+"        \n"
+"        \n"
+"        f = dot3F4( v, r );\n"
+"        if (f<maxDots.z)\n"
+"        {\n"
+"            maxDots.z = f;\n"
+"            contactIdx[0].z = ie;\n"
+"        }\n"
+"        \n"
+"        f = dot3F4( -v, r );\n"
+"        if (f<maxDots.w)\n"
+"        {\n"
+"            maxDots.w = f;\n"
+"            contactIdx[0].w = ie;\n"
+"        }\n"
+"        \n"
+"    }\n"
+"    \n"
+"    if (contactIdx[0].x != minIndex && contactIdx[0].y != minIndex && contactIdx[0].z != minIndex && contactIdx[0].w != minIndex)\n"
+"    {\n"
+"        //replace the first contact with minimum (todo: replace contact with least penetration)\n"
+"        contactIdx[0].x = minIndex;\n"
+"    }\n"
+"    \n"
+"    return 4;\n"
+"    \n"
+"}\n"
+"#define MAX_PLANE_CONVEX_POINTS 64\n"
+"int computeContactPlaneConvex(int pairIndex,\n"
+"								int bodyIndexA, int bodyIndexB, \n"
+"								int collidableIndexA, int collidableIndexB, \n"
+"								__global const BodyData* rigidBodies, \n"
+"								__global const btCollidableGpu*collidables,\n"
+"								__global const ConvexPolyhedronCL* convexShapes,\n"
+"								__global const float4* convexVertices,\n"
+"								__global const int* convexIndices,\n"
+"								__global const btGpuFace* faces,\n"
+"								__global struct b3Contact4Data* restrict globalContactsOut,\n"
+"								counter32_t nGlobalContactsOut,\n"
+"								int maxContactCapacity,\n"
+"								float4 posB,\n"
+"								Quaternion ornB\n"
+"								)\n"
+"{\n"
+"	int resultIndex=-1;\n"
+"		int shapeIndex = collidables[collidableIndexB].m_shapeIndex;\n"
+"	__global const ConvexPolyhedronCL* hullB = &convexShapes[shapeIndex];\n"
+"	\n"
+"	float4 posA;\n"
+"	posA = rigidBodies[bodyIndexA].m_pos;\n"
+"	Quaternion ornA;\n"
+"	ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"	int numContactsOut = 0;\n"
+"	int numWorldVertsB1= 0;\n"
+"	float4 planeEq;\n"
+"	 planeEq = faces[collidables[collidableIndexA].m_shapeIndex].m_plane;\n"
+"	float4 planeNormal = make_float4(planeEq.x,planeEq.y,planeEq.z,0.f);\n"
+"	float4 planeNormalWorld;\n"
+"	planeNormalWorld = qtRotate(ornA,planeNormal);\n"
+"	float planeConstant = planeEq.w;\n"
+"	\n"
+"	float4 invPosA;Quaternion invOrnA;\n"
+"	float4 convexInPlaneTransPos1; Quaternion convexInPlaneTransOrn1;\n"
+"	{\n"
+"		\n"
+"		trInverse(posA,ornA,&invPosA,&invOrnA);\n"
+"		trMul(invPosA,invOrnA,posB,ornB,&convexInPlaneTransPos1,&convexInPlaneTransOrn1);\n"
+"	}\n"
+"	float4 invPosB;Quaternion invOrnB;\n"
+"	float4 planeInConvexPos1;	Quaternion planeInConvexOrn1;\n"
+"	{\n"
+"		\n"
+"		trInverse(posB,ornB,&invPosB,&invOrnB);\n"
+"		trMul(invPosB,invOrnB,posA,ornA,&planeInConvexPos1,&planeInConvexOrn1);	\n"
+"	}\n"
+"	\n"
+"	float4 planeNormalInConvex = qtRotate(planeInConvexOrn1,-planeNormal);\n"
+"	float maxDot = -1e30;\n"
+"	int hitVertex=-1;\n"
+"	float4 hitVtx;\n"
+"	float4 contactPoints[MAX_PLANE_CONVEX_POINTS];\n"
+"	int numPoints = 0;\n"
+"	int4 contactIdx;\n"
+"	contactIdx=make_int4(0,1,2,3);\n"
+"    \n"
+"	\n"
+"	for (int i=0;i<hullB->m_numVertices;i++)\n"
+"	{\n"
+"		float4 vtx = convexVertices[hullB->m_vertexOffset+i];\n"
+"		float curDot = dot(vtx,planeNormalInConvex);\n"
+"		if (curDot>maxDot)\n"
+"		{\n"
+"			hitVertex=i;\n"
+"			maxDot=curDot;\n"
+"			hitVtx = vtx;\n"
+"			//make sure the deepest points is always included\n"
+"			if (numPoints==MAX_PLANE_CONVEX_POINTS)\n"
+"				numPoints--;\n"
+"		}\n"
+"		if (numPoints<MAX_PLANE_CONVEX_POINTS)\n"
+"		{\n"
+"			float4 vtxWorld = transform(&vtx, &posB, &ornB);\n"
+"			float4 vtxInPlane = transform(&vtxWorld, &invPosA, &invOrnA);//oplaneTransform.inverse()*vtxWorld;\n"
+"			float dist = dot(planeNormal,vtxInPlane)-planeConstant;\n"
+"			if (dist<0.f)\n"
+"			{\n"
+"				vtxWorld.w = dist;\n"
+"				contactPoints[numPoints] = vtxWorld;\n"
+"				numPoints++;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	int numReducedPoints  = numPoints;\n"
+"	if (numPoints>4)\n"
+"	{\n"
+"		numReducedPoints = extractManifoldSequential( contactPoints, numPoints, planeNormalInConvex, &contactIdx);\n"
+"	}\n"
+"	if (numReducedPoints>0)\n"
+"	{\n"
+"		int dstIdx;\n"
+"	    AppendInc( nGlobalContactsOut, dstIdx );\n"
+"		if (dstIdx < maxContactCapacity)\n"
+"		{\n"
+"			resultIndex = dstIdx;\n"
+"			__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];\n"
+"			c->m_worldNormalOnB = -planeNormalWorld;\n"
+"			//c->setFrictionCoeff(0.7);\n"
+"			//c->setRestituitionCoeff(0.f);\n"
+"			c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);\n"
+"			c->m_batchIdx = pairIndex;\n"
+"			c->m_bodyAPtrAndSignBit = rigidBodies[bodyIndexA].m_invMass==0?-bodyIndexA:bodyIndexA;\n"
+"			c->m_bodyBPtrAndSignBit = rigidBodies[bodyIndexB].m_invMass==0?-bodyIndexB:bodyIndexB;\n"
+"			c->m_childIndexA = -1;\n"
+"			c->m_childIndexB = -1;\n"
+"			switch (numReducedPoints)\n"
+"            {\n"
+"                case 4:\n"
+"                    c->m_worldPosB[3] = contactPoints[contactIdx.w];\n"
+"                case 3:\n"
+"                    c->m_worldPosB[2] = contactPoints[contactIdx.z];\n"
+"                case 2:\n"
+"                    c->m_worldPosB[1] = contactPoints[contactIdx.y];\n"
+"                case 1:\n"
+"                    c->m_worldPosB[0] = contactPoints[contactIdx.x];\n"
+"                default:\n"
+"                {\n"
+"                }\n"
+"            };\n"
+"			\n"
+"			GET_NPOINTS(*c) = numReducedPoints;\n"
+"		}//if (dstIdx < numPairs)\n"
+"	}	\n"
+"	return resultIndex;\n"
+"}\n"
+"void	computeContactPlaneSphere(int pairIndex,\n"
+"																int bodyIndexA, int bodyIndexB, \n"
+"																int collidableIndexA, int collidableIndexB, \n"
+"																__global const BodyData* rigidBodies, \n"
+"																__global const btCollidableGpu* collidables,\n"
+"																__global const btGpuFace* faces,\n"
+"																__global struct b3Contact4Data* restrict globalContactsOut,\n"
+"																counter32_t nGlobalContactsOut,\n"
+"																int maxContactCapacity)\n"
+"{\n"
+"	float4 planeEq = faces[collidables[collidableIndexA].m_shapeIndex].m_plane;\n"
+"	float radius = collidables[collidableIndexB].m_radius;\n"
+"	float4 posA1 = rigidBodies[bodyIndexA].m_pos;\n"
+"	float4 ornA1 = rigidBodies[bodyIndexA].m_quat;\n"
+"	float4 posB1 = rigidBodies[bodyIndexB].m_pos;\n"
+"	float4 ornB1 = rigidBodies[bodyIndexB].m_quat;\n"
+"	\n"
+"	bool hasCollision = false;\n"
+"	float4 planeNormal1 = make_float4(planeEq.x,planeEq.y,planeEq.z,0.f);\n"
+"	float planeConstant = planeEq.w;\n"
+"	float4 convexInPlaneTransPos1; Quaternion convexInPlaneTransOrn1;\n"
+"	{\n"
+"		float4 invPosA;Quaternion invOrnA;\n"
+"		trInverse(posA1,ornA1,&invPosA,&invOrnA);\n"
+"		trMul(invPosA,invOrnA,posB1,ornB1,&convexInPlaneTransPos1,&convexInPlaneTransOrn1);\n"
+"	}\n"
+"	float4 planeInConvexPos1;	Quaternion planeInConvexOrn1;\n"
+"	{\n"
+"		float4 invPosB;Quaternion invOrnB;\n"
+"		trInverse(posB1,ornB1,&invPosB,&invOrnB);\n"
+"		trMul(invPosB,invOrnB,posA1,ornA1,&planeInConvexPos1,&planeInConvexOrn1);	\n"
+"	}\n"
+"	float4 vtx1 = qtRotate(planeInConvexOrn1,-planeNormal1)*radius;\n"
+"	float4 vtxInPlane1 = transform(&vtx1,&convexInPlaneTransPos1,&convexInPlaneTransOrn1);\n"
+"	float distance = dot3F4(planeNormal1,vtxInPlane1) - planeConstant;\n"
+"	hasCollision = distance < 0.f;//m_manifoldPtr->getContactBreakingThreshold();\n"
+"	if (hasCollision)\n"
+"	{\n"
+"		float4 vtxInPlaneProjected1 = vtxInPlane1 -   distance*planeNormal1;\n"
+"		float4 vtxInPlaneWorld1 = transform(&vtxInPlaneProjected1,&posA1,&ornA1);\n"
+"		float4 normalOnSurfaceB1 = qtRotate(ornA1,planeNormal1);\n"
+"		float4 pOnB1 = vtxInPlaneWorld1+normalOnSurfaceB1*distance;\n"
+"		pOnB1.w = distance;\n"
+"		int dstIdx;\n"
+"    AppendInc( nGlobalContactsOut, dstIdx );\n"
+"		\n"
+"		if (dstIdx < maxContactCapacity)\n"
+"		{\n"
+"			__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];\n"
+"			c->m_worldNormalOnB = -normalOnSurfaceB1;\n"
+"			c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);\n"
+"			c->m_batchIdx = pairIndex;\n"
+"			c->m_bodyAPtrAndSignBit = rigidBodies[bodyIndexA].m_invMass==0?-bodyIndexA:bodyIndexA;\n"
+"			c->m_bodyBPtrAndSignBit = rigidBodies[bodyIndexB].m_invMass==0?-bodyIndexB:bodyIndexB;\n"
+"			c->m_worldPosB[0] = pOnB1;\n"
+"			c->m_childIndexA = -1;\n"
+"			c->m_childIndexB = -1;\n"
+"			GET_NPOINTS(*c) = 1;\n"
+"		}//if (dstIdx < numPairs)\n"
+"	}//if (hasCollision)\n"
+"}\n"
+"__kernel void   primitiveContactsKernel( __global int4* pairs, \n"
+"																					__global const BodyData* rigidBodies, \n"
+"																					__global const btCollidableGpu* collidables,\n"
+"																					__global const ConvexPolyhedronCL* convexShapes, \n"
+"																					__global const float4* vertices,\n"
+"																					__global const float4* uniqueEdges,\n"
+"																					__global const btGpuFace* faces,\n"
+"																					__global const int* indices,\n"
+"																					__global struct b3Contact4Data* restrict globalContactsOut,\n"
+"																					counter32_t nGlobalContactsOut,\n"
+"																					int numPairs, int maxContactCapacity)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	int pairIndex = i;\n"
+"	\n"
+"	float4 worldVertsB1[64];\n"
+"	float4 worldVertsB2[64];\n"
+"	int capacityWorldVerts = 64;	\n"
+"	float4 localContactsOut[64];\n"
+"	int localContactCapacity=64;\n"
+"	\n"
+"	float minDist = -1e30f;\n"
+"	float maxDist = 0.02f;\n"
+"	if (i<numPairs)\n"
+"	{\n"
+"		int bodyIndexA = pairs[i].x;\n"
+"		int bodyIndexB = pairs[i].y;\n"
+"			\n"
+"		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"	\n"
+"		if (collidables[collidableIndexA].m_shapeType == SHAPE_PLANE &&\n"
+"			collidables[collidableIndexB].m_shapeType == SHAPE_CONVEX_HULL)\n"
+"		{\n"
+"			float4 posB;\n"
+"			posB = rigidBodies[bodyIndexB].m_pos;\n"
+"			Quaternion ornB;\n"
+"			ornB = rigidBodies[bodyIndexB].m_quat;\n"
+"			int contactIndex = computeContactPlaneConvex(pairIndex, bodyIndexA, bodyIndexB, collidableIndexA, collidableIndexB, \n"
+"																rigidBodies,collidables,convexShapes,vertices,indices,\n"
+"																faces,	globalContactsOut, nGlobalContactsOut,maxContactCapacity, posB,ornB);\n"
+"			if (contactIndex>=0)\n"
+"				pairs[pairIndex].z = contactIndex;\n"
+"			return;\n"
+"		}\n"
+"		if (collidables[collidableIndexA].m_shapeType == SHAPE_CONVEX_HULL &&\n"
+"			collidables[collidableIndexB].m_shapeType == SHAPE_PLANE)\n"
+"		{\n"
+"			float4 posA;\n"
+"			posA = rigidBodies[bodyIndexA].m_pos;\n"
+"			Quaternion ornA;\n"
+"			ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"			int contactIndex = computeContactPlaneConvex( pairIndex, bodyIndexB,bodyIndexA,  collidableIndexB,collidableIndexA, \n"
+"																rigidBodies,collidables,convexShapes,vertices,indices,\n"
+"																faces,	globalContactsOut, nGlobalContactsOut,maxContactCapacity,posA,ornA);\n"
+"			if (contactIndex>=0)\n"
+"				pairs[pairIndex].z = contactIndex;\n"
+"			return;\n"
+"		}\n"
+"		if (collidables[collidableIndexA].m_shapeType == SHAPE_PLANE &&\n"
+"			collidables[collidableIndexB].m_shapeType == SHAPE_SPHERE)\n"
+"		{\n"
+"			computeContactPlaneSphere(pairIndex, bodyIndexA, bodyIndexB, collidableIndexA, collidableIndexB, \n"
+"																rigidBodies,collidables,faces,	globalContactsOut, nGlobalContactsOut,maxContactCapacity);\n"
+"			return;\n"
+"		}\n"
+"		if (collidables[collidableIndexA].m_shapeType == SHAPE_SPHERE &&\n"
+"			collidables[collidableIndexB].m_shapeType == SHAPE_PLANE)\n"
+"		{\n"
+"			computeContactPlaneSphere( pairIndex, bodyIndexB,bodyIndexA,  collidableIndexB,collidableIndexA, \n"
+"																rigidBodies,collidables,\n"
+"																faces,	globalContactsOut, nGlobalContactsOut,maxContactCapacity);\n"
+"			return;\n"
+"		}\n"
+"		\n"
+"	\n"
+"		if (collidables[collidableIndexA].m_shapeType == SHAPE_SPHERE &&\n"
+"			collidables[collidableIndexB].m_shapeType == SHAPE_CONVEX_HULL)\n"
+"		{\n"
+"		\n"
+"			float4 spherePos = rigidBodies[bodyIndexA].m_pos;\n"
+"			float sphereRadius = collidables[collidableIndexA].m_radius;\n"
+"			float4 convexPos = rigidBodies[bodyIndexB].m_pos;\n"
+"			float4 convexOrn = rigidBodies[bodyIndexB].m_quat;\n"
+"			computeContactSphereConvex(pairIndex, bodyIndexA, bodyIndexB, collidableIndexA, collidableIndexB, \n"
+"																rigidBodies,collidables,convexShapes,vertices,indices,faces, globalContactsOut, nGlobalContactsOut,maxContactCapacity,\n"
+"																spherePos,sphereRadius,convexPos,convexOrn);\n"
+"			return;\n"
+"		}\n"
+"		if (collidables[collidableIndexA].m_shapeType == SHAPE_CONVEX_HULL &&\n"
+"			collidables[collidableIndexB].m_shapeType == SHAPE_SPHERE)\n"
+"		{\n"
+"		\n"
+"			float4 spherePos = rigidBodies[bodyIndexB].m_pos;\n"
+"			float sphereRadius = collidables[collidableIndexB].m_radius;\n"
+"			float4 convexPos = rigidBodies[bodyIndexA].m_pos;\n"
+"			float4 convexOrn = rigidBodies[bodyIndexA].m_quat;\n"
+"			computeContactSphereConvex(pairIndex, bodyIndexB, bodyIndexA, collidableIndexB, collidableIndexA, \n"
+"																rigidBodies,collidables,convexShapes,vertices,indices,faces, globalContactsOut, nGlobalContactsOut,maxContactCapacity,\n"
+"																spherePos,sphereRadius,convexPos,convexOrn);\n"
+"			return;\n"
+"		}\n"
+"	\n"
+"	\n"
+"	\n"
+"		\n"
+"	\n"
+"	\n"
+"		if (collidables[collidableIndexA].m_shapeType == SHAPE_SPHERE &&\n"
+"			collidables[collidableIndexB].m_shapeType == SHAPE_SPHERE)\n"
+"		{\n"
+"			//sphere-sphere\n"
+"			float radiusA = collidables[collidableIndexA].m_radius;\n"
+"			float radiusB = collidables[collidableIndexB].m_radius;\n"
+"			float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"			float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"			float4 diff = posA-posB;\n"
+"			float len = length(diff);\n"
+"			\n"
+"			///iff distance positive, don't generate a new contact\n"
+"			if ( len <= (radiusA+radiusB))\n"
+"			{\n"
+"				///distance (negative means penetration)\n"
+"				float dist = len - (radiusA+radiusB);\n"
+"				float4 normalOnSurfaceB = make_float4(1.f,0.f,0.f,0.f);\n"
+"				if (len > 0.00001)\n"
+"				{\n"
+"					normalOnSurfaceB = diff / len;\n"
+"				}\n"
+"				float4 contactPosB = posB + normalOnSurfaceB*radiusB;\n"
+"				contactPosB.w = dist;\n"
+"								\n"
+"				int dstIdx;\n"
+"				 AppendInc( nGlobalContactsOut, dstIdx );\n"
+"				\n"
+"				if (dstIdx < maxContactCapacity)\n"
+"				{\n"
+"					__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];\n"
+"					c->m_worldNormalOnB = normalOnSurfaceB;\n"
+"					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);\n"
+"					c->m_batchIdx = pairIndex;\n"
+"					int bodyA = pairs[pairIndex].x;\n"
+"					int bodyB = pairs[pairIndex].y;\n"
+"					c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;\n"
+"					c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;\n"
+"					c->m_worldPosB[0] = contactPosB;\n"
+"					c->m_childIndexA = -1;\n"
+"					c->m_childIndexB = -1;\n"
+"					GET_NPOINTS(*c) = 1;\n"
+"				}//if (dstIdx < numPairs)\n"
+"			}//if ( len <= (radiusA+radiusB))\n"
+"			return;\n"
+"		}//SHAPE_SPHERE SHAPE_SPHERE\n"
+"	}//	if (i<numPairs)\n"
+"}\n"
+"// work-in-progress\n"
+"__kernel void   processCompoundPairsPrimitivesKernel( __global const int4* gpuCompoundPairs,\n"
+"													__global const BodyData* rigidBodies, \n"
+"													__global const btCollidableGpu* collidables,\n"
+"													__global const ConvexPolyhedronCL* convexShapes, \n"
+"													__global const float4* vertices,\n"
+"													__global const float4* uniqueEdges,\n"
+"													__global const btGpuFace* faces,\n"
+"													__global const int* indices,\n"
+"													__global btAabbCL* aabbs,\n"
+"													__global const btGpuChildShape* gpuChildShapes,\n"
+"													__global struct b3Contact4Data* restrict globalContactsOut,\n"
+"													counter32_t nGlobalContactsOut,\n"
+"													int numCompoundPairs, int maxContactCapacity\n"
+"													)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i<numCompoundPairs)\n"
+"	{\n"
+"		int bodyIndexA = gpuCompoundPairs[i].x;\n"
+"		int bodyIndexB = gpuCompoundPairs[i].y;\n"
+"		int childShapeIndexA = gpuCompoundPairs[i].z;\n"
+"		int childShapeIndexB = gpuCompoundPairs[i].w;\n"
+"		\n"
+"		int collidableIndexA = -1;\n"
+"		int collidableIndexB = -1;\n"
+"		\n"
+"		float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"		float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"		\n"
+"		float4 ornB = rigidBodies[bodyIndexB].m_quat;\n"
+"		float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"							\n"
+"		if (childShapeIndexA >= 0)\n"
+"		{\n"
+"			collidableIndexA = gpuChildShapes[childShapeIndexA].m_shapeIndex;\n"
+"			float4 childPosA = gpuChildShapes[childShapeIndexA].m_childPosition;\n"
+"			float4 childOrnA = gpuChildShapes[childShapeIndexA].m_childOrientation;\n"
+"			float4 newPosA = qtRotate(ornA,childPosA)+posA;\n"
+"			float4 newOrnA = qtMul(ornA,childOrnA);\n"
+"			posA = newPosA;\n"
+"			ornA = newOrnA;\n"
+"		} else\n"
+"		{\n"
+"			collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"		}\n"
+"		\n"
+"		if (childShapeIndexB>=0)\n"
+"		{\n"
+"			collidableIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;\n"
+"			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;\n"
+"			float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;\n"
+"			float4 newPosB = transform(&childPosB,&posB,&ornB);\n"
+"			float4 newOrnB = qtMul(ornB,childOrnB);\n"
+"			posB = newPosB;\n"
+"			ornB = newOrnB;\n"
+"		} else\n"
+"		{\n"
+"			collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;	\n"
+"		}\n"
+"	\n"
+"		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"	\n"
+"		int shapeTypeA = collidables[collidableIndexA].m_shapeType;\n"
+"		int shapeTypeB = collidables[collidableIndexB].m_shapeType;\n"
+"		int pairIndex = i;\n"
+"		if ((shapeTypeA == SHAPE_PLANE) && (shapeTypeB==SHAPE_CONVEX_HULL))\n"
+"		{\n"
+"			computeContactPlaneConvex( pairIndex, bodyIndexA,bodyIndexB,  collidableIndexA,collidableIndexB, \n"
+"																rigidBodies,collidables,convexShapes,vertices,indices,\n"
+"																faces,	globalContactsOut, nGlobalContactsOut,maxContactCapacity,posB,ornB);\n"
+"			return;\n"
+"		}\n"
+"		if ((shapeTypeA == SHAPE_CONVEX_HULL) && (shapeTypeB==SHAPE_PLANE))\n"
+"		{\n"
+"			computeContactPlaneConvex( pairIndex, bodyIndexB,bodyIndexA,  collidableIndexB,collidableIndexA, \n"
+"																rigidBodies,collidables,convexShapes,vertices,indices,\n"
+"																faces,	globalContactsOut, nGlobalContactsOut,maxContactCapacity,posA,ornA);\n"
+"			return;\n"
+"		}\n"
+"		if ((shapeTypeA == SHAPE_CONVEX_HULL) && (shapeTypeB == SHAPE_SPHERE))\n"
+"		{\n"
+"			float4 spherePos = rigidBodies[bodyIndexB].m_pos;\n"
+"			float sphereRadius = collidables[collidableIndexB].m_radius;\n"
+"			float4 convexPos = posA;\n"
+"			float4 convexOrn = ornA;\n"
+"			\n"
+"			computeContactSphereConvex(pairIndex, bodyIndexB, bodyIndexA , collidableIndexB,collidableIndexA, \n"
+"										rigidBodies,collidables,convexShapes,vertices,indices,faces, globalContactsOut, nGlobalContactsOut,maxContactCapacity,\n"
+"										spherePos,sphereRadius,convexPos,convexOrn);\n"
+"	\n"
+"			return;\n"
+"		}\n"
+"		if ((shapeTypeA == SHAPE_SPHERE) && (shapeTypeB == SHAPE_CONVEX_HULL))\n"
+"		{\n"
+"			float4 spherePos = rigidBodies[bodyIndexA].m_pos;\n"
+"			float sphereRadius = collidables[collidableIndexA].m_radius;\n"
+"			float4 convexPos = posB;\n"
+"			float4 convexOrn = ornB;\n"
+"			\n"
+"			computeContactSphereConvex(pairIndex, bodyIndexA, bodyIndexB, collidableIndexA, collidableIndexB, \n"
+"										rigidBodies,collidables,convexShapes,vertices,indices,faces, globalContactsOut, nGlobalContactsOut,maxContactCapacity,\n"
+"										spherePos,sphereRadius,convexPos,convexOrn);\n"
+"	\n"
+"			return;\n"
+"		}\n"
+"	}//	if (i<numCompoundPairs)\n"
+"}\n"
+"bool pointInTriangle(const float4* vertices, const float4* normal, float4 *p )\n"
+"{\n"
+"	const float4* p1 = &vertices[0];\n"
+"	const float4* p2 = &vertices[1];\n"
+"	const float4* p3 = &vertices[2];\n"
+"	float4 edge1;	edge1 = (*p2 - *p1);\n"
+"	float4 edge2;	edge2 = ( *p3 - *p2 );\n"
+"	float4 edge3;	edge3 = ( *p1 - *p3 );\n"
+"	\n"
+"	float4 p1_to_p; p1_to_p = ( *p - *p1 );\n"
+"	float4 p2_to_p; p2_to_p = ( *p - *p2 );\n"
+"	float4 p3_to_p; p3_to_p = ( *p - *p3 );\n"
+"	float4 edge1_normal; edge1_normal = ( cross(edge1,*normal));\n"
+"	float4 edge2_normal; edge2_normal = ( cross(edge2,*normal));\n"
+"	float4 edge3_normal; edge3_normal = ( cross(edge3,*normal));\n"
+"	\n"
+"	\n"
+"	float r1, r2, r3;\n"
+"	r1 = dot(edge1_normal,p1_to_p );\n"
+"	r2 = dot(edge2_normal,p2_to_p );\n"
+"	r3 = dot(edge3_normal,p3_to_p );\n"
+"	\n"
+"	if ( r1 > 0 && r2 > 0 && r3 > 0 )\n"
+"		return true;\n"
+"    if ( r1 <= 0 && r2 <= 0 && r3 <= 0 ) \n"
+"		return true;\n"
+"	return false;\n"
+"}\n"
+"float segmentSqrDistance(float4 from, float4 to,float4 p, float4* nearest) \n"
+"{\n"
+"	float4 diff = p - from;\n"
+"	float4 v = to - from;\n"
+"	float t = dot(v,diff);\n"
+"	\n"
+"	if (t > 0) \n"
+"	{\n"
+"		float dotVV = dot(v,v);\n"
+"		if (t < dotVV) \n"
+"		{\n"
+"			t /= dotVV;\n"
+"			diff -= t*v;\n"
+"		} else \n"
+"		{\n"
+"			t = 1;\n"
+"			diff -= v;\n"
+"		}\n"
+"	} else\n"
+"	{\n"
+"		t = 0;\n"
+"	}\n"
+"	*nearest = from + t*v;\n"
+"	return dot(diff,diff);	\n"
+"}\n"
+"void	computeContactSphereTriangle(int pairIndex,\n"
+"									int bodyIndexA, int bodyIndexB,\n"
+"									int collidableIndexA, int collidableIndexB, \n"
+"									__global const BodyData* rigidBodies, \n"
+"									__global const btCollidableGpu* collidables,\n"
+"									const float4* triangleVertices,\n"
+"									__global struct b3Contact4Data* restrict globalContactsOut,\n"
+"									counter32_t nGlobalContactsOut,\n"
+"									int maxContactCapacity,\n"
+"									float4 spherePos2,\n"
+"									float radius,\n"
+"									float4 pos,\n"
+"									float4 quat,\n"
+"									int faceIndex\n"
+"									)\n"
+"{\n"
+"	float4 invPos;\n"
+"	float4 invOrn;\n"
+"	trInverse(pos,quat, &invPos,&invOrn);\n"
+"	float4 spherePos = transform(&spherePos2,&invPos,&invOrn);\n"
+"	int numFaces = 3;\n"
+"	float4 closestPnt = (float4)(0, 0, 0, 0);\n"
+"	float4 hitNormalWorld = (float4)(0, 0, 0, 0);\n"
+"	float minDist = -1000000.f;\n"
+"	bool bCollide = false;\n"
+"	\n"
+"	//////////////////////////////////////\n"
+"	float4 sphereCenter;\n"
+"	sphereCenter = spherePos;\n"
+"	const float4* vertices = triangleVertices;\n"
+"	float contactBreakingThreshold = 0.f;//todo?\n"
+"	float radiusWithThreshold = radius + contactBreakingThreshold;\n"
+"	float4 edge10;\n"
+"	edge10 = vertices[1]-vertices[0];\n"
+"	edge10.w = 0.f;//is this needed?\n"
+"	float4 edge20;\n"
+"	edge20 = vertices[2]-vertices[0];\n"
+"	edge20.w = 0.f;//is this needed?\n"
+"	float4 normal = cross3(edge10,edge20);\n"
+"	normal = normalize(normal);\n"
+"	float4 p1ToCenter;\n"
+"	p1ToCenter = sphereCenter - vertices[0];\n"
+"	\n"
+"	float distanceFromPlane = dot(p1ToCenter,normal);\n"
+"	if (distanceFromPlane < 0.f)\n"
+"	{\n"
+"		//triangle facing the other way\n"
+"		distanceFromPlane *= -1.f;\n"
+"		normal *= -1.f;\n"
+"	}\n"
+"	hitNormalWorld = normal;\n"
+"	bool isInsideContactPlane = distanceFromPlane < radiusWithThreshold;\n"
+"	\n"
+"	// Check for contact / intersection\n"
+"	bool hasContact = false;\n"
+"	float4 contactPoint;\n"
+"	if (isInsideContactPlane) \n"
+"	{\n"
+"	\n"
+"		if (pointInTriangle(vertices,&normal, &sphereCenter)) \n"
+"		{\n"
+"			// Inside the contact wedge - touches a point on the shell plane\n"
+"			hasContact = true;\n"
+"			contactPoint = sphereCenter - normal*distanceFromPlane;\n"
+"			\n"
+"		} else {\n"
+"			// Could be inside one of the contact capsules\n"
+"			float contactCapsuleRadiusSqr = radiusWithThreshold*radiusWithThreshold;\n"
+"			float4 nearestOnEdge;\n"
+"			int numEdges = 3;\n"
+"			for (int i = 0; i < numEdges; i++) \n"
+"			{\n"
+"				float4 pa =vertices[i];\n"
+"				float4 pb = vertices[(i+1)%3];\n"
+"				float distanceSqr = segmentSqrDistance(pa,pb,sphereCenter, &nearestOnEdge);\n"
+"				if (distanceSqr < contactCapsuleRadiusSqr) \n"
+"				{\n"
+"					// Yep, we're inside a capsule\n"
+"					hasContact = true;\n"
+"					contactPoint = nearestOnEdge;\n"
+"					\n"
+"				}\n"
+"				\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	if (hasContact) \n"
+"	{\n"
+"		closestPnt = contactPoint;\n"
+"		float4 contactToCenter = sphereCenter - contactPoint;\n"
+"		minDist = length(contactToCenter);\n"
+"		if (minDist>FLT_EPSILON)\n"
+"		{\n"
+"			hitNormalWorld = normalize(contactToCenter);//*(1./minDist);\n"
+"			bCollide  = true;\n"
+"		}\n"
+"		\n"
+"	}\n"
+"	/////////////////////////////////////\n"
+"	if (bCollide && minDist > -10000)\n"
+"	{\n"
+"		\n"
+"		float4 normalOnSurfaceB1 = qtRotate(quat,-hitNormalWorld);\n"
+"		float4 pOnB1 = transform(&closestPnt,&pos,&quat);\n"
+"		float actualDepth = minDist-radius;\n"
+"		\n"
+"		if (actualDepth<=0.f)\n"
+"		{\n"
+"			pOnB1.w = actualDepth;\n"
+"			int dstIdx;\n"
+"			\n"
+"			float lenSqr = dot3F4(normalOnSurfaceB1,normalOnSurfaceB1);\n"
+"			if (lenSqr>FLT_EPSILON)\n"
+"			{\n"
+"				AppendInc( nGlobalContactsOut, dstIdx );\n"
+"			\n"
+"				if (dstIdx < maxContactCapacity)\n"
+"				{\n"
+"					__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];\n"
+"					c->m_worldNormalOnB = -normalOnSurfaceB1;\n"
+"					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);\n"
+"					c->m_batchIdx = pairIndex;\n"
+"					c->m_bodyAPtrAndSignBit = rigidBodies[bodyIndexA].m_invMass==0?-bodyIndexA:bodyIndexA;\n"
+"					c->m_bodyBPtrAndSignBit = rigidBodies[bodyIndexB].m_invMass==0?-bodyIndexB:bodyIndexB;\n"
+"					c->m_worldPosB[0] = pOnB1;\n"
+"					c->m_childIndexA = -1;\n"
+"					c->m_childIndexB = faceIndex;\n"
+"					GET_NPOINTS(*c) = 1;\n"
+"				} \n"
+"			}\n"
+"		}\n"
+"	}//if (hasCollision)\n"
+"}\n"
+"// work-in-progress\n"
+"__kernel void   findConcaveSphereContactsKernel( __global int4* concavePairs,\n"
+"												__global const BodyData* rigidBodies,\n"
+"												__global const btCollidableGpu* collidables,\n"
+"												__global const ConvexPolyhedronCL* convexShapes, \n"
+"												__global const float4* vertices,\n"
+"												__global const float4* uniqueEdges,\n"
+"												__global const btGpuFace* faces,\n"
+"												__global const int* indices,\n"
+"												__global btAabbCL* aabbs,\n"
+"												__global struct b3Contact4Data* restrict globalContactsOut,\n"
+"												counter32_t nGlobalContactsOut,\n"
+"													int numConcavePairs, int maxContactCapacity\n"
+"												)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numConcavePairs)\n"
+"		return;\n"
+"	int pairIdx = i;\n"
+"	int bodyIndexA = concavePairs[i].x;\n"
+"	int bodyIndexB = concavePairs[i].y;\n"
+"	int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"	int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"	int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"	int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"	if (collidables[collidableIndexB].m_shapeType==SHAPE_SPHERE)\n"
+"	{\n"
+"		int f = concavePairs[i].z;\n"
+"		btGpuFace face = faces[convexShapes[shapeIndexA].m_faceOffset+f];\n"
+"		\n"
+"		float4 verticesA[3];\n"
+"		for (int i=0;i<3;i++)\n"
+"		{\n"
+"			int index = indices[face.m_indexOffset+i];\n"
+"			float4 vert = vertices[convexShapes[shapeIndexA].m_vertexOffset+index];\n"
+"			verticesA[i] = vert;\n"
+"		}\n"
+"		float4 spherePos = rigidBodies[bodyIndexB].m_pos;\n"
+"		float sphereRadius = collidables[collidableIndexB].m_radius;\n"
+"		float4 convexPos = rigidBodies[bodyIndexA].m_pos;\n"
+"		float4 convexOrn = rigidBodies[bodyIndexA].m_quat;\n"
+"		computeContactSphereTriangle(i, bodyIndexB, bodyIndexA, collidableIndexB, collidableIndexA, \n"
+"																rigidBodies,collidables,\n"
+"																verticesA,\n"
+"																globalContactsOut, nGlobalContactsOut,maxContactCapacity,\n"
+"																spherePos,sphereRadius,convexPos,convexOrn, f);\n"
+"		return;\n"
+"	}\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/sat.cl b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/sat.cl
new file mode 100644
index 0000000000..a6565fd6fa
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/sat.cl
@@ -0,0 +1,2018 @@
+//keep this enum in sync with the CPU version (in btCollidable.h)
+//written by Erwin Coumans
+
+
+#define SHAPE_CONVEX_HULL 3
+#define SHAPE_CONCAVE_TRIMESH 5
+#define TRIANGLE_NUM_CONVEX_FACES 5
+#define SHAPE_COMPOUND_OF_CONVEX_HULLS 6
+
+#define B3_MAX_STACK_DEPTH 256
+
+
+typedef unsigned int u32;
+
+///keep this in sync with btCollidable.h
+typedef struct
+{
+	union {
+		int m_numChildShapes;
+		int m_bvhIndex;
+	};
+	union
+	{
+		float m_radius;
+		int	m_compoundBvhIndex;
+	};
+	
+	int m_shapeType;
+	int m_shapeIndex;
+	
+} btCollidableGpu;
+
+#define MAX_NUM_PARTS_IN_BITS 10
+
+///b3QuantizedBvhNode is a compressed aabb node, 16 bytes.
+///Node can be used for leafnode or internal node. Leafnodes can point to 32-bit triangle index (non-negative range).
+typedef struct
+{
+	//12 bytes
+	unsigned short int	m_quantizedAabbMin[3];
+	unsigned short int	m_quantizedAabbMax[3];
+	//4 bytes
+	int	m_escapeIndexOrTriangleIndex;
+} b3QuantizedBvhNode;
+
+typedef struct
+{
+	float4		m_aabbMin;
+	float4		m_aabbMax;
+	float4		m_quantization;
+	int			m_numNodes;
+	int			m_numSubTrees;
+	int			m_nodeOffset;
+	int			m_subTreeOffset;
+
+} b3BvhInfo;
+
+
+int	getTriangleIndex(const b3QuantizedBvhNode* rootNode)
+{
+	unsigned int x=0;
+	unsigned int y = (~(x&0))<<(31-MAX_NUM_PARTS_IN_BITS);
+	// Get only the lower bits where the triangle index is stored
+	return (rootNode->m_escapeIndexOrTriangleIndex&~(y));
+}
+
+int	getTriangleIndexGlobal(__global const b3QuantizedBvhNode* rootNode)
+{
+	unsigned int x=0;
+	unsigned int y = (~(x&0))<<(31-MAX_NUM_PARTS_IN_BITS);
+	// Get only the lower bits where the triangle index is stored
+	return (rootNode->m_escapeIndexOrTriangleIndex&~(y));
+}
+
+int isLeafNode(const b3QuantizedBvhNode* rootNode)
+{
+	//skipindex is negative (internal node), triangleindex >=0 (leafnode)
+	return (rootNode->m_escapeIndexOrTriangleIndex >= 0)? 1 : 0;
+}
+
+int isLeafNodeGlobal(__global const b3QuantizedBvhNode* rootNode)
+{
+	//skipindex is negative (internal node), triangleindex >=0 (leafnode)
+	return (rootNode->m_escapeIndexOrTriangleIndex >= 0)? 1 : 0;
+}
+	
+int getEscapeIndex(const b3QuantizedBvhNode* rootNode)
+{
+	return -rootNode->m_escapeIndexOrTriangleIndex;
+}
+
+int getEscapeIndexGlobal(__global const b3QuantizedBvhNode* rootNode)
+{
+	return -rootNode->m_escapeIndexOrTriangleIndex;
+}
+
+
+typedef struct
+{
+	//12 bytes
+	unsigned short int	m_quantizedAabbMin[3];
+	unsigned short int	m_quantizedAabbMax[3];
+	//4 bytes, points to the root of the subtree
+	int			m_rootNodeIndex;
+	//4 bytes
+	int			m_subtreeSize;
+	int			m_padding[3];
+} b3BvhSubtreeInfo;
+
+
+
+
+
+
+
+typedef struct
+{
+	float4	m_childPosition;
+	float4	m_childOrientation;
+	int m_shapeIndex;
+	int m_unused0;
+	int m_unused1;
+	int m_unused2;
+} btGpuChildShape;
+
+
+typedef struct
+{
+	float4 m_pos;
+	float4 m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_collidableIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} BodyData;
+
+
+typedef struct  
+{
+	float4		m_localCenter;
+	float4		m_extents;
+	float4		mC;
+	float4		mE;
+	
+	float			m_radius;
+	int	m_faceOffset;
+	int m_numFaces;
+	int	m_numVertices;
+
+	int m_vertexOffset;
+	int	m_uniqueEdgesOffset;
+	int	m_numUniqueEdges;
+	int m_unused;
+} ConvexPolyhedronCL;
+
+typedef struct 
+{
+	union
+	{
+		float4	m_min;
+		float   m_minElems[4];
+		int			m_minIndices[4];
+	};
+	union
+	{
+		float4	m_max;
+		float   m_maxElems[4];
+		int			m_maxIndices[4];
+	};
+} btAabbCL;
+
+#include "Bullet3Collision/BroadPhaseCollision/shared/b3Aabb.h"
+#include "Bullet3Common/shared/b3Int2.h"
+
+
+
+typedef struct
+{
+	float4 m_plane;
+	int m_indexOffset;
+	int m_numIndices;
+} btGpuFace;
+
+#define make_float4 (float4)
+
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+
+	
+//	float4 a1 = make_float4(a.xyz,0.f);
+//	float4 b1 = make_float4(b.xyz,0.f);
+
+//	return cross(a1,b1);
+
+//float4 c = make_float4(a.y*b.z - a.z*b.y,a.z*b.x - a.x*b.z,a.x*b.y - a.y*b.x,0.f);
+	
+	//	float4 c = make_float4(a.y*b.z - a.z*b.y,1.f,a.x*b.y - a.y*b.x,0.f);
+	
+	//return c;
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = make_float4(a.xyz,0.f);
+	float4 b1 = make_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	v = make_float4(v.xyz,0.f);
+	return fast_normalize(v);
+}
+
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b);
+
+__inline
+Quaternion qtNormalize(Quaternion in);
+
+__inline
+float4 qtRotate(Quaternion q, float4 vec);
+
+__inline
+Quaternion qtInvert(Quaternion q);
+
+
+
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross3( a, b );
+	ans += a.w*b+b.w*a;
+//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+Quaternion qtNormalize(Quaternion in)
+{
+	return fastNormalize4(in);
+//	in /= length( in );
+//	return in;
+}
+__inline
+float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(qtMul(q,vcpy),qInv);
+	return out;
+}
+
+__inline
+Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+__inline
+float4 qtInvRotate(const Quaternion q, float4 vec)
+{
+	return qtRotate( qtInvert( q ), vec );
+}
+
+__inline
+float4 transform(const float4* p, const float4* translation, const Quaternion* orientation)
+{
+	return qtRotate( *orientation, *p ) + (*translation);
+}
+
+
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = make_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+}
+
+inline void projectLocal(const ConvexPolyhedronCL* hull,  const float4 pos, const float4 orn, 
+const float4* dir, const float4* vertices, float* min, float* max)
+{
+	min[0] = FLT_MAX;
+	max[0] = -FLT_MAX;
+	int numVerts = hull->m_numVertices;
+
+	const float4 localDir = qtInvRotate(orn,*dir);
+	float offset = dot(pos,*dir);
+	for(int i=0;i<numVerts;i++)
+	{
+		float dp = dot(vertices[hull->m_vertexOffset+i],localDir);
+		if(dp < min[0])	
+			min[0] = dp;
+		if(dp > max[0])	
+			max[0] = dp;
+	}
+	if(min[0]>max[0])
+	{
+		float tmp = min[0];
+		min[0] = max[0];
+		max[0] = tmp;
+	}
+	min[0] += offset;
+	max[0] += offset;
+}
+
+inline void project(__global const ConvexPolyhedronCL* hull,  const float4 pos, const float4 orn, 
+const float4* dir, __global const float4* vertices, float* min, float* max)
+{
+	min[0] = FLT_MAX;
+	max[0] = -FLT_MAX;
+	int numVerts = hull->m_numVertices;
+
+	const float4 localDir = qtInvRotate(orn,*dir);
+	float offset = dot(pos,*dir);
+	for(int i=0;i<numVerts;i++)
+	{
+		float dp = dot(vertices[hull->m_vertexOffset+i],localDir);
+		if(dp < min[0])	
+			min[0] = dp;
+		if(dp > max[0])	
+			max[0] = dp;
+	}
+	if(min[0]>max[0])
+	{
+		float tmp = min[0];
+		min[0] = max[0];
+		max[0] = tmp;
+	}
+	min[0] += offset;
+	max[0] += offset;
+}
+
+inline bool TestSepAxisLocalA(const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, 
+	const float4 posA,const float4 ornA,
+	const float4 posB,const float4 ornB,
+	float4* sep_axis, const float4* verticesA, __global const float4* verticesB,float* depth)
+{
+	float Min0,Max0;
+	float Min1,Max1;
+	projectLocal(hullA,posA,ornA,sep_axis,verticesA, &Min0, &Max0);
+	project(hullB,posB,ornB, sep_axis,verticesB, &Min1, &Max1);
+
+	if(Max0<Min1 || Max1<Min0)
+		return false;
+
+	float d0 = Max0 - Min1;
+	float d1 = Max1 - Min0;
+	*depth = d0<d1 ? d0:d1;
+	return true;
+}
+
+
+
+
+inline bool IsAlmostZero(const float4 v)
+{
+	if(fabs(v.x)>1e-6f || fabs(v.y)>1e-6f || fabs(v.z)>1e-6f)
+		return false;
+	return true;
+}
+
+
+
+bool findSeparatingAxisLocalA(	const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, 
+	const float4 posA1,
+	const float4 ornA,
+	const float4 posB1,
+	const float4 ornB,
+	const float4 DeltaC2,
+	
+	const float4* verticesA, 
+	const float4* uniqueEdgesA, 
+	const btGpuFace* facesA,
+	const int*  indicesA,
+
+	__global const float4* verticesB, 
+	__global const float4* uniqueEdgesB, 
+	__global const btGpuFace* facesB,
+	__global const int*  indicesB,
+	float4* sep,
+	float* dmin)
+{
+	
+
+	float4 posA = posA1;
+	posA.w = 0.f;
+	float4 posB = posB1;
+	posB.w = 0.f;
+	int curPlaneTests=0;
+	{
+		int numFacesA = hullA->m_numFaces;
+		// Test normals from hullA
+		for(int i=0;i<numFacesA;i++)
+		{
+			const float4 normal = facesA[hullA->m_faceOffset+i].m_plane;
+			float4 faceANormalWS = qtRotate(ornA,normal);
+			if (dot3F4(DeltaC2,faceANormalWS)<0)
+				faceANormalWS*=-1.f;
+			curPlaneTests++;
+			float d;
+			if(!TestSepAxisLocalA( hullA, hullB, posA,ornA,posB,ornB,&faceANormalWS, verticesA, verticesB,&d))
+				return false;
+			if(d<*dmin)
+			{
+				*dmin = d;
+				*sep = faceANormalWS;
+			}
+		}
+	}
+	if((dot3F4(-DeltaC2,*sep))>0.0f)
+	{
+		*sep = -(*sep);
+	}
+	return true;
+}
+
+bool findSeparatingAxisLocalB(	__global const ConvexPolyhedronCL* hullA,  const ConvexPolyhedronCL* hullB, 
+	const float4 posA1,
+	const float4 ornA,
+	const float4 posB1,
+	const float4 ornB,
+	const float4 DeltaC2,
+	__global const float4* verticesA, 
+	__global const float4* uniqueEdgesA, 
+	__global const btGpuFace* facesA,
+	__global const int*  indicesA,
+	const float4* verticesB,
+	const float4* uniqueEdgesB, 
+	const btGpuFace* facesB,
+	const int*  indicesB,
+	float4* sep,
+	float* dmin)
+{
+
+
+	float4 posA = posA1;
+	posA.w = 0.f;
+	float4 posB = posB1;
+	posB.w = 0.f;
+	int curPlaneTests=0;
+	{
+		int numFacesA = hullA->m_numFaces;
+		// Test normals from hullA
+		for(int i=0;i<numFacesA;i++)
+		{
+			const float4 normal = facesA[hullA->m_faceOffset+i].m_plane;
+			float4 faceANormalWS = qtRotate(ornA,normal);
+			if (dot3F4(DeltaC2,faceANormalWS)<0)
+				faceANormalWS *= -1.f;
+			curPlaneTests++;
+			float d;
+			if(!TestSepAxisLocalA( hullB, hullA, posB,ornB,posA,ornA, &faceANormalWS, verticesB,verticesA, &d))
+				return false;
+			if(d<*dmin)
+			{
+				*dmin = d;
+				*sep = faceANormalWS;
+			}
+		}
+	}
+	if((dot3F4(-DeltaC2,*sep))>0.0f)
+	{
+		*sep = -(*sep);
+	}
+	return true;
+}
+
+
+
+bool findSeparatingAxisEdgeEdgeLocalA(	const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, 
+	const float4 posA1,
+	const float4 ornA,
+	const float4 posB1,
+	const float4 ornB,
+	const float4 DeltaC2,
+	const float4* verticesA, 
+	const float4* uniqueEdgesA, 
+	const btGpuFace* facesA,
+	const int*  indicesA,
+	__global const float4* verticesB, 
+	__global const float4* uniqueEdgesB, 
+	__global const btGpuFace* facesB,
+	__global const int*  indicesB,
+		float4* sep,
+	float* dmin)
+{
+
+
+	float4 posA = posA1;
+	posA.w = 0.f;
+	float4 posB = posB1;
+	posB.w = 0.f;
+
+	int curPlaneTests=0;
+
+	int curEdgeEdge = 0;
+	// Test edges
+	for(int e0=0;e0<hullA->m_numUniqueEdges;e0++)
+	{
+		const float4 edge0 = uniqueEdgesA[hullA->m_uniqueEdgesOffset+e0];
+		float4 edge0World = qtRotate(ornA,edge0);
+
+		for(int e1=0;e1<hullB->m_numUniqueEdges;e1++)
+		{
+			const float4 edge1 = uniqueEdgesB[hullB->m_uniqueEdgesOffset+e1];
+			float4 edge1World = qtRotate(ornB,edge1);
+
+
+			float4 crossje = cross3(edge0World,edge1World);
+
+			curEdgeEdge++;
+			if(!IsAlmostZero(crossje))
+			{
+				crossje = normalize3(crossje);
+				if (dot3F4(DeltaC2,crossje)<0)
+					crossje *= -1.f;
+
+				float dist;
+				bool result = true;
+				{
+					float Min0,Max0;
+					float Min1,Max1;
+					projectLocal(hullA,posA,ornA,&crossje,verticesA, &Min0, &Max0);
+					project(hullB,posB,ornB,&crossje,verticesB, &Min1, &Max1);
+				
+					if(Max0<Min1 || Max1<Min0)
+						result = false;
+				
+					float d0 = Max0 - Min1;
+					float d1 = Max1 - Min0;
+					dist = d0<d1 ? d0:d1;
+					result = true;
+
+				}
+				
+
+				if(dist<*dmin)
+				{
+					*dmin = dist;
+					*sep = crossje;
+				}
+			}
+		}
+
+	}
+
+	
+	if((dot3F4(-DeltaC2,*sep))>0.0f)
+	{
+		*sep = -(*sep);
+	}
+	return true;
+}
+
+
+inline bool TestSepAxis(__global const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, 
+	const float4 posA,const float4 ornA,
+	const float4 posB,const float4 ornB,
+	float4* sep_axis, __global const float4* vertices,float* depth)
+{
+	float Min0,Max0;
+	float Min1,Max1;
+	project(hullA,posA,ornA,sep_axis,vertices, &Min0, &Max0);
+	project(hullB,posB,ornB, sep_axis,vertices, &Min1, &Max1);
+
+	if(Max0<Min1 || Max1<Min0)
+		return false;
+
+	float d0 = Max0 - Min1;
+	float d1 = Max1 - Min0;
+	*depth = d0<d1 ? d0:d1;
+	return true;
+}
+
+
+bool findSeparatingAxis(	__global const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, 
+	const float4 posA1,
+	const float4 ornA,
+	const float4 posB1,
+	const float4 ornB,
+	const float4 DeltaC2,
+	__global const float4* vertices, 
+	__global const float4* uniqueEdges, 
+	__global const btGpuFace* faces,
+	__global const int*  indices,
+	float4* sep,
+	float* dmin)
+{
+	
+
+	float4 posA = posA1;
+	posA.w = 0.f;
+	float4 posB = posB1;
+	posB.w = 0.f;
+	
+	int curPlaneTests=0;
+
+	{
+		int numFacesA = hullA->m_numFaces;
+		// Test normals from hullA
+		for(int i=0;i<numFacesA;i++)
+		{
+			const float4 normal = faces[hullA->m_faceOffset+i].m_plane;
+			float4 faceANormalWS = qtRotate(ornA,normal);
+	
+			if (dot3F4(DeltaC2,faceANormalWS)<0)
+				faceANormalWS*=-1.f;
+				
+			curPlaneTests++;
+	
+			float d;
+			if(!TestSepAxis( hullA, hullB, posA,ornA,posB,ornB,&faceANormalWS, vertices,&d))
+				return false;
+	
+			if(d<*dmin)
+			{
+				*dmin = d;
+				*sep = faceANormalWS;
+			}
+		}
+	}
+
+
+		if((dot3F4(-DeltaC2,*sep))>0.0f)
+		{
+			*sep = -(*sep);
+		}
+	
+	return true;
+}
+
+
+
+
+bool findSeparatingAxisUnitSphere(	__global const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, 
+	const float4 posA1,
+	const float4 ornA,
+	const float4 posB1,
+	const float4 ornB,
+	const float4 DeltaC2,
+	__global const float4* vertices,
+	__global const float4* unitSphereDirections,
+	int numUnitSphereDirections,
+	float4* sep,
+	float* dmin)
+{
+	
+	float4 posA = posA1;
+	posA.w = 0.f;
+	float4 posB = posB1;
+	posB.w = 0.f;
+
+	int curPlaneTests=0;
+
+	int curEdgeEdge = 0;
+	// Test unit sphere directions
+	for (int i=0;i<numUnitSphereDirections;i++)
+	{
+
+		float4 crossje;
+		crossje = unitSphereDirections[i];	
+
+		if (dot3F4(DeltaC2,crossje)>0)
+			crossje *= -1.f;
+		{
+			float dist;
+			bool result = true;
+			float Min0,Max0;
+			float Min1,Max1;
+			project(hullA,posA,ornA,&crossje,vertices, &Min0, &Max0);
+			project(hullB,posB,ornB,&crossje,vertices, &Min1, &Max1);
+		
+			if(Max0<Min1 || Max1<Min0)
+				return false;
+		
+			float d0 = Max0 - Min1;
+			float d1 = Max1 - Min0;
+			dist = d0<d1 ? d0:d1;
+			result = true;
+	
+			if(dist<*dmin)
+			{
+				*dmin = dist;
+				*sep = crossje;
+			}
+		}
+	}
+
+	
+	if((dot3F4(-DeltaC2,*sep))>0.0f)
+	{
+		*sep = -(*sep);
+	}
+	return true;
+}
+
+
+bool findSeparatingAxisEdgeEdge(	__global const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, 
+	const float4 posA1,
+	const float4 ornA,
+	const float4 posB1,
+	const float4 ornB,
+	const float4 DeltaC2,
+	__global const float4* vertices, 
+	__global const float4* uniqueEdges, 
+	__global const btGpuFace* faces,
+	__global const int*  indices,
+	float4* sep,
+	float* dmin)
+{
+	
+
+	float4 posA = posA1;
+	posA.w = 0.f;
+	float4 posB = posB1;
+	posB.w = 0.f;
+
+	int curPlaneTests=0;
+
+	int curEdgeEdge = 0;
+	// Test edges
+	for(int e0=0;e0<hullA->m_numUniqueEdges;e0++)
+	{
+		const float4 edge0 = uniqueEdges[hullA->m_uniqueEdgesOffset+e0];
+		float4 edge0World = qtRotate(ornA,edge0);
+
+		for(int e1=0;e1<hullB->m_numUniqueEdges;e1++)
+		{
+			const float4 edge1 = uniqueEdges[hullB->m_uniqueEdgesOffset+e1];
+			float4 edge1World = qtRotate(ornB,edge1);
+
+
+			float4 crossje = cross3(edge0World,edge1World);
+
+			curEdgeEdge++;
+			if(!IsAlmostZero(crossje))
+			{
+				crossje = normalize3(crossje);
+				if (dot3F4(DeltaC2,crossje)<0)
+					crossje*=-1.f;
+					
+				float dist;
+				bool result = true;
+				{
+					float Min0,Max0;
+					float Min1,Max1;
+					project(hullA,posA,ornA,&crossje,vertices, &Min0, &Max0);
+					project(hullB,posB,ornB,&crossje,vertices, &Min1, &Max1);
+				
+					if(Max0<Min1 || Max1<Min0)
+						return false;
+				
+					float d0 = Max0 - Min1;
+					float d1 = Max1 - Min0;
+					dist = d0<d1 ? d0:d1;
+					result = true;
+
+				}
+				
+
+				if(dist<*dmin)
+				{
+					*dmin = dist;
+					*sep = crossje;
+				}
+			}
+		}
+
+	}
+
+	
+	if((dot3F4(-DeltaC2,*sep))>0.0f)
+	{
+		*sep = -(*sep);
+	}
+	return true;
+}
+
+
+// work-in-progress
+__kernel void   processCompoundPairsKernel( __global const int4* gpuCompoundPairs,
+																					__global const BodyData* rigidBodies, 
+																					__global const btCollidableGpu* collidables,
+																					__global const ConvexPolyhedronCL* convexShapes, 
+																					__global const float4* vertices,
+																					__global const float4* uniqueEdges,
+																					__global const btGpuFace* faces,
+																					__global const int* indices,
+																					__global btAabbCL* aabbs,
+																					__global const btGpuChildShape* gpuChildShapes,
+																					__global volatile float4* gpuCompoundSepNormalsOut,
+																					__global volatile int* gpuHasCompoundSepNormalsOut,
+																					int numCompoundPairs
+																					)
+{
+
+	int i = get_global_id(0);
+	if (i<numCompoundPairs)
+	{
+		int bodyIndexA = gpuCompoundPairs[i].x;
+		int bodyIndexB = gpuCompoundPairs[i].y;
+
+		int childShapeIndexA = gpuCompoundPairs[i].z;
+		int childShapeIndexB = gpuCompoundPairs[i].w;
+		
+		int collidableIndexA = -1;
+		int collidableIndexB = -1;
+		
+		float4 ornA = rigidBodies[bodyIndexA].m_quat;
+		float4 posA = rigidBodies[bodyIndexA].m_pos;
+		
+		float4 ornB = rigidBodies[bodyIndexB].m_quat;
+		float4 posB = rigidBodies[bodyIndexB].m_pos;
+							
+		if (childShapeIndexA >= 0)
+		{
+			collidableIndexA = gpuChildShapes[childShapeIndexA].m_shapeIndex;
+			float4 childPosA = gpuChildShapes[childShapeIndexA].m_childPosition;
+			float4 childOrnA = gpuChildShapes[childShapeIndexA].m_childOrientation;
+			float4 newPosA = qtRotate(ornA,childPosA)+posA;
+			float4 newOrnA = qtMul(ornA,childOrnA);
+			posA = newPosA;
+			ornA = newOrnA;
+		} else
+		{
+			collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+		}
+		
+		if (childShapeIndexB>=0)
+		{
+			collidableIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+			float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+			float4 newPosB = transform(&childPosB,&posB,&ornB);
+			float4 newOrnB = qtMul(ornB,childOrnB);
+			posB = newPosB;
+			ornB = newOrnB;
+		} else
+		{
+			collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;	
+		}
+	
+		gpuHasCompoundSepNormalsOut[i] = 0;
+	
+		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+	
+		int shapeTypeA = collidables[collidableIndexA].m_shapeType;
+		int shapeTypeB = collidables[collidableIndexB].m_shapeType;
+	
+
+		if ((shapeTypeA != SHAPE_CONVEX_HULL) || (shapeTypeB != SHAPE_CONVEX_HULL))
+		{
+			return;
+		}
+
+		int hasSeparatingAxis = 5;
+							
+		int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+		float dmin = FLT_MAX;
+		posA.w = 0.f;
+		posB.w = 0.f;
+		float4 c0local = convexShapes[shapeIndexA].m_localCenter;
+		float4 c0 = transform(&c0local, &posA, &ornA);
+		float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+		float4 c1 = transform(&c1local,&posB,&ornB);
+		const float4 DeltaC2 = c0 - c1;
+		float4 sepNormal = make_float4(1,0,0,0);
+		bool sepA = findSeparatingAxis(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,posB,ornB,DeltaC2,vertices,uniqueEdges,faces,indices,&sepNormal,&dmin);
+		hasSeparatingAxis = 4;
+		if (!sepA)
+		{
+			hasSeparatingAxis = 0;
+		} else
+		{
+			bool sepB = findSeparatingAxis(	&convexShapes[shapeIndexB],&convexShapes[shapeIndexA],posB,ornB,posA,ornA,DeltaC2,vertices,uniqueEdges,faces,indices,&sepNormal,&dmin);
+
+			if (!sepB)
+			{
+				hasSeparatingAxis = 0;
+			} else//(!sepB)
+			{
+				bool sepEE = findSeparatingAxisEdgeEdge(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,posB,ornB,DeltaC2,vertices,uniqueEdges,faces,indices,&sepNormal,&dmin);
+				if (sepEE)
+				{
+						gpuCompoundSepNormalsOut[i] = sepNormal;//fastNormalize4(sepNormal);
+						gpuHasCompoundSepNormalsOut[i] = 1;
+				}//sepEE
+			}//(!sepB)
+		}//(!sepA)
+		
+		
+	}
+		
+}
+
+
+inline b3Float4 MyUnQuantize(const unsigned short* vecIn, b3Float4 quantization, b3Float4 bvhAabbMin)
+{
+		b3Float4 vecOut;
+		vecOut = b3MakeFloat4(
+			(float)(vecIn[0]) / (quantization.x),
+			(float)(vecIn[1]) / (quantization.y),
+			(float)(vecIn[2]) / (quantization.z),
+			0.f);
+
+		vecOut += bvhAabbMin;
+		return vecOut;
+}
+
+inline b3Float4 MyUnQuantizeGlobal(__global const unsigned short* vecIn, b3Float4 quantization, b3Float4 bvhAabbMin)
+{
+		b3Float4 vecOut;
+		vecOut = b3MakeFloat4(
+			(float)(vecIn[0]) / (quantization.x),
+			(float)(vecIn[1]) / (quantization.y),
+			(float)(vecIn[2]) / (quantization.z),
+			0.f);
+
+		vecOut += bvhAabbMin;
+		return vecOut;
+}
+
+
+// work-in-progress
+__kernel void   findCompoundPairsKernel( __global const int4* pairs, 
+	__global const BodyData* rigidBodies, 
+	__global const btCollidableGpu* collidables,
+	__global const ConvexPolyhedronCL* convexShapes, 
+	__global const float4* vertices,
+	__global const float4* uniqueEdges,
+	__global const btGpuFace* faces,
+	__global const int* indices,
+	__global b3Aabb_t* aabbLocalSpace,
+	__global const btGpuChildShape* gpuChildShapes,
+	__global volatile int4* gpuCompoundPairsOut,
+	__global volatile int* numCompoundPairsOut,
+	__global const b3BvhSubtreeInfo* subtrees,
+	__global const b3QuantizedBvhNode* quantizedNodes,
+	__global const b3BvhInfo* bvhInfos,
+	int numPairs,
+	int maxNumCompoundPairsCapacity
+	)
+{
+
+	int i = get_global_id(0);
+
+	if (i<numPairs)
+	{
+		int bodyIndexA = pairs[i].x;
+		int bodyIndexB = pairs[i].y;
+
+		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+
+		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+
+
+		//once the broadphase avoids static-static pairs, we can remove this test
+		if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))
+		{
+			return;
+		}
+
+		if ((collidables[collidableIndexA].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS) &&(collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS))
+		{
+			int bvhA = collidables[collidableIndexA].m_compoundBvhIndex;
+			int bvhB = collidables[collidableIndexB].m_compoundBvhIndex;
+			int numSubTreesA = bvhInfos[bvhA].m_numSubTrees;
+			int subTreesOffsetA = bvhInfos[bvhA].m_subTreeOffset;
+			int subTreesOffsetB = bvhInfos[bvhB].m_subTreeOffset;
+
+
+			int numSubTreesB = bvhInfos[bvhB].m_numSubTrees;
+			
+			float4 posA = rigidBodies[bodyIndexA].m_pos;
+			b3Quat ornA = rigidBodies[bodyIndexA].m_quat;
+
+			b3Quat ornB = rigidBodies[bodyIndexB].m_quat;
+			float4 posB = rigidBodies[bodyIndexB].m_pos;
+
+			
+			for (int p=0;p<numSubTreesA;p++)
+			{
+				b3BvhSubtreeInfo subtreeA = subtrees[subTreesOffsetA+p];
+				//bvhInfos[bvhA].m_quantization
+				b3Float4 treeAminLocal = MyUnQuantize(subtreeA.m_quantizedAabbMin,bvhInfos[bvhA].m_quantization,bvhInfos[bvhA].m_aabbMin);
+				b3Float4 treeAmaxLocal = MyUnQuantize(subtreeA.m_quantizedAabbMax,bvhInfos[bvhA].m_quantization,bvhInfos[bvhA].m_aabbMin);
+
+				b3Float4 aabbAMinOut,aabbAMaxOut;
+				float margin=0.f;
+				b3TransformAabb2(treeAminLocal,treeAmaxLocal, margin,posA,ornA,&aabbAMinOut,&aabbAMaxOut);
+				
+				for (int q=0;q<numSubTreesB;q++)
+				{
+					b3BvhSubtreeInfo subtreeB = subtrees[subTreesOffsetB+q];
+
+					b3Float4 treeBminLocal = MyUnQuantize(subtreeB.m_quantizedAabbMin,bvhInfos[bvhB].m_quantization,bvhInfos[bvhB].m_aabbMin);
+					b3Float4 treeBmaxLocal = MyUnQuantize(subtreeB.m_quantizedAabbMax,bvhInfos[bvhB].m_quantization,bvhInfos[bvhB].m_aabbMin);
+
+					b3Float4 aabbBMinOut,aabbBMaxOut;
+					float margin=0.f;
+					b3TransformAabb2(treeBminLocal,treeBmaxLocal, margin,posB,ornB,&aabbBMinOut,&aabbBMaxOut);
+
+					
+					
+					bool aabbOverlap = b3TestAabbAgainstAabb(aabbAMinOut,aabbAMaxOut,aabbBMinOut,aabbBMaxOut);
+					if (aabbOverlap)
+					{
+						
+						int startNodeIndexA = subtreeA.m_rootNodeIndex+bvhInfos[bvhA].m_nodeOffset;
+						int endNodeIndexA = startNodeIndexA+subtreeA.m_subtreeSize;
+
+						int startNodeIndexB = subtreeB.m_rootNodeIndex+bvhInfos[bvhB].m_nodeOffset;
+						int endNodeIndexB = startNodeIndexB+subtreeB.m_subtreeSize;
+
+
+						b3Int2 nodeStack[B3_MAX_STACK_DEPTH];
+						b3Int2 node0;
+						node0.x = startNodeIndexA;
+						node0.y = startNodeIndexB;
+						int maxStackDepth = B3_MAX_STACK_DEPTH;
+						int depth=0;
+						nodeStack[depth++]=node0;
+
+						do
+						{
+							b3Int2 node = nodeStack[--depth];
+
+							b3Float4 aMinLocal = MyUnQuantizeGlobal(quantizedNodes[node.x].m_quantizedAabbMin,bvhInfos[bvhA].m_quantization,bvhInfos[bvhA].m_aabbMin);
+							b3Float4 aMaxLocal = MyUnQuantizeGlobal(quantizedNodes[node.x].m_quantizedAabbMax,bvhInfos[bvhA].m_quantization,bvhInfos[bvhA].m_aabbMin);
+
+							b3Float4 bMinLocal = MyUnQuantizeGlobal(quantizedNodes[node.y].m_quantizedAabbMin,bvhInfos[bvhB].m_quantization,bvhInfos[bvhB].m_aabbMin);
+							b3Float4 bMaxLocal = MyUnQuantizeGlobal(quantizedNodes[node.y].m_quantizedAabbMax,bvhInfos[bvhB].m_quantization,bvhInfos[bvhB].m_aabbMin);
+
+							float margin=0.f;
+							b3Float4 aabbAMinOut,aabbAMaxOut;
+							b3TransformAabb2(aMinLocal,aMaxLocal, margin,posA,ornA,&aabbAMinOut,&aabbAMaxOut);
+
+							b3Float4 aabbBMinOut,aabbBMaxOut;
+							b3TransformAabb2(bMinLocal,bMaxLocal, margin,posB,ornB,&aabbBMinOut,&aabbBMaxOut);
+
+							
+							bool nodeOverlap = b3TestAabbAgainstAabb(aabbAMinOut,aabbAMaxOut,aabbBMinOut,aabbBMaxOut);
+							if (nodeOverlap)
+							{
+								bool isLeafA = isLeafNodeGlobal(&quantizedNodes[node.x]);
+								bool isLeafB = isLeafNodeGlobal(&quantizedNodes[node.y]);
+								bool isInternalA = !isLeafA;
+								bool isInternalB = !isLeafB;
+
+								//fail, even though it might hit two leaf nodes
+								if (depth+4>maxStackDepth && !(isLeafA && isLeafB))
+								{
+									//printf("Error: traversal exceeded maxStackDepth");
+									continue;
+								}
+
+								if(isInternalA)
+								{
+									int nodeAleftChild = node.x+1;
+									bool isNodeALeftChildLeaf = isLeafNodeGlobal(&quantizedNodes[node.x+1]);
+									int nodeArightChild = isNodeALeftChildLeaf? node.x+2 : node.x+1 + getEscapeIndexGlobal(&quantizedNodes[node.x+1]);
+
+									if(isInternalB)
+									{					
+										int nodeBleftChild = node.y+1;
+										bool isNodeBLeftChildLeaf = isLeafNodeGlobal(&quantizedNodes[node.y+1]);
+										int nodeBrightChild = isNodeBLeftChildLeaf? node.y+2 : node.y+1 + getEscapeIndexGlobal(&quantizedNodes[node.y+1]);
+
+										nodeStack[depth++] = b3MakeInt2(nodeAleftChild, nodeBleftChild);
+										nodeStack[depth++] = b3MakeInt2(nodeArightChild, nodeBleftChild);
+										nodeStack[depth++] = b3MakeInt2(nodeAleftChild, nodeBrightChild);
+										nodeStack[depth++] = b3MakeInt2(nodeArightChild, nodeBrightChild);
+									}
+									else
+									{
+										nodeStack[depth++] = b3MakeInt2(nodeAleftChild,node.y);
+										nodeStack[depth++] = b3MakeInt2(nodeArightChild,node.y);
+									}
+								}
+								else
+								{
+									if(isInternalB)
+									{
+										int nodeBleftChild = node.y+1;
+										bool isNodeBLeftChildLeaf = isLeafNodeGlobal(&quantizedNodes[node.y+1]);
+										int nodeBrightChild = isNodeBLeftChildLeaf? node.y+2 : node.y+1 + getEscapeIndexGlobal(&quantizedNodes[node.y+1]);
+										nodeStack[depth++] = b3MakeInt2(node.x,nodeBleftChild);
+										nodeStack[depth++] = b3MakeInt2(node.x,nodeBrightChild);
+									}
+									else
+									{
+										int compoundPairIdx = atomic_inc(numCompoundPairsOut);
+										if (compoundPairIdx<maxNumCompoundPairsCapacity)
+										{
+											int childShapeIndexA = getTriangleIndexGlobal(&quantizedNodes[node.x]);
+											int childShapeIndexB = getTriangleIndexGlobal(&quantizedNodes[node.y]);
+											gpuCompoundPairsOut[compoundPairIdx]  = (int4)(bodyIndexA,bodyIndexB,childShapeIndexA,childShapeIndexB);
+										}
+									}
+								}
+							}
+						} while (depth);
+					}
+				}
+			}
+			
+			return;
+		}
+
+
+
+
+
+		if ((collidables[collidableIndexA].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS) ||(collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS))
+		{
+
+			if (collidables[collidableIndexA].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS) 
+			{
+
+				int numChildrenA = collidables[collidableIndexA].m_numChildShapes;
+				for (int c=0;c<numChildrenA;c++)
+				{
+					int childShapeIndexA = collidables[collidableIndexA].m_shapeIndex+c;
+					int childColIndexA = gpuChildShapes[childShapeIndexA].m_shapeIndex;
+
+					float4 posA = rigidBodies[bodyIndexA].m_pos;
+					float4 ornA = rigidBodies[bodyIndexA].m_quat;
+					float4 childPosA = gpuChildShapes[childShapeIndexA].m_childPosition;
+					float4 childOrnA = gpuChildShapes[childShapeIndexA].m_childOrientation;
+					float4 newPosA = qtRotate(ornA,childPosA)+posA;
+					float4 newOrnA = qtMul(ornA,childOrnA);
+
+					int shapeIndexA = collidables[childColIndexA].m_shapeIndex;
+					b3Aabb_t aabbAlocal = aabbLocalSpace[shapeIndexA];
+					float margin = 0.f;
+					
+					b3Float4 aabbAMinWS;
+					b3Float4 aabbAMaxWS;
+					
+					b3TransformAabb2(aabbAlocal.m_minVec,aabbAlocal.m_maxVec,margin,
+						newPosA,
+						newOrnA,
+						&aabbAMinWS,&aabbAMaxWS);
+						
+					
+					if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+					{
+						int numChildrenB = collidables[collidableIndexB].m_numChildShapes;
+						for (int b=0;b<numChildrenB;b++)
+						{
+							int childShapeIndexB = collidables[collidableIndexB].m_shapeIndex+b;
+							int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+							float4 ornB = rigidBodies[bodyIndexB].m_quat;
+							float4 posB = rigidBodies[bodyIndexB].m_pos;
+							float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+							float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+							float4 newPosB = transform(&childPosB,&posB,&ornB);
+							float4 newOrnB = qtMul(ornB,childOrnB);
+
+							int shapeIndexB = collidables[childColIndexB].m_shapeIndex;
+							b3Aabb_t aabbBlocal = aabbLocalSpace[shapeIndexB];
+							
+							b3Float4 aabbBMinWS;
+							b3Float4 aabbBMaxWS;
+							
+							b3TransformAabb2(aabbBlocal.m_minVec,aabbBlocal.m_maxVec,margin,
+								newPosB,
+								newOrnB,
+								&aabbBMinWS,&aabbBMaxWS);
+								
+								
+							
+							bool aabbOverlap = b3TestAabbAgainstAabb(aabbAMinWS,aabbAMaxWS,aabbBMinWS,aabbBMaxWS);
+							if (aabbOverlap)
+							{
+								int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+								float dmin = FLT_MAX;
+								float4 posA = newPosA;
+								posA.w = 0.f;
+								float4 posB = newPosB;
+								posB.w = 0.f;
+								float4 c0local = convexShapes[shapeIndexA].m_localCenter;
+								float4 ornA = newOrnA;
+								float4 c0 = transform(&c0local, &posA, &ornA);
+								float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+								float4 ornB =newOrnB;
+								float4 c1 = transform(&c1local,&posB,&ornB);
+								const float4 DeltaC2 = c0 - c1;
+
+								{//
+									int compoundPairIdx = atomic_inc(numCompoundPairsOut);
+									if (compoundPairIdx<maxNumCompoundPairsCapacity)
+									{
+										gpuCompoundPairsOut[compoundPairIdx]  = (int4)(bodyIndexA,bodyIndexB,childShapeIndexA,childShapeIndexB);
+									}
+								}//
+							}//fi(1)
+						} //for (int b=0
+					}//if (collidables[collidableIndexB].
+					else//if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+					{
+						if (1)
+						{
+							int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+							float dmin = FLT_MAX;
+							float4 posA = newPosA;
+							posA.w = 0.f;
+							float4 posB = rigidBodies[bodyIndexB].m_pos;
+							posB.w = 0.f;
+							float4 c0local = convexShapes[shapeIndexA].m_localCenter;
+							float4 ornA = newOrnA;
+							float4 c0 = transform(&c0local, &posA, &ornA);
+							float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+							float4 ornB = rigidBodies[bodyIndexB].m_quat;
+							float4 c1 = transform(&c1local,&posB,&ornB);
+							const float4 DeltaC2 = c0 - c1;
+
+							{
+								int compoundPairIdx = atomic_inc(numCompoundPairsOut);
+								if (compoundPairIdx<maxNumCompoundPairsCapacity)
+								{
+									gpuCompoundPairsOut[compoundPairIdx] = (int4)(bodyIndexA,bodyIndexB,childShapeIndexA,-1);
+								}//if (compoundPairIdx<maxNumCompoundPairsCapacity)
+							}//
+						}//fi (1)
+					}//if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+				}//for (int b=0;b<numChildrenB;b++)	
+				return;
+			}//if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+			if ((collidables[collidableIndexA].m_shapeType!=SHAPE_CONCAVE_TRIMESH) 
+				&& (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS))
+			{
+				int numChildrenB = collidables[collidableIndexB].m_numChildShapes;
+				for (int b=0;b<numChildrenB;b++)
+				{
+					int childShapeIndexB = collidables[collidableIndexB].m_shapeIndex+b;
+					int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+					float4 ornB = rigidBodies[bodyIndexB].m_quat;
+					float4 posB = rigidBodies[bodyIndexB].m_pos;
+					float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+					float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+					float4 newPosB = qtRotate(ornB,childPosB)+posB;
+					float4 newOrnB = qtMul(ornB,childOrnB);
+
+					int shapeIndexB = collidables[childColIndexB].m_shapeIndex;
+
+
+					//////////////////////////////////////
+
+					if (1)
+					{
+						int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+						float dmin = FLT_MAX;
+						float4 posA = rigidBodies[bodyIndexA].m_pos;
+						posA.w = 0.f;
+						float4 posB = newPosB;
+						posB.w = 0.f;
+						float4 c0local = convexShapes[shapeIndexA].m_localCenter;
+						float4 ornA = rigidBodies[bodyIndexA].m_quat;
+						float4 c0 = transform(&c0local, &posA, &ornA);
+						float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+						float4 ornB =newOrnB;
+						float4 c1 = transform(&c1local,&posB,&ornB);
+						const float4 DeltaC2 = c0 - c1;
+						{//
+							int compoundPairIdx = atomic_inc(numCompoundPairsOut);
+							if (compoundPairIdx<maxNumCompoundPairsCapacity)
+							{
+								gpuCompoundPairsOut[compoundPairIdx] = (int4)(bodyIndexA,bodyIndexB,-1,childShapeIndexB);
+							}//fi (compoundPairIdx<maxNumCompoundPairsCapacity)
+						}//
+					}//fi (1)	
+				}//for (int b=0;b<numChildrenB;b++)
+				return;
+			}//if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+			return;
+		}//fi ((collidables[collidableIndexA].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS) ||(collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS))
+	}//i<numPairs
+}
+
+// work-in-progress
+__kernel void   findSeparatingAxisKernel( __global const int4* pairs, 
+																					__global const BodyData* rigidBodies, 
+																					__global const btCollidableGpu* collidables,
+																					__global const ConvexPolyhedronCL* convexShapes, 
+																					__global const float4* vertices,
+																					__global const float4* uniqueEdges,
+																					__global const btGpuFace* faces,
+																					__global const int* indices,
+																					__global btAabbCL* aabbs,
+																					__global volatile float4* separatingNormals,
+																					__global volatile int* hasSeparatingAxis,
+																					int numPairs
+																					)
+{
+
+	int i = get_global_id(0);
+	
+	if (i<numPairs)
+	{
+
+	
+		int bodyIndexA = pairs[i].x;
+		int bodyIndexB = pairs[i].y;
+
+		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+	
+		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+		
+		
+		//once the broadphase avoids static-static pairs, we can remove this test
+		if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))
+		{
+			hasSeparatingAxis[i] = 0;
+			return;
+		}
+		
+
+		if ((collidables[collidableIndexA].m_shapeType!=SHAPE_CONVEX_HULL) ||(collidables[collidableIndexB].m_shapeType!=SHAPE_CONVEX_HULL))
+		{
+			hasSeparatingAxis[i] = 0;
+			return;
+		}
+			
+		if ((collidables[collidableIndexA].m_shapeType==SHAPE_CONCAVE_TRIMESH))
+		{
+			hasSeparatingAxis[i] = 0;
+			return;
+		}
+
+		int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+
+		float dmin = FLT_MAX;
+
+		float4 posA = rigidBodies[bodyIndexA].m_pos;
+		posA.w = 0.f;
+		float4 posB = rigidBodies[bodyIndexB].m_pos;
+		posB.w = 0.f;
+		float4 c0local = convexShapes[shapeIndexA].m_localCenter;
+		float4 ornA = rigidBodies[bodyIndexA].m_quat;
+		float4 c0 = transform(&c0local, &posA, &ornA);
+		float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+		float4 ornB =rigidBodies[bodyIndexB].m_quat;
+		float4 c1 = transform(&c1local,&posB,&ornB);
+		const float4 DeltaC2 = c0 - c1;
+		float4 sepNormal;
+		
+		bool sepA = findSeparatingAxis(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,
+																								posB,ornB,
+																								DeltaC2,
+																								vertices,uniqueEdges,faces,
+																								indices,&sepNormal,&dmin);
+		hasSeparatingAxis[i] = 4;
+		if (!sepA)
+		{
+			hasSeparatingAxis[i] = 0;
+		} else
+		{
+			bool sepB = findSeparatingAxis(	&convexShapes[shapeIndexB],&convexShapes[shapeIndexA],posB,ornB,
+																									posA,ornA,
+																									DeltaC2,
+																									vertices,uniqueEdges,faces,
+																									indices,&sepNormal,&dmin);
+
+			if (!sepB)
+			{
+				hasSeparatingAxis[i] = 0;
+			} else
+			{
+				bool sepEE = findSeparatingAxisEdgeEdge(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,
+																									posB,ornB,
+																									DeltaC2,
+																									vertices,uniqueEdges,faces,
+																									indices,&sepNormal,&dmin);
+				if (!sepEE)
+				{
+					hasSeparatingAxis[i] = 0;
+				} else
+				{
+					hasSeparatingAxis[i] = 1;
+					separatingNormals[i] = sepNormal;
+				}
+			}
+		}
+		
+	}
+
+}
+
+
+__kernel void   findSeparatingAxisVertexFaceKernel( __global const int4* pairs, 
+																					__global const BodyData* rigidBodies, 
+																					__global const btCollidableGpu* collidables,
+																					__global const ConvexPolyhedronCL* convexShapes, 
+																					__global const float4* vertices,
+																					__global const float4* uniqueEdges,
+																					__global const btGpuFace* faces,
+																					__global const int* indices,
+																					__global btAabbCL* aabbs,
+																					__global volatile float4* separatingNormals,
+																					__global volatile int* hasSeparatingAxis,
+																					__global  float* dmins,
+																					int numPairs
+																					)
+{
+
+	int i = get_global_id(0);
+	
+	if (i<numPairs)
+	{
+
+	
+		int bodyIndexA = pairs[i].x;
+		int bodyIndexB = pairs[i].y;
+
+		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+	
+		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+	
+		hasSeparatingAxis[i] = 0;	
+		
+		//once the broadphase avoids static-static pairs, we can remove this test
+		if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))
+		{
+			return;
+		}
+		
+
+		if ((collidables[collidableIndexA].m_shapeType!=SHAPE_CONVEX_HULL) ||(collidables[collidableIndexB].m_shapeType!=SHAPE_CONVEX_HULL))
+		{
+			return;
+		}
+			
+
+		int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+
+		float dmin = FLT_MAX;
+
+		dmins[i] = dmin;
+		
+		float4 posA = rigidBodies[bodyIndexA].m_pos;
+		posA.w = 0.f;
+		float4 posB = rigidBodies[bodyIndexB].m_pos;
+		posB.w = 0.f;
+		float4 c0local = convexShapes[shapeIndexA].m_localCenter;
+		float4 ornA = rigidBodies[bodyIndexA].m_quat;
+		float4 c0 = transform(&c0local, &posA, &ornA);
+		float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+		float4 ornB =rigidBodies[bodyIndexB].m_quat;
+		float4 c1 = transform(&c1local,&posB,&ornB);
+		const float4 DeltaC2 = c0 - c1;
+		float4 sepNormal;
+		
+		bool sepA = findSeparatingAxis(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,
+																								posB,ornB,
+																								DeltaC2,
+																								vertices,uniqueEdges,faces,
+																								indices,&sepNormal,&dmin);
+		hasSeparatingAxis[i] = 4;
+		if (!sepA)
+		{
+			hasSeparatingAxis[i] = 0;
+		} else
+		{
+			bool sepB = findSeparatingAxis(	&convexShapes[shapeIndexB],&convexShapes[shapeIndexA],posB,ornB,
+																									posA,ornA,
+																									DeltaC2,
+																									vertices,uniqueEdges,faces,
+																									indices,&sepNormal,&dmin);
+
+			if (sepB)
+			{
+				dmins[i] = dmin;
+				hasSeparatingAxis[i] = 1;
+				separatingNormals[i] = sepNormal;
+			}
+		}
+		
+	}
+
+}
+
+
+__kernel void   findSeparatingAxisEdgeEdgeKernel( __global const int4* pairs, 
+																					__global const BodyData* rigidBodies, 
+																					__global const btCollidableGpu* collidables,
+																					__global const ConvexPolyhedronCL* convexShapes, 
+																					__global const float4* vertices,
+																					__global const float4* uniqueEdges,
+																					__global const btGpuFace* faces,
+																					__global const int* indices,
+																					__global btAabbCL* aabbs,
+																					__global  float4* separatingNormals,
+																					__global  int* hasSeparatingAxis,
+																					__global  float* dmins,
+																					__global const float4* unitSphereDirections,
+																					int numUnitSphereDirections,
+																					int numPairs
+																					)
+{
+
+	int i = get_global_id(0);
+	
+	if (i<numPairs)
+	{
+
+		if (hasSeparatingAxis[i])
+		{
+	
+			int bodyIndexA = pairs[i].x;
+			int bodyIndexB = pairs[i].y;
+	
+			int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+			int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+		
+			int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+			int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+			
+			
+			int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+	
+			float dmin = dmins[i];
+	
+			float4 posA = rigidBodies[bodyIndexA].m_pos;
+			posA.w = 0.f;
+			float4 posB = rigidBodies[bodyIndexB].m_pos;
+			posB.w = 0.f;
+			float4 c0local = convexShapes[shapeIndexA].m_localCenter;
+			float4 ornA = rigidBodies[bodyIndexA].m_quat;
+			float4 c0 = transform(&c0local, &posA, &ornA);
+			float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+			float4 ornB =rigidBodies[bodyIndexB].m_quat;
+			float4 c1 = transform(&c1local,&posB,&ornB);
+			const float4 DeltaC2 = c0 - c1;
+			float4 sepNormal = separatingNormals[i];
+			
+			
+			
+			bool sepEE = false;
+			int numEdgeEdgeDirections = convexShapes[shapeIndexA].m_numUniqueEdges*convexShapes[shapeIndexB].m_numUniqueEdges;
+			if (numEdgeEdgeDirections<=numUnitSphereDirections)
+			{
+				sepEE = findSeparatingAxisEdgeEdge(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,
+																									posB,ornB,
+																									DeltaC2,
+																									vertices,uniqueEdges,faces,
+																									indices,&sepNormal,&dmin);
+																									
+					if (!sepEE)
+					{
+						hasSeparatingAxis[i] = 0;
+					} else
+					{
+						hasSeparatingAxis[i] = 1;
+						separatingNormals[i] = sepNormal;
+					}
+			}
+			/*
+			///else case is a separate kernel, to make Mac OSX OpenCL compiler happy
+			else
+			{
+				sepEE = findSeparatingAxisUnitSphere(&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,
+																									posB,ornB,
+																									DeltaC2,
+																									vertices,unitSphereDirections,numUnitSphereDirections,
+																									&sepNormal,&dmin);
+					if (!sepEE)
+					{
+						hasSeparatingAxis[i] = 0;
+					} else
+					{
+						hasSeparatingAxis[i] = 1;
+						separatingNormals[i] = sepNormal;
+					}
+			}
+			*/
+		}		//if (hasSeparatingAxis[i])
+	}//(i<numPairs)
+}
+
+
+
+
+
+inline int	findClippingFaces(const float4 separatingNormal,
+                      const ConvexPolyhedronCL* hullA, 
+					  __global const ConvexPolyhedronCL* hullB,
+                      const float4 posA, const Quaternion ornA,const float4 posB, const Quaternion ornB,
+                       __global float4* worldVertsA1,
+                      __global float4* worldNormalsA1,
+                      __global float4* worldVertsB1,
+                      int capacityWorldVerts,
+                      const float minDist, float maxDist,
+					  const float4* verticesA,
+                      const btGpuFace* facesA,
+                      const int* indicesA,
+					  __global const float4* verticesB,
+                      __global const btGpuFace* facesB,
+                      __global const int* indicesB,
+                      __global int4* clippingFaces, int pairIndex)
+{
+	int numContactsOut = 0;
+	int numWorldVertsB1= 0;
+    
+    
+	int closestFaceB=0;
+	float dmax = -FLT_MAX;
+    
+	{
+		for(int face=0;face<hullB->m_numFaces;face++)
+		{
+			const float4 Normal = make_float4(facesB[hullB->m_faceOffset+face].m_plane.x,
+                                              facesB[hullB->m_faceOffset+face].m_plane.y, facesB[hullB->m_faceOffset+face].m_plane.z,0.f);
+			const float4 WorldNormal = qtRotate(ornB, Normal);
+			float d = dot3F4(WorldNormal,separatingNormal);
+			if (d > dmax)
+			{
+				dmax = d;
+				closestFaceB = face;
+			}
+		}
+	}
+    
+	{
+		const btGpuFace polyB = facesB[hullB->m_faceOffset+closestFaceB];
+		int numVertices = polyB.m_numIndices;
+        if (numVertices>capacityWorldVerts)
+            numVertices = capacityWorldVerts;
+        
+		for(int e0=0;e0<numVertices;e0++)
+		{
+            if (e0<capacityWorldVerts)
+            {
+                const float4 b = verticesB[hullB->m_vertexOffset+indicesB[polyB.m_indexOffset+e0]];
+                worldVertsB1[pairIndex*capacityWorldVerts+numWorldVertsB1++] = transform(&b,&posB,&ornB);
+            }
+		}
+	}
+    
+    int closestFaceA=0;
+	{
+		float dmin = FLT_MAX;
+		for(int face=0;face<hullA->m_numFaces;face++)
+		{
+			const float4 Normal = make_float4(
+                                              facesA[hullA->m_faceOffset+face].m_plane.x,
+                                              facesA[hullA->m_faceOffset+face].m_plane.y,
+                                              facesA[hullA->m_faceOffset+face].m_plane.z,
+                                              0.f);
+			const float4 faceANormalWS = qtRotate(ornA,Normal);
+            
+			float d = dot3F4(faceANormalWS,separatingNormal);
+			if (d < dmin)
+			{
+				dmin = d;
+				closestFaceA = face;
+                worldNormalsA1[pairIndex] = faceANormalWS;
+			}
+		}
+	}
+    
+    int numVerticesA = facesA[hullA->m_faceOffset+closestFaceA].m_numIndices;
+    if (numVerticesA>capacityWorldVerts)
+       numVerticesA = capacityWorldVerts;
+    
+	for(int e0=0;e0<numVerticesA;e0++)
+	{
+        if (e0<capacityWorldVerts)
+        {
+            const float4 a = verticesA[hullA->m_vertexOffset+indicesA[facesA[hullA->m_faceOffset+closestFaceA].m_indexOffset+e0]];
+            worldVertsA1[pairIndex*capacityWorldVerts+e0] = transform(&a, &posA,&ornA);
+        }
+    }
+    
+    clippingFaces[pairIndex].x = closestFaceA;
+    clippingFaces[pairIndex].y = closestFaceB;
+    clippingFaces[pairIndex].z = numVerticesA;
+    clippingFaces[pairIndex].w = numWorldVertsB1;
+    
+    
+	return numContactsOut;
+}
+
+
+
+
+// work-in-progress
+__kernel void   findConcaveSeparatingAxisKernel( __global int4* concavePairs,
+																					__global const BodyData* rigidBodies,
+																					__global const btCollidableGpu* collidables,
+																					__global const ConvexPolyhedronCL* convexShapes, 
+																					__global const float4* vertices,
+																					__global const float4* uniqueEdges,
+																					__global const btGpuFace* faces,
+																					__global const int* indices,
+																					__global const btGpuChildShape* gpuChildShapes,
+																					__global btAabbCL* aabbs,
+																					__global float4* concaveSeparatingNormalsOut,
+																					__global int* concaveHasSeparatingNormals,
+																					__global int4* clippingFacesOut,
+																					__global float4* worldVertsA1GPU,
+																					__global float4*  worldNormalsAGPU,
+																					__global float4* worldVertsB1GPU,
+																					int vertexFaceCapacity,
+																					int numConcavePairs
+																					)
+{
+
+	int i = get_global_id(0);
+	if (i>=numConcavePairs)
+		return;
+
+	concaveHasSeparatingNormals[i] = 0;
+
+	int pairIdx = i;
+
+	int bodyIndexA = concavePairs[i].x;
+	int bodyIndexB = concavePairs[i].y;
+
+	int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+	int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+
+	int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+	int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+
+	if (collidables[collidableIndexB].m_shapeType!=SHAPE_CONVEX_HULL&&
+		collidables[collidableIndexB].m_shapeType!=SHAPE_COMPOUND_OF_CONVEX_HULLS)
+	{
+		concavePairs[pairIdx].w = -1;
+		return;
+	}
+
+
+
+	int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+	int numActualConcaveConvexTests = 0;
+	
+	int f = concavePairs[i].z;
+	
+	bool overlap = false;
+	
+	ConvexPolyhedronCL convexPolyhedronA;
+
+	//add 3 vertices of the triangle
+	convexPolyhedronA.m_numVertices = 3;
+	convexPolyhedronA.m_vertexOffset = 0;
+	float4	localCenter = make_float4(0.f,0.f,0.f,0.f);
+
+	btGpuFace face = faces[convexShapes[shapeIndexA].m_faceOffset+f];
+	float4 triMinAabb, triMaxAabb;
+	btAabbCL triAabb;
+	triAabb.m_min = make_float4(1e30f,1e30f,1e30f,0.f);
+	triAabb.m_max = make_float4(-1e30f,-1e30f,-1e30f,0.f);
+	
+	float4 verticesA[3];
+	for (int i=0;i<3;i++)
+	{
+		int index = indices[face.m_indexOffset+i];
+		float4 vert = vertices[convexShapes[shapeIndexA].m_vertexOffset+index];
+		verticesA[i] = vert;
+		localCenter += vert;
+			
+		triAabb.m_min = min(triAabb.m_min,vert);		
+		triAabb.m_max = max(triAabb.m_max,vert);		
+
+	}
+
+	overlap = true;
+	overlap = (triAabb.m_min.x > aabbs[bodyIndexB].m_max.x || triAabb.m_max.x < aabbs[bodyIndexB].m_min.x) ? false : overlap;
+	overlap = (triAabb.m_min.z > aabbs[bodyIndexB].m_max.z || triAabb.m_max.z < aabbs[bodyIndexB].m_min.z) ? false : overlap;
+	overlap = (triAabb.m_min.y > aabbs[bodyIndexB].m_max.y || triAabb.m_max.y < aabbs[bodyIndexB].m_min.y) ? false : overlap;
+		
+	if (overlap)
+	{
+		float dmin = FLT_MAX;
+		int hasSeparatingAxis=5;
+		float4 sepAxis=make_float4(1,2,3,4);
+
+		int localCC=0;
+		numActualConcaveConvexTests++;
+
+		//a triangle has 3 unique edges
+		convexPolyhedronA.m_numUniqueEdges = 3;
+		convexPolyhedronA.m_uniqueEdgesOffset = 0;
+		float4 uniqueEdgesA[3];
+		
+		uniqueEdgesA[0] = (verticesA[1]-verticesA[0]);
+		uniqueEdgesA[1] = (verticesA[2]-verticesA[1]);
+		uniqueEdgesA[2] = (verticesA[0]-verticesA[2]);
+
+
+		convexPolyhedronA.m_faceOffset = 0;
+                                  
+		float4 normal = make_float4(face.m_plane.x,face.m_plane.y,face.m_plane.z,0.f);
+                             
+		btGpuFace facesA[TRIANGLE_NUM_CONVEX_FACES];
+		int indicesA[3+3+2+2+2];
+		int curUsedIndices=0;
+		int fidx=0;
+
+		//front size of triangle
+		{
+			facesA[fidx].m_indexOffset=curUsedIndices;
+			indicesA[0] = 0;
+			indicesA[1] = 1;
+			indicesA[2] = 2;
+			curUsedIndices+=3;
+			float c = face.m_plane.w;
+			facesA[fidx].m_plane.x = normal.x;
+			facesA[fidx].m_plane.y = normal.y;
+			facesA[fidx].m_plane.z = normal.z;
+			facesA[fidx].m_plane.w = c;
+			facesA[fidx].m_numIndices=3;
+		}
+		fidx++;
+		//back size of triangle
+		{
+			facesA[fidx].m_indexOffset=curUsedIndices;
+			indicesA[3]=2;
+			indicesA[4]=1;
+			indicesA[5]=0;
+			curUsedIndices+=3;
+			float c = dot(normal,verticesA[0]);
+			float c1 = -face.m_plane.w;
+			facesA[fidx].m_plane.x = -normal.x;
+			facesA[fidx].m_plane.y = -normal.y;
+			facesA[fidx].m_plane.z = -normal.z;
+			facesA[fidx].m_plane.w = c;
+			facesA[fidx].m_numIndices=3;
+		}
+		fidx++;
+
+		bool addEdgePlanes = true;
+		if (addEdgePlanes)
+		{
+			int numVertices=3;
+			int prevVertex = numVertices-1;
+			for (int i=0;i<numVertices;i++)
+			{
+				float4 v0 = verticesA[i];
+				float4 v1 = verticesA[prevVertex];
+                                            
+				float4 edgeNormal = normalize(cross(normal,v1-v0));
+				float c = -dot(edgeNormal,v0);
+
+				facesA[fidx].m_numIndices = 2;
+				facesA[fidx].m_indexOffset=curUsedIndices;
+				indicesA[curUsedIndices++]=i;
+				indicesA[curUsedIndices++]=prevVertex;
+                                            
+				facesA[fidx].m_plane.x = edgeNormal.x;
+				facesA[fidx].m_plane.y = edgeNormal.y;
+				facesA[fidx].m_plane.z = edgeNormal.z;
+				facesA[fidx].m_plane.w = c;
+				fidx++;
+				prevVertex = i;
+			}
+		}
+		convexPolyhedronA.m_numFaces = TRIANGLE_NUM_CONVEX_FACES;
+		convexPolyhedronA.m_localCenter = localCenter*(1.f/3.f);
+
+
+		float4 posA = rigidBodies[bodyIndexA].m_pos;
+		posA.w = 0.f;
+		float4 posB = rigidBodies[bodyIndexB].m_pos;
+		posB.w = 0.f;
+
+		float4 ornA = rigidBodies[bodyIndexA].m_quat;
+		float4 ornB =rigidBodies[bodyIndexB].m_quat;
+
+		
+
+
+		///////////////////
+		///compound shape support
+
+		if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+		{
+			int compoundChild = concavePairs[pairIdx].w;
+			int childShapeIndexB = compoundChild;//collidables[collidableIndexB].m_shapeIndex+compoundChild;
+			int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+			float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+			float4 newPosB = transform(&childPosB,&posB,&ornB);
+			float4 newOrnB = qtMul(ornB,childOrnB);
+			posB = newPosB;
+			ornB = newOrnB;
+			shapeIndexB = collidables[childColIndexB].m_shapeIndex;
+		}
+		//////////////////
+
+		float4 c0local = convexPolyhedronA.m_localCenter;
+		float4 c0 = transform(&c0local, &posA, &ornA);
+		float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+		float4 c1 = transform(&c1local,&posB,&ornB);
+		const float4 DeltaC2 = c0 - c1;
+
+
+		bool sepA = findSeparatingAxisLocalA(	&convexPolyhedronA, &convexShapes[shapeIndexB],
+												posA,ornA,
+												posB,ornB,
+												DeltaC2,
+												verticesA,uniqueEdgesA,facesA,indicesA,
+												vertices,uniqueEdges,faces,indices,
+												&sepAxis,&dmin);
+		hasSeparatingAxis = 4;
+		if (!sepA)
+		{
+			hasSeparatingAxis = 0;
+		} else
+		{
+			bool sepB = findSeparatingAxisLocalB(	&convexShapes[shapeIndexB],&convexPolyhedronA,
+												posB,ornB,
+												posA,ornA,
+												DeltaC2,
+												vertices,uniqueEdges,faces,indices,
+												verticesA,uniqueEdgesA,facesA,indicesA,
+												&sepAxis,&dmin);
+
+			if (!sepB)
+			{
+				hasSeparatingAxis = 0;
+			} else
+			{
+				bool sepEE = findSeparatingAxisEdgeEdgeLocalA(	&convexPolyhedronA, &convexShapes[shapeIndexB],
+															posA,ornA,
+															posB,ornB,
+															DeltaC2,
+															verticesA,uniqueEdgesA,facesA,indicesA,
+															vertices,uniqueEdges,faces,indices,
+															&sepAxis,&dmin);
+	
+				if (!sepEE)
+				{
+					hasSeparatingAxis = 0;
+				} else
+				{
+					hasSeparatingAxis = 1;
+				}
+			}
+		}	
+		
+		if (hasSeparatingAxis)
+		{
+			sepAxis.w = dmin;
+			concaveSeparatingNormalsOut[pairIdx]=sepAxis;
+			concaveHasSeparatingNormals[i]=1;
+
+
+			float minDist = -1e30f;
+			float maxDist = 0.02f;
+
+		
+
+			findClippingFaces(sepAxis,
+                     &convexPolyhedronA,
+					 &convexShapes[shapeIndexB],
+					 posA,ornA,
+					 posB,ornB,
+                      worldVertsA1GPU,
+                      worldNormalsAGPU,
+                      worldVertsB1GPU,
+					  vertexFaceCapacity,
+                      minDist, maxDist,
+                      verticesA,
+                      facesA,
+                      indicesA,
+ 					  vertices,
+                      faces,
+                      indices,
+                      clippingFacesOut, pairIdx);
+
+
+		} else
+		{	
+			//mark this pair as in-active
+			concavePairs[pairIdx].w = -1;
+		}
+	}
+	else
+	{	
+		//mark this pair as in-active
+		concavePairs[pairIdx].w = -1;
+	}
+	
+	concavePairs[pairIdx].z = -1;//now z is used for existing/persistent contacts
+}
+
+
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satClipHullContacts.cl b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satClipHullContacts.cl
new file mode 100644
index 0000000000..f433971741
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satClipHullContacts.cl
@@ -0,0 +1,1888 @@
+
+#define TRIANGLE_NUM_CONVEX_FACES 5
+
+
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile __global int*
+#endif
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+#define max2 max
+#define min2 min
+
+typedef unsigned int u32;
+
+
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Contact4Data.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3ConvexPolyhedronData.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Collidable.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3RigidBodyData.h"
+
+
+
+#define GET_NPOINTS(x) (x).m_worldNormalOnB.w
+
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define make_float4 (float4)
+#define make_float2 (float2)
+#define make_uint4 (uint4)
+#define make_int4 (int4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+
+__inline
+float fastDiv(float numerator, float denominator)
+{
+	return native_divide(numerator, denominator);	
+//	return numerator/denominator;	
+}
+
+__inline
+float4 fastDiv4(float4 numerator, float4 denominator)
+{
+	return native_divide(numerator, denominator);	
+}
+
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+
+//#define dot3F4 dot
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = make_float4(a.xyz,0.f);
+	float4 b1 = make_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	return fast_normalize(v);
+}
+
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b);
+
+__inline
+Quaternion qtNormalize(Quaternion in);
+
+__inline
+float4 qtRotate(Quaternion q, float4 vec);
+
+__inline
+Quaternion qtInvert(Quaternion q);
+
+
+
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross3( a, b );
+	ans += a.w*b+b.w*a;
+//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+Quaternion qtNormalize(Quaternion in)
+{
+	return fastNormalize4(in);
+//	in /= length( in );
+//	return in;
+}
+__inline
+float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(qtMul(q,vcpy),qInv);
+	return out;
+}
+
+__inline
+Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+__inline
+float4 qtInvRotate(const Quaternion q, float4 vec)
+{
+	return qtRotate( qtInvert( q ), vec );
+}
+
+__inline
+float4 transform(const float4* p, const float4* translation, const Quaternion* orientation)
+{
+	return qtRotate( *orientation, *p ) + (*translation);
+}
+
+
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = make_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+}
+
+
+__inline float4 lerp3(const float4 a,const float4 b, float  t)
+{
+	return make_float4(	a.x + (b.x - a.x) * t,
+						a.y + (b.y - a.y) * t,
+						a.z + (b.z - a.z) * t,
+						0.f);
+}
+
+
+
+// Clips a face to the back of a plane, return the number of vertices out, stored in ppVtxOut
+int clipFaceGlobal(__global const float4* pVtxIn, int numVertsIn, float4 planeNormalWS,float planeEqWS, __global float4* ppVtxOut)
+{
+	
+	int ve;
+	float ds, de;
+	int numVertsOut = 0;
+    //double-check next test
+    	if (numVertsIn < 2)
+    		return 0;
+    
+	float4 firstVertex=pVtxIn[numVertsIn-1];
+	float4 endVertex = pVtxIn[0];
+	
+	ds = dot3F4(planeNormalWS,firstVertex)+planeEqWS;
+    
+	for (ve = 0; ve < numVertsIn; ve++)
+	{
+		endVertex=pVtxIn[ve];
+		de = dot3F4(planeNormalWS,endVertex)+planeEqWS;
+		if (ds<0)
+		{
+			if (de<0)
+			{
+				// Start < 0, end < 0, so output endVertex
+				ppVtxOut[numVertsOut++] = endVertex;
+			}
+			else
+			{
+				// Start < 0, end >= 0, so output intersection
+				ppVtxOut[numVertsOut++] = lerp3(firstVertex, endVertex,(ds * 1.f/(ds - de)) );
+			}
+		}
+		else
+		{
+			if (de<0)
+			{
+				// Start >= 0, end < 0 so output intersection and end
+				ppVtxOut[numVertsOut++] = lerp3(firstVertex, endVertex,(ds * 1.f/(ds - de)) );
+				ppVtxOut[numVertsOut++] = endVertex;
+			}
+		}
+		firstVertex = endVertex;
+		ds = de;
+	}
+	return numVertsOut;
+}
+
+
+
+// Clips a face to the back of a plane, return the number of vertices out, stored in ppVtxOut
+int clipFace(const float4* pVtxIn, int numVertsIn, float4 planeNormalWS,float planeEqWS, float4* ppVtxOut)
+{
+	
+	int ve;
+	float ds, de;
+	int numVertsOut = 0;
+//double-check next test
+	if (numVertsIn < 2)
+		return 0;
+
+	float4 firstVertex=pVtxIn[numVertsIn-1];
+	float4 endVertex = pVtxIn[0];
+	
+	ds = dot3F4(planeNormalWS,firstVertex)+planeEqWS;
+
+	for (ve = 0; ve < numVertsIn; ve++)
+	{
+		endVertex=pVtxIn[ve];
+
+		de = dot3F4(planeNormalWS,endVertex)+planeEqWS;
+
+		if (ds<0)
+		{
+			if (de<0)
+			{
+				// Start < 0, end < 0, so output endVertex
+				ppVtxOut[numVertsOut++] = endVertex;
+			}
+			else
+			{
+				// Start < 0, end >= 0, so output intersection
+				ppVtxOut[numVertsOut++] = lerp3(firstVertex, endVertex,(ds * 1.f/(ds - de)) );
+			}
+		}
+		else
+		{
+			if (de<0)
+			{
+				// Start >= 0, end < 0 so output intersection and end
+				ppVtxOut[numVertsOut++] = lerp3(firstVertex, endVertex,(ds * 1.f/(ds - de)) );
+				ppVtxOut[numVertsOut++] = endVertex;
+			}
+		}
+		firstVertex = endVertex;
+		ds = de;
+	}
+	return numVertsOut;
+}
+
+
+int clipFaceAgainstHull(const float4 separatingNormal, __global const b3ConvexPolyhedronData_t* hullA,  
+	const float4 posA, const Quaternion ornA, float4* worldVertsB1, int numWorldVertsB1,
+	float4* worldVertsB2, int capacityWorldVertsB2,
+	const float minDist, float maxDist,
+	__global const float4* vertices,
+	__global const b3GpuFace_t* faces,
+	__global const int* indices,
+	float4* contactsOut,
+	int contactCapacity)
+{
+	int numContactsOut = 0;
+
+	float4* pVtxIn = worldVertsB1;
+	float4* pVtxOut = worldVertsB2;
+	
+	int numVertsIn = numWorldVertsB1;
+	int numVertsOut = 0;
+
+	int closestFaceA=-1;
+	{
+		float dmin = FLT_MAX;
+		for(int face=0;face<hullA->m_numFaces;face++)
+		{
+			const float4 Normal = make_float4(
+				faces[hullA->m_faceOffset+face].m_plane.x, 
+				faces[hullA->m_faceOffset+face].m_plane.y, 
+				faces[hullA->m_faceOffset+face].m_plane.z,0.f);
+			const float4 faceANormalWS = qtRotate(ornA,Normal);
+		
+			float d = dot3F4(faceANormalWS,separatingNormal);
+			if (d < dmin)
+			{
+				dmin = d;
+				closestFaceA = face;
+			}
+		}
+	}
+	if (closestFaceA<0)
+		return numContactsOut;
+
+	b3GpuFace_t polyA = faces[hullA->m_faceOffset+closestFaceA];
+
+	// clip polygon to back of planes of all faces of hull A that are adjacent to witness face
+	int numVerticesA = polyA.m_numIndices;
+	for(int e0=0;e0<numVerticesA;e0++)
+	{
+		const float4 a = vertices[hullA->m_vertexOffset+indices[polyA.m_indexOffset+e0]];
+		const float4 b = vertices[hullA->m_vertexOffset+indices[polyA.m_indexOffset+((e0+1)%numVerticesA)]];
+		const float4 edge0 = a - b;
+		const float4 WorldEdge0 = qtRotate(ornA,edge0);
+		float4 planeNormalA = make_float4(polyA.m_plane.x,polyA.m_plane.y,polyA.m_plane.z,0.f);
+		float4 worldPlaneAnormal1 = qtRotate(ornA,planeNormalA);
+
+		float4 planeNormalWS1 = -cross3(WorldEdge0,worldPlaneAnormal1);
+		float4 worldA1 = transform(&a,&posA,&ornA);
+		float planeEqWS1 = -dot3F4(worldA1,planeNormalWS1);
+		
+		float4 planeNormalWS = planeNormalWS1;
+		float planeEqWS=planeEqWS1;
+		
+		//clip face
+		//clipFace(*pVtxIn, *pVtxOut,planeNormalWS,planeEqWS);
+		numVertsOut = clipFace(pVtxIn, numVertsIn, planeNormalWS,planeEqWS, pVtxOut);
+
+		//btSwap(pVtxIn,pVtxOut);
+		float4* tmp = pVtxOut;
+		pVtxOut = pVtxIn;
+		pVtxIn = tmp;
+		numVertsIn = numVertsOut;
+		numVertsOut = 0;
+	}
+
+	
+	// only keep points that are behind the witness face
+	{
+		float4 localPlaneNormal  = make_float4(polyA.m_plane.x,polyA.m_plane.y,polyA.m_plane.z,0.f);
+		float localPlaneEq = polyA.m_plane.w;
+		float4 planeNormalWS = qtRotate(ornA,localPlaneNormal);
+		float planeEqWS=localPlaneEq-dot3F4(planeNormalWS,posA);
+		for (int i=0;i<numVertsIn;i++)
+		{
+			float depth = dot3F4(planeNormalWS,pVtxIn[i])+planeEqWS;
+			if (depth <=minDist)
+			{
+				depth = minDist;
+			}
+
+			if (depth <=maxDist)
+			{
+				float4 pointInWorld = pVtxIn[i];
+				//resultOut.addContactPoint(separatingNormal,point,depth);
+				contactsOut[numContactsOut++] = make_float4(pointInWorld.x,pointInWorld.y,pointInWorld.z,depth);
+			}
+		}
+	}
+
+	return numContactsOut;
+}
+
+
+
+int clipFaceAgainstHullLocalA(const float4 separatingNormal, const b3ConvexPolyhedronData_t* hullA,  
+	const float4 posA, const Quaternion ornA, float4* worldVertsB1, int numWorldVertsB1,
+	float4* worldVertsB2, int capacityWorldVertsB2,
+	const float minDist, float maxDist,
+	const float4* verticesA,
+	const b3GpuFace_t* facesA,
+	const int* indicesA,
+	__global const float4* verticesB,
+	__global const b3GpuFace_t* facesB,
+	__global const int* indicesB,
+	float4* contactsOut,
+	int contactCapacity)
+{
+	int numContactsOut = 0;
+
+	float4* pVtxIn = worldVertsB1;
+	float4* pVtxOut = worldVertsB2;
+	
+	int numVertsIn = numWorldVertsB1;
+	int numVertsOut = 0;
+
+	int closestFaceA=-1;
+	{
+		float dmin = FLT_MAX;
+		for(int face=0;face<hullA->m_numFaces;face++)
+		{
+			const float4 Normal = make_float4(
+				facesA[hullA->m_faceOffset+face].m_plane.x, 
+				facesA[hullA->m_faceOffset+face].m_plane.y, 
+				facesA[hullA->m_faceOffset+face].m_plane.z,0.f);
+			const float4 faceANormalWS = qtRotate(ornA,Normal);
+		
+			float d = dot3F4(faceANormalWS,separatingNormal);
+			if (d < dmin)
+			{
+				dmin = d;
+				closestFaceA = face;
+			}
+		}
+	}
+	if (closestFaceA<0)
+		return numContactsOut;
+
+	b3GpuFace_t polyA = facesA[hullA->m_faceOffset+closestFaceA];
+
+	// clip polygon to back of planes of all faces of hull A that are adjacent to witness face
+	int numVerticesA = polyA.m_numIndices;
+	for(int e0=0;e0<numVerticesA;e0++)
+	{
+		const float4 a = verticesA[hullA->m_vertexOffset+indicesA[polyA.m_indexOffset+e0]];
+		const float4 b = verticesA[hullA->m_vertexOffset+indicesA[polyA.m_indexOffset+((e0+1)%numVerticesA)]];
+		const float4 edge0 = a - b;
+		const float4 WorldEdge0 = qtRotate(ornA,edge0);
+		float4 planeNormalA = make_float4(polyA.m_plane.x,polyA.m_plane.y,polyA.m_plane.z,0.f);
+		float4 worldPlaneAnormal1 = qtRotate(ornA,planeNormalA);
+
+		float4 planeNormalWS1 = -cross3(WorldEdge0,worldPlaneAnormal1);
+		float4 worldA1 = transform(&a,&posA,&ornA);
+		float planeEqWS1 = -dot3F4(worldA1,planeNormalWS1);
+		
+		float4 planeNormalWS = planeNormalWS1;
+		float planeEqWS=planeEqWS1;
+		
+		//clip face
+		//clipFace(*pVtxIn, *pVtxOut,planeNormalWS,planeEqWS);
+		numVertsOut = clipFace(pVtxIn, numVertsIn, planeNormalWS,planeEqWS, pVtxOut);
+
+		//btSwap(pVtxIn,pVtxOut);
+		float4* tmp = pVtxOut;
+		pVtxOut = pVtxIn;
+		pVtxIn = tmp;
+		numVertsIn = numVertsOut;
+		numVertsOut = 0;
+	}
+
+	
+	// only keep points that are behind the witness face
+	{
+		float4 localPlaneNormal  = make_float4(polyA.m_plane.x,polyA.m_plane.y,polyA.m_plane.z,0.f);
+		float localPlaneEq = polyA.m_plane.w;
+		float4 planeNormalWS = qtRotate(ornA,localPlaneNormal);
+		float planeEqWS=localPlaneEq-dot3F4(planeNormalWS,posA);
+		for (int i=0;i<numVertsIn;i++)
+		{
+			float depth = dot3F4(planeNormalWS,pVtxIn[i])+planeEqWS;
+			if (depth <=minDist)
+			{
+				depth = minDist;
+			}
+
+			if (depth <=maxDist)
+			{
+				float4 pointInWorld = pVtxIn[i];
+				//resultOut.addContactPoint(separatingNormal,point,depth);
+				contactsOut[numContactsOut++] = make_float4(pointInWorld.x,pointInWorld.y,pointInWorld.z,depth);
+			}
+		}
+	}
+
+	return numContactsOut;
+}
+
+int	clipHullAgainstHull(const float4 separatingNormal,
+	__global const b3ConvexPolyhedronData_t* hullA, __global const b3ConvexPolyhedronData_t* hullB, 
+	const float4 posA, const Quaternion ornA,const float4 posB, const Quaternion ornB, 
+	float4* worldVertsB1, float4* worldVertsB2, int capacityWorldVerts,
+	const float minDist, float maxDist,
+	__global const float4* vertices,
+	__global const b3GpuFace_t* faces,
+	__global const int* indices,
+	float4*	localContactsOut,
+	int localContactCapacity)
+{
+	int numContactsOut = 0;
+	int numWorldVertsB1= 0;
+
+
+	int closestFaceB=-1;
+	float dmax = -FLT_MAX;
+
+	{
+		for(int face=0;face<hullB->m_numFaces;face++)
+		{
+			const float4 Normal = make_float4(faces[hullB->m_faceOffset+face].m_plane.x, 
+				faces[hullB->m_faceOffset+face].m_plane.y, faces[hullB->m_faceOffset+face].m_plane.z,0.f);
+			const float4 WorldNormal = qtRotate(ornB, Normal);
+			float d = dot3F4(WorldNormal,separatingNormal);
+			if (d > dmax)
+			{
+				dmax = d;
+				closestFaceB = face;
+			}
+		}
+	}
+
+	{
+		const b3GpuFace_t polyB = faces[hullB->m_faceOffset+closestFaceB];
+		const int numVertices = polyB.m_numIndices;
+		for(int e0=0;e0<numVertices;e0++)
+		{
+			const float4 b = vertices[hullB->m_vertexOffset+indices[polyB.m_indexOffset+e0]];
+			worldVertsB1[numWorldVertsB1++] = transform(&b,&posB,&ornB);
+		}
+	}
+
+	if (closestFaceB>=0)
+	{
+		numContactsOut = clipFaceAgainstHull(separatingNormal, hullA, 
+				posA,ornA,
+				worldVertsB1,numWorldVertsB1,worldVertsB2,capacityWorldVerts, minDist, maxDist,vertices,
+				faces,
+				indices,localContactsOut,localContactCapacity);
+	}
+
+	return numContactsOut;
+}
+
+
+int	clipHullAgainstHullLocalA(const float4 separatingNormal,
+	const b3ConvexPolyhedronData_t* hullA, __global const b3ConvexPolyhedronData_t* hullB, 
+	const float4 posA, const Quaternion ornA,const float4 posB, const Quaternion ornB, 
+	float4* worldVertsB1, float4* worldVertsB2, int capacityWorldVerts,
+	const float minDist, float maxDist,
+	const float4* verticesA,
+	const b3GpuFace_t* facesA,
+	const int* indicesA,
+	__global const float4* verticesB,
+	__global const b3GpuFace_t* facesB,
+	__global const int* indicesB,
+	float4*	localContactsOut,
+	int localContactCapacity)
+{
+	int numContactsOut = 0;
+	int numWorldVertsB1= 0;
+
+
+	int closestFaceB=-1;
+	float dmax = -FLT_MAX;
+
+	{
+		for(int face=0;face<hullB->m_numFaces;face++)
+		{
+			const float4 Normal = make_float4(facesB[hullB->m_faceOffset+face].m_plane.x, 
+				facesB[hullB->m_faceOffset+face].m_plane.y, facesB[hullB->m_faceOffset+face].m_plane.z,0.f);
+			const float4 WorldNormal = qtRotate(ornB, Normal);
+			float d = dot3F4(WorldNormal,separatingNormal);
+			if (d > dmax)
+			{
+				dmax = d;
+				closestFaceB = face;
+			}
+		}
+	}
+
+	{
+		const b3GpuFace_t polyB = facesB[hullB->m_faceOffset+closestFaceB];
+		const int numVertices = polyB.m_numIndices;
+		for(int e0=0;e0<numVertices;e0++)
+		{
+			const float4 b = verticesB[hullB->m_vertexOffset+indicesB[polyB.m_indexOffset+e0]];
+			worldVertsB1[numWorldVertsB1++] = transform(&b,&posB,&ornB);
+		}
+	}
+
+	if (closestFaceB>=0)
+	{
+		numContactsOut = clipFaceAgainstHullLocalA(separatingNormal, hullA, 
+				posA,ornA,
+				worldVertsB1,numWorldVertsB1,worldVertsB2,capacityWorldVerts, minDist, maxDist,
+				verticesA,facesA,indicesA,
+				verticesB,facesB,indicesB,
+				localContactsOut,localContactCapacity);
+	}
+
+	return numContactsOut;
+}
+
+#define PARALLEL_SUM(v, n) for(int j=1; j<n; j++) v[0] += v[j];
+#define PARALLEL_DO(execution, n) for(int ie=0; ie<n; ie++){execution;}
+#define REDUCE_MAX(v, n) {int i=0;\
+for(int offset=0; offset<n; offset++) v[i] = (v[i].y > v[i+offset].y)? v[i]: v[i+offset]; }
+#define REDUCE_MIN(v, n) {int i=0;\
+for(int offset=0; offset<n; offset++) v[i] = (v[i].y < v[i+offset].y)? v[i]: v[i+offset]; }
+
+int extractManifoldSequentialGlobal(__global const float4* p, int nPoints, float4 nearNormal, int4* contactIdx)
+{
+	if( nPoints == 0 )
+        return 0;
+    
+    if (nPoints <=4)
+        return nPoints;
+    
+    
+    if (nPoints >64)
+        nPoints = 64;
+    
+	float4 center = make_float4(0.f);
+	{
+		
+		for (int i=0;i<nPoints;i++)
+			center += p[i];
+		center /= (float)nPoints;
+	}
+    
+	
+    
+	//	sample 4 directions
+    
+    float4 aVector = p[0] - center;
+    float4 u = cross3( nearNormal, aVector );
+    float4 v = cross3( nearNormal, u );
+    u = normalize3( u );
+    v = normalize3( v );
+    
+    
+    //keep point with deepest penetration
+    float minW= FLT_MAX;
+    
+    int minIndex=-1;
+    
+    float4 maxDots;
+    maxDots.x = FLT_MIN;
+    maxDots.y = FLT_MIN;
+    maxDots.z = FLT_MIN;
+    maxDots.w = FLT_MIN;
+    
+    //	idx, distance
+    for(int ie = 0; ie<nPoints; ie++ )
+    {
+        if (p[ie].w<minW)
+        {
+            minW = p[ie].w;
+            minIndex=ie;
+        }
+        float f;
+        float4 r = p[ie]-center;
+        f = dot3F4( u, r );
+        if (f<maxDots.x)
+        {
+            maxDots.x = f;
+            contactIdx[0].x = ie;
+        }
+        
+        f = dot3F4( -u, r );
+        if (f<maxDots.y)
+        {
+            maxDots.y = f;
+            contactIdx[0].y = ie;
+        }
+        
+        
+        f = dot3F4( v, r );
+        if (f<maxDots.z)
+        {
+            maxDots.z = f;
+            contactIdx[0].z = ie;
+        }
+        
+        f = dot3F4( -v, r );
+        if (f<maxDots.w)
+        {
+            maxDots.w = f;
+            contactIdx[0].w = ie;
+        }
+        
+    }
+    
+    if (contactIdx[0].x != minIndex && contactIdx[0].y != minIndex && contactIdx[0].z != minIndex && contactIdx[0].w != minIndex)
+    {
+        //replace the first contact with minimum (todo: replace contact with least penetration)
+        contactIdx[0].x = minIndex;
+    }
+    
+    return 4;
+    
+}
+
+
+int extractManifoldSequentialGlobalFake(__global const float4* p, int nPoints, float4 nearNormal, int* contactIdx)
+{
+    contactIdx[0] = 0;
+    contactIdx[1] = 1;
+    contactIdx[2] = 2;
+    contactIdx[3] = 3;
+    
+	if( nPoints == 0 ) return 0;
+    
+	nPoints = min2( nPoints, 4 );
+    return nPoints;
+    
+}
+
+
+
+int extractManifoldSequential(const float4* p, int nPoints, float4 nearNormal, int* contactIdx)
+{
+	if( nPoints == 0 ) return 0;
+
+	nPoints = min2( nPoints, 64 );
+
+	float4 center = make_float4(0.f);
+	{
+		float4 v[64];
+		for (int i=0;i<nPoints;i++)
+			v[i] = p[i];
+		//memcpy( v, p, nPoints*sizeof(float4) );
+		PARALLEL_SUM( v, nPoints );
+		center = v[0]/(float)nPoints;
+	}
+
+	
+
+	{	//	sample 4 directions
+		if( nPoints < 4 )
+		{
+			for(int i=0; i<nPoints; i++) 
+				contactIdx[i] = i;
+			return nPoints;
+		}
+
+		float4 aVector = p[0] - center;
+		float4 u = cross3( nearNormal, aVector );
+		float4 v = cross3( nearNormal, u );
+		u = normalize3( u );
+		v = normalize3( v );
+
+		int idx[4];
+
+		float2 max00 = make_float2(0,FLT_MAX);
+		{
+			//	idx, distance
+			{
+				{
+					int4 a[64];
+					for(int ie = 0; ie<nPoints; ie++ )
+					{
+						
+						
+						float f;
+						float4 r = p[ie]-center;
+						f = dot3F4( u, r );
+						a[ie].x = ((*(u32*)&f) & 0xffffff00) | (0xff & ie);
+
+						f = dot3F4( -u, r );
+						a[ie].y = ((*(u32*)&f) & 0xffffff00) | (0xff & ie);
+
+						f = dot3F4( v, r );
+						a[ie].z = ((*(u32*)&f) & 0xffffff00) | (0xff & ie);
+
+						f = dot3F4( -v, r );
+						a[ie].w = ((*(u32*)&f) & 0xffffff00) | (0xff & ie);
+					}
+
+					for(int ie=0; ie<nPoints; ie++)
+					{
+						a[0].x = (a[0].x > a[ie].x )? a[0].x: a[ie].x;
+						a[0].y = (a[0].y > a[ie].y )? a[0].y: a[ie].y;
+						a[0].z = (a[0].z > a[ie].z )? a[0].z: a[ie].z;
+						a[0].w = (a[0].w > a[ie].w )? a[0].w: a[ie].w;
+					}
+
+					idx[0] = (int)a[0].x & 0xff;
+					idx[1] = (int)a[0].y & 0xff;
+					idx[2] = (int)a[0].z & 0xff;
+					idx[3] = (int)a[0].w & 0xff;
+				}
+			}
+
+			{
+				float2 h[64];
+				PARALLEL_DO( h[ie] = make_float2((float)ie, p[ie].w), nPoints );
+				REDUCE_MIN( h, nPoints );
+				max00 = h[0];
+			}
+		}
+
+		contactIdx[0] = idx[0];
+		contactIdx[1] = idx[1];
+		contactIdx[2] = idx[2];
+		contactIdx[3] = idx[3];
+
+
+		return 4;
+	}
+}
+
+
+
+__kernel void   extractManifoldAndAddContactKernel(__global const int4* pairs, 
+																	__global const b3RigidBodyData_t* rigidBodies, 
+																	__global const float4* closestPointsWorld,
+																	__global const float4* separatingNormalsWorld,
+																	__global const int* contactCounts,
+																	__global const int* contactOffsets,
+																	__global struct b3Contact4Data* restrict contactsOut,
+																	counter32_t nContactsOut,
+																	int contactCapacity,
+																	int numPairs,
+																	int pairIndex
+																	)
+{
+	int idx = get_global_id(0);
+	
+	if (idx<numPairs)
+	{
+		float4 normal = separatingNormalsWorld[idx];
+		int nPoints = contactCounts[idx];
+		__global const float4* pointsIn = &closestPointsWorld[contactOffsets[idx]];
+		float4 localPoints[64];
+		for (int i=0;i<nPoints;i++)
+		{
+			localPoints[i] = pointsIn[i];
+		}
+
+		int contactIdx[4];// = {-1,-1,-1,-1};
+		contactIdx[0] = -1;
+		contactIdx[1] = -1;
+		contactIdx[2] = -1;
+		contactIdx[3] = -1;
+
+		int nContacts = extractManifoldSequential(localPoints, nPoints, normal, contactIdx);
+
+		int dstIdx;
+		AppendInc( nContactsOut, dstIdx );
+		if (dstIdx<contactCapacity)
+		{
+			__global struct b3Contact4Data* c = contactsOut + dstIdx;
+			c->m_worldNormalOnB = -normal;
+			c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+			c->m_batchIdx = idx;
+			int bodyA = pairs[pairIndex].x;
+			int bodyB = pairs[pairIndex].y;
+			c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0 ? -bodyA:bodyA;
+			c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0 ? -bodyB:bodyB;
+			c->m_childIndexA = -1;
+			c->m_childIndexB = -1;
+			for (int i=0;i<nContacts;i++)
+			{
+				c->m_worldPosB[i] = localPoints[contactIdx[i]];
+			}
+			GET_NPOINTS(*c) = nContacts;
+		}
+	}
+}
+
+
+void	trInverse(float4 translationIn, Quaternion orientationIn,
+		float4* translationOut, Quaternion* orientationOut)
+{
+	*orientationOut = qtInvert(orientationIn);
+	*translationOut = qtRotate(*orientationOut, -translationIn);
+}
+
+void	trMul(float4 translationA, Quaternion orientationA,
+						float4 translationB, Quaternion orientationB,
+		float4* translationOut, Quaternion* orientationOut)
+{
+	*orientationOut = qtMul(orientationA,orientationB);
+	*translationOut = transform(&translationB,&translationA,&orientationA);
+}
+
+
+
+
+__kernel void   clipHullHullKernel( __global int4* pairs, 
+																					__global const b3RigidBodyData_t* rigidBodies, 
+																					__global const b3Collidable_t* collidables,
+																					__global const b3ConvexPolyhedronData_t* convexShapes, 
+																					__global const float4* vertices,
+																					__global const float4* uniqueEdges,
+																					__global const b3GpuFace_t* faces,
+																					__global const int* indices,
+																					__global const float4* separatingNormals,
+																					__global const int* hasSeparatingAxis,
+																					__global struct b3Contact4Data* restrict globalContactsOut,
+																					counter32_t nGlobalContactsOut,
+																					int numPairs,
+																					int contactCapacity)
+{
+
+	int i = get_global_id(0);
+	int pairIndex = i;
+	
+	float4 worldVertsB1[64];
+	float4 worldVertsB2[64];
+	int capacityWorldVerts = 64;	
+
+	float4 localContactsOut[64];
+	int localContactCapacity=64;
+	
+	float minDist = -1e30f;
+	float maxDist = 0.02f;
+
+	if (i<numPairs)
+	{
+
+		int bodyIndexA = pairs[i].x;
+		int bodyIndexB = pairs[i].y;
+			
+		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+
+		if (hasSeparatingAxis[i])
+		{
+
+			
+			int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+			int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+			
+
+
+		
+			int numLocalContactsOut = clipHullAgainstHull(separatingNormals[i],
+														&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],
+														rigidBodies[bodyIndexA].m_pos,rigidBodies[bodyIndexA].m_quat,
+													  rigidBodies[bodyIndexB].m_pos,rigidBodies[bodyIndexB].m_quat,
+													  worldVertsB1,worldVertsB2,capacityWorldVerts,
+														minDist, maxDist,
+														vertices,faces,indices,
+														localContactsOut,localContactCapacity);
+												
+		if (numLocalContactsOut>0)
+		{
+				float4 normal = -separatingNormals[i];
+				int nPoints = numLocalContactsOut;
+				float4* pointsIn = localContactsOut;
+				int contactIdx[4];// = {-1,-1,-1,-1};
+
+				contactIdx[0] = -1;
+				contactIdx[1] = -1;
+				contactIdx[2] = -1;
+				contactIdx[3] = -1;
+		
+				int nReducedContacts = extractManifoldSequential(pointsIn, nPoints, normal, contactIdx);
+		
+				
+				int mprContactIndex = pairs[pairIndex].z;
+
+				int dstIdx = mprContactIndex;
+				if (dstIdx<0)
+				{
+					AppendInc( nGlobalContactsOut, dstIdx );
+				}
+
+				if (dstIdx<contactCapacity)
+				{
+					pairs[pairIndex].z = dstIdx;
+
+					__global struct b3Contact4Data* c = globalContactsOut+ dstIdx;
+					c->m_worldNormalOnB = -normal;
+					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+					c->m_batchIdx = pairIndex;
+					int bodyA = pairs[pairIndex].x;
+					int bodyB = pairs[pairIndex].y;
+					c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;
+					c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;
+					c->m_childIndexA = -1;
+					c->m_childIndexB = -1;
+
+					for (int i=0;i<nReducedContacts;i++)
+					{
+					//this condition means: overwrite contact point, unless at index i==0 we have a valid 'mpr' contact
+						if (i>0||(mprContactIndex<0))
+						{
+							c->m_worldPosB[i] = pointsIn[contactIdx[i]];
+						}
+					}
+					GET_NPOINTS(*c) = nReducedContacts;
+				}
+				
+			}//		if (numContactsOut>0)
+		}//		if (hasSeparatingAxis[i])
+	}//	if (i<numPairs)
+
+}
+
+
+__kernel void   clipCompoundsHullHullKernel( __global const int4* gpuCompoundPairs, 
+																					__global const b3RigidBodyData_t* rigidBodies, 
+																					__global const b3Collidable_t* collidables,
+																					__global const b3ConvexPolyhedronData_t* convexShapes, 
+																					__global const float4* vertices,
+																					__global const float4* uniqueEdges,
+																					__global const b3GpuFace_t* faces,
+																					__global const int* indices,
+																					__global const b3GpuChildShape_t* gpuChildShapes,
+																					__global const float4* gpuCompoundSepNormalsOut,
+																					__global const int* gpuHasCompoundSepNormalsOut,
+																					__global struct b3Contact4Data* restrict globalContactsOut,
+																					counter32_t nGlobalContactsOut,
+																					int numCompoundPairs, int maxContactCapacity)
+{
+
+	int i = get_global_id(0);
+	int pairIndex = i;
+	
+	float4 worldVertsB1[64];
+	float4 worldVertsB2[64];
+	int capacityWorldVerts = 64;	
+
+	float4 localContactsOut[64];
+	int localContactCapacity=64;
+	
+	float minDist = -1e30f;
+	float maxDist = 0.02f;
+
+	if (i<numCompoundPairs)
+	{
+
+		if (gpuHasCompoundSepNormalsOut[i])
+		{
+
+			int bodyIndexA = gpuCompoundPairs[i].x;
+			int bodyIndexB = gpuCompoundPairs[i].y;
+			
+			int childShapeIndexA = gpuCompoundPairs[i].z;
+			int childShapeIndexB = gpuCompoundPairs[i].w;
+			
+			int collidableIndexA = -1;
+			int collidableIndexB = -1;
+			
+			float4 ornA = rigidBodies[bodyIndexA].m_quat;
+			float4 posA = rigidBodies[bodyIndexA].m_pos;
+			
+			float4 ornB = rigidBodies[bodyIndexB].m_quat;
+			float4 posB = rigidBodies[bodyIndexB].m_pos;
+								
+			if (childShapeIndexA >= 0)
+			{
+				collidableIndexA = gpuChildShapes[childShapeIndexA].m_shapeIndex;
+				float4 childPosA = gpuChildShapes[childShapeIndexA].m_childPosition;
+				float4 childOrnA = gpuChildShapes[childShapeIndexA].m_childOrientation;
+				float4 newPosA = qtRotate(ornA,childPosA)+posA;
+				float4 newOrnA = qtMul(ornA,childOrnA);
+				posA = newPosA;
+				ornA = newOrnA;
+			} else
+			{
+				collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+			}
+			
+			if (childShapeIndexB>=0)
+			{
+				collidableIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+				float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+				float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+				float4 newPosB = transform(&childPosB,&posB,&ornB);
+				float4 newOrnB = qtMul(ornB,childOrnB);
+				posB = newPosB;
+				ornB = newOrnB;
+			} else
+			{
+				collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;	
+			}
+			
+			int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+			int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+		
+			int numLocalContactsOut = clipHullAgainstHull(gpuCompoundSepNormalsOut[i],
+														&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],
+														posA,ornA,
+													  posB,ornB,
+													  worldVertsB1,worldVertsB2,capacityWorldVerts,
+														minDist, maxDist,
+														vertices,faces,indices,
+														localContactsOut,localContactCapacity);
+												
+		if (numLocalContactsOut>0)
+		{
+				float4 normal = -gpuCompoundSepNormalsOut[i];
+				int nPoints = numLocalContactsOut;
+				float4* pointsIn = localContactsOut;
+				int contactIdx[4];// = {-1,-1,-1,-1};
+
+				contactIdx[0] = -1;
+				contactIdx[1] = -1;
+				contactIdx[2] = -1;
+				contactIdx[3] = -1;
+		
+				int nReducedContacts = extractManifoldSequential(pointsIn, nPoints, normal, contactIdx);
+		
+				int dstIdx;
+				AppendInc( nGlobalContactsOut, dstIdx );
+				if ((dstIdx+nReducedContacts) < maxContactCapacity)
+				{
+					__global struct b3Contact4Data* c = globalContactsOut+ dstIdx;
+					c->m_worldNormalOnB = -normal;
+					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+					c->m_batchIdx = pairIndex;
+					int bodyA = gpuCompoundPairs[pairIndex].x;
+					int bodyB = gpuCompoundPairs[pairIndex].y;
+					c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;
+					c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;
+					c->m_childIndexA = childShapeIndexA;
+					c->m_childIndexB = childShapeIndexB;
+					for (int i=0;i<nReducedContacts;i++)
+					{
+						c->m_worldPosB[i] = pointsIn[contactIdx[i]];
+					}
+					GET_NPOINTS(*c) = nReducedContacts;
+				}
+				
+			}//		if (numContactsOut>0)
+		}//		if (gpuHasCompoundSepNormalsOut[i])
+	}//	if (i<numCompoundPairs)
+
+}
+
+
+
+__kernel void   sphereSphereCollisionKernel( __global const int4* pairs, 
+																					__global const b3RigidBodyData_t* rigidBodies, 
+																					__global const b3Collidable_t* collidables,
+																					__global const float4* separatingNormals,
+																					__global const int* hasSeparatingAxis,
+																					__global struct b3Contact4Data* restrict globalContactsOut,
+																					counter32_t nGlobalContactsOut,
+																					int contactCapacity,
+																					int numPairs)
+{
+
+	int i = get_global_id(0);
+	int pairIndex = i;
+	
+	if (i<numPairs)
+	{
+		int bodyIndexA = pairs[i].x;
+		int bodyIndexB = pairs[i].y;
+			
+		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+
+		if (collidables[collidableIndexA].m_shapeType == SHAPE_SPHERE &&
+			collidables[collidableIndexB].m_shapeType == SHAPE_SPHERE)
+		{
+			//sphere-sphere
+			float radiusA = collidables[collidableIndexA].m_radius;
+			float radiusB = collidables[collidableIndexB].m_radius;
+			float4 posA = rigidBodies[bodyIndexA].m_pos;
+			float4 posB = rigidBodies[bodyIndexB].m_pos;
+
+			float4 diff = posA-posB;
+			float len = length(diff);
+			
+			///iff distance positive, don't generate a new contact
+			if ( len <= (radiusA+radiusB))
+			{
+				///distance (negative means penetration)
+				float dist = len - (radiusA+radiusB);
+				float4 normalOnSurfaceB = make_float4(1.f,0.f,0.f,0.f);
+				if (len > 0.00001)
+				{
+					normalOnSurfaceB = diff / len;
+				}
+				float4 contactPosB = posB + normalOnSurfaceB*radiusB;
+				contactPosB.w = dist;
+								
+				int dstIdx;
+				AppendInc( nGlobalContactsOut, dstIdx );
+				if (dstIdx < contactCapacity)
+				{
+					__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];
+					c->m_worldNormalOnB = -normalOnSurfaceB;
+					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+					c->m_batchIdx = pairIndex;
+					int bodyA = pairs[pairIndex].x;
+					int bodyB = pairs[pairIndex].y;
+					c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;
+					c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;
+					c->m_worldPosB[0] = contactPosB;
+					c->m_childIndexA = -1;
+					c->m_childIndexB = -1;
+
+					GET_NPOINTS(*c) = 1;
+				}//if (dstIdx < numPairs)
+			}//if ( len <= (radiusA+radiusB))
+		}//SHAPE_SPHERE SHAPE_SPHERE
+	}//if (i<numPairs)
+}				
+
+__kernel void   clipHullHullConcaveConvexKernel( __global int4* concavePairsIn,
+																					__global const b3RigidBodyData_t* rigidBodies, 
+																					__global const b3Collidable_t* collidables,
+																					__global const b3ConvexPolyhedronData_t* convexShapes, 
+																					__global const float4* vertices,
+																					__global const float4* uniqueEdges,
+																					__global const b3GpuFace_t* faces,
+																					__global const int* indices,
+																					__global const b3GpuChildShape_t* gpuChildShapes,
+																					__global const float4* separatingNormals,
+																					__global struct b3Contact4Data* restrict globalContactsOut,
+																					counter32_t nGlobalContactsOut,
+																					int contactCapacity,
+																					int numConcavePairs)
+{
+
+	int i = get_global_id(0);
+	int pairIndex = i;
+	
+	float4 worldVertsB1[64];
+	float4 worldVertsB2[64];
+	int capacityWorldVerts = 64;	
+
+	float4 localContactsOut[64];
+	int localContactCapacity=64;
+	
+	float minDist = -1e30f;
+	float maxDist = 0.02f;
+
+	if (i<numConcavePairs)
+	{
+		//negative value means that the pair is invalid
+		if (concavePairsIn[i].w<0)
+			return;
+
+		int bodyIndexA = concavePairsIn[i].x;
+		int bodyIndexB = concavePairsIn[i].y;
+		int f = concavePairsIn[i].z;
+		int childShapeIndexA = f;
+		
+		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+		
+		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+		
+		///////////////////////////////////////////////////////////////
+		
+	
+		bool overlap = false;
+		
+		b3ConvexPolyhedronData_t convexPolyhedronA;
+
+	//add 3 vertices of the triangle
+		convexPolyhedronA.m_numVertices = 3;
+		convexPolyhedronA.m_vertexOffset = 0;
+		float4	localCenter = make_float4(0.f,0.f,0.f,0.f);
+
+		b3GpuFace_t face = faces[convexShapes[shapeIndexA].m_faceOffset+f];
+		
+		float4 verticesA[3];
+		for (int i=0;i<3;i++)
+		{
+			int index = indices[face.m_indexOffset+i];
+			float4 vert = vertices[convexShapes[shapeIndexA].m_vertexOffset+index];
+			verticesA[i] = vert;
+			localCenter += vert;
+		}
+
+		float dmin = FLT_MAX;
+
+		int localCC=0;
+
+		//a triangle has 3 unique edges
+		convexPolyhedronA.m_numUniqueEdges = 3;
+		convexPolyhedronA.m_uniqueEdgesOffset = 0;
+		float4 uniqueEdgesA[3];
+		
+		uniqueEdgesA[0] = (verticesA[1]-verticesA[0]);
+		uniqueEdgesA[1] = (verticesA[2]-verticesA[1]);
+		uniqueEdgesA[2] = (verticesA[0]-verticesA[2]);
+
+
+		convexPolyhedronA.m_faceOffset = 0;
+                                  
+		float4 normal = make_float4(face.m_plane.x,face.m_plane.y,face.m_plane.z,0.f);
+                             
+		b3GpuFace_t facesA[TRIANGLE_NUM_CONVEX_FACES];
+		int indicesA[3+3+2+2+2];
+		int curUsedIndices=0;
+		int fidx=0;
+
+		//front size of triangle
+		{
+			facesA[fidx].m_indexOffset=curUsedIndices;
+			indicesA[0] = 0;
+			indicesA[1] = 1;
+			indicesA[2] = 2;
+			curUsedIndices+=3;
+			float c = face.m_plane.w;
+			facesA[fidx].m_plane.x = normal.x;
+			facesA[fidx].m_plane.y = normal.y;
+			facesA[fidx].m_plane.z = normal.z;
+			facesA[fidx].m_plane.w = c;
+			facesA[fidx].m_numIndices=3;
+		}
+		fidx++;
+		//back size of triangle
+		{
+			facesA[fidx].m_indexOffset=curUsedIndices;
+			indicesA[3]=2;
+			indicesA[4]=1;
+			indicesA[5]=0;
+			curUsedIndices+=3;
+			float c = dot3F4(normal,verticesA[0]);
+			float c1 = -face.m_plane.w;
+			facesA[fidx].m_plane.x = -normal.x;
+			facesA[fidx].m_plane.y = -normal.y;
+			facesA[fidx].m_plane.z = -normal.z;
+			facesA[fidx].m_plane.w = c;
+			facesA[fidx].m_numIndices=3;
+		}
+		fidx++;
+
+		bool addEdgePlanes = true;
+		if (addEdgePlanes)
+		{
+			int numVertices=3;
+			int prevVertex = numVertices-1;
+			for (int i=0;i<numVertices;i++)
+			{
+				float4 v0 = verticesA[i];
+				float4 v1 = verticesA[prevVertex];
+                                            
+				float4 edgeNormal = normalize(cross(normal,v1-v0));
+				float c = -dot3F4(edgeNormal,v0);
+
+				facesA[fidx].m_numIndices = 2;
+				facesA[fidx].m_indexOffset=curUsedIndices;
+				indicesA[curUsedIndices++]=i;
+				indicesA[curUsedIndices++]=prevVertex;
+                                            
+				facesA[fidx].m_plane.x = edgeNormal.x;
+				facesA[fidx].m_plane.y = edgeNormal.y;
+				facesA[fidx].m_plane.z = edgeNormal.z;
+				facesA[fidx].m_plane.w = c;
+				fidx++;
+				prevVertex = i;
+			}
+		}
+		convexPolyhedronA.m_numFaces = TRIANGLE_NUM_CONVEX_FACES;
+		convexPolyhedronA.m_localCenter = localCenter*(1.f/3.f);
+
+
+		float4 posA = rigidBodies[bodyIndexA].m_pos;
+		posA.w = 0.f;
+		float4 posB = rigidBodies[bodyIndexB].m_pos;
+		posB.w = 0.f;
+		float4 ornA = rigidBodies[bodyIndexA].m_quat;
+		float4 ornB =rigidBodies[bodyIndexB].m_quat;
+
+
+		float4 sepAxis = separatingNormals[i];
+		
+		int shapeTypeB = collidables[collidableIndexB].m_shapeType;
+		int childShapeIndexB =-1;
+		if (shapeTypeB==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+		{
+			///////////////////
+			///compound shape support
+			
+			childShapeIndexB = concavePairsIn[pairIndex].w;
+			int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+			shapeIndexB = collidables[childColIndexB].m_shapeIndex;
+			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+			float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+			float4 newPosB = transform(&childPosB,&posB,&ornB);
+			float4 newOrnB = qtMul(ornB,childOrnB);
+			posB = newPosB;
+			ornB = newOrnB;
+			
+		}
+		
+		////////////////////////////////////////
+		
+		
+		
+		int numLocalContactsOut = clipHullAgainstHullLocalA(sepAxis,
+														&convexPolyhedronA, &convexShapes[shapeIndexB],
+														posA,ornA,
+													  posB,ornB,
+													  worldVertsB1,worldVertsB2,capacityWorldVerts,
+														minDist, maxDist,
+														&verticesA,&facesA,&indicesA,
+														vertices,faces,indices,
+														localContactsOut,localContactCapacity);
+												
+		if (numLocalContactsOut>0)
+		{
+			float4 normal = -separatingNormals[i];
+			int nPoints = numLocalContactsOut;
+			float4* pointsIn = localContactsOut;
+			int contactIdx[4];// = {-1,-1,-1,-1};
+
+			contactIdx[0] = -1;
+			contactIdx[1] = -1;
+			contactIdx[2] = -1;
+			contactIdx[3] = -1;
+	
+			int nReducedContacts = extractManifoldSequential(pointsIn, nPoints, normal, contactIdx);
+	
+			int dstIdx;
+			AppendInc( nGlobalContactsOut, dstIdx );
+			if (dstIdx<contactCapacity)
+			{
+				__global struct b3Contact4Data* c = globalContactsOut+ dstIdx;
+				c->m_worldNormalOnB = -normal;
+				c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+				c->m_batchIdx = pairIndex;
+				int bodyA = concavePairsIn[pairIndex].x;
+				int bodyB = concavePairsIn[pairIndex].y;
+				c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;
+				c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;
+				c->m_childIndexA = childShapeIndexA;
+				c->m_childIndexB = childShapeIndexB;
+				for (int i=0;i<nReducedContacts;i++)
+				{
+					c->m_worldPosB[i] = pointsIn[contactIdx[i]];
+				}
+				GET_NPOINTS(*c) = nReducedContacts;
+			}
+				
+		}//		if (numContactsOut>0)
+	}//	if (i<numPairs)
+}
+
+
+
+
+
+
+int	findClippingFaces(const float4 separatingNormal,
+                      __global const b3ConvexPolyhedronData_t* hullA, __global const b3ConvexPolyhedronData_t* hullB,
+                      const float4 posA, const Quaternion ornA,const float4 posB, const Quaternion ornB,
+                       __global float4* worldVertsA1,
+                      __global float4* worldNormalsA1,
+                      __global float4* worldVertsB1,
+                      int capacityWorldVerts,
+                      const float minDist, float maxDist,
+                      __global const float4* vertices,
+                      __global const b3GpuFace_t* faces,
+                      __global const int* indices,
+                      __global int4* clippingFaces, int pairIndex)
+{
+	int numContactsOut = 0;
+	int numWorldVertsB1= 0;
+    
+    
+	int closestFaceB=-1;
+	float dmax = -FLT_MAX;
+    
+	{
+		for(int face=0;face<hullB->m_numFaces;face++)
+		{
+			const float4 Normal = make_float4(faces[hullB->m_faceOffset+face].m_plane.x,
+                                              faces[hullB->m_faceOffset+face].m_plane.y, faces[hullB->m_faceOffset+face].m_plane.z,0.f);
+			const float4 WorldNormal = qtRotate(ornB, Normal);
+			float d = dot3F4(WorldNormal,separatingNormal);
+			if (d > dmax)
+			{
+				dmax = d;
+				closestFaceB = face;
+			}
+		}
+	}
+    
+	{
+		const b3GpuFace_t polyB = faces[hullB->m_faceOffset+closestFaceB];
+		const int numVertices = polyB.m_numIndices;
+		for(int e0=0;e0<numVertices;e0++)
+		{
+			const float4 b = vertices[hullB->m_vertexOffset+indices[polyB.m_indexOffset+e0]];
+			worldVertsB1[pairIndex*capacityWorldVerts+numWorldVertsB1++] = transform(&b,&posB,&ornB);
+		}
+	}
+    
+    int closestFaceA=-1;
+	{
+		float dmin = FLT_MAX;
+		for(int face=0;face<hullA->m_numFaces;face++)
+		{
+			const float4 Normal = make_float4(
+                                              faces[hullA->m_faceOffset+face].m_plane.x,
+                                              faces[hullA->m_faceOffset+face].m_plane.y,
+                                              faces[hullA->m_faceOffset+face].m_plane.z,
+                                              0.f);
+			const float4 faceANormalWS = qtRotate(ornA,Normal);
+            
+			float d = dot3F4(faceANormalWS,separatingNormal);
+			if (d < dmin)
+			{
+				dmin = d;
+				closestFaceA = face;
+                worldNormalsA1[pairIndex] = faceANormalWS;
+			}
+		}
+	}
+    
+    int numVerticesA = faces[hullA->m_faceOffset+closestFaceA].m_numIndices;
+	for(int e0=0;e0<numVerticesA;e0++)
+	{
+        const float4 a = vertices[hullA->m_vertexOffset+indices[faces[hullA->m_faceOffset+closestFaceA].m_indexOffset+e0]];
+        worldVertsA1[pairIndex*capacityWorldVerts+e0] = transform(&a, &posA,&ornA);
+    }
+    
+    clippingFaces[pairIndex].x = closestFaceA;
+    clippingFaces[pairIndex].y = closestFaceB;
+    clippingFaces[pairIndex].z = numVerticesA;
+    clippingFaces[pairIndex].w = numWorldVertsB1;
+    
+    
+	return numContactsOut;
+}
+
+
+
+int clipFaces(__global float4* worldVertsA1,
+              __global float4* worldNormalsA1,
+              __global float4* worldVertsB1,
+              __global float4* worldVertsB2, 
+              int capacityWorldVertsB2,
+              const float minDist, float maxDist,
+              __global int4* clippingFaces,
+              int pairIndex)
+{
+	int numContactsOut = 0;
+    
+    int closestFaceA = clippingFaces[pairIndex].x;
+    int closestFaceB = clippingFaces[pairIndex].y;
+	int numVertsInA = clippingFaces[pairIndex].z;
+	int numVertsInB = clippingFaces[pairIndex].w;
+    
+	int numVertsOut = 0;
+    
+	if (closestFaceA<0)
+		return numContactsOut;
+    
+    __global float4* pVtxIn = &worldVertsB1[pairIndex*capacityWorldVertsB2];
+    __global float4* pVtxOut = &worldVertsB2[pairIndex*capacityWorldVertsB2];
+    
+    
+	
+	// clip polygon to back of planes of all faces of hull A that are adjacent to witness face
+    
+	for(int e0=0;e0<numVertsInA;e0++)
+	{
+		const float4 aw = worldVertsA1[pairIndex*capacityWorldVertsB2+e0];
+		const float4 bw = worldVertsA1[pairIndex*capacityWorldVertsB2+((e0+1)%numVertsInA)];
+		const float4 WorldEdge0 = aw - bw;
+		float4 worldPlaneAnormal1 = worldNormalsA1[pairIndex];
+		float4 planeNormalWS1 = -cross3(WorldEdge0,worldPlaneAnormal1);
+		float4 worldA1 = aw;
+		float planeEqWS1 = -dot3F4(worldA1,planeNormalWS1);
+		float4 planeNormalWS = planeNormalWS1;
+		float planeEqWS=planeEqWS1;
+		numVertsOut = clipFaceGlobal(pVtxIn, numVertsInB, planeNormalWS,planeEqWS, pVtxOut);
+		__global float4* tmp = pVtxOut;
+		pVtxOut = pVtxIn;
+		pVtxIn = tmp;
+		numVertsInB = numVertsOut;
+		numVertsOut = 0;
+	}
+    
+    //float4 planeNormalWS = worldNormalsA1[pairIndex];
+    //float planeEqWS=-dot3F4(planeNormalWS,worldVertsA1[pairIndex*capacityWorldVertsB2]);
+
+
+    
+    /*for (int i=0;i<numVertsInB;i++)
+    {
+        pVtxOut[i] = pVtxIn[i];
+    }*/
+    
+    
+    
+    
+    //numVertsInB=0;
+	
+    float4 planeNormalWS = worldNormalsA1[pairIndex];
+    float planeEqWS=-dot3F4(planeNormalWS,worldVertsA1[pairIndex*capacityWorldVertsB2]);
+
+    for (int i=0;i<numVertsInB;i++)
+    {
+        float depth = dot3F4(planeNormalWS,pVtxIn[i])+planeEqWS;
+        if (depth <=minDist)
+        {
+            depth = minDist;
+        }
+        
+        if (depth <=maxDist)
+        {
+            float4 pointInWorld = pVtxIn[i];
+            pVtxOut[numContactsOut++] = make_float4(pointInWorld.x,pointInWorld.y,pointInWorld.z,depth);
+        }
+    }
+   
+    clippingFaces[pairIndex].w =numContactsOut;
+   
+    
+	return numContactsOut;
+
+}
+
+
+
+
+__kernel void   findClippingFacesKernel(  __global const int4* pairs,
+                                        __global const b3RigidBodyData_t* rigidBodies,
+                                        __global const b3Collidable_t* collidables,
+                                        __global const b3ConvexPolyhedronData_t* convexShapes,
+                                        __global const float4* vertices,
+                                        __global const float4* uniqueEdges,
+                                        __global const b3GpuFace_t* faces,
+                                        __global const int* indices,
+                                        __global const float4* separatingNormals,
+                                        __global const int* hasSeparatingAxis,
+                                        __global int4* clippingFacesOut,
+                                        __global float4* worldVertsA1,
+                                        __global float4* worldNormalsA1,
+                                        __global float4* worldVertsB1,
+                                        int capacityWorldVerts,
+                                        int numPairs
+                                        )
+{
+    
+	int i = get_global_id(0);
+	int pairIndex = i;
+    
+	
+	float minDist = -1e30f;
+	float maxDist = 0.02f;
+    
+	if (i<numPairs)
+	{
+        
+		if (hasSeparatingAxis[i])
+		{
+            
+			int bodyIndexA = pairs[i].x;
+			int bodyIndexB = pairs[i].y;
+			
+			int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+			int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+			
+			int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+			int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+			
+            
+            
+			int numLocalContactsOut = findClippingFaces(separatingNormals[i],
+                                                        &convexShapes[shapeIndexA], &convexShapes[shapeIndexB],
+                                                        rigidBodies[bodyIndexA].m_pos,rigidBodies[bodyIndexA].m_quat,
+                                                        rigidBodies[bodyIndexB].m_pos,rigidBodies[bodyIndexB].m_quat,
+                                                        worldVertsA1,
+                                                        worldNormalsA1,
+                                                        worldVertsB1,capacityWorldVerts,
+                                                        minDist, maxDist,
+                                                        vertices,faces,indices,
+                                                        clippingFacesOut,i);
+            
+            
+		}//		if (hasSeparatingAxis[i])
+	}//	if (i<numPairs)
+    
+}
+
+
+
+
+__kernel void   clipFacesAndFindContactsKernel(    __global const float4* separatingNormals,
+                                                   __global const int* hasSeparatingAxis,
+                                                   __global int4* clippingFacesOut,
+                                                   __global float4* worldVertsA1,
+                                                   __global float4* worldNormalsA1,
+                                                   __global float4* worldVertsB1,
+                                                   __global float4* worldVertsB2,
+                                                    int vertexFaceCapacity,
+                                                   int numPairs,
+					                                        int debugMode
+                                                   )
+{
+    int i = get_global_id(0);
+	int pairIndex = i;
+	
+    
+	float minDist = -1e30f;
+	float maxDist = 0.02f;
+    
+	if (i<numPairs)
+	{
+        
+		if (hasSeparatingAxis[i])
+		{
+            
+//			int bodyIndexA = pairs[i].x;
+	//		int bodyIndexB = pairs[i].y;
+		    
+            int numLocalContactsOut = 0;
+
+            int capacityWorldVertsB2 = vertexFaceCapacity;
+            
+            __global float4* pVtxIn = &worldVertsB1[pairIndex*capacityWorldVertsB2];
+            __global float4* pVtxOut = &worldVertsB2[pairIndex*capacityWorldVertsB2];
+            
+
+            {
+                __global int4* clippingFaces = clippingFacesOut;
+            
+                
+                int closestFaceA = clippingFaces[pairIndex].x;
+                int closestFaceB = clippingFaces[pairIndex].y;
+                int numVertsInA = clippingFaces[pairIndex].z;
+                int numVertsInB = clippingFaces[pairIndex].w;
+                
+                int numVertsOut = 0;
+                
+                if (closestFaceA>=0)
+                {
+                    
+                    
+                    
+                    // clip polygon to back of planes of all faces of hull A that are adjacent to witness face
+                    
+                    for(int e0=0;e0<numVertsInA;e0++)
+                    {
+                        const float4 aw = worldVertsA1[pairIndex*capacityWorldVertsB2+e0];
+                        const float4 bw = worldVertsA1[pairIndex*capacityWorldVertsB2+((e0+1)%numVertsInA)];
+                        const float4 WorldEdge0 = aw - bw;
+                        float4 worldPlaneAnormal1 = worldNormalsA1[pairIndex];
+                        float4 planeNormalWS1 = -cross3(WorldEdge0,worldPlaneAnormal1);
+                        float4 worldA1 = aw;
+                        float planeEqWS1 = -dot3F4(worldA1,planeNormalWS1);
+                        float4 planeNormalWS = planeNormalWS1;
+                        float planeEqWS=planeEqWS1;
+                        numVertsOut = clipFaceGlobal(pVtxIn, numVertsInB, planeNormalWS,planeEqWS, pVtxOut);
+                        __global float4* tmp = pVtxOut;
+                        pVtxOut = pVtxIn;
+                        pVtxIn = tmp;
+                        numVertsInB = numVertsOut;
+                        numVertsOut = 0;
+                    }
+                    
+                    float4 planeNormalWS = worldNormalsA1[pairIndex];
+                    float planeEqWS=-dot3F4(planeNormalWS,worldVertsA1[pairIndex*capacityWorldVertsB2]);
+                    
+                    for (int i=0;i<numVertsInB;i++)
+                    {
+                        float depth = dot3F4(planeNormalWS,pVtxIn[i])+planeEqWS;
+                        if (depth <=minDist)
+                        {
+                            depth = minDist;
+                        }
+                        
+                        if (depth <=maxDist)
+                        {
+                            float4 pointInWorld = pVtxIn[i];
+                            pVtxOut[numLocalContactsOut++] = make_float4(pointInWorld.x,pointInWorld.y,pointInWorld.z,depth);
+                        }
+                    }
+                    
+                }
+                clippingFaces[pairIndex].w =numLocalContactsOut;
+                
+
+            }
+            
+            for (int i=0;i<numLocalContactsOut;i++)
+                pVtxIn[i] = pVtxOut[i];
+                
+		}//		if (hasSeparatingAxis[i])
+	}//	if (i<numPairs)
+    
+}
+
+
+
+
+
+__kernel void   newContactReductionKernel( __global int4* pairs,
+                                                   __global const b3RigidBodyData_t* rigidBodies,
+                                                   __global const float4* separatingNormals,
+                                                   __global const int* hasSeparatingAxis,
+                                                   __global struct b3Contact4Data* globalContactsOut,
+                                                   __global int4* clippingFaces,
+                                                   __global float4* worldVertsB2,
+                                                   volatile __global int* nGlobalContactsOut,
+                                                   int vertexFaceCapacity,
+												   int contactCapacity,
+                                                   int numPairs
+                                                   )
+{
+    int i = get_global_id(0);
+	int pairIndex = i;
+	
+    int4 contactIdx;
+    contactIdx=make_int4(0,1,2,3);
+    
+	if (i<numPairs)
+	{
+        
+		if (hasSeparatingAxis[i])
+		{
+            
+			
+            
+            
+			int nPoints = clippingFaces[pairIndex].w;
+           
+            if (nPoints>0)
+            {
+
+                 __global float4* pointsIn = &worldVertsB2[pairIndex*vertexFaceCapacity];
+                float4 normal = -separatingNormals[i];
+                
+                int nReducedContacts = extractManifoldSequentialGlobal(pointsIn, nPoints, normal, &contactIdx);
+            
+				int mprContactIndex = pairs[pairIndex].z;
+
+                int dstIdx = mprContactIndex;
+
+				if (dstIdx<0)
+				{
+	                AppendInc( nGlobalContactsOut, dstIdx );
+				}
+//#if 0
+                
+				if (dstIdx < contactCapacity)
+				{
+
+					__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];
+					c->m_worldNormalOnB = -normal;
+					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);
+					c->m_batchIdx = pairIndex;
+					int bodyA = pairs[pairIndex].x;
+					int bodyB = pairs[pairIndex].y;
+
+					pairs[pairIndex].w = dstIdx;
+
+					c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;
+					c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;
+                    c->m_childIndexA =-1;
+					c->m_childIndexB =-1;
+
+                    switch (nReducedContacts)
+                    {
+                        case 4:
+                            c->m_worldPosB[3] = pointsIn[contactIdx.w];
+                        case 3:
+                            c->m_worldPosB[2] = pointsIn[contactIdx.z];
+                        case 2:
+                            c->m_worldPosB[1] = pointsIn[contactIdx.y];
+                        case 1:
+							if (mprContactIndex<0)//test
+	                            c->m_worldPosB[0] = pointsIn[contactIdx.x];
+                        default:
+                        {
+                        }
+                    };
+                    
+					GET_NPOINTS(*c) = nReducedContacts;
+                    
+                 }
+                 
+                
+//#endif
+				
+			}//		if (numContactsOut>0)
+		}//		if (hasSeparatingAxis[i])
+	}//	if (i<numPairs)
+
+    
+    
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satClipHullContacts.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satClipHullContacts.h
new file mode 100644
index 0000000000..f0ecfc7851
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satClipHullContacts.h
@@ -0,0 +1,2099 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* satClipKernelsCL= \
+"#define TRIANGLE_NUM_CONVEX_FACES 5\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile __global int*\n"
+"#endif\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"typedef unsigned int u32;\n"
+"#ifndef B3_CONTACT4DATA_H\n"
+"#define B3_CONTACT4DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#define B3_FLOAT4_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#define B3_PLATFORM_DEFINITIONS_H\n"
+"struct MyTest\n"
+"{\n"
+"	int bla;\n"
+"};\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"//keep B3_LARGE_FLOAT*B3_LARGE_FLOAT < FLT_MAX\n"
+"#define B3_LARGE_FLOAT 1e18f\n"
+"#define B3_INFINITY 1e18f\n"
+"#define b3Assert(a)\n"
+"#define b3ConstArray(a) __global const a*\n"
+"#define b3AtomicInc atomic_inc\n"
+"#define b3AtomicAdd atomic_add\n"
+"#define b3Fabs fabs\n"
+"#define b3Sqrt native_sqrt\n"
+"#define b3Sin native_sin\n"
+"#define b3Cos native_cos\n"
+"#define B3_STATIC\n"
+"#endif\n"
+"#endif\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Float4;\n"
+"	#define b3Float4ConstArg const b3Float4\n"
+"	#define b3MakeFloat4 (float4)\n"
+"	float b3Dot3F4(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return dot(a1, b1);\n"
+"	}\n"
+"	b3Float4 b3Cross3(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return cross(a1, b1);\n"
+"	}\n"
+"	#define b3MinFloat4 min\n"
+"	#define b3MaxFloat4 max\n"
+"	#define b3Normalized(a) normalize(a)\n"
+"#endif \n"
+"		\n"
+"inline bool b3IsAlmostZero(b3Float4ConstArg v)\n"
+"{\n"
+"	if(b3Fabs(v.x)>1e-6 || b3Fabs(v.y)>1e-6 || b3Fabs(v.z)>1e-6)	\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"inline int    b3MaxDot( b3Float4ConstArg vec, __global const b3Float4* vecArray, int vecLen, float* dotOut )\n"
+"{\n"
+"    float maxDot = -B3_INFINITY;\n"
+"    int i = 0;\n"
+"    int ptIndex = -1;\n"
+"    for( i = 0; i < vecLen; i++ )\n"
+"    {\n"
+"        float dot = b3Dot3F4(vecArray[i],vec);\n"
+"            \n"
+"        if( dot > maxDot )\n"
+"        {\n"
+"            maxDot = dot;\n"
+"            ptIndex = i;\n"
+"        }\n"
+"    }\n"
+"	b3Assert(ptIndex>=0);\n"
+"    if (ptIndex<0)\n"
+"	{\n"
+"		ptIndex = 0;\n"
+"	}\n"
+"    *dotOut = maxDot;\n"
+"    return ptIndex;\n"
+"}\n"
+"#endif //B3_FLOAT4_H\n"
+"typedef  struct b3Contact4Data b3Contact4Data_t;\n"
+"struct b3Contact4Data\n"
+"{\n"
+"	b3Float4	m_worldPosB[4];\n"
+"//	b3Float4	m_localPosA[4];\n"
+"//	b3Float4	m_localPosB[4];\n"
+"	b3Float4	m_worldNormalOnB;	//	w: m_nPoints\n"
+"	unsigned short  m_restituitionCoeffCmp;\n"
+"	unsigned short  m_frictionCoeffCmp;\n"
+"	int m_batchIdx;\n"
+"	int m_bodyAPtrAndSignBit;//x:m_bodyAPtr, y:m_bodyBPtr\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"	int	m_childIndexA;\n"
+"	int	m_childIndexB;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"};\n"
+"inline int b3Contact4Data_getNumPoints(const struct b3Contact4Data* contact)\n"
+"{\n"
+"	return (int)contact->m_worldNormalOnB.w;\n"
+"};\n"
+"inline void b3Contact4Data_setNumPoints(struct b3Contact4Data* contact, int numPoints)\n"
+"{\n"
+"	contact->m_worldNormalOnB.w = (float)numPoints;\n"
+"};\n"
+"#endif //B3_CONTACT4DATA_H\n"
+"#ifndef B3_CONVEX_POLYHEDRON_DATA_H\n"
+"#define B3_CONVEX_POLYHEDRON_DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_QUAT_H\n"
+"#define B3_QUAT_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif\n"
+"#endif\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Quat;\n"
+"	#define b3QuatConstArg const b3Quat\n"
+"	\n"
+"	\n"
+"inline float4 b3FastNormalize4(float4 v)\n"
+"{\n"
+"	v = (float4)(v.xyz,0.f);\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"	\n"
+"inline b3Quat b3QuatMul(b3Quat a, b3Quat b);\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in);\n"
+"inline b3Quat b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec);\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatMul(b3QuatConstArg a, b3QuatConstArg b)\n"
+"{\n"
+"	b3Quat ans;\n"
+"	ans = b3Cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - b3Dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in)\n"
+"{\n"
+"	b3Quat q;\n"
+"	q=in;\n"
+"	//return b3FastNormalize4(in);\n"
+"	float len = native_sqrt(dot(q, q));\n"
+"	if(len > 0.f)\n"
+"	{\n"
+"		q *= 1.f / len;\n"
+"	}\n"
+"	else\n"
+"	{\n"
+"		q.x = q.y = q.z = 0.f;\n"
+"		q.w = 1.f;\n"
+"	}\n"
+"	return q;\n"
+"}\n"
+"inline float4 b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	b3Quat qInv = b3QuatInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = b3QuatMul(b3QuatMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline float4 b3QuatInvRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	return b3QuatRotate( b3QuatInvert( q ), vec );\n"
+"}\n"
+"inline b3Float4 b3TransformPoint(b3Float4ConstArg point, b3Float4ConstArg translation, b3QuatConstArg  orientation)\n"
+"{\n"
+"	return b3QuatRotate( orientation, point ) + (translation);\n"
+"}\n"
+"	\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"typedef struct b3GpuFace b3GpuFace_t;\n"
+"struct b3GpuFace\n"
+"{\n"
+"	b3Float4 m_plane;\n"
+"	int m_indexOffset;\n"
+"	int m_numIndices;\n"
+"	int m_unusedPadding1;\n"
+"	int m_unusedPadding2;\n"
+"};\n"
+"typedef struct b3ConvexPolyhedronData b3ConvexPolyhedronData_t;\n"
+"struct b3ConvexPolyhedronData\n"
+"{\n"
+"	b3Float4		m_localCenter;\n"
+"	b3Float4		m_extents;\n"
+"	b3Float4		mC;\n"
+"	b3Float4		mE;\n"
+"	float			m_radius;\n"
+"	int	m_faceOffset;\n"
+"	int m_numFaces;\n"
+"	int	m_numVertices;\n"
+"	int m_vertexOffset;\n"
+"	int	m_uniqueEdgesOffset;\n"
+"	int	m_numUniqueEdges;\n"
+"	int m_unused;\n"
+"};\n"
+"#endif //B3_CONVEX_POLYHEDRON_DATA_H\n"
+"#ifndef B3_COLLIDABLE_H\n"
+"#define B3_COLLIDABLE_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"enum b3ShapeTypes\n"
+"{\n"
+"	SHAPE_HEIGHT_FIELD=1,\n"
+"	SHAPE_CONVEX_HULL=3,\n"
+"	SHAPE_PLANE=4,\n"
+"	SHAPE_CONCAVE_TRIMESH=5,\n"
+"	SHAPE_COMPOUND_OF_CONVEX_HULLS=6,\n"
+"	SHAPE_SPHERE=7,\n"
+"	MAX_NUM_SHAPE_TYPES,\n"
+"};\n"
+"typedef struct b3Collidable b3Collidable_t;\n"
+"struct b3Collidable\n"
+"{\n"
+"	union {\n"
+"		int m_numChildShapes;\n"
+"		int m_bvhIndex;\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float m_radius;\n"
+"		int	m_compoundBvhIndex;\n"
+"	};\n"
+"	int m_shapeType;\n"
+"	int m_shapeIndex;\n"
+"};\n"
+"typedef struct b3GpuChildShape b3GpuChildShape_t;\n"
+"struct b3GpuChildShape\n"
+"{\n"
+"	b3Float4	m_childPosition;\n"
+"	b3Quat		m_childOrientation;\n"
+"	int m_shapeIndex;\n"
+"	int m_unused0;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"};\n"
+"struct b3CompoundOverlappingPair\n"
+"{\n"
+"	int m_bodyIndexA;\n"
+"	int m_bodyIndexB;\n"
+"//	int	m_pairType;\n"
+"	int m_childShapeIndexA;\n"
+"	int m_childShapeIndexB;\n"
+"};\n"
+"#endif //B3_COLLIDABLE_H\n"
+"#ifndef B3_RIGIDBODY_DATA_H\n"
+"#define B3_RIGIDBODY_DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"#ifndef B3_MAT3x3_H\n"
+"#define B3_MAT3x3_H\n"
+"#ifndef B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"typedef struct\n"
+"{\n"
+"	b3Float4 m_row[3];\n"
+"}b3Mat3x3;\n"
+"#define b3Mat3x3ConstArg const b3Mat3x3\n"
+"#define b3GetRow(m,row) (m.m_row[row])\n"
+"inline b3Mat3x3 b3QuatGetRotationMatrix(b3Quat quat)\n"
+"{\n"
+"	b3Float4 quat2 = (b3Float4)(quat.x*quat.x, quat.y*quat.y, quat.z*quat.z, 0.f);\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0].x=1-2*quat2.y-2*quat2.z;\n"
+"	out.m_row[0].y=2*quat.x*quat.y-2*quat.w*quat.z;\n"
+"	out.m_row[0].z=2*quat.x*quat.z+2*quat.w*quat.y;\n"
+"	out.m_row[0].w = 0.f;\n"
+"	out.m_row[1].x=2*quat.x*quat.y+2*quat.w*quat.z;\n"
+"	out.m_row[1].y=1-2*quat2.x-2*quat2.z;\n"
+"	out.m_row[1].z=2*quat.y*quat.z-2*quat.w*quat.x;\n"
+"	out.m_row[1].w = 0.f;\n"
+"	out.m_row[2].x=2*quat.x*quat.z-2*quat.w*quat.y;\n"
+"	out.m_row[2].y=2*quat.y*quat.z+2*quat.w*quat.x;\n"
+"	out.m_row[2].z=1-2*quat2.x-2*quat2.y;\n"
+"	out.m_row[2].w = 0.f;\n"
+"	return out;\n"
+"}\n"
+"inline b3Mat3x3 b3AbsoluteMat3x3(b3Mat3x3ConstArg matIn)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = fabs(matIn.m_row[0]);\n"
+"	out.m_row[1] = fabs(matIn.m_row[1]);\n"
+"	out.m_row[2] = fabs(matIn.m_row[2]);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtZero();\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity();\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m);\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b);\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Mat3x3 mtZero()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(0.f);\n"
+"	m.m_row[1] = (b3Float4)(0.f);\n"
+"	m.m_row[2] = (b3Float4)(0.f);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(1,0,0,0);\n"
+"	m.m_row[1] = (b3Float4)(0,1,0,0);\n"
+"	m.m_row[2] = (b3Float4)(0,0,1,0);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = (b3Float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (b3Float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (b3Float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Mat3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	b3Mat3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = b3Dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = b3Dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = b3Dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b)\n"
+"{\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a.m_row[0], b );\n"
+"	ans.y = b3Dot3F4( a.m_row[1], b );\n"
+"	ans.z = b3Dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Float4 colx = b3MakeFloat4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	b3Float4 coly = b3MakeFloat4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	b3Float4 colz = b3MakeFloat4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a, colx );\n"
+"	ans.y = b3Dot3F4( a, coly );\n"
+"	ans.z = b3Dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"#endif\n"
+"#endif //B3_MAT3x3_H\n"
+"typedef struct b3RigidBodyData b3RigidBodyData_t;\n"
+"struct b3RigidBodyData\n"
+"{\n"
+"	b3Float4				m_pos;\n"
+"	b3Quat					m_quat;\n"
+"	b3Float4				m_linVel;\n"
+"	b3Float4				m_angVel;\n"
+"	int 					m_collidableIdx;\n"
+"	float 				m_invMass;\n"
+"	float 				m_restituitionCoeff;\n"
+"	float 				m_frictionCoeff;\n"
+"};\n"
+"typedef struct b3InertiaData b3InertiaData_t;\n"
+"struct b3InertiaData\n"
+"{\n"
+"	b3Mat3x3 m_invInertiaWorld;\n"
+"	b3Mat3x3 m_initInvInertia;\n"
+"};\n"
+"#endif //B3_RIGIDBODY_DATA_H\n"
+"	\n"
+"#define GET_NPOINTS(x) (x).m_worldNormalOnB.w\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"#define make_float4 (float4)\n"
+"#define make_float2 (float2)\n"
+"#define make_uint4 (uint4)\n"
+"#define make_int4 (int4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"__inline\n"
+"float fastDiv(float numerator, float denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"//	return numerator/denominator;	\n"
+"}\n"
+"__inline\n"
+"float4 fastDiv4(float4 numerator, float4 denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"}\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"//#define dot3F4 dot\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = make_float4(a.xyz,0.f);\n"
+"	float4 b1 = make_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"typedef float4 Quaternion;\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b);\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in);\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec);\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q);\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in)\n"
+"{\n"
+"	return fastNormalize4(in);\n"
+"//	in /= length( in );\n"
+"//	return in;\n"
+"}\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(qtMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"__inline\n"
+"float4 qtInvRotate(const Quaternion q, float4 vec)\n"
+"{\n"
+"	return qtRotate( qtInvert( q ), vec );\n"
+"}\n"
+"__inline\n"
+"float4 transform(const float4* p, const float4* translation, const Quaternion* orientation)\n"
+"{\n"
+"	return qtRotate( *orientation, *p ) + (*translation);\n"
+"}\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = make_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"}\n"
+"__inline float4 lerp3(const float4 a,const float4 b, float  t)\n"
+"{\n"
+"	return make_float4(	a.x + (b.x - a.x) * t,\n"
+"						a.y + (b.y - a.y) * t,\n"
+"						a.z + (b.z - a.z) * t,\n"
+"						0.f);\n"
+"}\n"
+"// Clips a face to the back of a plane, return the number of vertices out, stored in ppVtxOut\n"
+"int clipFaceGlobal(__global const float4* pVtxIn, int numVertsIn, float4 planeNormalWS,float planeEqWS, __global float4* ppVtxOut)\n"
+"{\n"
+"	\n"
+"	int ve;\n"
+"	float ds, de;\n"
+"	int numVertsOut = 0;\n"
+"    //double-check next test\n"
+"    	if (numVertsIn < 2)\n"
+"    		return 0;\n"
+"    \n"
+"	float4 firstVertex=pVtxIn[numVertsIn-1];\n"
+"	float4 endVertex = pVtxIn[0];\n"
+"	\n"
+"	ds = dot3F4(planeNormalWS,firstVertex)+planeEqWS;\n"
+"    \n"
+"	for (ve = 0; ve < numVertsIn; ve++)\n"
+"	{\n"
+"		endVertex=pVtxIn[ve];\n"
+"		de = dot3F4(planeNormalWS,endVertex)+planeEqWS;\n"
+"		if (ds<0)\n"
+"		{\n"
+"			if (de<0)\n"
+"			{\n"
+"				// Start < 0, end < 0, so output endVertex\n"
+"				ppVtxOut[numVertsOut++] = endVertex;\n"
+"			}\n"
+"			else\n"
+"			{\n"
+"				// Start < 0, end >= 0, so output intersection\n"
+"				ppVtxOut[numVertsOut++] = lerp3(firstVertex, endVertex,(ds * 1.f/(ds - de)) );\n"
+"			}\n"
+"		}\n"
+"		else\n"
+"		{\n"
+"			if (de<0)\n"
+"			{\n"
+"				// Start >= 0, end < 0 so output intersection and end\n"
+"				ppVtxOut[numVertsOut++] = lerp3(firstVertex, endVertex,(ds * 1.f/(ds - de)) );\n"
+"				ppVtxOut[numVertsOut++] = endVertex;\n"
+"			}\n"
+"		}\n"
+"		firstVertex = endVertex;\n"
+"		ds = de;\n"
+"	}\n"
+"	return numVertsOut;\n"
+"}\n"
+"// Clips a face to the back of a plane, return the number of vertices out, stored in ppVtxOut\n"
+"int clipFace(const float4* pVtxIn, int numVertsIn, float4 planeNormalWS,float planeEqWS, float4* ppVtxOut)\n"
+"{\n"
+"	\n"
+"	int ve;\n"
+"	float ds, de;\n"
+"	int numVertsOut = 0;\n"
+"//double-check next test\n"
+"	if (numVertsIn < 2)\n"
+"		return 0;\n"
+"	float4 firstVertex=pVtxIn[numVertsIn-1];\n"
+"	float4 endVertex = pVtxIn[0];\n"
+"	\n"
+"	ds = dot3F4(planeNormalWS,firstVertex)+planeEqWS;\n"
+"	for (ve = 0; ve < numVertsIn; ve++)\n"
+"	{\n"
+"		endVertex=pVtxIn[ve];\n"
+"		de = dot3F4(planeNormalWS,endVertex)+planeEqWS;\n"
+"		if (ds<0)\n"
+"		{\n"
+"			if (de<0)\n"
+"			{\n"
+"				// Start < 0, end < 0, so output endVertex\n"
+"				ppVtxOut[numVertsOut++] = endVertex;\n"
+"			}\n"
+"			else\n"
+"			{\n"
+"				// Start < 0, end >= 0, so output intersection\n"
+"				ppVtxOut[numVertsOut++] = lerp3(firstVertex, endVertex,(ds * 1.f/(ds - de)) );\n"
+"			}\n"
+"		}\n"
+"		else\n"
+"		{\n"
+"			if (de<0)\n"
+"			{\n"
+"				// Start >= 0, end < 0 so output intersection and end\n"
+"				ppVtxOut[numVertsOut++] = lerp3(firstVertex, endVertex,(ds * 1.f/(ds - de)) );\n"
+"				ppVtxOut[numVertsOut++] = endVertex;\n"
+"			}\n"
+"		}\n"
+"		firstVertex = endVertex;\n"
+"		ds = de;\n"
+"	}\n"
+"	return numVertsOut;\n"
+"}\n"
+"int clipFaceAgainstHull(const float4 separatingNormal, __global const b3ConvexPolyhedronData_t* hullA,  \n"
+"	const float4 posA, const Quaternion ornA, float4* worldVertsB1, int numWorldVertsB1,\n"
+"	float4* worldVertsB2, int capacityWorldVertsB2,\n"
+"	const float minDist, float maxDist,\n"
+"	__global const float4* vertices,\n"
+"	__global const b3GpuFace_t* faces,\n"
+"	__global const int* indices,\n"
+"	float4* contactsOut,\n"
+"	int contactCapacity)\n"
+"{\n"
+"	int numContactsOut = 0;\n"
+"	float4* pVtxIn = worldVertsB1;\n"
+"	float4* pVtxOut = worldVertsB2;\n"
+"	\n"
+"	int numVertsIn = numWorldVertsB1;\n"
+"	int numVertsOut = 0;\n"
+"	int closestFaceA=-1;\n"
+"	{\n"
+"		float dmin = FLT_MAX;\n"
+"		for(int face=0;face<hullA->m_numFaces;face++)\n"
+"		{\n"
+"			const float4 Normal = make_float4(\n"
+"				faces[hullA->m_faceOffset+face].m_plane.x, \n"
+"				faces[hullA->m_faceOffset+face].m_plane.y, \n"
+"				faces[hullA->m_faceOffset+face].m_plane.z,0.f);\n"
+"			const float4 faceANormalWS = qtRotate(ornA,Normal);\n"
+"		\n"
+"			float d = dot3F4(faceANormalWS,separatingNormal);\n"
+"			if (d < dmin)\n"
+"			{\n"
+"				dmin = d;\n"
+"				closestFaceA = face;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	if (closestFaceA<0)\n"
+"		return numContactsOut;\n"
+"	b3GpuFace_t polyA = faces[hullA->m_faceOffset+closestFaceA];\n"
+"	// clip polygon to back of planes of all faces of hull A that are adjacent to witness face\n"
+"	int numVerticesA = polyA.m_numIndices;\n"
+"	for(int e0=0;e0<numVerticesA;e0++)\n"
+"	{\n"
+"		const float4 a = vertices[hullA->m_vertexOffset+indices[polyA.m_indexOffset+e0]];\n"
+"		const float4 b = vertices[hullA->m_vertexOffset+indices[polyA.m_indexOffset+((e0+1)%numVerticesA)]];\n"
+"		const float4 edge0 = a - b;\n"
+"		const float4 WorldEdge0 = qtRotate(ornA,edge0);\n"
+"		float4 planeNormalA = make_float4(polyA.m_plane.x,polyA.m_plane.y,polyA.m_plane.z,0.f);\n"
+"		float4 worldPlaneAnormal1 = qtRotate(ornA,planeNormalA);\n"
+"		float4 planeNormalWS1 = -cross3(WorldEdge0,worldPlaneAnormal1);\n"
+"		float4 worldA1 = transform(&a,&posA,&ornA);\n"
+"		float planeEqWS1 = -dot3F4(worldA1,planeNormalWS1);\n"
+"		\n"
+"		float4 planeNormalWS = planeNormalWS1;\n"
+"		float planeEqWS=planeEqWS1;\n"
+"		\n"
+"		//clip face\n"
+"		//clipFace(*pVtxIn, *pVtxOut,planeNormalWS,planeEqWS);\n"
+"		numVertsOut = clipFace(pVtxIn, numVertsIn, planeNormalWS,planeEqWS, pVtxOut);\n"
+"		//btSwap(pVtxIn,pVtxOut);\n"
+"		float4* tmp = pVtxOut;\n"
+"		pVtxOut = pVtxIn;\n"
+"		pVtxIn = tmp;\n"
+"		numVertsIn = numVertsOut;\n"
+"		numVertsOut = 0;\n"
+"	}\n"
+"	\n"
+"	// only keep points that are behind the witness face\n"
+"	{\n"
+"		float4 localPlaneNormal  = make_float4(polyA.m_plane.x,polyA.m_plane.y,polyA.m_plane.z,0.f);\n"
+"		float localPlaneEq = polyA.m_plane.w;\n"
+"		float4 planeNormalWS = qtRotate(ornA,localPlaneNormal);\n"
+"		float planeEqWS=localPlaneEq-dot3F4(planeNormalWS,posA);\n"
+"		for (int i=0;i<numVertsIn;i++)\n"
+"		{\n"
+"			float depth = dot3F4(planeNormalWS,pVtxIn[i])+planeEqWS;\n"
+"			if (depth <=minDist)\n"
+"			{\n"
+"				depth = minDist;\n"
+"			}\n"
+"			if (depth <=maxDist)\n"
+"			{\n"
+"				float4 pointInWorld = pVtxIn[i];\n"
+"				//resultOut.addContactPoint(separatingNormal,point,depth);\n"
+"				contactsOut[numContactsOut++] = make_float4(pointInWorld.x,pointInWorld.y,pointInWorld.z,depth);\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	return numContactsOut;\n"
+"}\n"
+"int clipFaceAgainstHullLocalA(const float4 separatingNormal, const b3ConvexPolyhedronData_t* hullA,  \n"
+"	const float4 posA, const Quaternion ornA, float4* worldVertsB1, int numWorldVertsB1,\n"
+"	float4* worldVertsB2, int capacityWorldVertsB2,\n"
+"	const float minDist, float maxDist,\n"
+"	const float4* verticesA,\n"
+"	const b3GpuFace_t* facesA,\n"
+"	const int* indicesA,\n"
+"	__global const float4* verticesB,\n"
+"	__global const b3GpuFace_t* facesB,\n"
+"	__global const int* indicesB,\n"
+"	float4* contactsOut,\n"
+"	int contactCapacity)\n"
+"{\n"
+"	int numContactsOut = 0;\n"
+"	float4* pVtxIn = worldVertsB1;\n"
+"	float4* pVtxOut = worldVertsB2;\n"
+"	\n"
+"	int numVertsIn = numWorldVertsB1;\n"
+"	int numVertsOut = 0;\n"
+"	int closestFaceA=-1;\n"
+"	{\n"
+"		float dmin = FLT_MAX;\n"
+"		for(int face=0;face<hullA->m_numFaces;face++)\n"
+"		{\n"
+"			const float4 Normal = make_float4(\n"
+"				facesA[hullA->m_faceOffset+face].m_plane.x, \n"
+"				facesA[hullA->m_faceOffset+face].m_plane.y, \n"
+"				facesA[hullA->m_faceOffset+face].m_plane.z,0.f);\n"
+"			const float4 faceANormalWS = qtRotate(ornA,Normal);\n"
+"		\n"
+"			float d = dot3F4(faceANormalWS,separatingNormal);\n"
+"			if (d < dmin)\n"
+"			{\n"
+"				dmin = d;\n"
+"				closestFaceA = face;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	if (closestFaceA<0)\n"
+"		return numContactsOut;\n"
+"	b3GpuFace_t polyA = facesA[hullA->m_faceOffset+closestFaceA];\n"
+"	// clip polygon to back of planes of all faces of hull A that are adjacent to witness face\n"
+"	int numVerticesA = polyA.m_numIndices;\n"
+"	for(int e0=0;e0<numVerticesA;e0++)\n"
+"	{\n"
+"		const float4 a = verticesA[hullA->m_vertexOffset+indicesA[polyA.m_indexOffset+e0]];\n"
+"		const float4 b = verticesA[hullA->m_vertexOffset+indicesA[polyA.m_indexOffset+((e0+1)%numVerticesA)]];\n"
+"		const float4 edge0 = a - b;\n"
+"		const float4 WorldEdge0 = qtRotate(ornA,edge0);\n"
+"		float4 planeNormalA = make_float4(polyA.m_plane.x,polyA.m_plane.y,polyA.m_plane.z,0.f);\n"
+"		float4 worldPlaneAnormal1 = qtRotate(ornA,planeNormalA);\n"
+"		float4 planeNormalWS1 = -cross3(WorldEdge0,worldPlaneAnormal1);\n"
+"		float4 worldA1 = transform(&a,&posA,&ornA);\n"
+"		float planeEqWS1 = -dot3F4(worldA1,planeNormalWS1);\n"
+"		\n"
+"		float4 planeNormalWS = planeNormalWS1;\n"
+"		float planeEqWS=planeEqWS1;\n"
+"		\n"
+"		//clip face\n"
+"		//clipFace(*pVtxIn, *pVtxOut,planeNormalWS,planeEqWS);\n"
+"		numVertsOut = clipFace(pVtxIn, numVertsIn, planeNormalWS,planeEqWS, pVtxOut);\n"
+"		//btSwap(pVtxIn,pVtxOut);\n"
+"		float4* tmp = pVtxOut;\n"
+"		pVtxOut = pVtxIn;\n"
+"		pVtxIn = tmp;\n"
+"		numVertsIn = numVertsOut;\n"
+"		numVertsOut = 0;\n"
+"	}\n"
+"	\n"
+"	// only keep points that are behind the witness face\n"
+"	{\n"
+"		float4 localPlaneNormal  = make_float4(polyA.m_plane.x,polyA.m_plane.y,polyA.m_plane.z,0.f);\n"
+"		float localPlaneEq = polyA.m_plane.w;\n"
+"		float4 planeNormalWS = qtRotate(ornA,localPlaneNormal);\n"
+"		float planeEqWS=localPlaneEq-dot3F4(planeNormalWS,posA);\n"
+"		for (int i=0;i<numVertsIn;i++)\n"
+"		{\n"
+"			float depth = dot3F4(planeNormalWS,pVtxIn[i])+planeEqWS;\n"
+"			if (depth <=minDist)\n"
+"			{\n"
+"				depth = minDist;\n"
+"			}\n"
+"			if (depth <=maxDist)\n"
+"			{\n"
+"				float4 pointInWorld = pVtxIn[i];\n"
+"				//resultOut.addContactPoint(separatingNormal,point,depth);\n"
+"				contactsOut[numContactsOut++] = make_float4(pointInWorld.x,pointInWorld.y,pointInWorld.z,depth);\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	return numContactsOut;\n"
+"}\n"
+"int	clipHullAgainstHull(const float4 separatingNormal,\n"
+"	__global const b3ConvexPolyhedronData_t* hullA, __global const b3ConvexPolyhedronData_t* hullB, \n"
+"	const float4 posA, const Quaternion ornA,const float4 posB, const Quaternion ornB, \n"
+"	float4* worldVertsB1, float4* worldVertsB2, int capacityWorldVerts,\n"
+"	const float minDist, float maxDist,\n"
+"	__global const float4* vertices,\n"
+"	__global const b3GpuFace_t* faces,\n"
+"	__global const int* indices,\n"
+"	float4*	localContactsOut,\n"
+"	int localContactCapacity)\n"
+"{\n"
+"	int numContactsOut = 0;\n"
+"	int numWorldVertsB1= 0;\n"
+"	int closestFaceB=-1;\n"
+"	float dmax = -FLT_MAX;\n"
+"	{\n"
+"		for(int face=0;face<hullB->m_numFaces;face++)\n"
+"		{\n"
+"			const float4 Normal = make_float4(faces[hullB->m_faceOffset+face].m_plane.x, \n"
+"				faces[hullB->m_faceOffset+face].m_plane.y, faces[hullB->m_faceOffset+face].m_plane.z,0.f);\n"
+"			const float4 WorldNormal = qtRotate(ornB, Normal);\n"
+"			float d = dot3F4(WorldNormal,separatingNormal);\n"
+"			if (d > dmax)\n"
+"			{\n"
+"				dmax = d;\n"
+"				closestFaceB = face;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	{\n"
+"		const b3GpuFace_t polyB = faces[hullB->m_faceOffset+closestFaceB];\n"
+"		const int numVertices = polyB.m_numIndices;\n"
+"		for(int e0=0;e0<numVertices;e0++)\n"
+"		{\n"
+"			const float4 b = vertices[hullB->m_vertexOffset+indices[polyB.m_indexOffset+e0]];\n"
+"			worldVertsB1[numWorldVertsB1++] = transform(&b,&posB,&ornB);\n"
+"		}\n"
+"	}\n"
+"	if (closestFaceB>=0)\n"
+"	{\n"
+"		numContactsOut = clipFaceAgainstHull(separatingNormal, hullA, \n"
+"				posA,ornA,\n"
+"				worldVertsB1,numWorldVertsB1,worldVertsB2,capacityWorldVerts, minDist, maxDist,vertices,\n"
+"				faces,\n"
+"				indices,localContactsOut,localContactCapacity);\n"
+"	}\n"
+"	return numContactsOut;\n"
+"}\n"
+"int	clipHullAgainstHullLocalA(const float4 separatingNormal,\n"
+"	const b3ConvexPolyhedronData_t* hullA, __global const b3ConvexPolyhedronData_t* hullB, \n"
+"	const float4 posA, const Quaternion ornA,const float4 posB, const Quaternion ornB, \n"
+"	float4* worldVertsB1, float4* worldVertsB2, int capacityWorldVerts,\n"
+"	const float minDist, float maxDist,\n"
+"	const float4* verticesA,\n"
+"	const b3GpuFace_t* facesA,\n"
+"	const int* indicesA,\n"
+"	__global const float4* verticesB,\n"
+"	__global const b3GpuFace_t* facesB,\n"
+"	__global const int* indicesB,\n"
+"	float4*	localContactsOut,\n"
+"	int localContactCapacity)\n"
+"{\n"
+"	int numContactsOut = 0;\n"
+"	int numWorldVertsB1= 0;\n"
+"	int closestFaceB=-1;\n"
+"	float dmax = -FLT_MAX;\n"
+"	{\n"
+"		for(int face=0;face<hullB->m_numFaces;face++)\n"
+"		{\n"
+"			const float4 Normal = make_float4(facesB[hullB->m_faceOffset+face].m_plane.x, \n"
+"				facesB[hullB->m_faceOffset+face].m_plane.y, facesB[hullB->m_faceOffset+face].m_plane.z,0.f);\n"
+"			const float4 WorldNormal = qtRotate(ornB, Normal);\n"
+"			float d = dot3F4(WorldNormal,separatingNormal);\n"
+"			if (d > dmax)\n"
+"			{\n"
+"				dmax = d;\n"
+"				closestFaceB = face;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	{\n"
+"		const b3GpuFace_t polyB = facesB[hullB->m_faceOffset+closestFaceB];\n"
+"		const int numVertices = polyB.m_numIndices;\n"
+"		for(int e0=0;e0<numVertices;e0++)\n"
+"		{\n"
+"			const float4 b = verticesB[hullB->m_vertexOffset+indicesB[polyB.m_indexOffset+e0]];\n"
+"			worldVertsB1[numWorldVertsB1++] = transform(&b,&posB,&ornB);\n"
+"		}\n"
+"	}\n"
+"	if (closestFaceB>=0)\n"
+"	{\n"
+"		numContactsOut = clipFaceAgainstHullLocalA(separatingNormal, hullA, \n"
+"				posA,ornA,\n"
+"				worldVertsB1,numWorldVertsB1,worldVertsB2,capacityWorldVerts, minDist, maxDist,\n"
+"				verticesA,facesA,indicesA,\n"
+"				verticesB,facesB,indicesB,\n"
+"				localContactsOut,localContactCapacity);\n"
+"	}\n"
+"	return numContactsOut;\n"
+"}\n"
+"#define PARALLEL_SUM(v, n) for(int j=1; j<n; j++) v[0] += v[j];\n"
+"#define PARALLEL_DO(execution, n) for(int ie=0; ie<n; ie++){execution;}\n"
+"#define REDUCE_MAX(v, n) {int i=0; for(int offset=0; offset<n; offset++) v[i] = (v[i].y > v[i+offset].y)? v[i]: v[i+offset]; }\n"
+"#define REDUCE_MIN(v, n) {int i=0; for(int offset=0; offset<n; offset++) v[i] = (v[i].y < v[i+offset].y)? v[i]: v[i+offset]; }\n"
+"int extractManifoldSequentialGlobal(__global const float4* p, int nPoints, float4 nearNormal, int4* contactIdx)\n"
+"{\n"
+"	if( nPoints == 0 )\n"
+"        return 0;\n"
+"    \n"
+"    if (nPoints <=4)\n"
+"        return nPoints;\n"
+"    \n"
+"    \n"
+"    if (nPoints >64)\n"
+"        nPoints = 64;\n"
+"    \n"
+"	float4 center = make_float4(0.f);\n"
+"	{\n"
+"		\n"
+"		for (int i=0;i<nPoints;i++)\n"
+"			center += p[i];\n"
+"		center /= (float)nPoints;\n"
+"	}\n"
+"    \n"
+"	\n"
+"    \n"
+"	//	sample 4 directions\n"
+"    \n"
+"    float4 aVector = p[0] - center;\n"
+"    float4 u = cross3( nearNormal, aVector );\n"
+"    float4 v = cross3( nearNormal, u );\n"
+"    u = normalize3( u );\n"
+"    v = normalize3( v );\n"
+"    \n"
+"    \n"
+"    //keep point with deepest penetration\n"
+"    float minW= FLT_MAX;\n"
+"    \n"
+"    int minIndex=-1;\n"
+"    \n"
+"    float4 maxDots;\n"
+"    maxDots.x = FLT_MIN;\n"
+"    maxDots.y = FLT_MIN;\n"
+"    maxDots.z = FLT_MIN;\n"
+"    maxDots.w = FLT_MIN;\n"
+"    \n"
+"    //	idx, distance\n"
+"    for(int ie = 0; ie<nPoints; ie++ )\n"
+"    {\n"
+"        if (p[ie].w<minW)\n"
+"        {\n"
+"            minW = p[ie].w;\n"
+"            minIndex=ie;\n"
+"        }\n"
+"        float f;\n"
+"        float4 r = p[ie]-center;\n"
+"        f = dot3F4( u, r );\n"
+"        if (f<maxDots.x)\n"
+"        {\n"
+"            maxDots.x = f;\n"
+"            contactIdx[0].x = ie;\n"
+"        }\n"
+"        \n"
+"        f = dot3F4( -u, r );\n"
+"        if (f<maxDots.y)\n"
+"        {\n"
+"            maxDots.y = f;\n"
+"            contactIdx[0].y = ie;\n"
+"        }\n"
+"        \n"
+"        \n"
+"        f = dot3F4( v, r );\n"
+"        if (f<maxDots.z)\n"
+"        {\n"
+"            maxDots.z = f;\n"
+"            contactIdx[0].z = ie;\n"
+"        }\n"
+"        \n"
+"        f = dot3F4( -v, r );\n"
+"        if (f<maxDots.w)\n"
+"        {\n"
+"            maxDots.w = f;\n"
+"            contactIdx[0].w = ie;\n"
+"        }\n"
+"        \n"
+"    }\n"
+"    \n"
+"    if (contactIdx[0].x != minIndex && contactIdx[0].y != minIndex && contactIdx[0].z != minIndex && contactIdx[0].w != minIndex)\n"
+"    {\n"
+"        //replace the first contact with minimum (todo: replace contact with least penetration)\n"
+"        contactIdx[0].x = minIndex;\n"
+"    }\n"
+"    \n"
+"    return 4;\n"
+"    \n"
+"}\n"
+"int extractManifoldSequentialGlobalFake(__global const float4* p, int nPoints, float4 nearNormal, int* contactIdx)\n"
+"{\n"
+"    contactIdx[0] = 0;\n"
+"    contactIdx[1] = 1;\n"
+"    contactIdx[2] = 2;\n"
+"    contactIdx[3] = 3;\n"
+"    \n"
+"	if( nPoints == 0 ) return 0;\n"
+"    \n"
+"	nPoints = min2( nPoints, 4 );\n"
+"    return nPoints;\n"
+"    \n"
+"}\n"
+"int extractManifoldSequential(const float4* p, int nPoints, float4 nearNormal, int* contactIdx)\n"
+"{\n"
+"	if( nPoints == 0 ) return 0;\n"
+"	nPoints = min2( nPoints, 64 );\n"
+"	float4 center = make_float4(0.f);\n"
+"	{\n"
+"		float4 v[64];\n"
+"		for (int i=0;i<nPoints;i++)\n"
+"			v[i] = p[i];\n"
+"		//memcpy( v, p, nPoints*sizeof(float4) );\n"
+"		PARALLEL_SUM( v, nPoints );\n"
+"		center = v[0]/(float)nPoints;\n"
+"	}\n"
+"	\n"
+"	{	//	sample 4 directions\n"
+"		if( nPoints < 4 )\n"
+"		{\n"
+"			for(int i=0; i<nPoints; i++) \n"
+"				contactIdx[i] = i;\n"
+"			return nPoints;\n"
+"		}\n"
+"		float4 aVector = p[0] - center;\n"
+"		float4 u = cross3( nearNormal, aVector );\n"
+"		float4 v = cross3( nearNormal, u );\n"
+"		u = normalize3( u );\n"
+"		v = normalize3( v );\n"
+"		int idx[4];\n"
+"		float2 max00 = make_float2(0,FLT_MAX);\n"
+"		{\n"
+"			//	idx, distance\n"
+"			{\n"
+"				{\n"
+"					int4 a[64];\n"
+"					for(int ie = 0; ie<nPoints; ie++ )\n"
+"					{\n"
+"						\n"
+"						\n"
+"						float f;\n"
+"						float4 r = p[ie]-center;\n"
+"						f = dot3F4( u, r );\n"
+"						a[ie].x = ((*(u32*)&f) & 0xffffff00) | (0xff & ie);\n"
+"						f = dot3F4( -u, r );\n"
+"						a[ie].y = ((*(u32*)&f) & 0xffffff00) | (0xff & ie);\n"
+"						f = dot3F4( v, r );\n"
+"						a[ie].z = ((*(u32*)&f) & 0xffffff00) | (0xff & ie);\n"
+"						f = dot3F4( -v, r );\n"
+"						a[ie].w = ((*(u32*)&f) & 0xffffff00) | (0xff & ie);\n"
+"					}\n"
+"					for(int ie=0; ie<nPoints; ie++)\n"
+"					{\n"
+"						a[0].x = (a[0].x > a[ie].x )? a[0].x: a[ie].x;\n"
+"						a[0].y = (a[0].y > a[ie].y )? a[0].y: a[ie].y;\n"
+"						a[0].z = (a[0].z > a[ie].z )? a[0].z: a[ie].z;\n"
+"						a[0].w = (a[0].w > a[ie].w )? a[0].w: a[ie].w;\n"
+"					}\n"
+"					idx[0] = (int)a[0].x & 0xff;\n"
+"					idx[1] = (int)a[0].y & 0xff;\n"
+"					idx[2] = (int)a[0].z & 0xff;\n"
+"					idx[3] = (int)a[0].w & 0xff;\n"
+"				}\n"
+"			}\n"
+"			{\n"
+"				float2 h[64];\n"
+"				PARALLEL_DO( h[ie] = make_float2((float)ie, p[ie].w), nPoints );\n"
+"				REDUCE_MIN( h, nPoints );\n"
+"				max00 = h[0];\n"
+"			}\n"
+"		}\n"
+"		contactIdx[0] = idx[0];\n"
+"		contactIdx[1] = idx[1];\n"
+"		contactIdx[2] = idx[2];\n"
+"		contactIdx[3] = idx[3];\n"
+"		return 4;\n"
+"	}\n"
+"}\n"
+"__kernel void   extractManifoldAndAddContactKernel(__global const int4* pairs, \n"
+"																	__global const b3RigidBodyData_t* rigidBodies, \n"
+"																	__global const float4* closestPointsWorld,\n"
+"																	__global const float4* separatingNormalsWorld,\n"
+"																	__global const int* contactCounts,\n"
+"																	__global const int* contactOffsets,\n"
+"																	__global struct b3Contact4Data* restrict contactsOut,\n"
+"																	counter32_t nContactsOut,\n"
+"																	int contactCapacity,\n"
+"																	int numPairs,\n"
+"																	int pairIndex\n"
+"																	)\n"
+"{\n"
+"	int idx = get_global_id(0);\n"
+"	\n"
+"	if (idx<numPairs)\n"
+"	{\n"
+"		float4 normal = separatingNormalsWorld[idx];\n"
+"		int nPoints = contactCounts[idx];\n"
+"		__global const float4* pointsIn = &closestPointsWorld[contactOffsets[idx]];\n"
+"		float4 localPoints[64];\n"
+"		for (int i=0;i<nPoints;i++)\n"
+"		{\n"
+"			localPoints[i] = pointsIn[i];\n"
+"		}\n"
+"		int contactIdx[4];// = {-1,-1,-1,-1};\n"
+"		contactIdx[0] = -1;\n"
+"		contactIdx[1] = -1;\n"
+"		contactIdx[2] = -1;\n"
+"		contactIdx[3] = -1;\n"
+"		int nContacts = extractManifoldSequential(localPoints, nPoints, normal, contactIdx);\n"
+"		int dstIdx;\n"
+"		AppendInc( nContactsOut, dstIdx );\n"
+"		if (dstIdx<contactCapacity)\n"
+"		{\n"
+"			__global struct b3Contact4Data* c = contactsOut + dstIdx;\n"
+"			c->m_worldNormalOnB = -normal;\n"
+"			c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);\n"
+"			c->m_batchIdx = idx;\n"
+"			int bodyA = pairs[pairIndex].x;\n"
+"			int bodyB = pairs[pairIndex].y;\n"
+"			c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0 ? -bodyA:bodyA;\n"
+"			c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0 ? -bodyB:bodyB;\n"
+"			c->m_childIndexA = -1;\n"
+"			c->m_childIndexB = -1;\n"
+"			for (int i=0;i<nContacts;i++)\n"
+"			{\n"
+"				c->m_worldPosB[i] = localPoints[contactIdx[i]];\n"
+"			}\n"
+"			GET_NPOINTS(*c) = nContacts;\n"
+"		}\n"
+"	}\n"
+"}\n"
+"void	trInverse(float4 translationIn, Quaternion orientationIn,\n"
+"		float4* translationOut, Quaternion* orientationOut)\n"
+"{\n"
+"	*orientationOut = qtInvert(orientationIn);\n"
+"	*translationOut = qtRotate(*orientationOut, -translationIn);\n"
+"}\n"
+"void	trMul(float4 translationA, Quaternion orientationA,\n"
+"						float4 translationB, Quaternion orientationB,\n"
+"		float4* translationOut, Quaternion* orientationOut)\n"
+"{\n"
+"	*orientationOut = qtMul(orientationA,orientationB);\n"
+"	*translationOut = transform(&translationB,&translationA,&orientationA);\n"
+"}\n"
+"__kernel void   clipHullHullKernel( __global int4* pairs, \n"
+"																					__global const b3RigidBodyData_t* rigidBodies, \n"
+"																					__global const b3Collidable_t* collidables,\n"
+"																					__global const b3ConvexPolyhedronData_t* convexShapes, \n"
+"																					__global const float4* vertices,\n"
+"																					__global const float4* uniqueEdges,\n"
+"																					__global const b3GpuFace_t* faces,\n"
+"																					__global const int* indices,\n"
+"																					__global const float4* separatingNormals,\n"
+"																					__global const int* hasSeparatingAxis,\n"
+"																					__global struct b3Contact4Data* restrict globalContactsOut,\n"
+"																					counter32_t nGlobalContactsOut,\n"
+"																					int numPairs,\n"
+"																					int contactCapacity)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	int pairIndex = i;\n"
+"	\n"
+"	float4 worldVertsB1[64];\n"
+"	float4 worldVertsB2[64];\n"
+"	int capacityWorldVerts = 64;	\n"
+"	float4 localContactsOut[64];\n"
+"	int localContactCapacity=64;\n"
+"	\n"
+"	float minDist = -1e30f;\n"
+"	float maxDist = 0.02f;\n"
+"	if (i<numPairs)\n"
+"	{\n"
+"		int bodyIndexA = pairs[i].x;\n"
+"		int bodyIndexB = pairs[i].y;\n"
+"			\n"
+"		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"		if (hasSeparatingAxis[i])\n"
+"		{\n"
+"			\n"
+"			int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"			int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"			\n"
+"		\n"
+"			int numLocalContactsOut = clipHullAgainstHull(separatingNormals[i],\n"
+"														&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],\n"
+"														rigidBodies[bodyIndexA].m_pos,rigidBodies[bodyIndexA].m_quat,\n"
+"													  rigidBodies[bodyIndexB].m_pos,rigidBodies[bodyIndexB].m_quat,\n"
+"													  worldVertsB1,worldVertsB2,capacityWorldVerts,\n"
+"														minDist, maxDist,\n"
+"														vertices,faces,indices,\n"
+"														localContactsOut,localContactCapacity);\n"
+"												\n"
+"		if (numLocalContactsOut>0)\n"
+"		{\n"
+"				float4 normal = -separatingNormals[i];\n"
+"				int nPoints = numLocalContactsOut;\n"
+"				float4* pointsIn = localContactsOut;\n"
+"				int contactIdx[4];// = {-1,-1,-1,-1};\n"
+"				contactIdx[0] = -1;\n"
+"				contactIdx[1] = -1;\n"
+"				contactIdx[2] = -1;\n"
+"				contactIdx[3] = -1;\n"
+"		\n"
+"				int nReducedContacts = extractManifoldSequential(pointsIn, nPoints, normal, contactIdx);\n"
+"		\n"
+"				\n"
+"				int mprContactIndex = pairs[pairIndex].z;\n"
+"				int dstIdx = mprContactIndex;\n"
+"				if (dstIdx<0)\n"
+"				{\n"
+"					AppendInc( nGlobalContactsOut, dstIdx );\n"
+"				}\n"
+"				if (dstIdx<contactCapacity)\n"
+"				{\n"
+"					pairs[pairIndex].z = dstIdx;\n"
+"					__global struct b3Contact4Data* c = globalContactsOut+ dstIdx;\n"
+"					c->m_worldNormalOnB = -normal;\n"
+"					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);\n"
+"					c->m_batchIdx = pairIndex;\n"
+"					int bodyA = pairs[pairIndex].x;\n"
+"					int bodyB = pairs[pairIndex].y;\n"
+"					c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;\n"
+"					c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;\n"
+"					c->m_childIndexA = -1;\n"
+"					c->m_childIndexB = -1;\n"
+"					for (int i=0;i<nReducedContacts;i++)\n"
+"					{\n"
+"					//this condition means: overwrite contact point, unless at index i==0 we have a valid 'mpr' contact\n"
+"						if (i>0||(mprContactIndex<0))\n"
+"						{\n"
+"							c->m_worldPosB[i] = pointsIn[contactIdx[i]];\n"
+"						}\n"
+"					}\n"
+"					GET_NPOINTS(*c) = nReducedContacts;\n"
+"				}\n"
+"				\n"
+"			}//		if (numContactsOut>0)\n"
+"		}//		if (hasSeparatingAxis[i])\n"
+"	}//	if (i<numPairs)\n"
+"}\n"
+"__kernel void   clipCompoundsHullHullKernel( __global const int4* gpuCompoundPairs, \n"
+"																					__global const b3RigidBodyData_t* rigidBodies, \n"
+"																					__global const b3Collidable_t* collidables,\n"
+"																					__global const b3ConvexPolyhedronData_t* convexShapes, \n"
+"																					__global const float4* vertices,\n"
+"																					__global const float4* uniqueEdges,\n"
+"																					__global const b3GpuFace_t* faces,\n"
+"																					__global const int* indices,\n"
+"																					__global const b3GpuChildShape_t* gpuChildShapes,\n"
+"																					__global const float4* gpuCompoundSepNormalsOut,\n"
+"																					__global const int* gpuHasCompoundSepNormalsOut,\n"
+"																					__global struct b3Contact4Data* restrict globalContactsOut,\n"
+"																					counter32_t nGlobalContactsOut,\n"
+"																					int numCompoundPairs, int maxContactCapacity)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	int pairIndex = i;\n"
+"	\n"
+"	float4 worldVertsB1[64];\n"
+"	float4 worldVertsB2[64];\n"
+"	int capacityWorldVerts = 64;	\n"
+"	float4 localContactsOut[64];\n"
+"	int localContactCapacity=64;\n"
+"	\n"
+"	float minDist = -1e30f;\n"
+"	float maxDist = 0.02f;\n"
+"	if (i<numCompoundPairs)\n"
+"	{\n"
+"		if (gpuHasCompoundSepNormalsOut[i])\n"
+"		{\n"
+"			int bodyIndexA = gpuCompoundPairs[i].x;\n"
+"			int bodyIndexB = gpuCompoundPairs[i].y;\n"
+"			\n"
+"			int childShapeIndexA = gpuCompoundPairs[i].z;\n"
+"			int childShapeIndexB = gpuCompoundPairs[i].w;\n"
+"			\n"
+"			int collidableIndexA = -1;\n"
+"			int collidableIndexB = -1;\n"
+"			\n"
+"			float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"			float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"			\n"
+"			float4 ornB = rigidBodies[bodyIndexB].m_quat;\n"
+"			float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"								\n"
+"			if (childShapeIndexA >= 0)\n"
+"			{\n"
+"				collidableIndexA = gpuChildShapes[childShapeIndexA].m_shapeIndex;\n"
+"				float4 childPosA = gpuChildShapes[childShapeIndexA].m_childPosition;\n"
+"				float4 childOrnA = gpuChildShapes[childShapeIndexA].m_childOrientation;\n"
+"				float4 newPosA = qtRotate(ornA,childPosA)+posA;\n"
+"				float4 newOrnA = qtMul(ornA,childOrnA);\n"
+"				posA = newPosA;\n"
+"				ornA = newOrnA;\n"
+"			} else\n"
+"			{\n"
+"				collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"			}\n"
+"			\n"
+"			if (childShapeIndexB>=0)\n"
+"			{\n"
+"				collidableIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;\n"
+"				float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;\n"
+"				float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;\n"
+"				float4 newPosB = transform(&childPosB,&posB,&ornB);\n"
+"				float4 newOrnB = qtMul(ornB,childOrnB);\n"
+"				posB = newPosB;\n"
+"				ornB = newOrnB;\n"
+"			} else\n"
+"			{\n"
+"				collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;	\n"
+"			}\n"
+"			\n"
+"			int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"			int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"		\n"
+"			int numLocalContactsOut = clipHullAgainstHull(gpuCompoundSepNormalsOut[i],\n"
+"														&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],\n"
+"														posA,ornA,\n"
+"													  posB,ornB,\n"
+"													  worldVertsB1,worldVertsB2,capacityWorldVerts,\n"
+"														minDist, maxDist,\n"
+"														vertices,faces,indices,\n"
+"														localContactsOut,localContactCapacity);\n"
+"												\n"
+"		if (numLocalContactsOut>0)\n"
+"		{\n"
+"				float4 normal = -gpuCompoundSepNormalsOut[i];\n"
+"				int nPoints = numLocalContactsOut;\n"
+"				float4* pointsIn = localContactsOut;\n"
+"				int contactIdx[4];// = {-1,-1,-1,-1};\n"
+"				contactIdx[0] = -1;\n"
+"				contactIdx[1] = -1;\n"
+"				contactIdx[2] = -1;\n"
+"				contactIdx[3] = -1;\n"
+"		\n"
+"				int nReducedContacts = extractManifoldSequential(pointsIn, nPoints, normal, contactIdx);\n"
+"		\n"
+"				int dstIdx;\n"
+"				AppendInc( nGlobalContactsOut, dstIdx );\n"
+"				if ((dstIdx+nReducedContacts) < maxContactCapacity)\n"
+"				{\n"
+"					__global struct b3Contact4Data* c = globalContactsOut+ dstIdx;\n"
+"					c->m_worldNormalOnB = -normal;\n"
+"					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);\n"
+"					c->m_batchIdx = pairIndex;\n"
+"					int bodyA = gpuCompoundPairs[pairIndex].x;\n"
+"					int bodyB = gpuCompoundPairs[pairIndex].y;\n"
+"					c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;\n"
+"					c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;\n"
+"					c->m_childIndexA = childShapeIndexA;\n"
+"					c->m_childIndexB = childShapeIndexB;\n"
+"					for (int i=0;i<nReducedContacts;i++)\n"
+"					{\n"
+"						c->m_worldPosB[i] = pointsIn[contactIdx[i]];\n"
+"					}\n"
+"					GET_NPOINTS(*c) = nReducedContacts;\n"
+"				}\n"
+"				\n"
+"			}//		if (numContactsOut>0)\n"
+"		}//		if (gpuHasCompoundSepNormalsOut[i])\n"
+"	}//	if (i<numCompoundPairs)\n"
+"}\n"
+"__kernel void   sphereSphereCollisionKernel( __global const int4* pairs, \n"
+"																					__global const b3RigidBodyData_t* rigidBodies, \n"
+"																					__global const b3Collidable_t* collidables,\n"
+"																					__global const float4* separatingNormals,\n"
+"																					__global const int* hasSeparatingAxis,\n"
+"																					__global struct b3Contact4Data* restrict globalContactsOut,\n"
+"																					counter32_t nGlobalContactsOut,\n"
+"																					int contactCapacity,\n"
+"																					int numPairs)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	int pairIndex = i;\n"
+"	\n"
+"	if (i<numPairs)\n"
+"	{\n"
+"		int bodyIndexA = pairs[i].x;\n"
+"		int bodyIndexB = pairs[i].y;\n"
+"			\n"
+"		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"		if (collidables[collidableIndexA].m_shapeType == SHAPE_SPHERE &&\n"
+"			collidables[collidableIndexB].m_shapeType == SHAPE_SPHERE)\n"
+"		{\n"
+"			//sphere-sphere\n"
+"			float radiusA = collidables[collidableIndexA].m_radius;\n"
+"			float radiusB = collidables[collidableIndexB].m_radius;\n"
+"			float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"			float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"			float4 diff = posA-posB;\n"
+"			float len = length(diff);\n"
+"			\n"
+"			///iff distance positive, don't generate a new contact\n"
+"			if ( len <= (radiusA+radiusB))\n"
+"			{\n"
+"				///distance (negative means penetration)\n"
+"				float dist = len - (radiusA+radiusB);\n"
+"				float4 normalOnSurfaceB = make_float4(1.f,0.f,0.f,0.f);\n"
+"				if (len > 0.00001)\n"
+"				{\n"
+"					normalOnSurfaceB = diff / len;\n"
+"				}\n"
+"				float4 contactPosB = posB + normalOnSurfaceB*radiusB;\n"
+"				contactPosB.w = dist;\n"
+"								\n"
+"				int dstIdx;\n"
+"				AppendInc( nGlobalContactsOut, dstIdx );\n"
+"				if (dstIdx < contactCapacity)\n"
+"				{\n"
+"					__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];\n"
+"					c->m_worldNormalOnB = -normalOnSurfaceB;\n"
+"					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);\n"
+"					c->m_batchIdx = pairIndex;\n"
+"					int bodyA = pairs[pairIndex].x;\n"
+"					int bodyB = pairs[pairIndex].y;\n"
+"					c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;\n"
+"					c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;\n"
+"					c->m_worldPosB[0] = contactPosB;\n"
+"					c->m_childIndexA = -1;\n"
+"					c->m_childIndexB = -1;\n"
+"					GET_NPOINTS(*c) = 1;\n"
+"				}//if (dstIdx < numPairs)\n"
+"			}//if ( len <= (radiusA+radiusB))\n"
+"		}//SHAPE_SPHERE SHAPE_SPHERE\n"
+"	}//if (i<numPairs)\n"
+"}				\n"
+"__kernel void   clipHullHullConcaveConvexKernel( __global int4* concavePairsIn,\n"
+"																					__global const b3RigidBodyData_t* rigidBodies, \n"
+"																					__global const b3Collidable_t* collidables,\n"
+"																					__global const b3ConvexPolyhedronData_t* convexShapes, \n"
+"																					__global const float4* vertices,\n"
+"																					__global const float4* uniqueEdges,\n"
+"																					__global const b3GpuFace_t* faces,\n"
+"																					__global const int* indices,\n"
+"																					__global const b3GpuChildShape_t* gpuChildShapes,\n"
+"																					__global const float4* separatingNormals,\n"
+"																					__global struct b3Contact4Data* restrict globalContactsOut,\n"
+"																					counter32_t nGlobalContactsOut,\n"
+"																					int contactCapacity,\n"
+"																					int numConcavePairs)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	int pairIndex = i;\n"
+"	\n"
+"	float4 worldVertsB1[64];\n"
+"	float4 worldVertsB2[64];\n"
+"	int capacityWorldVerts = 64;	\n"
+"	float4 localContactsOut[64];\n"
+"	int localContactCapacity=64;\n"
+"	\n"
+"	float minDist = -1e30f;\n"
+"	float maxDist = 0.02f;\n"
+"	if (i<numConcavePairs)\n"
+"	{\n"
+"		//negative value means that the pair is invalid\n"
+"		if (concavePairsIn[i].w<0)\n"
+"			return;\n"
+"		int bodyIndexA = concavePairsIn[i].x;\n"
+"		int bodyIndexB = concavePairsIn[i].y;\n"
+"		int f = concavePairsIn[i].z;\n"
+"		int childShapeIndexA = f;\n"
+"		\n"
+"		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"		\n"
+"		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"		\n"
+"		///////////////////////////////////////////////////////////////\n"
+"		\n"
+"	\n"
+"		bool overlap = false;\n"
+"		\n"
+"		b3ConvexPolyhedronData_t convexPolyhedronA;\n"
+"	//add 3 vertices of the triangle\n"
+"		convexPolyhedronA.m_numVertices = 3;\n"
+"		convexPolyhedronA.m_vertexOffset = 0;\n"
+"		float4	localCenter = make_float4(0.f,0.f,0.f,0.f);\n"
+"		b3GpuFace_t face = faces[convexShapes[shapeIndexA].m_faceOffset+f];\n"
+"		\n"
+"		float4 verticesA[3];\n"
+"		for (int i=0;i<3;i++)\n"
+"		{\n"
+"			int index = indices[face.m_indexOffset+i];\n"
+"			float4 vert = vertices[convexShapes[shapeIndexA].m_vertexOffset+index];\n"
+"			verticesA[i] = vert;\n"
+"			localCenter += vert;\n"
+"		}\n"
+"		float dmin = FLT_MAX;\n"
+"		int localCC=0;\n"
+"		//a triangle has 3 unique edges\n"
+"		convexPolyhedronA.m_numUniqueEdges = 3;\n"
+"		convexPolyhedronA.m_uniqueEdgesOffset = 0;\n"
+"		float4 uniqueEdgesA[3];\n"
+"		\n"
+"		uniqueEdgesA[0] = (verticesA[1]-verticesA[0]);\n"
+"		uniqueEdgesA[1] = (verticesA[2]-verticesA[1]);\n"
+"		uniqueEdgesA[2] = (verticesA[0]-verticesA[2]);\n"
+"		convexPolyhedronA.m_faceOffset = 0;\n"
+"                                  \n"
+"		float4 normal = make_float4(face.m_plane.x,face.m_plane.y,face.m_plane.z,0.f);\n"
+"                             \n"
+"		b3GpuFace_t facesA[TRIANGLE_NUM_CONVEX_FACES];\n"
+"		int indicesA[3+3+2+2+2];\n"
+"		int curUsedIndices=0;\n"
+"		int fidx=0;\n"
+"		//front size of triangle\n"
+"		{\n"
+"			facesA[fidx].m_indexOffset=curUsedIndices;\n"
+"			indicesA[0] = 0;\n"
+"			indicesA[1] = 1;\n"
+"			indicesA[2] = 2;\n"
+"			curUsedIndices+=3;\n"
+"			float c = face.m_plane.w;\n"
+"			facesA[fidx].m_plane.x = normal.x;\n"
+"			facesA[fidx].m_plane.y = normal.y;\n"
+"			facesA[fidx].m_plane.z = normal.z;\n"
+"			facesA[fidx].m_plane.w = c;\n"
+"			facesA[fidx].m_numIndices=3;\n"
+"		}\n"
+"		fidx++;\n"
+"		//back size of triangle\n"
+"		{\n"
+"			facesA[fidx].m_indexOffset=curUsedIndices;\n"
+"			indicesA[3]=2;\n"
+"			indicesA[4]=1;\n"
+"			indicesA[5]=0;\n"
+"			curUsedIndices+=3;\n"
+"			float c = dot3F4(normal,verticesA[0]);\n"
+"			float c1 = -face.m_plane.w;\n"
+"			facesA[fidx].m_plane.x = -normal.x;\n"
+"			facesA[fidx].m_plane.y = -normal.y;\n"
+"			facesA[fidx].m_plane.z = -normal.z;\n"
+"			facesA[fidx].m_plane.w = c;\n"
+"			facesA[fidx].m_numIndices=3;\n"
+"		}\n"
+"		fidx++;\n"
+"		bool addEdgePlanes = true;\n"
+"		if (addEdgePlanes)\n"
+"		{\n"
+"			int numVertices=3;\n"
+"			int prevVertex = numVertices-1;\n"
+"			for (int i=0;i<numVertices;i++)\n"
+"			{\n"
+"				float4 v0 = verticesA[i];\n"
+"				float4 v1 = verticesA[prevVertex];\n"
+"                                            \n"
+"				float4 edgeNormal = normalize(cross(normal,v1-v0));\n"
+"				float c = -dot3F4(edgeNormal,v0);\n"
+"				facesA[fidx].m_numIndices = 2;\n"
+"				facesA[fidx].m_indexOffset=curUsedIndices;\n"
+"				indicesA[curUsedIndices++]=i;\n"
+"				indicesA[curUsedIndices++]=prevVertex;\n"
+"                                            \n"
+"				facesA[fidx].m_plane.x = edgeNormal.x;\n"
+"				facesA[fidx].m_plane.y = edgeNormal.y;\n"
+"				facesA[fidx].m_plane.z = edgeNormal.z;\n"
+"				facesA[fidx].m_plane.w = c;\n"
+"				fidx++;\n"
+"				prevVertex = i;\n"
+"			}\n"
+"		}\n"
+"		convexPolyhedronA.m_numFaces = TRIANGLE_NUM_CONVEX_FACES;\n"
+"		convexPolyhedronA.m_localCenter = localCenter*(1.f/3.f);\n"
+"		float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"		posA.w = 0.f;\n"
+"		float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"		posB.w = 0.f;\n"
+"		float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"		float4 ornB =rigidBodies[bodyIndexB].m_quat;\n"
+"		float4 sepAxis = separatingNormals[i];\n"
+"		\n"
+"		int shapeTypeB = collidables[collidableIndexB].m_shapeType;\n"
+"		int childShapeIndexB =-1;\n"
+"		if (shapeTypeB==SHAPE_COMPOUND_OF_CONVEX_HULLS)\n"
+"		{\n"
+"			///////////////////\n"
+"			///compound shape support\n"
+"			\n"
+"			childShapeIndexB = concavePairsIn[pairIndex].w;\n"
+"			int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;\n"
+"			shapeIndexB = collidables[childColIndexB].m_shapeIndex;\n"
+"			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;\n"
+"			float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;\n"
+"			float4 newPosB = transform(&childPosB,&posB,&ornB);\n"
+"			float4 newOrnB = qtMul(ornB,childOrnB);\n"
+"			posB = newPosB;\n"
+"			ornB = newOrnB;\n"
+"			\n"
+"		}\n"
+"		\n"
+"		////////////////////////////////////////\n"
+"		\n"
+"		\n"
+"		\n"
+"		int numLocalContactsOut = clipHullAgainstHullLocalA(sepAxis,\n"
+"														&convexPolyhedronA, &convexShapes[shapeIndexB],\n"
+"														posA,ornA,\n"
+"													  posB,ornB,\n"
+"													  worldVertsB1,worldVertsB2,capacityWorldVerts,\n"
+"														minDist, maxDist,\n"
+"														&verticesA,&facesA,&indicesA,\n"
+"														vertices,faces,indices,\n"
+"														localContactsOut,localContactCapacity);\n"
+"												\n"
+"		if (numLocalContactsOut>0)\n"
+"		{\n"
+"			float4 normal = -separatingNormals[i];\n"
+"			int nPoints = numLocalContactsOut;\n"
+"			float4* pointsIn = localContactsOut;\n"
+"			int contactIdx[4];// = {-1,-1,-1,-1};\n"
+"			contactIdx[0] = -1;\n"
+"			contactIdx[1] = -1;\n"
+"			contactIdx[2] = -1;\n"
+"			contactIdx[3] = -1;\n"
+"	\n"
+"			int nReducedContacts = extractManifoldSequential(pointsIn, nPoints, normal, contactIdx);\n"
+"	\n"
+"			int dstIdx;\n"
+"			AppendInc( nGlobalContactsOut, dstIdx );\n"
+"			if (dstIdx<contactCapacity)\n"
+"			{\n"
+"				__global struct b3Contact4Data* c = globalContactsOut+ dstIdx;\n"
+"				c->m_worldNormalOnB = -normal;\n"
+"				c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);\n"
+"				c->m_batchIdx = pairIndex;\n"
+"				int bodyA = concavePairsIn[pairIndex].x;\n"
+"				int bodyB = concavePairsIn[pairIndex].y;\n"
+"				c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;\n"
+"				c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;\n"
+"				c->m_childIndexA = childShapeIndexA;\n"
+"				c->m_childIndexB = childShapeIndexB;\n"
+"				for (int i=0;i<nReducedContacts;i++)\n"
+"				{\n"
+"					c->m_worldPosB[i] = pointsIn[contactIdx[i]];\n"
+"				}\n"
+"				GET_NPOINTS(*c) = nReducedContacts;\n"
+"			}\n"
+"				\n"
+"		}//		if (numContactsOut>0)\n"
+"	}//	if (i<numPairs)\n"
+"}\n"
+"int	findClippingFaces(const float4 separatingNormal,\n"
+"                      __global const b3ConvexPolyhedronData_t* hullA, __global const b3ConvexPolyhedronData_t* hullB,\n"
+"                      const float4 posA, const Quaternion ornA,const float4 posB, const Quaternion ornB,\n"
+"                       __global float4* worldVertsA1,\n"
+"                      __global float4* worldNormalsA1,\n"
+"                      __global float4* worldVertsB1,\n"
+"                      int capacityWorldVerts,\n"
+"                      const float minDist, float maxDist,\n"
+"                      __global const float4* vertices,\n"
+"                      __global const b3GpuFace_t* faces,\n"
+"                      __global const int* indices,\n"
+"                      __global int4* clippingFaces, int pairIndex)\n"
+"{\n"
+"	int numContactsOut = 0;\n"
+"	int numWorldVertsB1= 0;\n"
+"    \n"
+"    \n"
+"	int closestFaceB=-1;\n"
+"	float dmax = -FLT_MAX;\n"
+"    \n"
+"	{\n"
+"		for(int face=0;face<hullB->m_numFaces;face++)\n"
+"		{\n"
+"			const float4 Normal = make_float4(faces[hullB->m_faceOffset+face].m_plane.x,\n"
+"                                              faces[hullB->m_faceOffset+face].m_plane.y, faces[hullB->m_faceOffset+face].m_plane.z,0.f);\n"
+"			const float4 WorldNormal = qtRotate(ornB, Normal);\n"
+"			float d = dot3F4(WorldNormal,separatingNormal);\n"
+"			if (d > dmax)\n"
+"			{\n"
+"				dmax = d;\n"
+"				closestFaceB = face;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"    \n"
+"	{\n"
+"		const b3GpuFace_t polyB = faces[hullB->m_faceOffset+closestFaceB];\n"
+"		const int numVertices = polyB.m_numIndices;\n"
+"		for(int e0=0;e0<numVertices;e0++)\n"
+"		{\n"
+"			const float4 b = vertices[hullB->m_vertexOffset+indices[polyB.m_indexOffset+e0]];\n"
+"			worldVertsB1[pairIndex*capacityWorldVerts+numWorldVertsB1++] = transform(&b,&posB,&ornB);\n"
+"		}\n"
+"	}\n"
+"    \n"
+"    int closestFaceA=-1;\n"
+"	{\n"
+"		float dmin = FLT_MAX;\n"
+"		for(int face=0;face<hullA->m_numFaces;face++)\n"
+"		{\n"
+"			const float4 Normal = make_float4(\n"
+"                                              faces[hullA->m_faceOffset+face].m_plane.x,\n"
+"                                              faces[hullA->m_faceOffset+face].m_plane.y,\n"
+"                                              faces[hullA->m_faceOffset+face].m_plane.z,\n"
+"                                              0.f);\n"
+"			const float4 faceANormalWS = qtRotate(ornA,Normal);\n"
+"            \n"
+"			float d = dot3F4(faceANormalWS,separatingNormal);\n"
+"			if (d < dmin)\n"
+"			{\n"
+"				dmin = d;\n"
+"				closestFaceA = face;\n"
+"                worldNormalsA1[pairIndex] = faceANormalWS;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"    \n"
+"    int numVerticesA = faces[hullA->m_faceOffset+closestFaceA].m_numIndices;\n"
+"	for(int e0=0;e0<numVerticesA;e0++)\n"
+"	{\n"
+"        const float4 a = vertices[hullA->m_vertexOffset+indices[faces[hullA->m_faceOffset+closestFaceA].m_indexOffset+e0]];\n"
+"        worldVertsA1[pairIndex*capacityWorldVerts+e0] = transform(&a, &posA,&ornA);\n"
+"    }\n"
+"    \n"
+"    clippingFaces[pairIndex].x = closestFaceA;\n"
+"    clippingFaces[pairIndex].y = closestFaceB;\n"
+"    clippingFaces[pairIndex].z = numVerticesA;\n"
+"    clippingFaces[pairIndex].w = numWorldVertsB1;\n"
+"    \n"
+"    \n"
+"	return numContactsOut;\n"
+"}\n"
+"int clipFaces(__global float4* worldVertsA1,\n"
+"              __global float4* worldNormalsA1,\n"
+"              __global float4* worldVertsB1,\n"
+"              __global float4* worldVertsB2, \n"
+"              int capacityWorldVertsB2,\n"
+"              const float minDist, float maxDist,\n"
+"              __global int4* clippingFaces,\n"
+"              int pairIndex)\n"
+"{\n"
+"	int numContactsOut = 0;\n"
+"    \n"
+"    int closestFaceA = clippingFaces[pairIndex].x;\n"
+"    int closestFaceB = clippingFaces[pairIndex].y;\n"
+"	int numVertsInA = clippingFaces[pairIndex].z;\n"
+"	int numVertsInB = clippingFaces[pairIndex].w;\n"
+"    \n"
+"	int numVertsOut = 0;\n"
+"    \n"
+"	if (closestFaceA<0)\n"
+"		return numContactsOut;\n"
+"    \n"
+"    __global float4* pVtxIn = &worldVertsB1[pairIndex*capacityWorldVertsB2];\n"
+"    __global float4* pVtxOut = &worldVertsB2[pairIndex*capacityWorldVertsB2];\n"
+"    \n"
+"    \n"
+"	\n"
+"	// clip polygon to back of planes of all faces of hull A that are adjacent to witness face\n"
+"    \n"
+"	for(int e0=0;e0<numVertsInA;e0++)\n"
+"	{\n"
+"		const float4 aw = worldVertsA1[pairIndex*capacityWorldVertsB2+e0];\n"
+"		const float4 bw = worldVertsA1[pairIndex*capacityWorldVertsB2+((e0+1)%numVertsInA)];\n"
+"		const float4 WorldEdge0 = aw - bw;\n"
+"		float4 worldPlaneAnormal1 = worldNormalsA1[pairIndex];\n"
+"		float4 planeNormalWS1 = -cross3(WorldEdge0,worldPlaneAnormal1);\n"
+"		float4 worldA1 = aw;\n"
+"		float planeEqWS1 = -dot3F4(worldA1,planeNormalWS1);\n"
+"		float4 planeNormalWS = planeNormalWS1;\n"
+"		float planeEqWS=planeEqWS1;\n"
+"		numVertsOut = clipFaceGlobal(pVtxIn, numVertsInB, planeNormalWS,planeEqWS, pVtxOut);\n"
+"		__global float4* tmp = pVtxOut;\n"
+"		pVtxOut = pVtxIn;\n"
+"		pVtxIn = tmp;\n"
+"		numVertsInB = numVertsOut;\n"
+"		numVertsOut = 0;\n"
+"	}\n"
+"    \n"
+"    //float4 planeNormalWS = worldNormalsA1[pairIndex];\n"
+"    //float planeEqWS=-dot3F4(planeNormalWS,worldVertsA1[pairIndex*capacityWorldVertsB2]);\n"
+"    \n"
+"    /*for (int i=0;i<numVertsInB;i++)\n"
+"    {\n"
+"        pVtxOut[i] = pVtxIn[i];\n"
+"    }*/\n"
+"    \n"
+"    \n"
+"    \n"
+"    \n"
+"    //numVertsInB=0;\n"
+"	\n"
+"    float4 planeNormalWS = worldNormalsA1[pairIndex];\n"
+"    float planeEqWS=-dot3F4(planeNormalWS,worldVertsA1[pairIndex*capacityWorldVertsB2]);\n"
+"    for (int i=0;i<numVertsInB;i++)\n"
+"    {\n"
+"        float depth = dot3F4(planeNormalWS,pVtxIn[i])+planeEqWS;\n"
+"        if (depth <=minDist)\n"
+"        {\n"
+"            depth = minDist;\n"
+"        }\n"
+"        \n"
+"        if (depth <=maxDist)\n"
+"        {\n"
+"            float4 pointInWorld = pVtxIn[i];\n"
+"            pVtxOut[numContactsOut++] = make_float4(pointInWorld.x,pointInWorld.y,pointInWorld.z,depth);\n"
+"        }\n"
+"    }\n"
+"   \n"
+"    clippingFaces[pairIndex].w =numContactsOut;\n"
+"   \n"
+"    \n"
+"	return numContactsOut;\n"
+"}\n"
+"__kernel void   findClippingFacesKernel(  __global const int4* pairs,\n"
+"                                        __global const b3RigidBodyData_t* rigidBodies,\n"
+"                                        __global const b3Collidable_t* collidables,\n"
+"                                        __global const b3ConvexPolyhedronData_t* convexShapes,\n"
+"                                        __global const float4* vertices,\n"
+"                                        __global const float4* uniqueEdges,\n"
+"                                        __global const b3GpuFace_t* faces,\n"
+"                                        __global const int* indices,\n"
+"                                        __global const float4* separatingNormals,\n"
+"                                        __global const int* hasSeparatingAxis,\n"
+"                                        __global int4* clippingFacesOut,\n"
+"                                        __global float4* worldVertsA1,\n"
+"                                        __global float4* worldNormalsA1,\n"
+"                                        __global float4* worldVertsB1,\n"
+"                                        int capacityWorldVerts,\n"
+"                                        int numPairs\n"
+"                                        )\n"
+"{\n"
+"    \n"
+"	int i = get_global_id(0);\n"
+"	int pairIndex = i;\n"
+"    \n"
+"	\n"
+"	float minDist = -1e30f;\n"
+"	float maxDist = 0.02f;\n"
+"    \n"
+"	if (i<numPairs)\n"
+"	{\n"
+"        \n"
+"		if (hasSeparatingAxis[i])\n"
+"		{\n"
+"            \n"
+"			int bodyIndexA = pairs[i].x;\n"
+"			int bodyIndexB = pairs[i].y;\n"
+"			\n"
+"			int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"			int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"			\n"
+"			int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"			int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"			\n"
+"            \n"
+"            \n"
+"			int numLocalContactsOut = findClippingFaces(separatingNormals[i],\n"
+"                                                        &convexShapes[shapeIndexA], &convexShapes[shapeIndexB],\n"
+"                                                        rigidBodies[bodyIndexA].m_pos,rigidBodies[bodyIndexA].m_quat,\n"
+"                                                        rigidBodies[bodyIndexB].m_pos,rigidBodies[bodyIndexB].m_quat,\n"
+"                                                        worldVertsA1,\n"
+"                                                        worldNormalsA1,\n"
+"                                                        worldVertsB1,capacityWorldVerts,\n"
+"                                                        minDist, maxDist,\n"
+"                                                        vertices,faces,indices,\n"
+"                                                        clippingFacesOut,i);\n"
+"            \n"
+"            \n"
+"		}//		if (hasSeparatingAxis[i])\n"
+"	}//	if (i<numPairs)\n"
+"    \n"
+"}\n"
+"__kernel void   clipFacesAndFindContactsKernel(    __global const float4* separatingNormals,\n"
+"                                                   __global const int* hasSeparatingAxis,\n"
+"                                                   __global int4* clippingFacesOut,\n"
+"                                                   __global float4* worldVertsA1,\n"
+"                                                   __global float4* worldNormalsA1,\n"
+"                                                   __global float4* worldVertsB1,\n"
+"                                                   __global float4* worldVertsB2,\n"
+"                                                    int vertexFaceCapacity,\n"
+"                                                   int numPairs,\n"
+"					                                        int debugMode\n"
+"                                                   )\n"
+"{\n"
+"    int i = get_global_id(0);\n"
+"	int pairIndex = i;\n"
+"	\n"
+"    \n"
+"	float minDist = -1e30f;\n"
+"	float maxDist = 0.02f;\n"
+"    \n"
+"	if (i<numPairs)\n"
+"	{\n"
+"        \n"
+"		if (hasSeparatingAxis[i])\n"
+"		{\n"
+"            \n"
+"//			int bodyIndexA = pairs[i].x;\n"
+"	//		int bodyIndexB = pairs[i].y;\n"
+"		    \n"
+"            int numLocalContactsOut = 0;\n"
+"            int capacityWorldVertsB2 = vertexFaceCapacity;\n"
+"            \n"
+"            __global float4* pVtxIn = &worldVertsB1[pairIndex*capacityWorldVertsB2];\n"
+"            __global float4* pVtxOut = &worldVertsB2[pairIndex*capacityWorldVertsB2];\n"
+"            \n"
+"            {\n"
+"                __global int4* clippingFaces = clippingFacesOut;\n"
+"            \n"
+"                \n"
+"                int closestFaceA = clippingFaces[pairIndex].x;\n"
+"                int closestFaceB = clippingFaces[pairIndex].y;\n"
+"                int numVertsInA = clippingFaces[pairIndex].z;\n"
+"                int numVertsInB = clippingFaces[pairIndex].w;\n"
+"                \n"
+"                int numVertsOut = 0;\n"
+"                \n"
+"                if (closestFaceA>=0)\n"
+"                {\n"
+"                    \n"
+"                    \n"
+"                    \n"
+"                    // clip polygon to back of planes of all faces of hull A that are adjacent to witness face\n"
+"                    \n"
+"                    for(int e0=0;e0<numVertsInA;e0++)\n"
+"                    {\n"
+"                        const float4 aw = worldVertsA1[pairIndex*capacityWorldVertsB2+e0];\n"
+"                        const float4 bw = worldVertsA1[pairIndex*capacityWorldVertsB2+((e0+1)%numVertsInA)];\n"
+"                        const float4 WorldEdge0 = aw - bw;\n"
+"                        float4 worldPlaneAnormal1 = worldNormalsA1[pairIndex];\n"
+"                        float4 planeNormalWS1 = -cross3(WorldEdge0,worldPlaneAnormal1);\n"
+"                        float4 worldA1 = aw;\n"
+"                        float planeEqWS1 = -dot3F4(worldA1,planeNormalWS1);\n"
+"                        float4 planeNormalWS = planeNormalWS1;\n"
+"                        float planeEqWS=planeEqWS1;\n"
+"                        numVertsOut = clipFaceGlobal(pVtxIn, numVertsInB, planeNormalWS,planeEqWS, pVtxOut);\n"
+"                        __global float4* tmp = pVtxOut;\n"
+"                        pVtxOut = pVtxIn;\n"
+"                        pVtxIn = tmp;\n"
+"                        numVertsInB = numVertsOut;\n"
+"                        numVertsOut = 0;\n"
+"                    }\n"
+"                    \n"
+"                    float4 planeNormalWS = worldNormalsA1[pairIndex];\n"
+"                    float planeEqWS=-dot3F4(planeNormalWS,worldVertsA1[pairIndex*capacityWorldVertsB2]);\n"
+"                    \n"
+"                    for (int i=0;i<numVertsInB;i++)\n"
+"                    {\n"
+"                        float depth = dot3F4(planeNormalWS,pVtxIn[i])+planeEqWS;\n"
+"                        if (depth <=minDist)\n"
+"                        {\n"
+"                            depth = minDist;\n"
+"                        }\n"
+"                        \n"
+"                        if (depth <=maxDist)\n"
+"                        {\n"
+"                            float4 pointInWorld = pVtxIn[i];\n"
+"                            pVtxOut[numLocalContactsOut++] = make_float4(pointInWorld.x,pointInWorld.y,pointInWorld.z,depth);\n"
+"                        }\n"
+"                    }\n"
+"                    \n"
+"                }\n"
+"                clippingFaces[pairIndex].w =numLocalContactsOut;\n"
+"                \n"
+"            }\n"
+"            \n"
+"            for (int i=0;i<numLocalContactsOut;i++)\n"
+"                pVtxIn[i] = pVtxOut[i];\n"
+"                \n"
+"		}//		if (hasSeparatingAxis[i])\n"
+"	}//	if (i<numPairs)\n"
+"    \n"
+"}\n"
+"__kernel void   newContactReductionKernel( __global int4* pairs,\n"
+"                                                   __global const b3RigidBodyData_t* rigidBodies,\n"
+"                                                   __global const float4* separatingNormals,\n"
+"                                                   __global const int* hasSeparatingAxis,\n"
+"                                                   __global struct b3Contact4Data* globalContactsOut,\n"
+"                                                   __global int4* clippingFaces,\n"
+"                                                   __global float4* worldVertsB2,\n"
+"                                                   volatile __global int* nGlobalContactsOut,\n"
+"                                                   int vertexFaceCapacity,\n"
+"												   int contactCapacity,\n"
+"                                                   int numPairs\n"
+"                                                   )\n"
+"{\n"
+"    int i = get_global_id(0);\n"
+"	int pairIndex = i;\n"
+"	\n"
+"    int4 contactIdx;\n"
+"    contactIdx=make_int4(0,1,2,3);\n"
+"    \n"
+"	if (i<numPairs)\n"
+"	{\n"
+"        \n"
+"		if (hasSeparatingAxis[i])\n"
+"		{\n"
+"            \n"
+"			\n"
+"            \n"
+"            \n"
+"			int nPoints = clippingFaces[pairIndex].w;\n"
+"           \n"
+"            if (nPoints>0)\n"
+"            {\n"
+"                 __global float4* pointsIn = &worldVertsB2[pairIndex*vertexFaceCapacity];\n"
+"                float4 normal = -separatingNormals[i];\n"
+"                \n"
+"                int nReducedContacts = extractManifoldSequentialGlobal(pointsIn, nPoints, normal, &contactIdx);\n"
+"            \n"
+"				int mprContactIndex = pairs[pairIndex].z;\n"
+"                int dstIdx = mprContactIndex;\n"
+"				if (dstIdx<0)\n"
+"				{\n"
+"	                AppendInc( nGlobalContactsOut, dstIdx );\n"
+"				}\n"
+"//#if 0\n"
+"                \n"
+"				if (dstIdx < contactCapacity)\n"
+"				{\n"
+"					__global struct b3Contact4Data* c = &globalContactsOut[dstIdx];\n"
+"					c->m_worldNormalOnB = -normal;\n"
+"					c->m_restituitionCoeffCmp = (0.f*0xffff);c->m_frictionCoeffCmp = (0.7f*0xffff);\n"
+"					c->m_batchIdx = pairIndex;\n"
+"					int bodyA = pairs[pairIndex].x;\n"
+"					int bodyB = pairs[pairIndex].y;\n"
+"					pairs[pairIndex].w = dstIdx;\n"
+"					c->m_bodyAPtrAndSignBit = rigidBodies[bodyA].m_invMass==0?-bodyA:bodyA;\n"
+"					c->m_bodyBPtrAndSignBit = rigidBodies[bodyB].m_invMass==0?-bodyB:bodyB;\n"
+"                    c->m_childIndexA =-1;\n"
+"					c->m_childIndexB =-1;\n"
+"                    switch (nReducedContacts)\n"
+"                    {\n"
+"                        case 4:\n"
+"                            c->m_worldPosB[3] = pointsIn[contactIdx.w];\n"
+"                        case 3:\n"
+"                            c->m_worldPosB[2] = pointsIn[contactIdx.z];\n"
+"                        case 2:\n"
+"                            c->m_worldPosB[1] = pointsIn[contactIdx.y];\n"
+"                        case 1:\n"
+"							if (mprContactIndex<0)//test\n"
+"	                            c->m_worldPosB[0] = pointsIn[contactIdx.x];\n"
+"                        default:\n"
+"                        {\n"
+"                        }\n"
+"                    };\n"
+"                    \n"
+"					GET_NPOINTS(*c) = nReducedContacts;\n"
+"                    \n"
+"                 }\n"
+"                 \n"
+"                \n"
+"//#endif\n"
+"				\n"
+"			}//		if (numContactsOut>0)\n"
+"		}//		if (hasSeparatingAxis[i])\n"
+"	}//	if (i<numPairs)\n"
+"    \n"
+"    \n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satConcave.cl b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satConcave.cl
new file mode 100644
index 0000000000..31ca43b8cd
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satConcave.cl
@@ -0,0 +1,1220 @@
+
+//keep this enum in sync with the CPU version (in btCollidable.h)
+//written by Erwin Coumans
+
+
+#define SHAPE_CONVEX_HULL 3
+#define SHAPE_CONCAVE_TRIMESH 5
+#define TRIANGLE_NUM_CONVEX_FACES 5
+#define SHAPE_COMPOUND_OF_CONVEX_HULLS 6
+
+#define B3_MAX_STACK_DEPTH 256
+
+
+typedef unsigned int u32;
+
+///keep this in sync with btCollidable.h
+typedef struct
+{
+	union {
+		int m_numChildShapes;
+		int m_bvhIndex;
+	};
+	union
+	{
+		float m_radius;
+		int	m_compoundBvhIndex;
+	};
+	
+	int m_shapeType;
+	int m_shapeIndex;
+	
+} btCollidableGpu;
+
+#define MAX_NUM_PARTS_IN_BITS 10
+
+///b3QuantizedBvhNode is a compressed aabb node, 16 bytes.
+///Node can be used for leafnode or internal node. Leafnodes can point to 32-bit triangle index (non-negative range).
+typedef struct
+{
+	//12 bytes
+	unsigned short int	m_quantizedAabbMin[3];
+	unsigned short int	m_quantizedAabbMax[3];
+	//4 bytes
+	int	m_escapeIndexOrTriangleIndex;
+} b3QuantizedBvhNode;
+
+typedef struct
+{
+	float4		m_aabbMin;
+	float4		m_aabbMax;
+	float4		m_quantization;
+	int			m_numNodes;
+	int			m_numSubTrees;
+	int			m_nodeOffset;
+	int			m_subTreeOffset;
+
+} b3BvhInfo;
+
+
+int	getTriangleIndex(const b3QuantizedBvhNode* rootNode)
+{
+	unsigned int x=0;
+	unsigned int y = (~(x&0))<<(31-MAX_NUM_PARTS_IN_BITS);
+	// Get only the lower bits where the triangle index is stored
+	return (rootNode->m_escapeIndexOrTriangleIndex&~(y));
+}
+
+int	getTriangleIndexGlobal(__global const b3QuantizedBvhNode* rootNode)
+{
+	unsigned int x=0;
+	unsigned int y = (~(x&0))<<(31-MAX_NUM_PARTS_IN_BITS);
+	// Get only the lower bits where the triangle index is stored
+	return (rootNode->m_escapeIndexOrTriangleIndex&~(y));
+}
+
+int isLeafNode(const b3QuantizedBvhNode* rootNode)
+{
+	//skipindex is negative (internal node), triangleindex >=0 (leafnode)
+	return (rootNode->m_escapeIndexOrTriangleIndex >= 0)? 1 : 0;
+}
+
+int isLeafNodeGlobal(__global const b3QuantizedBvhNode* rootNode)
+{
+	//skipindex is negative (internal node), triangleindex >=0 (leafnode)
+	return (rootNode->m_escapeIndexOrTriangleIndex >= 0)? 1 : 0;
+}
+	
+int getEscapeIndex(const b3QuantizedBvhNode* rootNode)
+{
+	return -rootNode->m_escapeIndexOrTriangleIndex;
+}
+
+int getEscapeIndexGlobal(__global const b3QuantizedBvhNode* rootNode)
+{
+	return -rootNode->m_escapeIndexOrTriangleIndex;
+}
+
+
+typedef struct
+{
+	//12 bytes
+	unsigned short int	m_quantizedAabbMin[3];
+	unsigned short int	m_quantizedAabbMax[3];
+	//4 bytes, points to the root of the subtree
+	int			m_rootNodeIndex;
+	//4 bytes
+	int			m_subtreeSize;
+	int			m_padding[3];
+} b3BvhSubtreeInfo;
+
+
+
+
+
+
+
+typedef struct
+{
+	float4	m_childPosition;
+	float4	m_childOrientation;
+	int m_shapeIndex;
+	int m_unused0;
+	int m_unused1;
+	int m_unused2;
+} btGpuChildShape;
+
+
+typedef struct
+{
+	float4 m_pos;
+	float4 m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_collidableIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} BodyData;
+
+
+typedef struct  
+{
+	float4		m_localCenter;
+	float4		m_extents;
+	float4		mC;
+	float4		mE;
+	
+	float			m_radius;
+	int	m_faceOffset;
+	int m_numFaces;
+	int	m_numVertices;
+
+	int m_vertexOffset;
+	int	m_uniqueEdgesOffset;
+	int	m_numUniqueEdges;
+	int m_unused;
+} ConvexPolyhedronCL;
+
+typedef struct 
+{
+	union
+	{
+		float4	m_min;
+		float   m_minElems[4];
+		int			m_minIndices[4];
+	};
+	union
+	{
+		float4	m_max;
+		float   m_maxElems[4];
+		int			m_maxIndices[4];
+	};
+} btAabbCL;
+
+#include "Bullet3Collision/BroadPhaseCollision/shared/b3Aabb.h"
+#include "Bullet3Common/shared/b3Int2.h"
+
+
+
+typedef struct
+{
+	float4 m_plane;
+	int m_indexOffset;
+	int m_numIndices;
+} btGpuFace;
+
+#define make_float4 (float4)
+
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+
+	
+//	float4 a1 = make_float4(a.xyz,0.f);
+//	float4 b1 = make_float4(b.xyz,0.f);
+
+//	return cross(a1,b1);
+
+//float4 c = make_float4(a.y*b.z - a.z*b.y,a.z*b.x - a.x*b.z,a.x*b.y - a.y*b.x,0.f);
+	
+	//	float4 c = make_float4(a.y*b.z - a.z*b.y,1.f,a.x*b.y - a.y*b.x,0.f);
+	
+	//return c;
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = make_float4(a.xyz,0.f);
+	float4 b1 = make_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	v = make_float4(v.xyz,0.f);
+	return fast_normalize(v);
+}
+
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b);
+
+__inline
+Quaternion qtNormalize(Quaternion in);
+
+__inline
+float4 qtRotate(Quaternion q, float4 vec);
+
+__inline
+Quaternion qtInvert(Quaternion q);
+
+
+
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross3( a, b );
+	ans += a.w*b+b.w*a;
+//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+Quaternion qtNormalize(Quaternion in)
+{
+	return fastNormalize4(in);
+//	in /= length( in );
+//	return in;
+}
+__inline
+float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(qtMul(q,vcpy),qInv);
+	return out;
+}
+
+__inline
+Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+__inline
+float4 qtInvRotate(const Quaternion q, float4 vec)
+{
+	return qtRotate( qtInvert( q ), vec );
+}
+
+__inline
+float4 transform(const float4* p, const float4* translation, const Quaternion* orientation)
+{
+	return qtRotate( *orientation, *p ) + (*translation);
+}
+
+
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = make_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+}
+
+inline void projectLocal(const ConvexPolyhedronCL* hull,  const float4 pos, const float4 orn, 
+const float4* dir, const float4* vertices, float* min, float* max)
+{
+	min[0] = FLT_MAX;
+	max[0] = -FLT_MAX;
+	int numVerts = hull->m_numVertices;
+
+	const float4 localDir = qtInvRotate(orn,*dir);
+	float offset = dot(pos,*dir);
+	for(int i=0;i<numVerts;i++)
+	{
+		float dp = dot(vertices[hull->m_vertexOffset+i],localDir);
+		if(dp < min[0])	
+			min[0] = dp;
+		if(dp > max[0])	
+			max[0] = dp;
+	}
+	if(min[0]>max[0])
+	{
+		float tmp = min[0];
+		min[0] = max[0];
+		max[0] = tmp;
+	}
+	min[0] += offset;
+	max[0] += offset;
+}
+
+inline void project(__global const ConvexPolyhedronCL* hull,  const float4 pos, const float4 orn, 
+const float4* dir, __global const float4* vertices, float* min, float* max)
+{
+	min[0] = FLT_MAX;
+	max[0] = -FLT_MAX;
+	int numVerts = hull->m_numVertices;
+
+	const float4 localDir = qtInvRotate(orn,*dir);
+	float offset = dot(pos,*dir);
+	for(int i=0;i<numVerts;i++)
+	{
+		float dp = dot(vertices[hull->m_vertexOffset+i],localDir);
+		if(dp < min[0])	
+			min[0] = dp;
+		if(dp > max[0])	
+			max[0] = dp;
+	}
+	if(min[0]>max[0])
+	{
+		float tmp = min[0];
+		min[0] = max[0];
+		max[0] = tmp;
+	}
+	min[0] += offset;
+	max[0] += offset;
+}
+
+inline bool TestSepAxisLocalA(const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, 
+	const float4 posA,const float4 ornA,
+	const float4 posB,const float4 ornB,
+	float4* sep_axis, const float4* verticesA, __global const float4* verticesB,float* depth)
+{
+	float Min0,Max0;
+	float Min1,Max1;
+	projectLocal(hullA,posA,ornA,sep_axis,verticesA, &Min0, &Max0);
+	project(hullB,posB,ornB, sep_axis,verticesB, &Min1, &Max1);
+
+	if(Max0<Min1 || Max1<Min0)
+		return false;
+
+	float d0 = Max0 - Min1;
+	float d1 = Max1 - Min0;
+	*depth = d0<d1 ? d0:d1;
+	return true;
+}
+
+
+
+
+inline bool IsAlmostZero(const float4 v)
+{
+	if(fabs(v.x)>1e-6f || fabs(v.y)>1e-6f || fabs(v.z)>1e-6f)
+		return false;
+	return true;
+}
+
+
+
+bool findSeparatingAxisLocalA(	const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, 
+	const float4 posA1,
+	const float4 ornA,
+	const float4 posB1,
+	const float4 ornB,
+	const float4 DeltaC2,
+	
+	const float4* verticesA, 
+	const float4* uniqueEdgesA, 
+	const btGpuFace* facesA,
+	const int*  indicesA,
+
+	__global const float4* verticesB, 
+	__global const float4* uniqueEdgesB, 
+	__global const btGpuFace* facesB,
+	__global const int*  indicesB,
+	float4* sep,
+	float* dmin)
+{
+	
+
+	float4 posA = posA1;
+	posA.w = 0.f;
+	float4 posB = posB1;
+	posB.w = 0.f;
+	int curPlaneTests=0;
+	{
+		int numFacesA = hullA->m_numFaces;
+		// Test normals from hullA
+		for(int i=0;i<numFacesA;i++)
+		{
+			const float4 normal = facesA[hullA->m_faceOffset+i].m_plane;
+			float4 faceANormalWS = qtRotate(ornA,normal);
+			if (dot3F4(DeltaC2,faceANormalWS)<0)
+				faceANormalWS*=-1.f;
+			curPlaneTests++;
+			float d;
+			if(!TestSepAxisLocalA( hullA, hullB, posA,ornA,posB,ornB,&faceANormalWS, verticesA, verticesB,&d))
+				return false;
+			if(d<*dmin)
+			{
+				*dmin = d;
+				*sep = faceANormalWS;
+			}
+		}
+	}
+	if((dot3F4(-DeltaC2,*sep))>0.0f)
+	{
+		*sep = -(*sep);
+	}
+	return true;
+}
+
+bool findSeparatingAxisLocalB(	__global const ConvexPolyhedronCL* hullA,  const ConvexPolyhedronCL* hullB, 
+	const float4 posA1,
+	const float4 ornA,
+	const float4 posB1,
+	const float4 ornB,
+	const float4 DeltaC2,
+	__global const float4* verticesA, 
+	__global const float4* uniqueEdgesA, 
+	__global const btGpuFace* facesA,
+	__global const int*  indicesA,
+	const float4* verticesB,
+	const float4* uniqueEdgesB, 
+	const btGpuFace* facesB,
+	const int*  indicesB,
+	float4* sep,
+	float* dmin)
+{
+
+
+	float4 posA = posA1;
+	posA.w = 0.f;
+	float4 posB = posB1;
+	posB.w = 0.f;
+	int curPlaneTests=0;
+	{
+		int numFacesA = hullA->m_numFaces;
+		// Test normals from hullA
+		for(int i=0;i<numFacesA;i++)
+		{
+			const float4 normal = facesA[hullA->m_faceOffset+i].m_plane;
+			float4 faceANormalWS = qtRotate(ornA,normal);
+			if (dot3F4(DeltaC2,faceANormalWS)<0)
+				faceANormalWS *= -1.f;
+			curPlaneTests++;
+			float d;
+			if(!TestSepAxisLocalA( hullB, hullA, posB,ornB,posA,ornA, &faceANormalWS, verticesB,verticesA, &d))
+				return false;
+			if(d<*dmin)
+			{
+				*dmin = d;
+				*sep = faceANormalWS;
+			}
+		}
+	}
+	if((dot3F4(-DeltaC2,*sep))>0.0f)
+	{
+		*sep = -(*sep);
+	}
+	return true;
+}
+
+
+
+bool findSeparatingAxisEdgeEdgeLocalA(	const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, 
+	const float4 posA1,
+	const float4 ornA,
+	const float4 posB1,
+	const float4 ornB,
+	const float4 DeltaC2,
+	const float4* verticesA, 
+	const float4* uniqueEdgesA, 
+	const btGpuFace* facesA,
+	const int*  indicesA,
+	__global const float4* verticesB, 
+	__global const float4* uniqueEdgesB, 
+	__global const btGpuFace* facesB,
+	__global const int*  indicesB,
+		float4* sep,
+	float* dmin)
+{
+
+
+	float4 posA = posA1;
+	posA.w = 0.f;
+	float4 posB = posB1;
+	posB.w = 0.f;
+
+	int curPlaneTests=0;
+
+	int curEdgeEdge = 0;
+	// Test edges
+	for(int e0=0;e0<hullA->m_numUniqueEdges;e0++)
+	{
+		const float4 edge0 = uniqueEdgesA[hullA->m_uniqueEdgesOffset+e0];
+		float4 edge0World = qtRotate(ornA,edge0);
+
+		for(int e1=0;e1<hullB->m_numUniqueEdges;e1++)
+		{
+			const float4 edge1 = uniqueEdgesB[hullB->m_uniqueEdgesOffset+e1];
+			float4 edge1World = qtRotate(ornB,edge1);
+
+
+			float4 crossje = cross3(edge0World,edge1World);
+
+			curEdgeEdge++;
+			if(!IsAlmostZero(crossje))
+			{
+				crossje = normalize3(crossje);
+				if (dot3F4(DeltaC2,crossje)<0)
+					crossje *= -1.f;
+
+				float dist;
+				bool result = true;
+				{
+					float Min0,Max0;
+					float Min1,Max1;
+					projectLocal(hullA,posA,ornA,&crossje,verticesA, &Min0, &Max0);
+					project(hullB,posB,ornB,&crossje,verticesB, &Min1, &Max1);
+				
+					if(Max0<Min1 || Max1<Min0)
+						result = false;
+				
+					float d0 = Max0 - Min1;
+					float d1 = Max1 - Min0;
+					dist = d0<d1 ? d0:d1;
+					result = true;
+
+				}
+				
+
+				if(dist<*dmin)
+				{
+					*dmin = dist;
+					*sep = crossje;
+				}
+			}
+		}
+
+	}
+
+	
+	if((dot3F4(-DeltaC2,*sep))>0.0f)
+	{
+		*sep = -(*sep);
+	}
+	return true;
+}
+
+
+
+inline int	findClippingFaces(const float4 separatingNormal,
+                      const ConvexPolyhedronCL* hullA, 
+					  __global const ConvexPolyhedronCL* hullB,
+                      const float4 posA, const Quaternion ornA,const float4 posB, const Quaternion ornB,
+                       __global float4* worldVertsA1,
+                      __global float4* worldNormalsA1,
+                      __global float4* worldVertsB1,
+                      int capacityWorldVerts,
+                      const float minDist, float maxDist,
+					  const float4* verticesA,
+                      const btGpuFace* facesA,
+                      const int* indicesA,
+					  __global const float4* verticesB,
+                      __global const btGpuFace* facesB,
+                      __global const int* indicesB,
+                      __global int4* clippingFaces, int pairIndex)
+{
+	int numContactsOut = 0;
+	int numWorldVertsB1= 0;
+    
+    
+	int closestFaceB=0;
+	float dmax = -FLT_MAX;
+    
+	{
+		for(int face=0;face<hullB->m_numFaces;face++)
+		{
+			const float4 Normal = make_float4(facesB[hullB->m_faceOffset+face].m_plane.x,
+                                              facesB[hullB->m_faceOffset+face].m_plane.y, facesB[hullB->m_faceOffset+face].m_plane.z,0.f);
+			const float4 WorldNormal = qtRotate(ornB, Normal);
+			float d = dot3F4(WorldNormal,separatingNormal);
+			if (d > dmax)
+			{
+				dmax = d;
+				closestFaceB = face;
+			}
+		}
+	}
+    
+	{
+		const btGpuFace polyB = facesB[hullB->m_faceOffset+closestFaceB];
+		int numVertices = polyB.m_numIndices;
+        if (numVertices>capacityWorldVerts)
+            numVertices = capacityWorldVerts;
+        if (numVertices<0)
+            numVertices = 0;
+        
+		for(int e0=0;e0<numVertices;e0++)
+		{
+            if (e0<capacityWorldVerts)
+            {
+                const float4 b = verticesB[hullB->m_vertexOffset+indicesB[polyB.m_indexOffset+e0]];
+                worldVertsB1[pairIndex*capacityWorldVerts+numWorldVertsB1++] = transform(&b,&posB,&ornB);
+            }
+		}
+	}
+    
+    int closestFaceA=0;
+	{
+		float dmin = FLT_MAX;
+		for(int face=0;face<hullA->m_numFaces;face++)
+		{
+			const float4 Normal = make_float4(
+                                              facesA[hullA->m_faceOffset+face].m_plane.x,
+                                              facesA[hullA->m_faceOffset+face].m_plane.y,
+                                              facesA[hullA->m_faceOffset+face].m_plane.z,
+                                              0.f);
+			const float4 faceANormalWS = qtRotate(ornA,Normal);
+            
+			float d = dot3F4(faceANormalWS,separatingNormal);
+			if (d < dmin)
+			{
+				dmin = d;
+				closestFaceA = face;
+                worldNormalsA1[pairIndex] = faceANormalWS;
+			}
+		}
+	}
+    
+    int numVerticesA = facesA[hullA->m_faceOffset+closestFaceA].m_numIndices;
+    if (numVerticesA>capacityWorldVerts)
+       numVerticesA = capacityWorldVerts;
+    if (numVerticesA<0)
+        numVerticesA=0;
+    
+	for(int e0=0;e0<numVerticesA;e0++)
+	{
+        if (e0<capacityWorldVerts)
+        {
+            const float4 a = verticesA[hullA->m_vertexOffset+indicesA[facesA[hullA->m_faceOffset+closestFaceA].m_indexOffset+e0]];
+            worldVertsA1[pairIndex*capacityWorldVerts+e0] = transform(&a, &posA,&ornA);
+        }
+    }
+    
+    clippingFaces[pairIndex].x = closestFaceA;
+    clippingFaces[pairIndex].y = closestFaceB;
+    clippingFaces[pairIndex].z = numVerticesA;
+    clippingFaces[pairIndex].w = numWorldVertsB1;
+    
+    
+	return numContactsOut;
+}
+
+
+
+
+// work-in-progress
+__kernel void   findConcaveSeparatingAxisVertexFaceKernel( __global int4* concavePairs,
+                                                __global const BodyData* rigidBodies,
+                                                __global const btCollidableGpu* collidables,
+                                                __global const ConvexPolyhedronCL* convexShapes,
+                                                __global const float4* vertices,
+                                                __global const float4* uniqueEdges,
+                                                __global const btGpuFace* faces,
+                                                __global const int* indices,
+                                                __global const btGpuChildShape* gpuChildShapes,
+                                                __global btAabbCL* aabbs,
+                                                __global float4* concaveSeparatingNormalsOut,
+                                                __global int* concaveHasSeparatingNormals,
+                                                __global int4* clippingFacesOut,
+                                                __global float4* worldVertsA1GPU,
+                                                __global float4*  worldNormalsAGPU,
+                                                __global float4* worldVertsB1GPU,
+                                                __global float* dmins,
+                                                int vertexFaceCapacity,
+                                                int numConcavePairs
+                                                )
+{
+    
+	int i = get_global_id(0);
+	if (i>=numConcavePairs)
+		return;
+    
+	concaveHasSeparatingNormals[i] = 0;
+    
+	int pairIdx = i;
+    
+	int bodyIndexA = concavePairs[i].x;
+	int bodyIndexB = concavePairs[i].y;
+    
+	int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+	int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+    
+	int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+	int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+    
+	if (collidables[collidableIndexB].m_shapeType!=SHAPE_CONVEX_HULL&&
+		collidables[collidableIndexB].m_shapeType!=SHAPE_COMPOUND_OF_CONVEX_HULLS)
+	{
+		concavePairs[pairIdx].w = -1;
+		return;
+	}
+    
+    
+    
+	int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+	int numActualConcaveConvexTests = 0;
+	
+	int f = concavePairs[i].z;
+	
+	bool overlap = false;
+	
+	ConvexPolyhedronCL convexPolyhedronA;
+    
+	//add 3 vertices of the triangle
+	convexPolyhedronA.m_numVertices = 3;
+	convexPolyhedronA.m_vertexOffset = 0;
+	float4	localCenter = make_float4(0.f,0.f,0.f,0.f);
+    
+	btGpuFace face = faces[convexShapes[shapeIndexA].m_faceOffset+f];
+	float4 triMinAabb, triMaxAabb;
+	btAabbCL triAabb;
+	triAabb.m_min = make_float4(1e30f,1e30f,1e30f,0.f);
+	triAabb.m_max = make_float4(-1e30f,-1e30f,-1e30f,0.f);
+	
+	float4 verticesA[3];
+	for (int i=0;i<3;i++)
+	{
+		int index = indices[face.m_indexOffset+i];
+		float4 vert = vertices[convexShapes[shapeIndexA].m_vertexOffset+index];
+		verticesA[i] = vert;
+		localCenter += vert;
+        
+		triAabb.m_min = min(triAabb.m_min,vert);
+		triAabb.m_max = max(triAabb.m_max,vert);
+        
+	}
+    
+	overlap = true;
+	overlap = (triAabb.m_min.x > aabbs[bodyIndexB].m_max.x || triAabb.m_max.x < aabbs[bodyIndexB].m_min.x) ? false : overlap;
+	overlap = (triAabb.m_min.z > aabbs[bodyIndexB].m_max.z || triAabb.m_max.z < aabbs[bodyIndexB].m_min.z) ? false : overlap;
+	overlap = (triAabb.m_min.y > aabbs[bodyIndexB].m_max.y || triAabb.m_max.y < aabbs[bodyIndexB].m_min.y) ? false : overlap;
+    
+	if (overlap)
+	{
+		float dmin = FLT_MAX;
+		int hasSeparatingAxis=5;
+		float4 sepAxis=make_float4(1,2,3,4);
+        
+		int localCC=0;
+		numActualConcaveConvexTests++;
+        
+		//a triangle has 3 unique edges
+		convexPolyhedronA.m_numUniqueEdges = 3;
+		convexPolyhedronA.m_uniqueEdgesOffset = 0;
+		float4 uniqueEdgesA[3];
+		
+		uniqueEdgesA[0] = (verticesA[1]-verticesA[0]);
+		uniqueEdgesA[1] = (verticesA[2]-verticesA[1]);
+		uniqueEdgesA[2] = (verticesA[0]-verticesA[2]);
+        
+        
+		convexPolyhedronA.m_faceOffset = 0;
+        
+		float4 normal = make_float4(face.m_plane.x,face.m_plane.y,face.m_plane.z,0.f);
+        
+		btGpuFace facesA[TRIANGLE_NUM_CONVEX_FACES];
+		int indicesA[3+3+2+2+2];
+		int curUsedIndices=0;
+		int fidx=0;
+        
+		//front size of triangle
+		{
+			facesA[fidx].m_indexOffset=curUsedIndices;
+			indicesA[0] = 0;
+			indicesA[1] = 1;
+			indicesA[2] = 2;
+			curUsedIndices+=3;
+			float c = face.m_plane.w;
+			facesA[fidx].m_plane.x = normal.x;
+			facesA[fidx].m_plane.y = normal.y;
+			facesA[fidx].m_plane.z = normal.z;
+			facesA[fidx].m_plane.w = c;
+			facesA[fidx].m_numIndices=3;
+		}
+		fidx++;
+		//back size of triangle
+		{
+			facesA[fidx].m_indexOffset=curUsedIndices;
+			indicesA[3]=2;
+			indicesA[4]=1;
+			indicesA[5]=0;
+			curUsedIndices+=3;
+			float c = dot(normal,verticesA[0]);
+			float c1 = -face.m_plane.w;
+			facesA[fidx].m_plane.x = -normal.x;
+			facesA[fidx].m_plane.y = -normal.y;
+			facesA[fidx].m_plane.z = -normal.z;
+			facesA[fidx].m_plane.w = c;
+			facesA[fidx].m_numIndices=3;
+		}
+		fidx++;
+        
+		bool addEdgePlanes = true;
+		if (addEdgePlanes)
+		{
+			int numVertices=3;
+			int prevVertex = numVertices-1;
+			for (int i=0;i<numVertices;i++)
+			{
+				float4 v0 = verticesA[i];
+				float4 v1 = verticesA[prevVertex];
+                
+				float4 edgeNormal = normalize(cross(normal,v1-v0));
+				float c = -dot(edgeNormal,v0);
+                
+				facesA[fidx].m_numIndices = 2;
+				facesA[fidx].m_indexOffset=curUsedIndices;
+				indicesA[curUsedIndices++]=i;
+				indicesA[curUsedIndices++]=prevVertex;
+                
+				facesA[fidx].m_plane.x = edgeNormal.x;
+				facesA[fidx].m_plane.y = edgeNormal.y;
+				facesA[fidx].m_plane.z = edgeNormal.z;
+				facesA[fidx].m_plane.w = c;
+				fidx++;
+				prevVertex = i;
+			}
+		}
+		convexPolyhedronA.m_numFaces = TRIANGLE_NUM_CONVEX_FACES;
+		convexPolyhedronA.m_localCenter = localCenter*(1.f/3.f);
+        
+        
+		float4 posA = rigidBodies[bodyIndexA].m_pos;
+		posA.w = 0.f;
+		float4 posB = rigidBodies[bodyIndexB].m_pos;
+		posB.w = 0.f;
+        
+		float4 ornA = rigidBodies[bodyIndexA].m_quat;
+		float4 ornB =rigidBodies[bodyIndexB].m_quat;
+        
+		
+        
+        
+		///////////////////
+		///compound shape support
+        
+		if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+		{
+			int compoundChild = concavePairs[pairIdx].w;
+			int childShapeIndexB = compoundChild;//collidables[collidableIndexB].m_shapeIndex+compoundChild;
+			int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+			float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+			float4 newPosB = transform(&childPosB,&posB,&ornB);
+			float4 newOrnB = qtMul(ornB,childOrnB);
+			posB = newPosB;
+			ornB = newOrnB;
+			shapeIndexB = collidables[childColIndexB].m_shapeIndex;
+		}
+		//////////////////
+        
+		float4 c0local = convexPolyhedronA.m_localCenter;
+		float4 c0 = transform(&c0local, &posA, &ornA);
+		float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+		float4 c1 = transform(&c1local,&posB,&ornB);
+		const float4 DeltaC2 = c0 - c1;
+        
+        
+		bool sepA = findSeparatingAxisLocalA(	&convexPolyhedronA, &convexShapes[shapeIndexB],
+                                             posA,ornA,
+                                             posB,ornB,
+                                             DeltaC2,
+                                             verticesA,uniqueEdgesA,facesA,indicesA,
+                                             vertices,uniqueEdges,faces,indices,
+                                             &sepAxis,&dmin);
+		hasSeparatingAxis = 4;
+		if (!sepA)
+		{
+			hasSeparatingAxis = 0;
+		} else
+		{
+			bool sepB = findSeparatingAxisLocalB(	&convexShapes[shapeIndexB],&convexPolyhedronA,
+                                                 posB,ornB,
+                                                 posA,ornA,
+                                                 DeltaC2,
+                                                 vertices,uniqueEdges,faces,indices,
+                                                 verticesA,uniqueEdgesA,facesA,indicesA,
+                                                 &sepAxis,&dmin);
+            
+			if (!sepB)
+			{
+				hasSeparatingAxis = 0;
+			} else
+			{
+				hasSeparatingAxis = 1;
+			}
+		}	
+		
+		if (hasSeparatingAxis)
+		{
+            dmins[i] = dmin;
+			concaveSeparatingNormalsOut[pairIdx]=sepAxis;
+			concaveHasSeparatingNormals[i]=1;
+            
+		} else
+		{	
+			//mark this pair as in-active
+			concavePairs[pairIdx].w = -1;
+		}
+	}
+	else
+	{	
+		//mark this pair as in-active
+		concavePairs[pairIdx].w = -1;
+	}
+}
+
+
+
+
+// work-in-progress
+__kernel void   findConcaveSeparatingAxisEdgeEdgeKernel( __global int4* concavePairs,
+                                                          __global const BodyData* rigidBodies,
+                                                          __global const btCollidableGpu* collidables,
+                                                          __global const ConvexPolyhedronCL* convexShapes,
+                                                          __global const float4* vertices,
+                                                          __global const float4* uniqueEdges,
+                                                          __global const btGpuFace* faces,
+                                                          __global const int* indices,
+                                                          __global const btGpuChildShape* gpuChildShapes,
+                                                          __global btAabbCL* aabbs,
+                                                          __global float4* concaveSeparatingNormalsOut,
+                                                          __global int* concaveHasSeparatingNormals,
+                                                          __global int4* clippingFacesOut,
+                                                          __global float4* worldVertsA1GPU,
+                                                          __global float4*  worldNormalsAGPU,
+                                                          __global float4* worldVertsB1GPU,
+                                                          __global float* dmins,
+                                                          int vertexFaceCapacity,
+                                                          int numConcavePairs
+                                                          )
+{
+    
+	int i = get_global_id(0);
+	if (i>=numConcavePairs)
+		return;
+    
+	if (!concaveHasSeparatingNormals[i])
+        return;
+    
+	int pairIdx = i;
+    
+	int bodyIndexA = concavePairs[i].x;
+	int bodyIndexB = concavePairs[i].y;
+    
+	int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;
+	int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;
+    
+	int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;
+	int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;
+    
+    
+	int numFacesA = convexShapes[shapeIndexA].m_numFaces;
+	int numActualConcaveConvexTests = 0;
+	
+	int f = concavePairs[i].z;
+	
+	bool overlap = false;
+	
+	ConvexPolyhedronCL convexPolyhedronA;
+    
+	//add 3 vertices of the triangle
+	convexPolyhedronA.m_numVertices = 3;
+	convexPolyhedronA.m_vertexOffset = 0;
+	float4	localCenter = make_float4(0.f,0.f,0.f,0.f);
+    
+	btGpuFace face = faces[convexShapes[shapeIndexA].m_faceOffset+f];
+	float4 triMinAabb, triMaxAabb;
+	btAabbCL triAabb;
+	triAabb.m_min = make_float4(1e30f,1e30f,1e30f,0.f);
+	triAabb.m_max = make_float4(-1e30f,-1e30f,-1e30f,0.f);
+	
+	float4 verticesA[3];
+	for (int i=0;i<3;i++)
+	{
+		int index = indices[face.m_indexOffset+i];
+		float4 vert = vertices[convexShapes[shapeIndexA].m_vertexOffset+index];
+		verticesA[i] = vert;
+		localCenter += vert;
+        
+		triAabb.m_min = min(triAabb.m_min,vert);
+		triAabb.m_max = max(triAabb.m_max,vert);
+        
+	}
+    
+	overlap = true;
+	overlap = (triAabb.m_min.x > aabbs[bodyIndexB].m_max.x || triAabb.m_max.x < aabbs[bodyIndexB].m_min.x) ? false : overlap;
+	overlap = (triAabb.m_min.z > aabbs[bodyIndexB].m_max.z || triAabb.m_max.z < aabbs[bodyIndexB].m_min.z) ? false : overlap;
+	overlap = (triAabb.m_min.y > aabbs[bodyIndexB].m_max.y || triAabb.m_max.y < aabbs[bodyIndexB].m_min.y) ? false : overlap;
+    
+	if (overlap)
+	{
+		float dmin = dmins[i];
+		int hasSeparatingAxis=5;
+		float4 sepAxis=make_float4(1,2,3,4);
+        sepAxis = concaveSeparatingNormalsOut[pairIdx];
+        
+		int localCC=0;
+		numActualConcaveConvexTests++;
+        
+		//a triangle has 3 unique edges
+		convexPolyhedronA.m_numUniqueEdges = 3;
+		convexPolyhedronA.m_uniqueEdgesOffset = 0;
+		float4 uniqueEdgesA[3];
+		
+		uniqueEdgesA[0] = (verticesA[1]-verticesA[0]);
+		uniqueEdgesA[1] = (verticesA[2]-verticesA[1]);
+		uniqueEdgesA[2] = (verticesA[0]-verticesA[2]);
+        
+        
+		convexPolyhedronA.m_faceOffset = 0;
+        
+		float4 normal = make_float4(face.m_plane.x,face.m_plane.y,face.m_plane.z,0.f);
+        
+		btGpuFace facesA[TRIANGLE_NUM_CONVEX_FACES];
+		int indicesA[3+3+2+2+2];
+		int curUsedIndices=0;
+		int fidx=0;
+        
+		//front size of triangle
+		{
+			facesA[fidx].m_indexOffset=curUsedIndices;
+			indicesA[0] = 0;
+			indicesA[1] = 1;
+			indicesA[2] = 2;
+			curUsedIndices+=3;
+			float c = face.m_plane.w;
+			facesA[fidx].m_plane.x = normal.x;
+			facesA[fidx].m_plane.y = normal.y;
+			facesA[fidx].m_plane.z = normal.z;
+			facesA[fidx].m_plane.w = c;
+			facesA[fidx].m_numIndices=3;
+		}
+		fidx++;
+		//back size of triangle
+		{
+			facesA[fidx].m_indexOffset=curUsedIndices;
+			indicesA[3]=2;
+			indicesA[4]=1;
+			indicesA[5]=0;
+			curUsedIndices+=3;
+			float c = dot(normal,verticesA[0]);
+			float c1 = -face.m_plane.w;
+			facesA[fidx].m_plane.x = -normal.x;
+			facesA[fidx].m_plane.y = -normal.y;
+			facesA[fidx].m_plane.z = -normal.z;
+			facesA[fidx].m_plane.w = c;
+			facesA[fidx].m_numIndices=3;
+		}
+		fidx++;
+        
+		bool addEdgePlanes = true;
+		if (addEdgePlanes)
+		{
+			int numVertices=3;
+			int prevVertex = numVertices-1;
+			for (int i=0;i<numVertices;i++)
+			{
+				float4 v0 = verticesA[i];
+				float4 v1 = verticesA[prevVertex];
+                
+				float4 edgeNormal = normalize(cross(normal,v1-v0));
+				float c = -dot(edgeNormal,v0);
+                
+				facesA[fidx].m_numIndices = 2;
+				facesA[fidx].m_indexOffset=curUsedIndices;
+				indicesA[curUsedIndices++]=i;
+				indicesA[curUsedIndices++]=prevVertex;
+                
+				facesA[fidx].m_plane.x = edgeNormal.x;
+				facesA[fidx].m_plane.y = edgeNormal.y;
+				facesA[fidx].m_plane.z = edgeNormal.z;
+				facesA[fidx].m_plane.w = c;
+				fidx++;
+				prevVertex = i;
+			}
+		}
+		convexPolyhedronA.m_numFaces = TRIANGLE_NUM_CONVEX_FACES;
+		convexPolyhedronA.m_localCenter = localCenter*(1.f/3.f);
+        
+        
+		float4 posA = rigidBodies[bodyIndexA].m_pos;
+		posA.w = 0.f;
+		float4 posB = rigidBodies[bodyIndexB].m_pos;
+		posB.w = 0.f;
+        
+		float4 ornA = rigidBodies[bodyIndexA].m_quat;
+		float4 ornB =rigidBodies[bodyIndexB].m_quat;
+        
+		
+        
+        
+		///////////////////
+		///compound shape support
+        
+		if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)
+		{
+			int compoundChild = concavePairs[pairIdx].w;
+			int childShapeIndexB = compoundChild;//collidables[collidableIndexB].m_shapeIndex+compoundChild;
+			int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;
+			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;
+			float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;
+			float4 newPosB = transform(&childPosB,&posB,&ornB);
+			float4 newOrnB = qtMul(ornB,childOrnB);
+			posB = newPosB;
+			ornB = newOrnB;
+			shapeIndexB = collidables[childColIndexB].m_shapeIndex;
+		}
+		//////////////////
+        
+		float4 c0local = convexPolyhedronA.m_localCenter;
+		float4 c0 = transform(&c0local, &posA, &ornA);
+		float4 c1local = convexShapes[shapeIndexB].m_localCenter;
+		float4 c1 = transform(&c1local,&posB,&ornB);
+		const float4 DeltaC2 = c0 - c1;
+        
+        
+		{
+			bool sepEE = findSeparatingAxisEdgeEdgeLocalA(	&convexPolyhedronA, &convexShapes[shapeIndexB],
+                                                              posA,ornA,
+                                                              posB,ornB,
+                                                              DeltaC2,
+                                                              verticesA,uniqueEdgesA,facesA,indicesA,
+                                                              vertices,uniqueEdges,faces,indices,
+                                                              &sepAxis,&dmin);
+                
+			if (!sepEE)
+			{
+				hasSeparatingAxis = 0;
+			} else
+			{
+				hasSeparatingAxis = 1;
+			}
+		}
+		
+		
+		if (hasSeparatingAxis)
+		{
+			sepAxis.w = dmin;
+            dmins[i] = dmin;
+			concaveSeparatingNormalsOut[pairIdx]=sepAxis;
+			concaveHasSeparatingNormals[i]=1;
+           
+ 	float minDist = -1e30f;
+			float maxDist = 0.02f;
+
+            
+            findClippingFaces(sepAxis,
+                              &convexPolyhedronA,
+                              &convexShapes[shapeIndexB],
+                              posA,ornA,
+                              posB,ornB,
+                              worldVertsA1GPU,
+                              worldNormalsAGPU,
+                              worldVertsB1GPU,
+                              vertexFaceCapacity,
+                              minDist, maxDist,
+                              verticesA,
+                              facesA,
+                              indicesA,
+                              vertices,
+                              faces,
+                              indices,
+                              clippingFacesOut, pairIdx);
+	           
+            
+		} else
+		{	
+			//mark this pair as in-active
+			concavePairs[pairIdx].w = -1;
+		}
+	}
+	else
+	{	
+		//mark this pair as in-active
+		concavePairs[pairIdx].w = -1;
+	}
+	
+	concavePairs[i].z = -1;//for the next stage, z is used to determine existing contact points
+}
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satConcaveKernels.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satConcaveKernels.h
new file mode 100644
index 0000000000..611569cacf
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satConcaveKernels.h
@@ -0,0 +1,1457 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* satConcaveKernelsCL= \
+"//keep this enum in sync with the CPU version (in btCollidable.h)\n"
+"//written by Erwin Coumans\n"
+"#define SHAPE_CONVEX_HULL 3\n"
+"#define SHAPE_CONCAVE_TRIMESH 5\n"
+"#define TRIANGLE_NUM_CONVEX_FACES 5\n"
+"#define SHAPE_COMPOUND_OF_CONVEX_HULLS 6\n"
+"#define B3_MAX_STACK_DEPTH 256\n"
+"typedef unsigned int u32;\n"
+"///keep this in sync with btCollidable.h\n"
+"typedef struct\n"
+"{\n"
+"	union {\n"
+"		int m_numChildShapes;\n"
+"		int m_bvhIndex;\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float m_radius;\n"
+"		int	m_compoundBvhIndex;\n"
+"	};\n"
+"	\n"
+"	int m_shapeType;\n"
+"	int m_shapeIndex;\n"
+"	\n"
+"} btCollidableGpu;\n"
+"#define MAX_NUM_PARTS_IN_BITS 10\n"
+"///b3QuantizedBvhNode is a compressed aabb node, 16 bytes.\n"
+"///Node can be used for leafnode or internal node. Leafnodes can point to 32-bit triangle index (non-negative range).\n"
+"typedef struct\n"
+"{\n"
+"	//12 bytes\n"
+"	unsigned short int	m_quantizedAabbMin[3];\n"
+"	unsigned short int	m_quantizedAabbMax[3];\n"
+"	//4 bytes\n"
+"	int	m_escapeIndexOrTriangleIndex;\n"
+"} b3QuantizedBvhNode;\n"
+"typedef struct\n"
+"{\n"
+"	float4		m_aabbMin;\n"
+"	float4		m_aabbMax;\n"
+"	float4		m_quantization;\n"
+"	int			m_numNodes;\n"
+"	int			m_numSubTrees;\n"
+"	int			m_nodeOffset;\n"
+"	int			m_subTreeOffset;\n"
+"} b3BvhInfo;\n"
+"int	getTriangleIndex(const b3QuantizedBvhNode* rootNode)\n"
+"{\n"
+"	unsigned int x=0;\n"
+"	unsigned int y = (~(x&0))<<(31-MAX_NUM_PARTS_IN_BITS);\n"
+"	// Get only the lower bits where the triangle index is stored\n"
+"	return (rootNode->m_escapeIndexOrTriangleIndex&~(y));\n"
+"}\n"
+"int	getTriangleIndexGlobal(__global const b3QuantizedBvhNode* rootNode)\n"
+"{\n"
+"	unsigned int x=0;\n"
+"	unsigned int y = (~(x&0))<<(31-MAX_NUM_PARTS_IN_BITS);\n"
+"	// Get only the lower bits where the triangle index is stored\n"
+"	return (rootNode->m_escapeIndexOrTriangleIndex&~(y));\n"
+"}\n"
+"int isLeafNode(const b3QuantizedBvhNode* rootNode)\n"
+"{\n"
+"	//skipindex is negative (internal node), triangleindex >=0 (leafnode)\n"
+"	return (rootNode->m_escapeIndexOrTriangleIndex >= 0)? 1 : 0;\n"
+"}\n"
+"int isLeafNodeGlobal(__global const b3QuantizedBvhNode* rootNode)\n"
+"{\n"
+"	//skipindex is negative (internal node), triangleindex >=0 (leafnode)\n"
+"	return (rootNode->m_escapeIndexOrTriangleIndex >= 0)? 1 : 0;\n"
+"}\n"
+"	\n"
+"int getEscapeIndex(const b3QuantizedBvhNode* rootNode)\n"
+"{\n"
+"	return -rootNode->m_escapeIndexOrTriangleIndex;\n"
+"}\n"
+"int getEscapeIndexGlobal(__global const b3QuantizedBvhNode* rootNode)\n"
+"{\n"
+"	return -rootNode->m_escapeIndexOrTriangleIndex;\n"
+"}\n"
+"typedef struct\n"
+"{\n"
+"	//12 bytes\n"
+"	unsigned short int	m_quantizedAabbMin[3];\n"
+"	unsigned short int	m_quantizedAabbMax[3];\n"
+"	//4 bytes, points to the root of the subtree\n"
+"	int			m_rootNodeIndex;\n"
+"	//4 bytes\n"
+"	int			m_subtreeSize;\n"
+"	int			m_padding[3];\n"
+"} b3BvhSubtreeInfo;\n"
+"typedef struct\n"
+"{\n"
+"	float4	m_childPosition;\n"
+"	float4	m_childOrientation;\n"
+"	int m_shapeIndex;\n"
+"	int m_unused0;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"} btGpuChildShape;\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	float4 m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"	u32 m_collidableIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} BodyData;\n"
+"typedef struct  \n"
+"{\n"
+"	float4		m_localCenter;\n"
+"	float4		m_extents;\n"
+"	float4		mC;\n"
+"	float4		mE;\n"
+"	\n"
+"	float			m_radius;\n"
+"	int	m_faceOffset;\n"
+"	int m_numFaces;\n"
+"	int	m_numVertices;\n"
+"	int m_vertexOffset;\n"
+"	int	m_uniqueEdgesOffset;\n"
+"	int	m_numUniqueEdges;\n"
+"	int m_unused;\n"
+"} ConvexPolyhedronCL;\n"
+"typedef struct \n"
+"{\n"
+"	union\n"
+"	{\n"
+"		float4	m_min;\n"
+"		float   m_minElems[4];\n"
+"		int			m_minIndices[4];\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float4	m_max;\n"
+"		float   m_maxElems[4];\n"
+"		int			m_maxIndices[4];\n"
+"	};\n"
+"} btAabbCL;\n"
+"#ifndef B3_AABB_H\n"
+"#define B3_AABB_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#define B3_FLOAT4_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#define B3_PLATFORM_DEFINITIONS_H\n"
+"struct MyTest\n"
+"{\n"
+"	int bla;\n"
+"};\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"//keep B3_LARGE_FLOAT*B3_LARGE_FLOAT < FLT_MAX\n"
+"#define B3_LARGE_FLOAT 1e18f\n"
+"#define B3_INFINITY 1e18f\n"
+"#define b3Assert(a)\n"
+"#define b3ConstArray(a) __global const a*\n"
+"#define b3AtomicInc atomic_inc\n"
+"#define b3AtomicAdd atomic_add\n"
+"#define b3Fabs fabs\n"
+"#define b3Sqrt native_sqrt\n"
+"#define b3Sin native_sin\n"
+"#define b3Cos native_cos\n"
+"#define B3_STATIC\n"
+"#endif\n"
+"#endif\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Float4;\n"
+"	#define b3Float4ConstArg const b3Float4\n"
+"	#define b3MakeFloat4 (float4)\n"
+"	float b3Dot3F4(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return dot(a1, b1);\n"
+"	}\n"
+"	b3Float4 b3Cross3(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return cross(a1, b1);\n"
+"	}\n"
+"	#define b3MinFloat4 min\n"
+"	#define b3MaxFloat4 max\n"
+"	#define b3Normalized(a) normalize(a)\n"
+"#endif \n"
+"		\n"
+"inline bool b3IsAlmostZero(b3Float4ConstArg v)\n"
+"{\n"
+"	if(b3Fabs(v.x)>1e-6 || b3Fabs(v.y)>1e-6 || b3Fabs(v.z)>1e-6)	\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"inline int    b3MaxDot( b3Float4ConstArg vec, __global const b3Float4* vecArray, int vecLen, float* dotOut )\n"
+"{\n"
+"    float maxDot = -B3_INFINITY;\n"
+"    int i = 0;\n"
+"    int ptIndex = -1;\n"
+"    for( i = 0; i < vecLen; i++ )\n"
+"    {\n"
+"        float dot = b3Dot3F4(vecArray[i],vec);\n"
+"            \n"
+"        if( dot > maxDot )\n"
+"        {\n"
+"            maxDot = dot;\n"
+"            ptIndex = i;\n"
+"        }\n"
+"    }\n"
+"	b3Assert(ptIndex>=0);\n"
+"    if (ptIndex<0)\n"
+"	{\n"
+"		ptIndex = 0;\n"
+"	}\n"
+"    *dotOut = maxDot;\n"
+"    return ptIndex;\n"
+"}\n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_MAT3x3_H\n"
+"#define B3_MAT3x3_H\n"
+"#ifndef B3_QUAT_H\n"
+"#define B3_QUAT_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif\n"
+"#endif\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Quat;\n"
+"	#define b3QuatConstArg const b3Quat\n"
+"	\n"
+"	\n"
+"inline float4 b3FastNormalize4(float4 v)\n"
+"{\n"
+"	v = (float4)(v.xyz,0.f);\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"	\n"
+"inline b3Quat b3QuatMul(b3Quat a, b3Quat b);\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in);\n"
+"inline b3Quat b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec);\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatMul(b3QuatConstArg a, b3QuatConstArg b)\n"
+"{\n"
+"	b3Quat ans;\n"
+"	ans = b3Cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - b3Dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in)\n"
+"{\n"
+"	b3Quat q;\n"
+"	q=in;\n"
+"	//return b3FastNormalize4(in);\n"
+"	float len = native_sqrt(dot(q, q));\n"
+"	if(len > 0.f)\n"
+"	{\n"
+"		q *= 1.f / len;\n"
+"	}\n"
+"	else\n"
+"	{\n"
+"		q.x = q.y = q.z = 0.f;\n"
+"		q.w = 1.f;\n"
+"	}\n"
+"	return q;\n"
+"}\n"
+"inline float4 b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	b3Quat qInv = b3QuatInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = b3QuatMul(b3QuatMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline float4 b3QuatInvRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	return b3QuatRotate( b3QuatInvert( q ), vec );\n"
+"}\n"
+"inline b3Float4 b3TransformPoint(b3Float4ConstArg point, b3Float4ConstArg translation, b3QuatConstArg  orientation)\n"
+"{\n"
+"	return b3QuatRotate( orientation, point ) + (translation);\n"
+"}\n"
+"	\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"typedef struct\n"
+"{\n"
+"	b3Float4 m_row[3];\n"
+"}b3Mat3x3;\n"
+"#define b3Mat3x3ConstArg const b3Mat3x3\n"
+"#define b3GetRow(m,row) (m.m_row[row])\n"
+"inline b3Mat3x3 b3QuatGetRotationMatrix(b3Quat quat)\n"
+"{\n"
+"	b3Float4 quat2 = (b3Float4)(quat.x*quat.x, quat.y*quat.y, quat.z*quat.z, 0.f);\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0].x=1-2*quat2.y-2*quat2.z;\n"
+"	out.m_row[0].y=2*quat.x*quat.y-2*quat.w*quat.z;\n"
+"	out.m_row[0].z=2*quat.x*quat.z+2*quat.w*quat.y;\n"
+"	out.m_row[0].w = 0.f;\n"
+"	out.m_row[1].x=2*quat.x*quat.y+2*quat.w*quat.z;\n"
+"	out.m_row[1].y=1-2*quat2.x-2*quat2.z;\n"
+"	out.m_row[1].z=2*quat.y*quat.z-2*quat.w*quat.x;\n"
+"	out.m_row[1].w = 0.f;\n"
+"	out.m_row[2].x=2*quat.x*quat.z-2*quat.w*quat.y;\n"
+"	out.m_row[2].y=2*quat.y*quat.z+2*quat.w*quat.x;\n"
+"	out.m_row[2].z=1-2*quat2.x-2*quat2.y;\n"
+"	out.m_row[2].w = 0.f;\n"
+"	return out;\n"
+"}\n"
+"inline b3Mat3x3 b3AbsoluteMat3x3(b3Mat3x3ConstArg matIn)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = fabs(matIn.m_row[0]);\n"
+"	out.m_row[1] = fabs(matIn.m_row[1]);\n"
+"	out.m_row[2] = fabs(matIn.m_row[2]);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtZero();\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity();\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m);\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b);\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Mat3x3 mtZero()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(0.f);\n"
+"	m.m_row[1] = (b3Float4)(0.f);\n"
+"	m.m_row[2] = (b3Float4)(0.f);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(1,0,0,0);\n"
+"	m.m_row[1] = (b3Float4)(0,1,0,0);\n"
+"	m.m_row[2] = (b3Float4)(0,0,1,0);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = (b3Float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (b3Float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (b3Float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Mat3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	b3Mat3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = b3Dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = b3Dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = b3Dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b)\n"
+"{\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a.m_row[0], b );\n"
+"	ans.y = b3Dot3F4( a.m_row[1], b );\n"
+"	ans.z = b3Dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Float4 colx = b3MakeFloat4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	b3Float4 coly = b3MakeFloat4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	b3Float4 colz = b3MakeFloat4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a, colx );\n"
+"	ans.y = b3Dot3F4( a, coly );\n"
+"	ans.z = b3Dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"#endif\n"
+"#endif //B3_MAT3x3_H\n"
+"typedef struct b3Aabb b3Aabb_t;\n"
+"struct b3Aabb\n"
+"{\n"
+"	union\n"
+"	{\n"
+"		float m_min[4];\n"
+"		b3Float4 m_minVec;\n"
+"		int m_minIndices[4];\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float	m_max[4];\n"
+"		b3Float4 m_maxVec;\n"
+"		int m_signedMaxIndices[4];\n"
+"	};\n"
+"};\n"
+"inline void b3TransformAabb2(b3Float4ConstArg localAabbMin,b3Float4ConstArg localAabbMax, float margin,\n"
+"						b3Float4ConstArg pos,\n"
+"						b3QuatConstArg orn,\n"
+"						b3Float4* aabbMinOut,b3Float4* aabbMaxOut)\n"
+"{\n"
+"		b3Float4 localHalfExtents = 0.5f*(localAabbMax-localAabbMin);\n"
+"		localHalfExtents+=b3MakeFloat4(margin,margin,margin,0.f);\n"
+"		b3Float4 localCenter = 0.5f*(localAabbMax+localAabbMin);\n"
+"		b3Mat3x3 m;\n"
+"		m = b3QuatGetRotationMatrix(orn);\n"
+"		b3Mat3x3 abs_b = b3AbsoluteMat3x3(m);\n"
+"		b3Float4 center = b3TransformPoint(localCenter,pos,orn);\n"
+"		\n"
+"		b3Float4 extent = b3MakeFloat4(b3Dot3F4(localHalfExtents,b3GetRow(abs_b,0)),\n"
+"										 b3Dot3F4(localHalfExtents,b3GetRow(abs_b,1)),\n"
+"										 b3Dot3F4(localHalfExtents,b3GetRow(abs_b,2)),\n"
+"										 0.f);\n"
+"		*aabbMinOut = center-extent;\n"
+"		*aabbMaxOut = center+extent;\n"
+"}\n"
+"/// conservative test for overlap between two aabbs\n"
+"inline bool b3TestAabbAgainstAabb(b3Float4ConstArg aabbMin1,b3Float4ConstArg aabbMax1,\n"
+"								b3Float4ConstArg aabbMin2, b3Float4ConstArg aabbMax2)\n"
+"{\n"
+"	bool overlap = true;\n"
+"	overlap = (aabbMin1.x > aabbMax2.x || aabbMax1.x < aabbMin2.x) ? false : overlap;\n"
+"	overlap = (aabbMin1.z > aabbMax2.z || aabbMax1.z < aabbMin2.z) ? false : overlap;\n"
+"	overlap = (aabbMin1.y > aabbMax2.y || aabbMax1.y < aabbMin2.y) ? false : overlap;\n"
+"	return overlap;\n"
+"}\n"
+"#endif //B3_AABB_H\n"
+"/*\n"
+"Bullet Continuous Collision Detection and Physics Library\n"
+"Copyright (c) 2003-2013 Erwin Coumans  http://bulletphysics.org\n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose,\n"
+"including commercial applications, and to alter it and redistribute it freely,\n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"#ifndef B3_INT2_H\n"
+"#define B3_INT2_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#define b3UnsignedInt2 uint2\n"
+"#define b3Int2 int2\n"
+"#define b3MakeInt2 (int2)\n"
+"#endif //__cplusplus\n"
+"#endif\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_plane;\n"
+"	int m_indexOffset;\n"
+"	int m_numIndices;\n"
+"} btGpuFace;\n"
+"#define make_float4 (float4)\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"	\n"
+"//	float4 a1 = make_float4(a.xyz,0.f);\n"
+"//	float4 b1 = make_float4(b.xyz,0.f);\n"
+"//	return cross(a1,b1);\n"
+"//float4 c = make_float4(a.y*b.z - a.z*b.y,a.z*b.x - a.x*b.z,a.x*b.y - a.y*b.x,0.f);\n"
+"	\n"
+"	//	float4 c = make_float4(a.y*b.z - a.z*b.y,1.f,a.x*b.y - a.y*b.x,0.f);\n"
+"	\n"
+"	//return c;\n"
+"}\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = make_float4(a.xyz,0.f);\n"
+"	float4 b1 = make_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	v = make_float4(v.xyz,0.f);\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"typedef float4 Quaternion;\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b);\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in);\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec);\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q);\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in)\n"
+"{\n"
+"	return fastNormalize4(in);\n"
+"//	in /= length( in );\n"
+"//	return in;\n"
+"}\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(qtMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"__inline\n"
+"float4 qtInvRotate(const Quaternion q, float4 vec)\n"
+"{\n"
+"	return qtRotate( qtInvert( q ), vec );\n"
+"}\n"
+"__inline\n"
+"float4 transform(const float4* p, const float4* translation, const Quaternion* orientation)\n"
+"{\n"
+"	return qtRotate( *orientation, *p ) + (*translation);\n"
+"}\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = make_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"}\n"
+"inline void projectLocal(const ConvexPolyhedronCL* hull,  const float4 pos, const float4 orn, \n"
+"const float4* dir, const float4* vertices, float* min, float* max)\n"
+"{\n"
+"	min[0] = FLT_MAX;\n"
+"	max[0] = -FLT_MAX;\n"
+"	int numVerts = hull->m_numVertices;\n"
+"	const float4 localDir = qtInvRotate(orn,*dir);\n"
+"	float offset = dot(pos,*dir);\n"
+"	for(int i=0;i<numVerts;i++)\n"
+"	{\n"
+"		float dp = dot(vertices[hull->m_vertexOffset+i],localDir);\n"
+"		if(dp < min[0])	\n"
+"			min[0] = dp;\n"
+"		if(dp > max[0])	\n"
+"			max[0] = dp;\n"
+"	}\n"
+"	if(min[0]>max[0])\n"
+"	{\n"
+"		float tmp = min[0];\n"
+"		min[0] = max[0];\n"
+"		max[0] = tmp;\n"
+"	}\n"
+"	min[0] += offset;\n"
+"	max[0] += offset;\n"
+"}\n"
+"inline void project(__global const ConvexPolyhedronCL* hull,  const float4 pos, const float4 orn, \n"
+"const float4* dir, __global const float4* vertices, float* min, float* max)\n"
+"{\n"
+"	min[0] = FLT_MAX;\n"
+"	max[0] = -FLT_MAX;\n"
+"	int numVerts = hull->m_numVertices;\n"
+"	const float4 localDir = qtInvRotate(orn,*dir);\n"
+"	float offset = dot(pos,*dir);\n"
+"	for(int i=0;i<numVerts;i++)\n"
+"	{\n"
+"		float dp = dot(vertices[hull->m_vertexOffset+i],localDir);\n"
+"		if(dp < min[0])	\n"
+"			min[0] = dp;\n"
+"		if(dp > max[0])	\n"
+"			max[0] = dp;\n"
+"	}\n"
+"	if(min[0]>max[0])\n"
+"	{\n"
+"		float tmp = min[0];\n"
+"		min[0] = max[0];\n"
+"		max[0] = tmp;\n"
+"	}\n"
+"	min[0] += offset;\n"
+"	max[0] += offset;\n"
+"}\n"
+"inline bool TestSepAxisLocalA(const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, \n"
+"	const float4 posA,const float4 ornA,\n"
+"	const float4 posB,const float4 ornB,\n"
+"	float4* sep_axis, const float4* verticesA, __global const float4* verticesB,float* depth)\n"
+"{\n"
+"	float Min0,Max0;\n"
+"	float Min1,Max1;\n"
+"	projectLocal(hullA,posA,ornA,sep_axis,verticesA, &Min0, &Max0);\n"
+"	project(hullB,posB,ornB, sep_axis,verticesB, &Min1, &Max1);\n"
+"	if(Max0<Min1 || Max1<Min0)\n"
+"		return false;\n"
+"	float d0 = Max0 - Min1;\n"
+"	float d1 = Max1 - Min0;\n"
+"	*depth = d0<d1 ? d0:d1;\n"
+"	return true;\n"
+"}\n"
+"inline bool IsAlmostZero(const float4 v)\n"
+"{\n"
+"	if(fabs(v.x)>1e-6f || fabs(v.y)>1e-6f || fabs(v.z)>1e-6f)\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"bool findSeparatingAxisLocalA(	const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, \n"
+"	const float4 posA1,\n"
+"	const float4 ornA,\n"
+"	const float4 posB1,\n"
+"	const float4 ornB,\n"
+"	const float4 DeltaC2,\n"
+"	\n"
+"	const float4* verticesA, \n"
+"	const float4* uniqueEdgesA, \n"
+"	const btGpuFace* facesA,\n"
+"	const int*  indicesA,\n"
+"	__global const float4* verticesB, \n"
+"	__global const float4* uniqueEdgesB, \n"
+"	__global const btGpuFace* facesB,\n"
+"	__global const int*  indicesB,\n"
+"	float4* sep,\n"
+"	float* dmin)\n"
+"{\n"
+"	\n"
+"	float4 posA = posA1;\n"
+"	posA.w = 0.f;\n"
+"	float4 posB = posB1;\n"
+"	posB.w = 0.f;\n"
+"	int curPlaneTests=0;\n"
+"	{\n"
+"		int numFacesA = hullA->m_numFaces;\n"
+"		// Test normals from hullA\n"
+"		for(int i=0;i<numFacesA;i++)\n"
+"		{\n"
+"			const float4 normal = facesA[hullA->m_faceOffset+i].m_plane;\n"
+"			float4 faceANormalWS = qtRotate(ornA,normal);\n"
+"			if (dot3F4(DeltaC2,faceANormalWS)<0)\n"
+"				faceANormalWS*=-1.f;\n"
+"			curPlaneTests++;\n"
+"			float d;\n"
+"			if(!TestSepAxisLocalA( hullA, hullB, posA,ornA,posB,ornB,&faceANormalWS, verticesA, verticesB,&d))\n"
+"				return false;\n"
+"			if(d<*dmin)\n"
+"			{\n"
+"				*dmin = d;\n"
+"				*sep = faceANormalWS;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	if((dot3F4(-DeltaC2,*sep))>0.0f)\n"
+"	{\n"
+"		*sep = -(*sep);\n"
+"	}\n"
+"	return true;\n"
+"}\n"
+"bool findSeparatingAxisLocalB(	__global const ConvexPolyhedronCL* hullA,  const ConvexPolyhedronCL* hullB, \n"
+"	const float4 posA1,\n"
+"	const float4 ornA,\n"
+"	const float4 posB1,\n"
+"	const float4 ornB,\n"
+"	const float4 DeltaC2,\n"
+"	__global const float4* verticesA, \n"
+"	__global const float4* uniqueEdgesA, \n"
+"	__global const btGpuFace* facesA,\n"
+"	__global const int*  indicesA,\n"
+"	const float4* verticesB,\n"
+"	const float4* uniqueEdgesB, \n"
+"	const btGpuFace* facesB,\n"
+"	const int*  indicesB,\n"
+"	float4* sep,\n"
+"	float* dmin)\n"
+"{\n"
+"	float4 posA = posA1;\n"
+"	posA.w = 0.f;\n"
+"	float4 posB = posB1;\n"
+"	posB.w = 0.f;\n"
+"	int curPlaneTests=0;\n"
+"	{\n"
+"		int numFacesA = hullA->m_numFaces;\n"
+"		// Test normals from hullA\n"
+"		for(int i=0;i<numFacesA;i++)\n"
+"		{\n"
+"			const float4 normal = facesA[hullA->m_faceOffset+i].m_plane;\n"
+"			float4 faceANormalWS = qtRotate(ornA,normal);\n"
+"			if (dot3F4(DeltaC2,faceANormalWS)<0)\n"
+"				faceANormalWS *= -1.f;\n"
+"			curPlaneTests++;\n"
+"			float d;\n"
+"			if(!TestSepAxisLocalA( hullB, hullA, posB,ornB,posA,ornA, &faceANormalWS, verticesB,verticesA, &d))\n"
+"				return false;\n"
+"			if(d<*dmin)\n"
+"			{\n"
+"				*dmin = d;\n"
+"				*sep = faceANormalWS;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	if((dot3F4(-DeltaC2,*sep))>0.0f)\n"
+"	{\n"
+"		*sep = -(*sep);\n"
+"	}\n"
+"	return true;\n"
+"}\n"
+"bool findSeparatingAxisEdgeEdgeLocalA(	const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, \n"
+"	const float4 posA1,\n"
+"	const float4 ornA,\n"
+"	const float4 posB1,\n"
+"	const float4 ornB,\n"
+"	const float4 DeltaC2,\n"
+"	const float4* verticesA, \n"
+"	const float4* uniqueEdgesA, \n"
+"	const btGpuFace* facesA,\n"
+"	const int*  indicesA,\n"
+"	__global const float4* verticesB, \n"
+"	__global const float4* uniqueEdgesB, \n"
+"	__global const btGpuFace* facesB,\n"
+"	__global const int*  indicesB,\n"
+"		float4* sep,\n"
+"	float* dmin)\n"
+"{\n"
+"	float4 posA = posA1;\n"
+"	posA.w = 0.f;\n"
+"	float4 posB = posB1;\n"
+"	posB.w = 0.f;\n"
+"	int curPlaneTests=0;\n"
+"	int curEdgeEdge = 0;\n"
+"	// Test edges\n"
+"	for(int e0=0;e0<hullA->m_numUniqueEdges;e0++)\n"
+"	{\n"
+"		const float4 edge0 = uniqueEdgesA[hullA->m_uniqueEdgesOffset+e0];\n"
+"		float4 edge0World = qtRotate(ornA,edge0);\n"
+"		for(int e1=0;e1<hullB->m_numUniqueEdges;e1++)\n"
+"		{\n"
+"			const float4 edge1 = uniqueEdgesB[hullB->m_uniqueEdgesOffset+e1];\n"
+"			float4 edge1World = qtRotate(ornB,edge1);\n"
+"			float4 crossje = cross3(edge0World,edge1World);\n"
+"			curEdgeEdge++;\n"
+"			if(!IsAlmostZero(crossje))\n"
+"			{\n"
+"				crossje = normalize3(crossje);\n"
+"				if (dot3F4(DeltaC2,crossje)<0)\n"
+"					crossje *= -1.f;\n"
+"				float dist;\n"
+"				bool result = true;\n"
+"				{\n"
+"					float Min0,Max0;\n"
+"					float Min1,Max1;\n"
+"					projectLocal(hullA,posA,ornA,&crossje,verticesA, &Min0, &Max0);\n"
+"					project(hullB,posB,ornB,&crossje,verticesB, &Min1, &Max1);\n"
+"				\n"
+"					if(Max0<Min1 || Max1<Min0)\n"
+"						result = false;\n"
+"				\n"
+"					float d0 = Max0 - Min1;\n"
+"					float d1 = Max1 - Min0;\n"
+"					dist = d0<d1 ? d0:d1;\n"
+"					result = true;\n"
+"				}\n"
+"				\n"
+"				if(dist<*dmin)\n"
+"				{\n"
+"					*dmin = dist;\n"
+"					*sep = crossje;\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	\n"
+"	if((dot3F4(-DeltaC2,*sep))>0.0f)\n"
+"	{\n"
+"		*sep = -(*sep);\n"
+"	}\n"
+"	return true;\n"
+"}\n"
+"inline int	findClippingFaces(const float4 separatingNormal,\n"
+"                      const ConvexPolyhedronCL* hullA, \n"
+"					  __global const ConvexPolyhedronCL* hullB,\n"
+"                      const float4 posA, const Quaternion ornA,const float4 posB, const Quaternion ornB,\n"
+"                       __global float4* worldVertsA1,\n"
+"                      __global float4* worldNormalsA1,\n"
+"                      __global float4* worldVertsB1,\n"
+"                      int capacityWorldVerts,\n"
+"                      const float minDist, float maxDist,\n"
+"					  const float4* verticesA,\n"
+"                      const btGpuFace* facesA,\n"
+"                      const int* indicesA,\n"
+"					  __global const float4* verticesB,\n"
+"                      __global const btGpuFace* facesB,\n"
+"                      __global const int* indicesB,\n"
+"                      __global int4* clippingFaces, int pairIndex)\n"
+"{\n"
+"	int numContactsOut = 0;\n"
+"	int numWorldVertsB1= 0;\n"
+"    \n"
+"    \n"
+"	int closestFaceB=0;\n"
+"	float dmax = -FLT_MAX;\n"
+"    \n"
+"	{\n"
+"		for(int face=0;face<hullB->m_numFaces;face++)\n"
+"		{\n"
+"			const float4 Normal = make_float4(facesB[hullB->m_faceOffset+face].m_plane.x,\n"
+"                                              facesB[hullB->m_faceOffset+face].m_plane.y, facesB[hullB->m_faceOffset+face].m_plane.z,0.f);\n"
+"			const float4 WorldNormal = qtRotate(ornB, Normal);\n"
+"			float d = dot3F4(WorldNormal,separatingNormal);\n"
+"			if (d > dmax)\n"
+"			{\n"
+"				dmax = d;\n"
+"				closestFaceB = face;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"    \n"
+"	{\n"
+"		const btGpuFace polyB = facesB[hullB->m_faceOffset+closestFaceB];\n"
+"		int numVertices = polyB.m_numIndices;\n"
+"        if (numVertices>capacityWorldVerts)\n"
+"            numVertices = capacityWorldVerts;\n"
+"        if (numVertices<0)\n"
+"            numVertices = 0;\n"
+"        \n"
+"		for(int e0=0;e0<numVertices;e0++)\n"
+"		{\n"
+"            if (e0<capacityWorldVerts)\n"
+"            {\n"
+"                const float4 b = verticesB[hullB->m_vertexOffset+indicesB[polyB.m_indexOffset+e0]];\n"
+"                worldVertsB1[pairIndex*capacityWorldVerts+numWorldVertsB1++] = transform(&b,&posB,&ornB);\n"
+"            }\n"
+"		}\n"
+"	}\n"
+"    \n"
+"    int closestFaceA=0;\n"
+"	{\n"
+"		float dmin = FLT_MAX;\n"
+"		for(int face=0;face<hullA->m_numFaces;face++)\n"
+"		{\n"
+"			const float4 Normal = make_float4(\n"
+"                                              facesA[hullA->m_faceOffset+face].m_plane.x,\n"
+"                                              facesA[hullA->m_faceOffset+face].m_plane.y,\n"
+"                                              facesA[hullA->m_faceOffset+face].m_plane.z,\n"
+"                                              0.f);\n"
+"			const float4 faceANormalWS = qtRotate(ornA,Normal);\n"
+"            \n"
+"			float d = dot3F4(faceANormalWS,separatingNormal);\n"
+"			if (d < dmin)\n"
+"			{\n"
+"				dmin = d;\n"
+"				closestFaceA = face;\n"
+"                worldNormalsA1[pairIndex] = faceANormalWS;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"    \n"
+"    int numVerticesA = facesA[hullA->m_faceOffset+closestFaceA].m_numIndices;\n"
+"    if (numVerticesA>capacityWorldVerts)\n"
+"       numVerticesA = capacityWorldVerts;\n"
+"    if (numVerticesA<0)\n"
+"        numVerticesA=0;\n"
+"    \n"
+"	for(int e0=0;e0<numVerticesA;e0++)\n"
+"	{\n"
+"        if (e0<capacityWorldVerts)\n"
+"        {\n"
+"            const float4 a = verticesA[hullA->m_vertexOffset+indicesA[facesA[hullA->m_faceOffset+closestFaceA].m_indexOffset+e0]];\n"
+"            worldVertsA1[pairIndex*capacityWorldVerts+e0] = transform(&a, &posA,&ornA);\n"
+"        }\n"
+"    }\n"
+"    \n"
+"    clippingFaces[pairIndex].x = closestFaceA;\n"
+"    clippingFaces[pairIndex].y = closestFaceB;\n"
+"    clippingFaces[pairIndex].z = numVerticesA;\n"
+"    clippingFaces[pairIndex].w = numWorldVertsB1;\n"
+"    \n"
+"    \n"
+"	return numContactsOut;\n"
+"}\n"
+"// work-in-progress\n"
+"__kernel void   findConcaveSeparatingAxisVertexFaceKernel( __global int4* concavePairs,\n"
+"                                                __global const BodyData* rigidBodies,\n"
+"                                                __global const btCollidableGpu* collidables,\n"
+"                                                __global const ConvexPolyhedronCL* convexShapes,\n"
+"                                                __global const float4* vertices,\n"
+"                                                __global const float4* uniqueEdges,\n"
+"                                                __global const btGpuFace* faces,\n"
+"                                                __global const int* indices,\n"
+"                                                __global const btGpuChildShape* gpuChildShapes,\n"
+"                                                __global btAabbCL* aabbs,\n"
+"                                                __global float4* concaveSeparatingNormalsOut,\n"
+"                                                __global int* concaveHasSeparatingNormals,\n"
+"                                                __global int4* clippingFacesOut,\n"
+"                                                __global float4* worldVertsA1GPU,\n"
+"                                                __global float4*  worldNormalsAGPU,\n"
+"                                                __global float4* worldVertsB1GPU,\n"
+"                                                __global float* dmins,\n"
+"                                                int vertexFaceCapacity,\n"
+"                                                int numConcavePairs\n"
+"                                                )\n"
+"{\n"
+"    \n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numConcavePairs)\n"
+"		return;\n"
+"    \n"
+"	concaveHasSeparatingNormals[i] = 0;\n"
+"    \n"
+"	int pairIdx = i;\n"
+"    \n"
+"	int bodyIndexA = concavePairs[i].x;\n"
+"	int bodyIndexB = concavePairs[i].y;\n"
+"    \n"
+"	int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"	int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"    \n"
+"	int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"	int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"    \n"
+"	if (collidables[collidableIndexB].m_shapeType!=SHAPE_CONVEX_HULL&&\n"
+"		collidables[collidableIndexB].m_shapeType!=SHAPE_COMPOUND_OF_CONVEX_HULLS)\n"
+"	{\n"
+"		concavePairs[pairIdx].w = -1;\n"
+"		return;\n"
+"	}\n"
+"    \n"
+"    \n"
+"    \n"
+"	int numFacesA = convexShapes[shapeIndexA].m_numFaces;\n"
+"	int numActualConcaveConvexTests = 0;\n"
+"	\n"
+"	int f = concavePairs[i].z;\n"
+"	\n"
+"	bool overlap = false;\n"
+"	\n"
+"	ConvexPolyhedronCL convexPolyhedronA;\n"
+"    \n"
+"	//add 3 vertices of the triangle\n"
+"	convexPolyhedronA.m_numVertices = 3;\n"
+"	convexPolyhedronA.m_vertexOffset = 0;\n"
+"	float4	localCenter = make_float4(0.f,0.f,0.f,0.f);\n"
+"    \n"
+"	btGpuFace face = faces[convexShapes[shapeIndexA].m_faceOffset+f];\n"
+"	float4 triMinAabb, triMaxAabb;\n"
+"	btAabbCL triAabb;\n"
+"	triAabb.m_min = make_float4(1e30f,1e30f,1e30f,0.f);\n"
+"	triAabb.m_max = make_float4(-1e30f,-1e30f,-1e30f,0.f);\n"
+"	\n"
+"	float4 verticesA[3];\n"
+"	for (int i=0;i<3;i++)\n"
+"	{\n"
+"		int index = indices[face.m_indexOffset+i];\n"
+"		float4 vert = vertices[convexShapes[shapeIndexA].m_vertexOffset+index];\n"
+"		verticesA[i] = vert;\n"
+"		localCenter += vert;\n"
+"        \n"
+"		triAabb.m_min = min(triAabb.m_min,vert);\n"
+"		triAabb.m_max = max(triAabb.m_max,vert);\n"
+"        \n"
+"	}\n"
+"    \n"
+"	overlap = true;\n"
+"	overlap = (triAabb.m_min.x > aabbs[bodyIndexB].m_max.x || triAabb.m_max.x < aabbs[bodyIndexB].m_min.x) ? false : overlap;\n"
+"	overlap = (triAabb.m_min.z > aabbs[bodyIndexB].m_max.z || triAabb.m_max.z < aabbs[bodyIndexB].m_min.z) ? false : overlap;\n"
+"	overlap = (triAabb.m_min.y > aabbs[bodyIndexB].m_max.y || triAabb.m_max.y < aabbs[bodyIndexB].m_min.y) ? false : overlap;\n"
+"    \n"
+"	if (overlap)\n"
+"	{\n"
+"		float dmin = FLT_MAX;\n"
+"		int hasSeparatingAxis=5;\n"
+"		float4 sepAxis=make_float4(1,2,3,4);\n"
+"        \n"
+"		int localCC=0;\n"
+"		numActualConcaveConvexTests++;\n"
+"        \n"
+"		//a triangle has 3 unique edges\n"
+"		convexPolyhedronA.m_numUniqueEdges = 3;\n"
+"		convexPolyhedronA.m_uniqueEdgesOffset = 0;\n"
+"		float4 uniqueEdgesA[3];\n"
+"		\n"
+"		uniqueEdgesA[0] = (verticesA[1]-verticesA[0]);\n"
+"		uniqueEdgesA[1] = (verticesA[2]-verticesA[1]);\n"
+"		uniqueEdgesA[2] = (verticesA[0]-verticesA[2]);\n"
+"        \n"
+"        \n"
+"		convexPolyhedronA.m_faceOffset = 0;\n"
+"        \n"
+"		float4 normal = make_float4(face.m_plane.x,face.m_plane.y,face.m_plane.z,0.f);\n"
+"        \n"
+"		btGpuFace facesA[TRIANGLE_NUM_CONVEX_FACES];\n"
+"		int indicesA[3+3+2+2+2];\n"
+"		int curUsedIndices=0;\n"
+"		int fidx=0;\n"
+"        \n"
+"		//front size of triangle\n"
+"		{\n"
+"			facesA[fidx].m_indexOffset=curUsedIndices;\n"
+"			indicesA[0] = 0;\n"
+"			indicesA[1] = 1;\n"
+"			indicesA[2] = 2;\n"
+"			curUsedIndices+=3;\n"
+"			float c = face.m_plane.w;\n"
+"			facesA[fidx].m_plane.x = normal.x;\n"
+"			facesA[fidx].m_plane.y = normal.y;\n"
+"			facesA[fidx].m_plane.z = normal.z;\n"
+"			facesA[fidx].m_plane.w = c;\n"
+"			facesA[fidx].m_numIndices=3;\n"
+"		}\n"
+"		fidx++;\n"
+"		//back size of triangle\n"
+"		{\n"
+"			facesA[fidx].m_indexOffset=curUsedIndices;\n"
+"			indicesA[3]=2;\n"
+"			indicesA[4]=1;\n"
+"			indicesA[5]=0;\n"
+"			curUsedIndices+=3;\n"
+"			float c = dot(normal,verticesA[0]);\n"
+"			float c1 = -face.m_plane.w;\n"
+"			facesA[fidx].m_plane.x = -normal.x;\n"
+"			facesA[fidx].m_plane.y = -normal.y;\n"
+"			facesA[fidx].m_plane.z = -normal.z;\n"
+"			facesA[fidx].m_plane.w = c;\n"
+"			facesA[fidx].m_numIndices=3;\n"
+"		}\n"
+"		fidx++;\n"
+"        \n"
+"		bool addEdgePlanes = true;\n"
+"		if (addEdgePlanes)\n"
+"		{\n"
+"			int numVertices=3;\n"
+"			int prevVertex = numVertices-1;\n"
+"			for (int i=0;i<numVertices;i++)\n"
+"			{\n"
+"				float4 v0 = verticesA[i];\n"
+"				float4 v1 = verticesA[prevVertex];\n"
+"                \n"
+"				float4 edgeNormal = normalize(cross(normal,v1-v0));\n"
+"				float c = -dot(edgeNormal,v0);\n"
+"                \n"
+"				facesA[fidx].m_numIndices = 2;\n"
+"				facesA[fidx].m_indexOffset=curUsedIndices;\n"
+"				indicesA[curUsedIndices++]=i;\n"
+"				indicesA[curUsedIndices++]=prevVertex;\n"
+"                \n"
+"				facesA[fidx].m_plane.x = edgeNormal.x;\n"
+"				facesA[fidx].m_plane.y = edgeNormal.y;\n"
+"				facesA[fidx].m_plane.z = edgeNormal.z;\n"
+"				facesA[fidx].m_plane.w = c;\n"
+"				fidx++;\n"
+"				prevVertex = i;\n"
+"			}\n"
+"		}\n"
+"		convexPolyhedronA.m_numFaces = TRIANGLE_NUM_CONVEX_FACES;\n"
+"		convexPolyhedronA.m_localCenter = localCenter*(1.f/3.f);\n"
+"        \n"
+"        \n"
+"		float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"		posA.w = 0.f;\n"
+"		float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"		posB.w = 0.f;\n"
+"        \n"
+"		float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"		float4 ornB =rigidBodies[bodyIndexB].m_quat;\n"
+"        \n"
+"		\n"
+"        \n"
+"        \n"
+"		///////////////////\n"
+"		///compound shape support\n"
+"        \n"
+"		if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)\n"
+"		{\n"
+"			int compoundChild = concavePairs[pairIdx].w;\n"
+"			int childShapeIndexB = compoundChild;//collidables[collidableIndexB].m_shapeIndex+compoundChild;\n"
+"			int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;\n"
+"			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;\n"
+"			float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;\n"
+"			float4 newPosB = transform(&childPosB,&posB,&ornB);\n"
+"			float4 newOrnB = qtMul(ornB,childOrnB);\n"
+"			posB = newPosB;\n"
+"			ornB = newOrnB;\n"
+"			shapeIndexB = collidables[childColIndexB].m_shapeIndex;\n"
+"		}\n"
+"		//////////////////\n"
+"        \n"
+"		float4 c0local = convexPolyhedronA.m_localCenter;\n"
+"		float4 c0 = transform(&c0local, &posA, &ornA);\n"
+"		float4 c1local = convexShapes[shapeIndexB].m_localCenter;\n"
+"		float4 c1 = transform(&c1local,&posB,&ornB);\n"
+"		const float4 DeltaC2 = c0 - c1;\n"
+"        \n"
+"        \n"
+"		bool sepA = findSeparatingAxisLocalA(	&convexPolyhedronA, &convexShapes[shapeIndexB],\n"
+"                                             posA,ornA,\n"
+"                                             posB,ornB,\n"
+"                                             DeltaC2,\n"
+"                                             verticesA,uniqueEdgesA,facesA,indicesA,\n"
+"                                             vertices,uniqueEdges,faces,indices,\n"
+"                                             &sepAxis,&dmin);\n"
+"		hasSeparatingAxis = 4;\n"
+"		if (!sepA)\n"
+"		{\n"
+"			hasSeparatingAxis = 0;\n"
+"		} else\n"
+"		{\n"
+"			bool sepB = findSeparatingAxisLocalB(	&convexShapes[shapeIndexB],&convexPolyhedronA,\n"
+"                                                 posB,ornB,\n"
+"                                                 posA,ornA,\n"
+"                                                 DeltaC2,\n"
+"                                                 vertices,uniqueEdges,faces,indices,\n"
+"                                                 verticesA,uniqueEdgesA,facesA,indicesA,\n"
+"                                                 &sepAxis,&dmin);\n"
+"            \n"
+"			if (!sepB)\n"
+"			{\n"
+"				hasSeparatingAxis = 0;\n"
+"			} else\n"
+"			{\n"
+"				hasSeparatingAxis = 1;\n"
+"			}\n"
+"		}	\n"
+"		\n"
+"		if (hasSeparatingAxis)\n"
+"		{\n"
+"            dmins[i] = dmin;\n"
+"			concaveSeparatingNormalsOut[pairIdx]=sepAxis;\n"
+"			concaveHasSeparatingNormals[i]=1;\n"
+"            \n"
+"		} else\n"
+"		{	\n"
+"			//mark this pair as in-active\n"
+"			concavePairs[pairIdx].w = -1;\n"
+"		}\n"
+"	}\n"
+"	else\n"
+"	{	\n"
+"		//mark this pair as in-active\n"
+"		concavePairs[pairIdx].w = -1;\n"
+"	}\n"
+"}\n"
+"// work-in-progress\n"
+"__kernel void   findConcaveSeparatingAxisEdgeEdgeKernel( __global int4* concavePairs,\n"
+"                                                          __global const BodyData* rigidBodies,\n"
+"                                                          __global const btCollidableGpu* collidables,\n"
+"                                                          __global const ConvexPolyhedronCL* convexShapes,\n"
+"                                                          __global const float4* vertices,\n"
+"                                                          __global const float4* uniqueEdges,\n"
+"                                                          __global const btGpuFace* faces,\n"
+"                                                          __global const int* indices,\n"
+"                                                          __global const btGpuChildShape* gpuChildShapes,\n"
+"                                                          __global btAabbCL* aabbs,\n"
+"                                                          __global float4* concaveSeparatingNormalsOut,\n"
+"                                                          __global int* concaveHasSeparatingNormals,\n"
+"                                                          __global int4* clippingFacesOut,\n"
+"                                                          __global float4* worldVertsA1GPU,\n"
+"                                                          __global float4*  worldNormalsAGPU,\n"
+"                                                          __global float4* worldVertsB1GPU,\n"
+"                                                          __global float* dmins,\n"
+"                                                          int vertexFaceCapacity,\n"
+"                                                          int numConcavePairs\n"
+"                                                          )\n"
+"{\n"
+"    \n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numConcavePairs)\n"
+"		return;\n"
+"    \n"
+"	if (!concaveHasSeparatingNormals[i])\n"
+"        return;\n"
+"    \n"
+"	int pairIdx = i;\n"
+"    \n"
+"	int bodyIndexA = concavePairs[i].x;\n"
+"	int bodyIndexB = concavePairs[i].y;\n"
+"    \n"
+"	int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"	int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"    \n"
+"	int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"	int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"    \n"
+"    \n"
+"	int numFacesA = convexShapes[shapeIndexA].m_numFaces;\n"
+"	int numActualConcaveConvexTests = 0;\n"
+"	\n"
+"	int f = concavePairs[i].z;\n"
+"	\n"
+"	bool overlap = false;\n"
+"	\n"
+"	ConvexPolyhedronCL convexPolyhedronA;\n"
+"    \n"
+"	//add 3 vertices of the triangle\n"
+"	convexPolyhedronA.m_numVertices = 3;\n"
+"	convexPolyhedronA.m_vertexOffset = 0;\n"
+"	float4	localCenter = make_float4(0.f,0.f,0.f,0.f);\n"
+"    \n"
+"	btGpuFace face = faces[convexShapes[shapeIndexA].m_faceOffset+f];\n"
+"	float4 triMinAabb, triMaxAabb;\n"
+"	btAabbCL triAabb;\n"
+"	triAabb.m_min = make_float4(1e30f,1e30f,1e30f,0.f);\n"
+"	triAabb.m_max = make_float4(-1e30f,-1e30f,-1e30f,0.f);\n"
+"	\n"
+"	float4 verticesA[3];\n"
+"	for (int i=0;i<3;i++)\n"
+"	{\n"
+"		int index = indices[face.m_indexOffset+i];\n"
+"		float4 vert = vertices[convexShapes[shapeIndexA].m_vertexOffset+index];\n"
+"		verticesA[i] = vert;\n"
+"		localCenter += vert;\n"
+"        \n"
+"		triAabb.m_min = min(triAabb.m_min,vert);\n"
+"		triAabb.m_max = max(triAabb.m_max,vert);\n"
+"        \n"
+"	}\n"
+"    \n"
+"	overlap = true;\n"
+"	overlap = (triAabb.m_min.x > aabbs[bodyIndexB].m_max.x || triAabb.m_max.x < aabbs[bodyIndexB].m_min.x) ? false : overlap;\n"
+"	overlap = (triAabb.m_min.z > aabbs[bodyIndexB].m_max.z || triAabb.m_max.z < aabbs[bodyIndexB].m_min.z) ? false : overlap;\n"
+"	overlap = (triAabb.m_min.y > aabbs[bodyIndexB].m_max.y || triAabb.m_max.y < aabbs[bodyIndexB].m_min.y) ? false : overlap;\n"
+"    \n"
+"	if (overlap)\n"
+"	{\n"
+"		float dmin = dmins[i];\n"
+"		int hasSeparatingAxis=5;\n"
+"		float4 sepAxis=make_float4(1,2,3,4);\n"
+"        sepAxis = concaveSeparatingNormalsOut[pairIdx];\n"
+"        \n"
+"		int localCC=0;\n"
+"		numActualConcaveConvexTests++;\n"
+"        \n"
+"		//a triangle has 3 unique edges\n"
+"		convexPolyhedronA.m_numUniqueEdges = 3;\n"
+"		convexPolyhedronA.m_uniqueEdgesOffset = 0;\n"
+"		float4 uniqueEdgesA[3];\n"
+"		\n"
+"		uniqueEdgesA[0] = (verticesA[1]-verticesA[0]);\n"
+"		uniqueEdgesA[1] = (verticesA[2]-verticesA[1]);\n"
+"		uniqueEdgesA[2] = (verticesA[0]-verticesA[2]);\n"
+"        \n"
+"        \n"
+"		convexPolyhedronA.m_faceOffset = 0;\n"
+"        \n"
+"		float4 normal = make_float4(face.m_plane.x,face.m_plane.y,face.m_plane.z,0.f);\n"
+"        \n"
+"		btGpuFace facesA[TRIANGLE_NUM_CONVEX_FACES];\n"
+"		int indicesA[3+3+2+2+2];\n"
+"		int curUsedIndices=0;\n"
+"		int fidx=0;\n"
+"        \n"
+"		//front size of triangle\n"
+"		{\n"
+"			facesA[fidx].m_indexOffset=curUsedIndices;\n"
+"			indicesA[0] = 0;\n"
+"			indicesA[1] = 1;\n"
+"			indicesA[2] = 2;\n"
+"			curUsedIndices+=3;\n"
+"			float c = face.m_plane.w;\n"
+"			facesA[fidx].m_plane.x = normal.x;\n"
+"			facesA[fidx].m_plane.y = normal.y;\n"
+"			facesA[fidx].m_plane.z = normal.z;\n"
+"			facesA[fidx].m_plane.w = c;\n"
+"			facesA[fidx].m_numIndices=3;\n"
+"		}\n"
+"		fidx++;\n"
+"		//back size of triangle\n"
+"		{\n"
+"			facesA[fidx].m_indexOffset=curUsedIndices;\n"
+"			indicesA[3]=2;\n"
+"			indicesA[4]=1;\n"
+"			indicesA[5]=0;\n"
+"			curUsedIndices+=3;\n"
+"			float c = dot(normal,verticesA[0]);\n"
+"			float c1 = -face.m_plane.w;\n"
+"			facesA[fidx].m_plane.x = -normal.x;\n"
+"			facesA[fidx].m_plane.y = -normal.y;\n"
+"			facesA[fidx].m_plane.z = -normal.z;\n"
+"			facesA[fidx].m_plane.w = c;\n"
+"			facesA[fidx].m_numIndices=3;\n"
+"		}\n"
+"		fidx++;\n"
+"        \n"
+"		bool addEdgePlanes = true;\n"
+"		if (addEdgePlanes)\n"
+"		{\n"
+"			int numVertices=3;\n"
+"			int prevVertex = numVertices-1;\n"
+"			for (int i=0;i<numVertices;i++)\n"
+"			{\n"
+"				float4 v0 = verticesA[i];\n"
+"				float4 v1 = verticesA[prevVertex];\n"
+"                \n"
+"				float4 edgeNormal = normalize(cross(normal,v1-v0));\n"
+"				float c = -dot(edgeNormal,v0);\n"
+"                \n"
+"				facesA[fidx].m_numIndices = 2;\n"
+"				facesA[fidx].m_indexOffset=curUsedIndices;\n"
+"				indicesA[curUsedIndices++]=i;\n"
+"				indicesA[curUsedIndices++]=prevVertex;\n"
+"                \n"
+"				facesA[fidx].m_plane.x = edgeNormal.x;\n"
+"				facesA[fidx].m_plane.y = edgeNormal.y;\n"
+"				facesA[fidx].m_plane.z = edgeNormal.z;\n"
+"				facesA[fidx].m_plane.w = c;\n"
+"				fidx++;\n"
+"				prevVertex = i;\n"
+"			}\n"
+"		}\n"
+"		convexPolyhedronA.m_numFaces = TRIANGLE_NUM_CONVEX_FACES;\n"
+"		convexPolyhedronA.m_localCenter = localCenter*(1.f/3.f);\n"
+"        \n"
+"        \n"
+"		float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"		posA.w = 0.f;\n"
+"		float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"		posB.w = 0.f;\n"
+"        \n"
+"		float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"		float4 ornB =rigidBodies[bodyIndexB].m_quat;\n"
+"        \n"
+"		\n"
+"        \n"
+"        \n"
+"		///////////////////\n"
+"		///compound shape support\n"
+"        \n"
+"		if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)\n"
+"		{\n"
+"			int compoundChild = concavePairs[pairIdx].w;\n"
+"			int childShapeIndexB = compoundChild;//collidables[collidableIndexB].m_shapeIndex+compoundChild;\n"
+"			int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;\n"
+"			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;\n"
+"			float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;\n"
+"			float4 newPosB = transform(&childPosB,&posB,&ornB);\n"
+"			float4 newOrnB = qtMul(ornB,childOrnB);\n"
+"			posB = newPosB;\n"
+"			ornB = newOrnB;\n"
+"			shapeIndexB = collidables[childColIndexB].m_shapeIndex;\n"
+"		}\n"
+"		//////////////////\n"
+"        \n"
+"		float4 c0local = convexPolyhedronA.m_localCenter;\n"
+"		float4 c0 = transform(&c0local, &posA, &ornA);\n"
+"		float4 c1local = convexShapes[shapeIndexB].m_localCenter;\n"
+"		float4 c1 = transform(&c1local,&posB,&ornB);\n"
+"		const float4 DeltaC2 = c0 - c1;\n"
+"        \n"
+"        \n"
+"		{\n"
+"			bool sepEE = findSeparatingAxisEdgeEdgeLocalA(	&convexPolyhedronA, &convexShapes[shapeIndexB],\n"
+"                                                              posA,ornA,\n"
+"                                                              posB,ornB,\n"
+"                                                              DeltaC2,\n"
+"                                                              verticesA,uniqueEdgesA,facesA,indicesA,\n"
+"                                                              vertices,uniqueEdges,faces,indices,\n"
+"                                                              &sepAxis,&dmin);\n"
+"                \n"
+"			if (!sepEE)\n"
+"			{\n"
+"				hasSeparatingAxis = 0;\n"
+"			} else\n"
+"			{\n"
+"				hasSeparatingAxis = 1;\n"
+"			}\n"
+"		}\n"
+"		\n"
+"		\n"
+"		if (hasSeparatingAxis)\n"
+"		{\n"
+"			sepAxis.w = dmin;\n"
+"            dmins[i] = dmin;\n"
+"			concaveSeparatingNormalsOut[pairIdx]=sepAxis;\n"
+"			concaveHasSeparatingNormals[i]=1;\n"
+"           \n"
+" 	float minDist = -1e30f;\n"
+"			float maxDist = 0.02f;\n"
+"            \n"
+"            findClippingFaces(sepAxis,\n"
+"                              &convexPolyhedronA,\n"
+"                              &convexShapes[shapeIndexB],\n"
+"                              posA,ornA,\n"
+"                              posB,ornB,\n"
+"                              worldVertsA1GPU,\n"
+"                              worldNormalsAGPU,\n"
+"                              worldVertsB1GPU,\n"
+"                              vertexFaceCapacity,\n"
+"                              minDist, maxDist,\n"
+"                              verticesA,\n"
+"                              facesA,\n"
+"                              indicesA,\n"
+"                              vertices,\n"
+"                              faces,\n"
+"                              indices,\n"
+"                              clippingFacesOut, pairIdx);\n"
+"	           \n"
+"            \n"
+"		} else\n"
+"		{	\n"
+"			//mark this pair as in-active\n"
+"			concavePairs[pairIdx].w = -1;\n"
+"		}\n"
+"	}\n"
+"	else\n"
+"	{	\n"
+"		//mark this pair as in-active\n"
+"		concavePairs[pairIdx].w = -1;\n"
+"	}\n"
+"	\n"
+"	concavePairs[i].z = -1;//for the next stage, z is used to determine existing contact points\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satKernels.h b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satKernels.h
new file mode 100644
index 0000000000..6f8b0a90db
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/NarrowphaseCollision/kernels/satKernels.h
@@ -0,0 +1,2104 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* satKernelsCL= \
+"//keep this enum in sync with the CPU version (in btCollidable.h)\n"
+"//written by Erwin Coumans\n"
+"#define SHAPE_CONVEX_HULL 3\n"
+"#define SHAPE_CONCAVE_TRIMESH 5\n"
+"#define TRIANGLE_NUM_CONVEX_FACES 5\n"
+"#define SHAPE_COMPOUND_OF_CONVEX_HULLS 6\n"
+"#define B3_MAX_STACK_DEPTH 256\n"
+"typedef unsigned int u32;\n"
+"///keep this in sync with btCollidable.h\n"
+"typedef struct\n"
+"{\n"
+"	union {\n"
+"		int m_numChildShapes;\n"
+"		int m_bvhIndex;\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float m_radius;\n"
+"		int	m_compoundBvhIndex;\n"
+"	};\n"
+"	\n"
+"	int m_shapeType;\n"
+"	int m_shapeIndex;\n"
+"	\n"
+"} btCollidableGpu;\n"
+"#define MAX_NUM_PARTS_IN_BITS 10\n"
+"///b3QuantizedBvhNode is a compressed aabb node, 16 bytes.\n"
+"///Node can be used for leafnode or internal node. Leafnodes can point to 32-bit triangle index (non-negative range).\n"
+"typedef struct\n"
+"{\n"
+"	//12 bytes\n"
+"	unsigned short int	m_quantizedAabbMin[3];\n"
+"	unsigned short int	m_quantizedAabbMax[3];\n"
+"	//4 bytes\n"
+"	int	m_escapeIndexOrTriangleIndex;\n"
+"} b3QuantizedBvhNode;\n"
+"typedef struct\n"
+"{\n"
+"	float4		m_aabbMin;\n"
+"	float4		m_aabbMax;\n"
+"	float4		m_quantization;\n"
+"	int			m_numNodes;\n"
+"	int			m_numSubTrees;\n"
+"	int			m_nodeOffset;\n"
+"	int			m_subTreeOffset;\n"
+"} b3BvhInfo;\n"
+"int	getTriangleIndex(const b3QuantizedBvhNode* rootNode)\n"
+"{\n"
+"	unsigned int x=0;\n"
+"	unsigned int y = (~(x&0))<<(31-MAX_NUM_PARTS_IN_BITS);\n"
+"	// Get only the lower bits where the triangle index is stored\n"
+"	return (rootNode->m_escapeIndexOrTriangleIndex&~(y));\n"
+"}\n"
+"int	getTriangleIndexGlobal(__global const b3QuantizedBvhNode* rootNode)\n"
+"{\n"
+"	unsigned int x=0;\n"
+"	unsigned int y = (~(x&0))<<(31-MAX_NUM_PARTS_IN_BITS);\n"
+"	// Get only the lower bits where the triangle index is stored\n"
+"	return (rootNode->m_escapeIndexOrTriangleIndex&~(y));\n"
+"}\n"
+"int isLeafNode(const b3QuantizedBvhNode* rootNode)\n"
+"{\n"
+"	//skipindex is negative (internal node), triangleindex >=0 (leafnode)\n"
+"	return (rootNode->m_escapeIndexOrTriangleIndex >= 0)? 1 : 0;\n"
+"}\n"
+"int isLeafNodeGlobal(__global const b3QuantizedBvhNode* rootNode)\n"
+"{\n"
+"	//skipindex is negative (internal node), triangleindex >=0 (leafnode)\n"
+"	return (rootNode->m_escapeIndexOrTriangleIndex >= 0)? 1 : 0;\n"
+"}\n"
+"	\n"
+"int getEscapeIndex(const b3QuantizedBvhNode* rootNode)\n"
+"{\n"
+"	return -rootNode->m_escapeIndexOrTriangleIndex;\n"
+"}\n"
+"int getEscapeIndexGlobal(__global const b3QuantizedBvhNode* rootNode)\n"
+"{\n"
+"	return -rootNode->m_escapeIndexOrTriangleIndex;\n"
+"}\n"
+"typedef struct\n"
+"{\n"
+"	//12 bytes\n"
+"	unsigned short int	m_quantizedAabbMin[3];\n"
+"	unsigned short int	m_quantizedAabbMax[3];\n"
+"	//4 bytes, points to the root of the subtree\n"
+"	int			m_rootNodeIndex;\n"
+"	//4 bytes\n"
+"	int			m_subtreeSize;\n"
+"	int			m_padding[3];\n"
+"} b3BvhSubtreeInfo;\n"
+"typedef struct\n"
+"{\n"
+"	float4	m_childPosition;\n"
+"	float4	m_childOrientation;\n"
+"	int m_shapeIndex;\n"
+"	int m_unused0;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"} btGpuChildShape;\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	float4 m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"	u32 m_collidableIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} BodyData;\n"
+"typedef struct  \n"
+"{\n"
+"	float4		m_localCenter;\n"
+"	float4		m_extents;\n"
+"	float4		mC;\n"
+"	float4		mE;\n"
+"	\n"
+"	float			m_radius;\n"
+"	int	m_faceOffset;\n"
+"	int m_numFaces;\n"
+"	int	m_numVertices;\n"
+"	int m_vertexOffset;\n"
+"	int	m_uniqueEdgesOffset;\n"
+"	int	m_numUniqueEdges;\n"
+"	int m_unused;\n"
+"} ConvexPolyhedronCL;\n"
+"typedef struct \n"
+"{\n"
+"	union\n"
+"	{\n"
+"		float4	m_min;\n"
+"		float   m_minElems[4];\n"
+"		int			m_minIndices[4];\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float4	m_max;\n"
+"		float   m_maxElems[4];\n"
+"		int			m_maxIndices[4];\n"
+"	};\n"
+"} btAabbCL;\n"
+"#ifndef B3_AABB_H\n"
+"#define B3_AABB_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#define B3_FLOAT4_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#define B3_PLATFORM_DEFINITIONS_H\n"
+"struct MyTest\n"
+"{\n"
+"	int bla;\n"
+"};\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"//keep B3_LARGE_FLOAT*B3_LARGE_FLOAT < FLT_MAX\n"
+"#define B3_LARGE_FLOAT 1e18f\n"
+"#define B3_INFINITY 1e18f\n"
+"#define b3Assert(a)\n"
+"#define b3ConstArray(a) __global const a*\n"
+"#define b3AtomicInc atomic_inc\n"
+"#define b3AtomicAdd atomic_add\n"
+"#define b3Fabs fabs\n"
+"#define b3Sqrt native_sqrt\n"
+"#define b3Sin native_sin\n"
+"#define b3Cos native_cos\n"
+"#define B3_STATIC\n"
+"#endif\n"
+"#endif\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Float4;\n"
+"	#define b3Float4ConstArg const b3Float4\n"
+"	#define b3MakeFloat4 (float4)\n"
+"	float b3Dot3F4(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return dot(a1, b1);\n"
+"	}\n"
+"	b3Float4 b3Cross3(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return cross(a1, b1);\n"
+"	}\n"
+"	#define b3MinFloat4 min\n"
+"	#define b3MaxFloat4 max\n"
+"	#define b3Normalized(a) normalize(a)\n"
+"#endif \n"
+"		\n"
+"inline bool b3IsAlmostZero(b3Float4ConstArg v)\n"
+"{\n"
+"	if(b3Fabs(v.x)>1e-6 || b3Fabs(v.y)>1e-6 || b3Fabs(v.z)>1e-6)	\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"inline int    b3MaxDot( b3Float4ConstArg vec, __global const b3Float4* vecArray, int vecLen, float* dotOut )\n"
+"{\n"
+"    float maxDot = -B3_INFINITY;\n"
+"    int i = 0;\n"
+"    int ptIndex = -1;\n"
+"    for( i = 0; i < vecLen; i++ )\n"
+"    {\n"
+"        float dot = b3Dot3F4(vecArray[i],vec);\n"
+"            \n"
+"        if( dot > maxDot )\n"
+"        {\n"
+"            maxDot = dot;\n"
+"            ptIndex = i;\n"
+"        }\n"
+"    }\n"
+"	b3Assert(ptIndex>=0);\n"
+"    if (ptIndex<0)\n"
+"	{\n"
+"		ptIndex = 0;\n"
+"	}\n"
+"    *dotOut = maxDot;\n"
+"    return ptIndex;\n"
+"}\n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_MAT3x3_H\n"
+"#define B3_MAT3x3_H\n"
+"#ifndef B3_QUAT_H\n"
+"#define B3_QUAT_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif\n"
+"#endif\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Quat;\n"
+"	#define b3QuatConstArg const b3Quat\n"
+"	\n"
+"	\n"
+"inline float4 b3FastNormalize4(float4 v)\n"
+"{\n"
+"	v = (float4)(v.xyz,0.f);\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"	\n"
+"inline b3Quat b3QuatMul(b3Quat a, b3Quat b);\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in);\n"
+"inline b3Quat b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec);\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatMul(b3QuatConstArg a, b3QuatConstArg b)\n"
+"{\n"
+"	b3Quat ans;\n"
+"	ans = b3Cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - b3Dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in)\n"
+"{\n"
+"	b3Quat q;\n"
+"	q=in;\n"
+"	//return b3FastNormalize4(in);\n"
+"	float len = native_sqrt(dot(q, q));\n"
+"	if(len > 0.f)\n"
+"	{\n"
+"		q *= 1.f / len;\n"
+"	}\n"
+"	else\n"
+"	{\n"
+"		q.x = q.y = q.z = 0.f;\n"
+"		q.w = 1.f;\n"
+"	}\n"
+"	return q;\n"
+"}\n"
+"inline float4 b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	b3Quat qInv = b3QuatInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = b3QuatMul(b3QuatMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline float4 b3QuatInvRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	return b3QuatRotate( b3QuatInvert( q ), vec );\n"
+"}\n"
+"inline b3Float4 b3TransformPoint(b3Float4ConstArg point, b3Float4ConstArg translation, b3QuatConstArg  orientation)\n"
+"{\n"
+"	return b3QuatRotate( orientation, point ) + (translation);\n"
+"}\n"
+"	\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"typedef struct\n"
+"{\n"
+"	b3Float4 m_row[3];\n"
+"}b3Mat3x3;\n"
+"#define b3Mat3x3ConstArg const b3Mat3x3\n"
+"#define b3GetRow(m,row) (m.m_row[row])\n"
+"inline b3Mat3x3 b3QuatGetRotationMatrix(b3Quat quat)\n"
+"{\n"
+"	b3Float4 quat2 = (b3Float4)(quat.x*quat.x, quat.y*quat.y, quat.z*quat.z, 0.f);\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0].x=1-2*quat2.y-2*quat2.z;\n"
+"	out.m_row[0].y=2*quat.x*quat.y-2*quat.w*quat.z;\n"
+"	out.m_row[0].z=2*quat.x*quat.z+2*quat.w*quat.y;\n"
+"	out.m_row[0].w = 0.f;\n"
+"	out.m_row[1].x=2*quat.x*quat.y+2*quat.w*quat.z;\n"
+"	out.m_row[1].y=1-2*quat2.x-2*quat2.z;\n"
+"	out.m_row[1].z=2*quat.y*quat.z-2*quat.w*quat.x;\n"
+"	out.m_row[1].w = 0.f;\n"
+"	out.m_row[2].x=2*quat.x*quat.z-2*quat.w*quat.y;\n"
+"	out.m_row[2].y=2*quat.y*quat.z+2*quat.w*quat.x;\n"
+"	out.m_row[2].z=1-2*quat2.x-2*quat2.y;\n"
+"	out.m_row[2].w = 0.f;\n"
+"	return out;\n"
+"}\n"
+"inline b3Mat3x3 b3AbsoluteMat3x3(b3Mat3x3ConstArg matIn)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = fabs(matIn.m_row[0]);\n"
+"	out.m_row[1] = fabs(matIn.m_row[1]);\n"
+"	out.m_row[2] = fabs(matIn.m_row[2]);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtZero();\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity();\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m);\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b);\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Mat3x3 mtZero()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(0.f);\n"
+"	m.m_row[1] = (b3Float4)(0.f);\n"
+"	m.m_row[2] = (b3Float4)(0.f);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(1,0,0,0);\n"
+"	m.m_row[1] = (b3Float4)(0,1,0,0);\n"
+"	m.m_row[2] = (b3Float4)(0,0,1,0);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = (b3Float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (b3Float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (b3Float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Mat3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	b3Mat3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = b3Dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = b3Dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = b3Dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b)\n"
+"{\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a.m_row[0], b );\n"
+"	ans.y = b3Dot3F4( a.m_row[1], b );\n"
+"	ans.z = b3Dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Float4 colx = b3MakeFloat4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	b3Float4 coly = b3MakeFloat4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	b3Float4 colz = b3MakeFloat4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a, colx );\n"
+"	ans.y = b3Dot3F4( a, coly );\n"
+"	ans.z = b3Dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"#endif\n"
+"#endif //B3_MAT3x3_H\n"
+"typedef struct b3Aabb b3Aabb_t;\n"
+"struct b3Aabb\n"
+"{\n"
+"	union\n"
+"	{\n"
+"		float m_min[4];\n"
+"		b3Float4 m_minVec;\n"
+"		int m_minIndices[4];\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float	m_max[4];\n"
+"		b3Float4 m_maxVec;\n"
+"		int m_signedMaxIndices[4];\n"
+"	};\n"
+"};\n"
+"inline void b3TransformAabb2(b3Float4ConstArg localAabbMin,b3Float4ConstArg localAabbMax, float margin,\n"
+"						b3Float4ConstArg pos,\n"
+"						b3QuatConstArg orn,\n"
+"						b3Float4* aabbMinOut,b3Float4* aabbMaxOut)\n"
+"{\n"
+"		b3Float4 localHalfExtents = 0.5f*(localAabbMax-localAabbMin);\n"
+"		localHalfExtents+=b3MakeFloat4(margin,margin,margin,0.f);\n"
+"		b3Float4 localCenter = 0.5f*(localAabbMax+localAabbMin);\n"
+"		b3Mat3x3 m;\n"
+"		m = b3QuatGetRotationMatrix(orn);\n"
+"		b3Mat3x3 abs_b = b3AbsoluteMat3x3(m);\n"
+"		b3Float4 center = b3TransformPoint(localCenter,pos,orn);\n"
+"		\n"
+"		b3Float4 extent = b3MakeFloat4(b3Dot3F4(localHalfExtents,b3GetRow(abs_b,0)),\n"
+"										 b3Dot3F4(localHalfExtents,b3GetRow(abs_b,1)),\n"
+"										 b3Dot3F4(localHalfExtents,b3GetRow(abs_b,2)),\n"
+"										 0.f);\n"
+"		*aabbMinOut = center-extent;\n"
+"		*aabbMaxOut = center+extent;\n"
+"}\n"
+"/// conservative test for overlap between two aabbs\n"
+"inline bool b3TestAabbAgainstAabb(b3Float4ConstArg aabbMin1,b3Float4ConstArg aabbMax1,\n"
+"								b3Float4ConstArg aabbMin2, b3Float4ConstArg aabbMax2)\n"
+"{\n"
+"	bool overlap = true;\n"
+"	overlap = (aabbMin1.x > aabbMax2.x || aabbMax1.x < aabbMin2.x) ? false : overlap;\n"
+"	overlap = (aabbMin1.z > aabbMax2.z || aabbMax1.z < aabbMin2.z) ? false : overlap;\n"
+"	overlap = (aabbMin1.y > aabbMax2.y || aabbMax1.y < aabbMin2.y) ? false : overlap;\n"
+"	return overlap;\n"
+"}\n"
+"#endif //B3_AABB_H\n"
+"/*\n"
+"Bullet Continuous Collision Detection and Physics Library\n"
+"Copyright (c) 2003-2013 Erwin Coumans  http://bulletphysics.org\n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose,\n"
+"including commercial applications, and to alter it and redistribute it freely,\n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"#ifndef B3_INT2_H\n"
+"#define B3_INT2_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#define b3UnsignedInt2 uint2\n"
+"#define b3Int2 int2\n"
+"#define b3MakeInt2 (int2)\n"
+"#endif //__cplusplus\n"
+"#endif\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_plane;\n"
+"	int m_indexOffset;\n"
+"	int m_numIndices;\n"
+"} btGpuFace;\n"
+"#define make_float4 (float4)\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"	\n"
+"//	float4 a1 = make_float4(a.xyz,0.f);\n"
+"//	float4 b1 = make_float4(b.xyz,0.f);\n"
+"//	return cross(a1,b1);\n"
+"//float4 c = make_float4(a.y*b.z - a.z*b.y,a.z*b.x - a.x*b.z,a.x*b.y - a.y*b.x,0.f);\n"
+"	\n"
+"	//	float4 c = make_float4(a.y*b.z - a.z*b.y,1.f,a.x*b.y - a.y*b.x,0.f);\n"
+"	\n"
+"	//return c;\n"
+"}\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = make_float4(a.xyz,0.f);\n"
+"	float4 b1 = make_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	v = make_float4(v.xyz,0.f);\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"typedef float4 Quaternion;\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b);\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in);\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec);\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q);\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in)\n"
+"{\n"
+"	return fastNormalize4(in);\n"
+"//	in /= length( in );\n"
+"//	return in;\n"
+"}\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(qtMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"__inline\n"
+"float4 qtInvRotate(const Quaternion q, float4 vec)\n"
+"{\n"
+"	return qtRotate( qtInvert( q ), vec );\n"
+"}\n"
+"__inline\n"
+"float4 transform(const float4* p, const float4* translation, const Quaternion* orientation)\n"
+"{\n"
+"	return qtRotate( *orientation, *p ) + (*translation);\n"
+"}\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = make_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"}\n"
+"inline void projectLocal(const ConvexPolyhedronCL* hull,  const float4 pos, const float4 orn, \n"
+"const float4* dir, const float4* vertices, float* min, float* max)\n"
+"{\n"
+"	min[0] = FLT_MAX;\n"
+"	max[0] = -FLT_MAX;\n"
+"	int numVerts = hull->m_numVertices;\n"
+"	const float4 localDir = qtInvRotate(orn,*dir);\n"
+"	float offset = dot(pos,*dir);\n"
+"	for(int i=0;i<numVerts;i++)\n"
+"	{\n"
+"		float dp = dot(vertices[hull->m_vertexOffset+i],localDir);\n"
+"		if(dp < min[0])	\n"
+"			min[0] = dp;\n"
+"		if(dp > max[0])	\n"
+"			max[0] = dp;\n"
+"	}\n"
+"	if(min[0]>max[0])\n"
+"	{\n"
+"		float tmp = min[0];\n"
+"		min[0] = max[0];\n"
+"		max[0] = tmp;\n"
+"	}\n"
+"	min[0] += offset;\n"
+"	max[0] += offset;\n"
+"}\n"
+"inline void project(__global const ConvexPolyhedronCL* hull,  const float4 pos, const float4 orn, \n"
+"const float4* dir, __global const float4* vertices, float* min, float* max)\n"
+"{\n"
+"	min[0] = FLT_MAX;\n"
+"	max[0] = -FLT_MAX;\n"
+"	int numVerts = hull->m_numVertices;\n"
+"	const float4 localDir = qtInvRotate(orn,*dir);\n"
+"	float offset = dot(pos,*dir);\n"
+"	for(int i=0;i<numVerts;i++)\n"
+"	{\n"
+"		float dp = dot(vertices[hull->m_vertexOffset+i],localDir);\n"
+"		if(dp < min[0])	\n"
+"			min[0] = dp;\n"
+"		if(dp > max[0])	\n"
+"			max[0] = dp;\n"
+"	}\n"
+"	if(min[0]>max[0])\n"
+"	{\n"
+"		float tmp = min[0];\n"
+"		min[0] = max[0];\n"
+"		max[0] = tmp;\n"
+"	}\n"
+"	min[0] += offset;\n"
+"	max[0] += offset;\n"
+"}\n"
+"inline bool TestSepAxisLocalA(const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, \n"
+"	const float4 posA,const float4 ornA,\n"
+"	const float4 posB,const float4 ornB,\n"
+"	float4* sep_axis, const float4* verticesA, __global const float4* verticesB,float* depth)\n"
+"{\n"
+"	float Min0,Max0;\n"
+"	float Min1,Max1;\n"
+"	projectLocal(hullA,posA,ornA,sep_axis,verticesA, &Min0, &Max0);\n"
+"	project(hullB,posB,ornB, sep_axis,verticesB, &Min1, &Max1);\n"
+"	if(Max0<Min1 || Max1<Min0)\n"
+"		return false;\n"
+"	float d0 = Max0 - Min1;\n"
+"	float d1 = Max1 - Min0;\n"
+"	*depth = d0<d1 ? d0:d1;\n"
+"	return true;\n"
+"}\n"
+"inline bool IsAlmostZero(const float4 v)\n"
+"{\n"
+"	if(fabs(v.x)>1e-6f || fabs(v.y)>1e-6f || fabs(v.z)>1e-6f)\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"bool findSeparatingAxisLocalA(	const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, \n"
+"	const float4 posA1,\n"
+"	const float4 ornA,\n"
+"	const float4 posB1,\n"
+"	const float4 ornB,\n"
+"	const float4 DeltaC2,\n"
+"	\n"
+"	const float4* verticesA, \n"
+"	const float4* uniqueEdgesA, \n"
+"	const btGpuFace* facesA,\n"
+"	const int*  indicesA,\n"
+"	__global const float4* verticesB, \n"
+"	__global const float4* uniqueEdgesB, \n"
+"	__global const btGpuFace* facesB,\n"
+"	__global const int*  indicesB,\n"
+"	float4* sep,\n"
+"	float* dmin)\n"
+"{\n"
+"	\n"
+"	float4 posA = posA1;\n"
+"	posA.w = 0.f;\n"
+"	float4 posB = posB1;\n"
+"	posB.w = 0.f;\n"
+"	int curPlaneTests=0;\n"
+"	{\n"
+"		int numFacesA = hullA->m_numFaces;\n"
+"		// Test normals from hullA\n"
+"		for(int i=0;i<numFacesA;i++)\n"
+"		{\n"
+"			const float4 normal = facesA[hullA->m_faceOffset+i].m_plane;\n"
+"			float4 faceANormalWS = qtRotate(ornA,normal);\n"
+"			if (dot3F4(DeltaC2,faceANormalWS)<0)\n"
+"				faceANormalWS*=-1.f;\n"
+"			curPlaneTests++;\n"
+"			float d;\n"
+"			if(!TestSepAxisLocalA( hullA, hullB, posA,ornA,posB,ornB,&faceANormalWS, verticesA, verticesB,&d))\n"
+"				return false;\n"
+"			if(d<*dmin)\n"
+"			{\n"
+"				*dmin = d;\n"
+"				*sep = faceANormalWS;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	if((dot3F4(-DeltaC2,*sep))>0.0f)\n"
+"	{\n"
+"		*sep = -(*sep);\n"
+"	}\n"
+"	return true;\n"
+"}\n"
+"bool findSeparatingAxisLocalB(	__global const ConvexPolyhedronCL* hullA,  const ConvexPolyhedronCL* hullB, \n"
+"	const float4 posA1,\n"
+"	const float4 ornA,\n"
+"	const float4 posB1,\n"
+"	const float4 ornB,\n"
+"	const float4 DeltaC2,\n"
+"	__global const float4* verticesA, \n"
+"	__global const float4* uniqueEdgesA, \n"
+"	__global const btGpuFace* facesA,\n"
+"	__global const int*  indicesA,\n"
+"	const float4* verticesB,\n"
+"	const float4* uniqueEdgesB, \n"
+"	const btGpuFace* facesB,\n"
+"	const int*  indicesB,\n"
+"	float4* sep,\n"
+"	float* dmin)\n"
+"{\n"
+"	float4 posA = posA1;\n"
+"	posA.w = 0.f;\n"
+"	float4 posB = posB1;\n"
+"	posB.w = 0.f;\n"
+"	int curPlaneTests=0;\n"
+"	{\n"
+"		int numFacesA = hullA->m_numFaces;\n"
+"		// Test normals from hullA\n"
+"		for(int i=0;i<numFacesA;i++)\n"
+"		{\n"
+"			const float4 normal = facesA[hullA->m_faceOffset+i].m_plane;\n"
+"			float4 faceANormalWS = qtRotate(ornA,normal);\n"
+"			if (dot3F4(DeltaC2,faceANormalWS)<0)\n"
+"				faceANormalWS *= -1.f;\n"
+"			curPlaneTests++;\n"
+"			float d;\n"
+"			if(!TestSepAxisLocalA( hullB, hullA, posB,ornB,posA,ornA, &faceANormalWS, verticesB,verticesA, &d))\n"
+"				return false;\n"
+"			if(d<*dmin)\n"
+"			{\n"
+"				*dmin = d;\n"
+"				*sep = faceANormalWS;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	if((dot3F4(-DeltaC2,*sep))>0.0f)\n"
+"	{\n"
+"		*sep = -(*sep);\n"
+"	}\n"
+"	return true;\n"
+"}\n"
+"bool findSeparatingAxisEdgeEdgeLocalA(	const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, \n"
+"	const float4 posA1,\n"
+"	const float4 ornA,\n"
+"	const float4 posB1,\n"
+"	const float4 ornB,\n"
+"	const float4 DeltaC2,\n"
+"	const float4* verticesA, \n"
+"	const float4* uniqueEdgesA, \n"
+"	const btGpuFace* facesA,\n"
+"	const int*  indicesA,\n"
+"	__global const float4* verticesB, \n"
+"	__global const float4* uniqueEdgesB, \n"
+"	__global const btGpuFace* facesB,\n"
+"	__global const int*  indicesB,\n"
+"		float4* sep,\n"
+"	float* dmin)\n"
+"{\n"
+"	float4 posA = posA1;\n"
+"	posA.w = 0.f;\n"
+"	float4 posB = posB1;\n"
+"	posB.w = 0.f;\n"
+"	int curPlaneTests=0;\n"
+"	int curEdgeEdge = 0;\n"
+"	// Test edges\n"
+"	for(int e0=0;e0<hullA->m_numUniqueEdges;e0++)\n"
+"	{\n"
+"		const float4 edge0 = uniqueEdgesA[hullA->m_uniqueEdgesOffset+e0];\n"
+"		float4 edge0World = qtRotate(ornA,edge0);\n"
+"		for(int e1=0;e1<hullB->m_numUniqueEdges;e1++)\n"
+"		{\n"
+"			const float4 edge1 = uniqueEdgesB[hullB->m_uniqueEdgesOffset+e1];\n"
+"			float4 edge1World = qtRotate(ornB,edge1);\n"
+"			float4 crossje = cross3(edge0World,edge1World);\n"
+"			curEdgeEdge++;\n"
+"			if(!IsAlmostZero(crossje))\n"
+"			{\n"
+"				crossje = normalize3(crossje);\n"
+"				if (dot3F4(DeltaC2,crossje)<0)\n"
+"					crossje *= -1.f;\n"
+"				float dist;\n"
+"				bool result = true;\n"
+"				{\n"
+"					float Min0,Max0;\n"
+"					float Min1,Max1;\n"
+"					projectLocal(hullA,posA,ornA,&crossje,verticesA, &Min0, &Max0);\n"
+"					project(hullB,posB,ornB,&crossje,verticesB, &Min1, &Max1);\n"
+"				\n"
+"					if(Max0<Min1 || Max1<Min0)\n"
+"						result = false;\n"
+"				\n"
+"					float d0 = Max0 - Min1;\n"
+"					float d1 = Max1 - Min0;\n"
+"					dist = d0<d1 ? d0:d1;\n"
+"					result = true;\n"
+"				}\n"
+"				\n"
+"				if(dist<*dmin)\n"
+"				{\n"
+"					*dmin = dist;\n"
+"					*sep = crossje;\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	\n"
+"	if((dot3F4(-DeltaC2,*sep))>0.0f)\n"
+"	{\n"
+"		*sep = -(*sep);\n"
+"	}\n"
+"	return true;\n"
+"}\n"
+"inline bool TestSepAxis(__global const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, \n"
+"	const float4 posA,const float4 ornA,\n"
+"	const float4 posB,const float4 ornB,\n"
+"	float4* sep_axis, __global const float4* vertices,float* depth)\n"
+"{\n"
+"	float Min0,Max0;\n"
+"	float Min1,Max1;\n"
+"	project(hullA,posA,ornA,sep_axis,vertices, &Min0, &Max0);\n"
+"	project(hullB,posB,ornB, sep_axis,vertices, &Min1, &Max1);\n"
+"	if(Max0<Min1 || Max1<Min0)\n"
+"		return false;\n"
+"	float d0 = Max0 - Min1;\n"
+"	float d1 = Max1 - Min0;\n"
+"	*depth = d0<d1 ? d0:d1;\n"
+"	return true;\n"
+"}\n"
+"bool findSeparatingAxis(	__global const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, \n"
+"	const float4 posA1,\n"
+"	const float4 ornA,\n"
+"	const float4 posB1,\n"
+"	const float4 ornB,\n"
+"	const float4 DeltaC2,\n"
+"	__global const float4* vertices, \n"
+"	__global const float4* uniqueEdges, \n"
+"	__global const btGpuFace* faces,\n"
+"	__global const int*  indices,\n"
+"	float4* sep,\n"
+"	float* dmin)\n"
+"{\n"
+"	\n"
+"	float4 posA = posA1;\n"
+"	posA.w = 0.f;\n"
+"	float4 posB = posB1;\n"
+"	posB.w = 0.f;\n"
+"	\n"
+"	int curPlaneTests=0;\n"
+"	{\n"
+"		int numFacesA = hullA->m_numFaces;\n"
+"		// Test normals from hullA\n"
+"		for(int i=0;i<numFacesA;i++)\n"
+"		{\n"
+"			const float4 normal = faces[hullA->m_faceOffset+i].m_plane;\n"
+"			float4 faceANormalWS = qtRotate(ornA,normal);\n"
+"	\n"
+"			if (dot3F4(DeltaC2,faceANormalWS)<0)\n"
+"				faceANormalWS*=-1.f;\n"
+"				\n"
+"			curPlaneTests++;\n"
+"	\n"
+"			float d;\n"
+"			if(!TestSepAxis( hullA, hullB, posA,ornA,posB,ornB,&faceANormalWS, vertices,&d))\n"
+"				return false;\n"
+"	\n"
+"			if(d<*dmin)\n"
+"			{\n"
+"				*dmin = d;\n"
+"				*sep = faceANormalWS;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"		if((dot3F4(-DeltaC2,*sep))>0.0f)\n"
+"		{\n"
+"			*sep = -(*sep);\n"
+"		}\n"
+"	\n"
+"	return true;\n"
+"}\n"
+"bool findSeparatingAxisUnitSphere(	__global const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, \n"
+"	const float4 posA1,\n"
+"	const float4 ornA,\n"
+"	const float4 posB1,\n"
+"	const float4 ornB,\n"
+"	const float4 DeltaC2,\n"
+"	__global const float4* vertices,\n"
+"	__global const float4* unitSphereDirections,\n"
+"	int numUnitSphereDirections,\n"
+"	float4* sep,\n"
+"	float* dmin)\n"
+"{\n"
+"	\n"
+"	float4 posA = posA1;\n"
+"	posA.w = 0.f;\n"
+"	float4 posB = posB1;\n"
+"	posB.w = 0.f;\n"
+"	int curPlaneTests=0;\n"
+"	int curEdgeEdge = 0;\n"
+"	// Test unit sphere directions\n"
+"	for (int i=0;i<numUnitSphereDirections;i++)\n"
+"	{\n"
+"		float4 crossje;\n"
+"		crossje = unitSphereDirections[i];	\n"
+"		if (dot3F4(DeltaC2,crossje)>0)\n"
+"			crossje *= -1.f;\n"
+"		{\n"
+"			float dist;\n"
+"			bool result = true;\n"
+"			float Min0,Max0;\n"
+"			float Min1,Max1;\n"
+"			project(hullA,posA,ornA,&crossje,vertices, &Min0, &Max0);\n"
+"			project(hullB,posB,ornB,&crossje,vertices, &Min1, &Max1);\n"
+"		\n"
+"			if(Max0<Min1 || Max1<Min0)\n"
+"				return false;\n"
+"		\n"
+"			float d0 = Max0 - Min1;\n"
+"			float d1 = Max1 - Min0;\n"
+"			dist = d0<d1 ? d0:d1;\n"
+"			result = true;\n"
+"	\n"
+"			if(dist<*dmin)\n"
+"			{\n"
+"				*dmin = dist;\n"
+"				*sep = crossje;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	\n"
+"	if((dot3F4(-DeltaC2,*sep))>0.0f)\n"
+"	{\n"
+"		*sep = -(*sep);\n"
+"	}\n"
+"	return true;\n"
+"}\n"
+"bool findSeparatingAxisEdgeEdge(	__global const ConvexPolyhedronCL* hullA, __global const ConvexPolyhedronCL* hullB, \n"
+"	const float4 posA1,\n"
+"	const float4 ornA,\n"
+"	const float4 posB1,\n"
+"	const float4 ornB,\n"
+"	const float4 DeltaC2,\n"
+"	__global const float4* vertices, \n"
+"	__global const float4* uniqueEdges, \n"
+"	__global const btGpuFace* faces,\n"
+"	__global const int*  indices,\n"
+"	float4* sep,\n"
+"	float* dmin)\n"
+"{\n"
+"	\n"
+"	float4 posA = posA1;\n"
+"	posA.w = 0.f;\n"
+"	float4 posB = posB1;\n"
+"	posB.w = 0.f;\n"
+"	int curPlaneTests=0;\n"
+"	int curEdgeEdge = 0;\n"
+"	// Test edges\n"
+"	for(int e0=0;e0<hullA->m_numUniqueEdges;e0++)\n"
+"	{\n"
+"		const float4 edge0 = uniqueEdges[hullA->m_uniqueEdgesOffset+e0];\n"
+"		float4 edge0World = qtRotate(ornA,edge0);\n"
+"		for(int e1=0;e1<hullB->m_numUniqueEdges;e1++)\n"
+"		{\n"
+"			const float4 edge1 = uniqueEdges[hullB->m_uniqueEdgesOffset+e1];\n"
+"			float4 edge1World = qtRotate(ornB,edge1);\n"
+"			float4 crossje = cross3(edge0World,edge1World);\n"
+"			curEdgeEdge++;\n"
+"			if(!IsAlmostZero(crossje))\n"
+"			{\n"
+"				crossje = normalize3(crossje);\n"
+"				if (dot3F4(DeltaC2,crossje)<0)\n"
+"					crossje*=-1.f;\n"
+"					\n"
+"				float dist;\n"
+"				bool result = true;\n"
+"				{\n"
+"					float Min0,Max0;\n"
+"					float Min1,Max1;\n"
+"					project(hullA,posA,ornA,&crossje,vertices, &Min0, &Max0);\n"
+"					project(hullB,posB,ornB,&crossje,vertices, &Min1, &Max1);\n"
+"				\n"
+"					if(Max0<Min1 || Max1<Min0)\n"
+"						return false;\n"
+"				\n"
+"					float d0 = Max0 - Min1;\n"
+"					float d1 = Max1 - Min0;\n"
+"					dist = d0<d1 ? d0:d1;\n"
+"					result = true;\n"
+"				}\n"
+"				\n"
+"				if(dist<*dmin)\n"
+"				{\n"
+"					*dmin = dist;\n"
+"					*sep = crossje;\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	\n"
+"	if((dot3F4(-DeltaC2,*sep))>0.0f)\n"
+"	{\n"
+"		*sep = -(*sep);\n"
+"	}\n"
+"	return true;\n"
+"}\n"
+"// work-in-progress\n"
+"__kernel void   processCompoundPairsKernel( __global const int4* gpuCompoundPairs,\n"
+"																					__global const BodyData* rigidBodies, \n"
+"																					__global const btCollidableGpu* collidables,\n"
+"																					__global const ConvexPolyhedronCL* convexShapes, \n"
+"																					__global const float4* vertices,\n"
+"																					__global const float4* uniqueEdges,\n"
+"																					__global const btGpuFace* faces,\n"
+"																					__global const int* indices,\n"
+"																					__global btAabbCL* aabbs,\n"
+"																					__global const btGpuChildShape* gpuChildShapes,\n"
+"																					__global volatile float4* gpuCompoundSepNormalsOut,\n"
+"																					__global volatile int* gpuHasCompoundSepNormalsOut,\n"
+"																					int numCompoundPairs\n"
+"																					)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i<numCompoundPairs)\n"
+"	{\n"
+"		int bodyIndexA = gpuCompoundPairs[i].x;\n"
+"		int bodyIndexB = gpuCompoundPairs[i].y;\n"
+"		int childShapeIndexA = gpuCompoundPairs[i].z;\n"
+"		int childShapeIndexB = gpuCompoundPairs[i].w;\n"
+"		\n"
+"		int collidableIndexA = -1;\n"
+"		int collidableIndexB = -1;\n"
+"		\n"
+"		float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"		float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"		\n"
+"		float4 ornB = rigidBodies[bodyIndexB].m_quat;\n"
+"		float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"							\n"
+"		if (childShapeIndexA >= 0)\n"
+"		{\n"
+"			collidableIndexA = gpuChildShapes[childShapeIndexA].m_shapeIndex;\n"
+"			float4 childPosA = gpuChildShapes[childShapeIndexA].m_childPosition;\n"
+"			float4 childOrnA = gpuChildShapes[childShapeIndexA].m_childOrientation;\n"
+"			float4 newPosA = qtRotate(ornA,childPosA)+posA;\n"
+"			float4 newOrnA = qtMul(ornA,childOrnA);\n"
+"			posA = newPosA;\n"
+"			ornA = newOrnA;\n"
+"		} else\n"
+"		{\n"
+"			collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"		}\n"
+"		\n"
+"		if (childShapeIndexB>=0)\n"
+"		{\n"
+"			collidableIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;\n"
+"			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;\n"
+"			float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;\n"
+"			float4 newPosB = transform(&childPosB,&posB,&ornB);\n"
+"			float4 newOrnB = qtMul(ornB,childOrnB);\n"
+"			posB = newPosB;\n"
+"			ornB = newOrnB;\n"
+"		} else\n"
+"		{\n"
+"			collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;	\n"
+"		}\n"
+"	\n"
+"		gpuHasCompoundSepNormalsOut[i] = 0;\n"
+"	\n"
+"		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"	\n"
+"		int shapeTypeA = collidables[collidableIndexA].m_shapeType;\n"
+"		int shapeTypeB = collidables[collidableIndexB].m_shapeType;\n"
+"	\n"
+"		if ((shapeTypeA != SHAPE_CONVEX_HULL) || (shapeTypeB != SHAPE_CONVEX_HULL))\n"
+"		{\n"
+"			return;\n"
+"		}\n"
+"		int hasSeparatingAxis = 5;\n"
+"							\n"
+"		int numFacesA = convexShapes[shapeIndexA].m_numFaces;\n"
+"		float dmin = FLT_MAX;\n"
+"		posA.w = 0.f;\n"
+"		posB.w = 0.f;\n"
+"		float4 c0local = convexShapes[shapeIndexA].m_localCenter;\n"
+"		float4 c0 = transform(&c0local, &posA, &ornA);\n"
+"		float4 c1local = convexShapes[shapeIndexB].m_localCenter;\n"
+"		float4 c1 = transform(&c1local,&posB,&ornB);\n"
+"		const float4 DeltaC2 = c0 - c1;\n"
+"		float4 sepNormal = make_float4(1,0,0,0);\n"
+"		bool sepA = findSeparatingAxis(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,posB,ornB,DeltaC2,vertices,uniqueEdges,faces,indices,&sepNormal,&dmin);\n"
+"		hasSeparatingAxis = 4;\n"
+"		if (!sepA)\n"
+"		{\n"
+"			hasSeparatingAxis = 0;\n"
+"		} else\n"
+"		{\n"
+"			bool sepB = findSeparatingAxis(	&convexShapes[shapeIndexB],&convexShapes[shapeIndexA],posB,ornB,posA,ornA,DeltaC2,vertices,uniqueEdges,faces,indices,&sepNormal,&dmin);\n"
+"			if (!sepB)\n"
+"			{\n"
+"				hasSeparatingAxis = 0;\n"
+"			} else//(!sepB)\n"
+"			{\n"
+"				bool sepEE = findSeparatingAxisEdgeEdge(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,posB,ornB,DeltaC2,vertices,uniqueEdges,faces,indices,&sepNormal,&dmin);\n"
+"				if (sepEE)\n"
+"				{\n"
+"						gpuCompoundSepNormalsOut[i] = sepNormal;//fastNormalize4(sepNormal);\n"
+"						gpuHasCompoundSepNormalsOut[i] = 1;\n"
+"				}//sepEE\n"
+"			}//(!sepB)\n"
+"		}//(!sepA)\n"
+"		\n"
+"		\n"
+"	}\n"
+"		\n"
+"}\n"
+"inline b3Float4 MyUnQuantize(const unsigned short* vecIn, b3Float4 quantization, b3Float4 bvhAabbMin)\n"
+"{\n"
+"		b3Float4 vecOut;\n"
+"		vecOut = b3MakeFloat4(\n"
+"			(float)(vecIn[0]) / (quantization.x),\n"
+"			(float)(vecIn[1]) / (quantization.y),\n"
+"			(float)(vecIn[2]) / (quantization.z),\n"
+"			0.f);\n"
+"		vecOut += bvhAabbMin;\n"
+"		return vecOut;\n"
+"}\n"
+"inline b3Float4 MyUnQuantizeGlobal(__global const unsigned short* vecIn, b3Float4 quantization, b3Float4 bvhAabbMin)\n"
+"{\n"
+"		b3Float4 vecOut;\n"
+"		vecOut = b3MakeFloat4(\n"
+"			(float)(vecIn[0]) / (quantization.x),\n"
+"			(float)(vecIn[1]) / (quantization.y),\n"
+"			(float)(vecIn[2]) / (quantization.z),\n"
+"			0.f);\n"
+"		vecOut += bvhAabbMin;\n"
+"		return vecOut;\n"
+"}\n"
+"// work-in-progress\n"
+"__kernel void   findCompoundPairsKernel( __global const int4* pairs, \n"
+"	__global const BodyData* rigidBodies, \n"
+"	__global const btCollidableGpu* collidables,\n"
+"	__global const ConvexPolyhedronCL* convexShapes, \n"
+"	__global const float4* vertices,\n"
+"	__global const float4* uniqueEdges,\n"
+"	__global const btGpuFace* faces,\n"
+"	__global const int* indices,\n"
+"	__global b3Aabb_t* aabbLocalSpace,\n"
+"	__global const btGpuChildShape* gpuChildShapes,\n"
+"	__global volatile int4* gpuCompoundPairsOut,\n"
+"	__global volatile int* numCompoundPairsOut,\n"
+"	__global const b3BvhSubtreeInfo* subtrees,\n"
+"	__global const b3QuantizedBvhNode* quantizedNodes,\n"
+"	__global const b3BvhInfo* bvhInfos,\n"
+"	int numPairs,\n"
+"	int maxNumCompoundPairsCapacity\n"
+"	)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i<numPairs)\n"
+"	{\n"
+"		int bodyIndexA = pairs[i].x;\n"
+"		int bodyIndexB = pairs[i].y;\n"
+"		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"		//once the broadphase avoids static-static pairs, we can remove this test\n"
+"		if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))\n"
+"		{\n"
+"			return;\n"
+"		}\n"
+"		if ((collidables[collidableIndexA].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS) &&(collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS))\n"
+"		{\n"
+"			int bvhA = collidables[collidableIndexA].m_compoundBvhIndex;\n"
+"			int bvhB = collidables[collidableIndexB].m_compoundBvhIndex;\n"
+"			int numSubTreesA = bvhInfos[bvhA].m_numSubTrees;\n"
+"			int subTreesOffsetA = bvhInfos[bvhA].m_subTreeOffset;\n"
+"			int subTreesOffsetB = bvhInfos[bvhB].m_subTreeOffset;\n"
+"			int numSubTreesB = bvhInfos[bvhB].m_numSubTrees;\n"
+"			\n"
+"			float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"			b3Quat ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"			b3Quat ornB = rigidBodies[bodyIndexB].m_quat;\n"
+"			float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"			\n"
+"			for (int p=0;p<numSubTreesA;p++)\n"
+"			{\n"
+"				b3BvhSubtreeInfo subtreeA = subtrees[subTreesOffsetA+p];\n"
+"				//bvhInfos[bvhA].m_quantization\n"
+"				b3Float4 treeAminLocal = MyUnQuantize(subtreeA.m_quantizedAabbMin,bvhInfos[bvhA].m_quantization,bvhInfos[bvhA].m_aabbMin);\n"
+"				b3Float4 treeAmaxLocal = MyUnQuantize(subtreeA.m_quantizedAabbMax,bvhInfos[bvhA].m_quantization,bvhInfos[bvhA].m_aabbMin);\n"
+"				b3Float4 aabbAMinOut,aabbAMaxOut;\n"
+"				float margin=0.f;\n"
+"				b3TransformAabb2(treeAminLocal,treeAmaxLocal, margin,posA,ornA,&aabbAMinOut,&aabbAMaxOut);\n"
+"				\n"
+"				for (int q=0;q<numSubTreesB;q++)\n"
+"				{\n"
+"					b3BvhSubtreeInfo subtreeB = subtrees[subTreesOffsetB+q];\n"
+"					b3Float4 treeBminLocal = MyUnQuantize(subtreeB.m_quantizedAabbMin,bvhInfos[bvhB].m_quantization,bvhInfos[bvhB].m_aabbMin);\n"
+"					b3Float4 treeBmaxLocal = MyUnQuantize(subtreeB.m_quantizedAabbMax,bvhInfos[bvhB].m_quantization,bvhInfos[bvhB].m_aabbMin);\n"
+"					b3Float4 aabbBMinOut,aabbBMaxOut;\n"
+"					float margin=0.f;\n"
+"					b3TransformAabb2(treeBminLocal,treeBmaxLocal, margin,posB,ornB,&aabbBMinOut,&aabbBMaxOut);\n"
+"					\n"
+"					\n"
+"					bool aabbOverlap = b3TestAabbAgainstAabb(aabbAMinOut,aabbAMaxOut,aabbBMinOut,aabbBMaxOut);\n"
+"					if (aabbOverlap)\n"
+"					{\n"
+"						\n"
+"						int startNodeIndexA = subtreeA.m_rootNodeIndex+bvhInfos[bvhA].m_nodeOffset;\n"
+"						int endNodeIndexA = startNodeIndexA+subtreeA.m_subtreeSize;\n"
+"						int startNodeIndexB = subtreeB.m_rootNodeIndex+bvhInfos[bvhB].m_nodeOffset;\n"
+"						int endNodeIndexB = startNodeIndexB+subtreeB.m_subtreeSize;\n"
+"						b3Int2 nodeStack[B3_MAX_STACK_DEPTH];\n"
+"						b3Int2 node0;\n"
+"						node0.x = startNodeIndexA;\n"
+"						node0.y = startNodeIndexB;\n"
+"						int maxStackDepth = B3_MAX_STACK_DEPTH;\n"
+"						int depth=0;\n"
+"						nodeStack[depth++]=node0;\n"
+"						do\n"
+"						{\n"
+"							b3Int2 node = nodeStack[--depth];\n"
+"							b3Float4 aMinLocal = MyUnQuantizeGlobal(quantizedNodes[node.x].m_quantizedAabbMin,bvhInfos[bvhA].m_quantization,bvhInfos[bvhA].m_aabbMin);\n"
+"							b3Float4 aMaxLocal = MyUnQuantizeGlobal(quantizedNodes[node.x].m_quantizedAabbMax,bvhInfos[bvhA].m_quantization,bvhInfos[bvhA].m_aabbMin);\n"
+"							b3Float4 bMinLocal = MyUnQuantizeGlobal(quantizedNodes[node.y].m_quantizedAabbMin,bvhInfos[bvhB].m_quantization,bvhInfos[bvhB].m_aabbMin);\n"
+"							b3Float4 bMaxLocal = MyUnQuantizeGlobal(quantizedNodes[node.y].m_quantizedAabbMax,bvhInfos[bvhB].m_quantization,bvhInfos[bvhB].m_aabbMin);\n"
+"							float margin=0.f;\n"
+"							b3Float4 aabbAMinOut,aabbAMaxOut;\n"
+"							b3TransformAabb2(aMinLocal,aMaxLocal, margin,posA,ornA,&aabbAMinOut,&aabbAMaxOut);\n"
+"							b3Float4 aabbBMinOut,aabbBMaxOut;\n"
+"							b3TransformAabb2(bMinLocal,bMaxLocal, margin,posB,ornB,&aabbBMinOut,&aabbBMaxOut);\n"
+"							\n"
+"							bool nodeOverlap = b3TestAabbAgainstAabb(aabbAMinOut,aabbAMaxOut,aabbBMinOut,aabbBMaxOut);\n"
+"							if (nodeOverlap)\n"
+"							{\n"
+"								bool isLeafA = isLeafNodeGlobal(&quantizedNodes[node.x]);\n"
+"								bool isLeafB = isLeafNodeGlobal(&quantizedNodes[node.y]);\n"
+"								bool isInternalA = !isLeafA;\n"
+"								bool isInternalB = !isLeafB;\n"
+"								//fail, even though it might hit two leaf nodes\n"
+"								if (depth+4>maxStackDepth && !(isLeafA && isLeafB))\n"
+"								{\n"
+"									//printf(\"Error: traversal exceeded maxStackDepth\");\n"
+"									continue;\n"
+"								}\n"
+"								if(isInternalA)\n"
+"								{\n"
+"									int nodeAleftChild = node.x+1;\n"
+"									bool isNodeALeftChildLeaf = isLeafNodeGlobal(&quantizedNodes[node.x+1]);\n"
+"									int nodeArightChild = isNodeALeftChildLeaf? node.x+2 : node.x+1 + getEscapeIndexGlobal(&quantizedNodes[node.x+1]);\n"
+"									if(isInternalB)\n"
+"									{					\n"
+"										int nodeBleftChild = node.y+1;\n"
+"										bool isNodeBLeftChildLeaf = isLeafNodeGlobal(&quantizedNodes[node.y+1]);\n"
+"										int nodeBrightChild = isNodeBLeftChildLeaf? node.y+2 : node.y+1 + getEscapeIndexGlobal(&quantizedNodes[node.y+1]);\n"
+"										nodeStack[depth++] = b3MakeInt2(nodeAleftChild, nodeBleftChild);\n"
+"										nodeStack[depth++] = b3MakeInt2(nodeArightChild, nodeBleftChild);\n"
+"										nodeStack[depth++] = b3MakeInt2(nodeAleftChild, nodeBrightChild);\n"
+"										nodeStack[depth++] = b3MakeInt2(nodeArightChild, nodeBrightChild);\n"
+"									}\n"
+"									else\n"
+"									{\n"
+"										nodeStack[depth++] = b3MakeInt2(nodeAleftChild,node.y);\n"
+"										nodeStack[depth++] = b3MakeInt2(nodeArightChild,node.y);\n"
+"									}\n"
+"								}\n"
+"								else\n"
+"								{\n"
+"									if(isInternalB)\n"
+"									{\n"
+"										int nodeBleftChild = node.y+1;\n"
+"										bool isNodeBLeftChildLeaf = isLeafNodeGlobal(&quantizedNodes[node.y+1]);\n"
+"										int nodeBrightChild = isNodeBLeftChildLeaf? node.y+2 : node.y+1 + getEscapeIndexGlobal(&quantizedNodes[node.y+1]);\n"
+"										nodeStack[depth++] = b3MakeInt2(node.x,nodeBleftChild);\n"
+"										nodeStack[depth++] = b3MakeInt2(node.x,nodeBrightChild);\n"
+"									}\n"
+"									else\n"
+"									{\n"
+"										int compoundPairIdx = atomic_inc(numCompoundPairsOut);\n"
+"										if (compoundPairIdx<maxNumCompoundPairsCapacity)\n"
+"										{\n"
+"											int childShapeIndexA = getTriangleIndexGlobal(&quantizedNodes[node.x]);\n"
+"											int childShapeIndexB = getTriangleIndexGlobal(&quantizedNodes[node.y]);\n"
+"											gpuCompoundPairsOut[compoundPairIdx]  = (int4)(bodyIndexA,bodyIndexB,childShapeIndexA,childShapeIndexB);\n"
+"										}\n"
+"									}\n"
+"								}\n"
+"							}\n"
+"						} while (depth);\n"
+"					}\n"
+"				}\n"
+"			}\n"
+"			\n"
+"			return;\n"
+"		}\n"
+"		if ((collidables[collidableIndexA].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS) ||(collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS))\n"
+"		{\n"
+"			if (collidables[collidableIndexA].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS) \n"
+"			{\n"
+"				int numChildrenA = collidables[collidableIndexA].m_numChildShapes;\n"
+"				for (int c=0;c<numChildrenA;c++)\n"
+"				{\n"
+"					int childShapeIndexA = collidables[collidableIndexA].m_shapeIndex+c;\n"
+"					int childColIndexA = gpuChildShapes[childShapeIndexA].m_shapeIndex;\n"
+"					float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"					float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"					float4 childPosA = gpuChildShapes[childShapeIndexA].m_childPosition;\n"
+"					float4 childOrnA = gpuChildShapes[childShapeIndexA].m_childOrientation;\n"
+"					float4 newPosA = qtRotate(ornA,childPosA)+posA;\n"
+"					float4 newOrnA = qtMul(ornA,childOrnA);\n"
+"					int shapeIndexA = collidables[childColIndexA].m_shapeIndex;\n"
+"					b3Aabb_t aabbAlocal = aabbLocalSpace[shapeIndexA];\n"
+"					float margin = 0.f;\n"
+"					\n"
+"					b3Float4 aabbAMinWS;\n"
+"					b3Float4 aabbAMaxWS;\n"
+"					\n"
+"					b3TransformAabb2(aabbAlocal.m_minVec,aabbAlocal.m_maxVec,margin,\n"
+"						newPosA,\n"
+"						newOrnA,\n"
+"						&aabbAMinWS,&aabbAMaxWS);\n"
+"						\n"
+"					\n"
+"					if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)\n"
+"					{\n"
+"						int numChildrenB = collidables[collidableIndexB].m_numChildShapes;\n"
+"						for (int b=0;b<numChildrenB;b++)\n"
+"						{\n"
+"							int childShapeIndexB = collidables[collidableIndexB].m_shapeIndex+b;\n"
+"							int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;\n"
+"							float4 ornB = rigidBodies[bodyIndexB].m_quat;\n"
+"							float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"							float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;\n"
+"							float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;\n"
+"							float4 newPosB = transform(&childPosB,&posB,&ornB);\n"
+"							float4 newOrnB = qtMul(ornB,childOrnB);\n"
+"							int shapeIndexB = collidables[childColIndexB].m_shapeIndex;\n"
+"							b3Aabb_t aabbBlocal = aabbLocalSpace[shapeIndexB];\n"
+"							\n"
+"							b3Float4 aabbBMinWS;\n"
+"							b3Float4 aabbBMaxWS;\n"
+"							\n"
+"							b3TransformAabb2(aabbBlocal.m_minVec,aabbBlocal.m_maxVec,margin,\n"
+"								newPosB,\n"
+"								newOrnB,\n"
+"								&aabbBMinWS,&aabbBMaxWS);\n"
+"								\n"
+"								\n"
+"							\n"
+"							bool aabbOverlap = b3TestAabbAgainstAabb(aabbAMinWS,aabbAMaxWS,aabbBMinWS,aabbBMaxWS);\n"
+"							if (aabbOverlap)\n"
+"							{\n"
+"								int numFacesA = convexShapes[shapeIndexA].m_numFaces;\n"
+"								float dmin = FLT_MAX;\n"
+"								float4 posA = newPosA;\n"
+"								posA.w = 0.f;\n"
+"								float4 posB = newPosB;\n"
+"								posB.w = 0.f;\n"
+"								float4 c0local = convexShapes[shapeIndexA].m_localCenter;\n"
+"								float4 ornA = newOrnA;\n"
+"								float4 c0 = transform(&c0local, &posA, &ornA);\n"
+"								float4 c1local = convexShapes[shapeIndexB].m_localCenter;\n"
+"								float4 ornB =newOrnB;\n"
+"								float4 c1 = transform(&c1local,&posB,&ornB);\n"
+"								const float4 DeltaC2 = c0 - c1;\n"
+"								{//\n"
+"									int compoundPairIdx = atomic_inc(numCompoundPairsOut);\n"
+"									if (compoundPairIdx<maxNumCompoundPairsCapacity)\n"
+"									{\n"
+"										gpuCompoundPairsOut[compoundPairIdx]  = (int4)(bodyIndexA,bodyIndexB,childShapeIndexA,childShapeIndexB);\n"
+"									}\n"
+"								}//\n"
+"							}//fi(1)\n"
+"						} //for (int b=0\n"
+"					}//if (collidables[collidableIndexB].\n"
+"					else//if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)\n"
+"					{\n"
+"						if (1)\n"
+"						{\n"
+"							int numFacesA = convexShapes[shapeIndexA].m_numFaces;\n"
+"							float dmin = FLT_MAX;\n"
+"							float4 posA = newPosA;\n"
+"							posA.w = 0.f;\n"
+"							float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"							posB.w = 0.f;\n"
+"							float4 c0local = convexShapes[shapeIndexA].m_localCenter;\n"
+"							float4 ornA = newOrnA;\n"
+"							float4 c0 = transform(&c0local, &posA, &ornA);\n"
+"							float4 c1local = convexShapes[shapeIndexB].m_localCenter;\n"
+"							float4 ornB = rigidBodies[bodyIndexB].m_quat;\n"
+"							float4 c1 = transform(&c1local,&posB,&ornB);\n"
+"							const float4 DeltaC2 = c0 - c1;\n"
+"							{\n"
+"								int compoundPairIdx = atomic_inc(numCompoundPairsOut);\n"
+"								if (compoundPairIdx<maxNumCompoundPairsCapacity)\n"
+"								{\n"
+"									gpuCompoundPairsOut[compoundPairIdx] = (int4)(bodyIndexA,bodyIndexB,childShapeIndexA,-1);\n"
+"								}//if (compoundPairIdx<maxNumCompoundPairsCapacity)\n"
+"							}//\n"
+"						}//fi (1)\n"
+"					}//if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)\n"
+"				}//for (int b=0;b<numChildrenB;b++)	\n"
+"				return;\n"
+"			}//if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)\n"
+"			if ((collidables[collidableIndexA].m_shapeType!=SHAPE_CONCAVE_TRIMESH) \n"
+"				&& (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS))\n"
+"			{\n"
+"				int numChildrenB = collidables[collidableIndexB].m_numChildShapes;\n"
+"				for (int b=0;b<numChildrenB;b++)\n"
+"				{\n"
+"					int childShapeIndexB = collidables[collidableIndexB].m_shapeIndex+b;\n"
+"					int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;\n"
+"					float4 ornB = rigidBodies[bodyIndexB].m_quat;\n"
+"					float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"					float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;\n"
+"					float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;\n"
+"					float4 newPosB = qtRotate(ornB,childPosB)+posB;\n"
+"					float4 newOrnB = qtMul(ornB,childOrnB);\n"
+"					int shapeIndexB = collidables[childColIndexB].m_shapeIndex;\n"
+"					//////////////////////////////////////\n"
+"					if (1)\n"
+"					{\n"
+"						int numFacesA = convexShapes[shapeIndexA].m_numFaces;\n"
+"						float dmin = FLT_MAX;\n"
+"						float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"						posA.w = 0.f;\n"
+"						float4 posB = newPosB;\n"
+"						posB.w = 0.f;\n"
+"						float4 c0local = convexShapes[shapeIndexA].m_localCenter;\n"
+"						float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"						float4 c0 = transform(&c0local, &posA, &ornA);\n"
+"						float4 c1local = convexShapes[shapeIndexB].m_localCenter;\n"
+"						float4 ornB =newOrnB;\n"
+"						float4 c1 = transform(&c1local,&posB,&ornB);\n"
+"						const float4 DeltaC2 = c0 - c1;\n"
+"						{//\n"
+"							int compoundPairIdx = atomic_inc(numCompoundPairsOut);\n"
+"							if (compoundPairIdx<maxNumCompoundPairsCapacity)\n"
+"							{\n"
+"								gpuCompoundPairsOut[compoundPairIdx] = (int4)(bodyIndexA,bodyIndexB,-1,childShapeIndexB);\n"
+"							}//fi (compoundPairIdx<maxNumCompoundPairsCapacity)\n"
+"						}//\n"
+"					}//fi (1)	\n"
+"				}//for (int b=0;b<numChildrenB;b++)\n"
+"				return;\n"
+"			}//if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)\n"
+"			return;\n"
+"		}//fi ((collidables[collidableIndexA].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS) ||(collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS))\n"
+"	}//i<numPairs\n"
+"}\n"
+"// work-in-progress\n"
+"__kernel void   findSeparatingAxisKernel( __global const int4* pairs, \n"
+"																					__global const BodyData* rigidBodies, \n"
+"																					__global const btCollidableGpu* collidables,\n"
+"																					__global const ConvexPolyhedronCL* convexShapes, \n"
+"																					__global const float4* vertices,\n"
+"																					__global const float4* uniqueEdges,\n"
+"																					__global const btGpuFace* faces,\n"
+"																					__global const int* indices,\n"
+"																					__global btAabbCL* aabbs,\n"
+"																					__global volatile float4* separatingNormals,\n"
+"																					__global volatile int* hasSeparatingAxis,\n"
+"																					int numPairs\n"
+"																					)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	\n"
+"	if (i<numPairs)\n"
+"	{\n"
+"	\n"
+"		int bodyIndexA = pairs[i].x;\n"
+"		int bodyIndexB = pairs[i].y;\n"
+"		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"	\n"
+"		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"		\n"
+"		\n"
+"		//once the broadphase avoids static-static pairs, we can remove this test\n"
+"		if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))\n"
+"		{\n"
+"			hasSeparatingAxis[i] = 0;\n"
+"			return;\n"
+"		}\n"
+"		\n"
+"		if ((collidables[collidableIndexA].m_shapeType!=SHAPE_CONVEX_HULL) ||(collidables[collidableIndexB].m_shapeType!=SHAPE_CONVEX_HULL))\n"
+"		{\n"
+"			hasSeparatingAxis[i] = 0;\n"
+"			return;\n"
+"		}\n"
+"			\n"
+"		if ((collidables[collidableIndexA].m_shapeType==SHAPE_CONCAVE_TRIMESH))\n"
+"		{\n"
+"			hasSeparatingAxis[i] = 0;\n"
+"			return;\n"
+"		}\n"
+"		int numFacesA = convexShapes[shapeIndexA].m_numFaces;\n"
+"		float dmin = FLT_MAX;\n"
+"		float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"		posA.w = 0.f;\n"
+"		float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"		posB.w = 0.f;\n"
+"		float4 c0local = convexShapes[shapeIndexA].m_localCenter;\n"
+"		float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"		float4 c0 = transform(&c0local, &posA, &ornA);\n"
+"		float4 c1local = convexShapes[shapeIndexB].m_localCenter;\n"
+"		float4 ornB =rigidBodies[bodyIndexB].m_quat;\n"
+"		float4 c1 = transform(&c1local,&posB,&ornB);\n"
+"		const float4 DeltaC2 = c0 - c1;\n"
+"		float4 sepNormal;\n"
+"		\n"
+"		bool sepA = findSeparatingAxis(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,\n"
+"																								posB,ornB,\n"
+"																								DeltaC2,\n"
+"																								vertices,uniqueEdges,faces,\n"
+"																								indices,&sepNormal,&dmin);\n"
+"		hasSeparatingAxis[i] = 4;\n"
+"		if (!sepA)\n"
+"		{\n"
+"			hasSeparatingAxis[i] = 0;\n"
+"		} else\n"
+"		{\n"
+"			bool sepB = findSeparatingAxis(	&convexShapes[shapeIndexB],&convexShapes[shapeIndexA],posB,ornB,\n"
+"																									posA,ornA,\n"
+"																									DeltaC2,\n"
+"																									vertices,uniqueEdges,faces,\n"
+"																									indices,&sepNormal,&dmin);\n"
+"			if (!sepB)\n"
+"			{\n"
+"				hasSeparatingAxis[i] = 0;\n"
+"			} else\n"
+"			{\n"
+"				bool sepEE = findSeparatingAxisEdgeEdge(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,\n"
+"																									posB,ornB,\n"
+"																									DeltaC2,\n"
+"																									vertices,uniqueEdges,faces,\n"
+"																									indices,&sepNormal,&dmin);\n"
+"				if (!sepEE)\n"
+"				{\n"
+"					hasSeparatingAxis[i] = 0;\n"
+"				} else\n"
+"				{\n"
+"					hasSeparatingAxis[i] = 1;\n"
+"					separatingNormals[i] = sepNormal;\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"		\n"
+"	}\n"
+"}\n"
+"__kernel void   findSeparatingAxisVertexFaceKernel( __global const int4* pairs, \n"
+"																					__global const BodyData* rigidBodies, \n"
+"																					__global const btCollidableGpu* collidables,\n"
+"																					__global const ConvexPolyhedronCL* convexShapes, \n"
+"																					__global const float4* vertices,\n"
+"																					__global const float4* uniqueEdges,\n"
+"																					__global const btGpuFace* faces,\n"
+"																					__global const int* indices,\n"
+"																					__global btAabbCL* aabbs,\n"
+"																					__global volatile float4* separatingNormals,\n"
+"																					__global volatile int* hasSeparatingAxis,\n"
+"																					__global  float* dmins,\n"
+"																					int numPairs\n"
+"																					)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	\n"
+"	if (i<numPairs)\n"
+"	{\n"
+"	\n"
+"		int bodyIndexA = pairs[i].x;\n"
+"		int bodyIndexB = pairs[i].y;\n"
+"		int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"		int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"	\n"
+"		int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"		int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"	\n"
+"		hasSeparatingAxis[i] = 0;	\n"
+"		\n"
+"		//once the broadphase avoids static-static pairs, we can remove this test\n"
+"		if ((rigidBodies[bodyIndexA].m_invMass==0) &&(rigidBodies[bodyIndexB].m_invMass==0))\n"
+"		{\n"
+"			return;\n"
+"		}\n"
+"		\n"
+"		if ((collidables[collidableIndexA].m_shapeType!=SHAPE_CONVEX_HULL) ||(collidables[collidableIndexB].m_shapeType!=SHAPE_CONVEX_HULL))\n"
+"		{\n"
+"			return;\n"
+"		}\n"
+"			\n"
+"		int numFacesA = convexShapes[shapeIndexA].m_numFaces;\n"
+"		float dmin = FLT_MAX;\n"
+"		dmins[i] = dmin;\n"
+"		\n"
+"		float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"		posA.w = 0.f;\n"
+"		float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"		posB.w = 0.f;\n"
+"		float4 c0local = convexShapes[shapeIndexA].m_localCenter;\n"
+"		float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"		float4 c0 = transform(&c0local, &posA, &ornA);\n"
+"		float4 c1local = convexShapes[shapeIndexB].m_localCenter;\n"
+"		float4 ornB =rigidBodies[bodyIndexB].m_quat;\n"
+"		float4 c1 = transform(&c1local,&posB,&ornB);\n"
+"		const float4 DeltaC2 = c0 - c1;\n"
+"		float4 sepNormal;\n"
+"		\n"
+"		bool sepA = findSeparatingAxis(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,\n"
+"																								posB,ornB,\n"
+"																								DeltaC2,\n"
+"																								vertices,uniqueEdges,faces,\n"
+"																								indices,&sepNormal,&dmin);\n"
+"		hasSeparatingAxis[i] = 4;\n"
+"		if (!sepA)\n"
+"		{\n"
+"			hasSeparatingAxis[i] = 0;\n"
+"		} else\n"
+"		{\n"
+"			bool sepB = findSeparatingAxis(	&convexShapes[shapeIndexB],&convexShapes[shapeIndexA],posB,ornB,\n"
+"																									posA,ornA,\n"
+"																									DeltaC2,\n"
+"																									vertices,uniqueEdges,faces,\n"
+"																									indices,&sepNormal,&dmin);\n"
+"			if (sepB)\n"
+"			{\n"
+"				dmins[i] = dmin;\n"
+"				hasSeparatingAxis[i] = 1;\n"
+"				separatingNormals[i] = sepNormal;\n"
+"			}\n"
+"		}\n"
+"		\n"
+"	}\n"
+"}\n"
+"__kernel void   findSeparatingAxisEdgeEdgeKernel( __global const int4* pairs, \n"
+"																					__global const BodyData* rigidBodies, \n"
+"																					__global const btCollidableGpu* collidables,\n"
+"																					__global const ConvexPolyhedronCL* convexShapes, \n"
+"																					__global const float4* vertices,\n"
+"																					__global const float4* uniqueEdges,\n"
+"																					__global const btGpuFace* faces,\n"
+"																					__global const int* indices,\n"
+"																					__global btAabbCL* aabbs,\n"
+"																					__global  float4* separatingNormals,\n"
+"																					__global  int* hasSeparatingAxis,\n"
+"																					__global  float* dmins,\n"
+"																					__global const float4* unitSphereDirections,\n"
+"																					int numUnitSphereDirections,\n"
+"																					int numPairs\n"
+"																					)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	\n"
+"	if (i<numPairs)\n"
+"	{\n"
+"		if (hasSeparatingAxis[i])\n"
+"		{\n"
+"	\n"
+"			int bodyIndexA = pairs[i].x;\n"
+"			int bodyIndexB = pairs[i].y;\n"
+"	\n"
+"			int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"			int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"		\n"
+"			int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"			int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"			\n"
+"			\n"
+"			int numFacesA = convexShapes[shapeIndexA].m_numFaces;\n"
+"	\n"
+"			float dmin = dmins[i];\n"
+"	\n"
+"			float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"			posA.w = 0.f;\n"
+"			float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"			posB.w = 0.f;\n"
+"			float4 c0local = convexShapes[shapeIndexA].m_localCenter;\n"
+"			float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"			float4 c0 = transform(&c0local, &posA, &ornA);\n"
+"			float4 c1local = convexShapes[shapeIndexB].m_localCenter;\n"
+"			float4 ornB =rigidBodies[bodyIndexB].m_quat;\n"
+"			float4 c1 = transform(&c1local,&posB,&ornB);\n"
+"			const float4 DeltaC2 = c0 - c1;\n"
+"			float4 sepNormal = separatingNormals[i];\n"
+"			\n"
+"			\n"
+"			\n"
+"			bool sepEE = false;\n"
+"			int numEdgeEdgeDirections = convexShapes[shapeIndexA].m_numUniqueEdges*convexShapes[shapeIndexB].m_numUniqueEdges;\n"
+"			if (numEdgeEdgeDirections<=numUnitSphereDirections)\n"
+"			{\n"
+"				sepEE = findSeparatingAxisEdgeEdge(	&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,\n"
+"																									posB,ornB,\n"
+"																									DeltaC2,\n"
+"																									vertices,uniqueEdges,faces,\n"
+"																									indices,&sepNormal,&dmin);\n"
+"																									\n"
+"					if (!sepEE)\n"
+"					{\n"
+"						hasSeparatingAxis[i] = 0;\n"
+"					} else\n"
+"					{\n"
+"						hasSeparatingAxis[i] = 1;\n"
+"						separatingNormals[i] = sepNormal;\n"
+"					}\n"
+"			}\n"
+"			/*\n"
+"			///else case is a separate kernel, to make Mac OSX OpenCL compiler happy\n"
+"			else\n"
+"			{\n"
+"				sepEE = findSeparatingAxisUnitSphere(&convexShapes[shapeIndexA], &convexShapes[shapeIndexB],posA,ornA,\n"
+"																									posB,ornB,\n"
+"																									DeltaC2,\n"
+"																									vertices,unitSphereDirections,numUnitSphereDirections,\n"
+"																									&sepNormal,&dmin);\n"
+"					if (!sepEE)\n"
+"					{\n"
+"						hasSeparatingAxis[i] = 0;\n"
+"					} else\n"
+"					{\n"
+"						hasSeparatingAxis[i] = 1;\n"
+"						separatingNormals[i] = sepNormal;\n"
+"					}\n"
+"			}\n"
+"			*/\n"
+"		}		//if (hasSeparatingAxis[i])\n"
+"	}//(i<numPairs)\n"
+"}\n"
+"inline int	findClippingFaces(const float4 separatingNormal,\n"
+"                      const ConvexPolyhedronCL* hullA, \n"
+"					  __global const ConvexPolyhedronCL* hullB,\n"
+"                      const float4 posA, const Quaternion ornA,const float4 posB, const Quaternion ornB,\n"
+"                       __global float4* worldVertsA1,\n"
+"                      __global float4* worldNormalsA1,\n"
+"                      __global float4* worldVertsB1,\n"
+"                      int capacityWorldVerts,\n"
+"                      const float minDist, float maxDist,\n"
+"					  const float4* verticesA,\n"
+"                      const btGpuFace* facesA,\n"
+"                      const int* indicesA,\n"
+"					  __global const float4* verticesB,\n"
+"                      __global const btGpuFace* facesB,\n"
+"                      __global const int* indicesB,\n"
+"                      __global int4* clippingFaces, int pairIndex)\n"
+"{\n"
+"	int numContactsOut = 0;\n"
+"	int numWorldVertsB1= 0;\n"
+"    \n"
+"    \n"
+"	int closestFaceB=0;\n"
+"	float dmax = -FLT_MAX;\n"
+"    \n"
+"	{\n"
+"		for(int face=0;face<hullB->m_numFaces;face++)\n"
+"		{\n"
+"			const float4 Normal = make_float4(facesB[hullB->m_faceOffset+face].m_plane.x,\n"
+"                                              facesB[hullB->m_faceOffset+face].m_plane.y, facesB[hullB->m_faceOffset+face].m_plane.z,0.f);\n"
+"			const float4 WorldNormal = qtRotate(ornB, Normal);\n"
+"			float d = dot3F4(WorldNormal,separatingNormal);\n"
+"			if (d > dmax)\n"
+"			{\n"
+"				dmax = d;\n"
+"				closestFaceB = face;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"    \n"
+"	{\n"
+"		const btGpuFace polyB = facesB[hullB->m_faceOffset+closestFaceB];\n"
+"		int numVertices = polyB.m_numIndices;\n"
+"        if (numVertices>capacityWorldVerts)\n"
+"            numVertices = capacityWorldVerts;\n"
+"        \n"
+"		for(int e0=0;e0<numVertices;e0++)\n"
+"		{\n"
+"            if (e0<capacityWorldVerts)\n"
+"            {\n"
+"                const float4 b = verticesB[hullB->m_vertexOffset+indicesB[polyB.m_indexOffset+e0]];\n"
+"                worldVertsB1[pairIndex*capacityWorldVerts+numWorldVertsB1++] = transform(&b,&posB,&ornB);\n"
+"            }\n"
+"		}\n"
+"	}\n"
+"    \n"
+"    int closestFaceA=0;\n"
+"	{\n"
+"		float dmin = FLT_MAX;\n"
+"		for(int face=0;face<hullA->m_numFaces;face++)\n"
+"		{\n"
+"			const float4 Normal = make_float4(\n"
+"                                              facesA[hullA->m_faceOffset+face].m_plane.x,\n"
+"                                              facesA[hullA->m_faceOffset+face].m_plane.y,\n"
+"                                              facesA[hullA->m_faceOffset+face].m_plane.z,\n"
+"                                              0.f);\n"
+"			const float4 faceANormalWS = qtRotate(ornA,Normal);\n"
+"            \n"
+"			float d = dot3F4(faceANormalWS,separatingNormal);\n"
+"			if (d < dmin)\n"
+"			{\n"
+"				dmin = d;\n"
+"				closestFaceA = face;\n"
+"                worldNormalsA1[pairIndex] = faceANormalWS;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"    \n"
+"    int numVerticesA = facesA[hullA->m_faceOffset+closestFaceA].m_numIndices;\n"
+"    if (numVerticesA>capacityWorldVerts)\n"
+"       numVerticesA = capacityWorldVerts;\n"
+"    \n"
+"	for(int e0=0;e0<numVerticesA;e0++)\n"
+"	{\n"
+"        if (e0<capacityWorldVerts)\n"
+"        {\n"
+"            const float4 a = verticesA[hullA->m_vertexOffset+indicesA[facesA[hullA->m_faceOffset+closestFaceA].m_indexOffset+e0]];\n"
+"            worldVertsA1[pairIndex*capacityWorldVerts+e0] = transform(&a, &posA,&ornA);\n"
+"        }\n"
+"    }\n"
+"    \n"
+"    clippingFaces[pairIndex].x = closestFaceA;\n"
+"    clippingFaces[pairIndex].y = closestFaceB;\n"
+"    clippingFaces[pairIndex].z = numVerticesA;\n"
+"    clippingFaces[pairIndex].w = numWorldVertsB1;\n"
+"    \n"
+"    \n"
+"	return numContactsOut;\n"
+"}\n"
+"// work-in-progress\n"
+"__kernel void   findConcaveSeparatingAxisKernel( __global int4* concavePairs,\n"
+"																					__global const BodyData* rigidBodies,\n"
+"																					__global const btCollidableGpu* collidables,\n"
+"																					__global const ConvexPolyhedronCL* convexShapes, \n"
+"																					__global const float4* vertices,\n"
+"																					__global const float4* uniqueEdges,\n"
+"																					__global const btGpuFace* faces,\n"
+"																					__global const int* indices,\n"
+"																					__global const btGpuChildShape* gpuChildShapes,\n"
+"																					__global btAabbCL* aabbs,\n"
+"																					__global float4* concaveSeparatingNormalsOut,\n"
+"																					__global int* concaveHasSeparatingNormals,\n"
+"																					__global int4* clippingFacesOut,\n"
+"																					__global float4* worldVertsA1GPU,\n"
+"																					__global float4*  worldNormalsAGPU,\n"
+"																					__global float4* worldVertsB1GPU,\n"
+"																					int vertexFaceCapacity,\n"
+"																					int numConcavePairs\n"
+"																					)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numConcavePairs)\n"
+"		return;\n"
+"	concaveHasSeparatingNormals[i] = 0;\n"
+"	int pairIdx = i;\n"
+"	int bodyIndexA = concavePairs[i].x;\n"
+"	int bodyIndexB = concavePairs[i].y;\n"
+"	int collidableIndexA = rigidBodies[bodyIndexA].m_collidableIdx;\n"
+"	int collidableIndexB = rigidBodies[bodyIndexB].m_collidableIdx;\n"
+"	int shapeIndexA = collidables[collidableIndexA].m_shapeIndex;\n"
+"	int shapeIndexB = collidables[collidableIndexB].m_shapeIndex;\n"
+"	if (collidables[collidableIndexB].m_shapeType!=SHAPE_CONVEX_HULL&&\n"
+"		collidables[collidableIndexB].m_shapeType!=SHAPE_COMPOUND_OF_CONVEX_HULLS)\n"
+"	{\n"
+"		concavePairs[pairIdx].w = -1;\n"
+"		return;\n"
+"	}\n"
+"	int numFacesA = convexShapes[shapeIndexA].m_numFaces;\n"
+"	int numActualConcaveConvexTests = 0;\n"
+"	\n"
+"	int f = concavePairs[i].z;\n"
+"	\n"
+"	bool overlap = false;\n"
+"	\n"
+"	ConvexPolyhedronCL convexPolyhedronA;\n"
+"	//add 3 vertices of the triangle\n"
+"	convexPolyhedronA.m_numVertices = 3;\n"
+"	convexPolyhedronA.m_vertexOffset = 0;\n"
+"	float4	localCenter = make_float4(0.f,0.f,0.f,0.f);\n"
+"	btGpuFace face = faces[convexShapes[shapeIndexA].m_faceOffset+f];\n"
+"	float4 triMinAabb, triMaxAabb;\n"
+"	btAabbCL triAabb;\n"
+"	triAabb.m_min = make_float4(1e30f,1e30f,1e30f,0.f);\n"
+"	triAabb.m_max = make_float4(-1e30f,-1e30f,-1e30f,0.f);\n"
+"	\n"
+"	float4 verticesA[3];\n"
+"	for (int i=0;i<3;i++)\n"
+"	{\n"
+"		int index = indices[face.m_indexOffset+i];\n"
+"		float4 vert = vertices[convexShapes[shapeIndexA].m_vertexOffset+index];\n"
+"		verticesA[i] = vert;\n"
+"		localCenter += vert;\n"
+"			\n"
+"		triAabb.m_min = min(triAabb.m_min,vert);		\n"
+"		triAabb.m_max = max(triAabb.m_max,vert);		\n"
+"	}\n"
+"	overlap = true;\n"
+"	overlap = (triAabb.m_min.x > aabbs[bodyIndexB].m_max.x || triAabb.m_max.x < aabbs[bodyIndexB].m_min.x) ? false : overlap;\n"
+"	overlap = (triAabb.m_min.z > aabbs[bodyIndexB].m_max.z || triAabb.m_max.z < aabbs[bodyIndexB].m_min.z) ? false : overlap;\n"
+"	overlap = (triAabb.m_min.y > aabbs[bodyIndexB].m_max.y || triAabb.m_max.y < aabbs[bodyIndexB].m_min.y) ? false : overlap;\n"
+"		\n"
+"	if (overlap)\n"
+"	{\n"
+"		float dmin = FLT_MAX;\n"
+"		int hasSeparatingAxis=5;\n"
+"		float4 sepAxis=make_float4(1,2,3,4);\n"
+"		int localCC=0;\n"
+"		numActualConcaveConvexTests++;\n"
+"		//a triangle has 3 unique edges\n"
+"		convexPolyhedronA.m_numUniqueEdges = 3;\n"
+"		convexPolyhedronA.m_uniqueEdgesOffset = 0;\n"
+"		float4 uniqueEdgesA[3];\n"
+"		\n"
+"		uniqueEdgesA[0] = (verticesA[1]-verticesA[0]);\n"
+"		uniqueEdgesA[1] = (verticesA[2]-verticesA[1]);\n"
+"		uniqueEdgesA[2] = (verticesA[0]-verticesA[2]);\n"
+"		convexPolyhedronA.m_faceOffset = 0;\n"
+"                                  \n"
+"		float4 normal = make_float4(face.m_plane.x,face.m_plane.y,face.m_plane.z,0.f);\n"
+"                             \n"
+"		btGpuFace facesA[TRIANGLE_NUM_CONVEX_FACES];\n"
+"		int indicesA[3+3+2+2+2];\n"
+"		int curUsedIndices=0;\n"
+"		int fidx=0;\n"
+"		//front size of triangle\n"
+"		{\n"
+"			facesA[fidx].m_indexOffset=curUsedIndices;\n"
+"			indicesA[0] = 0;\n"
+"			indicesA[1] = 1;\n"
+"			indicesA[2] = 2;\n"
+"			curUsedIndices+=3;\n"
+"			float c = face.m_plane.w;\n"
+"			facesA[fidx].m_plane.x = normal.x;\n"
+"			facesA[fidx].m_plane.y = normal.y;\n"
+"			facesA[fidx].m_plane.z = normal.z;\n"
+"			facesA[fidx].m_plane.w = c;\n"
+"			facesA[fidx].m_numIndices=3;\n"
+"		}\n"
+"		fidx++;\n"
+"		//back size of triangle\n"
+"		{\n"
+"			facesA[fidx].m_indexOffset=curUsedIndices;\n"
+"			indicesA[3]=2;\n"
+"			indicesA[4]=1;\n"
+"			indicesA[5]=0;\n"
+"			curUsedIndices+=3;\n"
+"			float c = dot(normal,verticesA[0]);\n"
+"			float c1 = -face.m_plane.w;\n"
+"			facesA[fidx].m_plane.x = -normal.x;\n"
+"			facesA[fidx].m_plane.y = -normal.y;\n"
+"			facesA[fidx].m_plane.z = -normal.z;\n"
+"			facesA[fidx].m_plane.w = c;\n"
+"			facesA[fidx].m_numIndices=3;\n"
+"		}\n"
+"		fidx++;\n"
+"		bool addEdgePlanes = true;\n"
+"		if (addEdgePlanes)\n"
+"		{\n"
+"			int numVertices=3;\n"
+"			int prevVertex = numVertices-1;\n"
+"			for (int i=0;i<numVertices;i++)\n"
+"			{\n"
+"				float4 v0 = verticesA[i];\n"
+"				float4 v1 = verticesA[prevVertex];\n"
+"                                            \n"
+"				float4 edgeNormal = normalize(cross(normal,v1-v0));\n"
+"				float c = -dot(edgeNormal,v0);\n"
+"				facesA[fidx].m_numIndices = 2;\n"
+"				facesA[fidx].m_indexOffset=curUsedIndices;\n"
+"				indicesA[curUsedIndices++]=i;\n"
+"				indicesA[curUsedIndices++]=prevVertex;\n"
+"                                            \n"
+"				facesA[fidx].m_plane.x = edgeNormal.x;\n"
+"				facesA[fidx].m_plane.y = edgeNormal.y;\n"
+"				facesA[fidx].m_plane.z = edgeNormal.z;\n"
+"				facesA[fidx].m_plane.w = c;\n"
+"				fidx++;\n"
+"				prevVertex = i;\n"
+"			}\n"
+"		}\n"
+"		convexPolyhedronA.m_numFaces = TRIANGLE_NUM_CONVEX_FACES;\n"
+"		convexPolyhedronA.m_localCenter = localCenter*(1.f/3.f);\n"
+"		float4 posA = rigidBodies[bodyIndexA].m_pos;\n"
+"		posA.w = 0.f;\n"
+"		float4 posB = rigidBodies[bodyIndexB].m_pos;\n"
+"		posB.w = 0.f;\n"
+"		float4 ornA = rigidBodies[bodyIndexA].m_quat;\n"
+"		float4 ornB =rigidBodies[bodyIndexB].m_quat;\n"
+"		\n"
+"		///////////////////\n"
+"		///compound shape support\n"
+"		if (collidables[collidableIndexB].m_shapeType==SHAPE_COMPOUND_OF_CONVEX_HULLS)\n"
+"		{\n"
+"			int compoundChild = concavePairs[pairIdx].w;\n"
+"			int childShapeIndexB = compoundChild;//collidables[collidableIndexB].m_shapeIndex+compoundChild;\n"
+"			int childColIndexB = gpuChildShapes[childShapeIndexB].m_shapeIndex;\n"
+"			float4 childPosB = gpuChildShapes[childShapeIndexB].m_childPosition;\n"
+"			float4 childOrnB = gpuChildShapes[childShapeIndexB].m_childOrientation;\n"
+"			float4 newPosB = transform(&childPosB,&posB,&ornB);\n"
+"			float4 newOrnB = qtMul(ornB,childOrnB);\n"
+"			posB = newPosB;\n"
+"			ornB = newOrnB;\n"
+"			shapeIndexB = collidables[childColIndexB].m_shapeIndex;\n"
+"		}\n"
+"		//////////////////\n"
+"		float4 c0local = convexPolyhedronA.m_localCenter;\n"
+"		float4 c0 = transform(&c0local, &posA, &ornA);\n"
+"		float4 c1local = convexShapes[shapeIndexB].m_localCenter;\n"
+"		float4 c1 = transform(&c1local,&posB,&ornB);\n"
+"		const float4 DeltaC2 = c0 - c1;\n"
+"		bool sepA = findSeparatingAxisLocalA(	&convexPolyhedronA, &convexShapes[shapeIndexB],\n"
+"												posA,ornA,\n"
+"												posB,ornB,\n"
+"												DeltaC2,\n"
+"												verticesA,uniqueEdgesA,facesA,indicesA,\n"
+"												vertices,uniqueEdges,faces,indices,\n"
+"												&sepAxis,&dmin);\n"
+"		hasSeparatingAxis = 4;\n"
+"		if (!sepA)\n"
+"		{\n"
+"			hasSeparatingAxis = 0;\n"
+"		} else\n"
+"		{\n"
+"			bool sepB = findSeparatingAxisLocalB(	&convexShapes[shapeIndexB],&convexPolyhedronA,\n"
+"												posB,ornB,\n"
+"												posA,ornA,\n"
+"												DeltaC2,\n"
+"												vertices,uniqueEdges,faces,indices,\n"
+"												verticesA,uniqueEdgesA,facesA,indicesA,\n"
+"												&sepAxis,&dmin);\n"
+"			if (!sepB)\n"
+"			{\n"
+"				hasSeparatingAxis = 0;\n"
+"			} else\n"
+"			{\n"
+"				bool sepEE = findSeparatingAxisEdgeEdgeLocalA(	&convexPolyhedronA, &convexShapes[shapeIndexB],\n"
+"															posA,ornA,\n"
+"															posB,ornB,\n"
+"															DeltaC2,\n"
+"															verticesA,uniqueEdgesA,facesA,indicesA,\n"
+"															vertices,uniqueEdges,faces,indices,\n"
+"															&sepAxis,&dmin);\n"
+"	\n"
+"				if (!sepEE)\n"
+"				{\n"
+"					hasSeparatingAxis = 0;\n"
+"				} else\n"
+"				{\n"
+"					hasSeparatingAxis = 1;\n"
+"				}\n"
+"			}\n"
+"		}	\n"
+"		\n"
+"		if (hasSeparatingAxis)\n"
+"		{\n"
+"			sepAxis.w = dmin;\n"
+"			concaveSeparatingNormalsOut[pairIdx]=sepAxis;\n"
+"			concaveHasSeparatingNormals[i]=1;\n"
+"			float minDist = -1e30f;\n"
+"			float maxDist = 0.02f;\n"
+"		\n"
+"			findClippingFaces(sepAxis,\n"
+"                     &convexPolyhedronA,\n"
+"					 &convexShapes[shapeIndexB],\n"
+"					 posA,ornA,\n"
+"					 posB,ornB,\n"
+"                      worldVertsA1GPU,\n"
+"                      worldNormalsAGPU,\n"
+"                      worldVertsB1GPU,\n"
+"					  vertexFaceCapacity,\n"
+"                      minDist, maxDist,\n"
+"                      verticesA,\n"
+"                      facesA,\n"
+"                      indicesA,\n"
+" 					  vertices,\n"
+"                      faces,\n"
+"                      indices,\n"
+"                      clippingFacesOut, pairIdx);\n"
+"		} else\n"
+"		{	\n"
+"			//mark this pair as in-active\n"
+"			concavePairs[pairIdx].w = -1;\n"
+"		}\n"
+"	}\n"
+"	else\n"
+"	{	\n"
+"		//mark this pair as in-active\n"
+"		concavePairs[pairIdx].w = -1;\n"
+"	}\n"
+"	\n"
+"	concavePairs[pairIdx].z = -1;//now z is used for existing/persistent contacts\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3BoundSearchCL.cpp b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3BoundSearchCL.cpp
new file mode 100644
index 0000000000..a4980f71e1
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3BoundSearchCL.cpp
@@ -0,0 +1,213 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+//Host-code rewritten by Erwin Coumans
+
+#define BOUNDSEARCH_PATH "src/Bullet3OpenCL/ParallelPrimitives/kernels/BoundSearchKernels.cl"
+#define KERNEL0 "SearchSortDataLowerKernel"
+#define KERNEL1 "SearchSortDataUpperKernel"
+#define KERNEL2 "SubtractKernel"
+
+
+#include "b3BoundSearchCL.h"
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "b3LauncherCL.h"
+#include "kernels/BoundSearchKernelsCL.h"
+
+b3BoundSearchCL::b3BoundSearchCL(cl_context ctx, cl_device_id device, cl_command_queue queue, int maxSize)
+	:m_context(ctx),
+	m_device(device),
+	m_queue(queue)
+{
+
+	const char* additionalMacros = "";
+	//const char* srcFileNameForCaching="";
+
+	cl_int pErrNum;
+	const char* kernelSource = boundSearchKernelsCL;
+
+	cl_program boundSearchProg = b3OpenCLUtils::compileCLProgramFromString( ctx, device, kernelSource, &pErrNum,additionalMacros, BOUNDSEARCH_PATH);
+	b3Assert(boundSearchProg);
+
+	m_lowerSortDataKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "SearchSortDataLowerKernel", &pErrNum, boundSearchProg,additionalMacros );
+	b3Assert(m_lowerSortDataKernel );
+
+	m_upperSortDataKernel= b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "SearchSortDataUpperKernel", &pErrNum, boundSearchProg,additionalMacros );
+	b3Assert(m_upperSortDataKernel);
+
+	m_subtractKernel = 0;
+
+	if( maxSize )
+	{
+		m_subtractKernel= b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "SubtractKernel", &pErrNum, boundSearchProg,additionalMacros );
+		b3Assert(m_subtractKernel);
+	}
+
+	//m_constBuffer = new b3OpenCLArray<b3Int4>( device, 1, BufferBase::BUFFER_CONST );
+	
+	m_lower = (maxSize == 0)? 0: new b3OpenCLArray<unsigned int>(ctx,queue,maxSize );
+	m_upper = (maxSize == 0)? 0: new b3OpenCLArray<unsigned int>(ctx,queue, maxSize );
+
+	m_filler = new b3FillCL(ctx,device,queue);
+}
+
+b3BoundSearchCL::~b3BoundSearchCL()
+{
+	
+	delete m_lower;
+	delete m_upper;
+	delete m_filler;
+			
+	clReleaseKernel(m_lowerSortDataKernel);
+	clReleaseKernel(m_upperSortDataKernel);
+	clReleaseKernel(m_subtractKernel);
+	
+
+}
+
+
+void b3BoundSearchCL::execute(b3OpenCLArray<b3SortData>& src, int nSrc, b3OpenCLArray<unsigned int>& dst, int nDst, Option option )
+{
+	b3Int4 constBuffer;
+	constBuffer.x = nSrc;
+	constBuffer.y = nDst;
+
+	if( option == BOUND_LOWER )
+	{
+		b3BufferInfoCL bInfo[] = { b3BufferInfoCL( src.getBufferCL(), true ), b3BufferInfoCL( dst.getBufferCL()) };
+
+		b3LauncherCL launcher( m_queue, m_lowerSortDataKernel,"m_lowerSortDataKernel" );
+		launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst( nSrc );
+        launcher.setConst( nDst );
+        
+		launcher.launch1D( nSrc, 64 );
+	}
+	else if( option == BOUND_UPPER )
+	{
+		b3BufferInfoCL bInfo[] = { b3BufferInfoCL( src.getBufferCL(), true ), b3BufferInfoCL( dst.getBufferCL() ) };
+
+		b3LauncherCL launcher(m_queue, m_upperSortDataKernel,"m_upperSortDataKernel" );
+		launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+        launcher.setConst( nSrc );
+        launcher.setConst( nDst );
+
+		launcher.launch1D( nSrc, 64 );
+	}
+	else if( option == COUNT )
+	{
+		b3Assert( m_lower );
+		b3Assert( m_upper );
+		b3Assert( m_lower->capacity() <= (int)nDst );
+		b3Assert( m_upper->capacity() <= (int)nDst );
+
+		int zero = 0;
+		m_filler->execute( *m_lower, zero, nDst );
+		m_filler->execute( *m_upper, zero, nDst );
+
+		execute( src, nSrc, *m_lower, nDst, BOUND_LOWER );
+		execute( src, nSrc, *m_upper, nDst, BOUND_UPPER );
+
+		{
+			b3BufferInfoCL bInfo[] = { b3BufferInfoCL( m_upper->getBufferCL(), true ), b3BufferInfoCL( m_lower->getBufferCL(), true ), b3BufferInfoCL( dst.getBufferCL() ) };
+
+			b3LauncherCL  launcher( m_queue, m_subtractKernel ,"m_subtractKernel");
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+            launcher.setConst( nSrc );
+            launcher.setConst( nDst );
+
+			launcher.launch1D( nDst, 64 );
+		}
+	}
+	else
+	{
+		b3Assert( 0 );
+	}
+
+}
+
+
+void b3BoundSearchCL::executeHost( b3AlignedObjectArray<b3SortData>& src, int nSrc, 
+	b3AlignedObjectArray<unsigned int>& dst,  int nDst, Option option )
+{
+
+
+	for(int i=0; i<nSrc-1; i++) 
+		b3Assert( src[i].m_key <= src[i+1].m_key );
+
+	b3SortData minData,zeroData,maxData;
+	minData.m_key = -1;
+	minData.m_value = -1;
+	zeroData.m_key=0;
+	zeroData.m_value=0;
+	maxData.m_key = nDst;
+	maxData.m_value = nDst;
+
+	if( option == BOUND_LOWER )
+	{
+		for(int i=0; i<nSrc; i++)
+		{
+			b3SortData& iData = (i==0)? minData: src[i-1];
+			b3SortData& jData = (i==nSrc)? maxData: src[i];
+
+			if( iData.m_key != jData.m_key )
+			{
+				int k = jData.m_key;
+				{
+					dst[k] = i;
+				}
+			}
+		}
+	}
+	else if( option == BOUND_UPPER )
+	{
+		for(int i=1; i<nSrc+1; i++)
+		{
+			b3SortData& iData = src[i-1];
+			b3SortData& jData = (i==nSrc)? maxData: src[i];
+
+			if( iData.m_key != jData.m_key )
+			{
+				int k = iData.m_key;
+				{
+					dst[k] = i;
+				}
+			}
+		}
+	}
+	else if( option == COUNT )
+	{
+		b3AlignedObjectArray<unsigned int> lower;
+		lower.resize(nDst );
+		b3AlignedObjectArray<unsigned int> upper;
+		upper.resize(nDst );
+
+		for(int i=0; i<nDst; i++) 
+		{ 
+			lower[i] = upper[i] = 0; 
+		}
+
+		executeHost( src, nSrc, lower, nDst, BOUND_LOWER );
+		executeHost( src, nSrc, upper, nDst, BOUND_UPPER );
+
+		for( int i=0; i<nDst; i++) 
+		{ 
+			dst[i] = upper[i] - lower[i]; 
+		}
+	}
+	else
+	{
+		b3Assert( 0 );
+	}
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3BoundSearchCL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3BoundSearchCL.h
new file mode 100644
index 0000000000..7e2940965c
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3BoundSearchCL.h
@@ -0,0 +1,67 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+#ifndef B3_BOUNDSEARCH_H
+#define B3_BOUNDSEARCH_H
+
+#pragma once
+
+/*#include <Adl/Adl.h>
+#include <AdlPrimitives/Math/Math.h>
+#include <AdlPrimitives/Sort/SortData.h>
+#include <AdlPrimitives/Fill/Fill.h>
+*/
+
+#include "b3OpenCLArray.h"
+#include "b3FillCL.h"
+#include "b3RadixSort32CL.h" //for b3SortData (perhaps move it?)
+class b3BoundSearchCL
+{
+	public:
+
+		enum Option
+		{
+			BOUND_LOWER,
+			BOUND_UPPER,
+			COUNT,
+		};
+
+		cl_context m_context;
+		cl_device_id m_device;
+		cl_command_queue m_queue;
+
+		
+		cl_kernel m_lowerSortDataKernel;
+		cl_kernel m_upperSortDataKernel;
+		cl_kernel m_subtractKernel;
+		
+		b3OpenCLArray<b3Int4>* m_constbtOpenCLArray;
+		b3OpenCLArray<unsigned int>* m_lower;
+		b3OpenCLArray<unsigned int>* m_upper;
+		
+		b3FillCL* m_filler;
+		
+		b3BoundSearchCL(cl_context context, cl_device_id device, cl_command_queue queue, int size);
+
+		virtual ~b3BoundSearchCL();
+
+		//	src has to be src[i].m_key <= src[i+1].m_key
+		void execute( b3OpenCLArray<b3SortData>& src, int nSrc, b3OpenCLArray<unsigned int>& dst, int nDst, Option option = BOUND_LOWER );
+
+		void executeHost( b3AlignedObjectArray<b3SortData>& src, int nSrc, b3AlignedObjectArray<unsigned int>& dst, int nDst, Option option = BOUND_LOWER);
+};
+
+
+#endif //B3_BOUNDSEARCH_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3BufferInfoCL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3BufferInfoCL.h
new file mode 100644
index 0000000000..52f219ae3f
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3BufferInfoCL.h
@@ -0,0 +1,19 @@
+
+#ifndef B3_BUFFER_INFO_CL_H
+#define B3_BUFFER_INFO_CL_H
+
+#include "b3OpenCLArray.h"
+
+
+struct b3BufferInfoCL
+{
+	//b3BufferInfoCL(){}
+
+//	template<typename T>
+	b3BufferInfoCL(cl_mem buff, bool isReadOnly = false): m_clBuffer(buff), m_isReadOnly(isReadOnly){}
+
+	cl_mem m_clBuffer;
+	bool m_isReadOnly;
+};
+
+#endif //B3_BUFFER_INFO_CL_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3FillCL.cpp b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3FillCL.cpp
new file mode 100644
index 0000000000..f05c2648f1
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3FillCL.cpp
@@ -0,0 +1,126 @@
+#include "b3FillCL.h"
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "b3BufferInfoCL.h"
+#include "b3LauncherCL.h"
+
+#define FILL_CL_PROGRAM_PATH "src/Bullet3OpenCL/ParallelPrimitives/kernels/FillKernels.cl"
+
+#include "kernels/FillKernelsCL.h"
+
+b3FillCL::b3FillCL(cl_context ctx, cl_device_id device, cl_command_queue queue)
+:m_commandQueue(queue)
+{
+	const char* kernelSource = fillKernelsCL;
+	cl_int pErrNum;
+	const char* additionalMacros = "";
+
+	cl_program fillProg = b3OpenCLUtils::compileCLProgramFromString( ctx, device, kernelSource, &pErrNum,additionalMacros, FILL_CL_PROGRAM_PATH);
+	b3Assert(fillProg);
+
+	m_fillIntKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "FillIntKernel", &pErrNum, fillProg,additionalMacros );
+	b3Assert(m_fillIntKernel);
+
+	m_fillUnsignedIntKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "FillUnsignedIntKernel", &pErrNum, fillProg,additionalMacros );
+	b3Assert(m_fillIntKernel);
+
+	m_fillFloatKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "FillFloatKernel", &pErrNum, fillProg,additionalMacros );
+	b3Assert(m_fillFloatKernel);
+
+	
+
+	m_fillKernelInt2 = b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "FillInt2Kernel", &pErrNum, fillProg,additionalMacros );
+	b3Assert(m_fillKernelInt2);
+	
+}
+
+b3FillCL::~b3FillCL()
+{
+	clReleaseKernel(m_fillKernelInt2);
+	clReleaseKernel(m_fillIntKernel);
+	clReleaseKernel(m_fillUnsignedIntKernel);
+	clReleaseKernel(m_fillFloatKernel);
+
+}
+
+void b3FillCL::execute(b3OpenCLArray<float>& src, const float value, int n, int offset)
+{
+	b3Assert( n>0 );
+
+	{
+		b3LauncherCL launcher( m_commandQueue, m_fillFloatKernel,"m_fillFloatKernel" );
+		launcher.setBuffer( src.getBufferCL());
+		launcher.setConst( n );
+		launcher.setConst( value );
+		launcher.setConst( offset);
+
+		launcher.launch1D( n );
+	}
+}
+
+void b3FillCL::execute(b3OpenCLArray<int>& src, const int value, int n, int offset)
+{
+	b3Assert( n>0 );
+	
+
+	{
+		b3LauncherCL launcher( m_commandQueue, m_fillIntKernel ,"m_fillIntKernel");
+		launcher.setBuffer(src.getBufferCL());
+		launcher.setConst( n);
+		launcher.setConst( value);
+		launcher.setConst( offset);
+		launcher.launch1D( n );
+	}
+}
+
+
+void b3FillCL::execute(b3OpenCLArray<unsigned int>& src, const unsigned int value, int n, int offset)
+{
+	b3Assert( n>0 );
+
+	{
+		b3BufferInfoCL bInfo[] = { b3BufferInfoCL( src.getBufferCL() ) };
+
+		b3LauncherCL launcher( m_commandQueue, m_fillUnsignedIntKernel,"m_fillUnsignedIntKernel" );
+		launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst( n );
+        launcher.setConst(value);
+		launcher.setConst(offset);
+
+		launcher.launch1D( n );
+	}
+}
+
+void b3FillCL::executeHost(b3AlignedObjectArray<b3Int2> &src, const b3Int2 &value, int n, int offset)
+{
+	for (int i=0;i<n;i++)
+	{
+		src[i+offset]=value;
+	}
+}
+
+void b3FillCL::executeHost(b3AlignedObjectArray<int> &src, const int value, int n, int offset)
+{
+	for (int i=0;i<n;i++)
+	{
+		src[i+offset]=value;
+	}
+}
+
+void b3FillCL::execute(b3OpenCLArray<b3Int2> &src, const b3Int2 &value, int n, int offset)
+{
+	b3Assert( n>0 );
+	
+
+	{
+		b3BufferInfoCL bInfo[] = { b3BufferInfoCL( src.getBufferCL() ) };
+
+		b3LauncherCL launcher(m_commandQueue, m_fillKernelInt2,"m_fillKernelInt2");
+		launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(n);
+		launcher.setConst(value);
+		launcher.setConst(offset);
+
+		//( constBuffer );
+		launcher.launch1D( n );
+	}
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3FillCL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3FillCL.h
new file mode 100644
index 0000000000..1609676b9d
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3FillCL.h
@@ -0,0 +1,63 @@
+#ifndef B3_FILL_CL_H
+#define B3_FILL_CL_H
+
+#include "b3OpenCLArray.h"
+#include "Bullet3Common/b3Scalar.h"
+
+#include "Bullet3Common/shared/b3Int2.h"
+#include "Bullet3Common/shared/b3Int4.h"
+
+
+class b3FillCL
+{
+	
+	cl_command_queue	m_commandQueue;
+	
+	cl_kernel			m_fillKernelInt2;
+	cl_kernel			m_fillIntKernel;
+	cl_kernel			m_fillUnsignedIntKernel;
+	cl_kernel			m_fillFloatKernel;
+
+	public:
+		
+		struct b3ConstData
+		{
+			union
+			{
+				b3Int4 m_data;
+				b3UnsignedInt4 m_UnsignedData;
+			};
+			int m_offset;
+			int m_n;
+			int m_padding[2];
+		};
+
+protected:
+
+public:
+
+		b3FillCL(cl_context ctx, cl_device_id device, cl_command_queue queue);
+
+		virtual ~b3FillCL();
+
+		void execute(b3OpenCLArray<unsigned int>& src, const unsigned int value, int n, int offset = 0);
+	
+		void execute(b3OpenCLArray<int>& src, const int value, int n, int offset = 0);
+
+		void execute(b3OpenCLArray<float>& src, const float value, int n, int offset = 0);
+
+		void execute(b3OpenCLArray<b3Int2>& src, const b3Int2& value, int n, int offset = 0);
+
+		void executeHost(b3AlignedObjectArray<b3Int2> &src, const b3Int2 &value, int n, int offset);
+
+		void executeHost(b3AlignedObjectArray<int> &src, const int value, int n, int offset);
+
+	//	void execute(b3OpenCLArray<b3Int4>& src, const b3Int4& value, int n, int offset = 0);
+
+};
+		
+		
+		
+	
+
+#endif //B3_FILL_CL_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.cpp b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.cpp
new file mode 100644
index 0000000000..94590d11ca
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.cpp
@@ -0,0 +1,308 @@
+#include "b3LauncherCL.h"
+
+bool gDebugLauncherCL = false;
+    
+b3LauncherCL::b3LauncherCL(cl_command_queue queue, cl_kernel kernel, const char* name)
+:m_commandQueue(queue),
+m_kernel(kernel),
+m_idx(0),
+m_enableSerialization(false),
+m_name(name)
+{
+	if (gDebugLauncherCL)
+	{
+		static int counter = 0;
+		printf("[%d] Prepare to launch OpenCL kernel %s\n", counter++, name);
+	}
+
+      m_serializationSizeInBytes = sizeof(int);
+}
+    
+b3LauncherCL::~b3LauncherCL()
+  {
+      for (int i=0;i<m_arrays.size();i++)
+      {
+		  delete (m_arrays[i]);
+      }
+
+	  m_arrays.clear();
+	  if (gDebugLauncherCL)
+	  {
+		static int counter = 0;
+		printf("[%d] Finished launching OpenCL kernel %s\n", counter++,m_name);
+	  }
+  }
+
+void b3LauncherCL::setBuffer( cl_mem clBuffer)
+{
+		if (m_enableSerialization)
+		{
+			b3KernelArgData kernelArg;
+			kernelArg.m_argIndex = m_idx;
+			kernelArg.m_isBuffer = 1;
+			kernelArg.m_clBuffer = clBuffer;
+		
+			cl_mem_info param_name = CL_MEM_SIZE;
+			size_t param_value;
+			size_t sizeInBytes = sizeof(size_t);
+			size_t actualSizeInBytes;
+			cl_int err;
+			err = clGetMemObjectInfo (	kernelArg.m_clBuffer,
+									  param_name,
+									  sizeInBytes,
+									  &param_value,
+									  &actualSizeInBytes);
+			
+			b3Assert( err == CL_SUCCESS );
+			kernelArg.m_argSizeInBytes = param_value;
+			
+			m_kernelArguments.push_back(kernelArg);
+			m_serializationSizeInBytes+= sizeof(b3KernelArgData);
+			m_serializationSizeInBytes+=param_value;
+            }
+            cl_int status = clSetKernelArg( m_kernel, m_idx++, sizeof(cl_mem), &clBuffer);
+		b3Assert( status == CL_SUCCESS );
+}
+
+
+void b3LauncherCL::setBuffers( b3BufferInfoCL* buffInfo, int n )
+{
+	for(int i=0; i<n; i++)
+	{
+		if (m_enableSerialization)
+		{
+			b3KernelArgData kernelArg;
+			kernelArg.m_argIndex = m_idx;
+			kernelArg.m_isBuffer = 1;
+			kernelArg.m_clBuffer = buffInfo[i].m_clBuffer;
+		
+			cl_mem_info param_name = CL_MEM_SIZE;
+			size_t param_value;
+			size_t sizeInBytes = sizeof(size_t);
+			size_t actualSizeInBytes;
+			cl_int err;
+			err = clGetMemObjectInfo (	kernelArg.m_clBuffer,
+									  param_name,
+									  sizeInBytes,
+									  &param_value,
+									  &actualSizeInBytes);
+			
+			b3Assert( err == CL_SUCCESS );
+			kernelArg.m_argSizeInBytes = param_value;
+			
+			m_kernelArguments.push_back(kernelArg);
+			m_serializationSizeInBytes+= sizeof(b3KernelArgData);
+			m_serializationSizeInBytes+=param_value;
+            }
+            cl_int status = clSetKernelArg( m_kernel, m_idx++, sizeof(cl_mem), &buffInfo[i].m_clBuffer);
+		b3Assert( status == CL_SUCCESS );
+        }
+}
+
+struct b3KernelArgDataUnaligned
+{
+    int m_isBuffer;
+    int m_argIndex;
+    int m_argSizeInBytes;
+	int m_unusedPadding;
+    union
+    {
+        cl_mem m_clBuffer;
+        unsigned char m_argData[B3_CL_MAX_ARG_SIZE];
+    };
+    
+};
+#include <string.h>
+
+
+
+int b3LauncherCL::deserializeArgs(unsigned char* buf, int bufSize, cl_context ctx)
+{
+    int index=0;
+    
+    int numArguments = *(int*) &buf[index];
+    index+=sizeof(int);
+    
+    for (int i=0;i<numArguments;i++)
+    {
+        b3KernelArgDataUnaligned* arg = (b3KernelArgDataUnaligned*)&buf[index];
+
+        index+=sizeof(b3KernelArgData);
+        if (arg->m_isBuffer)
+        {
+            b3OpenCLArray<unsigned char>* clData = new b3OpenCLArray<unsigned char>(ctx,m_commandQueue, arg->m_argSizeInBytes);
+            clData->resize(arg->m_argSizeInBytes);
+            
+            clData->copyFromHostPointer(&buf[index], arg->m_argSizeInBytes);
+            
+            arg->m_clBuffer = clData->getBufferCL();
+            
+            m_arrays.push_back(clData);
+            
+            cl_int status = clSetKernelArg( m_kernel, m_idx++, sizeof(cl_mem), &arg->m_clBuffer);
+		b3Assert( status == CL_SUCCESS );
+            index+=arg->m_argSizeInBytes;
+        } else 
+        {
+            cl_int status = clSetKernelArg( m_kernel, m_idx++, arg->m_argSizeInBytes, &arg->m_argData);
+		b3Assert( status == CL_SUCCESS );
+        }
+		b3KernelArgData b;
+		memcpy(&b,arg,sizeof(b3KernelArgDataUnaligned));
+	m_kernelArguments.push_back(b);
+    }
+m_serializationSizeInBytes = index;
+    return index;
+}
+
+int b3LauncherCL::validateResults(unsigned char* goldBuffer, int goldBufferCapacity, cl_context ctx)
+  {
+	 int index=0;
+      
+      int numArguments = *(int*) &goldBuffer[index];
+      index+=sizeof(int);
+
+	if (numArguments != m_kernelArguments.size())
+	{
+		printf("failed validation: expected %d arguments, found %d\n",numArguments, m_kernelArguments.size());
+		return -1;
+	}
+      
+      for (int ii=0;ii<numArguments;ii++)
+      {
+          b3KernelArgData* argGold = (b3KernelArgData*)&goldBuffer[index];
+
+		if (m_kernelArguments[ii].m_argSizeInBytes != argGold->m_argSizeInBytes)
+		{
+			printf("failed validation: argument %d sizeInBytes expected: %d, found %d\n",ii, argGold->m_argSizeInBytes, m_kernelArguments[ii].m_argSizeInBytes);
+			return -2;
+		}
+
+		{
+			int expected = argGold->m_isBuffer;
+			int found = m_kernelArguments[ii].m_isBuffer;
+
+			if (expected != found)
+			{
+				printf("failed validation: argument %d isBuffer expected: %d, found %d\n",ii,expected, found);
+				return -3;
+			}
+		}
+		index+=sizeof(b3KernelArgData);
+
+		if (argGold->m_isBuffer)
+          {
+
+			unsigned char* memBuf= (unsigned char*) malloc(m_kernelArguments[ii].m_argSizeInBytes);
+			unsigned char* goldBuf = &goldBuffer[index];
+			for (int j=0;j<m_kernelArguments[j].m_argSizeInBytes;j++)
+			{
+				memBuf[j] = 0xaa;
+			}
+
+			cl_int status = 0;
+			status = clEnqueueReadBuffer( m_commandQueue, m_kernelArguments[ii].m_clBuffer, CL_TRUE, 0, m_kernelArguments[ii].m_argSizeInBytes,
+                                           memBuf, 0,0,0 );
+              b3Assert( status==CL_SUCCESS );
+              clFinish(m_commandQueue);
+
+			for (int b=0;b<m_kernelArguments[ii].m_argSizeInBytes;b++)
+			{
+				int expected = goldBuf[b];
+				int found = memBuf[b];
+				if (expected != found)
+				{
+					printf("failed validation: argument %d OpenCL data at byte position %d expected: %d, found %d\n",
+						ii, b, expected, found);
+					return -4;
+				}
+			}
+
+              
+              index+=argGold->m_argSizeInBytes;
+          } else 
+          {
+			
+			//compare content
+			for (int b=0;b<m_kernelArguments[ii].m_argSizeInBytes;b++)
+			{
+				int expected = argGold->m_argData[b];
+				int found =m_kernelArguments[ii].m_argData[b];
+				if (expected != found)
+				{
+					printf("failed validation: argument %d const data at byte position %d expected: %d, found %d\n",
+						ii, b, expected, found);
+					return -5;
+				}
+			}
+
+          }
+      }
+      return index;
+
+}
+
+int b3LauncherCL::serializeArguments(unsigned char* destBuffer, int destBufferCapacity)
+{
+//initialize to known values
+for (int i=0;i<destBufferCapacity;i++)
+	destBuffer[i] = 0xec;
+
+    assert(destBufferCapacity>=m_serializationSizeInBytes);
+    
+    //todo: use the b3Serializer for this to allow for 32/64bit, endianness etc        
+    int numArguments = m_kernelArguments.size();
+    int curBufferSize = 0;
+    int* dest = (int*)&destBuffer[curBufferSize];
+    *dest = numArguments;
+    curBufferSize += sizeof(int);
+    
+    
+    
+    for (int i=0;i<this->m_kernelArguments.size();i++)
+    {
+        b3KernelArgData* arg = (b3KernelArgData*) &destBuffer[curBufferSize];
+        *arg = m_kernelArguments[i];
+        curBufferSize+=sizeof(b3KernelArgData);
+        if (arg->m_isBuffer==1)
+        {
+            //copy the OpenCL buffer content
+            cl_int status = 0;
+            status = clEnqueueReadBuffer( m_commandQueue, arg->m_clBuffer, 0, 0, arg->m_argSizeInBytes,
+                                         &destBuffer[curBufferSize], 0,0,0 );
+            b3Assert( status==CL_SUCCESS );
+            clFinish(m_commandQueue);
+            curBufferSize+=arg->m_argSizeInBytes;
+        }
+        
+    }
+    return curBufferSize;
+}
+
+void b3LauncherCL::serializeToFile(const char* fileName, int numWorkItems)
+{
+	int num = numWorkItems;
+	int buffSize = getSerializationBufferSize();
+	unsigned char* buf = new unsigned char[buffSize+sizeof(int)];
+	for (int i=0;i<buffSize+1;i++)
+	{
+		unsigned char* ptr = (unsigned char*)&buf[i];
+		*ptr = 0xff;
+	}
+//	int actualWrite = serializeArguments(buf,buffSize);
+              
+//	unsigned char* cptr = (unsigned char*)&buf[buffSize];
+//            printf("buf[buffSize] = %d\n",*cptr);
+              
+	assert(buf[buffSize]==0xff);//check for buffer overrun
+	int* ptr = (int*)&buf[buffSize];
+              
+	*ptr = num;
+              
+	FILE* f = fopen(fileName,"wb");
+	fwrite(buf,buffSize+sizeof(int),1,f);
+	fclose(f);
+
+	delete[] buf;
+}		
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h
new file mode 100644
index 0000000000..1b267b31ef
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h
@@ -0,0 +1,135 @@
+
+#ifndef B3_LAUNCHER_CL_H
+#define B3_LAUNCHER_CL_H
+
+#include "b3BufferInfoCL.h"
+#include "Bullet3Common/b3MinMax.h"
+#include "b3OpenCLArray.h"
+#include <stdio.h>
+
+#define B3_DEBUG_SERIALIZE_CL
+
+
+#ifdef _WIN32
+#pragma warning(disable :4996)
+#endif
+#define B3_CL_MAX_ARG_SIZE 16
+B3_ATTRIBUTE_ALIGNED16(struct) b3KernelArgData
+{
+    int m_isBuffer;
+    int m_argIndex;
+    int m_argSizeInBytes;
+	int m_unusedPadding;
+    union
+    {
+        cl_mem m_clBuffer;
+        unsigned char m_argData[B3_CL_MAX_ARG_SIZE];
+    };
+    
+};
+
+class b3LauncherCL
+{
+
+	cl_command_queue m_commandQueue;
+	cl_kernel m_kernel;
+	int m_idx;
+	
+    b3AlignedObjectArray<b3KernelArgData> m_kernelArguments;
+    int m_serializationSizeInBytes;
+	bool	m_enableSerialization;
+
+	const char* m_name;
+	public:
+
+     b3AlignedObjectArray<b3OpenCLArray<unsigned char>* > m_arrays;
+    
+		b3LauncherCL(cl_command_queue queue, cl_kernel kernel, const char* name);
+    
+    virtual ~b3LauncherCL();
+    
+		void setBuffer( cl_mem clBuffer);
+
+		void setBuffers( b3BufferInfoCL* buffInfo, int n );
+    
+    int getSerializationBufferSize() const 
+    {
+        return m_serializationSizeInBytes;
+    }
+    
+    int deserializeArgs(unsigned char* buf, int bufSize, cl_context ctx);
+
+	inline int validateResults(unsigned char* goldBuffer, int goldBufferCapacity, cl_context ctx);
+
+    int serializeArguments(unsigned char* destBuffer, int destBufferCapacity);
+    
+	int getNumArguments() const
+	{
+		return m_kernelArguments.size();
+	}
+
+	b3KernelArgData getArgument(int index)
+	{
+		return m_kernelArguments[index];
+	}
+
+	void serializeToFile(const char* fileName, int numWorkItems);
+
+	template<typename T>
+		inline void setConst( const T& consts )
+		{
+			int sz=sizeof(T);
+			b3Assert(sz<=B3_CL_MAX_ARG_SIZE);
+
+			if (m_enableSerialization)
+			{
+				b3KernelArgData kernelArg;
+				kernelArg.m_argIndex = m_idx;
+				kernelArg.m_isBuffer = 0;
+				T* destArg = (T*)kernelArg.m_argData;
+				*destArg = consts;
+				kernelArg.m_argSizeInBytes = sizeof(T);
+				m_kernelArguments.push_back(kernelArg);
+				m_serializationSizeInBytes+=sizeof(b3KernelArgData);
+			}
+            
+			cl_int status = clSetKernelArg( m_kernel, m_idx++, sz, &consts );
+			b3Assert( status == CL_SUCCESS );
+		}
+
+		inline void launch1D( int numThreads, int localSize = 64)
+		{
+			launch2D( numThreads, 1, localSize, 1 );
+		}
+
+		inline void launch2D( int numThreadsX, int numThreadsY, int localSizeX, int localSizeY )
+		{
+			size_t gRange[3] = {1,1,1};
+			size_t lRange[3] = {1,1,1};
+			lRange[0] = localSizeX;
+			lRange[1] = localSizeY;
+			gRange[0] = b3Max((size_t)1, (numThreadsX/lRange[0])+(!(numThreadsX%lRange[0])?0:1));
+			gRange[0] *= lRange[0];
+			gRange[1] = b3Max((size_t)1, (numThreadsY/lRange[1])+(!(numThreadsY%lRange[1])?0:1));
+			gRange[1] *= lRange[1];
+
+			cl_int status = clEnqueueNDRangeKernel( m_commandQueue, 
+				m_kernel, 2, NULL, gRange, lRange, 0,0,0 );
+            if (status != CL_SUCCESS)
+            {
+                printf("Error: OpenCL status = %d\n",status);
+            }
+			b3Assert( status == CL_SUCCESS );
+
+		}
+	
+		void	enableSerialization(bool serialize)
+		{
+			m_enableSerialization = serialize;
+		}
+		
+};
+
+
+
+#endif //B3_LAUNCHER_CL_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h
new file mode 100644
index 0000000000..d70c30f53f
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h
@@ -0,0 +1,306 @@
+#ifndef B3_OPENCL_ARRAY_H
+#define B3_OPENCL_ARRAY_H
+
+#include "Bullet3Common/b3AlignedObjectArray.h"
+#include "Bullet3OpenCL/Initialize/b3OpenCLInclude.h"
+
+template <typename T>
+class b3OpenCLArray
+{
+	size_t	m_size;
+	size_t	m_capacity;
+	cl_mem	m_clBuffer;
+
+	cl_context		 m_clContext;
+	cl_command_queue m_commandQueue;
+
+	bool	m_ownsMemory;
+
+	bool	m_allowGrowingCapacity;
+
+	void deallocate()
+	{
+		if (m_clBuffer && m_ownsMemory)
+		{
+			clReleaseMemObject(m_clBuffer);
+		}
+		m_clBuffer = 0;
+		m_capacity=0;
+	}
+
+	b3OpenCLArray<T>& operator=(const b3OpenCLArray<T>& src);
+
+	B3_FORCE_INLINE	size_t	allocSize(size_t size)
+		{
+			return (size ? size*2 : 1);
+		}
+
+public:
+
+	b3OpenCLArray(cl_context ctx, cl_command_queue queue, size_t initialCapacity=0, bool allowGrowingCapacity=true)
+	:m_size(0),  m_capacity(0),m_clBuffer(0),
+	m_clContext(ctx),m_commandQueue(queue),
+	m_ownsMemory(true),m_allowGrowingCapacity(true)
+	{
+		if (initialCapacity)
+		{
+			reserve(initialCapacity);
+		}
+		m_allowGrowingCapacity = allowGrowingCapacity;
+	}
+
+	///this is an error-prone method with no error checking, be careful!
+	void setFromOpenCLBuffer(cl_mem buffer, size_t sizeInElements)
+	{
+		deallocate();
+		m_ownsMemory = false;
+		m_allowGrowingCapacity = false;
+		m_clBuffer = buffer;
+		m_size = sizeInElements;
+		m_capacity = sizeInElements;
+	}
+
+// we could enable this assignment, but need to make sure to avoid accidental deep copies
+//	b3OpenCLArray<T>& operator=(const b3AlignedObjectArray<T>& src)
+//	{
+//		copyFromArray(src);
+//		return *this;
+//	}
+
+
+	cl_mem	getBufferCL() const
+	{
+		return m_clBuffer;
+	}
+
+
+	virtual ~b3OpenCLArray()
+	{
+		deallocate();
+		m_size=0;
+		m_capacity=0;
+	}
+
+	B3_FORCE_INLINE	bool push_back(const T& _Val,bool waitForCompletion=true)
+	{
+		bool result = true;
+		size_t sz = size();
+		if( sz == capacity() )
+		{
+			result = reserve( allocSize(size()) );
+		}
+		copyFromHostPointer(&_Val, 1, sz, waitForCompletion);
+		m_size++;
+		return result;
+	}
+
+	B3_FORCE_INLINE T forcedAt(size_t n) const
+	{
+		b3Assert(n>=0);
+		b3Assert(n<capacity());
+		T elem;
+		copyToHostPointer(&elem,1,n,true);
+		return elem;
+	}
+
+	B3_FORCE_INLINE T at(size_t n) const
+	{
+		b3Assert(n>=0);
+		b3Assert(n<size());
+		T elem;
+		copyToHostPointer(&elem,1,n,true);
+		return elem;
+	}
+
+	B3_FORCE_INLINE	bool resize(size_t newsize, bool copyOldContents=true)
+	{
+		bool result = true;
+		size_t curSize = size();
+
+		if (newsize < curSize)
+		{
+			//leave the OpenCL memory for now
+		} else
+		{
+			if (newsize > size())
+			{
+				result = reserve(newsize,copyOldContents);
+			}
+
+			//leave new data uninitialized (init in debug mode?)
+			//for (size_t i=curSize;i<newsize;i++) ...
+		}
+
+		if (result)
+		{
+			m_size = newsize;
+		} else
+		{
+			m_size = 0;
+		}
+		return result;
+	}
+
+	B3_FORCE_INLINE size_t size() const
+	{
+		return m_size;
+	}
+
+	B3_FORCE_INLINE	size_t capacity() const
+	{
+		return m_capacity;
+	}
+
+	B3_FORCE_INLINE	bool reserve(size_t _Count, bool copyOldContents=true)
+	{
+		bool result=true;
+		// determine new minimum length of allocated storage
+		if (capacity() < _Count)
+		{	// not enough room, reallocate
+
+			if (m_allowGrowingCapacity)
+			{
+				cl_int ciErrNum;
+				//create a new OpenCL buffer
+				size_t memSizeInBytes = sizeof(T)*_Count;
+				cl_mem buf = clCreateBuffer(m_clContext, CL_MEM_READ_WRITE, memSizeInBytes, NULL, &ciErrNum);
+				if (ciErrNum!=CL_SUCCESS)
+				{
+					b3Error("OpenCL out-of-memory\n");
+					_Count = 0;
+					result = false;
+				}
+//#define B3_ALWAYS_INITIALIZE_OPENCL_BUFFERS
+#ifdef B3_ALWAYS_INITIALIZE_OPENCL_BUFFERS
+				unsigned char* src = (unsigned char*)malloc(memSizeInBytes);
+				for (size_t i=0;i<memSizeInBytes;i++)
+					src[i] = 0xbb;
+				ciErrNum = clEnqueueWriteBuffer( m_commandQueue, buf, CL_TRUE, 0, memSizeInBytes, src, 0,0,0 );
+				b3Assert(ciErrNum==CL_SUCCESS);
+				clFinish(m_commandQueue);
+				free(src);
+#endif //B3_ALWAYS_INITIALIZE_OPENCL_BUFFERS
+
+				if (result)
+				{
+					if (copyOldContents)
+						copyToCL(buf, size());
+				}
+
+				//deallocate the old buffer
+				deallocate();
+
+				m_clBuffer = buf;
+
+				m_capacity = _Count;
+			} else
+			{
+				//fail: assert and
+				b3Assert(0);
+				deallocate();
+				result=false;
+			}
+		}
+		return result;
+	}
+
+
+	void copyToCL(cl_mem destination, size_t numElements, size_t firstElem=0, size_t dstOffsetInElems=0) const
+	{
+		if (numElements<=0)
+			return;
+
+		b3Assert(m_clBuffer);
+		b3Assert(destination);
+
+		//likely some error, destination is same as source
+		b3Assert(m_clBuffer != destination);
+
+		b3Assert((firstElem+numElements)<=m_size);
+
+		cl_int status = 0;
+
+
+		b3Assert(numElements>0);
+		b3Assert(numElements<=m_size);
+
+		size_t srcOffsetBytes = sizeof(T)*firstElem;
+		size_t dstOffsetInBytes = sizeof(T)*dstOffsetInElems;
+
+		status = clEnqueueCopyBuffer( m_commandQueue, m_clBuffer, destination,
+			srcOffsetBytes, dstOffsetInBytes, sizeof(T)*numElements, 0, 0, 0 );
+
+		b3Assert( status == CL_SUCCESS );
+	}
+
+	void copyFromHost(const b3AlignedObjectArray<T>& srcArray, bool waitForCompletion=true)
+	{
+		size_t newSize = srcArray.size();
+
+		bool copyOldContents = false;
+		resize (newSize,copyOldContents);
+		if (newSize)
+			copyFromHostPointer(&srcArray[0],newSize,0,waitForCompletion);
+
+	}
+
+	void copyFromHostPointer(const T* src, size_t numElems, size_t destFirstElem= 0, bool waitForCompletion=true)
+	{
+		b3Assert(numElems+destFirstElem <= capacity());
+
+		if (numElems+destFirstElem)
+		{
+			cl_int status = 0;
+			size_t sizeInBytes=sizeof(T)*numElems;
+			status = clEnqueueWriteBuffer( m_commandQueue, m_clBuffer, 0, sizeof(T)*destFirstElem, sizeInBytes,
+			src, 0,0,0 );
+			b3Assert(status == CL_SUCCESS );
+			if (waitForCompletion)
+				clFinish(m_commandQueue);
+		} else
+		{
+			b3Error("copyFromHostPointer invalid range\n");
+		}
+	}
+
+
+	void copyToHost(b3AlignedObjectArray<T>& destArray, bool waitForCompletion=true) const
+	{
+		destArray.resize(this->size());
+		if (size())
+			copyToHostPointer(&destArray[0], size(),0,waitForCompletion);
+	}
+
+	void copyToHostPointer(T* destPtr, size_t numElem, size_t srcFirstElem=0, bool waitForCompletion=true) const
+	{
+		b3Assert(numElem+srcFirstElem <= capacity());
+
+		if(numElem+srcFirstElem <= capacity())
+		{
+			cl_int status = 0;
+			status = clEnqueueReadBuffer( m_commandQueue, m_clBuffer, 0, sizeof(T)*srcFirstElem, sizeof(T)*numElem,
+			destPtr, 0,0,0 );
+			b3Assert( status==CL_SUCCESS );
+
+			if (waitForCompletion)
+				clFinish(m_commandQueue);
+		} else
+		{
+			b3Error("copyToHostPointer invalid range\n");
+		}
+	}
+
+	void copyFromOpenCLArray(const b3OpenCLArray& src)
+	{
+		size_t newSize = src.size();
+		resize(newSize);
+		if (size())
+		{
+			src.copyToCL(m_clBuffer,size());
+		}
+	}
+
+};
+
+
+#endif //B3_OPENCL_ARRAY_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanCL.cpp b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanCL.cpp
new file mode 100644
index 0000000000..42cd197740
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanCL.cpp
@@ -0,0 +1,126 @@
+#include "b3PrefixScanCL.h"
+#include "b3FillCL.h"
+#define B3_PREFIXSCAN_PROG_PATH "src/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanKernels.cl"
+
+#include "b3LauncherCL.h"
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "kernels/PrefixScanKernelsCL.h"
+
+b3PrefixScanCL::b3PrefixScanCL(cl_context ctx, cl_device_id device, cl_command_queue queue, int size)
+:m_commandQueue(queue)
+{
+	const char* scanKernelSource = prefixScanKernelsCL;
+	cl_int pErrNum;
+	char* additionalMacros=0;
+
+	m_workBuffer = new b3OpenCLArray<unsigned int>(ctx,queue,size);
+	cl_program scanProg = b3OpenCLUtils::compileCLProgramFromString( ctx, device, scanKernelSource, &pErrNum,additionalMacros, B3_PREFIXSCAN_PROG_PATH);
+	b3Assert(scanProg);
+
+	m_localScanKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, scanKernelSource, "LocalScanKernel", &pErrNum, scanProg,additionalMacros );
+	b3Assert(m_localScanKernel );
+	m_blockSumKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, scanKernelSource, "TopLevelScanKernel", &pErrNum, scanProg,additionalMacros );
+	b3Assert(m_blockSumKernel );
+	m_propagationKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, scanKernelSource, "AddOffsetKernel", &pErrNum, scanProg,additionalMacros );
+	b3Assert(m_propagationKernel );
+}
+
+
+b3PrefixScanCL::~b3PrefixScanCL()
+{
+	delete m_workBuffer;
+	clReleaseKernel(m_localScanKernel);
+	clReleaseKernel(m_blockSumKernel);
+	clReleaseKernel(m_propagationKernel);
+}
+
+template<class T>
+T b3NextPowerOf2(T n)
+{
+	n -= 1;
+	for(int i=0; i<sizeof(T)*8; i++)
+		n = n | (n>>i);
+	return n+1;
+}
+
+void b3PrefixScanCL::execute(b3OpenCLArray<unsigned int>& src, b3OpenCLArray<unsigned int>& dst, int n, unsigned int* sum)
+{
+	
+//	b3Assert( data->m_option == EXCLUSIVE );
+	const unsigned int numBlocks = (const unsigned int)( (n+BLOCK_SIZE*2-1)/(BLOCK_SIZE*2) );
+
+	dst.resize(src.size());
+	m_workBuffer->resize(src.size());
+
+	b3Int4 constBuffer;
+	constBuffer.x = n;
+	constBuffer.y = numBlocks;
+	constBuffer.z = (int)b3NextPowerOf2( numBlocks );
+
+	b3OpenCLArray<unsigned int>* srcNative = &src;
+	b3OpenCLArray<unsigned int>* dstNative = &dst;
+	
+	{
+		b3BufferInfoCL bInfo[] = { b3BufferInfoCL( dstNative->getBufferCL() ), b3BufferInfoCL( srcNative->getBufferCL() ), b3BufferInfoCL( m_workBuffer->getBufferCL() ) };
+
+		b3LauncherCL launcher( m_commandQueue, m_localScanKernel,"m_localScanKernel" );
+		launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(  constBuffer );
+		launcher.launch1D( numBlocks*BLOCK_SIZE, BLOCK_SIZE );
+	}
+
+	{
+		b3BufferInfoCL bInfo[] = { b3BufferInfoCL( m_workBuffer->getBufferCL() ) };
+
+		b3LauncherCL launcher( m_commandQueue, m_blockSumKernel,"m_blockSumKernel" );
+		launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst( constBuffer );
+		launcher.launch1D( BLOCK_SIZE, BLOCK_SIZE );
+	}
+	
+
+	if( numBlocks > 1 )
+	{
+		b3BufferInfoCL bInfo[] = { b3BufferInfoCL( dstNative->getBufferCL() ), b3BufferInfoCL( m_workBuffer->getBufferCL() ) };
+		b3LauncherCL launcher( m_commandQueue, m_propagationKernel,"m_propagationKernel" );
+		launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst( constBuffer );
+		launcher.launch1D( (numBlocks-1)*BLOCK_SIZE, BLOCK_SIZE );
+	}
+
+
+	if( sum )
+	{
+		clFinish(m_commandQueue);
+		dstNative->copyToHostPointer(sum,1,n-1,true);
+	}
+
+}
+
+
+void b3PrefixScanCL::executeHost(b3AlignedObjectArray<unsigned int>& src, b3AlignedObjectArray<unsigned int>& dst, int n, unsigned int* sum)
+{
+	unsigned int s = 0;
+	//if( data->m_option == EXCLUSIVE )
+	{
+		for(int i=0; i<n; i++)
+		{
+			dst[i] = s;
+			s += src[i];
+		}
+	}
+	/*else
+	{
+		for(int i=0; i<n; i++)
+		{
+			s += hSrc[i];
+			hDst[i] = s;
+		}
+	}
+	*/
+
+	if( sum )
+	{
+		*sum = dst[n-1];
+	}
+}
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanCL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanCL.h
new file mode 100644
index 0000000000..a9a2e61b9e
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanCL.h
@@ -0,0 +1,37 @@
+
+#ifndef B3_PREFIX_SCAN_CL_H
+#define B3_PREFIX_SCAN_CL_H
+
+#include "b3OpenCLArray.h"
+#include "b3BufferInfoCL.h"
+#include "Bullet3Common/b3AlignedObjectArray.h"
+
+class b3PrefixScanCL
+{
+	enum
+	{
+		BLOCK_SIZE = 128
+	};
+
+//	Option m_option;
+
+	cl_command_queue	m_commandQueue;
+
+	cl_kernel m_localScanKernel;
+	cl_kernel m_blockSumKernel;
+	cl_kernel m_propagationKernel;
+
+	b3OpenCLArray<unsigned int>* m_workBuffer;
+
+
+	public:
+		
+	b3PrefixScanCL(cl_context ctx, cl_device_id device, cl_command_queue queue,int size=0);
+
+	virtual ~b3PrefixScanCL();
+
+	void execute(b3OpenCLArray<unsigned int>& src, b3OpenCLArray<unsigned int>& dst, int n, unsigned int* sum = 0);
+	void executeHost(b3AlignedObjectArray<unsigned int>& src, b3AlignedObjectArray<unsigned int>& dst, int n, unsigned int* sum=0);
+};
+
+#endif //B3_PREFIX_SCAN_CL_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanFloat4CL.cpp b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanFloat4CL.cpp
new file mode 100644
index 0000000000..80560d793d
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanFloat4CL.cpp
@@ -0,0 +1,126 @@
+#include "b3PrefixScanFloat4CL.h"
+#include "b3FillCL.h"
+#define B3_PREFIXSCAN_FLOAT4_PROG_PATH "src/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanFloat4Kernels.cl"
+
+#include "b3LauncherCL.h"
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "kernels/PrefixScanKernelsFloat4CL.h"
+
+b3PrefixScanFloat4CL::b3PrefixScanFloat4CL(cl_context ctx, cl_device_id device, cl_command_queue queue, int size)
+:m_commandQueue(queue)
+{
+	const char* scanKernelSource = prefixScanKernelsFloat4CL;
+	cl_int pErrNum;
+	char* additionalMacros=0;
+
+	m_workBuffer = new b3OpenCLArray<b3Vector3>(ctx,queue,size);
+	cl_program scanProg = b3OpenCLUtils::compileCLProgramFromString( ctx, device, scanKernelSource, &pErrNum,additionalMacros, B3_PREFIXSCAN_FLOAT4_PROG_PATH);
+	b3Assert(scanProg);
+
+	m_localScanKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, scanKernelSource, "LocalScanKernel", &pErrNum, scanProg,additionalMacros );
+	b3Assert(m_localScanKernel );
+	m_blockSumKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, scanKernelSource, "TopLevelScanKernel", &pErrNum, scanProg,additionalMacros );
+	b3Assert(m_blockSumKernel );
+	m_propagationKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, scanKernelSource, "AddOffsetKernel", &pErrNum, scanProg,additionalMacros );
+	b3Assert(m_propagationKernel );
+}
+
+
+b3PrefixScanFloat4CL::~b3PrefixScanFloat4CL()
+{
+	delete m_workBuffer;
+	clReleaseKernel(m_localScanKernel);
+	clReleaseKernel(m_blockSumKernel);
+	clReleaseKernel(m_propagationKernel);
+}
+
+template<class T>
+T b3NextPowerOf2(T n)
+{
+	n -= 1;
+	for(int i=0; i<sizeof(T)*8; i++)
+		n = n | (n>>i);
+	return n+1;
+}
+
+void b3PrefixScanFloat4CL::execute(b3OpenCLArray<b3Vector3>& src, b3OpenCLArray<b3Vector3>& dst, int n, b3Vector3* sum)
+{
+	
+//	b3Assert( data->m_option == EXCLUSIVE );
+	const unsigned int numBlocks = (const unsigned int)( (n+BLOCK_SIZE*2-1)/(BLOCK_SIZE*2) );
+
+	dst.resize(src.size());
+	m_workBuffer->resize(src.size());
+
+	b3Int4 constBuffer;
+	constBuffer.x = n;
+	constBuffer.y = numBlocks;
+	constBuffer.z = (int)b3NextPowerOf2( numBlocks );
+
+	b3OpenCLArray<b3Vector3>* srcNative = &src;
+	b3OpenCLArray<b3Vector3>* dstNative = &dst;
+	
+	{
+		b3BufferInfoCL bInfo[] = { b3BufferInfoCL( dstNative->getBufferCL() ), b3BufferInfoCL( srcNative->getBufferCL() ), b3BufferInfoCL( m_workBuffer->getBufferCL() ) };
+
+		b3LauncherCL launcher( m_commandQueue, m_localScanKernel ,"m_localScanKernel");
+		launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst(  constBuffer );
+		launcher.launch1D( numBlocks*BLOCK_SIZE, BLOCK_SIZE );
+	}
+
+	{
+		b3BufferInfoCL bInfo[] = { b3BufferInfoCL( m_workBuffer->getBufferCL() ) };
+
+		b3LauncherCL launcher( m_commandQueue, m_blockSumKernel ,"m_blockSumKernel");
+		launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst( constBuffer );
+		launcher.launch1D( BLOCK_SIZE, BLOCK_SIZE );
+	}
+	
+
+	if( numBlocks > 1 )
+	{
+		b3BufferInfoCL bInfo[] = { b3BufferInfoCL( dstNative->getBufferCL() ), b3BufferInfoCL( m_workBuffer->getBufferCL() ) };
+		b3LauncherCL launcher( m_commandQueue, m_propagationKernel ,"m_propagationKernel");
+		launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+		launcher.setConst( constBuffer );
+		launcher.launch1D( (numBlocks-1)*BLOCK_SIZE, BLOCK_SIZE );
+	}
+
+
+	if( sum )
+	{
+		clFinish(m_commandQueue);
+		dstNative->copyToHostPointer(sum,1,n-1,true);
+	}
+
+}
+
+
+void b3PrefixScanFloat4CL::executeHost(b3AlignedObjectArray<b3Vector3>& src, b3AlignedObjectArray<b3Vector3>& dst, int n, b3Vector3* sum)
+{
+	b3Vector3 s=b3MakeVector3(0,0,0);
+	//if( data->m_option == EXCLUSIVE )
+	{
+		for(int i=0; i<n; i++)
+		{
+			dst[i] = s;
+			s += src[i];
+		}
+	}
+	/*else
+	{
+		for(int i=0; i<n; i++)
+		{
+			s += hSrc[i];
+			hDst[i] = s;
+		}
+	}
+	*/
+
+	if( sum )
+	{
+		*sum = dst[n-1];
+	}
+}
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanFloat4CL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanFloat4CL.h
new file mode 100644
index 0000000000..2c8003c1bb
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3PrefixScanFloat4CL.h
@@ -0,0 +1,38 @@
+
+#ifndef B3_PREFIX_SCAN_CL_H
+#define B3_PREFIX_SCAN_CL_H
+
+#include "b3OpenCLArray.h"
+#include "b3BufferInfoCL.h"
+#include "Bullet3Common/b3AlignedObjectArray.h"
+#include "Bullet3Common/b3Vector3.h"
+
+class b3PrefixScanFloat4CL
+{
+	enum
+	{
+		BLOCK_SIZE = 128
+	};
+
+//	Option m_option;
+
+	cl_command_queue	m_commandQueue;
+
+	cl_kernel m_localScanKernel;
+	cl_kernel m_blockSumKernel;
+	cl_kernel m_propagationKernel;
+
+	b3OpenCLArray<b3Vector3>* m_workBuffer;
+
+
+	public:
+		
+	b3PrefixScanFloat4CL(cl_context ctx, cl_device_id device, cl_command_queue queue,int size=0);
+
+	virtual ~b3PrefixScanFloat4CL();
+
+	void execute(b3OpenCLArray<b3Vector3>& src, b3OpenCLArray<b3Vector3>& dst, int n, b3Vector3* sum = 0);
+	void executeHost(b3AlignedObjectArray<b3Vector3>& src, b3AlignedObjectArray<b3Vector3>& dst, int n, b3Vector3* sum);
+};
+
+#endif //B3_PREFIX_SCAN_CL_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.cpp b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.cpp
new file mode 100644
index 0000000000..f11ae4bcdb
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.cpp
@@ -0,0 +1,710 @@
+
+#include "b3RadixSort32CL.h"
+#include "b3LauncherCL.h"
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "b3PrefixScanCL.h"
+#include "b3FillCL.h"
+
+#define RADIXSORT32_PATH "src/Bullet3OpenCL/ParallelPrimitives/kernels/RadixSort32Kernels.cl"
+
+#include "kernels/RadixSort32KernelsCL.h"
+
+b3RadixSort32CL::b3RadixSort32CL(cl_context ctx, cl_device_id device, cl_command_queue queue, int initialCapacity)
+:m_commandQueue(queue)
+{
+	b3OpenCLDeviceInfo info;
+	b3OpenCLUtils::getDeviceInfo(device,&info);
+	m_deviceCPU = (info.m_deviceType & CL_DEVICE_TYPE_CPU)!=0;
+
+	m_workBuffer1 = new b3OpenCLArray<unsigned int>(ctx,queue);
+	m_workBuffer2 = new b3OpenCLArray<unsigned int>(ctx,queue);
+	m_workBuffer3 = new b3OpenCLArray<b3SortData>(ctx,queue);
+	m_workBuffer3a = new b3OpenCLArray<unsigned int>(ctx,queue);
+	m_workBuffer4 = new b3OpenCLArray<b3SortData>(ctx,queue);
+	m_workBuffer4a = new b3OpenCLArray<unsigned int>(ctx,queue);
+
+
+	if (initialCapacity>0)
+	{
+		m_workBuffer1->resize(initialCapacity);
+		m_workBuffer3->resize(initialCapacity);
+		m_workBuffer3a->resize(initialCapacity);
+		m_workBuffer4->resize(initialCapacity);
+		m_workBuffer4a->resize(initialCapacity);
+	}
+
+	m_scan = new b3PrefixScanCL(ctx,device,queue);
+	m_fill = new b3FillCL(ctx,device,queue);
+	
+	const char* additionalMacros = "";
+
+	cl_int pErrNum;
+	const char* kernelSource = radixSort32KernelsCL;
+	
+	cl_program sortProg = b3OpenCLUtils::compileCLProgramFromString( ctx, device, kernelSource, &pErrNum,additionalMacros, RADIXSORT32_PATH);
+	b3Assert(sortProg);
+
+	m_streamCountSortDataKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "StreamCountSortDataKernel", &pErrNum, sortProg,additionalMacros );
+	b3Assert(m_streamCountSortDataKernel );
+
+
+	
+	m_streamCountKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "StreamCountKernel", &pErrNum, sortProg,additionalMacros );
+	b3Assert(m_streamCountKernel);
+
+
+	
+	if (m_deviceCPU)
+	{
+		
+		m_sortAndScatterSortDataKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "SortAndScatterSortDataKernelSerial", &pErrNum, sortProg,additionalMacros );
+		b3Assert(m_sortAndScatterSortDataKernel);
+		m_sortAndScatterKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "SortAndScatterKernelSerial", &pErrNum, sortProg,additionalMacros );
+		b3Assert(m_sortAndScatterKernel);
+	} else
+	{
+		m_sortAndScatterSortDataKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "SortAndScatterSortDataKernel", &pErrNum, sortProg,additionalMacros );
+		b3Assert(m_sortAndScatterSortDataKernel);
+		m_sortAndScatterKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "SortAndScatterKernel", &pErrNum, sortProg,additionalMacros );
+		b3Assert(m_sortAndScatterKernel);
+	}
+		
+	m_prefixScanKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, kernelSource, "PrefixScanKernel", &pErrNum, sortProg,additionalMacros );
+	b3Assert(m_prefixScanKernel);
+		
+}
+
+b3RadixSort32CL::~b3RadixSort32CL()
+{
+	delete m_scan;
+	delete m_fill;
+	delete m_workBuffer1;
+	delete m_workBuffer2;
+	delete m_workBuffer3;
+	delete m_workBuffer3a;
+	delete m_workBuffer4;
+	delete m_workBuffer4a;
+
+	clReleaseKernel(m_streamCountSortDataKernel);
+	clReleaseKernel(m_streamCountKernel);
+	clReleaseKernel(m_sortAndScatterSortDataKernel);
+	clReleaseKernel(m_sortAndScatterKernel);
+	clReleaseKernel(m_prefixScanKernel);
+}
+
+void b3RadixSort32CL::executeHost(b3AlignedObjectArray<b3SortData>& inout, int sortBits /* = 32 */)
+{
+	int n = inout.size();
+	const int BITS_PER_PASS = 8;
+	const int NUM_TABLES = (1<<BITS_PER_PASS);
+
+
+	int tables[NUM_TABLES];
+	int counter[NUM_TABLES];
+
+	b3SortData* src = &inout[0];
+	b3AlignedObjectArray<b3SortData> workbuffer;
+	workbuffer.resize(inout.size());
+	b3SortData* dst = &workbuffer[0];
+
+	int count=0;
+	for(int startBit=0; startBit<sortBits; startBit+=BITS_PER_PASS)
+	{
+		for(int i=0; i<NUM_TABLES; i++)
+		{
+			tables[i] = 0;
+		}
+
+		for(int i=0; i<n; i++)
+		{
+			int tableIdx = (src[i].m_key >> startBit) & (NUM_TABLES-1);
+			tables[tableIdx]++;
+		}
+//#define TEST
+#ifdef TEST
+		printf("histogram size=%d\n",NUM_TABLES);
+		for (int i=0;i<NUM_TABLES;i++)
+		{
+			if (tables[i]!=0)
+			{
+				printf("tables[%d]=%d]\n",i,tables[i]);
+			}
+
+		}
+#endif //TEST
+		//	prefix scan
+		int sum = 0;
+		for(int i=0; i<NUM_TABLES; i++)
+		{
+			int iData = tables[i];
+			tables[i] = sum;
+			sum += iData;
+			counter[i] = 0;
+		}
+
+		//	distribute
+		for(int i=0; i<n; i++)
+		{
+			int tableIdx = (src[i].m_key >> startBit) & (NUM_TABLES-1);
+			
+			dst[tables[tableIdx] + counter[tableIdx]] = src[i];
+			counter[tableIdx] ++;
+		}
+
+		b3Swap( src, dst );
+		count++;
+	}
+
+	if (count&1)
+	{
+		b3Assert(0);//need to copy 
+
+	}
+}
+
+void b3RadixSort32CL::executeHost(b3OpenCLArray<b3SortData>& keyValuesInOut, int sortBits /* = 32 */)
+{
+
+	b3AlignedObjectArray<b3SortData> inout;
+	keyValuesInOut.copyToHost(inout);
+
+	executeHost(inout,sortBits);
+
+	keyValuesInOut.copyFromHost(inout);
+}
+
+void b3RadixSort32CL::execute(b3OpenCLArray<unsigned int>& keysIn, b3OpenCLArray<unsigned int>& keysOut, b3OpenCLArray<unsigned int>& valuesIn, 
+								b3OpenCLArray<unsigned int>& valuesOut, int n, int sortBits)
+{
+
+}
+
+//#define DEBUG_RADIXSORT
+//#define DEBUG_RADIXSORT2
+
+
+void b3RadixSort32CL::execute(b3OpenCLArray<b3SortData>& keyValuesInOut, int sortBits /* = 32 */)
+{
+	
+	int originalSize = keyValuesInOut.size();
+	int workingSize = originalSize;
+	
+			
+	int dataAlignment = DATA_ALIGNMENT;
+
+#ifdef DEBUG_RADIXSORT2
+    b3AlignedObjectArray<b3SortData>   test2;
+    keyValuesInOut.copyToHost(test2);
+    printf("numElem = %d\n",test2.size());
+    for (int i=0;i<test2.size();i++)
+    {
+        printf("test2[%d].m_key=%d\n",i,test2[i].m_key);
+        printf("test2[%d].m_value=%d\n",i,test2[i].m_value);
+    }
+#endif //DEBUG_RADIXSORT2
+    
+	b3OpenCLArray<b3SortData>* src = 0;
+
+	if (workingSize%dataAlignment)
+	{
+		workingSize += dataAlignment-(workingSize%dataAlignment);
+		m_workBuffer4->copyFromOpenCLArray(keyValuesInOut);
+		m_workBuffer4->resize(workingSize);
+		b3SortData fillValue;
+		fillValue.m_key = 0xffffffff;
+		fillValue.m_value = 0xffffffff;
+
+#define USE_BTFILL
+#ifdef USE_BTFILL
+		m_fill->execute((b3OpenCLArray<b3Int2>&)*m_workBuffer4,(b3Int2&)fillValue,workingSize-originalSize,originalSize);
+#else
+		//fill the remaining bits (very slow way, todo: fill on GPU/OpenCL side)
+		
+		for (int i=originalSize; i<workingSize;i++)
+		{
+			m_workBuffer4->copyFromHostPointer(&fillValue,1,i);
+		}
+#endif//USE_BTFILL
+
+		src = m_workBuffer4;
+	} else
+	{
+		src = &keyValuesInOut;
+		m_workBuffer4->resize(0);
+	}
+		
+	b3Assert( workingSize%DATA_ALIGNMENT == 0 );
+	int minCap = NUM_BUCKET*NUM_WGS;
+
+
+	int n = workingSize;
+
+	m_workBuffer1->resize(minCap);
+	m_workBuffer3->resize(workingSize);
+	
+
+//	ADLASSERT( ELEMENTS_PER_WORK_ITEM == 4 );
+	b3Assert( BITS_PER_PASS == 4 );
+	b3Assert( WG_SIZE == 64 );
+	b3Assert( (sortBits&0x3) == 0 );
+
+	
+	
+	b3OpenCLArray<b3SortData>* dst = m_workBuffer3;
+
+	b3OpenCLArray<unsigned int>* srcHisto = m_workBuffer1;
+	b3OpenCLArray<unsigned int>* destHisto = m_workBuffer2;
+
+
+	int nWGs = NUM_WGS;
+	b3ConstData cdata;
+
+	{
+        int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;//set at 256
+     	int nBlocks = (n+blockSize-1)/(blockSize);
+		cdata.m_n = n;
+		cdata.m_nWGs = NUM_WGS;
+		cdata.m_startBit = 0;
+		cdata.m_nBlocksPerWG = (nBlocks + cdata.m_nWGs - 1)/cdata.m_nWGs;
+		if( nBlocks < NUM_WGS )
+		{
+			cdata.m_nBlocksPerWG = 1;
+			nWGs = nBlocks;
+		}
+	}
+
+	int count=0;
+	for(int ib=0; ib<sortBits; ib+=4)
+	{
+#ifdef DEBUG_RADIXSORT2
+        keyValuesInOut.copyToHost(test2);
+        printf("numElem = %d\n",test2.size());
+        for (int i=0;i<test2.size();i++)
+        {
+            if (test2[i].m_key != test2[i].m_value)
+            {
+                printf("test2[%d].m_key=%d\n",i,test2[i].m_key);
+                printf("test2[%d].m_value=%d\n",i,test2[i].m_value);
+            }
+        }
+#endif //DEBUG_RADIXSORT2
+        
+		cdata.m_startBit = ib;
+		
+		if (src->size())
+		{
+			b3BufferInfoCL bInfo[] = { b3BufferInfoCL( src->getBufferCL(), true ), b3BufferInfoCL( srcHisto->getBufferCL() ) };
+			b3LauncherCL launcher(m_commandQueue, m_streamCountSortDataKernel,"m_streamCountSortDataKernel");
+
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(  cdata );
+			
+			int num = NUM_WGS*WG_SIZE;
+			launcher.launch1D( num, WG_SIZE );
+		}
+
+        
+        
+#ifdef DEBUG_RADIXSORT
+		b3AlignedObjectArray<unsigned int> testHist;
+		srcHisto->copyToHost(testHist);
+		printf("ib = %d, testHist size = %d, non zero elements:\n",ib, testHist.size());
+		for (int i=0;i<testHist.size();i++)
+		{
+			if (testHist[i]!=0)
+				printf("testHist[%d]=%d\n",i,testHist[i]);
+		}
+#endif //DEBUG_RADIXSORT
+	
+	
+
+//fast prefix scan is not working properly on Mac OSX yet
+#ifdef __APPLE__
+	bool fastScan=false;
+#else
+	bool fastScan=!m_deviceCPU;//only use fast scan on GPU
+#endif
+
+		if (fastScan)
+		{//	prefix scan group histogram
+			b3BufferInfoCL bInfo[] = { b3BufferInfoCL( srcHisto->getBufferCL() ) };
+			b3LauncherCL launcher( m_commandQueue, m_prefixScanKernel,"m_prefixScanKernel" );
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(  cdata );
+			launcher.launch1D( 128, 128 );
+			destHisto = srcHisto;
+		}else
+		{
+			//unsigned int sum; //for debugging
+            m_scan->execute(*srcHisto,*destHisto,1920,0);//,&sum);
+		}
+
+
+#ifdef DEBUG_RADIXSORT
+		destHisto->copyToHost(testHist);
+		printf("ib = %d, testHist size = %d, non zero elements:\n",ib, testHist.size());
+		for (int i=0;i<testHist.size();i++)
+		{
+			if (testHist[i]!=0)
+				printf("testHist[%d]=%d\n",i,testHist[i]);
+		}
+        
+        for (int i=0;i<testHist.size();i+=NUM_WGS)
+		{
+				printf("testHist[%d]=%d\n",i/NUM_WGS,testHist[i]);
+		}
+
+#endif //DEBUG_RADIXSORT
+
+#define USE_GPU
+#ifdef USE_GPU
+        
+		if (src->size())
+		{//	local sort and distribute
+			b3BufferInfoCL bInfo[] = { b3BufferInfoCL( src->getBufferCL(), true ), b3BufferInfoCL( destHisto->getBufferCL(), true ), b3BufferInfoCL( dst->getBufferCL() )};
+			b3LauncherCL launcher( m_commandQueue, m_sortAndScatterSortDataKernel,"m_sortAndScatterSortDataKernel" );
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(  cdata );
+			launcher.launch1D( nWGs*WG_SIZE, WG_SIZE );
+            
+		}
+#else
+        {
+#define NUM_TABLES 16
+//#define SEQUENTIAL
+#ifdef SEQUENTIAL
+            int counter2[NUM_TABLES]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+            int tables[NUM_TABLES];
+            int startBit = ib;
+            
+            destHisto->copyToHost(testHist);
+            b3AlignedObjectArray<b3SortData> srcHost;
+            b3AlignedObjectArray<b3SortData> dstHost;
+            dstHost.resize(src->size());
+            
+            src->copyToHost(srcHost);
+            
+            for (int i=0;i<NUM_TABLES;i++)
+            {
+                tables[i] = testHist[i*NUM_WGS];
+            }
+            
+            //	distribute
+            for(int i=0; i<n; i++)
+            {
+                int tableIdx = (srcHost[i].m_key >> startBit) & (NUM_TABLES-1);
+                
+                dstHost[tables[tableIdx] + counter2[tableIdx]] = srcHost[i];
+                counter2[tableIdx] ++;
+            }
+            
+            
+#else
+          
+            int counter2[NUM_TABLES]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+            
+            int tables[NUM_TABLES];
+             b3AlignedObjectArray<b3SortData> dstHostOK;
+            dstHostOK.resize(src->size());
+
+            destHisto->copyToHost(testHist);
+            b3AlignedObjectArray<b3SortData> srcHost;
+            src->copyToHost(srcHost);
+        
+            int blockSize = 256;
+            int nBlocksPerWG = cdata.m_nBlocksPerWG;
+            int startBit = ib;
+
+            {
+                for (int i=0;i<NUM_TABLES;i++)
+                {
+                    tables[i] = testHist[i*NUM_WGS];
+                }
+                
+                //	distribute
+                for(int i=0; i<n; i++)
+                {
+                    int tableIdx = (srcHost[i].m_key >> startBit) & (NUM_TABLES-1);
+                    
+                    dstHostOK[tables[tableIdx] + counter2[tableIdx]] = srcHost[i];
+                    counter2[tableIdx] ++;
+                }
+
+            
+            }
+            
+            
+            b3AlignedObjectArray<b3SortData> dstHost;
+            dstHost.resize(src->size());
+            
+            
+            int counter[NUM_TABLES]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+            
+            
+            
+            for (int wgIdx=0;wgIdx<NUM_WGS;wgIdx++)
+            {
+              int counter[NUM_TABLES]={0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+
+              int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;
+                
+              for(int iblock=0; iblock<b3Min(cdata.m_nBlocksPerWG, nBlocks); iblock++)
+              {
+                for (int lIdx = 0;lIdx < 64;lIdx++)
+                {
+                    int addr = iblock*blockSize + blockSize*cdata.m_nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;
+                    
+                    //	MY_HISTOGRAM( localKeys.x ) ++ is much expensive than atomic add as it requires read and write while atomics can just add on AMD
+                    //	Using registers didn't perform well. It seems like use localKeys to address requires a lot of alu ops
+                    //	AMD: AtomInc performs better while NV prefers ++
+                    for(int j=0; j<ELEMENTS_PER_WORK_ITEM; j++)
+                    {
+                        if( addr+j < n )
+                        {
+                          //  printf ("addr+j=%d\n", addr+j);
+                            
+                            int i = addr+j;
+                            
+                            int tableIdx = (srcHost[i].m_key >> startBit) & (NUM_TABLES-1);
+                            
+                            int destIndex = testHist[tableIdx*NUM_WGS+wgIdx] + counter[tableIdx];
+                            
+                            b3SortData ok = dstHostOK[destIndex];
+                                                    
+                            if (ok.m_key != srcHost[i].m_key)
+                            {
+                                printf("ok.m_key = %d, srcHost[i].m_key = %d\n", ok.m_key,srcHost[i].m_key );
+                                printf("(ok.m_value = %d, srcHost[i].m_value = %d)\n", ok.m_value,srcHost[i].m_value );
+                            }
+                            if (ok.m_value != srcHost[i].m_value)
+                            {
+                                
+                               printf("ok.m_value = %d, srcHost[i].m_value = %d\n", ok.m_value,srcHost[i].m_value );
+                                printf("(ok.m_key = %d, srcHost[i].m_key = %d)\n", ok.m_key,srcHost[i].m_key );
+
+                            }
+                   
+                            dstHost[destIndex] = srcHost[i];
+                            counter[tableIdx] ++;
+                            
+                        }
+                    }
+                }
+              }
+            }
+            
+         
+#endif //SEQUENTIAL
+            
+            dst->copyFromHost(dstHost);
+        }
+#endif//USE_GPU
+        
+        
+        
+#ifdef DEBUG_RADIXSORT
+		destHisto->copyToHost(testHist);
+		printf("ib = %d, testHist size = %d, non zero elements:\n",ib, testHist.size());
+		for (int i=0;i<testHist.size();i++)
+		{
+			if (testHist[i]!=0)
+				printf("testHist[%d]=%d\n",i,testHist[i]);
+		}
+#endif //DEBUG_RADIXSORT
+		b3Swap(src, dst );
+		b3Swap(srcHisto,destHisto);
+
+#ifdef DEBUG_RADIXSORT2
+        keyValuesInOut.copyToHost(test2);
+        printf("numElem = %d\n",test2.size());
+        for (int i=0;i<test2.size();i++)
+        {
+            if (test2[i].m_key != test2[i].m_value)
+            {
+                printf("test2[%d].m_key=%d\n",i,test2[i].m_key);
+                printf("test2[%d].m_value=%d\n",i,test2[i].m_value);
+            }
+        }
+#endif //DEBUG_RADIXSORT2
+        
+        count++;
+                
+        
+	}
+	
+   
+    
+	if (count&1)
+	{
+		b3Assert(0);//need to copy from workbuffer to keyValuesInOut
+	}
+
+	if (m_workBuffer4->size())
+	{
+		m_workBuffer4->resize(originalSize);
+		keyValuesInOut.copyFromOpenCLArray(*m_workBuffer4);
+	}
+
+
+#ifdef DEBUG_RADIXSORT
+    keyValuesInOut.copyToHost(test2);
+   
+    printf("numElem = %d\n",test2.size());
+    for (int i=0;i<test2.size();i++)
+    {
+        printf("test2[%d].m_key=%d\n",i,test2[i].m_key);
+        printf("test2[%d].m_value=%d\n",i,test2[i].m_value);
+    }
+#endif    
+	
+}
+
+
+
+
+
+
+void b3RadixSort32CL::execute(b3OpenCLArray<unsigned int>& keysInOut, int sortBits /* = 32 */)
+{
+	int originalSize = keysInOut.size();
+	int workingSize = originalSize;
+	
+			
+	int dataAlignment = DATA_ALIGNMENT;
+
+	b3OpenCLArray<unsigned int>* src = 0;
+
+	if (workingSize%dataAlignment)
+	{
+		workingSize += dataAlignment-(workingSize%dataAlignment);
+		m_workBuffer4a->copyFromOpenCLArray(keysInOut);
+		m_workBuffer4a->resize(workingSize);
+		unsigned int fillValue = 0xffffffff;
+		
+		m_fill->execute(*m_workBuffer4a,fillValue,workingSize-originalSize,originalSize);
+
+		src = m_workBuffer4a;
+	} else
+	{
+		src = &keysInOut;
+		m_workBuffer4a->resize(0);
+	}
+	
+	
+
+	b3Assert( workingSize%DATA_ALIGNMENT == 0 );
+	int minCap = NUM_BUCKET*NUM_WGS;
+
+
+	int n = workingSize;
+
+	
+	m_workBuffer1->resize(minCap);
+	m_workBuffer3->resize(workingSize);
+	m_workBuffer3a->resize(workingSize);
+
+//	ADLASSERT( ELEMENTS_PER_WORK_ITEM == 4 );
+	b3Assert( BITS_PER_PASS == 4 );
+	b3Assert( WG_SIZE == 64 );
+	b3Assert( (sortBits&0x3) == 0 );
+
+	
+	
+	b3OpenCLArray<unsigned int>* dst = m_workBuffer3a;
+
+	b3OpenCLArray<unsigned int>* srcHisto = m_workBuffer1;
+	b3OpenCLArray<unsigned int>* destHisto = m_workBuffer2;
+
+
+	int nWGs = NUM_WGS;
+	b3ConstData cdata;
+
+	{
+        int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;//set at 256
+     	int nBlocks = (n+blockSize-1)/(blockSize);
+		cdata.m_n = n;
+		cdata.m_nWGs = NUM_WGS;
+		cdata.m_startBit = 0;
+		cdata.m_nBlocksPerWG = (nBlocks + cdata.m_nWGs - 1)/cdata.m_nWGs;
+		if( nBlocks < NUM_WGS )
+		{
+			cdata.m_nBlocksPerWG = 1;
+			nWGs = nBlocks;
+		}
+	}
+
+	int count=0;
+	for(int ib=0; ib<sortBits; ib+=4)
+	{
+		cdata.m_startBit = ib;
+		
+		if (src->size())
+		{
+			b3BufferInfoCL bInfo[] = { b3BufferInfoCL( src->getBufferCL(), true ), b3BufferInfoCL( srcHisto->getBufferCL() ) };
+			b3LauncherCL launcher(m_commandQueue, m_streamCountKernel,"m_streamCountKernel");
+
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(  cdata );
+			
+			int num = NUM_WGS*WG_SIZE;
+			launcher.launch1D( num, WG_SIZE );
+		}
+
+        
+
+//fast prefix scan is not working properly on Mac OSX yet
+#ifdef __APPLE__
+	bool fastScan=false;	
+#else
+	bool fastScan=!m_deviceCPU;
+#endif
+
+		if (fastScan)
+		{//	prefix scan group histogram
+			b3BufferInfoCL bInfo[] = { b3BufferInfoCL( srcHisto->getBufferCL() ) };
+			b3LauncherCL launcher( m_commandQueue, m_prefixScanKernel,"m_prefixScanKernel" );
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(  cdata );
+			launcher.launch1D( 128, 128 );
+			destHisto = srcHisto;
+		}else
+		{
+			//unsigned int sum; //for debugging
+            m_scan->execute(*srcHisto,*destHisto,1920,0);//,&sum);
+		}
+
+		if (src->size())
+		{//	local sort and distribute
+			b3BufferInfoCL bInfo[] = { b3BufferInfoCL( src->getBufferCL(), true ), b3BufferInfoCL( destHisto->getBufferCL(), true ), b3BufferInfoCL( dst->getBufferCL() )};
+			b3LauncherCL launcher( m_commandQueue, m_sortAndScatterKernel ,"m_sortAndScatterKernel");
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(  cdata );
+			launcher.launch1D( nWGs*WG_SIZE, WG_SIZE );
+            
+		}
+        
+		b3Swap(src, dst );
+		b3Swap(srcHisto,destHisto);
+
+        count++;
+	}
+    
+	if (count&1)
+	{
+		b3Assert(0);//need to copy from workbuffer to keyValuesInOut
+	}
+
+	if (m_workBuffer4a->size())
+	{
+		m_workBuffer4a->resize(originalSize);
+		keysInOut.copyFromOpenCLArray(*m_workBuffer4a);
+	}
+	
+}
+
+
+
+
+
+
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h
new file mode 100644
index 0000000000..975bd80e53
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h
@@ -0,0 +1,95 @@
+
+#ifndef B3_RADIXSORT32_H
+#define B3_RADIXSORT32_H
+
+#include "b3OpenCLArray.h"
+
+struct b3SortData
+{
+	union
+	{
+		unsigned int m_key;
+		unsigned int x;
+	};
+
+	union
+	{
+		unsigned int m_value;
+		unsigned int y;
+		
+	};
+};
+#include "b3BufferInfoCL.h"
+
+class  b3RadixSort32CL
+{
+
+		b3OpenCLArray<unsigned int>* m_workBuffer1;
+		b3OpenCLArray<unsigned int>* m_workBuffer2;
+		
+		b3OpenCLArray<b3SortData>*	m_workBuffer3;
+		b3OpenCLArray<b3SortData>*	m_workBuffer4;
+
+		b3OpenCLArray<unsigned int>* m_workBuffer3a;
+		b3OpenCLArray<unsigned int>* m_workBuffer4a;
+
+		cl_command_queue	m_commandQueue;
+
+		cl_kernel m_streamCountSortDataKernel;
+		cl_kernel m_streamCountKernel;
+
+		cl_kernel m_prefixScanKernel;
+		cl_kernel m_sortAndScatterSortDataKernel;
+		cl_kernel m_sortAndScatterKernel;
+
+
+		bool	m_deviceCPU;
+
+		class b3PrefixScanCL* m_scan;
+		class b3FillCL*	m_fill;
+
+public:
+	struct b3ConstData
+		{
+			int m_n;
+			int m_nWGs;
+			int m_startBit;
+			int m_nBlocksPerWG;
+		};
+	enum
+		{
+			DATA_ALIGNMENT = 256,
+			WG_SIZE = 64,
+            BLOCK_SIZE = 256,
+			ELEMENTS_PER_WORK_ITEM = (BLOCK_SIZE/WG_SIZE),
+			BITS_PER_PASS = 4,
+			NUM_BUCKET=(1<<BITS_PER_PASS),
+			//	if you change this, change nPerWI in kernel as well
+			NUM_WGS = 20*6,	//	cypress
+//			NUM_WGS = 24*6,	//	cayman
+//			NUM_WGS = 32*4,	//	nv
+		};
+
+
+private:
+		
+
+public:
+
+		b3RadixSort32CL(cl_context ctx, cl_device_id device, cl_command_queue queue, int initialCapacity =0);
+
+		virtual ~b3RadixSort32CL();
+
+		void execute(b3OpenCLArray<unsigned int>& keysIn, b3OpenCLArray<unsigned int>& keysOut, b3OpenCLArray<unsigned int>& valuesIn, 
+								b3OpenCLArray<unsigned int>& valuesOut, int n, int sortBits = 32);
+
+		///keys only
+		void execute(b3OpenCLArray<unsigned int>& keysInOut, int sortBits  = 32 );
+
+		void execute(b3OpenCLArray<b3SortData>& keyValuesInOut, int sortBits  = 32 );
+		void executeHost(b3OpenCLArray<b3SortData>& keyValuesInOut, int sortBits = 32);
+		void executeHost(b3AlignedObjectArray<b3SortData>& keyValuesInOut, int sortBits = 32);
+
+};
+#endif //B3_RADIXSORT32_H
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/BoundSearchKernels.cl b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/BoundSearchKernels.cl
new file mode 100644
index 0000000000..f3b4a1e8a7
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/BoundSearchKernels.cl
@@ -0,0 +1,106 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+typedef unsigned int u32;
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+
+typedef struct
+{
+	u32 m_key; 
+	u32 m_value;
+}SortData;
+
+
+
+typedef struct
+{
+	u32 m_nSrc;
+	u32 m_nDst;
+	u32 m_padding[2];
+} ConstBuffer;
+
+
+
+__attribute__((reqd_work_group_size(64,1,1)))
+__kernel
+void SearchSortDataLowerKernel(__global SortData* src, __global u32 *dst, 
+					unsigned int nSrc, unsigned int nDst)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < nSrc )
+	{
+		SortData first; first.m_key = (u32)(-1); first.m_value = (u32)(-1);
+		SortData end; end.m_key = nDst; end.m_value = nDst;
+
+		SortData iData = (gIdx==0)? first: src[gIdx-1];
+		SortData jData = (gIdx==nSrc)? end: src[gIdx];
+
+		if( iData.m_key != jData.m_key )
+		{
+//			for(u32 k=iData.m_key+1; k<=min(jData.m_key, nDst-1); k++)
+			u32 k = jData.m_key;
+			{
+				dst[k] = gIdx;
+			}
+		}
+	}
+}
+
+
+__attribute__((reqd_work_group_size(64,1,1)))
+__kernel
+void SearchSortDataUpperKernel(__global SortData* src, __global u32 *dst, 
+					unsigned int nSrc, unsigned int nDst)
+{
+	int gIdx = GET_GLOBAL_IDX+1;
+
+	if( gIdx < nSrc+1 )
+	{
+		SortData first; first.m_key = 0; first.m_value = 0;
+		SortData end; end.m_key = nDst; end.m_value = nDst;
+
+		SortData iData = src[gIdx-1];
+		SortData jData = (gIdx==nSrc)? end: src[gIdx];
+
+		if( iData.m_key != jData.m_key )
+		{
+			u32 k = iData.m_key;
+			{
+				dst[k] = gIdx;
+			}
+		}
+	}
+}
+
+__attribute__((reqd_work_group_size(64,1,1)))
+__kernel
+void SubtractKernel(__global u32* A, __global u32 *B, __global u32 *C, 
+					unsigned int nSrc, unsigned int nDst)
+{
+	int gIdx = GET_GLOBAL_IDX;
+	
+
+	if( gIdx < nDst )
+	{
+		C[gIdx] = A[gIdx] - B[gIdx];
+	}
+}
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/BoundSearchKernelsCL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/BoundSearchKernelsCL.h
new file mode 100644
index 0000000000..9c9e847138
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/BoundSearchKernelsCL.h
@@ -0,0 +1,87 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* boundSearchKernelsCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"typedef unsigned int u32;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"typedef struct\n"
+"{\n"
+"	u32 m_key; \n"
+"	u32 m_value;\n"
+"}SortData;\n"
+"typedef struct\n"
+"{\n"
+"	u32 m_nSrc;\n"
+"	u32 m_nDst;\n"
+"	u32 m_padding[2];\n"
+"} ConstBuffer;\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"__kernel\n"
+"void SearchSortDataLowerKernel(__global SortData* src, __global u32 *dst, \n"
+"					unsigned int nSrc, unsigned int nDst)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < nSrc )\n"
+"	{\n"
+"		SortData first; first.m_key = (u32)(-1); first.m_value = (u32)(-1);\n"
+"		SortData end; end.m_key = nDst; end.m_value = nDst;\n"
+"		SortData iData = (gIdx==0)? first: src[gIdx-1];\n"
+"		SortData jData = (gIdx==nSrc)? end: src[gIdx];\n"
+"		if( iData.m_key != jData.m_key )\n"
+"		{\n"
+"//			for(u32 k=iData.m_key+1; k<=min(jData.m_key, nDst-1); k++)\n"
+"			u32 k = jData.m_key;\n"
+"			{\n"
+"				dst[k] = gIdx;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"}\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"__kernel\n"
+"void SearchSortDataUpperKernel(__global SortData* src, __global u32 *dst, \n"
+"					unsigned int nSrc, unsigned int nDst)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX+1;\n"
+"	if( gIdx < nSrc+1 )\n"
+"	{\n"
+"		SortData first; first.m_key = 0; first.m_value = 0;\n"
+"		SortData end; end.m_key = nDst; end.m_value = nDst;\n"
+"		SortData iData = src[gIdx-1];\n"
+"		SortData jData = (gIdx==nSrc)? end: src[gIdx];\n"
+"		if( iData.m_key != jData.m_key )\n"
+"		{\n"
+"			u32 k = iData.m_key;\n"
+"			{\n"
+"				dst[k] = gIdx;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"}\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"__kernel\n"
+"void SubtractKernel(__global u32* A, __global u32 *B, __global u32 *C, \n"
+"					unsigned int nSrc, unsigned int nDst)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	\n"
+"	if( gIdx < nDst )\n"
+"	{\n"
+"		C[gIdx] = A[gIdx] - B[gIdx];\n"
+"	}\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/CopyKernels.cl b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/CopyKernels.cl
new file mode 100644
index 0000000000..2eee5752ec
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/CopyKernels.cl
@@ -0,0 +1,128 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+
+typedef unsigned int u32;
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+
+#define make_uint4 (uint4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+typedef struct
+{
+	int m_n;
+	int m_padding[3];
+} ConstBuffer;
+
+
+
+__kernel
+__attribute__((reqd_work_group_size(64,1,1)))
+void Copy1F4Kernel(__global float4* dst, __global float4* src, 
+					ConstBuffer cb)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < cb.m_n )
+	{
+		float4 a0 = src[gIdx];
+
+		dst[ gIdx ] = a0;
+	}
+}
+
+__kernel
+__attribute__((reqd_work_group_size(64,1,1)))
+void Copy2F4Kernel(__global float4* dst, __global float4* src, 
+					ConstBuffer cb)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( 2*gIdx <= cb.m_n )
+	{
+		float4 a0 = src[gIdx*2+0];
+		float4 a1 = src[gIdx*2+1];
+
+		dst[ gIdx*2+0 ] = a0;
+		dst[ gIdx*2+1 ] = a1;
+	}
+}
+
+__kernel
+__attribute__((reqd_work_group_size(64,1,1)))
+void Copy4F4Kernel(__global float4* dst, __global float4* src, 
+					ConstBuffer cb)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( 4*gIdx <= cb.m_n )
+	{
+		int idx0 = gIdx*4+0;
+		int idx1 = gIdx*4+1;
+		int idx2 = gIdx*4+2;
+		int idx3 = gIdx*4+3;
+
+		float4 a0 = src[idx0];
+		float4 a1 = src[idx1];
+		float4 a2 = src[idx2];
+		float4 a3 = src[idx3];
+
+		dst[ idx0 ] = a0;
+		dst[ idx1 ] = a1;
+		dst[ idx2 ] = a2;
+		dst[ idx3 ] = a3;
+	}
+}
+
+__kernel
+__attribute__((reqd_work_group_size(64,1,1)))
+void CopyF1Kernel(__global float* dstF1, __global float* srcF1, 
+					ConstBuffer cb)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < cb.m_n )
+	{
+		float a0 = srcF1[gIdx];
+
+		dstF1[ gIdx ] = a0;
+	}
+}
+
+__kernel
+__attribute__((reqd_work_group_size(64,1,1)))
+void CopyF2Kernel(__global float2* dstF2, __global float2* srcF2, 
+					ConstBuffer cb)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < cb.m_n )
+	{
+		float2 a0 = srcF2[gIdx];
+
+		dstF2[ gIdx ] = a0;
+	}
+}
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/CopyKernelsCL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/CopyKernelsCL.h
new file mode 100644
index 0000000000..e5670e3cd3
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/CopyKernelsCL.h
@@ -0,0 +1,132 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* copyKernelsCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"\n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"\n"
+"typedef unsigned int u32;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"\n"
+"#define make_uint4 (uint4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"\n"
+"typedef struct\n"
+"{\n"
+"	int m_n;\n"
+"	int m_padding[3];\n"
+"} ConstBuffer;\n"
+"\n"
+"\n"
+"\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"void Copy1F4Kernel(__global float4* dst, __global float4* src, \n"
+"					ConstBuffer cb)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"\n"
+"	if( gIdx < cb.m_n )\n"
+"	{\n"
+"		float4 a0 = src[gIdx];\n"
+"\n"
+"		dst[ gIdx ] = a0;\n"
+"	}\n"
+"}\n"
+"\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"void Copy2F4Kernel(__global float4* dst, __global float4* src, \n"
+"					ConstBuffer cb)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"\n"
+"	if( 2*gIdx <= cb.m_n )\n"
+"	{\n"
+"		float4 a0 = src[gIdx*2+0];\n"
+"		float4 a1 = src[gIdx*2+1];\n"
+"\n"
+"		dst[ gIdx*2+0 ] = a0;\n"
+"		dst[ gIdx*2+1 ] = a1;\n"
+"	}\n"
+"}\n"
+"\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"void Copy4F4Kernel(__global float4* dst, __global float4* src, \n"
+"					ConstBuffer cb)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"\n"
+"	if( 4*gIdx <= cb.m_n )\n"
+"	{\n"
+"		int idx0 = gIdx*4+0;\n"
+"		int idx1 = gIdx*4+1;\n"
+"		int idx2 = gIdx*4+2;\n"
+"		int idx3 = gIdx*4+3;\n"
+"\n"
+"		float4 a0 = src[idx0];\n"
+"		float4 a1 = src[idx1];\n"
+"		float4 a2 = src[idx2];\n"
+"		float4 a3 = src[idx3];\n"
+"\n"
+"		dst[ idx0 ] = a0;\n"
+"		dst[ idx1 ] = a1;\n"
+"		dst[ idx2 ] = a2;\n"
+"		dst[ idx3 ] = a3;\n"
+"	}\n"
+"}\n"
+"\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"void CopyF1Kernel(__global float* dstF1, __global float* srcF1, \n"
+"					ConstBuffer cb)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"\n"
+"	if( gIdx < cb.m_n )\n"
+"	{\n"
+"		float a0 = srcF1[gIdx];\n"
+"\n"
+"		dstF1[ gIdx ] = a0;\n"
+"	}\n"
+"}\n"
+"\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"void CopyF2Kernel(__global float2* dstF2, __global float2* srcF2, \n"
+"					ConstBuffer cb)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"\n"
+"	if( gIdx < cb.m_n )\n"
+"	{\n"
+"		float2 a0 = srcF2[gIdx];\n"
+"\n"
+"		dstF2[ gIdx ] = a0;\n"
+"	}\n"
+"}\n"
+"\n"
+"\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/FillKernels.cl b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/FillKernels.cl
new file mode 100644
index 0000000000..71c31075dd
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/FillKernels.cl
@@ -0,0 +1,107 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+
+typedef unsigned int u32;
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+
+#define make_uint4 (uint4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+typedef struct
+{
+	union
+	{
+		int4 m_data;
+		uint4 m_unsignedData;
+		float	m_floatData;
+	};
+	int m_offset;
+	int m_n;
+	int m_padding[2];
+} ConstBuffer;
+
+
+__kernel
+__attribute__((reqd_work_group_size(64,1,1)))
+void FillIntKernel(__global int* dstInt, 			int num_elements, int value, const int offset)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < num_elements )
+	{
+		dstInt[ offset+gIdx ] = value;
+	}
+}
+
+__kernel
+__attribute__((reqd_work_group_size(64,1,1)))
+void FillFloatKernel(__global float* dstFloat, 	int num_elements, float value, const int offset)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < num_elements )
+	{
+		dstFloat[ offset+gIdx ] = value;
+	}
+}
+
+__kernel
+__attribute__((reqd_work_group_size(64,1,1)))
+void FillUnsignedIntKernel(__global unsigned int* dstInt, const int num, const unsigned int value, const int offset)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < num )
+	{
+		dstInt[ offset+gIdx ] = value;
+	}
+}
+
+__kernel
+__attribute__((reqd_work_group_size(64,1,1)))
+void FillInt2Kernel(__global int2* dstInt2, 	const int num, const int2 value, const int offset)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < num )
+	{
+		dstInt2[ gIdx + offset] = make_int2( value.x, value.y );
+	}
+}
+
+__kernel
+__attribute__((reqd_work_group_size(64,1,1)))
+void FillInt4Kernel(__global int4* dstInt4, 		const int num, const int4 value, const int offset)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < num )
+	{
+		dstInt4[ offset+gIdx ] = value;
+	}
+}
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/FillKernelsCL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/FillKernelsCL.h
new file mode 100644
index 0000000000..4f8b96e489
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/FillKernelsCL.h
@@ -0,0 +1,91 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* fillKernelsCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"typedef unsigned int u32;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define make_uint4 (uint4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"typedef struct\n"
+"{\n"
+"	union\n"
+"	{\n"
+"		int4 m_data;\n"
+"		uint4 m_unsignedData;\n"
+"		float	m_floatData;\n"
+"	};\n"
+"	int m_offset;\n"
+"	int m_n;\n"
+"	int m_padding[2];\n"
+"} ConstBuffer;\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"void FillIntKernel(__global int* dstInt, 			int num_elements, int value, const int offset)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < num_elements )\n"
+"	{\n"
+"		dstInt[ offset+gIdx ] = value;\n"
+"	}\n"
+"}\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"void FillFloatKernel(__global float* dstFloat, 	int num_elements, float value, const int offset)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < num_elements )\n"
+"	{\n"
+"		dstFloat[ offset+gIdx ] = value;\n"
+"	}\n"
+"}\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"void FillUnsignedIntKernel(__global unsigned int* dstInt, const int num, const unsigned int value, const int offset)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < num )\n"
+"	{\n"
+"		dstInt[ offset+gIdx ] = value;\n"
+"	}\n"
+"}\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"void FillInt2Kernel(__global int2* dstInt2, 	const int num, const int2 value, const int offset)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < num )\n"
+"	{\n"
+"		dstInt2[ gIdx + offset] = make_int2( value.x, value.y );\n"
+"	}\n"
+"}\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(64,1,1)))\n"
+"void FillInt4Kernel(__global int4* dstInt4, 		const int num, const int4 value, const int offset)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < num )\n"
+"	{\n"
+"		dstInt4[ offset+gIdx ] = value;\n"
+"	}\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanFloat4Kernels.cl b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanFloat4Kernels.cl
new file mode 100644
index 0000000000..c9da79854a
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanFloat4Kernels.cl
@@ -0,0 +1,154 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+typedef unsigned int u32;
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+
+// takahiro end
+#define WG_SIZE 128 
+#define m_numElems x
+#define m_numBlocks y
+#define m_numScanBlocks z
+
+/*typedef struct
+{
+	uint m_numElems;
+	uint m_numBlocks;
+	uint m_numScanBlocks;
+	uint m_padding[1];
+} ConstBuffer;
+*/
+
+float4 ScanExclusiveFloat4(__local float4* data, u32 n, int lIdx, int lSize)
+{
+	float4 blocksum;
+    int offset = 1;
+    for(int nActive=n>>1; nActive>0; nActive>>=1, offset<<=1)
+    {
+        GROUP_LDS_BARRIER;
+        for(int iIdx=lIdx; iIdx<nActive; iIdx+=lSize)
+        {
+            int ai = offset*(2*iIdx+1)-1;
+            int bi = offset*(2*iIdx+2)-1;
+            data[bi] += data[ai];
+        }
+	}
+
+    GROUP_LDS_BARRIER;
+
+    if( lIdx == 0 )
+	{
+		blocksum = data[ n-1 ];
+    data[ n-1 ] = 0;
+	}
+
+	GROUP_LDS_BARRIER;
+
+	offset >>= 1;
+    for(int nActive=1; nActive<n; nActive<<=1, offset>>=1 )
+    {
+        GROUP_LDS_BARRIER;
+        for( int iIdx = lIdx; iIdx<nActive; iIdx += lSize )
+        {
+            int ai = offset*(2*iIdx+1)-1;
+            int bi = offset*(2*iIdx+2)-1;
+            float4 temp = data[ai];
+            data[ai] = data[bi];
+            data[bi] += temp;
+        }
+	}
+	GROUP_LDS_BARRIER;
+
+	return blocksum;
+}
+
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+__kernel
+void LocalScanKernel(__global float4* dst, __global float4* src, __global float4* sumBuffer,	uint4 cb)
+{
+	__local float4 ldsData[WG_SIZE*2];
+
+	int gIdx = GET_GLOBAL_IDX;
+	int lIdx = GET_LOCAL_IDX;
+
+	ldsData[2*lIdx]     = ( 2*gIdx < cb.m_numElems )? src[2*gIdx]: 0;
+	ldsData[2*lIdx + 1] = ( 2*gIdx+1 < cb.m_numElems )? src[2*gIdx + 1]: 0;
+
+	float4 sum = ScanExclusiveFloat4(ldsData, WG_SIZE*2, GET_LOCAL_IDX, GET_GROUP_SIZE);
+
+	if( lIdx == 0 ) 
+		sumBuffer[GET_GROUP_IDX] = sum;
+
+	if( (2*gIdx) < cb.m_numElems )
+    {
+        dst[2*gIdx]     = ldsData[2*lIdx];
+	}
+	if( (2*gIdx + 1) < cb.m_numElems )
+	{
+        dst[2*gIdx + 1] = ldsData[2*lIdx + 1];
+    }
+}
+
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+__kernel
+void AddOffsetKernel(__global float4* dst, __global float4* blockSum, uint4 cb)
+{
+	const u32 blockSize = WG_SIZE*2;
+
+	int myIdx = GET_GROUP_IDX+1;
+	int lIdx = GET_LOCAL_IDX;
+
+	float4 iBlockSum = blockSum[myIdx];
+
+	int endValue = min((myIdx+1)*(blockSize), cb.m_numElems);
+	for(int i=myIdx*blockSize+lIdx; i<endValue; i+=GET_GROUP_SIZE)
+	{
+		dst[i] += iBlockSum;
+	}
+}
+
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+__kernel
+void TopLevelScanKernel(__global float4* dst, uint4 cb)
+{
+	__local float4 ldsData[2048];
+	int gIdx = GET_GLOBAL_IDX;
+	int lIdx = GET_LOCAL_IDX;
+	int lSize = GET_GROUP_SIZE;
+
+	for(int i=lIdx; i<cb.m_numScanBlocks; i+=lSize )
+	{
+		ldsData[i] = (i<cb.m_numBlocks)? dst[i]:0;
+	}
+
+	GROUP_LDS_BARRIER;
+
+	float4 sum = ScanExclusiveFloat4(ldsData, cb.m_numScanBlocks, GET_LOCAL_IDX, GET_GROUP_SIZE);
+
+	for(int i=lIdx; i<cb.m_numBlocks; i+=lSize )
+	{
+		dst[i] = ldsData[i];
+	}
+
+	if( gIdx == 0 )
+	{
+		dst[cb.m_numBlocks] = sum;
+	}
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanKernels.cl b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanKernels.cl
new file mode 100644
index 0000000000..963cc1e48e
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanKernels.cl
@@ -0,0 +1,154 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+typedef unsigned int u32;
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+
+// takahiro end
+#define WG_SIZE 128 
+#define m_numElems x
+#define m_numBlocks y
+#define m_numScanBlocks z
+
+/*typedef struct
+{
+	uint m_numElems;
+	uint m_numBlocks;
+	uint m_numScanBlocks;
+	uint m_padding[1];
+} ConstBuffer;
+*/
+
+u32 ScanExclusive(__local u32* data, u32 n, int lIdx, int lSize)
+{
+	u32 blocksum;
+    int offset = 1;
+    for(int nActive=n>>1; nActive>0; nActive>>=1, offset<<=1)
+    {
+        GROUP_LDS_BARRIER;
+        for(int iIdx=lIdx; iIdx<nActive; iIdx+=lSize)
+        {
+            int ai = offset*(2*iIdx+1)-1;
+            int bi = offset*(2*iIdx+2)-1;
+            data[bi] += data[ai];
+        }
+	}
+
+    GROUP_LDS_BARRIER;
+
+    if( lIdx == 0 )
+	{
+		blocksum = data[ n-1 ];
+        data[ n-1 ] = 0;
+	}
+
+	GROUP_LDS_BARRIER;
+
+	offset >>= 1;
+    for(int nActive=1; nActive<n; nActive<<=1, offset>>=1 )
+    {
+        GROUP_LDS_BARRIER;
+        for( int iIdx = lIdx; iIdx<nActive; iIdx += lSize )
+        {
+            int ai = offset*(2*iIdx+1)-1;
+            int bi = offset*(2*iIdx+2)-1;
+            u32 temp = data[ai];
+            data[ai] = data[bi];
+            data[bi] += temp;
+        }
+	}
+	GROUP_LDS_BARRIER;
+
+	return blocksum;
+}
+
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+__kernel
+void LocalScanKernel(__global u32* dst, __global u32 *src, __global u32 *sumBuffer,
+		uint4 cb)
+{
+	__local u32 ldsData[WG_SIZE*2];
+
+	int gIdx = GET_GLOBAL_IDX;
+	int lIdx = GET_LOCAL_IDX;
+
+	ldsData[2*lIdx]     = ( 2*gIdx < cb.m_numElems )? src[2*gIdx]: 0;
+	ldsData[2*lIdx + 1] = ( 2*gIdx+1 < cb.m_numElems )? src[2*gIdx + 1]: 0;
+
+	u32 sum = ScanExclusive(ldsData, WG_SIZE*2, GET_LOCAL_IDX, GET_GROUP_SIZE);
+
+	if( lIdx == 0 ) sumBuffer[GET_GROUP_IDX] = sum;
+
+	if( (2*gIdx) < cb.m_numElems )
+    {
+        dst[2*gIdx]     = ldsData[2*lIdx];
+	}
+	if( (2*gIdx + 1) < cb.m_numElems )
+	{
+        dst[2*gIdx + 1] = ldsData[2*lIdx + 1];
+    }
+}
+
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+__kernel
+void AddOffsetKernel(__global u32 *dst, __global u32 *blockSum, uint4 cb)
+{
+	const u32 blockSize = WG_SIZE*2;
+
+	int myIdx = GET_GROUP_IDX+1;
+	int lIdx = GET_LOCAL_IDX;
+
+	u32 iBlockSum = blockSum[myIdx];
+
+	int endValue = min((myIdx+1)*(blockSize), cb.m_numElems);
+	for(int i=myIdx*blockSize+lIdx; i<endValue; i+=GET_GROUP_SIZE)
+	{
+		dst[i] += iBlockSum;
+	}
+}
+
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+__kernel
+void TopLevelScanKernel(__global u32* dst, uint4 cb)
+{
+	__local u32 ldsData[2048];
+	int gIdx = GET_GLOBAL_IDX;
+	int lIdx = GET_LOCAL_IDX;
+	int lSize = GET_GROUP_SIZE;
+
+	for(int i=lIdx; i<cb.m_numScanBlocks; i+=lSize )
+	{
+		ldsData[i] = (i<cb.m_numBlocks)? dst[i]:0;
+	}
+
+	GROUP_LDS_BARRIER;
+
+	u32 sum = ScanExclusive(ldsData, cb.m_numScanBlocks, GET_LOCAL_IDX, GET_GROUP_SIZE);
+
+	for(int i=lIdx; i<cb.m_numBlocks; i+=lSize )
+	{
+		dst[i] = ldsData[i];
+	}
+
+	if( gIdx == 0 )
+	{
+		dst[cb.m_numBlocks] = sum;
+	}
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanKernelsCL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanKernelsCL.h
new file mode 100644
index 0000000000..27baab8331
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanKernelsCL.h
@@ -0,0 +1,129 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* prefixScanKernelsCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"typedef unsigned int u32;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"// takahiro end\n"
+"#define WG_SIZE 128 \n"
+"#define m_numElems x\n"
+"#define m_numBlocks y\n"
+"#define m_numScanBlocks z\n"
+"/*typedef struct\n"
+"{\n"
+"	uint m_numElems;\n"
+"	uint m_numBlocks;\n"
+"	uint m_numScanBlocks;\n"
+"	uint m_padding[1];\n"
+"} ConstBuffer;\n"
+"*/\n"
+"u32 ScanExclusive(__local u32* data, u32 n, int lIdx, int lSize)\n"
+"{\n"
+"	u32 blocksum;\n"
+"    int offset = 1;\n"
+"    for(int nActive=n>>1; nActive>0; nActive>>=1, offset<<=1)\n"
+"    {\n"
+"        GROUP_LDS_BARRIER;\n"
+"        for(int iIdx=lIdx; iIdx<nActive; iIdx+=lSize)\n"
+"        {\n"
+"            int ai = offset*(2*iIdx+1)-1;\n"
+"            int bi = offset*(2*iIdx+2)-1;\n"
+"            data[bi] += data[ai];\n"
+"        }\n"
+"	}\n"
+"    GROUP_LDS_BARRIER;\n"
+"    if( lIdx == 0 )\n"
+"	{\n"
+"		blocksum = data[ n-1 ];\n"
+"        data[ n-1 ] = 0;\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	offset >>= 1;\n"
+"    for(int nActive=1; nActive<n; nActive<<=1, offset>>=1 )\n"
+"    {\n"
+"        GROUP_LDS_BARRIER;\n"
+"        for( int iIdx = lIdx; iIdx<nActive; iIdx += lSize )\n"
+"        {\n"
+"            int ai = offset*(2*iIdx+1)-1;\n"
+"            int bi = offset*(2*iIdx+2)-1;\n"
+"            u32 temp = data[ai];\n"
+"            data[ai] = data[bi];\n"
+"            data[bi] += temp;\n"
+"        }\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	return blocksum;\n"
+"}\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"__kernel\n"
+"void LocalScanKernel(__global u32* dst, __global u32 *src, __global u32 *sumBuffer,\n"
+"		uint4 cb)\n"
+"{\n"
+"	__local u32 ldsData[WG_SIZE*2];\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	ldsData[2*lIdx]     = ( 2*gIdx < cb.m_numElems )? src[2*gIdx]: 0;\n"
+"	ldsData[2*lIdx + 1] = ( 2*gIdx+1 < cb.m_numElems )? src[2*gIdx + 1]: 0;\n"
+"	u32 sum = ScanExclusive(ldsData, WG_SIZE*2, GET_LOCAL_IDX, GET_GROUP_SIZE);\n"
+"	if( lIdx == 0 ) sumBuffer[GET_GROUP_IDX] = sum;\n"
+"	if( (2*gIdx) < cb.m_numElems )\n"
+"    {\n"
+"        dst[2*gIdx]     = ldsData[2*lIdx];\n"
+"	}\n"
+"	if( (2*gIdx + 1) < cb.m_numElems )\n"
+"	{\n"
+"        dst[2*gIdx + 1] = ldsData[2*lIdx + 1];\n"
+"    }\n"
+"}\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"__kernel\n"
+"void AddOffsetKernel(__global u32 *dst, __global u32 *blockSum, uint4 cb)\n"
+"{\n"
+"	const u32 blockSize = WG_SIZE*2;\n"
+"	int myIdx = GET_GROUP_IDX+1;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	u32 iBlockSum = blockSum[myIdx];\n"
+"	int endValue = min((myIdx+1)*(blockSize), cb.m_numElems);\n"
+"	for(int i=myIdx*blockSize+lIdx; i<endValue; i+=GET_GROUP_SIZE)\n"
+"	{\n"
+"		dst[i] += iBlockSum;\n"
+"	}\n"
+"}\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"__kernel\n"
+"void TopLevelScanKernel(__global u32* dst, uint4 cb)\n"
+"{\n"
+"	__local u32 ldsData[2048];\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	int lSize = GET_GROUP_SIZE;\n"
+"	for(int i=lIdx; i<cb.m_numScanBlocks; i+=lSize )\n"
+"	{\n"
+"		ldsData[i] = (i<cb.m_numBlocks)? dst[i]:0;\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	u32 sum = ScanExclusive(ldsData, cb.m_numScanBlocks, GET_LOCAL_IDX, GET_GROUP_SIZE);\n"
+"	for(int i=lIdx; i<cb.m_numBlocks; i+=lSize )\n"
+"	{\n"
+"		dst[i] = ldsData[i];\n"
+"	}\n"
+"	if( gIdx == 0 )\n"
+"	{\n"
+"		dst[cb.m_numBlocks] = sum;\n"
+"	}\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanKernelsFloat4CL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanKernelsFloat4CL.h
new file mode 100644
index 0000000000..5b13254796
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/PrefixScanKernelsFloat4CL.h
@@ -0,0 +1,129 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* prefixScanKernelsFloat4CL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"typedef unsigned int u32;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"// takahiro end\n"
+"#define WG_SIZE 128 \n"
+"#define m_numElems x\n"
+"#define m_numBlocks y\n"
+"#define m_numScanBlocks z\n"
+"/*typedef struct\n"
+"{\n"
+"	uint m_numElems;\n"
+"	uint m_numBlocks;\n"
+"	uint m_numScanBlocks;\n"
+"	uint m_padding[1];\n"
+"} ConstBuffer;\n"
+"*/\n"
+"float4 ScanExclusiveFloat4(__local float4* data, u32 n, int lIdx, int lSize)\n"
+"{\n"
+"	float4 blocksum;\n"
+"    int offset = 1;\n"
+"    for(int nActive=n>>1; nActive>0; nActive>>=1, offset<<=1)\n"
+"    {\n"
+"        GROUP_LDS_BARRIER;\n"
+"        for(int iIdx=lIdx; iIdx<nActive; iIdx+=lSize)\n"
+"        {\n"
+"            int ai = offset*(2*iIdx+1)-1;\n"
+"            int bi = offset*(2*iIdx+2)-1;\n"
+"            data[bi] += data[ai];\n"
+"        }\n"
+"	}\n"
+"    GROUP_LDS_BARRIER;\n"
+"    if( lIdx == 0 )\n"
+"	{\n"
+"		blocksum = data[ n-1 ];\n"
+"    data[ n-1 ] = 0;\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	offset >>= 1;\n"
+"    for(int nActive=1; nActive<n; nActive<<=1, offset>>=1 )\n"
+"    {\n"
+"        GROUP_LDS_BARRIER;\n"
+"        for( int iIdx = lIdx; iIdx<nActive; iIdx += lSize )\n"
+"        {\n"
+"            int ai = offset*(2*iIdx+1)-1;\n"
+"            int bi = offset*(2*iIdx+2)-1;\n"
+"            float4 temp = data[ai];\n"
+"            data[ai] = data[bi];\n"
+"            data[bi] += temp;\n"
+"        }\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	return blocksum;\n"
+"}\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"__kernel\n"
+"void LocalScanKernel(__global float4* dst, __global float4* src, __global float4* sumBuffer,	uint4 cb)\n"
+"{\n"
+"	__local float4 ldsData[WG_SIZE*2];\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	ldsData[2*lIdx]     = ( 2*gIdx < cb.m_numElems )? src[2*gIdx]: 0;\n"
+"	ldsData[2*lIdx + 1] = ( 2*gIdx+1 < cb.m_numElems )? src[2*gIdx + 1]: 0;\n"
+"	float4 sum = ScanExclusiveFloat4(ldsData, WG_SIZE*2, GET_LOCAL_IDX, GET_GROUP_SIZE);\n"
+"	if( lIdx == 0 ) \n"
+"		sumBuffer[GET_GROUP_IDX] = sum;\n"
+"	if( (2*gIdx) < cb.m_numElems )\n"
+"    {\n"
+"        dst[2*gIdx]     = ldsData[2*lIdx];\n"
+"	}\n"
+"	if( (2*gIdx + 1) < cb.m_numElems )\n"
+"	{\n"
+"        dst[2*gIdx + 1] = ldsData[2*lIdx + 1];\n"
+"    }\n"
+"}\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"__kernel\n"
+"void AddOffsetKernel(__global float4* dst, __global float4* blockSum, uint4 cb)\n"
+"{\n"
+"	const u32 blockSize = WG_SIZE*2;\n"
+"	int myIdx = GET_GROUP_IDX+1;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	float4 iBlockSum = blockSum[myIdx];\n"
+"	int endValue = min((myIdx+1)*(blockSize), cb.m_numElems);\n"
+"	for(int i=myIdx*blockSize+lIdx; i<endValue; i+=GET_GROUP_SIZE)\n"
+"	{\n"
+"		dst[i] += iBlockSum;\n"
+"	}\n"
+"}\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"__kernel\n"
+"void TopLevelScanKernel(__global float4* dst, uint4 cb)\n"
+"{\n"
+"	__local float4 ldsData[2048];\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	int lSize = GET_GROUP_SIZE;\n"
+"	for(int i=lIdx; i<cb.m_numScanBlocks; i+=lSize )\n"
+"	{\n"
+"		ldsData[i] = (i<cb.m_numBlocks)? dst[i]:0;\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	float4 sum = ScanExclusiveFloat4(ldsData, cb.m_numScanBlocks, GET_LOCAL_IDX, GET_GROUP_SIZE);\n"
+"	for(int i=lIdx; i<cb.m_numBlocks; i+=lSize )\n"
+"	{\n"
+"		dst[i] = ldsData[i];\n"
+"	}\n"
+"	if( gIdx == 0 )\n"
+"	{\n"
+"		dst[cb.m_numBlocks] = sum;\n"
+"	}\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/RadixSort32Kernels.cl b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/RadixSort32Kernels.cl
new file mode 100644
index 0000000000..7402e2f3b3
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/RadixSort32Kernels.cl
@@ -0,0 +1,1071 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2011 Advanced Micro Devices, Inc.  http://bulletphysics.org
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Author Takahiro Harada
+
+
+//#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+
+typedef unsigned int u32;
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AtomAdd(x, value) atom_add(&(x), value)
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+
+#define make_uint4 (uint4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+#define WG_SIZE 64
+#define ELEMENTS_PER_WORK_ITEM (256/WG_SIZE)
+#define BITS_PER_PASS 4
+#define NUM_BUCKET (1<<BITS_PER_PASS)
+typedef uchar u8;
+
+//	this isn't optimization for VLIW. But just reducing writes. 
+#define USE_2LEVEL_REDUCE 1
+
+//#define CHECK_BOUNDARY 1
+
+//#define NV_GPU 1
+
+
+//	Cypress
+#define nPerWI 16
+//	Cayman
+//#define nPerWI 20
+
+#define m_n x
+#define m_nWGs y
+#define m_startBit z
+#define m_nBlocksPerWG w
+
+/*
+typedef struct
+{
+	int m_n;
+	int m_nWGs;
+	int m_startBit;
+	int m_nBlocksPerWG;
+} ConstBuffer;
+*/
+
+typedef struct
+{
+	unsigned int m_key;
+	unsigned int m_value;
+} SortDataCL;
+
+
+uint prefixScanVectorEx( uint4* data )
+{
+	u32 sum = 0;
+	u32 tmp = data[0].x;
+	data[0].x = sum;
+	sum += tmp;
+	tmp = data[0].y;
+	data[0].y = sum;
+	sum += tmp;
+	tmp = data[0].z;
+	data[0].z = sum;
+	sum += tmp;
+	tmp = data[0].w;
+	data[0].w = sum;
+	sum += tmp;
+	return sum;
+}
+
+u32 localPrefixSum( u32 pData, uint lIdx, uint* totalSum, __local u32* sorterSharedMemory, int wgSize /*64 or 128*/ )
+{
+	{	//	Set data
+		sorterSharedMemory[lIdx] = 0;
+		sorterSharedMemory[lIdx+wgSize] = pData;
+	}
+
+	GROUP_LDS_BARRIER;
+
+	{	//	Prefix sum
+		int idx = 2*lIdx + (wgSize+1);
+#if defined(USE_2LEVEL_REDUCE)
+		if( lIdx < 64 )
+		{
+			u32 u0, u1, u2;
+			u0 = sorterSharedMemory[idx-3];
+			u1 = sorterSharedMemory[idx-2];
+			u2 = sorterSharedMemory[idx-1];
+			AtomAdd( sorterSharedMemory[idx], u0+u1+u2 );			
+			GROUP_MEM_FENCE;
+
+			u0 = sorterSharedMemory[idx-12];
+			u1 = sorterSharedMemory[idx-8];
+			u2 = sorterSharedMemory[idx-4];
+			AtomAdd( sorterSharedMemory[idx], u0+u1+u2 );			
+			GROUP_MEM_FENCE;
+
+			u0 = sorterSharedMemory[idx-48];
+			u1 = sorterSharedMemory[idx-32];
+			u2 = sorterSharedMemory[idx-16];
+			AtomAdd( sorterSharedMemory[idx], u0+u1+u2 );			
+			GROUP_MEM_FENCE;
+			if( wgSize > 64 )
+			{
+				sorterSharedMemory[idx] += sorterSharedMemory[idx-64];
+				GROUP_MEM_FENCE;
+			}
+
+			sorterSharedMemory[idx-1] += sorterSharedMemory[idx-2];
+			GROUP_MEM_FENCE;
+		}
+#else
+		if( lIdx < 64 )
+		{
+			sorterSharedMemory[idx] += sorterSharedMemory[idx-1];
+			GROUP_MEM_FENCE;
+			sorterSharedMemory[idx] += sorterSharedMemory[idx-2];			
+			GROUP_MEM_FENCE;
+			sorterSharedMemory[idx] += sorterSharedMemory[idx-4];
+			GROUP_MEM_FENCE;
+			sorterSharedMemory[idx] += sorterSharedMemory[idx-8];
+			GROUP_MEM_FENCE;
+			sorterSharedMemory[idx] += sorterSharedMemory[idx-16];
+			GROUP_MEM_FENCE;
+			sorterSharedMemory[idx] += sorterSharedMemory[idx-32];
+			GROUP_MEM_FENCE;
+			if( wgSize > 64 )
+			{
+				sorterSharedMemory[idx] += sorterSharedMemory[idx-64];
+				GROUP_MEM_FENCE;
+			}
+
+			sorterSharedMemory[idx-1] += sorterSharedMemory[idx-2];
+			GROUP_MEM_FENCE;
+		}
+#endif
+	}
+
+	GROUP_LDS_BARRIER;
+
+	*totalSum = sorterSharedMemory[wgSize*2-1];
+	u32 addValue = sorterSharedMemory[lIdx+wgSize-1];
+	return addValue;
+}
+
+//__attribute__((reqd_work_group_size(128,1,1)))
+uint4 localPrefixSum128V( uint4 pData, uint lIdx, uint* totalSum, __local u32* sorterSharedMemory )
+{
+	u32 s4 = prefixScanVectorEx( &pData );
+	u32 rank = localPrefixSum( s4, lIdx, totalSum, sorterSharedMemory, 128 );
+	return pData + make_uint4( rank, rank, rank, rank );
+}
+
+
+//__attribute__((reqd_work_group_size(64,1,1)))
+uint4 localPrefixSum64V( uint4 pData, uint lIdx, uint* totalSum, __local u32* sorterSharedMemory )
+{
+	u32 s4 = prefixScanVectorEx( &pData );
+	u32 rank = localPrefixSum( s4, lIdx, totalSum, sorterSharedMemory, 64 );
+	return pData + make_uint4( rank, rank, rank, rank );
+}
+
+u32 unpack4Key( u32 key, int keyIdx ){ return (key>>(keyIdx*8)) & 0xff;}
+
+u32 bit8Scan(u32 v)
+{
+	return (v<<8) + (v<<16) + (v<<24);
+}
+
+//===
+
+
+
+
+#define MY_HISTOGRAM(idx) localHistogramMat[(idx)*WG_SIZE+lIdx]
+
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void StreamCountKernel( __global u32* gSrc, __global u32* histogramOut, int4 cb )
+{
+	__local u32 localHistogramMat[NUM_BUCKET*WG_SIZE];
+
+	u32 gIdx = GET_GLOBAL_IDX;
+	u32 lIdx = GET_LOCAL_IDX;
+	u32 wgIdx = GET_GROUP_IDX;
+	u32 wgSize = GET_GROUP_SIZE;
+	const int startBit = cb.m_startBit;
+	const int n = cb.m_n;
+	const int nWGs = cb.m_nWGs;
+	const int nBlocksPerWG = cb.m_nBlocksPerWG;
+
+	for(int i=0; i<NUM_BUCKET; i++)
+	{
+		MY_HISTOGRAM(i) = 0;
+	}
+
+	GROUP_LDS_BARRIER;
+
+	const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;
+	u32 localKey;
+
+	int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;
+
+	int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;
+
+	for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)
+	{
+		//	MY_HISTOGRAM( localKeys.x ) ++ is much expensive than atomic add as it requires read and write while atomics can just add on AMD
+		//	Using registers didn't perform well. It seems like use localKeys to address requires a lot of alu ops
+		//	AMD: AtomInc performs better while NV prefers ++
+		for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)
+		{
+#if defined(CHECK_BOUNDARY)
+			if( addr+i < n )
+#endif
+			{
+				localKey = (gSrc[addr+i]>>startBit) & 0xf;
+#if defined(NV_GPU)
+				MY_HISTOGRAM( localKey )++;
+#else
+				AtomInc( MY_HISTOGRAM( localKey ) );
+#endif
+			}
+		}
+	}
+
+	GROUP_LDS_BARRIER;
+	
+	if( lIdx < NUM_BUCKET )
+	{
+		u32 sum = 0;
+		for(int i=0; i<GET_GROUP_SIZE; i++)
+		{
+			sum += localHistogramMat[lIdx*WG_SIZE+(i+lIdx)%GET_GROUP_SIZE];
+		}
+		histogramOut[lIdx*nWGs+wgIdx] = sum;
+	}
+}
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void StreamCountSortDataKernel( __global SortDataCL* gSrc, __global u32* histogramOut, int4  cb )
+{
+	__local u32 localHistogramMat[NUM_BUCKET*WG_SIZE];
+
+	u32 gIdx = GET_GLOBAL_IDX;
+	u32 lIdx = GET_LOCAL_IDX;
+	u32 wgIdx = GET_GROUP_IDX;
+	u32 wgSize = GET_GROUP_SIZE;
+	const int startBit = cb.m_startBit;
+	const int n = cb.m_n;
+	const int nWGs = cb.m_nWGs;
+	const int nBlocksPerWG = cb.m_nBlocksPerWG;
+
+	for(int i=0; i<NUM_BUCKET; i++)
+	{
+		MY_HISTOGRAM(i) = 0;
+	}
+
+	GROUP_LDS_BARRIER;
+
+	const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;
+	u32 localKey;
+
+	int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;
+
+	int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;
+
+	for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)
+	{
+		//	MY_HISTOGRAM( localKeys.x ) ++ is much expensive than atomic add as it requires read and write while atomics can just add on AMD
+		//	Using registers didn't perform well. It seems like use localKeys to address requires a lot of alu ops
+		//	AMD: AtomInc performs better while NV prefers ++
+		for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)
+		{
+#if defined(CHECK_BOUNDARY)
+			if( addr+i < n )
+#endif
+			{
+				localKey = (gSrc[addr+i].m_key>>startBit) & 0xf;
+#if defined(NV_GPU)
+				MY_HISTOGRAM( localKey )++;
+#else
+				AtomInc( MY_HISTOGRAM( localKey ) );
+#endif
+			}
+		}
+	}
+
+	GROUP_LDS_BARRIER;
+	
+	if( lIdx < NUM_BUCKET )
+	{
+		u32 sum = 0;
+		for(int i=0; i<GET_GROUP_SIZE; i++)
+		{
+			sum += localHistogramMat[lIdx*WG_SIZE+(i+lIdx)%GET_GROUP_SIZE];
+		}
+		histogramOut[lIdx*nWGs+wgIdx] = sum;
+	}
+}
+
+#define nPerLane (nPerWI/4)
+
+//	NUM_BUCKET*nWGs < 128*nPerWI
+__kernel
+__attribute__((reqd_work_group_size(128,1,1)))
+void PrefixScanKernel( __global u32* wHistogram1, int4  cb )
+{
+	__local u32 ldsTopScanData[128*2];
+
+	u32 lIdx = GET_LOCAL_IDX;
+	u32 wgIdx = GET_GROUP_IDX;
+	const int nWGs = cb.m_nWGs;
+
+	u32 data[nPerWI];
+	for(int i=0; i<nPerWI; i++)
+	{
+		data[i] = 0;
+		if( (nPerWI*lIdx+i) < NUM_BUCKET*nWGs )
+			data[i] = wHistogram1[nPerWI*lIdx+i];
+	}
+
+	uint4 myData = make_uint4(0,0,0,0);
+
+	for(int i=0; i<nPerLane; i++)
+	{
+		myData.x += data[nPerLane*0+i];
+		myData.y += data[nPerLane*1+i];
+		myData.z += data[nPerLane*2+i];
+		myData.w += data[nPerLane*3+i];
+	}
+
+	uint totalSum;
+	uint4 scanned = localPrefixSum128V( myData, lIdx, &totalSum, ldsTopScanData );
+
+//	for(int j=0; j<4; j++) //	somehow it introduces a lot of branches
+	{	int j = 0;
+		u32 sum = 0;
+		for(int i=0; i<nPerLane; i++)
+		{
+			u32 tmp = data[nPerLane*j+i];
+			data[nPerLane*j+i] = sum;
+			sum += tmp;
+		}
+	}
+	{	int j = 1;
+		u32 sum = 0;
+		for(int i=0; i<nPerLane; i++)
+		{
+			u32 tmp = data[nPerLane*j+i];
+			data[nPerLane*j+i] = sum;
+			sum += tmp;
+		}
+	}
+	{	int j = 2;
+		u32 sum = 0;
+		for(int i=0; i<nPerLane; i++)
+		{
+			u32 tmp = data[nPerLane*j+i];
+			data[nPerLane*j+i] = sum;
+			sum += tmp;
+		}
+	}
+	{	int j = 3;
+		u32 sum = 0;
+		for(int i=0; i<nPerLane; i++)
+		{
+			u32 tmp = data[nPerLane*j+i];
+			data[nPerLane*j+i] = sum;
+			sum += tmp;
+		}
+	}
+
+	for(int i=0; i<nPerLane; i++)
+	{
+		data[nPerLane*0+i] += scanned.x;
+		data[nPerLane*1+i] += scanned.y;
+		data[nPerLane*2+i] += scanned.z;
+		data[nPerLane*3+i] += scanned.w;
+	}
+
+	for(int i=0; i<nPerWI; i++)
+	{
+		int index = nPerWI*lIdx+i;
+		if (index < NUM_BUCKET*nWGs)
+			wHistogram1[nPerWI*lIdx+i] = data[i];
+	}
+}
+
+//	4 scan, 4 exchange
+void sort4Bits(u32 sortData[4], int startBit, int lIdx, __local u32* ldsSortData)
+{
+	for(int bitIdx=0; bitIdx<BITS_PER_PASS; bitIdx++)
+	{
+		u32 mask = (1<<bitIdx);
+		uint4 cmpResult = make_uint4( (sortData[0]>>startBit) & mask, (sortData[1]>>startBit) & mask, (sortData[2]>>startBit) & mask, (sortData[3]>>startBit) & mask );
+		uint4 prefixSum = SELECT_UINT4( make_uint4(1,1,1,1), make_uint4(0,0,0,0), cmpResult != make_uint4(0,0,0,0) );
+		u32 total;
+		prefixSum = localPrefixSum64V( prefixSum, lIdx, &total, ldsSortData );
+		{
+			uint4 localAddr = make_uint4(lIdx*4+0,lIdx*4+1,lIdx*4+2,lIdx*4+3);
+			uint4 dstAddr = localAddr - prefixSum + make_uint4( total, total, total, total );
+			dstAddr = SELECT_UINT4( prefixSum, dstAddr, cmpResult != make_uint4(0, 0, 0, 0) );
+
+			GROUP_LDS_BARRIER;
+
+			ldsSortData[dstAddr.x] = sortData[0];
+			ldsSortData[dstAddr.y] = sortData[1];
+			ldsSortData[dstAddr.z] = sortData[2];
+			ldsSortData[dstAddr.w] = sortData[3];
+
+			GROUP_LDS_BARRIER;
+
+			sortData[0] = ldsSortData[localAddr.x];
+			sortData[1] = ldsSortData[localAddr.y];
+			sortData[2] = ldsSortData[localAddr.z];
+			sortData[3] = ldsSortData[localAddr.w];
+
+			GROUP_LDS_BARRIER;
+		}
+	}
+}
+
+//	2 scan, 2 exchange
+void sort4Bits1(u32 sortData[4], int startBit, int lIdx, __local u32* ldsSortData)
+{
+	for(uint ibit=0; ibit<BITS_PER_PASS; ibit+=2)
+	{
+		uint4 b = make_uint4((sortData[0]>>(startBit+ibit)) & 0x3, 
+			(sortData[1]>>(startBit+ibit)) & 0x3, 
+			(sortData[2]>>(startBit+ibit)) & 0x3, 
+			(sortData[3]>>(startBit+ibit)) & 0x3);
+
+		u32 key4;
+		u32 sKeyPacked[4] = { 0, 0, 0, 0 };
+		{
+			sKeyPacked[0] |= 1<<(8*b.x);
+			sKeyPacked[1] |= 1<<(8*b.y);
+			sKeyPacked[2] |= 1<<(8*b.z);
+			sKeyPacked[3] |= 1<<(8*b.w);
+
+			key4 = sKeyPacked[0] + sKeyPacked[1] + sKeyPacked[2] + sKeyPacked[3];
+		}
+
+		u32 rankPacked;
+		u32 sumPacked;
+		{
+			rankPacked = localPrefixSum( key4, lIdx, &sumPacked, ldsSortData, WG_SIZE );
+		}
+
+		GROUP_LDS_BARRIER;
+
+		u32 newOffset[4] = { 0,0,0,0 };
+		{
+			u32 sumScanned = bit8Scan( sumPacked );
+
+			u32 scannedKeys[4];
+			scannedKeys[0] = 1<<(8*b.x);
+			scannedKeys[1] = 1<<(8*b.y);
+			scannedKeys[2] = 1<<(8*b.z);
+			scannedKeys[3] = 1<<(8*b.w);
+			{	//	4 scans at once
+				u32 sum4 = 0;
+				for(int ie=0; ie<4; ie++)
+				{
+					u32 tmp = scannedKeys[ie];
+					scannedKeys[ie] = sum4;
+					sum4 += tmp;
+				}
+			}
+
+			{
+				u32 sumPlusRank = sumScanned + rankPacked;
+				{	u32 ie = b.x;
+					scannedKeys[0] += sumPlusRank;
+					newOffset[0] = unpack4Key( scannedKeys[0], ie );
+				}
+				{	u32 ie = b.y;
+					scannedKeys[1] += sumPlusRank;
+					newOffset[1] = unpack4Key( scannedKeys[1], ie );
+				}
+				{	u32 ie = b.z;
+					scannedKeys[2] += sumPlusRank;
+					newOffset[2] = unpack4Key( scannedKeys[2], ie );
+				}
+				{	u32 ie = b.w;
+					scannedKeys[3] += sumPlusRank;
+					newOffset[3] = unpack4Key( scannedKeys[3], ie );
+				}
+			}
+		}
+
+
+		GROUP_LDS_BARRIER;
+
+		{
+			ldsSortData[newOffset[0]] = sortData[0];
+			ldsSortData[newOffset[1]] = sortData[1];
+			ldsSortData[newOffset[2]] = sortData[2];
+			ldsSortData[newOffset[3]] = sortData[3];
+
+			GROUP_LDS_BARRIER;
+
+			u32 dstAddr = 4*lIdx;
+			sortData[0] = ldsSortData[dstAddr+0];
+			sortData[1] = ldsSortData[dstAddr+1];
+			sortData[2] = ldsSortData[dstAddr+2];
+			sortData[3] = ldsSortData[dstAddr+3];
+
+			GROUP_LDS_BARRIER;
+		}
+	}
+}
+
+#define SET_HISTOGRAM(setIdx, key) ldsSortData[(setIdx)*NUM_BUCKET+key]
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void SortAndScatterKernel( __global const u32* restrict gSrc, __global const u32* rHistogram, __global u32* restrict gDst, int4  cb )
+{
+	__local u32 ldsSortData[WG_SIZE*ELEMENTS_PER_WORK_ITEM+16];
+	__local u32 localHistogramToCarry[NUM_BUCKET];
+	__local u32 localHistogram[NUM_BUCKET*2];
+
+	u32 gIdx = GET_GLOBAL_IDX;
+	u32 lIdx = GET_LOCAL_IDX;
+	u32 wgIdx = GET_GROUP_IDX;
+	u32 wgSize = GET_GROUP_SIZE;
+
+	const int n = cb.m_n;
+	const int nWGs = cb.m_nWGs;
+	const int startBit = cb.m_startBit;
+	const int nBlocksPerWG = cb.m_nBlocksPerWG;
+
+	if( lIdx < (NUM_BUCKET) )
+	{
+		localHistogramToCarry[lIdx] = rHistogram[lIdx*nWGs + wgIdx];
+	}
+
+	GROUP_LDS_BARRIER;
+
+	const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;
+
+	int nBlocks = n/blockSize - nBlocksPerWG*wgIdx;
+
+	int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;
+
+	for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)
+	{
+		u32 myHistogram = 0;
+
+		u32 sortData[ELEMENTS_PER_WORK_ITEM];
+		for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)
+#if defined(CHECK_BOUNDARY)
+			sortData[i] = ( addr+i < n )? gSrc[ addr+i ] : 0xffffffff;
+#else
+			sortData[i] = gSrc[ addr+i ];
+#endif
+
+		sort4Bits(sortData, startBit, lIdx, ldsSortData);
+
+		u32 keys[ELEMENTS_PER_WORK_ITEM];
+		for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)
+			keys[i] = (sortData[i]>>startBit) & 0xf;
+
+		{	//	create histogram
+			u32 setIdx = lIdx/16;
+			if( lIdx < NUM_BUCKET )
+			{
+				localHistogram[lIdx] = 0;
+			}
+			ldsSortData[lIdx] = 0;
+			GROUP_LDS_BARRIER;
+
+			for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)
+#if defined(CHECK_BOUNDARY)
+				if( addr+i < n )
+#endif
+
+#if defined(NV_GPU)
+				SET_HISTOGRAM( setIdx, keys[i] )++;
+#else
+				AtomInc( SET_HISTOGRAM( setIdx, keys[i] ) );
+#endif
+			
+			GROUP_LDS_BARRIER;
+			
+			uint hIdx = NUM_BUCKET+lIdx;
+			if( lIdx < NUM_BUCKET )
+			{
+				u32 sum = 0;
+				for(int i=0; i<WG_SIZE/16; i++)
+				{
+					sum += SET_HISTOGRAM( i, lIdx );
+				}
+				myHistogram = sum;
+				localHistogram[hIdx] = sum;
+			}
+			GROUP_LDS_BARRIER;
+
+#if defined(USE_2LEVEL_REDUCE)
+			if( lIdx < NUM_BUCKET )
+			{
+				localHistogram[hIdx] = localHistogram[hIdx-1];
+				GROUP_MEM_FENCE;
+
+				u32 u0, u1, u2;
+				u0 = localHistogram[hIdx-3];
+				u1 = localHistogram[hIdx-2];
+				u2 = localHistogram[hIdx-1];
+				AtomAdd( localHistogram[hIdx], u0 + u1 + u2 );
+				GROUP_MEM_FENCE;
+				u0 = localHistogram[hIdx-12];
+				u1 = localHistogram[hIdx-8];
+				u2 = localHistogram[hIdx-4];
+				AtomAdd( localHistogram[hIdx], u0 + u1 + u2 );
+				GROUP_MEM_FENCE;
+			}
+#else
+			if( lIdx < NUM_BUCKET )
+			{
+				localHistogram[hIdx] = localHistogram[hIdx-1];
+				GROUP_MEM_FENCE;
+				localHistogram[hIdx] += localHistogram[hIdx-1];
+				GROUP_MEM_FENCE;
+				localHistogram[hIdx] += localHistogram[hIdx-2];
+				GROUP_MEM_FENCE;
+				localHistogram[hIdx] += localHistogram[hIdx-4];
+				GROUP_MEM_FENCE;
+				localHistogram[hIdx] += localHistogram[hIdx-8];
+				GROUP_MEM_FENCE;
+			}
+#endif
+			GROUP_LDS_BARRIER;
+		}
+
+		{
+			for(int ie=0; ie<ELEMENTS_PER_WORK_ITEM; ie++)
+			{
+				int dataIdx = ELEMENTS_PER_WORK_ITEM*lIdx+ie;
+				int binIdx = keys[ie];
+				int groupOffset = localHistogramToCarry[binIdx];
+				int myIdx = dataIdx - localHistogram[NUM_BUCKET+binIdx];
+#if defined(CHECK_BOUNDARY)
+				if( addr+ie < n )
+#endif
+				gDst[ groupOffset + myIdx ] = sortData[ie];
+			}
+		}
+
+		GROUP_LDS_BARRIER;
+
+		if( lIdx < NUM_BUCKET )
+		{
+			localHistogramToCarry[lIdx] += myHistogram;
+		}
+		GROUP_LDS_BARRIER;
+	}
+}
+
+//	2 scan, 2 exchange
+void sort4Bits1KeyValue(u32 sortData[4], int sortVal[4], int startBit, int lIdx, __local u32* ldsSortData, __local int *ldsSortVal)
+{
+	for(uint ibit=0; ibit<BITS_PER_PASS; ibit+=2)
+	{
+		uint4 b = make_uint4((sortData[0]>>(startBit+ibit)) & 0x3, 
+			(sortData[1]>>(startBit+ibit)) & 0x3, 
+			(sortData[2]>>(startBit+ibit)) & 0x3, 
+			(sortData[3]>>(startBit+ibit)) & 0x3);
+
+		u32 key4;
+		u32 sKeyPacked[4] = { 0, 0, 0, 0 };
+		{
+			sKeyPacked[0] |= 1<<(8*b.x);
+			sKeyPacked[1] |= 1<<(8*b.y);
+			sKeyPacked[2] |= 1<<(8*b.z);
+			sKeyPacked[3] |= 1<<(8*b.w);
+
+			key4 = sKeyPacked[0] + sKeyPacked[1] + sKeyPacked[2] + sKeyPacked[3];
+		}
+
+		u32 rankPacked;
+		u32 sumPacked;
+		{
+			rankPacked = localPrefixSum( key4, lIdx, &sumPacked, ldsSortData, WG_SIZE );
+		}
+
+		GROUP_LDS_BARRIER;
+
+		u32 newOffset[4] = { 0,0,0,0 };
+		{
+			u32 sumScanned = bit8Scan( sumPacked );
+
+			u32 scannedKeys[4];
+			scannedKeys[0] = 1<<(8*b.x);
+			scannedKeys[1] = 1<<(8*b.y);
+			scannedKeys[2] = 1<<(8*b.z);
+			scannedKeys[3] = 1<<(8*b.w);
+			{	//	4 scans at once
+				u32 sum4 = 0;
+				for(int ie=0; ie<4; ie++)
+				{
+					u32 tmp = scannedKeys[ie];
+					scannedKeys[ie] = sum4;
+					sum4 += tmp;
+				}
+			}
+
+			{
+				u32 sumPlusRank = sumScanned + rankPacked;
+				{	u32 ie = b.x;
+					scannedKeys[0] += sumPlusRank;
+					newOffset[0] = unpack4Key( scannedKeys[0], ie );
+				}
+				{	u32 ie = b.y;
+					scannedKeys[1] += sumPlusRank;
+					newOffset[1] = unpack4Key( scannedKeys[1], ie );
+				}
+				{	u32 ie = b.z;
+					scannedKeys[2] += sumPlusRank;
+					newOffset[2] = unpack4Key( scannedKeys[2], ie );
+				}
+				{	u32 ie = b.w;
+					scannedKeys[3] += sumPlusRank;
+					newOffset[3] = unpack4Key( scannedKeys[3], ie );
+				}
+			}
+		}
+
+
+		GROUP_LDS_BARRIER;
+
+		{
+			ldsSortData[newOffset[0]] = sortData[0];
+			ldsSortData[newOffset[1]] = sortData[1];
+			ldsSortData[newOffset[2]] = sortData[2];
+			ldsSortData[newOffset[3]] = sortData[3];
+
+			ldsSortVal[newOffset[0]] = sortVal[0];
+			ldsSortVal[newOffset[1]] = sortVal[1];
+			ldsSortVal[newOffset[2]] = sortVal[2];
+			ldsSortVal[newOffset[3]] = sortVal[3];
+
+			GROUP_LDS_BARRIER;
+
+			u32 dstAddr = 4*lIdx;
+			sortData[0] = ldsSortData[dstAddr+0];
+			sortData[1] = ldsSortData[dstAddr+1];
+			sortData[2] = ldsSortData[dstAddr+2];
+			sortData[3] = ldsSortData[dstAddr+3];
+
+			sortVal[0] = ldsSortVal[dstAddr+0];
+			sortVal[1] = ldsSortVal[dstAddr+1];
+			sortVal[2] = ldsSortVal[dstAddr+2];
+			sortVal[3] = ldsSortVal[dstAddr+3];
+
+			GROUP_LDS_BARRIER;
+		}
+	}
+}
+
+
+
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void SortAndScatterSortDataKernel( __global const SortDataCL* restrict gSrc, __global const u32* rHistogram, __global SortDataCL* restrict gDst, int4 cb)
+{
+	__local int ldsSortData[WG_SIZE*ELEMENTS_PER_WORK_ITEM+16];
+	__local int ldsSortVal[WG_SIZE*ELEMENTS_PER_WORK_ITEM+16];
+	__local u32 localHistogramToCarry[NUM_BUCKET];
+	__local u32 localHistogram[NUM_BUCKET*2];
+
+	u32 gIdx = GET_GLOBAL_IDX;
+	u32 lIdx = GET_LOCAL_IDX;
+	u32 wgIdx = GET_GROUP_IDX;
+	u32 wgSize = GET_GROUP_SIZE;
+
+	const int n = cb.m_n;
+	const int nWGs = cb.m_nWGs;
+	const int startBit = cb.m_startBit;
+	const int nBlocksPerWG = cb.m_nBlocksPerWG;
+
+	if( lIdx < (NUM_BUCKET) )
+	{
+		localHistogramToCarry[lIdx] = rHistogram[lIdx*nWGs + wgIdx];
+	}
+
+	GROUP_LDS_BARRIER;
+    
+
+	const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;
+
+	int nBlocks = n/blockSize - nBlocksPerWG*wgIdx;
+
+	int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;
+
+	for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)
+	{
+
+		u32 myHistogram = 0;
+
+		int sortData[ELEMENTS_PER_WORK_ITEM];
+		int sortVal[ELEMENTS_PER_WORK_ITEM];
+
+		for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)
+#if defined(CHECK_BOUNDARY)
+		{
+			sortData[i] = ( addr+i < n )? gSrc[ addr+i ].m_key : 0xffffffff;
+			sortVal[i] = ( addr+i < n )? gSrc[ addr+i ].m_value : 0xffffffff;
+		}
+#else
+		{
+			sortData[i] = gSrc[ addr+i ].m_key;
+			sortVal[i] = gSrc[ addr+i ].m_value;
+		}
+#endif
+
+		sort4Bits1KeyValue(sortData, sortVal, startBit, lIdx, ldsSortData, ldsSortVal);
+
+		u32 keys[ELEMENTS_PER_WORK_ITEM];
+		for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)
+			keys[i] = (sortData[i]>>startBit) & 0xf;
+
+		{	//	create histogram
+			u32 setIdx = lIdx/16;
+			if( lIdx < NUM_BUCKET )
+			{
+				localHistogram[lIdx] = 0;
+			}
+			ldsSortData[lIdx] = 0;
+			GROUP_LDS_BARRIER;
+
+			for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)
+#if defined(CHECK_BOUNDARY)
+				if( addr+i < n )
+#endif
+
+#if defined(NV_GPU)
+				SET_HISTOGRAM( setIdx, keys[i] )++;
+#else
+				AtomInc( SET_HISTOGRAM( setIdx, keys[i] ) );
+#endif
+			
+			GROUP_LDS_BARRIER;
+			
+			uint hIdx = NUM_BUCKET+lIdx;
+			if( lIdx < NUM_BUCKET )
+			{
+				u32 sum = 0;
+				for(int i=0; i<WG_SIZE/16; i++)
+				{
+					sum += SET_HISTOGRAM( i, lIdx );
+				}
+				myHistogram = sum;
+				localHistogram[hIdx] = sum;
+			}
+			GROUP_LDS_BARRIER;
+
+#if defined(USE_2LEVEL_REDUCE)
+			if( lIdx < NUM_BUCKET )
+			{
+				localHistogram[hIdx] = localHistogram[hIdx-1];
+				GROUP_MEM_FENCE;
+
+				u32 u0, u1, u2;
+				u0 = localHistogram[hIdx-3];
+				u1 = localHistogram[hIdx-2];
+				u2 = localHistogram[hIdx-1];
+				AtomAdd( localHistogram[hIdx], u0 + u1 + u2 );
+				GROUP_MEM_FENCE;
+				u0 = localHistogram[hIdx-12];
+				u1 = localHistogram[hIdx-8];
+				u2 = localHistogram[hIdx-4];
+				AtomAdd( localHistogram[hIdx], u0 + u1 + u2 );
+				GROUP_MEM_FENCE;
+			}
+#else
+			if( lIdx < NUM_BUCKET )
+			{
+				localHistogram[hIdx] = localHistogram[hIdx-1];
+				GROUP_MEM_FENCE;
+				localHistogram[hIdx] += localHistogram[hIdx-1];
+				GROUP_MEM_FENCE;
+				localHistogram[hIdx] += localHistogram[hIdx-2];
+				GROUP_MEM_FENCE;
+				localHistogram[hIdx] += localHistogram[hIdx-4];
+				GROUP_MEM_FENCE;
+				localHistogram[hIdx] += localHistogram[hIdx-8];
+				GROUP_MEM_FENCE;
+			}
+#endif
+			GROUP_LDS_BARRIER;
+		}
+
+    	{
+			for(int ie=0; ie<ELEMENTS_PER_WORK_ITEM; ie++)
+			{
+				int dataIdx = ELEMENTS_PER_WORK_ITEM*lIdx+ie;
+				int binIdx = keys[ie];
+				int groupOffset = localHistogramToCarry[binIdx];
+				int myIdx = dataIdx - localHistogram[NUM_BUCKET+binIdx];
+#if defined(CHECK_BOUNDARY)
+				if( addr+ie < n )
+				{
+                    if ((groupOffset + myIdx)<n)
+                    {
+                        if (sortData[ie]==sortVal[ie])
+                        {
+                            
+                            SortDataCL tmp;
+                            tmp.m_key = sortData[ie];
+                            tmp.m_value = sortVal[ie];
+                            if (tmp.m_key == tmp.m_value)
+                                gDst[groupOffset + myIdx ] = tmp;
+                        }
+                        
+                    }
+				}
+#else
+                if ((groupOffset + myIdx)<n)
+                {
+                    gDst[ groupOffset + myIdx ].m_key = sortData[ie];
+                    gDst[ groupOffset + myIdx ].m_value = sortVal[ie];
+                }
+#endif
+			}
+		}
+
+		GROUP_LDS_BARRIER;
+
+		if( lIdx < NUM_BUCKET )
+		{
+			localHistogramToCarry[lIdx] += myHistogram;
+		}
+		GROUP_LDS_BARRIER;
+	}
+}
+
+
+
+
+
+
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void SortAndScatterSortDataKernelSerial( __global const SortDataCL* restrict gSrc, __global const u32* rHistogram, __global SortDataCL* restrict gDst, int4 cb)
+{
+    
+	u32 gIdx = GET_GLOBAL_IDX;
+	u32 realLocalIdx = GET_LOCAL_IDX;
+	u32 wgIdx = GET_GROUP_IDX;
+	u32 wgSize = GET_GROUP_SIZE;
+	const int startBit = cb.m_startBit;
+	const int n = cb.m_n;
+	const int nWGs = cb.m_nWGs;
+	const int nBlocksPerWG = cb.m_nBlocksPerWG;
+
+    int counter[NUM_BUCKET];
+    
+    if (realLocalIdx>0)
+        return;
+    
+    for (int c=0;c<NUM_BUCKET;c++)
+        counter[c]=0;
+
+    const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;
+	
+	int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;
+
+   for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++)
+  {
+     for (int lIdx=0;lIdx<WG_SIZE;lIdx++)
+ 	{
+        int addr2 = iblock*blockSize + blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;
+        
+		for(int j=0; j<ELEMENTS_PER_WORK_ITEM; j++)
+		{
+            int i = addr2+j;
+			if( i < n )
+			{
+                int tableIdx;
+				tableIdx = (gSrc[i].m_key>>startBit) & 0xf;//0xf = NUM_TABLES-1
+                gDst[rHistogram[tableIdx*nWGs+wgIdx] + counter[tableIdx]] = gSrc[i];
+                counter[tableIdx] ++;
+			}
+		}
+	}
+  }
+    
+}
+
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void SortAndScatterKernelSerial( __global const u32* restrict gSrc, __global const u32* rHistogram, __global u32* restrict gDst, int4  cb )
+{
+    
+	u32 gIdx = GET_GLOBAL_IDX;
+	u32 realLocalIdx = GET_LOCAL_IDX;
+	u32 wgIdx = GET_GROUP_IDX;
+	u32 wgSize = GET_GROUP_SIZE;
+	const int startBit = cb.m_startBit;
+	const int n = cb.m_n;
+	const int nWGs = cb.m_nWGs;
+	const int nBlocksPerWG = cb.m_nBlocksPerWG;
+
+    int counter[NUM_BUCKET];
+    
+    if (realLocalIdx>0)
+        return;
+    
+    for (int c=0;c<NUM_BUCKET;c++)
+        counter[c]=0;
+
+    const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;
+	
+	int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;
+
+   for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++)
+  {
+     for (int lIdx=0;lIdx<WG_SIZE;lIdx++)
+ 	{
+        int addr2 = iblock*blockSize + blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;
+        
+		for(int j=0; j<ELEMENTS_PER_WORK_ITEM; j++)
+		{
+            int i = addr2+j;
+			if( i < n )
+			{
+                int tableIdx;
+				tableIdx = (gSrc[i]>>startBit) & 0xf;//0xf = NUM_TABLES-1
+                gDst[rHistogram[tableIdx*nWGs+wgIdx] + counter[tableIdx]] = gSrc[i];
+                counter[tableIdx] ++;
+			}
+		}
+	}
+  }
+    
+}
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/RadixSort32KernelsCL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/RadixSort32KernelsCL.h
new file mode 100644
index 0000000000..8876c16aa6
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/kernels/RadixSort32KernelsCL.h
@@ -0,0 +1,910 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* radixSort32KernelsCL= \
+"/*\n"
+"Bullet Continuous Collision Detection and Physics Library\n"
+"Copyright (c) 2011 Advanced Micro Devices, Inc.  http://bulletphysics.org\n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Author Takahiro Harada\n"
+"//#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"typedef unsigned int u32;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"#define make_uint4 (uint4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"#define WG_SIZE 64\n"
+"#define ELEMENTS_PER_WORK_ITEM (256/WG_SIZE)\n"
+"#define BITS_PER_PASS 4\n"
+"#define NUM_BUCKET (1<<BITS_PER_PASS)\n"
+"typedef uchar u8;\n"
+"//	this isn't optimization for VLIW. But just reducing writes. \n"
+"#define USE_2LEVEL_REDUCE 1\n"
+"//#define CHECK_BOUNDARY 1\n"
+"//#define NV_GPU 1\n"
+"//	Cypress\n"
+"#define nPerWI 16\n"
+"//	Cayman\n"
+"//#define nPerWI 20\n"
+"#define m_n x\n"
+"#define m_nWGs y\n"
+"#define m_startBit z\n"
+"#define m_nBlocksPerWG w\n"
+"/*\n"
+"typedef struct\n"
+"{\n"
+"	int m_n;\n"
+"	int m_nWGs;\n"
+"	int m_startBit;\n"
+"	int m_nBlocksPerWG;\n"
+"} ConstBuffer;\n"
+"*/\n"
+"typedef struct\n"
+"{\n"
+"	unsigned int m_key;\n"
+"	unsigned int m_value;\n"
+"} SortDataCL;\n"
+"uint prefixScanVectorEx( uint4* data )\n"
+"{\n"
+"	u32 sum = 0;\n"
+"	u32 tmp = data[0].x;\n"
+"	data[0].x = sum;\n"
+"	sum += tmp;\n"
+"	tmp = data[0].y;\n"
+"	data[0].y = sum;\n"
+"	sum += tmp;\n"
+"	tmp = data[0].z;\n"
+"	data[0].z = sum;\n"
+"	sum += tmp;\n"
+"	tmp = data[0].w;\n"
+"	data[0].w = sum;\n"
+"	sum += tmp;\n"
+"	return sum;\n"
+"}\n"
+"u32 localPrefixSum( u32 pData, uint lIdx, uint* totalSum, __local u32* sorterSharedMemory, int wgSize /*64 or 128*/ )\n"
+"{\n"
+"	{	//	Set data\n"
+"		sorterSharedMemory[lIdx] = 0;\n"
+"		sorterSharedMemory[lIdx+wgSize] = pData;\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	{	//	Prefix sum\n"
+"		int idx = 2*lIdx + (wgSize+1);\n"
+"#if defined(USE_2LEVEL_REDUCE)\n"
+"		if( lIdx < 64 )\n"
+"		{\n"
+"			u32 u0, u1, u2;\n"
+"			u0 = sorterSharedMemory[idx-3];\n"
+"			u1 = sorterSharedMemory[idx-2];\n"
+"			u2 = sorterSharedMemory[idx-1];\n"
+"			AtomAdd( sorterSharedMemory[idx], u0+u1+u2 );			\n"
+"			GROUP_MEM_FENCE;\n"
+"			u0 = sorterSharedMemory[idx-12];\n"
+"			u1 = sorterSharedMemory[idx-8];\n"
+"			u2 = sorterSharedMemory[idx-4];\n"
+"			AtomAdd( sorterSharedMemory[idx], u0+u1+u2 );			\n"
+"			GROUP_MEM_FENCE;\n"
+"			u0 = sorterSharedMemory[idx-48];\n"
+"			u1 = sorterSharedMemory[idx-32];\n"
+"			u2 = sorterSharedMemory[idx-16];\n"
+"			AtomAdd( sorterSharedMemory[idx], u0+u1+u2 );			\n"
+"			GROUP_MEM_FENCE;\n"
+"			if( wgSize > 64 )\n"
+"			{\n"
+"				sorterSharedMemory[idx] += sorterSharedMemory[idx-64];\n"
+"				GROUP_MEM_FENCE;\n"
+"			}\n"
+"			sorterSharedMemory[idx-1] += sorterSharedMemory[idx-2];\n"
+"			GROUP_MEM_FENCE;\n"
+"		}\n"
+"#else\n"
+"		if( lIdx < 64 )\n"
+"		{\n"
+"			sorterSharedMemory[idx] += sorterSharedMemory[idx-1];\n"
+"			GROUP_MEM_FENCE;\n"
+"			sorterSharedMemory[idx] += sorterSharedMemory[idx-2];			\n"
+"			GROUP_MEM_FENCE;\n"
+"			sorterSharedMemory[idx] += sorterSharedMemory[idx-4];\n"
+"			GROUP_MEM_FENCE;\n"
+"			sorterSharedMemory[idx] += sorterSharedMemory[idx-8];\n"
+"			GROUP_MEM_FENCE;\n"
+"			sorterSharedMemory[idx] += sorterSharedMemory[idx-16];\n"
+"			GROUP_MEM_FENCE;\n"
+"			sorterSharedMemory[idx] += sorterSharedMemory[idx-32];\n"
+"			GROUP_MEM_FENCE;\n"
+"			if( wgSize > 64 )\n"
+"			{\n"
+"				sorterSharedMemory[idx] += sorterSharedMemory[idx-64];\n"
+"				GROUP_MEM_FENCE;\n"
+"			}\n"
+"			sorterSharedMemory[idx-1] += sorterSharedMemory[idx-2];\n"
+"			GROUP_MEM_FENCE;\n"
+"		}\n"
+"#endif\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	*totalSum = sorterSharedMemory[wgSize*2-1];\n"
+"	u32 addValue = sorterSharedMemory[lIdx+wgSize-1];\n"
+"	return addValue;\n"
+"}\n"
+"//__attribute__((reqd_work_group_size(128,1,1)))\n"
+"uint4 localPrefixSum128V( uint4 pData, uint lIdx, uint* totalSum, __local u32* sorterSharedMemory )\n"
+"{\n"
+"	u32 s4 = prefixScanVectorEx( &pData );\n"
+"	u32 rank = localPrefixSum( s4, lIdx, totalSum, sorterSharedMemory, 128 );\n"
+"	return pData + make_uint4( rank, rank, rank, rank );\n"
+"}\n"
+"//__attribute__((reqd_work_group_size(64,1,1)))\n"
+"uint4 localPrefixSum64V( uint4 pData, uint lIdx, uint* totalSum, __local u32* sorterSharedMemory )\n"
+"{\n"
+"	u32 s4 = prefixScanVectorEx( &pData );\n"
+"	u32 rank = localPrefixSum( s4, lIdx, totalSum, sorterSharedMemory, 64 );\n"
+"	return pData + make_uint4( rank, rank, rank, rank );\n"
+"}\n"
+"u32 unpack4Key( u32 key, int keyIdx ){ return (key>>(keyIdx*8)) & 0xff;}\n"
+"u32 bit8Scan(u32 v)\n"
+"{\n"
+"	return (v<<8) + (v<<16) + (v<<24);\n"
+"}\n"
+"//===\n"
+"#define MY_HISTOGRAM(idx) localHistogramMat[(idx)*WG_SIZE+lIdx]\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void StreamCountKernel( __global u32* gSrc, __global u32* histogramOut, int4 cb )\n"
+"{\n"
+"	__local u32 localHistogramMat[NUM_BUCKET*WG_SIZE];\n"
+"	u32 gIdx = GET_GLOBAL_IDX;\n"
+"	u32 lIdx = GET_LOCAL_IDX;\n"
+"	u32 wgIdx = GET_GROUP_IDX;\n"
+"	u32 wgSize = GET_GROUP_SIZE;\n"
+"	const int startBit = cb.m_startBit;\n"
+"	const int n = cb.m_n;\n"
+"	const int nWGs = cb.m_nWGs;\n"
+"	const int nBlocksPerWG = cb.m_nBlocksPerWG;\n"
+"	for(int i=0; i<NUM_BUCKET; i++)\n"
+"	{\n"
+"		MY_HISTOGRAM(i) = 0;\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n"
+"	u32 localKey;\n"
+"	int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;\n"
+"	int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n"
+"	for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)\n"
+"	{\n"
+"		//	MY_HISTOGRAM( localKeys.x ) ++ is much expensive than atomic add as it requires read and write while atomics can just add on AMD\n"
+"		//	Using registers didn't perform well. It seems like use localKeys to address requires a lot of alu ops\n"
+"		//	AMD: AtomInc performs better while NV prefers ++\n"
+"		for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
+"		{\n"
+"#if defined(CHECK_BOUNDARY)\n"
+"			if( addr+i < n )\n"
+"#endif\n"
+"			{\n"
+"				localKey = (gSrc[addr+i]>>startBit) & 0xf;\n"
+"#if defined(NV_GPU)\n"
+"				MY_HISTOGRAM( localKey )++;\n"
+"#else\n"
+"				AtomInc( MY_HISTOGRAM( localKey ) );\n"
+"#endif\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	\n"
+"	if( lIdx < NUM_BUCKET )\n"
+"	{\n"
+"		u32 sum = 0;\n"
+"		for(int i=0; i<GET_GROUP_SIZE; i++)\n"
+"		{\n"
+"			sum += localHistogramMat[lIdx*WG_SIZE+(i+lIdx)%GET_GROUP_SIZE];\n"
+"		}\n"
+"		histogramOut[lIdx*nWGs+wgIdx] = sum;\n"
+"	}\n"
+"}\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void StreamCountSortDataKernel( __global SortDataCL* gSrc, __global u32* histogramOut, int4  cb )\n"
+"{\n"
+"	__local u32 localHistogramMat[NUM_BUCKET*WG_SIZE];\n"
+"	u32 gIdx = GET_GLOBAL_IDX;\n"
+"	u32 lIdx = GET_LOCAL_IDX;\n"
+"	u32 wgIdx = GET_GROUP_IDX;\n"
+"	u32 wgSize = GET_GROUP_SIZE;\n"
+"	const int startBit = cb.m_startBit;\n"
+"	const int n = cb.m_n;\n"
+"	const int nWGs = cb.m_nWGs;\n"
+"	const int nBlocksPerWG = cb.m_nBlocksPerWG;\n"
+"	for(int i=0; i<NUM_BUCKET; i++)\n"
+"	{\n"
+"		MY_HISTOGRAM(i) = 0;\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n"
+"	u32 localKey;\n"
+"	int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;\n"
+"	int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n"
+"	for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)\n"
+"	{\n"
+"		//	MY_HISTOGRAM( localKeys.x ) ++ is much expensive than atomic add as it requires read and write while atomics can just add on AMD\n"
+"		//	Using registers didn't perform well. It seems like use localKeys to address requires a lot of alu ops\n"
+"		//	AMD: AtomInc performs better while NV prefers ++\n"
+"		for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
+"		{\n"
+"#if defined(CHECK_BOUNDARY)\n"
+"			if( addr+i < n )\n"
+"#endif\n"
+"			{\n"
+"				localKey = (gSrc[addr+i].m_key>>startBit) & 0xf;\n"
+"#if defined(NV_GPU)\n"
+"				MY_HISTOGRAM( localKey )++;\n"
+"#else\n"
+"				AtomInc( MY_HISTOGRAM( localKey ) );\n"
+"#endif\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	\n"
+"	if( lIdx < NUM_BUCKET )\n"
+"	{\n"
+"		u32 sum = 0;\n"
+"		for(int i=0; i<GET_GROUP_SIZE; i++)\n"
+"		{\n"
+"			sum += localHistogramMat[lIdx*WG_SIZE+(i+lIdx)%GET_GROUP_SIZE];\n"
+"		}\n"
+"		histogramOut[lIdx*nWGs+wgIdx] = sum;\n"
+"	}\n"
+"}\n"
+"#define nPerLane (nPerWI/4)\n"
+"//	NUM_BUCKET*nWGs < 128*nPerWI\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(128,1,1)))\n"
+"void PrefixScanKernel( __global u32* wHistogram1, int4  cb )\n"
+"{\n"
+"	__local u32 ldsTopScanData[128*2];\n"
+"	u32 lIdx = GET_LOCAL_IDX;\n"
+"	u32 wgIdx = GET_GROUP_IDX;\n"
+"	const int nWGs = cb.m_nWGs;\n"
+"	u32 data[nPerWI];\n"
+"	for(int i=0; i<nPerWI; i++)\n"
+"	{\n"
+"		data[i] = 0;\n"
+"		if( (nPerWI*lIdx+i) < NUM_BUCKET*nWGs )\n"
+"			data[i] = wHistogram1[nPerWI*lIdx+i];\n"
+"	}\n"
+"	uint4 myData = make_uint4(0,0,0,0);\n"
+"	for(int i=0; i<nPerLane; i++)\n"
+"	{\n"
+"		myData.x += data[nPerLane*0+i];\n"
+"		myData.y += data[nPerLane*1+i];\n"
+"		myData.z += data[nPerLane*2+i];\n"
+"		myData.w += data[nPerLane*3+i];\n"
+"	}\n"
+"	uint totalSum;\n"
+"	uint4 scanned = localPrefixSum128V( myData, lIdx, &totalSum, ldsTopScanData );\n"
+"//	for(int j=0; j<4; j++) //	somehow it introduces a lot of branches\n"
+"	{	int j = 0;\n"
+"		u32 sum = 0;\n"
+"		for(int i=0; i<nPerLane; i++)\n"
+"		{\n"
+"			u32 tmp = data[nPerLane*j+i];\n"
+"			data[nPerLane*j+i] = sum;\n"
+"			sum += tmp;\n"
+"		}\n"
+"	}\n"
+"	{	int j = 1;\n"
+"		u32 sum = 0;\n"
+"		for(int i=0; i<nPerLane; i++)\n"
+"		{\n"
+"			u32 tmp = data[nPerLane*j+i];\n"
+"			data[nPerLane*j+i] = sum;\n"
+"			sum += tmp;\n"
+"		}\n"
+"	}\n"
+"	{	int j = 2;\n"
+"		u32 sum = 0;\n"
+"		for(int i=0; i<nPerLane; i++)\n"
+"		{\n"
+"			u32 tmp = data[nPerLane*j+i];\n"
+"			data[nPerLane*j+i] = sum;\n"
+"			sum += tmp;\n"
+"		}\n"
+"	}\n"
+"	{	int j = 3;\n"
+"		u32 sum = 0;\n"
+"		for(int i=0; i<nPerLane; i++)\n"
+"		{\n"
+"			u32 tmp = data[nPerLane*j+i];\n"
+"			data[nPerLane*j+i] = sum;\n"
+"			sum += tmp;\n"
+"		}\n"
+"	}\n"
+"	for(int i=0; i<nPerLane; i++)\n"
+"	{\n"
+"		data[nPerLane*0+i] += scanned.x;\n"
+"		data[nPerLane*1+i] += scanned.y;\n"
+"		data[nPerLane*2+i] += scanned.z;\n"
+"		data[nPerLane*3+i] += scanned.w;\n"
+"	}\n"
+"	for(int i=0; i<nPerWI; i++)\n"
+"	{\n"
+"		int index = nPerWI*lIdx+i;\n"
+"		if (index < NUM_BUCKET*nWGs)\n"
+"			wHistogram1[nPerWI*lIdx+i] = data[i];\n"
+"	}\n"
+"}\n"
+"//	4 scan, 4 exchange\n"
+"void sort4Bits(u32 sortData[4], int startBit, int lIdx, __local u32* ldsSortData)\n"
+"{\n"
+"	for(int bitIdx=0; bitIdx<BITS_PER_PASS; bitIdx++)\n"
+"	{\n"
+"		u32 mask = (1<<bitIdx);\n"
+"		uint4 cmpResult = make_uint4( (sortData[0]>>startBit) & mask, (sortData[1]>>startBit) & mask, (sortData[2]>>startBit) & mask, (sortData[3]>>startBit) & mask );\n"
+"		uint4 prefixSum = SELECT_UINT4( make_uint4(1,1,1,1), make_uint4(0,0,0,0), cmpResult != make_uint4(0,0,0,0) );\n"
+"		u32 total;\n"
+"		prefixSum = localPrefixSum64V( prefixSum, lIdx, &total, ldsSortData );\n"
+"		{\n"
+"			uint4 localAddr = make_uint4(lIdx*4+0,lIdx*4+1,lIdx*4+2,lIdx*4+3);\n"
+"			uint4 dstAddr = localAddr - prefixSum + make_uint4( total, total, total, total );\n"
+"			dstAddr = SELECT_UINT4( prefixSum, dstAddr, cmpResult != make_uint4(0, 0, 0, 0) );\n"
+"			GROUP_LDS_BARRIER;\n"
+"			ldsSortData[dstAddr.x] = sortData[0];\n"
+"			ldsSortData[dstAddr.y] = sortData[1];\n"
+"			ldsSortData[dstAddr.z] = sortData[2];\n"
+"			ldsSortData[dstAddr.w] = sortData[3];\n"
+"			GROUP_LDS_BARRIER;\n"
+"			sortData[0] = ldsSortData[localAddr.x];\n"
+"			sortData[1] = ldsSortData[localAddr.y];\n"
+"			sortData[2] = ldsSortData[localAddr.z];\n"
+"			sortData[3] = ldsSortData[localAddr.w];\n"
+"			GROUP_LDS_BARRIER;\n"
+"		}\n"
+"	}\n"
+"}\n"
+"//	2 scan, 2 exchange\n"
+"void sort4Bits1(u32 sortData[4], int startBit, int lIdx, __local u32* ldsSortData)\n"
+"{\n"
+"	for(uint ibit=0; ibit<BITS_PER_PASS; ibit+=2)\n"
+"	{\n"
+"		uint4 b = make_uint4((sortData[0]>>(startBit+ibit)) & 0x3, \n"
+"			(sortData[1]>>(startBit+ibit)) & 0x3, \n"
+"			(sortData[2]>>(startBit+ibit)) & 0x3, \n"
+"			(sortData[3]>>(startBit+ibit)) & 0x3);\n"
+"		u32 key4;\n"
+"		u32 sKeyPacked[4] = { 0, 0, 0, 0 };\n"
+"		{\n"
+"			sKeyPacked[0] |= 1<<(8*b.x);\n"
+"			sKeyPacked[1] |= 1<<(8*b.y);\n"
+"			sKeyPacked[2] |= 1<<(8*b.z);\n"
+"			sKeyPacked[3] |= 1<<(8*b.w);\n"
+"			key4 = sKeyPacked[0] + sKeyPacked[1] + sKeyPacked[2] + sKeyPacked[3];\n"
+"		}\n"
+"		u32 rankPacked;\n"
+"		u32 sumPacked;\n"
+"		{\n"
+"			rankPacked = localPrefixSum( key4, lIdx, &sumPacked, ldsSortData, WG_SIZE );\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"		u32 newOffset[4] = { 0,0,0,0 };\n"
+"		{\n"
+"			u32 sumScanned = bit8Scan( sumPacked );\n"
+"			u32 scannedKeys[4];\n"
+"			scannedKeys[0] = 1<<(8*b.x);\n"
+"			scannedKeys[1] = 1<<(8*b.y);\n"
+"			scannedKeys[2] = 1<<(8*b.z);\n"
+"			scannedKeys[3] = 1<<(8*b.w);\n"
+"			{	//	4 scans at once\n"
+"				u32 sum4 = 0;\n"
+"				for(int ie=0; ie<4; ie++)\n"
+"				{\n"
+"					u32 tmp = scannedKeys[ie];\n"
+"					scannedKeys[ie] = sum4;\n"
+"					sum4 += tmp;\n"
+"				}\n"
+"			}\n"
+"			{\n"
+"				u32 sumPlusRank = sumScanned + rankPacked;\n"
+"				{	u32 ie = b.x;\n"
+"					scannedKeys[0] += sumPlusRank;\n"
+"					newOffset[0] = unpack4Key( scannedKeys[0], ie );\n"
+"				}\n"
+"				{	u32 ie = b.y;\n"
+"					scannedKeys[1] += sumPlusRank;\n"
+"					newOffset[1] = unpack4Key( scannedKeys[1], ie );\n"
+"				}\n"
+"				{	u32 ie = b.z;\n"
+"					scannedKeys[2] += sumPlusRank;\n"
+"					newOffset[2] = unpack4Key( scannedKeys[2], ie );\n"
+"				}\n"
+"				{	u32 ie = b.w;\n"
+"					scannedKeys[3] += sumPlusRank;\n"
+"					newOffset[3] = unpack4Key( scannedKeys[3], ie );\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"		{\n"
+"			ldsSortData[newOffset[0]] = sortData[0];\n"
+"			ldsSortData[newOffset[1]] = sortData[1];\n"
+"			ldsSortData[newOffset[2]] = sortData[2];\n"
+"			ldsSortData[newOffset[3]] = sortData[3];\n"
+"			GROUP_LDS_BARRIER;\n"
+"			u32 dstAddr = 4*lIdx;\n"
+"			sortData[0] = ldsSortData[dstAddr+0];\n"
+"			sortData[1] = ldsSortData[dstAddr+1];\n"
+"			sortData[2] = ldsSortData[dstAddr+2];\n"
+"			sortData[3] = ldsSortData[dstAddr+3];\n"
+"			GROUP_LDS_BARRIER;\n"
+"		}\n"
+"	}\n"
+"}\n"
+"#define SET_HISTOGRAM(setIdx, key) ldsSortData[(setIdx)*NUM_BUCKET+key]\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void SortAndScatterKernel( __global const u32* restrict gSrc, __global const u32* rHistogram, __global u32* restrict gDst, int4  cb )\n"
+"{\n"
+"	__local u32 ldsSortData[WG_SIZE*ELEMENTS_PER_WORK_ITEM+16];\n"
+"	__local u32 localHistogramToCarry[NUM_BUCKET];\n"
+"	__local u32 localHistogram[NUM_BUCKET*2];\n"
+"	u32 gIdx = GET_GLOBAL_IDX;\n"
+"	u32 lIdx = GET_LOCAL_IDX;\n"
+"	u32 wgIdx = GET_GROUP_IDX;\n"
+"	u32 wgSize = GET_GROUP_SIZE;\n"
+"	const int n = cb.m_n;\n"
+"	const int nWGs = cb.m_nWGs;\n"
+"	const int startBit = cb.m_startBit;\n"
+"	const int nBlocksPerWG = cb.m_nBlocksPerWG;\n"
+"	if( lIdx < (NUM_BUCKET) )\n"
+"	{\n"
+"		localHistogramToCarry[lIdx] = rHistogram[lIdx*nWGs + wgIdx];\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n"
+"	int nBlocks = n/blockSize - nBlocksPerWG*wgIdx;\n"
+"	int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n"
+"	for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)\n"
+"	{\n"
+"		u32 myHistogram = 0;\n"
+"		u32 sortData[ELEMENTS_PER_WORK_ITEM];\n"
+"		for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
+"#if defined(CHECK_BOUNDARY)\n"
+"			sortData[i] = ( addr+i < n )? gSrc[ addr+i ] : 0xffffffff;\n"
+"#else\n"
+"			sortData[i] = gSrc[ addr+i ];\n"
+"#endif\n"
+"		sort4Bits(sortData, startBit, lIdx, ldsSortData);\n"
+"		u32 keys[ELEMENTS_PER_WORK_ITEM];\n"
+"		for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
+"			keys[i] = (sortData[i]>>startBit) & 0xf;\n"
+"		{	//	create histogram\n"
+"			u32 setIdx = lIdx/16;\n"
+"			if( lIdx < NUM_BUCKET )\n"
+"			{\n"
+"				localHistogram[lIdx] = 0;\n"
+"			}\n"
+"			ldsSortData[lIdx] = 0;\n"
+"			GROUP_LDS_BARRIER;\n"
+"			for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
+"#if defined(CHECK_BOUNDARY)\n"
+"				if( addr+i < n )\n"
+"#endif\n"
+"#if defined(NV_GPU)\n"
+"				SET_HISTOGRAM( setIdx, keys[i] )++;\n"
+"#else\n"
+"				AtomInc( SET_HISTOGRAM( setIdx, keys[i] ) );\n"
+"#endif\n"
+"			\n"
+"			GROUP_LDS_BARRIER;\n"
+"			\n"
+"			uint hIdx = NUM_BUCKET+lIdx;\n"
+"			if( lIdx < NUM_BUCKET )\n"
+"			{\n"
+"				u32 sum = 0;\n"
+"				for(int i=0; i<WG_SIZE/16; i++)\n"
+"				{\n"
+"					sum += SET_HISTOGRAM( i, lIdx );\n"
+"				}\n"
+"				myHistogram = sum;\n"
+"				localHistogram[hIdx] = sum;\n"
+"			}\n"
+"			GROUP_LDS_BARRIER;\n"
+"#if defined(USE_2LEVEL_REDUCE)\n"
+"			if( lIdx < NUM_BUCKET )\n"
+"			{\n"
+"				localHistogram[hIdx] = localHistogram[hIdx-1];\n"
+"				GROUP_MEM_FENCE;\n"
+"				u32 u0, u1, u2;\n"
+"				u0 = localHistogram[hIdx-3];\n"
+"				u1 = localHistogram[hIdx-2];\n"
+"				u2 = localHistogram[hIdx-1];\n"
+"				AtomAdd( localHistogram[hIdx], u0 + u1 + u2 );\n"
+"				GROUP_MEM_FENCE;\n"
+"				u0 = localHistogram[hIdx-12];\n"
+"				u1 = localHistogram[hIdx-8];\n"
+"				u2 = localHistogram[hIdx-4];\n"
+"				AtomAdd( localHistogram[hIdx], u0 + u1 + u2 );\n"
+"				GROUP_MEM_FENCE;\n"
+"			}\n"
+"#else\n"
+"			if( lIdx < NUM_BUCKET )\n"
+"			{\n"
+"				localHistogram[hIdx] = localHistogram[hIdx-1];\n"
+"				GROUP_MEM_FENCE;\n"
+"				localHistogram[hIdx] += localHistogram[hIdx-1];\n"
+"				GROUP_MEM_FENCE;\n"
+"				localHistogram[hIdx] += localHistogram[hIdx-2];\n"
+"				GROUP_MEM_FENCE;\n"
+"				localHistogram[hIdx] += localHistogram[hIdx-4];\n"
+"				GROUP_MEM_FENCE;\n"
+"				localHistogram[hIdx] += localHistogram[hIdx-8];\n"
+"				GROUP_MEM_FENCE;\n"
+"			}\n"
+"#endif\n"
+"			GROUP_LDS_BARRIER;\n"
+"		}\n"
+"		{\n"
+"			for(int ie=0; ie<ELEMENTS_PER_WORK_ITEM; ie++)\n"
+"			{\n"
+"				int dataIdx = ELEMENTS_PER_WORK_ITEM*lIdx+ie;\n"
+"				int binIdx = keys[ie];\n"
+"				int groupOffset = localHistogramToCarry[binIdx];\n"
+"				int myIdx = dataIdx - localHistogram[NUM_BUCKET+binIdx];\n"
+"#if defined(CHECK_BOUNDARY)\n"
+"				if( addr+ie < n )\n"
+"#endif\n"
+"				gDst[ groupOffset + myIdx ] = sortData[ie];\n"
+"			}\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"		if( lIdx < NUM_BUCKET )\n"
+"		{\n"
+"			localHistogramToCarry[lIdx] += myHistogram;\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"	}\n"
+"}\n"
+"//	2 scan, 2 exchange\n"
+"void sort4Bits1KeyValue(u32 sortData[4], int sortVal[4], int startBit, int lIdx, __local u32* ldsSortData, __local int *ldsSortVal)\n"
+"{\n"
+"	for(uint ibit=0; ibit<BITS_PER_PASS; ibit+=2)\n"
+"	{\n"
+"		uint4 b = make_uint4((sortData[0]>>(startBit+ibit)) & 0x3, \n"
+"			(sortData[1]>>(startBit+ibit)) & 0x3, \n"
+"			(sortData[2]>>(startBit+ibit)) & 0x3, \n"
+"			(sortData[3]>>(startBit+ibit)) & 0x3);\n"
+"		u32 key4;\n"
+"		u32 sKeyPacked[4] = { 0, 0, 0, 0 };\n"
+"		{\n"
+"			sKeyPacked[0] |= 1<<(8*b.x);\n"
+"			sKeyPacked[1] |= 1<<(8*b.y);\n"
+"			sKeyPacked[2] |= 1<<(8*b.z);\n"
+"			sKeyPacked[3] |= 1<<(8*b.w);\n"
+"			key4 = sKeyPacked[0] + sKeyPacked[1] + sKeyPacked[2] + sKeyPacked[3];\n"
+"		}\n"
+"		u32 rankPacked;\n"
+"		u32 sumPacked;\n"
+"		{\n"
+"			rankPacked = localPrefixSum( key4, lIdx, &sumPacked, ldsSortData, WG_SIZE );\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"		u32 newOffset[4] = { 0,0,0,0 };\n"
+"		{\n"
+"			u32 sumScanned = bit8Scan( sumPacked );\n"
+"			u32 scannedKeys[4];\n"
+"			scannedKeys[0] = 1<<(8*b.x);\n"
+"			scannedKeys[1] = 1<<(8*b.y);\n"
+"			scannedKeys[2] = 1<<(8*b.z);\n"
+"			scannedKeys[3] = 1<<(8*b.w);\n"
+"			{	//	4 scans at once\n"
+"				u32 sum4 = 0;\n"
+"				for(int ie=0; ie<4; ie++)\n"
+"				{\n"
+"					u32 tmp = scannedKeys[ie];\n"
+"					scannedKeys[ie] = sum4;\n"
+"					sum4 += tmp;\n"
+"				}\n"
+"			}\n"
+"			{\n"
+"				u32 sumPlusRank = sumScanned + rankPacked;\n"
+"				{	u32 ie = b.x;\n"
+"					scannedKeys[0] += sumPlusRank;\n"
+"					newOffset[0] = unpack4Key( scannedKeys[0], ie );\n"
+"				}\n"
+"				{	u32 ie = b.y;\n"
+"					scannedKeys[1] += sumPlusRank;\n"
+"					newOffset[1] = unpack4Key( scannedKeys[1], ie );\n"
+"				}\n"
+"				{	u32 ie = b.z;\n"
+"					scannedKeys[2] += sumPlusRank;\n"
+"					newOffset[2] = unpack4Key( scannedKeys[2], ie );\n"
+"				}\n"
+"				{	u32 ie = b.w;\n"
+"					scannedKeys[3] += sumPlusRank;\n"
+"					newOffset[3] = unpack4Key( scannedKeys[3], ie );\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"		{\n"
+"			ldsSortData[newOffset[0]] = sortData[0];\n"
+"			ldsSortData[newOffset[1]] = sortData[1];\n"
+"			ldsSortData[newOffset[2]] = sortData[2];\n"
+"			ldsSortData[newOffset[3]] = sortData[3];\n"
+"			ldsSortVal[newOffset[0]] = sortVal[0];\n"
+"			ldsSortVal[newOffset[1]] = sortVal[1];\n"
+"			ldsSortVal[newOffset[2]] = sortVal[2];\n"
+"			ldsSortVal[newOffset[3]] = sortVal[3];\n"
+"			GROUP_LDS_BARRIER;\n"
+"			u32 dstAddr = 4*lIdx;\n"
+"			sortData[0] = ldsSortData[dstAddr+0];\n"
+"			sortData[1] = ldsSortData[dstAddr+1];\n"
+"			sortData[2] = ldsSortData[dstAddr+2];\n"
+"			sortData[3] = ldsSortData[dstAddr+3];\n"
+"			sortVal[0] = ldsSortVal[dstAddr+0];\n"
+"			sortVal[1] = ldsSortVal[dstAddr+1];\n"
+"			sortVal[2] = ldsSortVal[dstAddr+2];\n"
+"			sortVal[3] = ldsSortVal[dstAddr+3];\n"
+"			GROUP_LDS_BARRIER;\n"
+"		}\n"
+"	}\n"
+"}\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void SortAndScatterSortDataKernel( __global const SortDataCL* restrict gSrc, __global const u32* rHistogram, __global SortDataCL* restrict gDst, int4 cb)\n"
+"{\n"
+"	__local int ldsSortData[WG_SIZE*ELEMENTS_PER_WORK_ITEM+16];\n"
+"	__local int ldsSortVal[WG_SIZE*ELEMENTS_PER_WORK_ITEM+16];\n"
+"	__local u32 localHistogramToCarry[NUM_BUCKET];\n"
+"	__local u32 localHistogram[NUM_BUCKET*2];\n"
+"	u32 gIdx = GET_GLOBAL_IDX;\n"
+"	u32 lIdx = GET_LOCAL_IDX;\n"
+"	u32 wgIdx = GET_GROUP_IDX;\n"
+"	u32 wgSize = GET_GROUP_SIZE;\n"
+"	const int n = cb.m_n;\n"
+"	const int nWGs = cb.m_nWGs;\n"
+"	const int startBit = cb.m_startBit;\n"
+"	const int nBlocksPerWG = cb.m_nBlocksPerWG;\n"
+"	if( lIdx < (NUM_BUCKET) )\n"
+"	{\n"
+"		localHistogramToCarry[lIdx] = rHistogram[lIdx*nWGs + wgIdx];\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"    \n"
+"	const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n"
+"	int nBlocks = n/blockSize - nBlocksPerWG*wgIdx;\n"
+"	int addr = blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n"
+"	for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++, addr+=blockSize)\n"
+"	{\n"
+"		u32 myHistogram = 0;\n"
+"		int sortData[ELEMENTS_PER_WORK_ITEM];\n"
+"		int sortVal[ELEMENTS_PER_WORK_ITEM];\n"
+"		for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
+"#if defined(CHECK_BOUNDARY)\n"
+"		{\n"
+"			sortData[i] = ( addr+i < n )? gSrc[ addr+i ].m_key : 0xffffffff;\n"
+"			sortVal[i] = ( addr+i < n )? gSrc[ addr+i ].m_value : 0xffffffff;\n"
+"		}\n"
+"#else\n"
+"		{\n"
+"			sortData[i] = gSrc[ addr+i ].m_key;\n"
+"			sortVal[i] = gSrc[ addr+i ].m_value;\n"
+"		}\n"
+"#endif\n"
+"		sort4Bits1KeyValue(sortData, sortVal, startBit, lIdx, ldsSortData, ldsSortVal);\n"
+"		u32 keys[ELEMENTS_PER_WORK_ITEM];\n"
+"		for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
+"			keys[i] = (sortData[i]>>startBit) & 0xf;\n"
+"		{	//	create histogram\n"
+"			u32 setIdx = lIdx/16;\n"
+"			if( lIdx < NUM_BUCKET )\n"
+"			{\n"
+"				localHistogram[lIdx] = 0;\n"
+"			}\n"
+"			ldsSortData[lIdx] = 0;\n"
+"			GROUP_LDS_BARRIER;\n"
+"			for(int i=0; i<ELEMENTS_PER_WORK_ITEM; i++)\n"
+"#if defined(CHECK_BOUNDARY)\n"
+"				if( addr+i < n )\n"
+"#endif\n"
+"#if defined(NV_GPU)\n"
+"				SET_HISTOGRAM( setIdx, keys[i] )++;\n"
+"#else\n"
+"				AtomInc( SET_HISTOGRAM( setIdx, keys[i] ) );\n"
+"#endif\n"
+"			\n"
+"			GROUP_LDS_BARRIER;\n"
+"			\n"
+"			uint hIdx = NUM_BUCKET+lIdx;\n"
+"			if( lIdx < NUM_BUCKET )\n"
+"			{\n"
+"				u32 sum = 0;\n"
+"				for(int i=0; i<WG_SIZE/16; i++)\n"
+"				{\n"
+"					sum += SET_HISTOGRAM( i, lIdx );\n"
+"				}\n"
+"				myHistogram = sum;\n"
+"				localHistogram[hIdx] = sum;\n"
+"			}\n"
+"			GROUP_LDS_BARRIER;\n"
+"#if defined(USE_2LEVEL_REDUCE)\n"
+"			if( lIdx < NUM_BUCKET )\n"
+"			{\n"
+"				localHistogram[hIdx] = localHistogram[hIdx-1];\n"
+"				GROUP_MEM_FENCE;\n"
+"				u32 u0, u1, u2;\n"
+"				u0 = localHistogram[hIdx-3];\n"
+"				u1 = localHistogram[hIdx-2];\n"
+"				u2 = localHistogram[hIdx-1];\n"
+"				AtomAdd( localHistogram[hIdx], u0 + u1 + u2 );\n"
+"				GROUP_MEM_FENCE;\n"
+"				u0 = localHistogram[hIdx-12];\n"
+"				u1 = localHistogram[hIdx-8];\n"
+"				u2 = localHistogram[hIdx-4];\n"
+"				AtomAdd( localHistogram[hIdx], u0 + u1 + u2 );\n"
+"				GROUP_MEM_FENCE;\n"
+"			}\n"
+"#else\n"
+"			if( lIdx < NUM_BUCKET )\n"
+"			{\n"
+"				localHistogram[hIdx] = localHistogram[hIdx-1];\n"
+"				GROUP_MEM_FENCE;\n"
+"				localHistogram[hIdx] += localHistogram[hIdx-1];\n"
+"				GROUP_MEM_FENCE;\n"
+"				localHistogram[hIdx] += localHistogram[hIdx-2];\n"
+"				GROUP_MEM_FENCE;\n"
+"				localHistogram[hIdx] += localHistogram[hIdx-4];\n"
+"				GROUP_MEM_FENCE;\n"
+"				localHistogram[hIdx] += localHistogram[hIdx-8];\n"
+"				GROUP_MEM_FENCE;\n"
+"			}\n"
+"#endif\n"
+"			GROUP_LDS_BARRIER;\n"
+"		}\n"
+"    	{\n"
+"			for(int ie=0; ie<ELEMENTS_PER_WORK_ITEM; ie++)\n"
+"			{\n"
+"				int dataIdx = ELEMENTS_PER_WORK_ITEM*lIdx+ie;\n"
+"				int binIdx = keys[ie];\n"
+"				int groupOffset = localHistogramToCarry[binIdx];\n"
+"				int myIdx = dataIdx - localHistogram[NUM_BUCKET+binIdx];\n"
+"#if defined(CHECK_BOUNDARY)\n"
+"				if( addr+ie < n )\n"
+"				{\n"
+"                    if ((groupOffset + myIdx)<n)\n"
+"                    {\n"
+"                        if (sortData[ie]==sortVal[ie])\n"
+"                        {\n"
+"                            \n"
+"                            SortDataCL tmp;\n"
+"                            tmp.m_key = sortData[ie];\n"
+"                            tmp.m_value = sortVal[ie];\n"
+"                            if (tmp.m_key == tmp.m_value)\n"
+"                                gDst[groupOffset + myIdx ] = tmp;\n"
+"                        }\n"
+"                        \n"
+"                    }\n"
+"				}\n"
+"#else\n"
+"                if ((groupOffset + myIdx)<n)\n"
+"                {\n"
+"                    gDst[ groupOffset + myIdx ].m_key = sortData[ie];\n"
+"                    gDst[ groupOffset + myIdx ].m_value = sortVal[ie];\n"
+"                }\n"
+"#endif\n"
+"			}\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"		if( lIdx < NUM_BUCKET )\n"
+"		{\n"
+"			localHistogramToCarry[lIdx] += myHistogram;\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"	}\n"
+"}\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void SortAndScatterSortDataKernelSerial( __global const SortDataCL* restrict gSrc, __global const u32* rHistogram, __global SortDataCL* restrict gDst, int4 cb)\n"
+"{\n"
+"    \n"
+"	u32 gIdx = GET_GLOBAL_IDX;\n"
+"	u32 realLocalIdx = GET_LOCAL_IDX;\n"
+"	u32 wgIdx = GET_GROUP_IDX;\n"
+"	u32 wgSize = GET_GROUP_SIZE;\n"
+"	const int startBit = cb.m_startBit;\n"
+"	const int n = cb.m_n;\n"
+"	const int nWGs = cb.m_nWGs;\n"
+"	const int nBlocksPerWG = cb.m_nBlocksPerWG;\n"
+"    int counter[NUM_BUCKET];\n"
+"    \n"
+"    if (realLocalIdx>0)\n"
+"        return;\n"
+"    \n"
+"    for (int c=0;c<NUM_BUCKET;c++)\n"
+"        counter[c]=0;\n"
+"    const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n"
+"	\n"
+"	int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;\n"
+"   for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++)\n"
+"  {\n"
+"     for (int lIdx=0;lIdx<WG_SIZE;lIdx++)\n"
+" 	{\n"
+"        int addr2 = iblock*blockSize + blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n"
+"        \n"
+"		for(int j=0; j<ELEMENTS_PER_WORK_ITEM; j++)\n"
+"		{\n"
+"            int i = addr2+j;\n"
+"			if( i < n )\n"
+"			{\n"
+"                int tableIdx;\n"
+"				tableIdx = (gSrc[i].m_key>>startBit) & 0xf;//0xf = NUM_TABLES-1\n"
+"                gDst[rHistogram[tableIdx*nWGs+wgIdx] + counter[tableIdx]] = gSrc[i];\n"
+"                counter[tableIdx] ++;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"  }\n"
+"    \n"
+"}\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void SortAndScatterKernelSerial( __global const u32* restrict gSrc, __global const u32* rHistogram, __global u32* restrict gDst, int4  cb )\n"
+"{\n"
+"    \n"
+"	u32 gIdx = GET_GLOBAL_IDX;\n"
+"	u32 realLocalIdx = GET_LOCAL_IDX;\n"
+"	u32 wgIdx = GET_GROUP_IDX;\n"
+"	u32 wgSize = GET_GROUP_SIZE;\n"
+"	const int startBit = cb.m_startBit;\n"
+"	const int n = cb.m_n;\n"
+"	const int nWGs = cb.m_nWGs;\n"
+"	const int nBlocksPerWG = cb.m_nBlocksPerWG;\n"
+"    int counter[NUM_BUCKET];\n"
+"    \n"
+"    if (realLocalIdx>0)\n"
+"        return;\n"
+"    \n"
+"    for (int c=0;c<NUM_BUCKET;c++)\n"
+"        counter[c]=0;\n"
+"    const int blockSize = ELEMENTS_PER_WORK_ITEM*WG_SIZE;\n"
+"	\n"
+"	int nBlocks = (n)/blockSize - nBlocksPerWG*wgIdx;\n"
+"   for(int iblock=0; iblock<min(nBlocksPerWG, nBlocks); iblock++)\n"
+"  {\n"
+"     for (int lIdx=0;lIdx<WG_SIZE;lIdx++)\n"
+" 	{\n"
+"        int addr2 = iblock*blockSize + blockSize*nBlocksPerWG*wgIdx + ELEMENTS_PER_WORK_ITEM*lIdx;\n"
+"        \n"
+"		for(int j=0; j<ELEMENTS_PER_WORK_ITEM; j++)\n"
+"		{\n"
+"            int i = addr2+j;\n"
+"			if( i < n )\n"
+"			{\n"
+"                int tableIdx;\n"
+"				tableIdx = (gSrc[i]>>startBit) & 0xf;//0xf = NUM_TABLES-1\n"
+"                gDst[rHistogram[tableIdx*nWGs+wgIdx] + counter[tableIdx]] = gSrc[i];\n"
+"                counter[tableIdx] ++;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"  }\n"
+"    \n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/Raycast/b3GpuRaycast.cpp b/thirdparty/bullet/Bullet3OpenCL/Raycast/b3GpuRaycast.cpp
new file mode 100644
index 0000000000..161e304f09
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/Raycast/b3GpuRaycast.cpp
@@ -0,0 +1,391 @@
+
+#include "b3GpuRaycast.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Collidable.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3RigidBodyData.h"
+#include "Bullet3OpenCL/RigidBody/b3GpuNarrowPhaseInternalData.h"
+
+
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3FillCL.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h"
+#include "Bullet3OpenCL/BroadphaseCollision/b3GpuBroadphaseInterface.h"
+#include "Bullet3OpenCL/BroadphaseCollision/b3GpuParallelLinearBvh.h"
+
+#include "Bullet3OpenCL/Raycast/kernels/rayCastKernels.h"
+
+
+#define B3_RAYCAST_PATH "src/Bullet3OpenCL/Raycast/kernels/rayCastKernels.cl"
+
+
+
+struct b3GpuRaycastInternalData
+{
+	cl_context m_context;
+	cl_device_id m_device;
+	cl_command_queue  m_q;
+	cl_kernel m_raytraceKernel;
+	cl_kernel m_raytracePairsKernel;
+	cl_kernel m_findRayRigidPairIndexRanges;
+	
+	b3GpuParallelLinearBvh* m_plbvh;
+	b3RadixSort32CL* m_radixSorter;
+	b3FillCL* m_fill;
+	
+	//1 element per ray
+	b3OpenCLArray<b3RayInfo>* m_gpuRays;
+	b3OpenCLArray<b3RayHit>* m_gpuHitResults;
+	b3OpenCLArray<int>* m_firstRayRigidPairIndexPerRay;
+	b3OpenCLArray<int>* m_numRayRigidPairsPerRay;
+	
+	//1 element per (ray index, rigid index) pair, where the ray intersects with the rigid's AABB
+	b3OpenCLArray<int>* m_gpuNumRayRigidPairs;
+	b3OpenCLArray<b3Int2>* m_gpuRayRigidPairs;	//x == ray index, y == rigid index
+	
+	int m_test;
+};
+
+b3GpuRaycast::b3GpuRaycast(cl_context ctx,cl_device_id device, cl_command_queue  q)
+{
+	m_data = new b3GpuRaycastInternalData;
+	m_data->m_context = ctx;
+	m_data->m_device = device;
+	m_data->m_q = q;
+	m_data->m_raytraceKernel = 0;
+	m_data->m_raytracePairsKernel = 0;
+	m_data->m_findRayRigidPairIndexRanges = 0;
+
+	m_data->m_plbvh = new b3GpuParallelLinearBvh(ctx, device, q);
+	m_data->m_radixSorter = new b3RadixSort32CL(ctx, device, q);
+	m_data->m_fill = new b3FillCL(ctx, device, q);
+	
+	m_data->m_gpuRays = new b3OpenCLArray<b3RayInfo>(ctx, q);
+	m_data->m_gpuHitResults = new b3OpenCLArray<b3RayHit>(ctx, q);
+	m_data->m_firstRayRigidPairIndexPerRay = new b3OpenCLArray<int>(ctx, q);
+	m_data->m_numRayRigidPairsPerRay = new b3OpenCLArray<int>(ctx, q);
+	m_data->m_gpuNumRayRigidPairs = new b3OpenCLArray<int>(ctx, q);
+	m_data->m_gpuRayRigidPairs = new b3OpenCLArray<b3Int2>(ctx, q);
+
+	{
+		cl_int errNum=0;
+		cl_program prog = b3OpenCLUtils::compileCLProgramFromString(m_data->m_context,m_data->m_device,rayCastKernelCL,&errNum,"",B3_RAYCAST_PATH);
+		b3Assert(errNum==CL_SUCCESS);
+		m_data->m_raytraceKernel = b3OpenCLUtils::compileCLKernelFromString(m_data->m_context, m_data->m_device,rayCastKernelCL, "rayCastKernel",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+		m_data->m_raytracePairsKernel = b3OpenCLUtils::compileCLKernelFromString(m_data->m_context, m_data->m_device,rayCastKernelCL, "rayCastPairsKernel",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+		m_data->m_findRayRigidPairIndexRanges = b3OpenCLUtils::compileCLKernelFromString(m_data->m_context, m_data->m_device,rayCastKernelCL, "findRayRigidPairIndexRanges",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+		clReleaseProgram(prog);
+	}
+
+
+}
+
+b3GpuRaycast::~b3GpuRaycast()
+{
+	clReleaseKernel(m_data->m_raytraceKernel);
+	clReleaseKernel(m_data->m_raytracePairsKernel);
+	clReleaseKernel(m_data->m_findRayRigidPairIndexRanges);
+	
+	delete m_data->m_plbvh;
+	delete m_data->m_radixSorter;
+	delete m_data->m_fill;
+	
+	delete m_data->m_gpuRays;
+	delete m_data->m_gpuHitResults;
+	delete m_data->m_firstRayRigidPairIndexPerRay;
+	delete m_data->m_numRayRigidPairsPerRay;
+	delete m_data->m_gpuNumRayRigidPairs;
+	delete m_data->m_gpuRayRigidPairs;
+	
+	delete m_data;
+}
+
+bool sphere_intersect(const b3Vector3& spherePos,  b3Scalar radius, const b3Vector3& rayFrom, const b3Vector3& rayTo, float& hitFraction)
+{
+    b3Vector3 rs = rayFrom - spherePos;
+	b3Vector3 rayDir = rayTo-rayFrom;
+	
+	float A = b3Dot(rayDir,rayDir);
+    float B = b3Dot(rs, rayDir);
+    float C = b3Dot(rs, rs) - (radius * radius);
+    
+	float D = B * B - A*C;
+
+    if (D > 0.0)
+    {
+        float t = (-B - sqrt(D))/A;
+
+        if ( (t >= 0.0f) && (t < hitFraction) )
+        {
+			hitFraction = t;
+            return true;
+		}
+	}
+	return false;
+}
+
+bool rayConvex(const b3Vector3& rayFromLocal, const b3Vector3& rayToLocal, const b3ConvexPolyhedronData& poly,
+	const b3AlignedObjectArray<b3GpuFace>& faces,  float& hitFraction, b3Vector3& hitNormal)
+{
+	float exitFraction = hitFraction;
+	float enterFraction = -0.1f;
+	b3Vector3 curHitNormal=b3MakeVector3(0,0,0);
+	for (int i=0;i<poly.m_numFaces;i++)
+	{
+		const b3GpuFace& face = faces[poly.m_faceOffset+i];
+		float fromPlaneDist = b3Dot(rayFromLocal,face.m_plane)+face.m_plane.w;
+		float toPlaneDist = b3Dot(rayToLocal,face.m_plane)+face.m_plane.w;
+		if (fromPlaneDist<0.f)
+		{
+			if (toPlaneDist >= 0.f)
+			{
+				float fraction = fromPlaneDist / (fromPlaneDist-toPlaneDist);
+				if (exitFraction>fraction)
+				{
+					exitFraction = fraction;
+				}
+			} 			
+		} else
+		{
+			if (toPlaneDist<0.f)
+			{
+				float fraction = fromPlaneDist / (fromPlaneDist-toPlaneDist);
+				if (enterFraction <= fraction)
+				{
+					enterFraction = fraction;
+					curHitNormal = face.m_plane;
+					curHitNormal.w = 0.f;
+				}
+			} else
+			{
+				return false;
+			}
+		}
+		if (exitFraction <= enterFraction)
+			return false;
+	}
+
+	if (enterFraction < 0.f)
+		return false;
+
+	hitFraction = enterFraction;
+	hitNormal = curHitNormal;
+	return true;
+}
+
+void b3GpuRaycast::castRaysHost(const b3AlignedObjectArray<b3RayInfo>& rays,	b3AlignedObjectArray<b3RayHit>& hitResults,
+		int numBodies,const struct b3RigidBodyData* bodies, int numCollidables,const struct b3Collidable* collidables, const struct b3GpuNarrowPhaseInternalData* narrowphaseData)
+{
+
+//	return castRays(rays,hitResults,numBodies,bodies,numCollidables,collidables);
+
+	B3_PROFILE("castRaysHost");
+	for (int r=0;r<rays.size();r++)
+	{
+		b3Vector3 rayFrom = rays[r].m_from;
+		b3Vector3 rayTo = rays[r].m_to;
+		float hitFraction = hitResults[r].m_hitFraction;
+
+		int hitBodyIndex= -1;
+		b3Vector3 hitNormal;
+
+		for (int b=0;b<numBodies;b++)
+		{
+				
+			const b3Vector3& pos = bodies[b].m_pos;
+			//const b3Quaternion& orn = bodies[b].m_quat;
+			
+			switch (collidables[bodies[b].m_collidableIdx].m_shapeType)
+			{
+			case SHAPE_SPHERE:
+				{
+					b3Scalar radius = collidables[bodies[b].m_collidableIdx].m_radius;
+					if (sphere_intersect(pos,  radius, rayFrom, rayTo,hitFraction))
+					{
+						hitBodyIndex = b;
+						b3Vector3 hitPoint;
+						hitPoint.setInterpolate3(rays[r].m_from, rays[r].m_to,hitFraction);
+						hitNormal = (hitPoint-bodies[b].m_pos).normalize();
+					}
+				}
+			case SHAPE_CONVEX_HULL:
+				{
+
+					b3Transform convexWorldTransform;
+					convexWorldTransform.setIdentity();
+					convexWorldTransform.setOrigin(bodies[b].m_pos);
+					convexWorldTransform.setRotation(bodies[b].m_quat);
+					b3Transform convexWorld2Local = convexWorldTransform.inverse();
+
+					b3Vector3 rayFromLocal = convexWorld2Local(rayFrom);
+					b3Vector3 rayToLocal = convexWorld2Local(rayTo);
+					
+					
+					int shapeIndex = collidables[bodies[b].m_collidableIdx].m_shapeIndex;
+					const b3ConvexPolyhedronData& poly = narrowphaseData->m_convexPolyhedra[shapeIndex];
+					if (rayConvex(rayFromLocal, rayToLocal,poly,narrowphaseData->m_convexFaces, hitFraction, hitNormal))
+					{
+						hitBodyIndex = b;
+					}
+
+					
+					break;
+				}
+			default:
+				{
+					static bool once=true;
+					if (once)
+					{
+						once=false;
+						b3Warning("Raytest: unsupported shape type\n");
+					}
+				}
+			}
+		}
+		if (hitBodyIndex>=0)
+		{
+
+			hitResults[r].m_hitFraction = hitFraction;
+			hitResults[r].m_hitPoint.setInterpolate3(rays[r].m_from, rays[r].m_to,hitFraction);
+			hitResults[r].m_hitNormal = hitNormal;
+			hitResults[r].m_hitBody = hitBodyIndex;
+		}
+
+	}
+}
+///todo: add some acceleration structure (AABBs, tree etc)
+void b3GpuRaycast::castRays(const b3AlignedObjectArray<b3RayInfo>& rays,	b3AlignedObjectArray<b3RayHit>& hitResults,
+		int numBodies,const struct b3RigidBodyData* bodies, int numCollidables, const struct b3Collidable* collidables, 
+		const struct b3GpuNarrowPhaseInternalData* narrowphaseData,	class b3GpuBroadphaseInterface* broadphase)
+{
+	//castRaysHost(rays,hitResults,numBodies,bodies,numCollidables,collidables,narrowphaseData);
+
+	B3_PROFILE("castRaysGPU");
+	
+	{
+		B3_PROFILE("raycast copyFromHost");
+		m_data->m_gpuRays->copyFromHost(rays);
+		m_data->m_gpuHitResults->copyFromHost(hitResults);
+		
+	}
+	
+	int numRays = hitResults.size();
+	{
+		m_data->m_firstRayRigidPairIndexPerRay->resize(numRays);
+		m_data->m_numRayRigidPairsPerRay->resize(numRays);
+		
+		m_data->m_gpuNumRayRigidPairs->resize(1);
+		m_data->m_gpuRayRigidPairs->resize(numRays * 16);
+	}
+	
+	//run kernel
+	const bool USE_BRUTE_FORCE_RAYCAST = false;
+	if(USE_BRUTE_FORCE_RAYCAST)
+	{
+		B3_PROFILE("raycast launch1D");
+
+		b3LauncherCL launcher(m_data->m_q,m_data->m_raytraceKernel,"m_raytraceKernel");
+		int numRays = rays.size();
+		launcher.setConst(numRays);
+
+		launcher.setBuffer(m_data->m_gpuRays->getBufferCL());
+		launcher.setBuffer(m_data->m_gpuHitResults->getBufferCL());
+
+		launcher.setConst(numBodies);
+		launcher.setBuffer(narrowphaseData->m_bodyBufferGPU->getBufferCL());
+		launcher.setBuffer(narrowphaseData->m_collidablesGPU->getBufferCL());
+		launcher.setBuffer(narrowphaseData->m_convexFacesGPU->getBufferCL());
+		launcher.setBuffer(narrowphaseData->m_convexPolyhedraGPU->getBufferCL());
+		
+		launcher.launch1D(numRays);
+		clFinish(m_data->m_q);
+	}
+	else
+	{
+		m_data->m_plbvh->build( broadphase->getAllAabbsGPU(), broadphase->getSmallAabbIndicesGPU(), broadphase->getLargeAabbIndicesGPU() );
+
+		m_data->m_plbvh->testRaysAgainstBvhAabbs(*m_data->m_gpuRays, *m_data->m_gpuNumRayRigidPairs, *m_data->m_gpuRayRigidPairs);
+		
+		int numRayRigidPairs = -1;
+		m_data->m_gpuNumRayRigidPairs->copyToHostPointer(&numRayRigidPairs, 1);
+		if( numRayRigidPairs > m_data->m_gpuRayRigidPairs->size() )
+		{
+			numRayRigidPairs = m_data->m_gpuRayRigidPairs->size();
+			m_data->m_gpuNumRayRigidPairs->copyFromHostPointer(&numRayRigidPairs, 1);
+		}
+		
+		m_data->m_gpuRayRigidPairs->resize(numRayRigidPairs);	//Radix sort needs b3OpenCLArray::size() to be correct
+		
+		//Sort ray-rigid pairs by ray index
+		{
+			B3_PROFILE("sort ray-rigid pairs");
+			m_data->m_radixSorter->execute( *reinterpret_cast< b3OpenCLArray<b3SortData>* >(m_data->m_gpuRayRigidPairs) );
+		}
+		
+		//detect start,count of each ray pair
+		{
+			B3_PROFILE("detect ray-rigid pair index ranges");
+			
+			{
+				B3_PROFILE("reset ray-rigid pair index ranges");
+				
+				m_data->m_fill->execute(*m_data->m_firstRayRigidPairIndexPerRay, numRayRigidPairs, numRays);	//atomic_min used to find first index
+				m_data->m_fill->execute(*m_data->m_numRayRigidPairsPerRay, 0, numRays);
+				clFinish(m_data->m_q);
+			}
+			
+			b3BufferInfoCL bufferInfo[] = 
+			{
+				b3BufferInfoCL( m_data->m_gpuRayRigidPairs->getBufferCL() ),
+				
+				b3BufferInfoCL( m_data->m_firstRayRigidPairIndexPerRay->getBufferCL() ),
+				b3BufferInfoCL( m_data->m_numRayRigidPairsPerRay->getBufferCL() )
+			};
+			
+			b3LauncherCL launcher(m_data->m_q, m_data->m_findRayRigidPairIndexRanges, "m_findRayRigidPairIndexRanges");
+			launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(numRayRigidPairs);
+			
+			launcher.launch1D(numRayRigidPairs);
+			clFinish(m_data->m_q);
+		}
+		
+		{
+			B3_PROFILE("ray-rigid intersection");
+			
+			b3BufferInfoCL bufferInfo[] = 
+			{
+				b3BufferInfoCL( m_data->m_gpuRays->getBufferCL() ),
+				b3BufferInfoCL( m_data->m_gpuHitResults->getBufferCL() ),
+				b3BufferInfoCL( m_data->m_firstRayRigidPairIndexPerRay->getBufferCL() ),
+				b3BufferInfoCL( m_data->m_numRayRigidPairsPerRay->getBufferCL() ),
+				
+				b3BufferInfoCL( narrowphaseData->m_bodyBufferGPU->getBufferCL() ),
+				b3BufferInfoCL( narrowphaseData->m_collidablesGPU->getBufferCL() ),
+				b3BufferInfoCL( narrowphaseData->m_convexFacesGPU->getBufferCL() ),
+				b3BufferInfoCL( narrowphaseData->m_convexPolyhedraGPU->getBufferCL() ),
+				
+				b3BufferInfoCL( m_data->m_gpuRayRigidPairs->getBufferCL() )
+			};
+			
+			b3LauncherCL launcher(m_data->m_q, m_data->m_raytracePairsKernel, "m_raytracePairsKernel");
+			launcher.setBuffers( bufferInfo, sizeof(bufferInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(numRays);
+			
+			launcher.launch1D(numRays);
+			clFinish(m_data->m_q);
+		}
+	}
+	
+	
+
+	//copy results
+	{
+		B3_PROFILE("raycast copyToHost");
+		m_data->m_gpuHitResults->copyToHost(hitResults);
+	}
+
+}
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/Raycast/b3GpuRaycast.h b/thirdparty/bullet/Bullet3OpenCL/Raycast/b3GpuRaycast.h
new file mode 100644
index 0000000000..3a5cf44b79
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/Raycast/b3GpuRaycast.h
@@ -0,0 +1,32 @@
+#ifndef B3_GPU_RAYCAST_H
+#define B3_GPU_RAYCAST_H
+
+#include "Bullet3Common/b3Vector3.h"
+#include "Bullet3OpenCL/Initialize/b3OpenCLInclude.h"
+
+#include "Bullet3Common/b3AlignedObjectArray.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3RaycastInfo.h"
+
+
+
+class b3GpuRaycast
+{
+protected:
+	struct b3GpuRaycastInternalData* m_data;
+public:
+	b3GpuRaycast(cl_context ctx,cl_device_id device, cl_command_queue  q);
+	virtual ~b3GpuRaycast();
+
+	void castRaysHost(const b3AlignedObjectArray<b3RayInfo>& raysIn,	b3AlignedObjectArray<b3RayHit>& hitResults,
+		int numBodies, const struct b3RigidBodyData* bodies, int numCollidables, const struct b3Collidable* collidables,
+		const struct b3GpuNarrowPhaseInternalData* narrowphaseData);
+
+	void castRays(const b3AlignedObjectArray<b3RayInfo>& rays,	b3AlignedObjectArray<b3RayHit>& hitResults,
+		int numBodies,const struct b3RigidBodyData* bodies, int numCollidables, const struct b3Collidable* collidables,
+		const struct b3GpuNarrowPhaseInternalData* narrowphaseData, class b3GpuBroadphaseInterface* broadphase);
+	
+
+		
+};
+
+#endif //B3_GPU_RAYCAST_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/Raycast/kernels/rayCastKernels.cl b/thirdparty/bullet/Bullet3OpenCL/Raycast/kernels/rayCastKernels.cl
new file mode 100644
index 0000000000..e72d96876b
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/Raycast/kernels/rayCastKernels.cl
@@ -0,0 +1,439 @@
+
+#define SHAPE_CONVEX_HULL 3
+#define SHAPE_PLANE 4
+#define SHAPE_CONCAVE_TRIMESH 5
+#define SHAPE_COMPOUND_OF_CONVEX_HULLS 6
+#define SHAPE_SPHERE 7
+
+
+typedef struct
+{
+	float4 m_from;
+	float4 m_to;
+} b3RayInfo;
+
+typedef struct
+{
+	float m_hitFraction;
+	int	m_hitResult0;
+	int	m_hitResult1;
+	int	m_hitResult2;
+	float4	m_hitPoint;
+	float4	m_hitNormal;
+} b3RayHit;
+
+typedef struct
+{
+	float4 m_pos;
+	float4 m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	unsigned int m_collidableIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} Body;
+
+typedef struct Collidable
+{
+	union {
+		int m_numChildShapes;
+		int m_bvhIndex;
+	};
+	float m_radius;
+	int m_shapeType;
+	int m_shapeIndex;
+} Collidable;
+
+
+typedef struct  
+{
+	float4		m_localCenter;
+	float4		m_extents;
+	float4		mC;
+	float4		mE;
+
+	float			m_radius;
+	int	m_faceOffset;
+	int m_numFaces;
+	int	m_numVertices;
+
+	int m_vertexOffset;
+	int	m_uniqueEdgesOffset;
+	int	m_numUniqueEdges;
+	int m_unused;
+
+} ConvexPolyhedronCL;
+
+typedef struct
+{
+	float4 m_plane;
+	int m_indexOffset;
+	int m_numIndices;
+} b3GpuFace;
+
+
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+__inline
+	Quaternion qtMul(Quaternion a, Quaternion b);
+
+__inline
+	Quaternion qtNormalize(Quaternion in);
+
+
+__inline
+	Quaternion qtInvert(Quaternion q);
+
+
+__inline
+	float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = (float4)(a.xyz,0.f);
+	float4 b1 = (float4)(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+
+__inline
+	Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross( a, b );
+	ans += a.w*b+b.w*a;
+	//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+	Quaternion qtNormalize(Quaternion in)
+{
+	return fast_normalize(in);
+	//	in /= length( in );
+	//	return in;
+}
+__inline
+	float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(q,vcpy);
+	out = qtMul(out,qInv);
+	return out;
+}
+
+__inline
+	Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+__inline
+	float4 qtInvRotate(const Quaternion q, float4 vec)
+{
+	return qtRotate( qtInvert( q ), vec );
+}
+
+
+
+void	trInverse(float4 translationIn, Quaternion orientationIn,
+	float4* translationOut, Quaternion* orientationOut)
+{
+	*orientationOut = qtInvert(orientationIn);
+	*translationOut = qtRotate(*orientationOut, -translationIn);
+}
+
+
+
+
+
+bool rayConvex(float4 rayFromLocal, float4 rayToLocal, int numFaces, int faceOffset,
+	__global const b3GpuFace* faces, float* hitFraction, float4* hitNormal)
+{
+	rayFromLocal.w = 0.f;
+	rayToLocal.w = 0.f;
+	bool result = true;
+
+	float exitFraction = hitFraction[0];
+	float enterFraction = -0.3f;
+	float4 curHitNormal = (float4)(0,0,0,0);
+	for (int i=0;i<numFaces && result;i++)
+	{
+		b3GpuFace face = faces[faceOffset+i];
+		float fromPlaneDist = dot(rayFromLocal,face.m_plane)+face.m_plane.w;
+		float toPlaneDist = dot(rayToLocal,face.m_plane)+face.m_plane.w;
+		if (fromPlaneDist<0.f)
+		{
+			if (toPlaneDist >= 0.f)
+			{
+				float fraction = fromPlaneDist / (fromPlaneDist-toPlaneDist);
+				if (exitFraction>fraction)
+				{
+					exitFraction = fraction;
+				}
+			} 			
+		} else
+		{
+			if (toPlaneDist<0.f)
+			{
+				float fraction = fromPlaneDist / (fromPlaneDist-toPlaneDist);
+				if (enterFraction <= fraction)
+				{
+					enterFraction = fraction;
+					curHitNormal = face.m_plane;
+					curHitNormal.w = 0.f;
+				}
+			} else
+			{
+				result = false;
+			}
+		}
+		if (exitFraction <= enterFraction)
+			result = false;
+	}
+
+	if (enterFraction < 0.f)
+	{
+		result = false;
+	}
+
+	if (result)
+	{	
+		hitFraction[0] = enterFraction;
+		hitNormal[0] = curHitNormal;
+	}
+	return result;
+}
+
+
+
+
+
+
+bool sphere_intersect(float4 spherePos,  float radius, float4 rayFrom, float4 rayTo, float* hitFraction)
+{
+	float4 rs = rayFrom - spherePos;
+	rs.w = 0.f;
+	float4 rayDir = rayTo-rayFrom;
+	rayDir.w = 0.f;
+	float A = dot(rayDir,rayDir);
+	float B = dot(rs, rayDir);
+	float C = dot(rs, rs) - (radius * radius);
+
+	float D = B * B - A*C;
+
+	if (D > 0.0f)
+	{
+		float t = (-B - sqrt(D))/A;
+
+		if ( (t >= 0.0f) && (t < (*hitFraction)) )
+		{
+			*hitFraction = t;
+			return true;
+		}
+	}
+	return false;
+}
+
+float4 setInterpolate3(float4 from, float4 to, float t)
+{
+	float s = 1.0f - t;
+	float4 result;
+	result = s * from + t * to;
+	result.w = 0.f;	
+	return result;	
+}
+
+__kernel void rayCastKernel(  
+	int numRays, 
+	const __global b3RayInfo* rays, 
+	__global b3RayHit* hitResults, 
+	const int numBodies, 
+	__global Body* bodies,
+	__global Collidable* collidables,
+	__global const b3GpuFace* faces,
+	__global const ConvexPolyhedronCL* convexShapes	)
+{
+
+	int i = get_global_id(0);
+	if (i>=numRays)
+		return;
+
+	hitResults[i].m_hitFraction = 1.f;
+
+	float4 rayFrom = rays[i].m_from;
+	float4 rayTo = rays[i].m_to;
+	float hitFraction = 1.f;
+	float4 hitPoint;
+	float4 hitNormal;
+	int hitBodyIndex= -1;
+
+	int cachedCollidableIndex = -1;
+	Collidable cachedCollidable;
+
+	for (int b=0;b<numBodies;b++)
+	{
+		if (hitResults[i].m_hitResult2==b)
+			continue;
+		Body body = bodies[b];
+		float4 pos = body.m_pos;
+		float4 orn = body.m_quat;
+		if (cachedCollidableIndex != body.m_collidableIdx)
+		{
+			cachedCollidableIndex = body.m_collidableIdx;
+			cachedCollidable = collidables[cachedCollidableIndex];
+		}
+		if (cachedCollidable.m_shapeType == SHAPE_CONVEX_HULL)
+		{
+
+			float4 invPos = (float4)(0,0,0,0);
+			float4 invOrn = (float4)(0,0,0,0);
+			float4 rayFromLocal = (float4)(0,0,0,0);
+			float4 rayToLocal = (float4)(0,0,0,0);
+			invOrn = qtInvert(orn);
+			invPos = qtRotate(invOrn, -pos);
+			rayFromLocal = qtRotate( invOrn, rayFrom ) + invPos;
+			rayToLocal = qtRotate( invOrn, rayTo) + invPos;
+			rayFromLocal.w = 0.f;
+			rayToLocal.w = 0.f;
+			int numFaces = convexShapes[cachedCollidable.m_shapeIndex].m_numFaces;
+			int faceOffset = convexShapes[cachedCollidable.m_shapeIndex].m_faceOffset;
+			if (numFaces)
+			{
+				if (rayConvex(rayFromLocal, rayToLocal, numFaces, faceOffset,faces, &hitFraction, &hitNormal))
+				{
+					hitBodyIndex = b;
+					
+				}
+			}
+		}
+		if (cachedCollidable.m_shapeType == SHAPE_SPHERE)
+		{
+			float radius = cachedCollidable.m_radius;
+		
+			if (sphere_intersect(pos,  radius, rayFrom, rayTo, &hitFraction))
+			{
+				hitBodyIndex = b;
+				hitNormal = (float4) (hitPoint-bodies[b].m_pos);
+			}
+		}
+	}
+
+	if (hitBodyIndex>=0)
+	{
+		hitPoint = setInterpolate3(rayFrom, rayTo,hitFraction);
+		hitResults[i].m_hitFraction = hitFraction;
+		hitResults[i].m_hitPoint = hitPoint;
+		hitResults[i].m_hitNormal = normalize(hitNormal);
+		hitResults[i].m_hitResult0 = hitBodyIndex;
+	}
+
+}
+
+
+__kernel void findRayRigidPairIndexRanges(__global int2* rayRigidPairs, 
+											__global int* out_firstRayRigidPairIndexPerRay,
+											__global int* out_numRayRigidPairsPerRay,
+											int numRayRigidPairs)
+{
+	int rayRigidPairIndex = get_global_id(0);
+	if (rayRigidPairIndex >= numRayRigidPairs) return;
+	
+	int rayIndex = rayRigidPairs[rayRigidPairIndex].x;
+	
+	atomic_min(&out_firstRayRigidPairIndexPerRay[rayIndex], rayRigidPairIndex);
+	atomic_inc(&out_numRayRigidPairsPerRay[rayIndex]);
+}
+
+__kernel void rayCastPairsKernel(const __global b3RayInfo* rays, 
+								__global b3RayHit* hitResults, 
+								__global int* firstRayRigidPairIndexPerRay,
+								__global int* numRayRigidPairsPerRay,
+									
+								__global Body* bodies,
+								__global Collidable* collidables,
+								__global const b3GpuFace* faces,
+								__global const ConvexPolyhedronCL* convexShapes,
+								
+								__global int2* rayRigidPairs,
+								int numRays)
+{
+	int i = get_global_id(0);
+	if (i >= numRays) return;
+	
+	float4 rayFrom = rays[i].m_from;
+	float4 rayTo = rays[i].m_to;
+		
+	hitResults[i].m_hitFraction = 1.f;
+		
+	float hitFraction = 1.f;
+	float4 hitPoint;
+	float4 hitNormal;
+	int hitBodyIndex = -1;
+		
+	//
+	for(int pair = 0; pair < numRayRigidPairsPerRay[i]; ++pair)
+	{
+		int rayRigidPairIndex = pair + firstRayRigidPairIndexPerRay[i];
+		int b = rayRigidPairs[rayRigidPairIndex].y;
+		
+		if (hitResults[i].m_hitResult2 == b) continue;
+		
+		Body body = bodies[b];
+		Collidable rigidCollidable = collidables[body.m_collidableIdx];
+		
+		float4 pos = body.m_pos;
+		float4 orn = body.m_quat;
+		
+		if (rigidCollidable.m_shapeType == SHAPE_CONVEX_HULL)
+		{
+			float4 invPos = (float4)(0,0,0,0);
+			float4 invOrn = (float4)(0,0,0,0);
+			float4 rayFromLocal = (float4)(0,0,0,0);
+			float4 rayToLocal = (float4)(0,0,0,0);
+			invOrn = qtInvert(orn);
+			invPos = qtRotate(invOrn, -pos);
+			rayFromLocal = qtRotate( invOrn, rayFrom ) + invPos;
+			rayToLocal = qtRotate( invOrn, rayTo) + invPos;
+			rayFromLocal.w = 0.f;
+			rayToLocal.w = 0.f;
+			int numFaces = convexShapes[rigidCollidable.m_shapeIndex].m_numFaces;
+			int faceOffset = convexShapes[rigidCollidable.m_shapeIndex].m_faceOffset;
+			
+			if (numFaces && rayConvex(rayFromLocal, rayToLocal, numFaces, faceOffset,faces, &hitFraction, &hitNormal))
+			{
+				hitBodyIndex = b;
+				hitPoint = setInterpolate3(rayFrom, rayTo, hitFraction);
+			}
+		}
+		
+		if (rigidCollidable.m_shapeType == SHAPE_SPHERE)
+		{
+			float radius = rigidCollidable.m_radius;
+		
+			if (sphere_intersect(pos, radius, rayFrom, rayTo, &hitFraction))
+			{
+				hitBodyIndex = b;
+				hitPoint = setInterpolate3(rayFrom, rayTo, hitFraction);
+				hitNormal = (float4) (hitPoint - bodies[b].m_pos);
+			}
+		}
+	}
+	
+	if (hitBodyIndex >= 0)
+	{
+		hitResults[i].m_hitFraction = hitFraction;
+		hitResults[i].m_hitPoint = hitPoint;
+		hitResults[i].m_hitNormal = normalize(hitNormal);
+		hitResults[i].m_hitResult0 = hitBodyIndex;
+	}
+	
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/Raycast/kernels/rayCastKernels.h b/thirdparty/bullet/Bullet3OpenCL/Raycast/kernels/rayCastKernels.h
new file mode 100644
index 0000000000..6257909a4d
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/Raycast/kernels/rayCastKernels.h
@@ -0,0 +1,381 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* rayCastKernelCL= \
+"#define SHAPE_CONVEX_HULL 3\n"
+"#define SHAPE_PLANE 4\n"
+"#define SHAPE_CONCAVE_TRIMESH 5\n"
+"#define SHAPE_COMPOUND_OF_CONVEX_HULLS 6\n"
+"#define SHAPE_SPHERE 7\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_from;\n"
+"	float4 m_to;\n"
+"} b3RayInfo;\n"
+"typedef struct\n"
+"{\n"
+"	float m_hitFraction;\n"
+"	int	m_hitResult0;\n"
+"	int	m_hitResult1;\n"
+"	int	m_hitResult2;\n"
+"	float4	m_hitPoint;\n"
+"	float4	m_hitNormal;\n"
+"} b3RayHit;\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	float4 m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"	unsigned int m_collidableIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} Body;\n"
+"typedef struct Collidable\n"
+"{\n"
+"	union {\n"
+"		int m_numChildShapes;\n"
+"		int m_bvhIndex;\n"
+"	};\n"
+"	float m_radius;\n"
+"	int m_shapeType;\n"
+"	int m_shapeIndex;\n"
+"} Collidable;\n"
+"typedef struct  \n"
+"{\n"
+"	float4		m_localCenter;\n"
+"	float4		m_extents;\n"
+"	float4		mC;\n"
+"	float4		mE;\n"
+"	float			m_radius;\n"
+"	int	m_faceOffset;\n"
+"	int m_numFaces;\n"
+"	int	m_numVertices;\n"
+"	int m_vertexOffset;\n"
+"	int	m_uniqueEdgesOffset;\n"
+"	int	m_numUniqueEdges;\n"
+"	int m_unused;\n"
+"} ConvexPolyhedronCL;\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_plane;\n"
+"	int m_indexOffset;\n"
+"	int m_numIndices;\n"
+"} b3GpuFace;\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"typedef float4 Quaternion;\n"
+"__inline\n"
+"	Quaternion qtMul(Quaternion a, Quaternion b);\n"
+"__inline\n"
+"	Quaternion qtNormalize(Quaternion in);\n"
+"__inline\n"
+"	Quaternion qtInvert(Quaternion q);\n"
+"__inline\n"
+"	float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = (float4)(a.xyz,0.f);\n"
+"	float4 b1 = (float4)(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"__inline\n"
+"	Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"	//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"	Quaternion qtNormalize(Quaternion in)\n"
+"{\n"
+"	return fast_normalize(in);\n"
+"	//	in /= length( in );\n"
+"	//	return in;\n"
+"}\n"
+"__inline\n"
+"	float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(q,vcpy);\n"
+"	out = qtMul(out,qInv);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"	Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"__inline\n"
+"	float4 qtInvRotate(const Quaternion q, float4 vec)\n"
+"{\n"
+"	return qtRotate( qtInvert( q ), vec );\n"
+"}\n"
+"void	trInverse(float4 translationIn, Quaternion orientationIn,\n"
+"	float4* translationOut, Quaternion* orientationOut)\n"
+"{\n"
+"	*orientationOut = qtInvert(orientationIn);\n"
+"	*translationOut = qtRotate(*orientationOut, -translationIn);\n"
+"}\n"
+"bool rayConvex(float4 rayFromLocal, float4 rayToLocal, int numFaces, int faceOffset,\n"
+"	__global const b3GpuFace* faces, float* hitFraction, float4* hitNormal)\n"
+"{\n"
+"	rayFromLocal.w = 0.f;\n"
+"	rayToLocal.w = 0.f;\n"
+"	bool result = true;\n"
+"	float exitFraction = hitFraction[0];\n"
+"	float enterFraction = -0.3f;\n"
+"	float4 curHitNormal = (float4)(0,0,0,0);\n"
+"	for (int i=0;i<numFaces && result;i++)\n"
+"	{\n"
+"		b3GpuFace face = faces[faceOffset+i];\n"
+"		float fromPlaneDist = dot(rayFromLocal,face.m_plane)+face.m_plane.w;\n"
+"		float toPlaneDist = dot(rayToLocal,face.m_plane)+face.m_plane.w;\n"
+"		if (fromPlaneDist<0.f)\n"
+"		{\n"
+"			if (toPlaneDist >= 0.f)\n"
+"			{\n"
+"				float fraction = fromPlaneDist / (fromPlaneDist-toPlaneDist);\n"
+"				if (exitFraction>fraction)\n"
+"				{\n"
+"					exitFraction = fraction;\n"
+"				}\n"
+"			} 			\n"
+"		} else\n"
+"		{\n"
+"			if (toPlaneDist<0.f)\n"
+"			{\n"
+"				float fraction = fromPlaneDist / (fromPlaneDist-toPlaneDist);\n"
+"				if (enterFraction <= fraction)\n"
+"				{\n"
+"					enterFraction = fraction;\n"
+"					curHitNormal = face.m_plane;\n"
+"					curHitNormal.w = 0.f;\n"
+"				}\n"
+"			} else\n"
+"			{\n"
+"				result = false;\n"
+"			}\n"
+"		}\n"
+"		if (exitFraction <= enterFraction)\n"
+"			result = false;\n"
+"	}\n"
+"	if (enterFraction < 0.f)\n"
+"	{\n"
+"		result = false;\n"
+"	}\n"
+"	if (result)\n"
+"	{	\n"
+"		hitFraction[0] = enterFraction;\n"
+"		hitNormal[0] = curHitNormal;\n"
+"	}\n"
+"	return result;\n"
+"}\n"
+"bool sphere_intersect(float4 spherePos,  float radius, float4 rayFrom, float4 rayTo, float* hitFraction)\n"
+"{\n"
+"	float4 rs = rayFrom - spherePos;\n"
+"	rs.w = 0.f;\n"
+"	float4 rayDir = rayTo-rayFrom;\n"
+"	rayDir.w = 0.f;\n"
+"	float A = dot(rayDir,rayDir);\n"
+"	float B = dot(rs, rayDir);\n"
+"	float C = dot(rs, rs) - (radius * radius);\n"
+"	float D = B * B - A*C;\n"
+"	if (D > 0.0f)\n"
+"	{\n"
+"		float t = (-B - sqrt(D))/A;\n"
+"		if ( (t >= 0.0f) && (t < (*hitFraction)) )\n"
+"		{\n"
+"			*hitFraction = t;\n"
+"			return true;\n"
+"		}\n"
+"	}\n"
+"	return false;\n"
+"}\n"
+"float4 setInterpolate3(float4 from, float4 to, float t)\n"
+"{\n"
+"	float s = 1.0f - t;\n"
+"	float4 result;\n"
+"	result = s * from + t * to;\n"
+"	result.w = 0.f;	\n"
+"	return result;	\n"
+"}\n"
+"__kernel void rayCastKernel(  \n"
+"	int numRays, \n"
+"	const __global b3RayInfo* rays, \n"
+"	__global b3RayHit* hitResults, \n"
+"	const int numBodies, \n"
+"	__global Body* bodies,\n"
+"	__global Collidable* collidables,\n"
+"	__global const b3GpuFace* faces,\n"
+"	__global const ConvexPolyhedronCL* convexShapes	)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numRays)\n"
+"		return;\n"
+"	hitResults[i].m_hitFraction = 1.f;\n"
+"	float4 rayFrom = rays[i].m_from;\n"
+"	float4 rayTo = rays[i].m_to;\n"
+"	float hitFraction = 1.f;\n"
+"	float4 hitPoint;\n"
+"	float4 hitNormal;\n"
+"	int hitBodyIndex= -1;\n"
+"	int cachedCollidableIndex = -1;\n"
+"	Collidable cachedCollidable;\n"
+"	for (int b=0;b<numBodies;b++)\n"
+"	{\n"
+"		if (hitResults[i].m_hitResult2==b)\n"
+"			continue;\n"
+"		Body body = bodies[b];\n"
+"		float4 pos = body.m_pos;\n"
+"		float4 orn = body.m_quat;\n"
+"		if (cachedCollidableIndex != body.m_collidableIdx)\n"
+"		{\n"
+"			cachedCollidableIndex = body.m_collidableIdx;\n"
+"			cachedCollidable = collidables[cachedCollidableIndex];\n"
+"		}\n"
+"		if (cachedCollidable.m_shapeType == SHAPE_CONVEX_HULL)\n"
+"		{\n"
+"			float4 invPos = (float4)(0,0,0,0);\n"
+"			float4 invOrn = (float4)(0,0,0,0);\n"
+"			float4 rayFromLocal = (float4)(0,0,0,0);\n"
+"			float4 rayToLocal = (float4)(0,0,0,0);\n"
+"			invOrn = qtInvert(orn);\n"
+"			invPos = qtRotate(invOrn, -pos);\n"
+"			rayFromLocal = qtRotate( invOrn, rayFrom ) + invPos;\n"
+"			rayToLocal = qtRotate( invOrn, rayTo) + invPos;\n"
+"			rayFromLocal.w = 0.f;\n"
+"			rayToLocal.w = 0.f;\n"
+"			int numFaces = convexShapes[cachedCollidable.m_shapeIndex].m_numFaces;\n"
+"			int faceOffset = convexShapes[cachedCollidable.m_shapeIndex].m_faceOffset;\n"
+"			if (numFaces)\n"
+"			{\n"
+"				if (rayConvex(rayFromLocal, rayToLocal, numFaces, faceOffset,faces, &hitFraction, &hitNormal))\n"
+"				{\n"
+"					hitBodyIndex = b;\n"
+"					\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"		if (cachedCollidable.m_shapeType == SHAPE_SPHERE)\n"
+"		{\n"
+"			float radius = cachedCollidable.m_radius;\n"
+"		\n"
+"			if (sphere_intersect(pos,  radius, rayFrom, rayTo, &hitFraction))\n"
+"			{\n"
+"				hitBodyIndex = b;\n"
+"				hitNormal = (float4) (hitPoint-bodies[b].m_pos);\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	if (hitBodyIndex>=0)\n"
+"	{\n"
+"		hitPoint = setInterpolate3(rayFrom, rayTo,hitFraction);\n"
+"		hitResults[i].m_hitFraction = hitFraction;\n"
+"		hitResults[i].m_hitPoint = hitPoint;\n"
+"		hitResults[i].m_hitNormal = normalize(hitNormal);\n"
+"		hitResults[i].m_hitResult0 = hitBodyIndex;\n"
+"	}\n"
+"}\n"
+"__kernel void findRayRigidPairIndexRanges(__global int2* rayRigidPairs, \n"
+"											__global int* out_firstRayRigidPairIndexPerRay,\n"
+"											__global int* out_numRayRigidPairsPerRay,\n"
+"											int numRayRigidPairs)\n"
+"{\n"
+"	int rayRigidPairIndex = get_global_id(0);\n"
+"	if (rayRigidPairIndex >= numRayRigidPairs) return;\n"
+"	\n"
+"	int rayIndex = rayRigidPairs[rayRigidPairIndex].x;\n"
+"	\n"
+"	atomic_min(&out_firstRayRigidPairIndexPerRay[rayIndex], rayRigidPairIndex);\n"
+"	atomic_inc(&out_numRayRigidPairsPerRay[rayIndex]);\n"
+"}\n"
+"__kernel void rayCastPairsKernel(const __global b3RayInfo* rays, \n"
+"								__global b3RayHit* hitResults, \n"
+"								__global int* firstRayRigidPairIndexPerRay,\n"
+"								__global int* numRayRigidPairsPerRay,\n"
+"									\n"
+"								__global Body* bodies,\n"
+"								__global Collidable* collidables,\n"
+"								__global const b3GpuFace* faces,\n"
+"								__global const ConvexPolyhedronCL* convexShapes,\n"
+"								\n"
+"								__global int2* rayRigidPairs,\n"
+"								int numRays)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i >= numRays) return;\n"
+"	\n"
+"	float4 rayFrom = rays[i].m_from;\n"
+"	float4 rayTo = rays[i].m_to;\n"
+"		\n"
+"	hitResults[i].m_hitFraction = 1.f;\n"
+"		\n"
+"	float hitFraction = 1.f;\n"
+"	float4 hitPoint;\n"
+"	float4 hitNormal;\n"
+"	int hitBodyIndex = -1;\n"
+"		\n"
+"	//\n"
+"	for(int pair = 0; pair < numRayRigidPairsPerRay[i]; ++pair)\n"
+"	{\n"
+"		int rayRigidPairIndex = pair + firstRayRigidPairIndexPerRay[i];\n"
+"		int b = rayRigidPairs[rayRigidPairIndex].y;\n"
+"		\n"
+"		if (hitResults[i].m_hitResult2 == b) continue;\n"
+"		\n"
+"		Body body = bodies[b];\n"
+"		Collidable rigidCollidable = collidables[body.m_collidableIdx];\n"
+"		\n"
+"		float4 pos = body.m_pos;\n"
+"		float4 orn = body.m_quat;\n"
+"		\n"
+"		if (rigidCollidable.m_shapeType == SHAPE_CONVEX_HULL)\n"
+"		{\n"
+"			float4 invPos = (float4)(0,0,0,0);\n"
+"			float4 invOrn = (float4)(0,0,0,0);\n"
+"			float4 rayFromLocal = (float4)(0,0,0,0);\n"
+"			float4 rayToLocal = (float4)(0,0,0,0);\n"
+"			invOrn = qtInvert(orn);\n"
+"			invPos = qtRotate(invOrn, -pos);\n"
+"			rayFromLocal = qtRotate( invOrn, rayFrom ) + invPos;\n"
+"			rayToLocal = qtRotate( invOrn, rayTo) + invPos;\n"
+"			rayFromLocal.w = 0.f;\n"
+"			rayToLocal.w = 0.f;\n"
+"			int numFaces = convexShapes[rigidCollidable.m_shapeIndex].m_numFaces;\n"
+"			int faceOffset = convexShapes[rigidCollidable.m_shapeIndex].m_faceOffset;\n"
+"			\n"
+"			if (numFaces && rayConvex(rayFromLocal, rayToLocal, numFaces, faceOffset,faces, &hitFraction, &hitNormal))\n"
+"			{\n"
+"				hitBodyIndex = b;\n"
+"				hitPoint = setInterpolate3(rayFrom, rayTo, hitFraction);\n"
+"			}\n"
+"		}\n"
+"		\n"
+"		if (rigidCollidable.m_shapeType == SHAPE_SPHERE)\n"
+"		{\n"
+"			float radius = rigidCollidable.m_radius;\n"
+"		\n"
+"			if (sphere_intersect(pos, radius, rayFrom, rayTo, &hitFraction))\n"
+"			{\n"
+"				hitBodyIndex = b;\n"
+"				hitPoint = setInterpolate3(rayFrom, rayTo, hitFraction);\n"
+"				hitNormal = (float4) (hitPoint - bodies[b].m_pos);\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"	\n"
+"	if (hitBodyIndex >= 0)\n"
+"	{\n"
+"		hitResults[i].m_hitFraction = hitFraction;\n"
+"		hitResults[i].m_hitPoint = hitPoint;\n"
+"		hitResults[i].m_hitNormal = normalize(hitNormal);\n"
+"		hitResults[i].m_hitResult0 = hitBodyIndex;\n"
+"	}\n"
+"	\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuConstraint4.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuConstraint4.h
new file mode 100644
index 0000000000..c7478f54a1
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuConstraint4.h
@@ -0,0 +1,18 @@
+
+#ifndef B3_CONSTRAINT4_h
+#define B3_CONSTRAINT4_h
+#include "Bullet3Common/b3Vector3.h"
+
+#include "Bullet3Dynamics/shared/b3ContactConstraint4.h"
+
+
+B3_ATTRIBUTE_ALIGNED16(struct) b3GpuConstraint4 : public b3ContactConstraint4
+{
+	B3_DECLARE_ALIGNED_ALLOCATOR();
+
+	inline	void setFrictionCoeff(float value) { m_linear[3] = value; }
+	inline	float getFrictionCoeff() const { return m_linear[3]; }
+};
+
+#endif //B3_CONSTRAINT4_h
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuGenericConstraint.cpp b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuGenericConstraint.cpp
new file mode 100644
index 0000000000..af687b54e9
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuGenericConstraint.cpp
@@ -0,0 +1,137 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+#include "b3GpuGenericConstraint.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3RigidBodyData.h"
+
+#include <new>
+#include "Bullet3Common/b3Transform.h"
+
+void b3GpuGenericConstraint::getInfo1 (unsigned int* info,const b3RigidBodyData* bodies)
+{
+	switch (m_constraintType)
+	{
+	case B3_GPU_POINT2POINT_CONSTRAINT_TYPE:
+		{
+			*info = 3;
+			break;
+		};
+		default:
+		{
+			b3Assert(0);
+		}
+	};
+}
+
+void getInfo2Point2Point(b3GpuGenericConstraint* constraint, b3GpuConstraintInfo2* info,  const b3RigidBodyData* bodies)
+{
+	b3Transform trA;
+	trA.setIdentity();
+	trA.setOrigin(bodies[constraint->m_rbA].m_pos);
+	trA.setRotation(bodies[constraint->m_rbA].m_quat);
+
+	b3Transform trB;
+	trB.setIdentity();
+	trB.setOrigin(bodies[constraint->m_rbB].m_pos);
+	trB.setRotation(bodies[constraint->m_rbB].m_quat);
+
+		// anchor points in global coordinates with respect to body PORs.
+   
+    // set jacobian
+    info->m_J1linearAxis[0] = 1;
+	info->m_J1linearAxis[info->rowskip+1] = 1;
+	info->m_J1linearAxis[2*info->rowskip+2] = 1;
+
+	b3Vector3 a1 = trA.getBasis()*constraint->getPivotInA();
+	//b3Vector3 a1a = b3QuatRotate(trA.getRotation(),constraint->getPivotInA());
+
+	{
+		b3Vector3* angular0 = (b3Vector3*)(info->m_J1angularAxis);
+		b3Vector3* angular1 = (b3Vector3*)(info->m_J1angularAxis+info->rowskip);
+		b3Vector3* angular2 = (b3Vector3*)(info->m_J1angularAxis+2*info->rowskip);
+		b3Vector3 a1neg = -a1;
+		a1neg.getSkewSymmetricMatrix(angular0,angular1,angular2);
+	}
+    
+	if (info->m_J2linearAxis)
+	{
+		info->m_J2linearAxis[0] = -1;
+		info->m_J2linearAxis[info->rowskip+1] = -1;
+		info->m_J2linearAxis[2*info->rowskip+2] = -1;
+	}
+	
+	b3Vector3 a2 = trB.getBasis()*constraint->getPivotInB();
+   
+	{
+	//	b3Vector3 a2n = -a2;
+		b3Vector3* angular0 = (b3Vector3*)(info->m_J2angularAxis);
+		b3Vector3* angular1 = (b3Vector3*)(info->m_J2angularAxis+info->rowskip);
+		b3Vector3* angular2 = (b3Vector3*)(info->m_J2angularAxis+2*info->rowskip);
+		a2.getSkewSymmetricMatrix(angular0,angular1,angular2);
+	}
+    
+
+
+    // set right hand side
+//	b3Scalar currERP = (m_flags & B3_P2P_FLAGS_ERP) ? m_erp : info->erp;
+	b3Scalar currERP = info->erp;
+
+	b3Scalar k = info->fps * currERP;
+    int j;
+	for (j=0; j<3; j++)
+    {
+        info->m_constraintError[j*info->rowskip] = k * (a2[j] + trB.getOrigin()[j] - a1[j] - trA.getOrigin()[j]);
+		//printf("info->m_constraintError[%d]=%f\n",j,info->m_constraintError[j]);
+    }
+#if 0
+	if(m_flags & B3_P2P_FLAGS_CFM)
+	{
+		for (j=0; j<3; j++)
+		{
+			info->cfm[j*info->rowskip] = m_cfm;
+		}
+	}
+#endif
+
+#if 0
+	b3Scalar impulseClamp = m_setting.m_impulseClamp;//
+	for (j=0; j<3; j++)
+    {
+		if (m_setting.m_impulseClamp > 0)
+		{
+			info->m_lowerLimit[j*info->rowskip] = -impulseClamp;
+			info->m_upperLimit[j*info->rowskip] = impulseClamp;
+		}
+	}
+	info->m_damping = m_setting.m_damping;
+#endif
+
+}
+
+void b3GpuGenericConstraint::getInfo2 (b3GpuConstraintInfo2* info,  const b3RigidBodyData* bodies)
+{
+	switch (m_constraintType)
+	{
+	case B3_GPU_POINT2POINT_CONSTRAINT_TYPE:
+		{
+			getInfo2Point2Point(this,info,bodies);
+			break;
+		};
+		default:
+			{
+				b3Assert(0);
+			}
+	};
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuGenericConstraint.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuGenericConstraint.h
new file mode 100644
index 0000000000..14b3ba7fec
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuGenericConstraint.h
@@ -0,0 +1,132 @@
+/*
+Copyright (c) 2013 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+#ifndef B3_GPU_GENERIC_CONSTRAINT_H
+#define B3_GPU_GENERIC_CONSTRAINT_H
+
+#include "Bullet3Common/b3Quaternion.h"
+struct b3RigidBodyData;
+enum B3_CONSTRAINT_FLAGS
+{
+	B3_CONSTRAINT_FLAG_ENABLED=1,
+};
+
+enum b3GpuGenericConstraintType
+{
+	B3_GPU_POINT2POINT_CONSTRAINT_TYPE=3,
+	B3_GPU_FIXED_CONSTRAINT_TYPE=4,
+//	B3_HINGE_CONSTRAINT_TYPE,
+//	B3_CONETWIST_CONSTRAINT_TYPE,
+//	B3_D6_CONSTRAINT_TYPE,
+//	B3_SLIDER_CONSTRAINT_TYPE,
+//	B3_CONTACT_CONSTRAINT_TYPE,
+//	B3_D6_SPRING_CONSTRAINT_TYPE,
+//	B3_GEAR_CONSTRAINT_TYPE,
+	
+	B3_GPU_MAX_CONSTRAINT_TYPE
+};
+
+
+
+struct b3GpuConstraintInfo2 
+{
+	// integrator parameters: frames per second (1/stepsize), default error
+	// reduction parameter (0..1).
+	b3Scalar fps,erp;
+
+	// for the first and second body, pointers to two (linear and angular)
+	// n*3 jacobian sub matrices, stored by rows. these matrices will have
+	// been initialized to 0 on entry. if the second body is zero then the
+	// J2xx pointers may be 0.
+	b3Scalar *m_J1linearAxis,*m_J1angularAxis,*m_J2linearAxis,*m_J2angularAxis;
+
+	// elements to jump from one row to the next in J's
+	int rowskip;
+
+	// right hand sides of the equation J*v = c + cfm * lambda. cfm is the
+	// "constraint force mixing" vector. c is set to zero on entry, cfm is
+	// set to a constant value (typically very small or zero) value on entry.
+	b3Scalar *m_constraintError,*cfm;
+
+	// lo and hi limits for variables (set to -/+ infinity on entry).
+	b3Scalar *m_lowerLimit,*m_upperLimit;
+
+	// findex vector for variables. see the LCP solver interface for a
+	// description of what this does. this is set to -1 on entry.
+	// note that the returned indexes are relative to the first index of
+	// the constraint.
+	int *findex;
+	// number of solver iterations
+	int m_numIterations;
+
+	//damping of the velocity
+	b3Scalar	m_damping;
+};
+
+
+B3_ATTRIBUTE_ALIGNED16(struct) b3GpuGenericConstraint
+{
+	int				m_constraintType;
+	int				m_rbA;
+	int				m_rbB;
+	float			m_breakingImpulseThreshold;
+
+	b3Vector3 m_pivotInA;
+	b3Vector3 m_pivotInB;
+	b3Quaternion m_relTargetAB;
+
+	int	m_flags;
+	int m_uid;
+	int m_padding[2];
+
+	int	getRigidBodyA() const
+	{
+		return m_rbA;
+	}
+	int	getRigidBodyB() const
+	{
+		return m_rbB;
+	}
+
+	const b3Vector3& getPivotInA() const
+	{
+		return m_pivotInA;
+	}
+
+	const b3Vector3& getPivotInB() const
+	{
+		return m_pivotInB;
+	}
+
+	int isEnabled() const
+	{
+		return m_flags & B3_CONSTRAINT_FLAG_ENABLED;
+	}
+
+	float getBreakingImpulseThreshold() const
+	{
+		return m_breakingImpulseThreshold;
+	}
+
+	///internal method used by the constraint solver, don't use them directly
+	void getInfo1 (unsigned int* info,const b3RigidBodyData* bodies);
+
+	///internal method used by the constraint solver, don't use them directly
+	void getInfo2 (b3GpuConstraintInfo2* info,  const b3RigidBodyData* bodies);
+
+
+};
+
+#endif //B3_GPU_GENERIC_CONSTRAINT_H
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuJacobiContactSolver.cpp b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuJacobiContactSolver.cpp
new file mode 100644
index 0000000000..179dfc4f26
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuJacobiContactSolver.cpp
@@ -0,0 +1,1382 @@
+
+#include "b3GpuJacobiContactSolver.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3Contact4.h"
+#include "Bullet3Common/b3AlignedObjectArray.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3FillCL.h" //b3Int2
+class b3Vector3;
+#include "Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3PrefixScanCL.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h"
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "Bullet3OpenCL/RigidBody/kernels/solverUtils.h"
+#include "Bullet3Common/b3Logging.h"
+#include "b3GpuConstraint4.h"
+#include "Bullet3Common/shared/b3Int2.h"
+#include "Bullet3Common/shared/b3Int4.h"
+#define SOLVER_UTILS_KERNEL_PATH "src/Bullet3OpenCL/RigidBody/kernels/solverUtils.cl"
+
+
+struct b3GpuJacobiSolverInternalData
+{
+		//btRadixSort32CL*	m_sort32;
+		//btBoundSearchCL*	m_search;
+		b3PrefixScanCL*	m_scan;
+
+		b3OpenCLArray<unsigned int>* m_bodyCount;
+		b3OpenCLArray<b3Int2>*		m_contactConstraintOffsets;
+		b3OpenCLArray<unsigned int>* m_offsetSplitBodies;
+
+		b3OpenCLArray<b3Vector3>*	m_deltaLinearVelocities;
+		b3OpenCLArray<b3Vector3>*	m_deltaAngularVelocities;
+
+		b3AlignedObjectArray<b3Vector3>	m_deltaLinearVelocitiesCPU;
+		b3AlignedObjectArray<b3Vector3>	m_deltaAngularVelocitiesCPU;
+
+
+
+		b3OpenCLArray<b3GpuConstraint4>* m_contactConstraints;
+
+		b3FillCL*	m_filler;
+		
+
+		cl_kernel	m_countBodiesKernel;
+		cl_kernel	m_contactToConstraintSplitKernel;
+		cl_kernel	m_clearVelocitiesKernel;
+		cl_kernel	m_averageVelocitiesKernel;
+		cl_kernel	m_updateBodyVelocitiesKernel;
+		cl_kernel	m_solveContactKernel;
+		cl_kernel	m_solveFrictionKernel;
+
+
+
+};
+
+
+b3GpuJacobiContactSolver::b3GpuJacobiContactSolver(cl_context ctx, cl_device_id device, cl_command_queue queue, int pairCapacity)
+	:m_context(ctx),
+	m_device(device),
+	m_queue(queue)
+{
+	m_data = new b3GpuJacobiSolverInternalData;
+	m_data->m_scan = new b3PrefixScanCL(m_context,m_device,m_queue);
+	m_data->m_bodyCount = new b3OpenCLArray<unsigned int>(m_context,m_queue);
+	m_data->m_filler = new b3FillCL(m_context,m_device,m_queue);
+	m_data->m_contactConstraintOffsets = new b3OpenCLArray<b3Int2>(m_context,m_queue);
+	m_data->m_offsetSplitBodies = new b3OpenCLArray<unsigned int>(m_context,m_queue);
+	m_data->m_contactConstraints = new b3OpenCLArray<b3GpuConstraint4>(m_context,m_queue);
+	m_data->m_deltaLinearVelocities = new b3OpenCLArray<b3Vector3>(m_context,m_queue);
+	m_data->m_deltaAngularVelocities = new b3OpenCLArray<b3Vector3>(m_context,m_queue);
+
+	cl_int pErrNum;
+	const char* additionalMacros="";
+	const char* solverUtilsSource = solverUtilsCL;
+	{
+		cl_program solverUtilsProg= b3OpenCLUtils::compileCLProgramFromString( ctx, device, solverUtilsSource, &pErrNum,additionalMacros, SOLVER_UTILS_KERNEL_PATH);
+		b3Assert(solverUtilsProg);
+		m_data->m_countBodiesKernel =  b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverUtilsSource, "CountBodiesKernel", &pErrNum, solverUtilsProg,additionalMacros );
+		b3Assert(m_data->m_countBodiesKernel);
+
+		m_data->m_contactToConstraintSplitKernel  = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverUtilsSource, "ContactToConstraintSplitKernel", &pErrNum, solverUtilsProg,additionalMacros );
+		b3Assert(m_data->m_contactToConstraintSplitKernel);
+		m_data->m_clearVelocitiesKernel  = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverUtilsSource, "ClearVelocitiesKernel", &pErrNum, solverUtilsProg,additionalMacros );
+		b3Assert(m_data->m_clearVelocitiesKernel);
+
+		m_data->m_averageVelocitiesKernel  = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverUtilsSource, "AverageVelocitiesKernel", &pErrNum, solverUtilsProg,additionalMacros );
+		b3Assert(m_data->m_averageVelocitiesKernel);
+
+		m_data->m_updateBodyVelocitiesKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverUtilsSource, "UpdateBodyVelocitiesKernel", &pErrNum, solverUtilsProg,additionalMacros );
+		b3Assert(m_data->m_updateBodyVelocitiesKernel);
+
+		
+		m_data->m_solveContactKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverUtilsSource, "SolveContactJacobiKernel", &pErrNum, solverUtilsProg,additionalMacros );
+		b3Assert(m_data->m_solveContactKernel );
+
+		m_data->m_solveFrictionKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverUtilsSource, "SolveFrictionJacobiKernel", &pErrNum, solverUtilsProg,additionalMacros );
+		b3Assert(m_data->m_solveFrictionKernel);
+	}
+
+}
+
+
+b3GpuJacobiContactSolver::~b3GpuJacobiContactSolver()
+{
+	clReleaseKernel(m_data->m_solveContactKernel);
+	clReleaseKernel(m_data->m_solveFrictionKernel);
+	clReleaseKernel(m_data->m_countBodiesKernel);
+	clReleaseKernel(m_data->m_contactToConstraintSplitKernel);
+	clReleaseKernel(m_data->m_averageVelocitiesKernel);
+	clReleaseKernel(m_data->m_updateBodyVelocitiesKernel);
+	clReleaseKernel(m_data->m_clearVelocitiesKernel );
+
+	delete m_data->m_deltaLinearVelocities;
+	delete m_data->m_deltaAngularVelocities;
+	delete m_data->m_contactConstraints;
+	delete m_data->m_offsetSplitBodies;
+	delete m_data->m_contactConstraintOffsets;
+	delete m_data->m_bodyCount;
+	delete m_data->m_filler;
+	delete m_data->m_scan;
+	delete m_data;
+}
+
+
+
+b3Vector3 make_float4(float v)
+{
+	return b3MakeVector3 (v,v,v);
+}
+
+b3Vector4 make_float4(float x,float y, float z, float w)
+{
+	return b3MakeVector4 (x,y,z,w);
+}
+
+
+	static
+	inline
+	float calcRelVel(const b3Vector3& l0, const b3Vector3& l1, const b3Vector3& a0, const b3Vector3& a1, 
+					 const b3Vector3& linVel0, const b3Vector3& angVel0, const b3Vector3& linVel1, const b3Vector3& angVel1)
+	{
+		return b3Dot(l0, linVel0) + b3Dot(a0, angVel0) + b3Dot(l1, linVel1) + b3Dot(a1, angVel1);
+	}
+
+
+	static
+	inline
+	void setLinearAndAngular(const b3Vector3& n, const b3Vector3& r0, const b3Vector3& r1,
+							 b3Vector3& linear, b3Vector3& angular0, b3Vector3& angular1)
+	{
+		linear = n;
+		angular0 = b3Cross(r0, n);
+		angular1 = -b3Cross(r1, n);
+	}
+
+
+static __inline void solveContact(b3GpuConstraint4& cs, 
+	const b3Vector3& posA, const b3Vector3& linVelARO, const b3Vector3& angVelARO, float invMassA, const b3Matrix3x3& invInertiaA,
+	const b3Vector3& posB, const b3Vector3& linVelBRO, const b3Vector3& angVelBRO, float invMassB, const b3Matrix3x3& invInertiaB, 
+	float maxRambdaDt[4], float minRambdaDt[4], b3Vector3& dLinVelA, b3Vector3& dAngVelA, b3Vector3& dLinVelB, b3Vector3& dAngVelB)
+{
+
+
+	for(int ic=0; ic<4; ic++)
+	{
+		//	dont necessary because this makes change to 0
+		if( cs.m_jacCoeffInv[ic] == 0.f ) continue;
+
+		{
+			b3Vector3 angular0, angular1, linear;
+			b3Vector3 r0 = cs.m_worldPos[ic] - (b3Vector3&)posA;
+			b3Vector3 r1 = cs.m_worldPos[ic] - (b3Vector3&)posB;
+			setLinearAndAngular( (const b3Vector3 &)cs.m_linear, (const b3Vector3 &)r0, (const b3Vector3 &)r1, linear, angular0, angular1 );
+
+			float rambdaDt = calcRelVel((const b3Vector3 &)cs.m_linear,(const b3Vector3 &) -cs.m_linear, angular0, angular1,
+				linVelARO+dLinVelA, angVelARO+dAngVelA, linVelBRO+dLinVelB, angVelBRO+dAngVelB ) + cs.m_b[ic];
+			rambdaDt *= cs.m_jacCoeffInv[ic];
+
+			{
+				float prevSum = cs.m_appliedRambdaDt[ic];
+				float updated = prevSum;
+				updated += rambdaDt;
+				updated = b3Max( updated, minRambdaDt[ic] );
+				updated = b3Min( updated, maxRambdaDt[ic] );
+				rambdaDt = updated - prevSum;
+				cs.m_appliedRambdaDt[ic] = updated;
+			}
+
+			b3Vector3 linImp0 = invMassA*linear*rambdaDt;
+			b3Vector3 linImp1 = invMassB*(-linear)*rambdaDt;
+			b3Vector3 angImp0 = (invInertiaA* angular0)*rambdaDt;
+			b3Vector3 angImp1 = (invInertiaB* angular1)*rambdaDt;
+#ifdef _WIN32
+            b3Assert(_finite(linImp0.getX()));
+			b3Assert(_finite(linImp1.getX()));
+#endif
+			
+			if (invMassA)
+			{
+				dLinVelA += linImp0;
+				dAngVelA += angImp0;
+			}
+			if (invMassB)
+			{
+				dLinVelB += linImp1;
+				dAngVelB += angImp1;
+			}
+		}
+	}
+}
+
+
+
+void solveContact3(b3GpuConstraint4* cs,
+			b3Vector3* posAPtr, b3Vector3* linVelA, b3Vector3* angVelA, float invMassA, const b3Matrix3x3& invInertiaA,
+			b3Vector3* posBPtr, b3Vector3* linVelB, b3Vector3* angVelB, float invMassB, const b3Matrix3x3& invInertiaB,
+			b3Vector3* dLinVelA, b3Vector3* dAngVelA, b3Vector3* dLinVelB, b3Vector3* dAngVelB)
+{
+	float minRambdaDt = 0;
+	float maxRambdaDt = FLT_MAX;
+
+	for(int ic=0; ic<4; ic++)
+	{
+		if( cs->m_jacCoeffInv[ic] == 0.f ) continue;
+
+		b3Vector3 angular0, angular1, linear;
+		b3Vector3 r0 = cs->m_worldPos[ic] - *posAPtr;
+		b3Vector3 r1 = cs->m_worldPos[ic] - *posBPtr;
+		setLinearAndAngular( cs->m_linear, r0, r1, linear, angular0, angular1 );
+
+		float rambdaDt = calcRelVel( cs->m_linear, -cs->m_linear, angular0, angular1, 
+			*linVelA+*dLinVelA, *angVelA+*dAngVelA, *linVelB+*dLinVelB, *angVelB+*dAngVelB ) + cs->m_b[ic];
+		rambdaDt *= cs->m_jacCoeffInv[ic];
+
+		{
+			float prevSum = cs->m_appliedRambdaDt[ic];
+			float updated = prevSum;
+			updated += rambdaDt;
+			updated = b3Max( updated, minRambdaDt );
+			updated = b3Min( updated, maxRambdaDt );
+			rambdaDt = updated - prevSum;
+			cs->m_appliedRambdaDt[ic] = updated;
+		}
+
+		b3Vector3 linImp0 = invMassA*linear*rambdaDt;
+		b3Vector3 linImp1 = invMassB*(-linear)*rambdaDt;
+		b3Vector3 angImp0 = (invInertiaA* angular0)*rambdaDt;
+		b3Vector3 angImp1 = (invInertiaB* angular1)*rambdaDt;
+
+		if (invMassA)
+		{
+			*dLinVelA += linImp0;
+			*dAngVelA += angImp0;
+		}
+		if (invMassB)
+		{
+			*dLinVelB += linImp1;
+			*dAngVelB += angImp1;
+		}
+	}
+}
+
+
+static inline void solveFriction(b3GpuConstraint4& cs, 
+	const b3Vector3& posA, const b3Vector3& linVelARO, const b3Vector3& angVelARO, float invMassA, const b3Matrix3x3& invInertiaA,
+	const b3Vector3& posB, const b3Vector3& linVelBRO, const b3Vector3& angVelBRO, float invMassB, const b3Matrix3x3& invInertiaB, 
+	float maxRambdaDt[4], float minRambdaDt[4], b3Vector3& dLinVelA, b3Vector3& dAngVelA, b3Vector3& dLinVelB, b3Vector3& dAngVelB)
+{
+
+	b3Vector3 linVelA = linVelARO+dLinVelA;
+	b3Vector3 linVelB = linVelBRO+dLinVelB;
+	b3Vector3 angVelA = angVelARO+dAngVelA;
+	b3Vector3 angVelB = angVelBRO+dAngVelB;
+
+	if( cs.m_fJacCoeffInv[0] == 0 && cs.m_fJacCoeffInv[0] == 0 ) return;
+	const b3Vector3& center = (const b3Vector3&)cs.m_center;
+
+	b3Vector3 n = -(const b3Vector3&)cs.m_linear;
+
+	b3Vector3 tangent[2];
+#if 1		
+	b3PlaneSpace1 (n, tangent[0],tangent[1]);
+#else
+	b3Vector3 r = cs.m_worldPos[0]-center;
+	tangent[0] = cross3( n, r );
+	tangent[1] = cross3( tangent[0], n );
+	tangent[0] = normalize3( tangent[0] );
+	tangent[1] = normalize3( tangent[1] );
+#endif
+
+	b3Vector3 angular0, angular1, linear;
+	b3Vector3 r0 = center - posA;
+	b3Vector3 r1 = center - posB;
+	for(int i=0; i<2; i++)
+	{
+		setLinearAndAngular( tangent[i], r0, r1, linear, angular0, angular1 );
+		float rambdaDt = calcRelVel(linear, -linear, angular0, angular1,
+			linVelA, angVelA, linVelB, angVelB );
+		rambdaDt *= cs.m_fJacCoeffInv[i];
+
+			{
+				float prevSum = cs.m_fAppliedRambdaDt[i];
+				float updated = prevSum;
+				updated += rambdaDt;
+				updated = b3Max( updated, minRambdaDt[i] );
+				updated = b3Min( updated, maxRambdaDt[i] );
+				rambdaDt = updated - prevSum;
+				cs.m_fAppliedRambdaDt[i] = updated;
+			}
+
+		b3Vector3 linImp0 = invMassA*linear*rambdaDt;
+		b3Vector3 linImp1 = invMassB*(-linear)*rambdaDt;
+		b3Vector3 angImp0 = (invInertiaA* angular0)*rambdaDt;
+		b3Vector3 angImp1 = (invInertiaB* angular1)*rambdaDt;
+#ifdef _WIN32
+		b3Assert(_finite(linImp0.getX()));
+		b3Assert(_finite(linImp1.getX()));
+#endif
+		if (invMassA)
+		{
+			dLinVelA += linImp0;
+			dAngVelA += angImp0;
+		}
+		if (invMassB)
+		{
+			dLinVelB += linImp1;
+			dAngVelB += angImp1;
+		}
+	}
+
+	{	//	angular damping for point constraint
+		b3Vector3 ab = ( posB - posA ).normalized();
+		b3Vector3 ac = ( center - posA ).normalized();
+		if( b3Dot( ab, ac ) > 0.95f || (invMassA == 0.f || invMassB == 0.f))
+		{
+			float angNA = b3Dot( n, angVelA );
+			float angNB = b3Dot( n, angVelB );
+
+			if (invMassA)
+				dAngVelA -= (angNA*0.1f)*n;
+			if (invMassB)
+				dAngVelB -= (angNB*0.1f)*n;
+		}
+	}
+
+}
+
+
+
+
+float calcJacCoeff(const b3Vector3& linear0, const b3Vector3& linear1, const b3Vector3& angular0, const b3Vector3& angular1,
+					float invMass0, const b3Matrix3x3* invInertia0, float invMass1, const b3Matrix3x3* invInertia1, float countA, float countB)
+{
+	//	linear0,1 are normlized
+	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;
+	
+	float jmj1 = b3Dot(mtMul3(angular0,*invInertia0), angular0);
+	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;
+	float jmj3 = b3Dot(mtMul3(angular1,*invInertia1), angular1);
+	return -1.f/((jmj0+jmj1)*countA+(jmj2+jmj3)*countB);
+//	return -1.f/((jmj0+jmj1)+(jmj2+jmj3));
+
+}
+
+
+void setConstraint4( const b3Vector3& posA, const b3Vector3& linVelA, const b3Vector3& angVelA, float invMassA, const b3Matrix3x3& invInertiaA,
+	const b3Vector3& posB, const b3Vector3& linVelB, const b3Vector3& angVelB, float invMassB, const b3Matrix3x3& invInertiaB, 
+	 b3Contact4* src, float dt, float positionDrift, float positionConstraintCoeff, float countA, float countB,
+	b3GpuConstraint4* dstC )
+{
+	dstC->m_bodyA = abs(src->m_bodyAPtrAndSignBit);
+	dstC->m_bodyB = abs(src->m_bodyBPtrAndSignBit);
+
+	float dtInv = 1.f/dt;
+	for(int ic=0; ic<4; ic++)
+	{
+		dstC->m_appliedRambdaDt[ic] = 0.f;
+	}
+	dstC->m_fJacCoeffInv[0] = dstC->m_fJacCoeffInv[1] = 0.f;
+
+
+	dstC->m_linear = src->m_worldNormalOnB;
+	dstC->m_linear[3] = 0.7f ;//src->getFrictionCoeff() );
+	for(int ic=0; ic<4; ic++)
+	{
+		b3Vector3 r0 = src->m_worldPosB[ic] - posA;
+		b3Vector3 r1 = src->m_worldPosB[ic] - posB;
+
+		if( ic >= src->m_worldNormalOnB[3] )//npoints
+		{
+			dstC->m_jacCoeffInv[ic] = 0.f;
+			continue;
+		}
+
+		float relVelN;
+		{
+			b3Vector3 linear, angular0, angular1;
+			setLinearAndAngular(src->m_worldNormalOnB, r0, r1, linear, angular0, angular1);
+
+			dstC->m_jacCoeffInv[ic] = calcJacCoeff(linear, -linear, angular0, angular1,
+				invMassA, &invInertiaA, invMassB, &invInertiaB ,countA,countB);
+
+			relVelN = calcRelVel(linear, -linear, angular0, angular1,
+				linVelA, angVelA, linVelB, angVelB);
+
+			float e = 0.f;//src->getRestituitionCoeff();
+			if( relVelN*relVelN < 0.004f ) 
+			{
+				e = 0.f;
+			}
+
+			dstC->m_b[ic] = e*relVelN;
+			//float penetration = src->m_worldPos[ic].w;
+			dstC->m_b[ic] += (src->m_worldPosB[ic][3] + positionDrift)*positionConstraintCoeff*dtInv;
+			dstC->m_appliedRambdaDt[ic] = 0.f;
+		}
+	}
+
+	if( src->m_worldNormalOnB[3] > 0 )//npoints
+	{	//	prepare friction
+		b3Vector3 center = make_float4(0.f);
+		for(int i=0; i<src->m_worldNormalOnB[3]; i++) 
+			center += src->m_worldPosB[i];
+		center /= (float)src->m_worldNormalOnB[3];
+
+		b3Vector3 tangent[2];
+		b3PlaneSpace1(src->m_worldNormalOnB,tangent[0],tangent[1]);
+		
+		b3Vector3 r[2];
+		r[0] = center - posA;
+		r[1] = center - posB;
+
+		for(int i=0; i<2; i++)
+		{
+			b3Vector3 linear, angular0, angular1;
+			setLinearAndAngular(tangent[i], r[0], r[1], linear, angular0, angular1);
+
+			dstC->m_fJacCoeffInv[i] = calcJacCoeff(linear, -linear, angular0, angular1,
+				invMassA, &invInertiaA, invMassB, &invInertiaB ,countA,countB);
+			dstC->m_fAppliedRambdaDt[i] = 0.f;
+		}
+		dstC->m_center = center;
+	}
+
+	for(int i=0; i<4; i++)
+	{
+		if( i<src->m_worldNormalOnB[3] )
+		{
+			dstC->m_worldPos[i] = src->m_worldPosB[i];
+		}
+		else
+		{
+			dstC->m_worldPos[i] = make_float4(0.f);
+		}
+	}
+}
+
+
+
+void ContactToConstraintKernel(b3Contact4* gContact, b3RigidBodyData* gBodies, b3InertiaData* gShapes, b3GpuConstraint4* gConstraintOut, int nContacts,
+float dt,
+float positionDrift,
+float positionConstraintCoeff, int gIdx, b3AlignedObjectArray<unsigned int>& bodyCount
+)
+{
+	//int gIdx = 0;//GET_GLOBAL_IDX;
+	
+	if( gIdx < nContacts )
+	{
+		int aIdx = abs(gContact[gIdx].m_bodyAPtrAndSignBit);
+		int bIdx = abs(gContact[gIdx].m_bodyBPtrAndSignBit);
+
+		b3Vector3 posA = gBodies[aIdx].m_pos;
+		b3Vector3 linVelA = gBodies[aIdx].m_linVel;
+		b3Vector3 angVelA = gBodies[aIdx].m_angVel;
+		float invMassA = gBodies[aIdx].m_invMass;
+		b3Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertiaWorld;//.m_invInertia;
+
+		b3Vector3 posB = gBodies[bIdx].m_pos;
+		b3Vector3 linVelB = gBodies[bIdx].m_linVel;
+		b3Vector3 angVelB = gBodies[bIdx].m_angVel;
+		float invMassB = gBodies[bIdx].m_invMass;
+		b3Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertiaWorld;//m_invInertia;
+
+		b3GpuConstraint4 cs;
+		float countA = invMassA ? (float)(bodyCount[aIdx]) : 1;
+		float countB = invMassB ? (float)(bodyCount[bIdx]) : 1;
+    	setConstraint4( posA, linVelA, angVelA, invMassA, invInertiaA, posB, linVelB, angVelB, invMassB, invInertiaB,
+			&gContact[gIdx], dt, positionDrift, positionConstraintCoeff,countA,countB,
+			&cs );
+		
+
+		
+		cs.m_batchIdx = gContact[gIdx].m_batchIdx;
+
+		gConstraintOut[gIdx] = cs;
+	}
+}
+
+
+void b3GpuJacobiContactSolver::solveGroupHost(b3RigidBodyData* bodies,b3InertiaData* inertias,int numBodies,b3Contact4* manifoldPtr, int numManifolds,const b3JacobiSolverInfo& solverInfo)
+{
+	B3_PROFILE("b3GpuJacobiContactSolver::solveGroup");
+
+	b3AlignedObjectArray<unsigned int> bodyCount;
+	bodyCount.resize(numBodies);
+	for (int i=0;i<numBodies;i++)
+		bodyCount[i] = 0;
+
+	b3AlignedObjectArray<b3Int2> contactConstraintOffsets;
+	contactConstraintOffsets.resize(numManifolds);
+
+
+	for (int i=0;i<numManifolds;i++)
+	{
+		int pa = manifoldPtr[i].m_bodyAPtrAndSignBit;
+		int pb = manifoldPtr[i].m_bodyBPtrAndSignBit;
+
+		bool isFixedA = (pa <0) || (pa == solverInfo.m_fixedBodyIndex);
+		bool isFixedB = (pb <0) || (pb == solverInfo.m_fixedBodyIndex);
+
+		int bodyIndexA = manifoldPtr[i].getBodyA();
+		int bodyIndexB = manifoldPtr[i].getBodyB();
+
+		if (!isFixedA)
+		{
+			contactConstraintOffsets[i].x = bodyCount[bodyIndexA];
+			bodyCount[bodyIndexA]++;
+		}
+		if (!isFixedB)
+		{
+			contactConstraintOffsets[i].y = bodyCount[bodyIndexB];
+			bodyCount[bodyIndexB]++;
+		} 
+	}
+
+	b3AlignedObjectArray<unsigned int> offsetSplitBodies;
+	offsetSplitBodies.resize(numBodies);
+	unsigned int totalNumSplitBodies;
+	m_data->m_scan->executeHost(bodyCount,offsetSplitBodies,numBodies,&totalNumSplitBodies);
+	int numlastBody = bodyCount[numBodies-1];
+	totalNumSplitBodies += numlastBody;
+	printf("totalNumSplitBodies = %d\n",totalNumSplitBodies);
+
+	
+
+
+
+	b3AlignedObjectArray<b3GpuConstraint4> contactConstraints;
+	contactConstraints.resize(numManifolds);
+
+	for (int i=0;i<numManifolds;i++)
+	{
+		ContactToConstraintKernel(&manifoldPtr[0],bodies,inertias,&contactConstraints[0],numManifolds,
+			solverInfo.m_deltaTime,
+			solverInfo.m_positionDrift,
+			solverInfo.m_positionConstraintCoeff,
+			i, bodyCount);
+	}
+	int maxIter = solverInfo.m_numIterations;
+
+
+	b3AlignedObjectArray<b3Vector3> deltaLinearVelocities;
+	b3AlignedObjectArray<b3Vector3> deltaAngularVelocities;
+	deltaLinearVelocities.resize(totalNumSplitBodies);
+	deltaAngularVelocities.resize(totalNumSplitBodies);
+	for (unsigned int i=0;i<totalNumSplitBodies;i++)
+	{
+		deltaLinearVelocities[i].setZero();
+		deltaAngularVelocities[i].setZero();
+	}
+
+
+
+	for (int iter = 0;iter<maxIter;iter++)
+	{
+		int i=0;
+		for( i=0; i<numManifolds; i++)
+		{
+
+			//float frictionCoeff = contactConstraints[i].getFrictionCoeff();
+			int aIdx = (int)contactConstraints[i].m_bodyA;
+			int bIdx = (int)contactConstraints[i].m_bodyB;
+			b3RigidBodyData& bodyA = bodies[aIdx];
+			b3RigidBodyData& bodyB = bodies[bIdx];
+
+			b3Vector3 zero = b3MakeVector3(0,0,0);
+			
+			b3Vector3* dlvAPtr=&zero;
+			b3Vector3* davAPtr=&zero;
+			b3Vector3* dlvBPtr=&zero;
+			b3Vector3* davBPtr=&zero;
+			
+			if (bodyA.m_invMass)
+			{
+				int bodyOffsetA = offsetSplitBodies[aIdx];
+				int constraintOffsetA = contactConstraintOffsets[i].x;
+				int splitIndexA = bodyOffsetA+constraintOffsetA;
+				dlvAPtr = &deltaLinearVelocities[splitIndexA];
+				davAPtr = &deltaAngularVelocities[splitIndexA];
+			}
+
+			if (bodyB.m_invMass)
+			{
+				int bodyOffsetB = offsetSplitBodies[bIdx];
+				int constraintOffsetB = contactConstraintOffsets[i].y;
+				int splitIndexB= bodyOffsetB+constraintOffsetB;
+				dlvBPtr =&deltaLinearVelocities[splitIndexB];
+				davBPtr = &deltaAngularVelocities[splitIndexB];
+			}
+
+
+
+			{
+				float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};
+				float minRambdaDt[4] = {0.f,0.f,0.f,0.f};
+
+				solveContact( contactConstraints[i], (b3Vector3&)bodyA.m_pos, (b3Vector3&)bodyA.m_linVel, (b3Vector3&)bodyA.m_angVel, bodyA.m_invMass, inertias[aIdx].m_invInertiaWorld, 
+					(b3Vector3&)bodyB.m_pos, (b3Vector3&)bodyB.m_linVel, (b3Vector3&)bodyB.m_angVel, bodyB.m_invMass, inertias[bIdx].m_invInertiaWorld,
+					maxRambdaDt, minRambdaDt , *dlvAPtr,*davAPtr,*dlvBPtr,*davBPtr		);
+
+
+			}
+			
+		}
+
+		
+		//easy
+		for (int i=0;i<numBodies;i++)
+		{
+			if (bodies[i].m_invMass)
+			{
+				int bodyOffset = offsetSplitBodies[i];
+				int count = bodyCount[i];
+				float factor = 1.f/float(count);
+				b3Vector3 averageLinVel;
+				averageLinVel.setZero();
+				b3Vector3 averageAngVel;
+				averageAngVel.setZero();
+				for (int j=0;j<count;j++)
+				{
+					averageLinVel += deltaLinearVelocities[bodyOffset+j]*factor;
+					averageAngVel += deltaAngularVelocities[bodyOffset+j]*factor;
+				}
+				for (int j=0;j<count;j++)
+				{
+					deltaLinearVelocities[bodyOffset+j] = averageLinVel;
+					deltaAngularVelocities[bodyOffset+j] = averageAngVel;
+				}
+			}
+		}
+	}
+	for (int iter = 0;iter<maxIter;iter++)
+	{
+		//int i=0;
+	
+		//solve friction
+
+		for(int i=0; i<numManifolds; i++)
+		{
+			float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};
+			float minRambdaDt[4] = {0.f,0.f,0.f,0.f};
+
+			float sum = 0;
+			for(int j=0; j<4; j++)
+			{
+				sum +=contactConstraints[i].m_appliedRambdaDt[j];
+			}
+			float frictionCoeff = contactConstraints[i].getFrictionCoeff();
+			int aIdx = (int)contactConstraints[i].m_bodyA;
+			int bIdx = (int)contactConstraints[i].m_bodyB;
+			b3RigidBodyData& bodyA = bodies[aIdx];
+			b3RigidBodyData& bodyB = bodies[bIdx];
+
+			b3Vector3 zero = b3MakeVector3(0,0,0);
+			
+			b3Vector3* dlvAPtr=&zero;
+			b3Vector3* davAPtr=&zero;
+			b3Vector3* dlvBPtr=&zero;
+			b3Vector3* davBPtr=&zero;
+			
+			if (bodyA.m_invMass)
+			{
+				int bodyOffsetA = offsetSplitBodies[aIdx];
+				int constraintOffsetA = contactConstraintOffsets[i].x;
+				int splitIndexA = bodyOffsetA+constraintOffsetA;
+				dlvAPtr = &deltaLinearVelocities[splitIndexA];
+				davAPtr = &deltaAngularVelocities[splitIndexA];
+			}
+
+			if (bodyB.m_invMass)
+			{
+				int bodyOffsetB = offsetSplitBodies[bIdx];
+				int constraintOffsetB = contactConstraintOffsets[i].y;
+				int splitIndexB= bodyOffsetB+constraintOffsetB;
+				dlvBPtr =&deltaLinearVelocities[splitIndexB];
+				davBPtr = &deltaAngularVelocities[splitIndexB];
+			}
+
+			for(int j=0; j<4; j++)
+			{
+				maxRambdaDt[j] = frictionCoeff*sum;
+				minRambdaDt[j] = -maxRambdaDt[j];
+			}
+
+			solveFriction( contactConstraints[i], (b3Vector3&)bodyA.m_pos, (b3Vector3&)bodyA.m_linVel, (b3Vector3&)bodyA.m_angVel, bodyA.m_invMass,inertias[aIdx].m_invInertiaWorld, 
+				(b3Vector3&)bodyB.m_pos, (b3Vector3&)bodyB.m_linVel, (b3Vector3&)bodyB.m_angVel, bodyB.m_invMass, inertias[bIdx].m_invInertiaWorld,
+				maxRambdaDt, minRambdaDt , *dlvAPtr,*davAPtr,*dlvBPtr,*davBPtr);
+
+		}
+
+		//easy
+		for (int i=0;i<numBodies;i++)
+		{
+			if (bodies[i].m_invMass)
+			{
+				int bodyOffset = offsetSplitBodies[i];
+				int count = bodyCount[i];
+				float factor = 1.f/float(count);
+				b3Vector3 averageLinVel;
+				averageLinVel.setZero();
+				b3Vector3 averageAngVel;
+				averageAngVel.setZero();
+				for (int j=0;j<count;j++)
+				{
+					averageLinVel += deltaLinearVelocities[bodyOffset+j]*factor;
+					averageAngVel += deltaAngularVelocities[bodyOffset+j]*factor;
+				}
+				for (int j=0;j<count;j++)
+				{
+					deltaLinearVelocities[bodyOffset+j] = averageLinVel;
+					deltaAngularVelocities[bodyOffset+j] = averageAngVel;
+				}
+			}
+		}
+
+
+
+	}
+
+
+	//easy
+	for (int i=0;i<numBodies;i++)
+	{
+		if (bodies[i].m_invMass)
+		{
+			int bodyOffset = offsetSplitBodies[i];
+			int count = bodyCount[i];
+			if (count)
+			{
+				bodies[i].m_linVel += deltaLinearVelocities[bodyOffset];
+				bodies[i].m_angVel += deltaAngularVelocities[bodyOffset];
+			}
+		}
+	}
+}
+
+
+
+void b3GpuJacobiContactSolver::solveContacts(int numBodies, cl_mem bodyBuf, cl_mem inertiaBuf, int numContacts, cl_mem contactBuf, const struct b3Config& config, int static0Index)
+//
+//
+//void  b3GpuJacobiContactSolver::solveGroup(b3OpenCLArray<b3RigidBodyData>* bodies,b3OpenCLArray<b3InertiaData>* inertias,b3OpenCLArray<b3Contact4>* manifoldPtr,const btJacobiSolverInfo& solverInfo)
+{
+	b3JacobiSolverInfo solverInfo;
+	solverInfo.m_fixedBodyIndex = static0Index;
+	
+
+	B3_PROFILE("b3GpuJacobiContactSolver::solveGroup");
+
+	//int numBodies = bodies->size();
+	int numManifolds = numContacts;//manifoldPtr->size();
+
+	{
+		B3_PROFILE("resize");
+		m_data->m_bodyCount->resize(numBodies);
+	}
+	
+	unsigned int val=0;
+	b3Int2 val2;
+	val2.x=0;
+	val2.y=0;
+
+	 {
+		B3_PROFILE("m_filler");
+		m_data->m_contactConstraintOffsets->resize(numManifolds);
+		m_data->m_filler->execute(*m_data->m_bodyCount,val,numBodies);
+		
+	
+		m_data->m_filler->execute(*m_data->m_contactConstraintOffsets,val2,numManifolds);
+	}
+
+	{
+		B3_PROFILE("m_countBodiesKernel");
+		b3LauncherCL launcher(this->m_queue,m_data->m_countBodiesKernel,"m_countBodiesKernel");
+		launcher.setBuffer(contactBuf);//manifoldPtr->getBufferCL());
+		launcher.setBuffer(m_data->m_bodyCount->getBufferCL());
+		launcher.setBuffer(m_data->m_contactConstraintOffsets->getBufferCL());
+		launcher.setConst(numManifolds);
+		launcher.setConst(solverInfo.m_fixedBodyIndex);
+		launcher.launch1D(numManifolds);
+	}
+	unsigned int totalNumSplitBodies=0;
+	{
+		B3_PROFILE("m_scan->execute");
+		
+		m_data->m_offsetSplitBodies->resize(numBodies);
+		m_data->m_scan->execute(*m_data->m_bodyCount,*m_data->m_offsetSplitBodies,numBodies,&totalNumSplitBodies);
+		totalNumSplitBodies+=m_data->m_bodyCount->at(numBodies-1);
+	}
+
+	{
+		B3_PROFILE("m_data->m_contactConstraints->resize");
+		//int numContacts = manifoldPtr->size();
+		m_data->m_contactConstraints->resize(numContacts);
+	}
+	
+	{
+		B3_PROFILE("contactToConstraintSplitKernel");
+		b3LauncherCL launcher( m_queue, m_data->m_contactToConstraintSplitKernel,"m_contactToConstraintSplitKernel");
+		launcher.setBuffer(contactBuf);
+		launcher.setBuffer(bodyBuf);
+		launcher.setBuffer(inertiaBuf);
+		launcher.setBuffer(m_data->m_contactConstraints->getBufferCL());
+		launcher.setBuffer(m_data->m_bodyCount->getBufferCL());
+        launcher.setConst(numContacts);
+		launcher.setConst(solverInfo.m_deltaTime);
+		launcher.setConst(solverInfo.m_positionDrift);
+		launcher.setConst(solverInfo.m_positionConstraintCoeff);
+		launcher.launch1D( numContacts, 64 );
+		
+	}
+
+	
+	{
+		B3_PROFILE("m_data->m_deltaLinearVelocities->resize");
+		m_data->m_deltaLinearVelocities->resize(totalNumSplitBodies);
+		m_data->m_deltaAngularVelocities->resize(totalNumSplitBodies);
+	}
+
+
+	
+	{
+		B3_PROFILE("m_clearVelocitiesKernel");
+		b3LauncherCL launch(m_queue,m_data->m_clearVelocitiesKernel,"m_clearVelocitiesKernel");
+		launch.setBuffer(m_data->m_deltaAngularVelocities->getBufferCL());
+		launch.setBuffer(m_data->m_deltaLinearVelocities->getBufferCL());
+		launch.setConst(totalNumSplitBodies);
+		launch.launch1D(totalNumSplitBodies);
+		clFinish(m_queue);
+	}
+	
+	
+	int maxIter = solverInfo.m_numIterations;
+
+	for (int iter = 0;iter<maxIter;iter++)
+	{
+		{
+			B3_PROFILE("m_solveContactKernel");
+			b3LauncherCL launcher( m_queue, m_data->m_solveContactKernel,"m_solveContactKernel" );
+			launcher.setBuffer(m_data->m_contactConstraints->getBufferCL());
+			launcher.setBuffer(bodyBuf);
+			launcher.setBuffer(inertiaBuf);
+			launcher.setBuffer(m_data->m_contactConstraintOffsets->getBufferCL());
+			launcher.setBuffer(m_data->m_offsetSplitBodies->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaLinearVelocities->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaAngularVelocities->getBufferCL());
+			launcher.setConst(solverInfo.m_deltaTime);
+			launcher.setConst(solverInfo.m_positionDrift);
+			launcher.setConst(solverInfo.m_positionConstraintCoeff);
+			launcher.setConst(solverInfo.m_fixedBodyIndex);
+			launcher.setConst(numManifolds);
+
+			launcher.launch1D(numManifolds);
+			clFinish(m_queue);
+		}
+
+		
+
+		{
+			B3_PROFILE("average velocities");
+			b3LauncherCL launcher( m_queue, m_data->m_averageVelocitiesKernel,"m_averageVelocitiesKernel");
+			launcher.setBuffer(bodyBuf);
+			launcher.setBuffer(m_data->m_offsetSplitBodies->getBufferCL());
+			launcher.setBuffer(m_data->m_bodyCount->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaLinearVelocities->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaAngularVelocities->getBufferCL());
+			launcher.setConst(numBodies);
+			launcher.launch1D(numBodies);
+			clFinish(m_queue);
+		}
+
+		
+		{
+			B3_PROFILE("m_solveFrictionKernel");
+			b3LauncherCL launcher( m_queue, m_data->m_solveFrictionKernel,"m_solveFrictionKernel");
+			launcher.setBuffer(m_data->m_contactConstraints->getBufferCL());
+			launcher.setBuffer(bodyBuf);
+			launcher.setBuffer(inertiaBuf);
+			launcher.setBuffer(m_data->m_contactConstraintOffsets->getBufferCL());
+			launcher.setBuffer(m_data->m_offsetSplitBodies->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaLinearVelocities->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaAngularVelocities->getBufferCL());
+			launcher.setConst(solverInfo.m_deltaTime);
+			launcher.setConst(solverInfo.m_positionDrift);
+			launcher.setConst(solverInfo.m_positionConstraintCoeff);
+			launcher.setConst(solverInfo.m_fixedBodyIndex);
+			launcher.setConst(numManifolds);
+
+			launcher.launch1D(numManifolds);
+			clFinish(m_queue);
+		}
+
+		
+		{
+			B3_PROFILE("average velocities");
+			b3LauncherCL launcher( m_queue, m_data->m_averageVelocitiesKernel,"m_averageVelocitiesKernel");
+			launcher.setBuffer(bodyBuf);
+			launcher.setBuffer(m_data->m_offsetSplitBodies->getBufferCL());
+			launcher.setBuffer(m_data->m_bodyCount->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaLinearVelocities->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaAngularVelocities->getBufferCL());
+			launcher.setConst(numBodies);
+			launcher.launch1D(numBodies);
+			clFinish(m_queue);
+		}
+
+		
+
+	}
+
+	
+	{
+			B3_PROFILE("update body velocities");
+			b3LauncherCL launcher( m_queue, m_data->m_updateBodyVelocitiesKernel,"m_updateBodyVelocitiesKernel");
+			launcher.setBuffer(bodyBuf);
+			launcher.setBuffer(m_data->m_offsetSplitBodies->getBufferCL());
+			launcher.setBuffer(m_data->m_bodyCount->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaLinearVelocities->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaAngularVelocities->getBufferCL());
+			launcher.setConst(numBodies);
+			launcher.launch1D(numBodies);
+			clFinish(m_queue);
+		}
+
+
+
+}
+
+#if 0
+
+void  b3GpuJacobiContactSolver::solveGroupMixed(b3OpenCLArray<b3RigidBodyData>* bodiesGPU,b3OpenCLArray<b3InertiaData>* inertiasGPU,b3OpenCLArray<b3Contact4>* manifoldPtrGPU,const btJacobiSolverInfo& solverInfo)
+{
+
+	b3AlignedObjectArray<b3RigidBodyData> bodiesCPU;
+	bodiesGPU->copyToHost(bodiesCPU);
+	b3AlignedObjectArray<b3InertiaData> inertiasCPU;
+	inertiasGPU->copyToHost(inertiasCPU);
+	b3AlignedObjectArray<b3Contact4> manifoldPtrCPU;
+	manifoldPtrGPU->copyToHost(manifoldPtrCPU);
+	
+	int numBodiesCPU = bodiesGPU->size();
+	int numManifoldsCPU = manifoldPtrGPU->size();
+	B3_PROFILE("b3GpuJacobiContactSolver::solveGroupMixed");
+
+	b3AlignedObjectArray<unsigned int> bodyCount;
+	bodyCount.resize(numBodiesCPU);
+	for (int i=0;i<numBodiesCPU;i++)
+		bodyCount[i] = 0;
+
+	b3AlignedObjectArray<b3Int2> contactConstraintOffsets;
+	contactConstraintOffsets.resize(numManifoldsCPU);
+
+
+	for (int i=0;i<numManifoldsCPU;i++)
+	{
+		int pa = manifoldPtrCPU[i].m_bodyAPtrAndSignBit;
+		int pb = manifoldPtrCPU[i].m_bodyBPtrAndSignBit;
+
+		bool isFixedA = (pa <0) || (pa == solverInfo.m_fixedBodyIndex);
+		bool isFixedB = (pb <0) || (pb == solverInfo.m_fixedBodyIndex);
+
+		int bodyIndexA = manifoldPtrCPU[i].getBodyA();
+		int bodyIndexB = manifoldPtrCPU[i].getBodyB();
+
+		if (!isFixedA)
+		{
+			contactConstraintOffsets[i].x = bodyCount[bodyIndexA];
+			bodyCount[bodyIndexA]++;
+		}
+		if (!isFixedB)
+		{
+			contactConstraintOffsets[i].y = bodyCount[bodyIndexB];
+			bodyCount[bodyIndexB]++;
+		} 
+	}
+
+	b3AlignedObjectArray<unsigned int> offsetSplitBodies;
+	offsetSplitBodies.resize(numBodiesCPU);
+	unsigned int totalNumSplitBodiesCPU;
+	m_data->m_scan->executeHost(bodyCount,offsetSplitBodies,numBodiesCPU,&totalNumSplitBodiesCPU);
+	int numlastBody = bodyCount[numBodiesCPU-1];
+	totalNumSplitBodiesCPU += numlastBody;
+
+		int numBodies = bodiesGPU->size();
+	int numManifolds = manifoldPtrGPU->size();
+
+	m_data->m_bodyCount->resize(numBodies);
+	
+	unsigned int val=0;
+	b3Int2 val2;
+	val2.x=0;
+	val2.y=0;
+
+	 {
+		B3_PROFILE("m_filler");
+		m_data->m_contactConstraintOffsets->resize(numManifolds);
+		m_data->m_filler->execute(*m_data->m_bodyCount,val,numBodies);
+		
+	
+		m_data->m_filler->execute(*m_data->m_contactConstraintOffsets,val2,numManifolds);
+	}
+
+	{
+		B3_PROFILE("m_countBodiesKernel");
+		b3LauncherCL launcher(this->m_queue,m_data->m_countBodiesKernel);
+		launcher.setBuffer(manifoldPtrGPU->getBufferCL());
+		launcher.setBuffer(m_data->m_bodyCount->getBufferCL());
+		launcher.setBuffer(m_data->m_contactConstraintOffsets->getBufferCL());
+		launcher.setConst(numManifolds);
+		launcher.setConst(solverInfo.m_fixedBodyIndex);
+		launcher.launch1D(numManifolds);
+	}
+
+	unsigned int totalNumSplitBodies=0;
+	m_data->m_offsetSplitBodies->resize(numBodies);
+	m_data->m_scan->execute(*m_data->m_bodyCount,*m_data->m_offsetSplitBodies,numBodies,&totalNumSplitBodies);
+	totalNumSplitBodies+=m_data->m_bodyCount->at(numBodies-1);
+
+	if (totalNumSplitBodies != totalNumSplitBodiesCPU)
+	{
+		printf("error in totalNumSplitBodies!\n");
+	}
+
+	int numContacts = manifoldPtrGPU->size();
+	m_data->m_contactConstraints->resize(numContacts);
+
+	
+	{
+		B3_PROFILE("contactToConstraintSplitKernel");
+		b3LauncherCL launcher( m_queue, m_data->m_contactToConstraintSplitKernel);
+		launcher.setBuffer(manifoldPtrGPU->getBufferCL());
+		launcher.setBuffer(bodiesGPU->getBufferCL());
+		launcher.setBuffer(inertiasGPU->getBufferCL());
+		launcher.setBuffer(m_data->m_contactConstraints->getBufferCL());
+		launcher.setBuffer(m_data->m_bodyCount->getBufferCL());
+        launcher.setConst(numContacts);
+		launcher.setConst(solverInfo.m_deltaTime);
+		launcher.setConst(solverInfo.m_positionDrift);
+		launcher.setConst(solverInfo.m_positionConstraintCoeff);
+		launcher.launch1D( numContacts, 64 );
+		clFinish(m_queue);
+	}
+
+
+
+	b3AlignedObjectArray<b3GpuConstraint4> contactConstraints;
+	contactConstraints.resize(numManifoldsCPU);
+
+	for (int i=0;i<numManifoldsCPU;i++)
+	{
+		ContactToConstraintKernel(&manifoldPtrCPU[0],&bodiesCPU[0],&inertiasCPU[0],&contactConstraints[0],numManifoldsCPU,
+			solverInfo.m_deltaTime,
+			solverInfo.m_positionDrift,
+			solverInfo.m_positionConstraintCoeff,
+			i, bodyCount);
+	}
+	int maxIter = solverInfo.m_numIterations;
+
+
+	b3AlignedObjectArray<b3Vector3> deltaLinearVelocities;
+	b3AlignedObjectArray<b3Vector3> deltaAngularVelocities;
+	deltaLinearVelocities.resize(totalNumSplitBodiesCPU);
+	deltaAngularVelocities.resize(totalNumSplitBodiesCPU);
+	for (int i=0;i<totalNumSplitBodiesCPU;i++)
+	{
+		deltaLinearVelocities[i].setZero();
+		deltaAngularVelocities[i].setZero();
+	}
+
+	m_data->m_deltaLinearVelocities->resize(totalNumSplitBodies);
+	m_data->m_deltaAngularVelocities->resize(totalNumSplitBodies);
+
+
+	
+	{
+		B3_PROFILE("m_clearVelocitiesKernel");
+		b3LauncherCL launch(m_queue,m_data->m_clearVelocitiesKernel);
+		launch.setBuffer(m_data->m_deltaAngularVelocities->getBufferCL());
+		launch.setBuffer(m_data->m_deltaLinearVelocities->getBufferCL());
+		launch.setConst(totalNumSplitBodies);
+		launch.launch1D(totalNumSplitBodies);
+	}
+	
+
+		///!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+
+
+	m_data->m_contactConstraints->copyToHost(contactConstraints);
+	m_data->m_offsetSplitBodies->copyToHost(offsetSplitBodies);
+	m_data->m_contactConstraintOffsets->copyToHost(contactConstraintOffsets);
+	m_data->m_deltaLinearVelocities->copyToHost(deltaLinearVelocities);
+	m_data->m_deltaAngularVelocities->copyToHost(deltaAngularVelocities);
+
+	for (int iter = 0;iter<maxIter;iter++)
+	{
+
+				{
+			B3_PROFILE("m_solveContactKernel");
+			b3LauncherCL launcher( m_queue, m_data->m_solveContactKernel );
+			launcher.setBuffer(m_data->m_contactConstraints->getBufferCL());
+			launcher.setBuffer(bodiesGPU->getBufferCL());
+			launcher.setBuffer(inertiasGPU->getBufferCL());
+			launcher.setBuffer(m_data->m_contactConstraintOffsets->getBufferCL());
+			launcher.setBuffer(m_data->m_offsetSplitBodies->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaLinearVelocities->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaAngularVelocities->getBufferCL());
+			launcher.setConst(solverInfo.m_deltaTime);
+			launcher.setConst(solverInfo.m_positionDrift);
+			launcher.setConst(solverInfo.m_positionConstraintCoeff);
+			launcher.setConst(solverInfo.m_fixedBodyIndex);
+			launcher.setConst(numManifolds);
+
+			launcher.launch1D(numManifolds);
+			clFinish(m_queue);
+		}
+
+
+		int i=0;
+		for( i=0; i<numManifoldsCPU; i++)
+		{
+
+			float frictionCoeff = contactConstraints[i].getFrictionCoeff();
+			int aIdx = (int)contactConstraints[i].m_bodyA;
+			int bIdx = (int)contactConstraints[i].m_bodyB;
+			b3RigidBodyData& bodyA = bodiesCPU[aIdx];
+			b3RigidBodyData& bodyB = bodiesCPU[bIdx];
+
+			b3Vector3 zero(0,0,0);
+			
+			b3Vector3* dlvAPtr=&zero;
+			b3Vector3* davAPtr=&zero;
+			b3Vector3* dlvBPtr=&zero;
+			b3Vector3* davBPtr=&zero;
+			
+			if (bodyA.m_invMass)
+			{
+				int bodyOffsetA = offsetSplitBodies[aIdx];
+				int constraintOffsetA = contactConstraintOffsets[i].x;
+				int splitIndexA = bodyOffsetA+constraintOffsetA;
+				dlvAPtr = &deltaLinearVelocities[splitIndexA];
+				davAPtr = &deltaAngularVelocities[splitIndexA];
+			}
+
+			if (bodyB.m_invMass)
+			{
+				int bodyOffsetB = offsetSplitBodies[bIdx];
+				int constraintOffsetB = contactConstraintOffsets[i].y;
+				int splitIndexB= bodyOffsetB+constraintOffsetB;
+				dlvBPtr =&deltaLinearVelocities[splitIndexB];
+				davBPtr = &deltaAngularVelocities[splitIndexB];
+			}
+
+
+
+			{
+				float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};
+				float minRambdaDt[4] = {0.f,0.f,0.f,0.f};
+
+				solveContact( contactConstraints[i], (b3Vector3&)bodyA.m_pos, (b3Vector3&)bodyA.m_linVel, (b3Vector3&)bodyA.m_angVel, bodyA.m_invMass, inertiasCPU[aIdx].m_invInertiaWorld, 
+					(b3Vector3&)bodyB.m_pos, (b3Vector3&)bodyB.m_linVel, (b3Vector3&)bodyB.m_angVel, bodyB.m_invMass, inertiasCPU[bIdx].m_invInertiaWorld,
+					maxRambdaDt, minRambdaDt , *dlvAPtr,*davAPtr,*dlvBPtr,*davBPtr		);
+
+
+			}
+		}
+
+		
+		{
+			B3_PROFILE("average velocities");
+			b3LauncherCL launcher( m_queue, m_data->m_averageVelocitiesKernel);
+			launcher.setBuffer(bodiesGPU->getBufferCL());
+			launcher.setBuffer(m_data->m_offsetSplitBodies->getBufferCL());
+			launcher.setBuffer(m_data->m_bodyCount->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaLinearVelocities->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaAngularVelocities->getBufferCL());
+			launcher.setConst(numBodies);
+			launcher.launch1D(numBodies);
+			clFinish(m_queue);
+		}
+
+		//easy
+		for (int i=0;i<numBodiesCPU;i++)
+		{
+			if (bodiesCPU[i].m_invMass)
+			{
+				int bodyOffset = offsetSplitBodies[i];
+				int count = bodyCount[i];
+				float factor = 1.f/float(count);
+				b3Vector3 averageLinVel;
+				averageLinVel.setZero();
+				b3Vector3 averageAngVel;
+				averageAngVel.setZero();
+				for (int j=0;j<count;j++)
+				{
+					averageLinVel += deltaLinearVelocities[bodyOffset+j]*factor;
+					averageAngVel += deltaAngularVelocities[bodyOffset+j]*factor;
+				}
+				for (int j=0;j<count;j++)
+				{
+					deltaLinearVelocities[bodyOffset+j] = averageLinVel;
+					deltaAngularVelocities[bodyOffset+j] = averageAngVel;
+				}
+			}
+		}
+//	m_data->m_deltaAngularVelocities->copyFromHost(deltaAngularVelocities);
+	//m_data->m_deltaLinearVelocities->copyFromHost(deltaLinearVelocities);
+	m_data->m_deltaAngularVelocities->copyToHost(deltaAngularVelocities);
+	m_data->m_deltaLinearVelocities->copyToHost(deltaLinearVelocities);
+
+#if 0
+
+		{
+			B3_PROFILE("m_solveFrictionKernel");
+			b3LauncherCL launcher( m_queue, m_data->m_solveFrictionKernel);
+			launcher.setBuffer(m_data->m_contactConstraints->getBufferCL());
+			launcher.setBuffer(bodiesGPU->getBufferCL());
+			launcher.setBuffer(inertiasGPU->getBufferCL());
+			launcher.setBuffer(m_data->m_contactConstraintOffsets->getBufferCL());
+			launcher.setBuffer(m_data->m_offsetSplitBodies->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaLinearVelocities->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaAngularVelocities->getBufferCL());
+			launcher.setConst(solverInfo.m_deltaTime);
+			launcher.setConst(solverInfo.m_positionDrift);
+			launcher.setConst(solverInfo.m_positionConstraintCoeff);
+			launcher.setConst(solverInfo.m_fixedBodyIndex);
+			launcher.setConst(numManifolds);
+
+			launcher.launch1D(numManifolds);
+			clFinish(m_queue);
+		}
+
+		//solve friction
+
+		for(int i=0; i<numManifoldsCPU; i++)
+		{
+			float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};
+			float minRambdaDt[4] = {0.f,0.f,0.f,0.f};
+
+			float sum = 0;
+			for(int j=0; j<4; j++)
+			{
+				sum +=contactConstraints[i].m_appliedRambdaDt[j];
+			}
+			float frictionCoeff = contactConstraints[i].getFrictionCoeff();
+			int aIdx = (int)contactConstraints[i].m_bodyA;
+			int bIdx = (int)contactConstraints[i].m_bodyB;
+			b3RigidBodyData& bodyA = bodiesCPU[aIdx];
+			b3RigidBodyData& bodyB = bodiesCPU[bIdx];
+
+			b3Vector3 zero(0,0,0);
+			
+			b3Vector3* dlvAPtr=&zero;
+			b3Vector3* davAPtr=&zero;
+			b3Vector3* dlvBPtr=&zero;
+			b3Vector3* davBPtr=&zero;
+			
+			if (bodyA.m_invMass)
+			{
+				int bodyOffsetA = offsetSplitBodies[aIdx];
+				int constraintOffsetA = contactConstraintOffsets[i].x;
+				int splitIndexA = bodyOffsetA+constraintOffsetA;
+				dlvAPtr = &deltaLinearVelocities[splitIndexA];
+				davAPtr = &deltaAngularVelocities[splitIndexA];
+			}
+
+			if (bodyB.m_invMass)
+			{
+				int bodyOffsetB = offsetSplitBodies[bIdx];
+				int constraintOffsetB = contactConstraintOffsets[i].y;
+				int splitIndexB= bodyOffsetB+constraintOffsetB;
+				dlvBPtr =&deltaLinearVelocities[splitIndexB];
+				davBPtr = &deltaAngularVelocities[splitIndexB];
+			}
+
+			for(int j=0; j<4; j++)
+			{
+				maxRambdaDt[j] = frictionCoeff*sum;
+				minRambdaDt[j] = -maxRambdaDt[j];
+			}
+
+			solveFriction( contactConstraints[i], (b3Vector3&)bodyA.m_pos, (b3Vector3&)bodyA.m_linVel, (b3Vector3&)bodyA.m_angVel, bodyA.m_invMass,inertiasCPU[aIdx].m_invInertiaWorld, 
+				(b3Vector3&)bodyB.m_pos, (b3Vector3&)bodyB.m_linVel, (b3Vector3&)bodyB.m_angVel, bodyB.m_invMass, inertiasCPU[bIdx].m_invInertiaWorld,
+				maxRambdaDt, minRambdaDt , *dlvAPtr,*davAPtr,*dlvBPtr,*davBPtr);
+
+		}
+
+		{
+			B3_PROFILE("average velocities");
+			b3LauncherCL launcher( m_queue, m_data->m_averageVelocitiesKernel);
+			launcher.setBuffer(bodiesGPU->getBufferCL());
+			launcher.setBuffer(m_data->m_offsetSplitBodies->getBufferCL());
+			launcher.setBuffer(m_data->m_bodyCount->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaLinearVelocities->getBufferCL());
+			launcher.setBuffer(m_data->m_deltaAngularVelocities->getBufferCL());
+			launcher.setConst(numBodies);
+			launcher.launch1D(numBodies);
+			clFinish(m_queue);
+		}
+
+		//easy
+		for (int i=0;i<numBodiesCPU;i++)
+		{
+			if (bodiesCPU[i].m_invMass)
+			{
+				int bodyOffset = offsetSplitBodies[i];
+				int count = bodyCount[i];
+				float factor = 1.f/float(count);
+				b3Vector3 averageLinVel;
+				averageLinVel.setZero();
+				b3Vector3 averageAngVel;
+				averageAngVel.setZero();
+				for (int j=0;j<count;j++)
+				{
+					averageLinVel += deltaLinearVelocities[bodyOffset+j]*factor;
+					averageAngVel += deltaAngularVelocities[bodyOffset+j]*factor;
+				}
+				for (int j=0;j<count;j++)
+				{
+					deltaLinearVelocities[bodyOffset+j] = averageLinVel;
+					deltaAngularVelocities[bodyOffset+j] = averageAngVel;
+				}
+			}
+		}
+
+#endif
+
+	}
+
+	{
+		B3_PROFILE("update body velocities");
+		b3LauncherCL launcher( m_queue, m_data->m_updateBodyVelocitiesKernel);
+		launcher.setBuffer(bodiesGPU->getBufferCL());
+		launcher.setBuffer(m_data->m_offsetSplitBodies->getBufferCL());
+		launcher.setBuffer(m_data->m_bodyCount->getBufferCL());
+		launcher.setBuffer(m_data->m_deltaLinearVelocities->getBufferCL());
+		launcher.setBuffer(m_data->m_deltaAngularVelocities->getBufferCL());
+		launcher.setConst(numBodies);
+		launcher.launch1D(numBodies);
+		clFinish(m_queue);
+	}
+
+
+	//easy
+	for (int i=0;i<numBodiesCPU;i++)
+	{
+		if (bodiesCPU[i].m_invMass)
+		{
+			int bodyOffset = offsetSplitBodies[i];
+			int count = bodyCount[i];
+			if (count)
+			{
+				bodiesCPU[i].m_linVel += deltaLinearVelocities[bodyOffset];
+				bodiesCPU[i].m_angVel += deltaAngularVelocities[bodyOffset];
+			}
+		}
+	}
+
+
+//	bodiesGPU->copyFromHost(bodiesCPU);
+
+
+}
+#endif
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuJacobiContactSolver.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuJacobiContactSolver.h
new file mode 100644
index 0000000000..b418f29ec4
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuJacobiContactSolver.h
@@ -0,0 +1,62 @@
+
+#ifndef B3_GPU_JACOBI_CONTACT_SOLVER_H
+#define B3_GPU_JACOBI_CONTACT_SOLVER_H
+#include "Bullet3OpenCL/Initialize/b3OpenCLInclude.h"
+//#include "Bullet3Collision/NarrowPhaseCollision/shared/b3RigidBodyData.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3RigidBodyData.h"
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Contact4Data.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
+
+
+//struct b3InertiaData;
+//b3InertiaData
+
+class b3TypedConstraint;
+
+struct b3JacobiSolverInfo
+{
+	int m_fixedBodyIndex;
+
+	float m_deltaTime;
+	float m_positionDrift;
+	float m_positionConstraintCoeff;
+	int	m_numIterations;
+
+	b3JacobiSolverInfo()
+		:m_fixedBodyIndex(0),
+		m_deltaTime(1./60.f),
+		m_positionDrift( 0.005f ), 
+		m_positionConstraintCoeff( 0.99f ),
+		m_numIterations(7)
+	{
+	}
+};
+class b3GpuJacobiContactSolver
+{
+protected:
+
+	struct b3GpuJacobiSolverInternalData* m_data;
+
+	cl_context m_context;
+	cl_device_id m_device;
+	cl_command_queue m_queue;
+
+public:
+
+	b3GpuJacobiContactSolver(cl_context ctx, cl_device_id device, cl_command_queue queue, int pairCapacity);
+	virtual ~b3GpuJacobiContactSolver();
+
+
+	void solveContacts(int numBodies, cl_mem bodyBuf, cl_mem inertiaBuf, int numContacts, cl_mem contactBuf, const struct b3Config& config, int static0Index);
+	void solveGroupHost(b3RigidBodyData* bodies,b3InertiaData* inertias,int numBodies,struct b3Contact4* manifoldPtr, int numManifolds,const b3JacobiSolverInfo& solverInfo);
+	//void  solveGroupHost(btRigidBodyCL* bodies,b3InertiaData* inertias,int numBodies,btContact4* manifoldPtr, int numManifolds,btTypedConstraint** constraints,int numConstraints,const btJacobiSolverInfo& solverInfo);
+
+	//b3Scalar solveGroup(b3OpenCLArray<b3RigidBodyData>* gpuBodies,b3OpenCLArray<b3InertiaData>* gpuInertias, int numBodies,b3OpenCLArray<b3GpuGenericConstraint>* gpuConstraints,int numConstraints,const b3ContactSolverInfo& infoGlobal);
+
+	//void  solveGroup(btOpenCLArray<btRigidBodyCL>* bodies,btOpenCLArray<btInertiaCL>* inertias,btOpenCLArray<btContact4>* manifoldPtr,const btJacobiSolverInfo& solverInfo);
+	//void  solveGroupMixed(btOpenCLArray<btRigidBodyCL>* bodies,btOpenCLArray<btInertiaCL>* inertias,btOpenCLArray<btContact4>* manifoldPtr,const btJacobiSolverInfo& solverInfo);
+
+};
+#endif //B3_GPU_JACOBI_CONTACT_SOLVER_H
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuNarrowPhase.cpp b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuNarrowPhase.cpp
new file mode 100644
index 0000000000..698fa15f96
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuNarrowPhase.cpp
@@ -0,0 +1,1107 @@
+#include "b3GpuNarrowPhase.h"
+
+
+#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3ConvexPolyhedronData.h"
+#include "Bullet3OpenCL/NarrowphaseCollision/b3ConvexHullContact.h"
+#include "Bullet3OpenCL/BroadphaseCollision/b3SapAabb.h"
+#include <string.h>
+#include "Bullet3Collision/NarrowPhaseCollision/b3Config.h"
+#include "Bullet3OpenCL/NarrowphaseCollision/b3OptimizedBvh.h"
+#include "Bullet3OpenCL/NarrowphaseCollision/b3TriangleIndexVertexArray.h"
+#include "Bullet3Geometry/b3AabbUtil.h"
+#include "Bullet3OpenCL/NarrowphaseCollision/b3BvhInfo.h"
+
+#include "b3GpuNarrowPhaseInternalData.h"
+#include "Bullet3OpenCL/NarrowphaseCollision/b3QuantizedBvh.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3ConvexUtility.h"
+
+
+
+
+b3GpuNarrowPhase::b3GpuNarrowPhase(cl_context ctx, cl_device_id device, cl_command_queue queue, const b3Config& config)
+:m_data(0) ,m_planeBodyIndex(-1),m_static0Index(-1),
+m_context(ctx),
+m_device(device),
+m_queue(queue)
+{
+    
+	m_data = new b3GpuNarrowPhaseInternalData();
+	m_data->m_currentContactBuffer = 0;
+
+	memset(m_data,0,sizeof(b3GpuNarrowPhaseInternalData));
+    
+
+	m_data->m_config = config;
+	
+	m_data->m_gpuSatCollision = new GpuSatCollision(ctx,device,queue);
+	
+	
+	m_data->m_triangleConvexPairs = new b3OpenCLArray<b3Int4>(m_context,m_queue, config.m_maxTriConvexPairCapacity);
+
+
+	//m_data->m_convexPairsOutGPU = new b3OpenCLArray<b3Int2>(ctx,queue,config.m_maxBroadphasePairs,false);
+	//m_data->m_planePairs = new b3OpenCLArray<b3Int2>(ctx,queue,config.m_maxBroadphasePairs,false);
+    
+	m_data->m_pBufContactOutCPU = new b3AlignedObjectArray<b3Contact4>();
+	m_data->m_pBufContactOutCPU->resize(config.m_maxBroadphasePairs);
+	m_data->m_bodyBufferCPU = new b3AlignedObjectArray<b3RigidBodyData>();
+	m_data->m_bodyBufferCPU->resize(config.m_maxConvexBodies);
+    
+	m_data->m_inertiaBufferCPU = new b3AlignedObjectArray<b3InertiaData>();
+	m_data->m_inertiaBufferCPU->resize(config.m_maxConvexBodies);
+	
+	m_data->m_pBufContactBuffersGPU[0] = new b3OpenCLArray<b3Contact4>(ctx,queue, config.m_maxContactCapacity,true);
+	m_data->m_pBufContactBuffersGPU[1] = new b3OpenCLArray<b3Contact4>(ctx,queue, config.m_maxContactCapacity,true);
+	
+	m_data->m_inertiaBufferGPU = new b3OpenCLArray<b3InertiaData>(ctx,queue,config.m_maxConvexBodies,false);
+	m_data->m_collidablesGPU = new b3OpenCLArray<b3Collidable>(ctx,queue,config.m_maxConvexShapes);
+	m_data->m_collidablesCPU.reserve(config.m_maxConvexShapes);
+
+	m_data->m_localShapeAABBCPU = new b3AlignedObjectArray<b3SapAabb>;
+	m_data->m_localShapeAABBGPU = new b3OpenCLArray<b3SapAabb>(ctx,queue,config.m_maxConvexShapes);
+    
+    
+	//m_data->m_solverDataGPU = adl::Solver<adl::TYPE_CL>::allocate(ctx,queue, config.m_maxBroadphasePairs,false);
+	m_data->m_bodyBufferGPU = new b3OpenCLArray<b3RigidBodyData>(ctx,queue, config.m_maxConvexBodies,false);
+
+	m_data->m_convexFacesGPU = new b3OpenCLArray<b3GpuFace>(ctx,queue,config.m_maxConvexShapes*config.m_maxFacesPerShape,false);
+	m_data->m_convexFaces.reserve(config.m_maxConvexShapes*config.m_maxFacesPerShape);
+
+	m_data->m_gpuChildShapes = new b3OpenCLArray<b3GpuChildShape>(ctx,queue,config.m_maxCompoundChildShapes,false);
+	
+	m_data->m_convexPolyhedraGPU = new b3OpenCLArray<b3ConvexPolyhedronData>(ctx,queue,config.m_maxConvexShapes,false);
+	m_data->m_convexPolyhedra.reserve(config.m_maxConvexShapes);
+
+	m_data->m_uniqueEdgesGPU = new b3OpenCLArray<b3Vector3>(ctx,queue,config.m_maxConvexUniqueEdges,true);
+	m_data->m_uniqueEdges.reserve(config.m_maxConvexUniqueEdges);
+	
+	
+
+	m_data->m_convexVerticesGPU = new b3OpenCLArray<b3Vector3>(ctx,queue,config.m_maxConvexVertices,true);
+	m_data->m_convexVertices.reserve(config.m_maxConvexVertices);
+
+	m_data->m_convexIndicesGPU = new b3OpenCLArray<int>(ctx,queue,config.m_maxConvexIndices,true);
+    m_data->m_convexIndices.reserve(config.m_maxConvexIndices);
+    
+	m_data->m_worldVertsB1GPU = new b3OpenCLArray<b3Vector3>(ctx,queue,config.m_maxConvexBodies*config.m_maxVerticesPerFace);
+    m_data->m_clippingFacesOutGPU = new  b3OpenCLArray<b3Int4>(ctx,queue,config.m_maxConvexBodies);
+    m_data->m_worldNormalsAGPU = new  b3OpenCLArray<b3Vector3>(ctx,queue,config.m_maxConvexBodies);
+	m_data->m_worldVertsA1GPU = new b3OpenCLArray<b3Vector3>(ctx,queue,config.m_maxConvexBodies*config.m_maxVerticesPerFace);
+    m_data->m_worldVertsB2GPU = new  b3OpenCLArray<b3Vector3>(ctx,queue,config.m_maxConvexBodies*config.m_maxVerticesPerFace);
+    
+    
+
+	m_data->m_convexData = new b3AlignedObjectArray<b3ConvexUtility* >();
+
+	m_data->m_convexData->resize(config.m_maxConvexShapes);
+	m_data->m_convexPolyhedra.resize(config.m_maxConvexShapes);
+    
+	m_data->m_numAcceleratedShapes = 0;
+	m_data->m_numAcceleratedRigidBodies = 0;
+    
+		
+	m_data->m_subTreesGPU = new b3OpenCLArray<b3BvhSubtreeInfo>(this->m_context,this->m_queue);
+	m_data->m_treeNodesGPU = new b3OpenCLArray<b3QuantizedBvhNode>(this->m_context,this->m_queue);
+	m_data->m_bvhInfoGPU = new b3OpenCLArray<b3BvhInfo>(this->m_context,this->m_queue);
+
+	//m_data->m_contactCGPU = new b3OpenCLArray<Constraint4>(ctx,queue,config.m_maxBroadphasePairs,false);
+	//m_data->m_frictionCGPU = new b3OpenCLArray<adl::Solver<adl::TYPE_CL>::allocateFrictionConstraint( m_data->m_deviceCL, config.m_maxBroadphasePairs);
+ 
+	
+
+}
+
+
+b3GpuNarrowPhase::~b3GpuNarrowPhase()
+{
+	delete m_data->m_gpuSatCollision;
+	
+	delete m_data->m_triangleConvexPairs;
+	//delete m_data->m_convexPairsOutGPU;
+	//delete m_data->m_planePairs;
+	delete m_data->m_pBufContactOutCPU;
+	delete m_data->m_bodyBufferCPU;
+	delete m_data->m_inertiaBufferCPU;
+	delete m_data->m_pBufContactBuffersGPU[0];
+	delete m_data->m_pBufContactBuffersGPU[1];
+
+
+	delete m_data->m_inertiaBufferGPU;
+	delete m_data->m_collidablesGPU;
+	delete m_data->m_localShapeAABBCPU;
+	delete m_data->m_localShapeAABBGPU;
+	delete m_data->m_bodyBufferGPU;
+	delete m_data->m_convexFacesGPU;
+	delete m_data->m_gpuChildShapes;
+	delete m_data->m_convexPolyhedraGPU;
+	delete m_data->m_uniqueEdgesGPU;
+	delete m_data->m_convexVerticesGPU;
+	delete m_data->m_convexIndicesGPU;
+	delete m_data->m_worldVertsB1GPU;
+    delete m_data->m_clippingFacesOutGPU;
+    delete m_data->m_worldNormalsAGPU;
+	delete m_data->m_worldVertsA1GPU;
+    delete m_data->m_worldVertsB2GPU;
+    
+	delete m_data->m_bvhInfoGPU;
+
+	for (int i=0;i<m_data->m_bvhData.size();i++)
+	{
+		delete m_data->m_bvhData[i];
+	}
+	for (int i=0;i<m_data->m_meshInterfaces.size();i++)
+	{
+		delete m_data->m_meshInterfaces[i];
+	}
+	m_data->m_meshInterfaces.clear();
+	m_data->m_bvhData.clear();
+	delete m_data->m_treeNodesGPU;
+	delete m_data->m_subTreesGPU;
+
+    
+    delete m_data->m_convexData;
+	delete m_data;
+}
+
+
+int	b3GpuNarrowPhase::allocateCollidable()
+{
+	int curSize = m_data->m_collidablesCPU.size();
+	if (curSize<m_data->m_config.m_maxConvexShapes)
+	{
+		m_data->m_collidablesCPU.expand();
+		return curSize;
+	}
+	else
+	{
+		b3Error("allocateCollidable out-of-range %d\n",m_data->m_config.m_maxConvexShapes);
+	}
+	return -1;
+
+}
+
+
+
+
+
+int		b3GpuNarrowPhase::registerSphereShape(float radius)
+{
+	int collidableIndex = allocateCollidable();
+	if (collidableIndex<0)
+		return collidableIndex;
+
+
+	b3Collidable& col = getCollidableCpu(collidableIndex);
+	col.m_shapeType = SHAPE_SPHERE;
+	col.m_shapeIndex = 0;
+	col.m_radius = radius;
+	
+	if (col.m_shapeIndex>=0)
+	{
+		b3SapAabb aabb;
+		b3Vector3 myAabbMin=b3MakeVector3(-radius,-radius,-radius);
+		b3Vector3 myAabbMax=b3MakeVector3(radius,radius,radius);
+
+		aabb.m_min[0] = myAabbMin[0];//s_convexHeightField->m_aabb.m_min.x;
+		aabb.m_min[1] = myAabbMin[1];//s_convexHeightField->m_aabb.m_min.y;
+		aabb.m_min[2] = myAabbMin[2];//s_convexHeightField->m_aabb.m_min.z;
+		aabb.m_minIndices[3] = 0;
+
+		aabb.m_max[0] = myAabbMax[0];//s_convexHeightField->m_aabb.m_max.x;
+		aabb.m_max[1] = myAabbMax[1];//s_convexHeightField->m_aabb.m_max.y;
+		aabb.m_max[2] = myAabbMax[2];//s_convexHeightField->m_aabb.m_max.z;
+		aabb.m_signedMaxIndices[3] = 0;
+
+		m_data->m_localShapeAABBCPU->push_back(aabb);
+//		m_data->m_localShapeAABBGPU->push_back(aabb);
+		clFinish(m_queue);
+	}
+	
+	return collidableIndex;
+}
+
+
+int b3GpuNarrowPhase::registerFace(const b3Vector3& faceNormal, float faceConstant)
+{
+	int faceOffset = m_data->m_convexFaces.size();
+	b3GpuFace& face = m_data->m_convexFaces.expand();
+	face.m_plane = b3MakeVector3(faceNormal.x,faceNormal.y,faceNormal.z,faceConstant);
+	return faceOffset;
+}
+
+int		b3GpuNarrowPhase::registerPlaneShape(const b3Vector3& planeNormal, float planeConstant)
+{
+	int collidableIndex = allocateCollidable();
+	if (collidableIndex<0)
+		return collidableIndex;
+
+
+	b3Collidable& col = getCollidableCpu(collidableIndex);
+	col.m_shapeType = SHAPE_PLANE;
+	col.m_shapeIndex = registerFace(planeNormal,planeConstant);
+	col.m_radius = planeConstant;
+	
+	if (col.m_shapeIndex>=0)
+	{
+		b3SapAabb aabb;
+		aabb.m_min[0] = -1e30f;
+		aabb.m_min[1] = -1e30f;
+		aabb.m_min[2] = -1e30f;
+		aabb.m_minIndices[3] = 0;
+		
+		aabb.m_max[0] = 1e30f;
+		aabb.m_max[1] = 1e30f;
+		aabb.m_max[2] = 1e30f;
+		aabb.m_signedMaxIndices[3] = 0;
+
+		m_data->m_localShapeAABBCPU->push_back(aabb);
+//		m_data->m_localShapeAABBGPU->push_back(aabb);
+		clFinish(m_queue);
+	}
+	
+	return collidableIndex;
+}
+
+
+int b3GpuNarrowPhase::registerConvexHullShapeInternal(b3ConvexUtility* convexPtr,b3Collidable& col)
+{
+
+	m_data->m_convexData->resize(m_data->m_numAcceleratedShapes+1);
+	m_data->m_convexPolyhedra.resize(m_data->m_numAcceleratedShapes+1);
+	
+    
+	b3ConvexPolyhedronData& convex = m_data->m_convexPolyhedra.at(m_data->m_convexPolyhedra.size()-1);
+	convex.mC = convexPtr->mC;
+	convex.mE = convexPtr->mE;
+	convex.m_extents= convexPtr->m_extents;
+	convex.m_localCenter = convexPtr->m_localCenter;
+	convex.m_radius = convexPtr->m_radius;
+	
+	convex.m_numUniqueEdges = convexPtr->m_uniqueEdges.size();
+	int edgeOffset = m_data->m_uniqueEdges.size();
+	convex.m_uniqueEdgesOffset = edgeOffset;
+	
+	m_data->m_uniqueEdges.resize(edgeOffset+convex.m_numUniqueEdges);
+    
+	//convex data here
+	int i;
+	for ( i=0;i<convexPtr->m_uniqueEdges.size();i++)
+	{
+		m_data->m_uniqueEdges[edgeOffset+i] = convexPtr->m_uniqueEdges[i];
+	}
+    
+	int faceOffset = m_data->m_convexFaces.size();
+	convex.m_faceOffset = faceOffset;
+	convex.m_numFaces = convexPtr->m_faces.size();
+
+	m_data->m_convexFaces.resize(faceOffset+convex.m_numFaces);
+	
+
+	for (i=0;i<convexPtr->m_faces.size();i++)
+	{
+		m_data->m_convexFaces[convex.m_faceOffset+i].m_plane = b3MakeVector3(convexPtr->m_faces[i].m_plane[0],
+																			convexPtr->m_faces[i].m_plane[1],
+																			convexPtr->m_faces[i].m_plane[2],
+																			convexPtr->m_faces[i].m_plane[3]);
+
+		
+		int indexOffset = m_data->m_convexIndices.size();
+		int numIndices = convexPtr->m_faces[i].m_indices.size();
+		m_data->m_convexFaces[convex.m_faceOffset+i].m_numIndices = numIndices;
+		m_data->m_convexFaces[convex.m_faceOffset+i].m_indexOffset = indexOffset;
+		m_data->m_convexIndices.resize(indexOffset+numIndices);
+		for (int p=0;p<numIndices;p++)
+		{
+			m_data->m_convexIndices[indexOffset+p] = convexPtr->m_faces[i].m_indices[p];
+		}
+	}
+    
+	convex.m_numVertices = convexPtr->m_vertices.size();
+	int vertexOffset = m_data->m_convexVertices.size();
+	convex.m_vertexOffset =vertexOffset;
+	
+	m_data->m_convexVertices.resize(vertexOffset+convex.m_numVertices);
+	for (int i=0;i<convexPtr->m_vertices.size();i++)
+	{
+		m_data->m_convexVertices[vertexOffset+i] = convexPtr->m_vertices[i];
+	}
+
+	(*m_data->m_convexData)[m_data->m_numAcceleratedShapes] = convexPtr;
+	
+    
+    
+	return m_data->m_numAcceleratedShapes++;
+}
+
+
+int		b3GpuNarrowPhase::registerConvexHullShape(const float* vertices, int strideInBytes, int numVertices, const float* scaling)
+{
+	b3AlignedObjectArray<b3Vector3> verts;
+
+	unsigned char* vts = (unsigned char*) vertices;
+	for (int i=0;i<numVertices;i++)
+	{
+		float* vertex = (float*) &vts[i*strideInBytes];
+		verts.push_back(b3MakeVector3(vertex[0]*scaling[0],vertex[1]*scaling[1],vertex[2]*scaling[2]));
+	}
+
+	b3ConvexUtility* utilPtr = new b3ConvexUtility();
+	bool merge = true;
+	if (numVertices)
+	{
+		utilPtr->initializePolyhedralFeatures(&verts[0],verts.size(),merge);
+	}
+
+	int collidableIndex = registerConvexHullShape(utilPtr);
+	delete utilPtr;
+	return collidableIndex;
+}
+
+int		b3GpuNarrowPhase::registerConvexHullShape(b3ConvexUtility* utilPtr)
+{
+	int collidableIndex = allocateCollidable();
+	if (collidableIndex<0)
+		return collidableIndex;
+
+	b3Collidable& col = getCollidableCpu(collidableIndex);
+	col.m_shapeType = SHAPE_CONVEX_HULL;
+	col.m_shapeIndex = -1;
+	
+	
+	{
+		b3Vector3 localCenter=b3MakeVector3(0,0,0);
+		for (int i=0;i<utilPtr->m_vertices.size();i++)
+			localCenter+=utilPtr->m_vertices[i];
+		localCenter*= (1.f/utilPtr->m_vertices.size());
+		utilPtr->m_localCenter = localCenter;
+
+		col.m_shapeIndex = registerConvexHullShapeInternal(utilPtr,col);
+	}
+
+	if (col.m_shapeIndex>=0)
+	{
+		b3SapAabb aabb;
+		
+		b3Vector3 myAabbMin=b3MakeVector3(1e30f,1e30f,1e30f);
+		b3Vector3 myAabbMax=b3MakeVector3(-1e30f,-1e30f,-1e30f);
+
+		for (int i=0;i<utilPtr->m_vertices.size();i++)
+		{
+			myAabbMin.setMin(utilPtr->m_vertices[i]);
+			myAabbMax.setMax(utilPtr->m_vertices[i]);
+		}
+		aabb.m_min[0] = myAabbMin[0];
+		aabb.m_min[1] = myAabbMin[1];
+		aabb.m_min[2] = myAabbMin[2];
+		aabb.m_minIndices[3] = 0;
+
+		aabb.m_max[0] = myAabbMax[0];
+		aabb.m_max[1] = myAabbMax[1];
+		aabb.m_max[2] = myAabbMax[2];
+		aabb.m_signedMaxIndices[3] = 0;
+
+		m_data->m_localShapeAABBCPU->push_back(aabb);
+//		m_data->m_localShapeAABBGPU->push_back(aabb);
+	}
+	
+	return collidableIndex;
+
+}
+
+int		b3GpuNarrowPhase::registerCompoundShape(b3AlignedObjectArray<b3GpuChildShape>* childShapes)
+{
+	
+	int collidableIndex = allocateCollidable();
+	if (collidableIndex<0)
+		return collidableIndex;
+
+	b3Collidable& col = getCollidableCpu(collidableIndex);
+	col.m_shapeType = SHAPE_COMPOUND_OF_CONVEX_HULLS;
+	col.m_shapeIndex = m_data->m_cpuChildShapes.size();
+	col.m_compoundBvhIndex = m_data->m_bvhInfoCPU.size();
+
+	{
+		b3Assert(col.m_shapeIndex+childShapes->size()<m_data->m_config.m_maxCompoundChildShapes);
+		for (int i=0;i<childShapes->size();i++)
+		{
+			m_data->m_cpuChildShapes.push_back(childShapes->at(i));
+		}
+	}
+
+
+
+	col.m_numChildShapes = childShapes->size();
+	
+	
+	b3SapAabb aabbLocalSpace;
+	b3Vector3 myAabbMin=b3MakeVector3(1e30f,1e30f,1e30f);
+	b3Vector3 myAabbMax=b3MakeVector3(-1e30f,-1e30f,-1e30f);
+	
+	b3AlignedObjectArray<b3Aabb> childLocalAabbs;
+	childLocalAabbs.resize(childShapes->size());
+
+	//compute local AABB of the compound of all children
+	for (int i=0;i<childShapes->size();i++)
+	{
+		int childColIndex = childShapes->at(i).m_shapeIndex;
+		//b3Collidable& childCol = getCollidableCpu(childColIndex);
+		b3SapAabb aabbLoc =m_data->m_localShapeAABBCPU->at(childColIndex);
+
+		b3Vector3 childLocalAabbMin=b3MakeVector3(aabbLoc.m_min[0],aabbLoc.m_min[1],aabbLoc.m_min[2]);
+		b3Vector3 childLocalAabbMax=b3MakeVector3(aabbLoc.m_max[0],aabbLoc.m_max[1],aabbLoc.m_max[2]);
+		b3Vector3 aMin,aMax;
+		b3Scalar margin(0.f);
+		b3Transform childTr;
+		childTr.setIdentity();
+
+		childTr.setOrigin(childShapes->at(i).m_childPosition);
+		childTr.setRotation(b3Quaternion(childShapes->at(i).m_childOrientation));
+		b3TransformAabb(childLocalAabbMin,childLocalAabbMax,margin,childTr,aMin,aMax);
+		myAabbMin.setMin(aMin);
+		myAabbMax.setMax(aMax);		
+		childLocalAabbs[i].m_min[0] = aMin[0];
+		childLocalAabbs[i].m_min[1] = aMin[1];
+		childLocalAabbs[i].m_min[2] = aMin[2];
+		childLocalAabbs[i].m_min[3] = 0;
+		childLocalAabbs[i].m_max[0] = aMax[0];
+		childLocalAabbs[i].m_max[1] = aMax[1];
+		childLocalAabbs[i].m_max[2] = aMax[2];
+		childLocalAabbs[i].m_max[3] = 0;
+	}
+	
+	aabbLocalSpace.m_min[0] = myAabbMin[0];//s_convexHeightField->m_aabb.m_min.x;
+	aabbLocalSpace.m_min[1]= myAabbMin[1];//s_convexHeightField->m_aabb.m_min.y;
+	aabbLocalSpace.m_min[2]= myAabbMin[2];//s_convexHeightField->m_aabb.m_min.z;
+	aabbLocalSpace.m_minIndices[3] = 0;
+	
+	aabbLocalSpace.m_max[0] = myAabbMax[0];//s_convexHeightField->m_aabb.m_max.x;
+	aabbLocalSpace.m_max[1]= myAabbMax[1];//s_convexHeightField->m_aabb.m_max.y;
+	aabbLocalSpace.m_max[2]= myAabbMax[2];//s_convexHeightField->m_aabb.m_max.z;
+	aabbLocalSpace.m_signedMaxIndices[3] = 0;
+	
+	m_data->m_localShapeAABBCPU->push_back(aabbLocalSpace);
+
+
+	b3QuantizedBvh* bvh = new b3QuantizedBvh;
+	bvh->setQuantizationValues(myAabbMin,myAabbMax);
+	QuantizedNodeArray&	nodes = bvh->getLeafNodeArray();
+	int numNodes = childShapes->size();
+
+	for (int i=0;i<numNodes;i++)
+	{
+		b3QuantizedBvhNode node;
+		b3Vector3 aabbMin,aabbMax;
+		aabbMin = (b3Vector3&) childLocalAabbs[i].m_min;
+		aabbMax = (b3Vector3&) childLocalAabbs[i].m_max;
+
+		bvh->quantize(&node.m_quantizedAabbMin[0],aabbMin,0);
+		bvh->quantize(&node.m_quantizedAabbMax[0],aabbMax,1);
+		int partId = 0;
+		node.m_escapeIndexOrTriangleIndex = (partId<<(31-MAX_NUM_PARTS_IN_BITS)) | i;
+		nodes.push_back(node);
+	}
+	bvh->buildInternal();
+
+	int numSubTrees = bvh->getSubtreeInfoArray().size();
+
+	//void	setQuantizationValues(const b3Vector3& bvhAabbMin,const b3Vector3& bvhAabbMax,b3Scalar quantizationMargin=b3Scalar(1.0));
+	//QuantizedNodeArray&	getLeafNodeArray() {			return	m_quantizedLeafNodes;	}
+	///buildInternal is expert use only: assumes that setQuantizationValues and LeafNodeArray are initialized
+	//void	buildInternal();
+
+	b3BvhInfo bvhInfo;
+	
+	bvhInfo.m_aabbMin = bvh->m_bvhAabbMin;
+	bvhInfo.m_aabbMax = bvh->m_bvhAabbMax;
+	bvhInfo.m_quantization = bvh->m_bvhQuantization;
+	bvhInfo.m_numNodes = numNodes;
+	bvhInfo.m_numSubTrees = numSubTrees;
+	bvhInfo.m_nodeOffset = m_data->m_treeNodesCPU.size();
+	bvhInfo.m_subTreeOffset = m_data->m_subTreesCPU.size();
+
+	int numNewNodes = 		bvh->getQuantizedNodeArray().size();
+
+	for (int i=0;i<numNewNodes-1;i++)
+	{
+
+		if (bvh->getQuantizedNodeArray()[i].isLeafNode())
+		{
+			int orgIndex = bvh->getQuantizedNodeArray()[i].getTriangleIndex();
+
+			b3Vector3 nodeMinVec = bvh->unQuantize(bvh->getQuantizedNodeArray()[i].m_quantizedAabbMin);
+			b3Vector3 nodeMaxVec = bvh->unQuantize(bvh->getQuantizedNodeArray()[i].m_quantizedAabbMax);
+		
+			for (int c=0;c<3;c++)
+			{
+				if (childLocalAabbs[orgIndex].m_min[c] < nodeMinVec[c])
+				{
+					printf("min org (%f) and new (%f) ? at i:%d,c:%d\n",childLocalAabbs[i].m_min[c],nodeMinVec[c],i,c);
+				}
+				if (childLocalAabbs[orgIndex].m_max[c] > nodeMaxVec[c])
+				{
+					printf("max org (%f) and new (%f) ? at i:%d,c:%d\n",childLocalAabbs[i].m_max[c],nodeMaxVec[c],i,c);
+				}
+
+			}
+		}
+
+	}
+
+	m_data->m_bvhInfoCPU.push_back(bvhInfo);
+
+	int numNewSubtrees = bvh->getSubtreeInfoArray().size();
+	m_data->m_subTreesCPU.reserve(m_data->m_subTreesCPU.size()+numNewSubtrees);
+	for (int i=0;i<numNewSubtrees;i++)
+	{
+		m_data->m_subTreesCPU.push_back(bvh->getSubtreeInfoArray()[i]);
+	}
+	int numNewTreeNodes = bvh->getQuantizedNodeArray().size();
+
+	for (int i=0;i<numNewTreeNodes;i++)
+	{
+		m_data->m_treeNodesCPU.push_back(bvh->getQuantizedNodeArray()[i]);
+	}
+
+//	m_data->m_localShapeAABBGPU->push_back(aabbWS);
+	clFinish(m_queue);
+	return collidableIndex;
+	
+}
+
+
+int		b3GpuNarrowPhase::registerConcaveMesh(b3AlignedObjectArray<b3Vector3>* vertices, b3AlignedObjectArray<int>* indices,const float* scaling1)
+{
+	
+
+	b3Vector3 scaling=b3MakeVector3(scaling1[0],scaling1[1],scaling1[2]);
+
+	int collidableIndex = allocateCollidable();
+	if (collidableIndex<0)
+		return collidableIndex;
+
+	b3Collidable& col = getCollidableCpu(collidableIndex);
+	
+	col.m_shapeType = SHAPE_CONCAVE_TRIMESH;
+	col.m_shapeIndex = registerConcaveMeshShape(vertices,indices,col,scaling);
+	col.m_bvhIndex = m_data->m_bvhInfoCPU.size();
+		
+
+	b3SapAabb aabb;
+	b3Vector3 myAabbMin=b3MakeVector3(1e30f,1e30f,1e30f);
+	b3Vector3 myAabbMax=b3MakeVector3(-1e30f,-1e30f,-1e30f);
+
+	for (int i=0;i<vertices->size();i++)
+	{
+		b3Vector3 vtx(vertices->at(i)*scaling);
+		myAabbMin.setMin(vtx);
+		myAabbMax.setMax(vtx);
+	}
+	aabb.m_min[0] = myAabbMin[0];
+	aabb.m_min[1] = myAabbMin[1];
+	aabb.m_min[2] = myAabbMin[2];
+	aabb.m_minIndices[3] = 0;
+
+	aabb.m_max[0] = myAabbMax[0];
+	aabb.m_max[1]= myAabbMax[1];
+	aabb.m_max[2]= myAabbMax[2];
+	aabb.m_signedMaxIndices[3]= 0;
+
+	m_data->m_localShapeAABBCPU->push_back(aabb);
+//	m_data->m_localShapeAABBGPU->push_back(aabb);
+
+	b3OptimizedBvh* bvh = new b3OptimizedBvh();
+	//void b3OptimizedBvh::build(b3StridingMeshInterface* triangles, bool useQuantizedAabbCompression, const b3Vector3& bvhAabbMin, const b3Vector3& bvhAabbMax)
+	
+	bool useQuantizedAabbCompression = true;
+	b3TriangleIndexVertexArray* meshInterface=new b3TriangleIndexVertexArray();
+	m_data->m_meshInterfaces.push_back(meshInterface);
+	b3IndexedMesh mesh;
+	mesh.m_numTriangles = indices->size()/3;
+	mesh.m_numVertices = vertices->size();
+	mesh.m_vertexBase = (const unsigned char *)&vertices->at(0).x;
+	mesh.m_vertexStride = sizeof(b3Vector3);
+	mesh.m_triangleIndexStride = 3 * sizeof(int);// or sizeof(int)
+	mesh.m_triangleIndexBase = (const unsigned char *)&indices->at(0);
+	
+	meshInterface->addIndexedMesh(mesh);
+	bvh->build(meshInterface, useQuantizedAabbCompression, (b3Vector3&)aabb.m_min, (b3Vector3&)aabb.m_max);
+	m_data->m_bvhData.push_back(bvh);
+	int numNodes = bvh->getQuantizedNodeArray().size();
+	//b3OpenCLArray<b3QuantizedBvhNode>*	treeNodesGPU = new b3OpenCLArray<b3QuantizedBvhNode>(this->m_context,this->m_queue,numNodes);
+	int numSubTrees = bvh->getSubtreeInfoArray().size();
+
+	b3BvhInfo bvhInfo;
+	
+	bvhInfo.m_aabbMin = bvh->m_bvhAabbMin;
+	bvhInfo.m_aabbMax = bvh->m_bvhAabbMax;
+	bvhInfo.m_quantization = bvh->m_bvhQuantization;
+	bvhInfo.m_numNodes = numNodes;
+	bvhInfo.m_numSubTrees = numSubTrees;
+	bvhInfo.m_nodeOffset = m_data->m_treeNodesCPU.size();
+	bvhInfo.m_subTreeOffset = m_data->m_subTreesCPU.size();
+
+	m_data->m_bvhInfoCPU.push_back(bvhInfo);
+
+
+	int numNewSubtrees = bvh->getSubtreeInfoArray().size();
+	m_data->m_subTreesCPU.reserve(m_data->m_subTreesCPU.size()+numNewSubtrees);
+	for (int i=0;i<numNewSubtrees;i++)
+	{
+		m_data->m_subTreesCPU.push_back(bvh->getSubtreeInfoArray()[i]);
+	}
+	int numNewTreeNodes = bvh->getQuantizedNodeArray().size();
+
+	for (int i=0;i<numNewTreeNodes;i++)
+	{
+		m_data->m_treeNodesCPU.push_back(bvh->getQuantizedNodeArray()[i]);
+	}
+
+
+
+
+	return collidableIndex;
+}
+
+int b3GpuNarrowPhase::registerConcaveMeshShape(b3AlignedObjectArray<b3Vector3>* vertices, b3AlignedObjectArray<int>* indices,b3Collidable& col, const float* scaling1)
+{
+
+
+	b3Vector3 scaling=b3MakeVector3(scaling1[0],scaling1[1],scaling1[2]);
+
+	m_data->m_convexData->resize(m_data->m_numAcceleratedShapes+1);
+	m_data->m_convexPolyhedra.resize(m_data->m_numAcceleratedShapes+1);
+	
+    
+	b3ConvexPolyhedronData& convex = m_data->m_convexPolyhedra.at(m_data->m_convexPolyhedra.size()-1);
+	convex.mC = b3MakeVector3(0,0,0);
+	convex.mE = b3MakeVector3(0,0,0);
+	convex.m_extents= b3MakeVector3(0,0,0);
+	convex.m_localCenter = b3MakeVector3(0,0,0);
+	convex.m_radius = 0.f;
+	
+	convex.m_numUniqueEdges = 0;
+	int edgeOffset = m_data->m_uniqueEdges.size();
+	convex.m_uniqueEdgesOffset = edgeOffset;
+	
+	int faceOffset = m_data->m_convexFaces.size();
+	convex.m_faceOffset = faceOffset;
+	
+	convex.m_numFaces = indices->size()/3;
+	m_data->m_convexFaces.resize(faceOffset+convex.m_numFaces);
+	m_data->m_convexIndices.reserve(convex.m_numFaces*3);
+	for (int i=0;i<convex.m_numFaces;i++)
+	{
+		if (i%256==0)
+		{
+			//printf("i=%d out of %d", i,convex.m_numFaces);
+		}
+		b3Vector3 vert0(vertices->at(indices->at(i*3))*scaling);
+		b3Vector3 vert1(vertices->at(indices->at(i*3+1))*scaling);
+		b3Vector3 vert2(vertices->at(indices->at(i*3+2))*scaling);
+
+		b3Vector3 normal = ((vert1-vert0).cross(vert2-vert0)).normalize();
+		b3Scalar c = -(normal.dot(vert0));
+
+		m_data->m_convexFaces[convex.m_faceOffset+i].m_plane = b3MakeVector4(normal.x,normal.y,normal.z,c);
+		int indexOffset = m_data->m_convexIndices.size();
+		int numIndices = 3;
+		m_data->m_convexFaces[convex.m_faceOffset+i].m_numIndices = numIndices;
+		m_data->m_convexFaces[convex.m_faceOffset+i].m_indexOffset = indexOffset;
+		m_data->m_convexIndices.resize(indexOffset+numIndices);
+		for (int p=0;p<numIndices;p++)
+		{
+			int vi = indices->at(i*3+p);
+			m_data->m_convexIndices[indexOffset+p] = vi;//convexPtr->m_faces[i].m_indices[p];
+		}
+	}
+    
+	convex.m_numVertices = vertices->size();
+	int vertexOffset = m_data->m_convexVertices.size();
+	convex.m_vertexOffset =vertexOffset;
+	m_data->m_convexVertices.resize(vertexOffset+convex.m_numVertices);
+	for (int i=0;i<vertices->size();i++)
+	{
+		m_data->m_convexVertices[vertexOffset+i] = vertices->at(i)*scaling;
+	}
+
+	(*m_data->m_convexData)[m_data->m_numAcceleratedShapes] = 0;
+	
+  
+	return m_data->m_numAcceleratedShapes++;
+}
+
+
+
+cl_mem	b3GpuNarrowPhase::getBodiesGpu()
+{
+	return (cl_mem)m_data->m_bodyBufferGPU->getBufferCL();
+}
+
+const struct b3RigidBodyData* b3GpuNarrowPhase::getBodiesCpu() const
+{
+	return &m_data->m_bodyBufferCPU->at(0);
+};
+
+
+
+
+int	b3GpuNarrowPhase::getNumBodiesGpu() const
+{
+	return m_data->m_bodyBufferGPU->size();
+}
+
+cl_mem	b3GpuNarrowPhase::getBodyInertiasGpu()
+{
+	return (cl_mem)m_data->m_inertiaBufferGPU->getBufferCL();
+}
+
+int	b3GpuNarrowPhase::getNumBodyInertiasGpu() const
+{
+	return m_data->m_inertiaBufferGPU->size();
+}
+
+
+b3Collidable& b3GpuNarrowPhase::getCollidableCpu(int collidableIndex)
+{
+	return m_data->m_collidablesCPU[collidableIndex];
+}
+
+const b3Collidable& b3GpuNarrowPhase::getCollidableCpu(int collidableIndex) const
+{
+	return m_data->m_collidablesCPU[collidableIndex];
+}
+
+cl_mem b3GpuNarrowPhase::getCollidablesGpu()
+{
+	return m_data->m_collidablesGPU->getBufferCL();
+}
+
+const struct b3Collidable* b3GpuNarrowPhase::getCollidablesCpu() const
+{
+	if (m_data->m_collidablesCPU.size())
+		return &m_data->m_collidablesCPU[0];
+	return 0;
+}
+
+const struct b3SapAabb* b3GpuNarrowPhase::getLocalSpaceAabbsCpu() const
+{
+	if (m_data->m_localShapeAABBCPU->size())
+	{
+		return &m_data->m_localShapeAABBCPU->at(0);
+	} 
+	return 0;
+}
+
+
+cl_mem	b3GpuNarrowPhase::getAabbLocalSpaceBufferGpu()
+{
+	return m_data->m_localShapeAABBGPU->getBufferCL();
+}
+int	b3GpuNarrowPhase::getNumCollidablesGpu() const
+{
+	return m_data->m_collidablesGPU->size();
+}
+
+
+
+
+
+int	b3GpuNarrowPhase::getNumContactsGpu() const
+{
+	return m_data->m_pBufContactBuffersGPU[m_data->m_currentContactBuffer]->size();
+}
+cl_mem b3GpuNarrowPhase::getContactsGpu()
+{
+	return m_data->m_pBufContactBuffersGPU[m_data->m_currentContactBuffer]->getBufferCL();
+}
+
+const b3Contact4* b3GpuNarrowPhase::getContactsCPU() const
+{
+	m_data->m_pBufContactBuffersGPU[m_data->m_currentContactBuffer]->copyToHost(*m_data->m_pBufContactOutCPU);
+	return &m_data->m_pBufContactOutCPU->at(0);
+}
+
+void b3GpuNarrowPhase::computeContacts(cl_mem broadphasePairs, int numBroadphasePairs, cl_mem aabbsWorldSpace, int numObjects)
+{
+
+	cl_mem aabbsLocalSpace = m_data->m_localShapeAABBGPU->getBufferCL();
+
+	int nContactOut = 0;
+
+	//swap buffer
+	m_data->m_currentContactBuffer=1-m_data->m_currentContactBuffer;
+
+	//int curSize = m_data->m_pBufContactBuffersGPU[m_data->m_currentContactBuffer]->size();
+
+	int maxTriConvexPairCapacity = m_data->m_config.m_maxTriConvexPairCapacity;
+	int numTriConvexPairsOut=0;
+	
+	b3OpenCLArray<b3Int4> broadphasePairsGPU(m_context,m_queue);
+	broadphasePairsGPU.setFromOpenCLBuffer(broadphasePairs,numBroadphasePairs);
+
+	
+
+
+	b3OpenCLArray<b3Aabb> clAabbArrayWorldSpace(this->m_context,this->m_queue);
+	clAabbArrayWorldSpace.setFromOpenCLBuffer(aabbsWorldSpace,numObjects);
+
+	b3OpenCLArray<b3Aabb> clAabbArrayLocalSpace(this->m_context,this->m_queue);
+	clAabbArrayLocalSpace.setFromOpenCLBuffer(aabbsLocalSpace,numObjects);
+
+	m_data->m_gpuSatCollision->computeConvexConvexContactsGPUSAT(
+		&broadphasePairsGPU, numBroadphasePairs,
+		m_data->m_bodyBufferGPU,
+		m_data->m_pBufContactBuffersGPU[m_data->m_currentContactBuffer],
+		nContactOut,
+		m_data->m_pBufContactBuffersGPU[1-m_data->m_currentContactBuffer],
+		m_data->m_config.m_maxContactCapacity,
+		m_data->m_config.m_compoundPairCapacity,
+		*m_data->m_convexPolyhedraGPU,
+		*m_data->m_convexVerticesGPU,
+		*m_data->m_uniqueEdgesGPU,
+		*m_data->m_convexFacesGPU,
+		*m_data->m_convexIndicesGPU,
+		*m_data->m_collidablesGPU,
+		*m_data->m_gpuChildShapes,
+		clAabbArrayWorldSpace,
+		clAabbArrayLocalSpace,
+		*m_data->m_worldVertsB1GPU,
+		*m_data->m_clippingFacesOutGPU,
+		*m_data->m_worldNormalsAGPU,
+		*m_data->m_worldVertsA1GPU,
+		*m_data->m_worldVertsB2GPU,
+		m_data->m_bvhData,
+		m_data->m_treeNodesGPU,
+		m_data->m_subTreesGPU,
+		m_data->m_bvhInfoGPU,
+		numObjects,
+		maxTriConvexPairCapacity,
+		*m_data->m_triangleConvexPairs,
+		numTriConvexPairsOut
+		);
+
+	/*b3AlignedObjectArray<b3Int4> broadphasePairsCPU;
+	broadphasePairsGPU.copyToHost(broadphasePairsCPU);
+	printf("checking pairs\n");
+	*/
+}
+
+const b3SapAabb& b3GpuNarrowPhase::getLocalSpaceAabb(int collidableIndex) const
+{
+	return m_data->m_localShapeAABBCPU->at(collidableIndex);
+}
+
+
+
+
+
+int b3GpuNarrowPhase::registerRigidBody(int collidableIndex, float mass, const float* position, const float* orientation , const float* aabbMinPtr, const float* aabbMaxPtr,bool writeToGpu)
+{
+	b3Vector3 aabbMin=b3MakeVector3(aabbMinPtr[0],aabbMinPtr[1],aabbMinPtr[2]);
+	b3Vector3 aabbMax=b3MakeVector3(aabbMaxPtr[0],aabbMaxPtr[1],aabbMaxPtr[2]);
+	
+
+	if (m_data->m_numAcceleratedRigidBodies >= (m_data->m_config.m_maxConvexBodies))
+	{
+		b3Error("registerRigidBody: exceeding the number of rigid bodies, %d > %d \n",m_data->m_numAcceleratedRigidBodies,m_data->m_config.m_maxConvexBodies);
+		return -1;
+	}
+    
+	m_data->m_bodyBufferCPU->resize(m_data->m_numAcceleratedRigidBodies+1);
+    
+	b3RigidBodyData& body = m_data->m_bodyBufferCPU->at(m_data->m_numAcceleratedRigidBodies);
+    
+	float friction = 1.f;
+	float restitution = 0.f;
+    
+	body.m_frictionCoeff = friction;
+	body.m_restituitionCoeff = restitution;
+	body.m_angVel = b3MakeVector3(0,0,0);
+	body.m_linVel=b3MakeVector3(0,0,0);//.setZero();
+	body.m_pos =b3MakeVector3(position[0],position[1],position[2]);
+	body.m_quat.setValue(orientation[0],orientation[1],orientation[2],orientation[3]);
+	body.m_collidableIdx = collidableIndex;
+	if (collidableIndex>=0)
+	{
+//		body.m_shapeType = m_data->m_collidablesCPU.at(collidableIndex).m_shapeType;
+	} else
+	{
+	//	body.m_shapeType = CollisionShape::SHAPE_PLANE;
+		m_planeBodyIndex = m_data->m_numAcceleratedRigidBodies;
+	}
+	//body.m_shapeType = shapeType;
+	
+	
+	body.m_invMass = mass? 1.f/mass : 0.f;
+    
+	if (writeToGpu)
+	{
+		m_data->m_bodyBufferGPU->copyFromHostPointer(&body,1,m_data->m_numAcceleratedRigidBodies);
+	}
+    
+	b3InertiaData& shapeInfo = m_data->m_inertiaBufferCPU->at(m_data->m_numAcceleratedRigidBodies);
+    
+	if (mass==0.f)
+	{
+		if (m_data->m_numAcceleratedRigidBodies==0)
+			m_static0Index = 0;
+        
+		shapeInfo.m_initInvInertia.setValue(0,0,0,0,0,0,0,0,0);
+		shapeInfo.m_invInertiaWorld.setValue(0,0,0,0,0,0,0,0,0);
+	} else
+	{
+        
+		b3Assert(body.m_collidableIdx>=0);
+        
+		//approximate using the aabb of the shape
+        
+		//Aabb aabb = (*m_data->m_shapePointers)[shapeIndex]->m_aabb;
+		b3Vector3 halfExtents = (aabbMax-aabbMin);//*0.5f;//fake larger inertia makes demos more stable ;-)
+        
+		b3Vector3 localInertia;
+        
+		float lx=2.f*halfExtents[0];
+		float ly=2.f*halfExtents[1];
+		float lz=2.f*halfExtents[2];
+        
+		localInertia.setValue( (mass/12.0f) * (ly*ly + lz*lz),
+                                   (mass/12.0f) * (lx*lx + lz*lz),
+                                   (mass/12.0f) * (lx*lx + ly*ly));
+        
+		b3Vector3 invLocalInertia;
+		invLocalInertia[0] = 1.f/localInertia[0];
+		invLocalInertia[1] = 1.f/localInertia[1];
+		invLocalInertia[2] = 1.f/localInertia[2];
+		invLocalInertia[3] = 0.f;
+        
+		shapeInfo.m_initInvInertia.setValue(
+			invLocalInertia[0],		0,						0,
+			0,						invLocalInertia[1],		0,
+			0,						0,						invLocalInertia[2]);
+
+		b3Matrix3x3 m (body.m_quat);
+
+		shapeInfo.m_invInertiaWorld = m.scaled(invLocalInertia) * m.transpose();
+        
+	}
+    
+	if (writeToGpu)
+		m_data->m_inertiaBufferGPU->copyFromHostPointer(&shapeInfo,1,m_data->m_numAcceleratedRigidBodies);
+    
+    
+    
+	return m_data->m_numAcceleratedRigidBodies++;
+}
+
+int b3GpuNarrowPhase::getNumRigidBodies() const
+{
+	return m_data->m_numAcceleratedRigidBodies;
+}
+
+void	b3GpuNarrowPhase::writeAllBodiesToGpu()
+{
+
+	if (m_data->m_localShapeAABBCPU->size())
+	{
+		m_data->m_localShapeAABBGPU->copyFromHost(*m_data->m_localShapeAABBCPU);
+	}
+	
+	
+	m_data->m_gpuChildShapes->copyFromHost(m_data->m_cpuChildShapes);
+	m_data->m_convexFacesGPU->copyFromHost(m_data->m_convexFaces);
+	m_data->m_convexPolyhedraGPU->copyFromHost(m_data->m_convexPolyhedra);
+	m_data->m_uniqueEdgesGPU->copyFromHost(m_data->m_uniqueEdges);
+	m_data->m_convexVerticesGPU->copyFromHost(m_data->m_convexVertices);
+	m_data->m_convexIndicesGPU->copyFromHost(m_data->m_convexIndices);
+	m_data->m_bvhInfoGPU->copyFromHost(m_data->m_bvhInfoCPU);
+	m_data->m_treeNodesGPU->copyFromHost(m_data->m_treeNodesCPU);
+	m_data->m_subTreesGPU->copyFromHost(m_data->m_subTreesCPU);
+
+
+	m_data->m_bodyBufferGPU->resize(m_data->m_numAcceleratedRigidBodies);
+	m_data->m_inertiaBufferGPU->resize(m_data->m_numAcceleratedRigidBodies);
+    
+	if (m_data->m_numAcceleratedRigidBodies)
+	{
+		m_data->m_bodyBufferGPU->copyFromHostPointer(&m_data->m_bodyBufferCPU->at(0),m_data->m_numAcceleratedRigidBodies);
+		m_data->m_inertiaBufferGPU->copyFromHostPointer(&m_data->m_inertiaBufferCPU->at(0),m_data->m_numAcceleratedRigidBodies);
+	}
+    if (m_data->m_collidablesCPU.size())
+	{
+		m_data->m_collidablesGPU->copyFromHost(m_data->m_collidablesCPU);
+	}
+	
+    
+}
+
+
+void	b3GpuNarrowPhase::reset()
+{
+	m_data->m_numAcceleratedShapes = 0;
+	m_data->m_numAcceleratedRigidBodies = 0;
+	this->m_static0Index = -1;
+	m_data->m_uniqueEdges.resize(0);
+	m_data->m_convexVertices.resize(0);
+	m_data->m_convexPolyhedra.resize(0);
+	m_data->m_convexIndices.resize(0);
+	m_data->m_cpuChildShapes.resize(0);
+	m_data->m_convexFaces.resize(0);
+	m_data->m_collidablesCPU.resize(0);
+	m_data->m_localShapeAABBCPU->resize(0);
+	m_data->m_bvhData.resize(0);
+	m_data->m_treeNodesCPU.resize(0);
+	m_data->m_subTreesCPU.resize(0);
+	m_data->m_bvhInfoCPU.resize(0);
+
+}
+
+
+void	b3GpuNarrowPhase::readbackAllBodiesToCpu()
+{
+	m_data->m_bodyBufferGPU->copyToHostPointer(&m_data->m_bodyBufferCPU->at(0),m_data->m_numAcceleratedRigidBodies);
+}
+
+void b3GpuNarrowPhase::setObjectTransformCpu(float* position, float* orientation , int bodyIndex)
+{
+	if (bodyIndex>=0 && bodyIndex<m_data->m_bodyBufferCPU->size())
+	{
+		m_data->m_bodyBufferCPU->at(bodyIndex).m_pos=b3MakeVector3(position[0],position[1],position[2]);
+		m_data->m_bodyBufferCPU->at(bodyIndex).m_quat.setValue(orientation[0],orientation[1],orientation[2],orientation[3]);
+	}
+	else
+	{
+		b3Warning("setObjectVelocityCpu out of range.\n");
+	}
+}
+void b3GpuNarrowPhase::setObjectVelocityCpu(float* linVel, float* angVel, int bodyIndex)
+{
+	if (bodyIndex>=0 && bodyIndex<m_data->m_bodyBufferCPU->size())
+	{
+		m_data->m_bodyBufferCPU->at(bodyIndex).m_linVel=b3MakeVector3(linVel[0],linVel[1],linVel[2]);
+		m_data->m_bodyBufferCPU->at(bodyIndex).m_angVel=b3MakeVector3(angVel[0],angVel[1],angVel[2]);
+	} else
+	{
+		b3Warning("setObjectVelocityCpu out of range.\n");
+	}
+}
+
+bool b3GpuNarrowPhase::getObjectTransformFromCpu(float* position, float* orientation , int bodyIndex) const
+{
+	if (bodyIndex>=0 && bodyIndex<m_data->m_bodyBufferCPU->size())
+	{
+		position[0] = m_data->m_bodyBufferCPU->at(bodyIndex).m_pos.x;
+		position[1] = m_data->m_bodyBufferCPU->at(bodyIndex).m_pos.y;
+		position[2] = m_data->m_bodyBufferCPU->at(bodyIndex).m_pos.z;
+		position[3] = 1.f;//or 1
+
+		orientation[0] = m_data->m_bodyBufferCPU->at(bodyIndex).m_quat.x;
+		orientation[1] = m_data->m_bodyBufferCPU->at(bodyIndex).m_quat.y;
+		orientation[2] = m_data->m_bodyBufferCPU->at(bodyIndex).m_quat.z;
+		orientation[3] = m_data->m_bodyBufferCPU->at(bodyIndex).m_quat.w;
+		return true;
+	}
+
+	b3Warning("getObjectTransformFromCpu out of range.\n");
+	return false;
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuNarrowPhase.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuNarrowPhase.h
new file mode 100644
index 0000000000..05ff3fd09e
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuNarrowPhase.h
@@ -0,0 +1,109 @@
+#ifndef B3_GPU_NARROWPHASE_H
+#define B3_GPU_NARROWPHASE_H
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Collidable.h"
+#include "Bullet3OpenCL/Initialize/b3OpenCLInclude.h"
+#include "Bullet3Common/b3AlignedObjectArray.h"
+#include "Bullet3Common/b3Vector3.h"
+
+class b3GpuNarrowPhase
+{
+protected:
+
+	struct b3GpuNarrowPhaseInternalData*	m_data;
+	int m_acceleratedCompanionShapeIndex;
+	int m_planeBodyIndex;
+	int	m_static0Index;
+
+	cl_context m_context;
+	cl_device_id m_device;
+	cl_command_queue m_queue;
+
+	int registerConvexHullShapeInternal(class b3ConvexUtility* convexPtr, b3Collidable& col);
+	int registerConcaveMeshShape(b3AlignedObjectArray<b3Vector3>* vertices, b3AlignedObjectArray<int>* indices, b3Collidable& col, const float* scaling);
+
+public:
+
+	
+
+
+	b3GpuNarrowPhase(cl_context vtx, cl_device_id dev, cl_command_queue q, const struct b3Config& config);
+
+	virtual ~b3GpuNarrowPhase(void);
+
+	int		registerSphereShape(float radius);
+	int		registerPlaneShape(const b3Vector3& planeNormal, float planeConstant);
+
+	int registerCompoundShape(b3AlignedObjectArray<b3GpuChildShape>* childShapes);
+	int registerFace(const b3Vector3& faceNormal, float faceConstant);
+	
+	int	registerConcaveMesh(b3AlignedObjectArray<b3Vector3>* vertices, b3AlignedObjectArray<int>* indices,const float* scaling);
+	
+	//do they need to be merged?
+	
+	int	registerConvexHullShape(b3ConvexUtility* utilPtr);
+	int	registerConvexHullShape(const float* vertices, int strideInBytes, int numVertices, const float* scaling);
+
+	int registerRigidBody(int collidableIndex, float mass, const float* position, const float* orientation, const float* aabbMin, const float* aabbMax,bool writeToGpu);
+	void setObjectTransform(const float* position, const float* orientation , int bodyIndex);
+
+	void	writeAllBodiesToGpu();
+	void  reset();
+	void	readbackAllBodiesToCpu();
+	bool	getObjectTransformFromCpu(float* position, float* orientation , int bodyIndex) const;
+
+	void setObjectTransformCpu(float* position, float* orientation , int bodyIndex);
+	void setObjectVelocityCpu(float* linVel, float* angVel, int bodyIndex);
+
+	
+	virtual void computeContacts(cl_mem broadphasePairs, int numBroadphasePairs, cl_mem aabbsWorldSpace, int numObjects);
+	
+
+	cl_mem	getBodiesGpu();
+	const struct b3RigidBodyData* getBodiesCpu() const;
+	//struct b3RigidBodyData* getBodiesCpu();
+
+	int	getNumBodiesGpu() const;
+
+	cl_mem	getBodyInertiasGpu();
+	int	getNumBodyInertiasGpu() const;
+
+	cl_mem	getCollidablesGpu();
+	const struct b3Collidable* getCollidablesCpu() const;
+	int		getNumCollidablesGpu() const;
+
+	const struct b3SapAabb* getLocalSpaceAabbsCpu() const;
+
+	const struct b3Contact4* getContactsCPU() const;
+
+	cl_mem	getContactsGpu();
+	int	getNumContactsGpu() const;
+
+	cl_mem	getAabbLocalSpaceBufferGpu();
+	
+	int getNumRigidBodies() const;
+
+	int allocateCollidable();
+
+	int getStatic0Index() const
+	{
+		return m_static0Index;
+	}
+	b3Collidable& getCollidableCpu(int collidableIndex);
+	const b3Collidable& getCollidableCpu(int collidableIndex) const;
+
+	const b3GpuNarrowPhaseInternalData*	getInternalData() const
+	{
+			return m_data;
+	}
+
+	b3GpuNarrowPhaseInternalData*	getInternalData()
+	{
+			return m_data;
+	}
+
+	const struct b3SapAabb& getLocalSpaceAabb(int collidableIndex) const;
+};
+
+#endif //B3_GPU_NARROWPHASE_H
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuNarrowPhaseInternalData.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuNarrowPhaseInternalData.h
new file mode 100644
index 0000000000..8a7f1ea859
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuNarrowPhaseInternalData.h
@@ -0,0 +1,95 @@
+
+#ifndef B3_GPU_NARROWPHASE_INTERNAL_DATA_H
+#define B3_GPU_NARROWPHASE_INTERNAL_DATA_H
+
+#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3ConvexPolyhedronData.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3Config.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Collidable.h"
+
+#include "Bullet3OpenCL/Initialize/b3OpenCLInclude.h"
+#include "Bullet3Common/b3AlignedObjectArray.h"
+#include "Bullet3Common/b3Vector3.h"
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3RigidBodyData.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3Contact4.h"
+#include "Bullet3OpenCL/BroadphaseCollision/b3SapAabb.h"
+
+#include "Bullet3OpenCL/NarrowphaseCollision/b3QuantizedBvh.h"
+#include "Bullet3OpenCL/NarrowphaseCollision/b3BvhInfo.h"
+#include "Bullet3Common/shared/b3Int4.h"
+#include "Bullet3Common/shared/b3Int2.h"
+
+
+class b3ConvexUtility;
+
+struct b3GpuNarrowPhaseInternalData
+{
+	b3AlignedObjectArray<b3ConvexUtility*>* m_convexData;
+    
+	b3AlignedObjectArray<b3ConvexPolyhedronData> m_convexPolyhedra;
+	b3AlignedObjectArray<b3Vector3> m_uniqueEdges;
+	b3AlignedObjectArray<b3Vector3> m_convexVertices;
+	b3AlignedObjectArray<int> m_convexIndices;
+    
+	b3OpenCLArray<b3ConvexPolyhedronData>* m_convexPolyhedraGPU;
+	b3OpenCLArray<b3Vector3>* m_uniqueEdgesGPU;
+	b3OpenCLArray<b3Vector3>* m_convexVerticesGPU;
+	b3OpenCLArray<int>* m_convexIndicesGPU;
+    
+    b3OpenCLArray<b3Vector3>* m_worldVertsB1GPU;
+    b3OpenCLArray<b3Int4>* m_clippingFacesOutGPU;
+    b3OpenCLArray<b3Vector3>* m_worldNormalsAGPU;
+    b3OpenCLArray<b3Vector3>* m_worldVertsA1GPU;
+    b3OpenCLArray<b3Vector3>* m_worldVertsB2GPU;
+    
+	b3AlignedObjectArray<b3GpuChildShape> m_cpuChildShapes;
+	b3OpenCLArray<b3GpuChildShape>*	m_gpuChildShapes;
+    
+	b3AlignedObjectArray<b3GpuFace> m_convexFaces;
+	b3OpenCLArray<b3GpuFace>* m_convexFacesGPU;
+    
+	struct GpuSatCollision*	m_gpuSatCollision;
+	    
+	
+	b3OpenCLArray<b3Int4>*			m_triangleConvexPairs;
+    
+	
+	b3OpenCLArray<b3Contact4>* m_pBufContactBuffersGPU[2];
+	int	m_currentContactBuffer;
+	b3AlignedObjectArray<b3Contact4>* m_pBufContactOutCPU;
+	
+    
+	b3AlignedObjectArray<b3RigidBodyData>* m_bodyBufferCPU;
+	b3OpenCLArray<b3RigidBodyData>* m_bodyBufferGPU;
+    
+	b3AlignedObjectArray<b3InertiaData>*	m_inertiaBufferCPU;
+	b3OpenCLArray<b3InertiaData>*	m_inertiaBufferGPU;
+    
+	int m_numAcceleratedShapes;
+	int m_numAcceleratedRigidBodies;
+    
+	b3AlignedObjectArray<b3Collidable>	m_collidablesCPU;
+	b3OpenCLArray<b3Collidable>*	m_collidablesGPU;
+
+	b3OpenCLArray<b3SapAabb>* m_localShapeAABBGPU;
+	b3AlignedObjectArray<b3SapAabb>* m_localShapeAABBCPU;
+
+	b3AlignedObjectArray<class b3OptimizedBvh*> m_bvhData;
+	b3AlignedObjectArray<class b3TriangleIndexVertexArray*> m_meshInterfaces;
+
+	b3AlignedObjectArray<b3QuantizedBvhNode>	m_treeNodesCPU;
+	b3AlignedObjectArray<b3BvhSubtreeInfo>	m_subTreesCPU;
+
+	b3AlignedObjectArray<b3BvhInfo>	m_bvhInfoCPU;
+	b3OpenCLArray<b3BvhInfo>*			m_bvhInfoGPU;
+	
+	b3OpenCLArray<b3QuantizedBvhNode>*	m_treeNodesGPU;
+	b3OpenCLArray<b3BvhSubtreeInfo>*	m_subTreesGPU;
+	
+
+	b3Config	m_config;
+    
+};
+
+#endif //B3_GPU_NARROWPHASE_INTERNAL_DATA_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuPgsConstraintSolver.cpp b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuPgsConstraintSolver.cpp
new file mode 100644
index 0000000000..0d3d50c548
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuPgsConstraintSolver.cpp
@@ -0,0 +1,1158 @@
+
+/*
+Copyright (c) 2013 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+
+bool useGpuInitSolverBodies = true;
+bool useGpuInfo1 = true;
+bool useGpuInfo2= true;
+bool useGpuSolveJointConstraintRows=true;
+bool useGpuWriteBackVelocities = true;
+bool gpuBreakConstraints = true;
+
+#include "b3GpuPgsConstraintSolver.h"
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3RigidBodyData.h"
+
+#include "Bullet3Dynamics/ConstraintSolver/b3TypedConstraint.h"
+#include <new>
+#include "Bullet3Common/b3AlignedObjectArray.h"
+#include <string.h> //for memset
+#include "Bullet3Collision/NarrowPhaseCollision/b3Contact4.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h"
+
+#include "Bullet3OpenCL/ParallelPrimitives/b3PrefixScanCL.h"
+
+#include "Bullet3OpenCL/RigidBody/kernels/jointSolver.h" //solveConstraintRowsCL
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+
+#define B3_JOINT_SOLVER_PATH "src/Bullet3OpenCL/RigidBody/kernels/jointSolver.cl"
+
+
+struct b3GpuPgsJacobiSolverInternalData
+{
+
+	cl_context m_context;
+	cl_device_id m_device;
+	cl_command_queue m_queue;
+
+	b3PrefixScanCL*	m_prefixScan;
+
+	cl_kernel m_solveJointConstraintRowsKernels;
+	cl_kernel m_initSolverBodiesKernel;
+	cl_kernel m_getInfo1Kernel;
+	cl_kernel m_initBatchConstraintsKernel;
+	cl_kernel m_getInfo2Kernel;
+	cl_kernel m_writeBackVelocitiesKernel;
+	cl_kernel m_breakViolatedConstraintsKernel;
+
+	b3OpenCLArray<unsigned int>*	m_gpuConstraintRowOffsets;
+
+	b3OpenCLArray<b3GpuSolverBody>*			m_gpuSolverBodies;
+	b3OpenCLArray<b3BatchConstraint>*		m_gpuBatchConstraints;
+	b3OpenCLArray<b3GpuSolverConstraint>*		m_gpuConstraintRows;
+	b3OpenCLArray<unsigned int>*			m_gpuConstraintInfo1;
+
+//	b3AlignedObjectArray<b3GpuSolverBody>		m_cpuSolverBodies;
+	b3AlignedObjectArray<b3BatchConstraint>		m_cpuBatchConstraints;
+	b3AlignedObjectArray<b3GpuSolverConstraint>	m_cpuConstraintRows;
+	b3AlignedObjectArray<unsigned int>			m_cpuConstraintInfo1;
+	b3AlignedObjectArray<unsigned int>			m_cpuConstraintRowOffsets;
+
+	b3AlignedObjectArray<b3RigidBodyData>			m_cpuBodies;
+	b3AlignedObjectArray<b3InertiaData>			m_cpuInertias;
+
+	
+	b3AlignedObjectArray<b3GpuGenericConstraint> m_cpuConstraints;
+
+	b3AlignedObjectArray<int>		m_batchSizes;
+
+
+};
+
+
+/*
+static b3Transform	getWorldTransform(b3RigidBodyData* rb)
+{
+	b3Transform newTrans;
+	newTrans.setOrigin(rb->m_pos);
+	newTrans.setRotation(rb->m_quat);
+	return newTrans;
+}
+
+static const b3Matrix3x3&	getInvInertiaTensorWorld(b3InertiaData* inertia)
+{
+	return inertia->m_invInertiaWorld;
+}
+
+*/
+
+static const b3Vector3&	getLinearVelocity(b3RigidBodyData* rb)
+{
+	return rb->m_linVel;
+}
+
+static const b3Vector3&	getAngularVelocity(b3RigidBodyData* rb)
+{
+	return rb->m_angVel;
+}
+
+b3Vector3 getVelocityInLocalPoint(b3RigidBodyData* rb, const b3Vector3& rel_pos)
+{
+	//we also calculate lin/ang velocity for kinematic objects
+	return getLinearVelocity(rb) + getAngularVelocity(rb).cross(rel_pos);
+	
+}
+
+
+
+b3GpuPgsConstraintSolver::b3GpuPgsConstraintSolver (cl_context ctx, cl_device_id device, cl_command_queue queue,bool usePgs)
+{
+	m_usePgs = usePgs;
+	m_gpuData = new b3GpuPgsJacobiSolverInternalData();
+	m_gpuData->m_context = ctx;
+	m_gpuData->m_device = device;
+	m_gpuData->m_queue = queue;
+
+	m_gpuData->m_prefixScan = new b3PrefixScanCL(ctx,device,queue);
+
+	m_gpuData->m_gpuConstraintRowOffsets = new b3OpenCLArray<unsigned int>(m_gpuData->m_context,m_gpuData->m_queue);
+
+	m_gpuData->m_gpuSolverBodies = new b3OpenCLArray<b3GpuSolverBody>(m_gpuData->m_context,m_gpuData->m_queue);
+	m_gpuData->m_gpuBatchConstraints = new b3OpenCLArray<b3BatchConstraint>(m_gpuData->m_context,m_gpuData->m_queue);
+	m_gpuData->m_gpuConstraintRows = new b3OpenCLArray<b3GpuSolverConstraint>(m_gpuData->m_context,m_gpuData->m_queue);
+	m_gpuData->m_gpuConstraintInfo1 = new b3OpenCLArray<unsigned int>(m_gpuData->m_context,m_gpuData->m_queue);
+	cl_int errNum=0;
+
+	{
+		cl_program prog = b3OpenCLUtils::compileCLProgramFromString(m_gpuData->m_context,m_gpuData->m_device,solveConstraintRowsCL,&errNum,"",B3_JOINT_SOLVER_PATH);
+		//cl_program prog = b3OpenCLUtils::compileCLProgramFromString(m_gpuData->m_context,m_gpuData->m_device,0,&errNum,"",B3_JOINT_SOLVER_PATH,true);
+		b3Assert(errNum==CL_SUCCESS);
+		m_gpuData->m_solveJointConstraintRowsKernels = b3OpenCLUtils::compileCLKernelFromString(m_gpuData->m_context, m_gpuData->m_device,solveConstraintRowsCL, "solveJointConstraintRows",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+		m_gpuData->m_initSolverBodiesKernel = b3OpenCLUtils::compileCLKernelFromString(m_gpuData->m_context,m_gpuData->m_device,solveConstraintRowsCL,"initSolverBodies",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+		m_gpuData->m_getInfo1Kernel = b3OpenCLUtils::compileCLKernelFromString(m_gpuData->m_context,m_gpuData->m_device,solveConstraintRowsCL,"getInfo1Kernel",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+		m_gpuData->m_initBatchConstraintsKernel = b3OpenCLUtils::compileCLKernelFromString(m_gpuData->m_context,m_gpuData->m_device,solveConstraintRowsCL,"initBatchConstraintsKernel",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+		m_gpuData->m_getInfo2Kernel= b3OpenCLUtils::compileCLKernelFromString(m_gpuData->m_context,m_gpuData->m_device,solveConstraintRowsCL,"getInfo2Kernel",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+		m_gpuData->m_writeBackVelocitiesKernel = b3OpenCLUtils::compileCLKernelFromString(m_gpuData->m_context,m_gpuData->m_device,solveConstraintRowsCL,"writeBackVelocitiesKernel",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+		m_gpuData->m_breakViolatedConstraintsKernel = b3OpenCLUtils::compileCLKernelFromString(m_gpuData->m_context,m_gpuData->m_device,solveConstraintRowsCL,"breakViolatedConstraintsKernel",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+
+		
+		
+
+		clReleaseProgram(prog);
+	}
+
+
+}
+
+b3GpuPgsConstraintSolver::~b3GpuPgsConstraintSolver ()
+{
+	clReleaseKernel(m_gpuData->m_solveJointConstraintRowsKernels);
+	clReleaseKernel(m_gpuData->m_initSolverBodiesKernel);
+	clReleaseKernel(m_gpuData->m_getInfo1Kernel);
+	clReleaseKernel(m_gpuData->m_initBatchConstraintsKernel);
+	clReleaseKernel(m_gpuData->m_getInfo2Kernel);
+	clReleaseKernel(m_gpuData->m_writeBackVelocitiesKernel);
+	clReleaseKernel(m_gpuData->m_breakViolatedConstraintsKernel);
+
+	delete m_gpuData->m_prefixScan;
+	delete m_gpuData->m_gpuConstraintRowOffsets;
+	delete m_gpuData->m_gpuSolverBodies;
+	delete m_gpuData->m_gpuBatchConstraints;
+	delete m_gpuData->m_gpuConstraintRows;
+	delete m_gpuData->m_gpuConstraintInfo1;
+
+	delete m_gpuData;
+}
+
+struct b3BatchConstraint
+{
+	int m_bodyAPtrAndSignBit;
+	int m_bodyBPtrAndSignBit;
+	int m_originalConstraintIndex;
+	int m_batchId;
+};
+
+static b3AlignedObjectArray<b3BatchConstraint> batchConstraints;
+
+
+void	b3GpuPgsConstraintSolver::recomputeBatches()
+{
+	m_gpuData->m_batchSizes.clear();
+}
+
+
+
+
+b3Scalar b3GpuPgsConstraintSolver::solveGroupCacheFriendlySetup(b3OpenCLArray<b3RigidBodyData>* gpuBodies, b3OpenCLArray<b3InertiaData>* gpuInertias, int numBodies, b3OpenCLArray<b3GpuGenericConstraint>* gpuConstraints,int numConstraints,const b3ContactSolverInfo& infoGlobal)
+{
+	B3_PROFILE("GPU solveGroupCacheFriendlySetup");
+	batchConstraints.resize(numConstraints);
+	m_gpuData->m_gpuBatchConstraints->resize(numConstraints);
+	m_staticIdx = -1;
+	m_maxOverrideNumSolverIterations = 0;
+
+
+	/*	m_gpuData->m_gpuBodies->resize(numBodies);
+	m_gpuData->m_gpuBodies->copyFromHostPointer(bodies,numBodies);
+
+	b3OpenCLArray<b3InertiaData> gpuInertias(m_gpuData->m_context,m_gpuData->m_queue);
+	gpuInertias.resize(numBodies);
+	gpuInertias.copyFromHostPointer(inertias,numBodies);
+	*/
+
+	m_gpuData->m_gpuSolverBodies->resize(numBodies);
+
+
+	m_tmpSolverBodyPool.resize(numBodies);
+	{
+		
+		if (useGpuInitSolverBodies)
+		{
+			B3_PROFILE("m_initSolverBodiesKernel");
+
+			b3LauncherCL launcher(m_gpuData->m_queue,m_gpuData->m_initSolverBodiesKernel,"m_initSolverBodiesKernel");
+			launcher.setBuffer(m_gpuData->m_gpuSolverBodies->getBufferCL());
+			launcher.setBuffer(gpuBodies->getBufferCL());
+			launcher.setConst(numBodies);
+			launcher.launch1D(numBodies);
+			clFinish(m_gpuData->m_queue);
+
+			//			m_gpuData->m_gpuSolverBodies->copyToHost(m_tmpSolverBodyPool);
+		} else
+		{
+			gpuBodies->copyToHost(m_gpuData->m_cpuBodies);
+			for (int i=0;i<numBodies;i++)
+			{
+
+				b3RigidBodyData& body = m_gpuData->m_cpuBodies[i];
+				b3GpuSolverBody& solverBody = m_tmpSolverBodyPool[i];
+				initSolverBody(i,&solverBody,&body);
+				solverBody.m_originalBodyIndex = i;
+			}
+			m_gpuData->m_gpuSolverBodies->copyFromHost(m_tmpSolverBodyPool);
+		}
+	}
+
+//	int totalBodies = 0;
+	int totalNumRows = 0;
+	//b3RigidBody* rb0=0,*rb1=0;
+	//if (1)
+	{
+		{
+
+
+			//			int i;
+
+			m_tmpConstraintSizesPool.resizeNoInitialize(numConstraints);
+
+			//			b3OpenCLArray<b3GpuGenericConstraint> gpuConstraints(m_gpuData->m_context,m_gpuData->m_queue);
+
+
+			if (useGpuInfo1)
+			{
+				B3_PROFILE("info1 and init batchConstraint");
+				
+				m_gpuData->m_gpuConstraintInfo1->resize(numConstraints);
+
+				
+				if (1)
+				{
+					B3_PROFILE("getInfo1Kernel");
+
+					b3LauncherCL launcher(m_gpuData->m_queue,m_gpuData->m_getInfo1Kernel,"m_getInfo1Kernel");
+					launcher.setBuffer(m_gpuData->m_gpuConstraintInfo1->getBufferCL());
+					launcher.setBuffer(gpuConstraints->getBufferCL());
+					launcher.setConst(numConstraints);
+					launcher.launch1D(numConstraints);
+					clFinish(m_gpuData->m_queue);
+				}
+
+				if (m_gpuData->m_batchSizes.size()==0)
+				{
+					B3_PROFILE("initBatchConstraintsKernel");
+
+					m_gpuData->m_gpuConstraintRowOffsets->resize(numConstraints);
+					unsigned int total=0;
+					m_gpuData->m_prefixScan->execute(*m_gpuData->m_gpuConstraintInfo1,*m_gpuData->m_gpuConstraintRowOffsets,numConstraints,&total);
+					unsigned int lastElem = m_gpuData->m_gpuConstraintInfo1->at(numConstraints-1);
+					totalNumRows = total+lastElem;
+
+					{
+						B3_PROFILE("init batch constraints");
+						b3LauncherCL launcher(m_gpuData->m_queue,m_gpuData->m_initBatchConstraintsKernel,"m_initBatchConstraintsKernel");
+						launcher.setBuffer(m_gpuData->m_gpuConstraintInfo1->getBufferCL());
+						launcher.setBuffer(m_gpuData->m_gpuConstraintRowOffsets->getBufferCL());
+						launcher.setBuffer(m_gpuData->m_gpuBatchConstraints->getBufferCL());
+						launcher.setBuffer(gpuConstraints->getBufferCL());
+						launcher.setBuffer(gpuBodies->getBufferCL());
+						launcher.setConst(numConstraints);
+						launcher.launch1D(numConstraints);
+						clFinish(m_gpuData->m_queue);
+					}
+					//assume the batching happens on CPU, so copy the data
+					m_gpuData->m_gpuBatchConstraints->copyToHost(batchConstraints);
+				}
+			} 
+			else
+			{
+				totalNumRows  = 0;
+				gpuConstraints->copyToHost(m_gpuData->m_cpuConstraints);
+				//calculate the total number of contraint rows
+				for (int i=0;i<numConstraints;i++)
+				{
+					unsigned int& info1= m_tmpConstraintSizesPool[i];
+					//					unsigned int info1;
+					if (m_gpuData->m_cpuConstraints[i].isEnabled())
+					{
+
+						m_gpuData->m_cpuConstraints[i].getInfo1(&info1,&m_gpuData->m_cpuBodies[0]);
+					} else
+					{
+						info1 = 0;
+					}
+					
+					totalNumRows += info1;
+				}
+
+				m_gpuData->m_gpuBatchConstraints->copyFromHost(batchConstraints);
+				m_gpuData->m_gpuConstraintInfo1->copyFromHost(m_tmpConstraintSizesPool);
+
+			}
+			m_tmpSolverNonContactConstraintPool.resizeNoInitialize(totalNumRows);
+			m_gpuData->m_gpuConstraintRows->resize(totalNumRows);
+			
+			//			b3GpuConstraintArray		verify;
+
+			if (useGpuInfo2)
+			{
+				{
+						B3_PROFILE("getInfo2Kernel");
+						b3LauncherCL launcher(m_gpuData->m_queue,m_gpuData->m_getInfo2Kernel,"m_getInfo2Kernel");
+						launcher.setBuffer(m_gpuData->m_gpuConstraintRows->getBufferCL());
+						launcher.setBuffer(m_gpuData->m_gpuConstraintInfo1->getBufferCL());
+						launcher.setBuffer(m_gpuData->m_gpuConstraintRowOffsets->getBufferCL());
+						launcher.setBuffer(gpuConstraints->getBufferCL());
+						launcher.setBuffer(m_gpuData->m_gpuBatchConstraints->getBufferCL());
+						launcher.setBuffer(gpuBodies->getBufferCL());
+						launcher.setBuffer(gpuInertias->getBufferCL());
+						launcher.setBuffer(m_gpuData->m_gpuSolverBodies->getBufferCL());
+						launcher.setConst(infoGlobal.m_timeStep);
+						launcher.setConst(infoGlobal.m_erp);
+						launcher.setConst(infoGlobal.m_globalCfm);
+						launcher.setConst(infoGlobal.m_damping);
+						launcher.setConst(infoGlobal.m_numIterations);
+						launcher.setConst(numConstraints);
+						launcher.launch1D(numConstraints);
+						clFinish(m_gpuData->m_queue);
+
+						if (m_gpuData->m_batchSizes.size()==0)
+							m_gpuData->m_gpuBatchConstraints->copyToHost(batchConstraints);
+						//m_gpuData->m_gpuConstraintRows->copyToHost(verify);
+						//m_gpuData->m_gpuConstraintRows->copyToHost(m_tmpSolverNonContactConstraintPool);
+
+						
+
+					}
+			} 
+			else
+			{
+			
+				gpuInertias->copyToHost(m_gpuData->m_cpuInertias);
+
+					///setup the b3SolverConstraints
+			
+				for (int i=0;i<numConstraints;i++)
+				{
+					const int& info1 = m_tmpConstraintSizesPool[i];
+				
+					if (info1)
+					{
+						int constraintIndex = batchConstraints[i].m_originalConstraintIndex;
+						int constraintRowOffset = m_gpuData->m_cpuConstraintRowOffsets[constraintIndex];
+
+						b3GpuSolverConstraint* currentConstraintRow = &m_tmpSolverNonContactConstraintPool[constraintRowOffset];
+						b3GpuGenericConstraint& constraint = m_gpuData->m_cpuConstraints[i];
+
+						b3RigidBodyData& rbA = m_gpuData->m_cpuBodies[ constraint.getRigidBodyA()];
+						//b3RigidBody& rbA = constraint.getRigidBodyA();
+		//				b3RigidBody& rbB = constraint.getRigidBodyB();
+						b3RigidBodyData& rbB = m_gpuData->m_cpuBodies[ constraint.getRigidBodyB()];
+					
+					
+
+						int solverBodyIdA = constraint.getRigidBodyA();//getOrInitSolverBody(constraint.getRigidBodyA(),bodies,inertias);
+						int solverBodyIdB = constraint.getRigidBodyB();//getOrInitSolverBody(constraint.getRigidBodyB(),bodies,inertias);
+
+						b3GpuSolverBody* bodyAPtr = &m_tmpSolverBodyPool[solverBodyIdA];
+						b3GpuSolverBody* bodyBPtr = &m_tmpSolverBodyPool[solverBodyIdB];
+
+						if (rbA.m_invMass)
+						{
+							batchConstraints[i].m_bodyAPtrAndSignBit = solverBodyIdA;
+						} else
+						{
+							if (!solverBodyIdA)
+								m_staticIdx = 0;
+							batchConstraints[i].m_bodyAPtrAndSignBit = -solverBodyIdA;
+						}
+
+						if (rbB.m_invMass)
+						{
+							batchConstraints[i].m_bodyBPtrAndSignBit = solverBodyIdB;
+						} else
+						{
+							if (!solverBodyIdB)
+								m_staticIdx = 0;
+							batchConstraints[i].m_bodyBPtrAndSignBit = -solverBodyIdB;
+						}
+
+
+						int overrideNumSolverIterations = 0;//constraint->getOverrideNumSolverIterations() > 0 ? constraint->getOverrideNumSolverIterations() : infoGlobal.m_numIterations;
+						if (overrideNumSolverIterations>m_maxOverrideNumSolverIterations)
+							m_maxOverrideNumSolverIterations = overrideNumSolverIterations;
+
+
+						int j;
+						for ( j=0;j<info1;j++)
+						{
+							memset(&currentConstraintRow[j],0,sizeof(b3GpuSolverConstraint));
+							currentConstraintRow[j].m_angularComponentA.setValue(0,0,0);
+							currentConstraintRow[j].m_angularComponentB.setValue(0,0,0);
+							currentConstraintRow[j].m_appliedImpulse = 0.f;
+							currentConstraintRow[j].m_appliedPushImpulse = 0.f;
+							currentConstraintRow[j].m_cfm = 0.f;
+							currentConstraintRow[j].m_contactNormal.setValue(0,0,0);
+							currentConstraintRow[j].m_friction = 0.f;
+							currentConstraintRow[j].m_frictionIndex = 0;
+							currentConstraintRow[j].m_jacDiagABInv = 0.f;
+							currentConstraintRow[j].m_lowerLimit = 0.f;
+							currentConstraintRow[j].m_upperLimit = 0.f;
+
+							currentConstraintRow[j].m_originalContactPoint = 0;
+							currentConstraintRow[j].m_overrideNumSolverIterations = 0;
+							currentConstraintRow[j].m_relpos1CrossNormal.setValue(0,0,0);
+							currentConstraintRow[j].m_relpos2CrossNormal.setValue(0,0,0);
+							currentConstraintRow[j].m_rhs = 0.f;
+							currentConstraintRow[j].m_rhsPenetration = 0.f;
+							currentConstraintRow[j].m_solverBodyIdA = 0;
+							currentConstraintRow[j].m_solverBodyIdB = 0;
+							
+							currentConstraintRow[j].m_lowerLimit = -B3_INFINITY;
+							currentConstraintRow[j].m_upperLimit = B3_INFINITY;
+							currentConstraintRow[j].m_appliedImpulse = 0.f;
+							currentConstraintRow[j].m_appliedPushImpulse = 0.f;
+							currentConstraintRow[j].m_solverBodyIdA = solverBodyIdA;
+							currentConstraintRow[j].m_solverBodyIdB = solverBodyIdB;
+							currentConstraintRow[j].m_overrideNumSolverIterations = overrideNumSolverIterations;
+						}
+
+						bodyAPtr->internalGetDeltaLinearVelocity().setValue(0.f,0.f,0.f);
+						bodyAPtr->internalGetDeltaAngularVelocity().setValue(0.f,0.f,0.f);
+						bodyAPtr->internalGetPushVelocity().setValue(0.f,0.f,0.f);
+						bodyAPtr->internalGetTurnVelocity().setValue(0.f,0.f,0.f);
+						bodyBPtr->internalGetDeltaLinearVelocity().setValue(0.f,0.f,0.f);
+						bodyBPtr->internalGetDeltaAngularVelocity().setValue(0.f,0.f,0.f);
+						bodyBPtr->internalGetPushVelocity().setValue(0.f,0.f,0.f);
+						bodyBPtr->internalGetTurnVelocity().setValue(0.f,0.f,0.f);
+
+
+						b3GpuConstraintInfo2 info2;
+						info2.fps = 1.f/infoGlobal.m_timeStep;
+						info2.erp = infoGlobal.m_erp;
+						info2.m_J1linearAxis = currentConstraintRow->m_contactNormal;
+						info2.m_J1angularAxis = currentConstraintRow->m_relpos1CrossNormal;
+						info2.m_J2linearAxis = 0;
+						info2.m_J2angularAxis = currentConstraintRow->m_relpos2CrossNormal;
+						info2.rowskip = sizeof(b3GpuSolverConstraint)/sizeof(b3Scalar);//check this
+						///the size of b3GpuSolverConstraint needs be a multiple of b3Scalar
+						b3Assert(info2.rowskip*sizeof(b3Scalar)== sizeof(b3GpuSolverConstraint));
+						info2.m_constraintError = &currentConstraintRow->m_rhs;
+						currentConstraintRow->m_cfm = infoGlobal.m_globalCfm;
+						info2.m_damping = infoGlobal.m_damping;
+						info2.cfm = &currentConstraintRow->m_cfm;
+						info2.m_lowerLimit = &currentConstraintRow->m_lowerLimit;
+						info2.m_upperLimit = &currentConstraintRow->m_upperLimit;
+						info2.m_numIterations = infoGlobal.m_numIterations;
+						m_gpuData->m_cpuConstraints[i].getInfo2(&info2,&m_gpuData->m_cpuBodies[0]);
+
+						///finalize the constraint setup
+						for ( j=0;j<info1;j++)
+						{
+							b3GpuSolverConstraint& solverConstraint = currentConstraintRow[j];
+
+							if (solverConstraint.m_upperLimit>=m_gpuData->m_cpuConstraints[i].getBreakingImpulseThreshold())
+							{
+								solverConstraint.m_upperLimit = m_gpuData->m_cpuConstraints[i].getBreakingImpulseThreshold();
+							}
+
+							if (solverConstraint.m_lowerLimit<=-m_gpuData->m_cpuConstraints[i].getBreakingImpulseThreshold())
+							{
+								solverConstraint.m_lowerLimit = -m_gpuData->m_cpuConstraints[i].getBreakingImpulseThreshold();
+							}
+
+	//						solverConstraint.m_originalContactPoint = constraint;
+							
+							b3Matrix3x3& invInertiaWorldA= m_gpuData->m_cpuInertias[constraint.getRigidBodyA()].m_invInertiaWorld;
+							{
+
+								//b3Vector3 angularFactorA(1,1,1);
+								const b3Vector3& ftorqueAxis1 = solverConstraint.m_relpos1CrossNormal;
+								solverConstraint.m_angularComponentA = invInertiaWorldA*ftorqueAxis1;//*angularFactorA;
+							}
+						
+							b3Matrix3x3& invInertiaWorldB= m_gpuData->m_cpuInertias[constraint.getRigidBodyB()].m_invInertiaWorld;
+							{
+
+								const b3Vector3& ftorqueAxis2 = solverConstraint.m_relpos2CrossNormal;
+								solverConstraint.m_angularComponentB = invInertiaWorldB*ftorqueAxis2;//*constraint.getRigidBodyB().getAngularFactor();
+							}
+
+							{
+								//it is ok to use solverConstraint.m_contactNormal instead of -solverConstraint.m_contactNormal
+								//because it gets multiplied iMJlB
+								b3Vector3 iMJlA = solverConstraint.m_contactNormal*rbA.m_invMass;
+								b3Vector3 iMJaA = invInertiaWorldA*solverConstraint.m_relpos1CrossNormal;
+								b3Vector3 iMJlB = solverConstraint.m_contactNormal*rbB.m_invMass;//sign of normal?
+								b3Vector3 iMJaB = invInertiaWorldB*solverConstraint.m_relpos2CrossNormal;
+
+								b3Scalar sum = iMJlA.dot(solverConstraint.m_contactNormal);
+								sum += iMJaA.dot(solverConstraint.m_relpos1CrossNormal);
+								sum += iMJlB.dot(solverConstraint.m_contactNormal);
+								sum += iMJaB.dot(solverConstraint.m_relpos2CrossNormal);
+								b3Scalar fsum = b3Fabs(sum);
+								b3Assert(fsum > B3_EPSILON);
+								solverConstraint.m_jacDiagABInv = fsum>B3_EPSILON?b3Scalar(1.)/sum : 0.f;
+							}
+
+
+							///fix rhs
+							///todo: add force/torque accelerators
+							{
+								b3Scalar rel_vel;
+								b3Scalar vel1Dotn = solverConstraint.m_contactNormal.dot(rbA.m_linVel) + solverConstraint.m_relpos1CrossNormal.dot(rbA.m_angVel);
+								b3Scalar vel2Dotn = -solverConstraint.m_contactNormal.dot(rbB.m_linVel) + solverConstraint.m_relpos2CrossNormal.dot(rbB.m_angVel);
+
+								rel_vel = vel1Dotn+vel2Dotn;
+
+								b3Scalar restitution = 0.f;
+								b3Scalar positionalError = solverConstraint.m_rhs;//already filled in by getConstraintInfo2
+								b3Scalar	velocityError = restitution - rel_vel * info2.m_damping;
+								b3Scalar	penetrationImpulse = positionalError*solverConstraint.m_jacDiagABInv;
+								b3Scalar	velocityImpulse = velocityError *solverConstraint.m_jacDiagABInv;
+								solverConstraint.m_rhs = penetrationImpulse+velocityImpulse;
+								solverConstraint.m_appliedImpulse = 0.f;
+
+							}
+						}
+
+					}
+				}
+
+
+
+				m_gpuData->m_gpuConstraintRows->copyFromHost(m_tmpSolverNonContactConstraintPool);
+				m_gpuData->m_gpuConstraintInfo1->copyFromHost(m_tmpConstraintSizesPool);
+
+				if (m_gpuData->m_batchSizes.size()==0)
+					m_gpuData->m_gpuBatchConstraints->copyFromHost(batchConstraints);
+				else
+					m_gpuData->m_gpuBatchConstraints->copyToHost(batchConstraints);
+
+				m_gpuData->m_gpuSolverBodies->copyFromHost(m_tmpSolverBodyPool);
+
+
+
+			}//end useGpuInfo2
+
+
+		}
+
+#ifdef B3_SUPPORT_CONTACT_CONSTRAINTS
+		{
+			int i;
+
+			for (i=0;i<numManifolds;i++)
+			{
+				b3Contact4& manifold = manifoldPtr[i];
+				convertContact(bodies,inertias,&manifold,infoGlobal);
+			}
+		}
+#endif //B3_SUPPORT_CONTACT_CONSTRAINTS
+	}
+
+//	b3ContactSolverInfo info = infoGlobal;
+
+
+//	int numNonContactPool = m_tmpSolverNonContactConstraintPool.size();
+//	int numConstraintPool = m_tmpSolverContactConstraintPool.size();
+//	int numFrictionPool = m_tmpSolverContactFrictionConstraintPool.size();
+
+
+	return 0.f;
+
+}
+
+
+
+///a straight copy from GPU/OpenCL kernel, for debugging
+__inline void internalApplyImpulse( b3GpuSolverBody* body,  const b3Vector3& linearComponent, const b3Vector3& angularComponent,float impulseMagnitude)
+{
+	body->m_deltaLinearVelocity += linearComponent*impulseMagnitude*body->m_linearFactor;
+	body->m_deltaAngularVelocity += angularComponent*(impulseMagnitude*body->m_angularFactor);
+}
+
+
+void resolveSingleConstraintRowGeneric2( b3GpuSolverBody* body1,  b3GpuSolverBody* body2,  b3GpuSolverConstraint* c)
+{
+	float deltaImpulse = c->m_rhs-b3Scalar(c->m_appliedImpulse)*c->m_cfm;
+	float deltaVel1Dotn	=	b3Dot(c->m_contactNormal,body1->m_deltaLinearVelocity) 	+ b3Dot(c->m_relpos1CrossNormal,body1->m_deltaAngularVelocity);
+	float deltaVel2Dotn	=	-b3Dot(c->m_contactNormal,body2->m_deltaLinearVelocity) + b3Dot(c->m_relpos2CrossNormal,body2->m_deltaAngularVelocity);
+
+	deltaImpulse	-=	deltaVel1Dotn*c->m_jacDiagABInv;
+	deltaImpulse	-=	deltaVel2Dotn*c->m_jacDiagABInv;
+
+	float sum = b3Scalar(c->m_appliedImpulse) + deltaImpulse;
+	if (sum < c->m_lowerLimit)
+	{
+		deltaImpulse = c->m_lowerLimit-b3Scalar(c->m_appliedImpulse);
+		c->m_appliedImpulse = c->m_lowerLimit;
+	}
+	else if (sum > c->m_upperLimit) 
+	{
+		deltaImpulse = c->m_upperLimit-b3Scalar(c->m_appliedImpulse);
+		c->m_appliedImpulse = c->m_upperLimit;
+	}
+	else
+	{
+		c->m_appliedImpulse = sum;
+	}
+
+	internalApplyImpulse(body1,c->m_contactNormal*body1->m_invMass,c->m_angularComponentA,deltaImpulse);
+	internalApplyImpulse(body2,-c->m_contactNormal*body2->m_invMass,c->m_angularComponentB,deltaImpulse);
+
+}
+
+
+
+void	b3GpuPgsConstraintSolver::initSolverBody(int bodyIndex, b3GpuSolverBody* solverBody, b3RigidBodyData* rb)
+{
+
+	solverBody->m_deltaLinearVelocity.setValue(0.f,0.f,0.f);
+	solverBody->m_deltaAngularVelocity.setValue(0.f,0.f,0.f);
+	solverBody->internalGetPushVelocity().setValue(0.f,0.f,0.f);
+	solverBody->internalGetTurnVelocity().setValue(0.f,0.f,0.f);
+
+	b3Assert(rb);
+//	solverBody->m_worldTransform = getWorldTransform(rb);
+	solverBody->internalSetInvMass(b3MakeVector3(rb->m_invMass,rb->m_invMass,rb->m_invMass));
+	solverBody->m_originalBodyIndex = bodyIndex;
+	solverBody->m_angularFactor = b3MakeVector3(1,1,1);
+	solverBody->m_linearFactor = b3MakeVector3(1,1,1);
+	solverBody->m_linearVelocity = getLinearVelocity(rb);
+	solverBody->m_angularVelocity = getAngularVelocity(rb);
+}
+
+
+void	b3GpuPgsConstraintSolver::averageVelocities()
+{
+}
+
+
+b3Scalar b3GpuPgsConstraintSolver::solveGroupCacheFriendlyIterations(b3OpenCLArray<b3GpuGenericConstraint>* gpuConstraints1,int numConstraints,const b3ContactSolverInfo& infoGlobal)
+{
+	//only create the batches once.
+	//@todo: incrementally update batches when constraints are added/activated and/or removed/deactivated
+	B3_PROFILE("GpuSolveGroupCacheFriendlyIterations");
+
+	bool createBatches = m_gpuData->m_batchSizes.size()==0;
+	{
+		
+		if (createBatches)
+		{
+			
+			m_gpuData->m_batchSizes.resize(0);
+
+			{
+				m_gpuData->m_gpuBatchConstraints->copyToHost(batchConstraints);
+
+				B3_PROFILE("batch joints");
+				b3Assert(batchConstraints.size()==numConstraints);
+				int simdWidth =numConstraints+1;
+				int numBodies = m_tmpSolverBodyPool.size();
+				sortConstraintByBatch3( &batchConstraints[0], numConstraints, simdWidth , m_staticIdx,  numBodies);
+
+				m_gpuData->m_gpuBatchConstraints->copyFromHost(batchConstraints);
+
+			}
+		} else
+		{
+			/*b3AlignedObjectArray<b3BatchConstraint> cpuCheckBatches;
+			m_gpuData->m_gpuBatchConstraints->copyToHost(cpuCheckBatches);
+			b3Assert(cpuCheckBatches.size()==batchConstraints.size());
+			printf(".\n");
+			*/
+			//>copyFromHost(batchConstraints);
+		}
+		int maxIterations = infoGlobal.m_numIterations;
+	
+		bool useBatching = true;
+
+		if (useBatching )
+		{
+			
+			if (!useGpuSolveJointConstraintRows)
+			{
+				B3_PROFILE("copy to host");
+				m_gpuData->m_gpuSolverBodies->copyToHost(m_tmpSolverBodyPool);
+				m_gpuData->m_gpuBatchConstraints->copyToHost(batchConstraints);
+				m_gpuData->m_gpuConstraintRows->copyToHost(m_tmpSolverNonContactConstraintPool);
+				m_gpuData->m_gpuConstraintInfo1->copyToHost(m_gpuData->m_cpuConstraintInfo1);
+				m_gpuData->m_gpuConstraintRowOffsets->copyToHost(m_gpuData->m_cpuConstraintRowOffsets);
+				gpuConstraints1->copyToHost(m_gpuData->m_cpuConstraints);
+
+			}
+
+			for ( int iteration = 0 ; iteration< maxIterations ; iteration++)
+			{
+				
+				int batchOffset = 0;
+				int constraintOffset=0;
+				int numBatches = m_gpuData->m_batchSizes.size();
+				for (int bb=0;bb<numBatches;bb++)
+				{
+					int numConstraintsInBatch = m_gpuData->m_batchSizes[bb];
+
+					
+					if (useGpuSolveJointConstraintRows)
+					{
+						B3_PROFILE("solveJointConstraintRowsKernels");
+					
+						/*
+						__kernel void solveJointConstraintRows(__global b3GpuSolverBody* solverBodies,
+					  __global b3BatchConstraint* batchConstraints,
+					  	__global b3SolverConstraint* rows,
+						__global unsigned int* numConstraintRowsInfo1, 
+						__global unsigned int* rowOffsets,
+						__global b3GpuGenericConstraint* constraints,
+						int batchOffset,
+						int numConstraintsInBatch*/
+
+
+						b3LauncherCL launcher(m_gpuData->m_queue,m_gpuData->m_solveJointConstraintRowsKernels,"m_solveJointConstraintRowsKernels");
+						launcher.setBuffer(m_gpuData->m_gpuSolverBodies->getBufferCL());
+						launcher.setBuffer(m_gpuData->m_gpuBatchConstraints->getBufferCL());
+						launcher.setBuffer(m_gpuData->m_gpuConstraintRows->getBufferCL());
+						launcher.setBuffer(m_gpuData->m_gpuConstraintInfo1->getBufferCL());
+						launcher.setBuffer(m_gpuData->m_gpuConstraintRowOffsets->getBufferCL());
+						launcher.setBuffer(gpuConstraints1->getBufferCL());//to detect disabled constraints
+						launcher.setConst(batchOffset);
+						launcher.setConst(numConstraintsInBatch);
+
+						launcher.launch1D(numConstraintsInBatch);
+						
+
+					} else//useGpu
+					{
+						
+
+						
+						for (int b=0;b<numConstraintsInBatch;b++)
+						{
+							const b3BatchConstraint& c = batchConstraints[batchOffset+b];
+							/*printf("-----------\n");
+							printf("bb=%d\n",bb);
+							printf("c.batchId = %d\n", c.m_batchId);
+							*/
+							b3Assert(c.m_batchId==bb);
+							b3GpuGenericConstraint* constraint = &m_gpuData->m_cpuConstraints[c.m_originalConstraintIndex];
+							if (constraint->m_flags&B3_CONSTRAINT_FLAG_ENABLED)
+							{
+								int numConstraintRows = m_gpuData->m_cpuConstraintInfo1[c.m_originalConstraintIndex];
+								int constraintOffset = m_gpuData->m_cpuConstraintRowOffsets[c.m_originalConstraintIndex];
+								
+								for (int jj=0;jj<numConstraintRows;jj++)
+								{
+	//							
+									b3GpuSolverConstraint& constraint = m_tmpSolverNonContactConstraintPool[constraintOffset+jj];
+									//resolveSingleConstraintRowGenericSIMD(m_tmpSolverBodyPool[constraint.m_solverBodyIdA],m_tmpSolverBodyPool[constraint.m_solverBodyIdB],constraint);
+									resolveSingleConstraintRowGeneric2(&m_tmpSolverBodyPool[constraint.m_solverBodyIdA],&m_tmpSolverBodyPool[constraint.m_solverBodyIdB],&constraint);
+								}
+							}
+						}
+					}//useGpu
+					batchOffset+=numConstraintsInBatch;
+					constraintOffset+=numConstraintsInBatch;
+				}
+			}//for (int iteration...
+
+			if (!useGpuSolveJointConstraintRows)
+			{
+				{
+					B3_PROFILE("copy from host");
+					m_gpuData->m_gpuSolverBodies->copyFromHost(m_tmpSolverBodyPool);
+					m_gpuData->m_gpuBatchConstraints->copyFromHost(batchConstraints);
+					m_gpuData->m_gpuConstraintRows->copyFromHost(m_tmpSolverNonContactConstraintPool);
+				}
+
+				//B3_PROFILE("copy to host");
+				//m_gpuData->m_gpuSolverBodies->copyToHost(m_tmpSolverBodyPool);
+			}
+			//int sz = sizeof(b3GpuSolverBody);
+			//printf("cpu sizeof(b3GpuSolverBody)=%d\n",sz);
+
+			
+			
+
+
+		} else
+		{
+			for ( int iteration = 0 ; iteration< maxIterations ; iteration++)
+			{			
+				int numJoints =			m_tmpSolverNonContactConstraintPool.size();
+				for (int j=0;j<numJoints;j++)
+				{
+					b3GpuSolverConstraint& constraint = m_tmpSolverNonContactConstraintPool[j];
+					resolveSingleConstraintRowGeneric2(&m_tmpSolverBodyPool[constraint.m_solverBodyIdA],&m_tmpSolverBodyPool[constraint.m_solverBodyIdB],&constraint);
+				}
+
+				if (!m_usePgs)
+				{
+					averageVelocities();
+				}
+			}
+		}
+		
+	}
+	clFinish(m_gpuData->m_queue);
+	return 0.f;
+}
+
+
+
+
+static b3AlignedObjectArray<int> bodyUsed;
+static b3AlignedObjectArray<int> curUsed;
+
+
+
+inline int b3GpuPgsConstraintSolver::sortConstraintByBatch3( b3BatchConstraint* cs, int numConstraints, int simdWidth , int staticIdx, int numBodies)
+{
+	//int sz = sizeof(b3BatchConstraint);
+
+	B3_PROFILE("sortConstraintByBatch3");
+	
+	static int maxSwaps = 0;
+	int numSwaps = 0;
+
+	curUsed.resize(2*simdWidth);
+
+	static int maxNumConstraints = 0;
+	if (maxNumConstraints<numConstraints)
+	{
+		maxNumConstraints = numConstraints;
+		//printf("maxNumConstraints  = %d\n",maxNumConstraints );
+	}
+
+	int numUsedArray = numBodies/32+1;
+	bodyUsed.resize(numUsedArray);
+
+	for (int q=0;q<numUsedArray;q++)
+		bodyUsed[q]=0;
+
+	
+	int curBodyUsed = 0;
+
+	int numIter = 0;
+    
+		
+#if defined(_DEBUG)
+	for(int i=0; i<numConstraints; i++)
+		cs[i].m_batchId = -1;
+#endif
+	
+	int numValidConstraints = 0;
+//	int unprocessedConstraintIndex = 0;
+
+	int batchIdx = 0;
+    
+
+	{
+		B3_PROFILE("cpu batch innerloop");
+		
+		while( numValidConstraints < numConstraints)
+		{
+			numIter++;
+			int nCurrentBatch = 0;
+			//	clear flag
+			for(int i=0; i<curBodyUsed; i++) 
+				bodyUsed[curUsed[i]/32] = 0;
+
+            curBodyUsed = 0;
+
+			for(int i=numValidConstraints; i<numConstraints; i++)
+			{
+				int idx = i;
+				b3Assert( idx < numConstraints );
+				//	check if it can go
+				int bodyAS = cs[idx].m_bodyAPtrAndSignBit;
+				int bodyBS = cs[idx].m_bodyBPtrAndSignBit;
+				int bodyA = abs(bodyAS);
+				int bodyB = abs(bodyBS);
+				bool aIsStatic = (bodyAS<0) || bodyAS==staticIdx;
+				bool bIsStatic = (bodyBS<0) || bodyBS==staticIdx;
+				int aUnavailable = 0;
+				int bUnavailable = 0;
+				if (!aIsStatic)
+				{
+					aUnavailable = bodyUsed[ bodyA/32 ] & (1<<(bodyA&31));
+				}
+				if (!aUnavailable)
+				if (!bIsStatic)
+				{
+					bUnavailable = bodyUsed[ bodyB/32 ] & (1<<(bodyB&31));
+				}
+                
+				if( aUnavailable==0 && bUnavailable==0 ) // ok
+				{
+					if (!aIsStatic)
+					{
+						bodyUsed[ bodyA/32 ] |= (1<<(bodyA&31));
+						curUsed[curBodyUsed++]=bodyA;
+					}
+					if (!bIsStatic)
+					{
+						bodyUsed[ bodyB/32 ] |= (1<<(bodyB&31));
+						curUsed[curBodyUsed++]=bodyB;
+					}
+
+					cs[idx].m_batchId = batchIdx;
+
+					if (i!=numValidConstraints)
+					{
+						b3Swap(cs[i],cs[numValidConstraints]);
+						numSwaps++;
+					}
+
+					numValidConstraints++;
+					{
+						nCurrentBatch++;
+						if( nCurrentBatch == simdWidth )
+						{
+							nCurrentBatch = 0;
+							for(int i=0; i<curBodyUsed; i++) 
+								bodyUsed[curUsed[i]/32] = 0;
+							curBodyUsed = 0;
+						}
+					}
+				}
+			}
+			m_gpuData->m_batchSizes.push_back(nCurrentBatch);
+			batchIdx ++;
+		}
+	}
+	
+#if defined(_DEBUG)
+    //		debugPrintf( "nBatches: %d\n", batchIdx );
+	for(int i=0; i<numConstraints; i++)
+    {
+        b3Assert( cs[i].m_batchId != -1 );
+    }
+#endif
+
+	if (maxSwaps<numSwaps)
+	{
+		maxSwaps = numSwaps;
+		//printf("maxSwaps = %d\n", maxSwaps);
+	}
+	
+	return batchIdx;
+}
+
+
+/// b3PgsJacobiSolver Sequentially applies impulses
+b3Scalar b3GpuPgsConstraintSolver::solveGroup(b3OpenCLArray<b3RigidBodyData>* gpuBodies, b3OpenCLArray<b3InertiaData>* gpuInertias, 
+				int numBodies, b3OpenCLArray<b3GpuGenericConstraint>* gpuConstraints,int numConstraints, const b3ContactSolverInfo& infoGlobal)
+{
+
+	B3_PROFILE("solveJoints");
+	//you need to provide at least some bodies
+	
+	solveGroupCacheFriendlySetup( gpuBodies, gpuInertias,numBodies,gpuConstraints, numConstraints,infoGlobal);
+
+	solveGroupCacheFriendlyIterations(gpuConstraints, numConstraints,infoGlobal);
+
+	solveGroupCacheFriendlyFinish(gpuBodies, gpuInertias,numBodies, gpuConstraints, numConstraints, infoGlobal);
+	
+	return 0.f;
+}
+
+void	b3GpuPgsConstraintSolver::solveJoints(int numBodies, b3OpenCLArray<b3RigidBodyData>* gpuBodies, b3OpenCLArray<b3InertiaData>* gpuInertias, 
+				int numConstraints, b3OpenCLArray<b3GpuGenericConstraint>* gpuConstraints)
+{
+	b3ContactSolverInfo infoGlobal;
+	infoGlobal.m_splitImpulse = false;
+	infoGlobal.m_timeStep = 1.f/60.f;
+	infoGlobal.m_numIterations = 4;//4;
+//	infoGlobal.m_solverMode|=B3_SOLVER_USE_2_FRICTION_DIRECTIONS|B3_SOLVER_INTERLEAVE_CONTACT_AND_FRICTION_CONSTRAINTS|B3_SOLVER_DISABLE_VELOCITY_DEPENDENT_FRICTION_DIRECTION;
+	//infoGlobal.m_solverMode|=B3_SOLVER_USE_2_FRICTION_DIRECTIONS|B3_SOLVER_INTERLEAVE_CONTACT_AND_FRICTION_CONSTRAINTS;
+	infoGlobal.m_solverMode|=B3_SOLVER_USE_2_FRICTION_DIRECTIONS;
+
+	//if (infoGlobal.m_solverMode & B3_SOLVER_INTERLEAVE_CONTACT_AND_FRICTION_CONSTRAINTS)
+	//if ((infoGlobal.m_solverMode & B3_SOLVER_USE_2_FRICTION_DIRECTIONS) && (infoGlobal.m_solverMode & B3_SOLVER_DISABLE_VELOCITY_DEPENDENT_FRICTION_DIRECTION))
+				
+
+	solveGroup(gpuBodies,gpuInertias,numBodies,gpuConstraints,numConstraints,infoGlobal);
+
+}
+
+//b3AlignedObjectArray<b3RigidBodyData> testBodies;
+
+
+b3Scalar b3GpuPgsConstraintSolver::solveGroupCacheFriendlyFinish(b3OpenCLArray<b3RigidBodyData>* gpuBodies,b3OpenCLArray<b3InertiaData>* gpuInertias,int numBodies,b3OpenCLArray<b3GpuGenericConstraint>* gpuConstraints,int numConstraints,const b3ContactSolverInfo& infoGlobal)
+{
+	B3_PROFILE("solveGroupCacheFriendlyFinish");
+//	int numPoolConstraints = m_tmpSolverContactConstraintPool.size();
+//	int i,j;
+
+
+	{
+		if (gpuBreakConstraints)
+		{
+			B3_PROFILE("breakViolatedConstraintsKernel");
+			b3LauncherCL launcher(m_gpuData->m_queue,m_gpuData->m_breakViolatedConstraintsKernel,"m_breakViolatedConstraintsKernel");
+			launcher.setBuffer(gpuConstraints->getBufferCL());
+			launcher.setBuffer(m_gpuData->m_gpuConstraintInfo1->getBufferCL());
+			launcher.setBuffer(m_gpuData->m_gpuConstraintRowOffsets->getBufferCL());
+			launcher.setBuffer(m_gpuData->m_gpuConstraintRows->getBufferCL());
+			launcher.setConst(numConstraints);
+			launcher.launch1D(numConstraints);
+		} else
+		{
+			gpuConstraints->copyToHost(m_gpuData->m_cpuConstraints);
+			m_gpuData->m_gpuBatchConstraints->copyToHost(m_gpuData->m_cpuBatchConstraints);
+			m_gpuData->m_gpuConstraintRows->copyToHost(m_gpuData->m_cpuConstraintRows);
+			gpuConstraints->copyToHost(m_gpuData->m_cpuConstraints);
+			m_gpuData->m_gpuConstraintInfo1->copyToHost(m_gpuData->m_cpuConstraintInfo1);
+			m_gpuData->m_gpuConstraintRowOffsets->copyToHost(m_gpuData->m_cpuConstraintRowOffsets);
+
+			for (int cid=0;cid<numConstraints;cid++)
+			{
+				int originalConstraintIndex = batchConstraints[cid].m_originalConstraintIndex;
+				int constraintRowOffset = m_gpuData->m_cpuConstraintRowOffsets[originalConstraintIndex];
+				int numRows = m_gpuData->m_cpuConstraintInfo1[originalConstraintIndex];
+				if (numRows)
+				{
+		
+				//	printf("cid=%d, breakingThreshold =%f\n",cid,breakingThreshold);
+					for (int i=0;i<numRows;i++)
+					{
+						int rowIndex =constraintRowOffset+i;
+						int orgConstraintIndex = m_gpuData->m_cpuConstraintRows[rowIndex].m_originalConstraintIndex;
+						float breakingThreshold = m_gpuData->m_cpuConstraints[orgConstraintIndex].m_breakingImpulseThreshold;
+					//	printf("rows[%d].m_appliedImpulse=%f\n",rowIndex,rows[rowIndex].m_appliedImpulse);
+						if (b3Fabs(m_gpuData->m_cpuConstraintRows[rowIndex].m_appliedImpulse) >= breakingThreshold)
+						{
+
+							m_gpuData->m_cpuConstraints[orgConstraintIndex].m_flags =0;//&= ~B3_CONSTRAINT_FLAG_ENABLED;
+						}
+					}
+				}
+			}
+
+
+			gpuConstraints->copyFromHost(m_gpuData->m_cpuConstraints);
+		}
+	}
+
+	{
+		if (useGpuWriteBackVelocities)
+		{
+			B3_PROFILE("GPU write back velocities and transforms");
+
+			b3LauncherCL launcher(m_gpuData->m_queue,m_gpuData->m_writeBackVelocitiesKernel,"m_writeBackVelocitiesKernel");
+			launcher.setBuffer(gpuBodies->getBufferCL());
+			launcher.setBuffer(m_gpuData->m_gpuSolverBodies->getBufferCL());
+			launcher.setConst(numBodies);
+			launcher.launch1D(numBodies);
+			clFinish(m_gpuData->m_queue);
+//			m_gpuData->m_gpuSolverBodies->copyToHost(m_tmpSolverBodyPool);
+//			m_gpuData->m_gpuBodies->copyToHostPointer(bodies,numBodies);
+			//m_gpuData->m_gpuBodies->copyToHost(testBodies);
+
+		} 
+		else
+		{
+			B3_PROFILE("CPU write back velocities and transforms");
+
+			m_gpuData->m_gpuSolverBodies->copyToHost(m_tmpSolverBodyPool);
+			gpuBodies->copyToHost(m_gpuData->m_cpuBodies);
+			for ( int i=0;i<m_tmpSolverBodyPool.size();i++)
+			{
+				int bodyIndex = m_tmpSolverBodyPool[i].m_originalBodyIndex;
+				//printf("bodyIndex=%d\n",bodyIndex);
+				b3Assert(i==bodyIndex);
+
+				b3RigidBodyData* body = &m_gpuData->m_cpuBodies[bodyIndex];
+				if (body->m_invMass)
+				{
+					if (infoGlobal.m_splitImpulse)
+						m_tmpSolverBodyPool[i].writebackVelocityAndTransform(infoGlobal.m_timeStep, infoGlobal.m_splitImpulseTurnErp);
+					else
+						m_tmpSolverBodyPool[i].writebackVelocity();
+
+					if (m_usePgs)
+					{
+						body->m_linVel = m_tmpSolverBodyPool[i].m_linearVelocity;
+						body->m_angVel = m_tmpSolverBodyPool[i].m_angularVelocity;
+					} else
+					{
+						b3Assert(0);
+					}
+	/*			
+					if (infoGlobal.m_splitImpulse)
+					{
+						body->m_pos = m_tmpSolverBodyPool[i].m_worldTransform.getOrigin();
+						b3Quaternion orn;
+						orn = m_tmpSolverBodyPool[i].m_worldTransform.getRotation();
+						body->m_quat = orn;
+					}
+					*/
+				}
+			}//for
+
+			gpuBodies->copyFromHost(m_gpuData->m_cpuBodies);
+
+		}
+	}
+
+	clFinish(m_gpuData->m_queue);
+
+	m_tmpSolverContactConstraintPool.resizeNoInitialize(0);
+	m_tmpSolverNonContactConstraintPool.resizeNoInitialize(0);
+	m_tmpSolverContactFrictionConstraintPool.resizeNoInitialize(0);
+	m_tmpSolverContactRollingFrictionConstraintPool.resizeNoInitialize(0);
+
+	m_tmpSolverBodyPool.resizeNoInitialize(0);
+	return 0.f;
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuPgsConstraintSolver.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuPgsConstraintSolver.h
new file mode 100644
index 0000000000..ec0e3f73d6
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuPgsConstraintSolver.h
@@ -0,0 +1,78 @@
+/*
+Copyright (c) 2013 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+#ifndef B3_GPU_PGS_CONSTRAINT_SOLVER_H
+#define B3_GPU_PGS_CONSTRAINT_SOLVER_H
+
+struct b3Contact4;
+struct b3ContactPoint;
+
+
+class b3Dispatcher;
+
+#include "Bullet3Dynamics/ConstraintSolver/b3TypedConstraint.h"
+#include "Bullet3Dynamics/ConstraintSolver/b3ContactSolverInfo.h"
+#include "b3GpuSolverBody.h"
+#include "b3GpuSolverConstraint.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
+struct b3RigidBodyData;
+struct b3InertiaData;
+
+#include "Bullet3OpenCL/Initialize/b3OpenCLInclude.h"
+#include "b3GpuGenericConstraint.h"
+
+class b3GpuPgsConstraintSolver
+{
+protected:
+	int m_staticIdx;
+	struct b3GpuPgsJacobiSolverInternalData* m_gpuData;
+	protected:
+	b3AlignedObjectArray<b3GpuSolverBody>      m_tmpSolverBodyPool;
+	b3GpuConstraintArray			m_tmpSolverContactConstraintPool;
+	b3GpuConstraintArray			m_tmpSolverNonContactConstraintPool;
+	b3GpuConstraintArray			m_tmpSolverContactFrictionConstraintPool;
+	b3GpuConstraintArray			m_tmpSolverContactRollingFrictionConstraintPool;
+
+	b3AlignedObjectArray<unsigned int> m_tmpConstraintSizesPool;
+	
+
+	bool						m_usePgs;
+	void						averageVelocities();
+
+	int							m_maxOverrideNumSolverIterations;
+
+	int							m_numSplitImpulseRecoveries;
+
+//	int	getOrInitSolverBody(int bodyIndex, b3RigidBodyData* bodies,b3InertiaData* inertias);
+	void	initSolverBody(int bodyIndex, b3GpuSolverBody* solverBody, b3RigidBodyData* rb);
+
+public:
+	b3GpuPgsConstraintSolver (cl_context ctx, cl_device_id device, cl_command_queue queue,bool usePgs);
+	virtual~b3GpuPgsConstraintSolver ();
+
+	virtual b3Scalar solveGroupCacheFriendlyIterations(b3OpenCLArray<b3GpuGenericConstraint>* gpuConstraints1,int numConstraints,const b3ContactSolverInfo& infoGlobal);
+	virtual b3Scalar solveGroupCacheFriendlySetup(b3OpenCLArray<b3RigidBodyData>* gpuBodies, b3OpenCLArray<b3InertiaData>* gpuInertias, int numBodies,b3OpenCLArray<b3GpuGenericConstraint>* gpuConstraints,int numConstraints,const b3ContactSolverInfo& infoGlobal);
+	b3Scalar solveGroupCacheFriendlyFinish(b3OpenCLArray<b3RigidBodyData>* gpuBodies,b3OpenCLArray<b3InertiaData>* gpuInertias,int numBodies,b3OpenCLArray<b3GpuGenericConstraint>* gpuConstraints,int numConstraints,const b3ContactSolverInfo& infoGlobal);
+
+
+	b3Scalar solveGroup(b3OpenCLArray<b3RigidBodyData>* gpuBodies,b3OpenCLArray<b3InertiaData>* gpuInertias, int numBodies,b3OpenCLArray<b3GpuGenericConstraint>* gpuConstraints,int numConstraints,const b3ContactSolverInfo& infoGlobal);
+	void	solveJoints(int numBodies, b3OpenCLArray<b3RigidBodyData>* gpuBodies, b3OpenCLArray<b3InertiaData>* gpuInertias, 
+				int numConstraints, b3OpenCLArray<b3GpuGenericConstraint>* gpuConstraints);
+
+	int sortConstraintByBatch3( struct b3BatchConstraint* cs, int numConstraints, int simdWidth , int staticIdx, int numBodies);
+	void	recomputeBatches();
+};
+
+#endif //B3_GPU_PGS_CONSTRAINT_SOLVER_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuPgsContactSolver.cpp b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuPgsContactSolver.cpp
new file mode 100644
index 0000000000..f0b0abd5e0
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuPgsContactSolver.cpp
@@ -0,0 +1,1708 @@
+
+bool gUseLargeBatches = false;
+bool gCpuBatchContacts = false;
+bool gCpuSolveConstraint = false;
+bool gCpuRadixSort=false;
+bool gCpuSetSortData = false;
+bool gCpuSortContactsDeterminism = false;
+bool gUseCpuCopyConstraints = false;
+bool gUseScanHost = false;
+bool gReorderContactsOnCpu = false;
+
+bool optionalSortContactsDeterminism = true;
+
+
+#include "b3GpuPgsContactSolver.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h"
+
+#include "Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3BoundSearchCL.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3PrefixScanCL.h"
+#include <string.h>
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3Config.h"
+#include "b3Solver.h"
+
+
+#define B3_SOLVER_SETUP_KERNEL_PATH "src/Bullet3OpenCL/RigidBody/kernels/solverSetup.cl"
+#define B3_SOLVER_SETUP2_KERNEL_PATH "src/Bullet3OpenCL/RigidBody/kernels/solverSetup2.cl"
+#define B3_SOLVER_CONTACT_KERNEL_PATH "src/Bullet3OpenCL/RigidBody/kernels/solveContact.cl"
+#define B3_SOLVER_FRICTION_KERNEL_PATH "src/Bullet3OpenCL/RigidBody/kernels/solveFriction.cl"
+#define B3_BATCHING_PATH "src/Bullet3OpenCL/RigidBody/kernels/batchingKernels.cl"
+#define B3_BATCHING_NEW_PATH "src/Bullet3OpenCL/RigidBody/kernels/batchingKernelsNew.cl"
+
+#include "kernels/solverSetup.h"
+#include "kernels/solverSetup2.h"
+#include "kernels/solveContact.h"
+#include "kernels/solveFriction.h"
+#include "kernels/batchingKernels.h"
+#include "kernels/batchingKernelsNew.h"
+
+
+
+
+
+struct	b3GpuBatchingPgsSolverInternalData
+{
+	cl_context m_context;
+	cl_device_id m_device;
+	cl_command_queue m_queue;
+	int m_pairCapacity;
+	int m_nIterations;
+
+	b3OpenCLArray<b3GpuConstraint4>* m_contactCGPU;
+	b3OpenCLArray<unsigned int>* m_numConstraints;
+	b3OpenCLArray<unsigned int>* m_offsets;
+		
+	b3Solver*		m_solverGPU;		
+	
+	cl_kernel m_batchingKernel;
+	cl_kernel m_batchingKernelNew;
+	cl_kernel m_solveContactKernel;
+	cl_kernel m_solveSingleContactKernel;
+	cl_kernel m_solveSingleFrictionKernel;
+	cl_kernel m_solveFrictionKernel;
+	cl_kernel m_contactToConstraintKernel;
+	cl_kernel m_setSortDataKernel;
+	cl_kernel m_reorderContactKernel;
+	cl_kernel m_copyConstraintKernel;
+
+	cl_kernel	m_setDeterminismSortDataBodyAKernel;
+	cl_kernel	m_setDeterminismSortDataBodyBKernel;
+	cl_kernel	m_setDeterminismSortDataChildShapeAKernel;
+	cl_kernel	m_setDeterminismSortDataChildShapeBKernel;
+
+
+
+
+	class b3RadixSort32CL*	m_sort32;
+	class b3BoundSearchCL*	m_search;
+	class b3PrefixScanCL*	m_scan;
+
+	b3OpenCLArray<b3SortData>* m_sortDataBuffer;
+	b3OpenCLArray<b3Contact4>* m_contactBuffer;
+
+	b3OpenCLArray<b3RigidBodyData>* m_bodyBufferGPU;
+	b3OpenCLArray<b3InertiaData>* m_inertiaBufferGPU;
+	b3OpenCLArray<b3Contact4>* m_pBufContactOutGPU;
+	
+	b3OpenCLArray<b3Contact4>* m_pBufContactOutGPUCopy;
+	b3OpenCLArray<b3SortData>*	m_contactKeyValues;
+
+
+	b3AlignedObjectArray<unsigned int> m_idxBuffer;
+	b3AlignedObjectArray<b3SortData> m_sortData;
+	b3AlignedObjectArray<b3Contact4> m_old;
+
+	b3AlignedObjectArray<int>	m_batchSizes;
+	b3OpenCLArray<int>*	m_batchSizesGpu;
+
+};
+
+
+
+b3GpuPgsContactSolver::b3GpuPgsContactSolver(cl_context ctx,cl_device_id device, cl_command_queue  q,int pairCapacity)
+{
+	m_debugOutput=0;
+	m_data = new b3GpuBatchingPgsSolverInternalData;
+	m_data->m_context = ctx;
+	m_data->m_device = device;
+	m_data->m_queue = q;
+	m_data->m_pairCapacity = pairCapacity;
+	m_data->m_nIterations = 4;
+	m_data->m_batchSizesGpu = new b3OpenCLArray<int>(ctx,q);
+	m_data->m_bodyBufferGPU = new b3OpenCLArray<b3RigidBodyData>(ctx,q);
+	m_data->m_inertiaBufferGPU = new b3OpenCLArray<b3InertiaData>(ctx,q);
+	m_data->m_pBufContactOutGPU = new b3OpenCLArray<b3Contact4>(ctx,q);
+
+	m_data->m_pBufContactOutGPUCopy = new b3OpenCLArray<b3Contact4>(ctx,q);
+	m_data->m_contactKeyValues = new b3OpenCLArray<b3SortData>(ctx,q);
+
+
+	m_data->m_solverGPU = new b3Solver(ctx,device,q,512*1024);
+
+	m_data->m_sort32 = new b3RadixSort32CL(ctx,device,m_data->m_queue);
+	m_data->m_scan = new b3PrefixScanCL(ctx,device,m_data->m_queue,B3_SOLVER_N_CELLS);
+	m_data->m_search = new b3BoundSearchCL(ctx,device,m_data->m_queue,B3_SOLVER_N_CELLS);
+
+	const int sortSize = B3NEXTMULTIPLEOF( pairCapacity, 512 );
+
+	m_data->m_sortDataBuffer = new b3OpenCLArray<b3SortData>(ctx,m_data->m_queue,sortSize);
+	m_data->m_contactBuffer = new b3OpenCLArray<b3Contact4>(ctx,m_data->m_queue);
+
+	m_data->m_numConstraints = new b3OpenCLArray<unsigned int>(ctx,m_data->m_queue,B3_SOLVER_N_CELLS);
+	m_data->m_numConstraints->resize(B3_SOLVER_N_CELLS);
+
+	m_data->m_contactCGPU = new b3OpenCLArray<b3GpuConstraint4>(ctx,q,pairCapacity);
+
+	m_data->m_offsets = new b3OpenCLArray<unsigned int>( ctx,m_data->m_queue,B3_SOLVER_N_CELLS);
+	m_data->m_offsets->resize(B3_SOLVER_N_CELLS);
+	const char* additionalMacros = "";
+	//const char* srcFileNameForCaching="";
+
+
+
+	cl_int pErrNum;
+	const char* batchKernelSource = batchingKernelsCL;
+	const char* batchKernelNewSource = batchingKernelsNewCL;
+	const char* solverSetupSource = solverSetupCL;
+	const char* solverSetup2Source = solverSetup2CL;
+	const char* solveContactSource = solveContactCL;
+	const char* solveFrictionSource = solveFrictionCL;
+	
+		
+	{
+		
+		cl_program solveContactProg= b3OpenCLUtils::compileCLProgramFromString( ctx, device, solveContactSource, &pErrNum,additionalMacros, B3_SOLVER_CONTACT_KERNEL_PATH);
+		b3Assert(solveContactProg);
+		
+		cl_program solveFrictionProg= b3OpenCLUtils::compileCLProgramFromString( ctx, device, solveFrictionSource, &pErrNum,additionalMacros, B3_SOLVER_FRICTION_KERNEL_PATH);
+		b3Assert(solveFrictionProg);
+
+		cl_program solverSetup2Prog= b3OpenCLUtils::compileCLProgramFromString( ctx, device, solverSetup2Source, &pErrNum,additionalMacros, B3_SOLVER_SETUP2_KERNEL_PATH);
+		
+		
+		b3Assert(solverSetup2Prog);
+
+		
+		cl_program solverSetupProg= b3OpenCLUtils::compileCLProgramFromString( ctx, device, solverSetupSource, &pErrNum,additionalMacros, B3_SOLVER_SETUP_KERNEL_PATH);
+		b3Assert(solverSetupProg);
+		
+		
+		m_data->m_solveFrictionKernel= b3OpenCLUtils::compileCLKernelFromString( ctx, device, solveFrictionSource, "BatchSolveKernelFriction", &pErrNum, solveFrictionProg,additionalMacros );
+		b3Assert(m_data->m_solveFrictionKernel);
+
+		m_data->m_solveContactKernel= b3OpenCLUtils::compileCLKernelFromString( ctx, device, solveContactSource, "BatchSolveKernelContact", &pErrNum, solveContactProg,additionalMacros );
+		b3Assert(m_data->m_solveContactKernel);
+
+		m_data->m_solveSingleContactKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solveContactSource, "solveSingleContactKernel", &pErrNum, solveContactProg,additionalMacros );
+		b3Assert(m_data->m_solveSingleContactKernel);
+
+		m_data->m_solveSingleFrictionKernel =b3OpenCLUtils::compileCLKernelFromString( ctx, device, solveFrictionSource, "solveSingleFrictionKernel", &pErrNum, solveFrictionProg,additionalMacros );
+		b3Assert(m_data->m_solveSingleFrictionKernel);
+		
+		m_data->m_contactToConstraintKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverSetupSource, "ContactToConstraintKernel", &pErrNum, solverSetupProg,additionalMacros );
+		b3Assert(m_data->m_contactToConstraintKernel);
+			
+		m_data->m_setSortDataKernel =  b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverSetup2Source, "SetSortDataKernel", &pErrNum, solverSetup2Prog,additionalMacros );
+		b3Assert(m_data->m_setSortDataKernel);
+
+		m_data->m_setDeterminismSortDataBodyAKernel =  b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverSetup2Source, "SetDeterminismSortDataBodyA", &pErrNum, solverSetup2Prog,additionalMacros );
+		b3Assert(m_data->m_setDeterminismSortDataBodyAKernel);
+
+		m_data->m_setDeterminismSortDataBodyBKernel =  b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverSetup2Source, "SetDeterminismSortDataBodyB", &pErrNum, solverSetup2Prog,additionalMacros );
+		b3Assert(m_data->m_setDeterminismSortDataBodyBKernel);
+
+		m_data->m_setDeterminismSortDataChildShapeAKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverSetup2Source, "SetDeterminismSortDataChildShapeA", &pErrNum, solverSetup2Prog,additionalMacros );
+		b3Assert(m_data->m_setDeterminismSortDataChildShapeAKernel);
+
+		m_data->m_setDeterminismSortDataChildShapeBKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverSetup2Source, "SetDeterminismSortDataChildShapeB", &pErrNum, solverSetup2Prog,additionalMacros );
+		b3Assert(m_data->m_setDeterminismSortDataChildShapeBKernel);
+
+		
+		m_data->m_reorderContactKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverSetup2Source, "ReorderContactKernel", &pErrNum, solverSetup2Prog,additionalMacros );
+		b3Assert(m_data->m_reorderContactKernel);
+		
+
+		m_data->m_copyConstraintKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverSetup2Source, "CopyConstraintKernel", &pErrNum, solverSetup2Prog,additionalMacros );
+		b3Assert(m_data->m_copyConstraintKernel);
+		
+	}
+
+	{
+		cl_program batchingProg = b3OpenCLUtils::compileCLProgramFromString( ctx, device, batchKernelSource, &pErrNum,additionalMacros, B3_BATCHING_PATH);
+		b3Assert(batchingProg);
+		
+		m_data->m_batchingKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, batchKernelSource, "CreateBatches", &pErrNum, batchingProg,additionalMacros );
+		b3Assert(m_data->m_batchingKernel);
+	}
+			
+	{
+		cl_program batchingNewProg = b3OpenCLUtils::compileCLProgramFromString( ctx, device, batchKernelNewSource, &pErrNum,additionalMacros, B3_BATCHING_NEW_PATH);
+		b3Assert(batchingNewProg);
+		
+		m_data->m_batchingKernelNew = b3OpenCLUtils::compileCLKernelFromString( ctx, device, batchKernelNewSource, "CreateBatchesNew", &pErrNum, batchingNewProg,additionalMacros );
+		b3Assert(m_data->m_batchingKernelNew);
+	}
+		
+
+
+
+
+
+
+}
+
+b3GpuPgsContactSolver::~b3GpuPgsContactSolver()
+{
+	delete m_data->m_batchSizesGpu;
+	delete m_data->m_bodyBufferGPU;
+	delete m_data->m_inertiaBufferGPU;
+	delete m_data->m_pBufContactOutGPU;
+	delete m_data->m_pBufContactOutGPUCopy;
+	delete m_data->m_contactKeyValues;
+
+
+
+	delete m_data->m_contactCGPU;
+	delete m_data->m_numConstraints;
+	delete m_data->m_offsets;
+	delete m_data->m_sortDataBuffer;
+	delete m_data->m_contactBuffer;
+
+	delete m_data->m_sort32;
+	delete m_data->m_scan;
+	delete m_data->m_search;
+	delete m_data->m_solverGPU;
+
+	clReleaseKernel(m_data->m_batchingKernel);
+	clReleaseKernel(m_data->m_batchingKernelNew);
+	clReleaseKernel(m_data->m_solveSingleContactKernel);
+	clReleaseKernel(m_data->m_solveSingleFrictionKernel);
+	clReleaseKernel( m_data->m_solveContactKernel);
+	clReleaseKernel( m_data->m_solveFrictionKernel);
+
+	clReleaseKernel( m_data->m_contactToConstraintKernel);
+	clReleaseKernel( m_data->m_setSortDataKernel);
+	clReleaseKernel( m_data->m_reorderContactKernel);
+	clReleaseKernel( m_data->m_copyConstraintKernel);
+
+	clReleaseKernel(m_data->m_setDeterminismSortDataBodyAKernel);
+	clReleaseKernel(m_data->m_setDeterminismSortDataBodyBKernel);
+	clReleaseKernel(m_data->m_setDeterminismSortDataChildShapeAKernel);
+	clReleaseKernel(m_data->m_setDeterminismSortDataChildShapeBKernel);
+
+
+
+	delete m_data;
+}
+
+
+
+struct b3ConstraintCfg
+{
+	b3ConstraintCfg( float dt = 0.f ): m_positionDrift( 0.005f ), m_positionConstraintCoeff( 0.2f ), m_dt(dt), m_staticIdx(0) {}
+
+	float m_positionDrift;
+	float m_positionConstraintCoeff;
+	float m_dt;
+	bool m_enableParallelSolve;
+	float m_batchCellSize;
+	int m_staticIdx;
+};
+
+
+
+void b3GpuPgsContactSolver::solveContactConstraintBatchSizes(  const b3OpenCLArray<b3RigidBodyData>* bodyBuf, const b3OpenCLArray<b3InertiaData>* shapeBuf, 
+			b3OpenCLArray<b3GpuConstraint4>* constraint, void* additionalData, int n ,int maxNumBatches,int numIterations, const b3AlignedObjectArray<int>* batchSizes)//const b3OpenCLArray<int>* gpuBatchSizes)
+{
+	B3_PROFILE("solveContactConstraintBatchSizes");
+	int numBatches = batchSizes->size()/B3_MAX_NUM_BATCHES;
+	for(int iter=0; iter<numIterations; iter++)
+	{
+		
+		for (int cellId=0;cellId<numBatches;cellId++)
+		{
+			int offset = 0;
+			for (int ii=0;ii<B3_MAX_NUM_BATCHES;ii++)
+			{
+				int numInBatch = batchSizes->at(cellId*B3_MAX_NUM_BATCHES+ii);
+				if (!numInBatch)
+					break;
+
+				{
+					b3LauncherCL launcher( m_data->m_queue, m_data->m_solveSingleContactKernel,"m_solveSingleContactKernel" );
+					launcher.setBuffer(bodyBuf->getBufferCL() );
+					launcher.setBuffer(shapeBuf->getBufferCL() );
+					launcher.setBuffer(	constraint->getBufferCL() );
+					launcher.setConst(cellId);
+					launcher.setConst(offset);
+					launcher.setConst(numInBatch);
+					launcher.launch1D(numInBatch);
+					offset+=numInBatch;
+				}
+			}
+		}
+	}
+
+
+	for(int iter=0; iter<numIterations; iter++)
+	{
+		for (int cellId=0;cellId<numBatches;cellId++)
+		{
+			int offset = 0;
+			for (int ii=0;ii<B3_MAX_NUM_BATCHES;ii++)
+			{
+				int numInBatch = batchSizes->at(cellId*B3_MAX_NUM_BATCHES+ii);
+				if (!numInBatch)
+					break;
+
+				{
+					b3LauncherCL launcher( m_data->m_queue, m_data->m_solveSingleFrictionKernel,"m_solveSingleFrictionKernel" );
+					launcher.setBuffer(bodyBuf->getBufferCL() );
+					launcher.setBuffer(shapeBuf->getBufferCL() );
+					launcher.setBuffer(	constraint->getBufferCL() );
+					launcher.setConst(cellId);
+					launcher.setConst(offset);
+					launcher.setConst(numInBatch);
+					launcher.launch1D(numInBatch);
+					offset+=numInBatch;
+				}
+			}
+		}
+	}
+}
+
+void b3GpuPgsContactSolver::solveContactConstraint(  const b3OpenCLArray<b3RigidBodyData>* bodyBuf, const b3OpenCLArray<b3InertiaData>* shapeBuf, 
+			b3OpenCLArray<b3GpuConstraint4>* constraint, void* additionalData, int n ,int maxNumBatches,int numIterations, const b3AlignedObjectArray<int>* batchSizes)//,const b3OpenCLArray<int>* gpuBatchSizes)
+{
+	
+	//sort the contacts
+
+	
+	b3Int4 cdata = b3MakeInt4( n, 0, 0, 0 );
+	{
+		
+		const int nn = B3_SOLVER_N_CELLS;
+
+		cdata.x = 0;
+		cdata.y = maxNumBatches;//250;
+
+
+		int numWorkItems = 64*nn/B3_SOLVER_N_BATCHES;
+#ifdef DEBUG_ME
+		SolverDebugInfo* debugInfo = new  SolverDebugInfo[numWorkItems];
+		adl::b3OpenCLArray<SolverDebugInfo> gpuDebugInfo(data->m_device,numWorkItems);
+#endif
+
+
+
+		{
+
+			B3_PROFILE("m_batchSolveKernel iterations");
+			for(int iter=0; iter<numIterations; iter++)
+			{
+				for(int ib=0; ib<B3_SOLVER_N_BATCHES; ib++)
+				{
+#ifdef DEBUG_ME
+					memset(debugInfo,0,sizeof(SolverDebugInfo)*numWorkItems);
+					gpuDebugInfo.write(debugInfo,numWorkItems);
+#endif
+
+
+					cdata.z = ib;
+					
+
+				b3LauncherCL launcher( m_data->m_queue, m_data->m_solveContactKernel,"m_solveContactKernel" );
+#if 1
+                    
+					b3BufferInfoCL bInfo[] = { 
+
+						b3BufferInfoCL( bodyBuf->getBufferCL() ), 
+						b3BufferInfoCL( shapeBuf->getBufferCL() ), 
+						b3BufferInfoCL( constraint->getBufferCL() ),
+						b3BufferInfoCL( m_data->m_solverGPU->m_numConstraints->getBufferCL() ), 
+						b3BufferInfoCL( m_data->m_solverGPU->m_offsets->getBufferCL() ) 
+#ifdef DEBUG_ME
+						,	b3BufferInfoCL(&gpuDebugInfo)
+#endif
+						};
+
+					
+
+                    launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+					launcher.setBuffer( m_data->m_solverGPU->m_batchSizes.getBufferCL());
+					//launcher.setConst(  cdata.x );
+                    launcher.setConst(  cdata.y );
+                    launcher.setConst(  cdata.z );
+					b3Int4 nSplit;
+					nSplit.x = B3_SOLVER_N_SPLIT_X;
+					nSplit.y = B3_SOLVER_N_SPLIT_Y;
+					nSplit.z = B3_SOLVER_N_SPLIT_Z;
+
+                    launcher.setConst(  nSplit );
+                    launcher.launch1D( numWorkItems, 64 );
+
+                    
+#else
+                    const char* fileName = "m_batchSolveKernel.bin";
+                    FILE* f = fopen(fileName,"rb");
+                    if (f)
+                    {
+                        int sizeInBytes=0;
+                        if (fseek(f, 0, SEEK_END) || (sizeInBytes = ftell(f)) == EOF || fseek(f, 0, SEEK_SET))
+                        {
+                            printf("error, cannot get file size\n");
+                            exit(0);
+                        }
+                        
+                        unsigned char* buf = (unsigned char*) malloc(sizeInBytes);
+                        fread(buf,sizeInBytes,1,f);
+                        int serializedBytes = launcher.deserializeArgs(buf, sizeInBytes,m_context);
+                        int num = *(int*)&buf[serializedBytes];
+                        
+                        launcher.launch1D( num);
+
+                        //this clFinish is for testing on errors
+                        clFinish(m_queue);
+                    }
+
+#endif
+					
+
+#ifdef DEBUG_ME
+					clFinish(m_queue);
+					gpuDebugInfo.read(debugInfo,numWorkItems);
+					clFinish(m_queue);
+					for (int i=0;i<numWorkItems;i++)
+					{
+						if (debugInfo[i].m_valInt2>0)
+						{
+							printf("debugInfo[i].m_valInt2 = %d\n",i,debugInfo[i].m_valInt2);
+						}
+
+						if (debugInfo[i].m_valInt3>0)
+						{
+							printf("debugInfo[i].m_valInt3 = %d\n",i,debugInfo[i].m_valInt3);
+						}
+					}
+#endif //DEBUG_ME
+
+
+				}
+			}
+		
+			clFinish(m_data->m_queue);
+
+
+		}
+
+		cdata.x = 1;
+		bool applyFriction=true;
+		if (applyFriction)
+    	{
+			B3_PROFILE("m_batchSolveKernel iterations2");
+			for(int iter=0; iter<numIterations; iter++)
+			{
+				for(int ib=0; ib<B3_SOLVER_N_BATCHES; ib++)
+				{
+					cdata.z = ib;
+					
+
+					b3BufferInfoCL bInfo[] = { 
+						b3BufferInfoCL( bodyBuf->getBufferCL() ), 
+						b3BufferInfoCL( shapeBuf->getBufferCL() ), 
+						b3BufferInfoCL( constraint->getBufferCL() ),
+						b3BufferInfoCL( m_data->m_solverGPU->m_numConstraints->getBufferCL() ), 
+						b3BufferInfoCL( m_data->m_solverGPU->m_offsets->getBufferCL() )
+#ifdef DEBUG_ME
+						,b3BufferInfoCL(&gpuDebugInfo)
+#endif //DEBUG_ME
+					};
+					b3LauncherCL launcher( m_data->m_queue, m_data->m_solveFrictionKernel,"m_solveFrictionKernel" );
+					launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+					launcher.setBuffer( m_data->m_solverGPU->m_batchSizes.getBufferCL());
+					//launcher.setConst(  cdata.x );
+                    launcher.setConst(  cdata.y );
+                    launcher.setConst(  cdata.z );
+
+                    b3Int4 nSplit;
+					nSplit.x = B3_SOLVER_N_SPLIT_X;
+					nSplit.y = B3_SOLVER_N_SPLIT_Y;
+					nSplit.z = B3_SOLVER_N_SPLIT_Z;
+
+                    launcher.setConst(  nSplit );
+                    
+					launcher.launch1D( 64*nn/B3_SOLVER_N_BATCHES, 64 );
+				}
+			}
+			clFinish(m_data->m_queue);
+			
+		}
+#ifdef DEBUG_ME
+		delete[] debugInfo;
+#endif //DEBUG_ME
+	}
+
+	
+}
+
+
+
+
+
+
+
+
+
+
+
+static bool sortfnc(const b3SortData& a,const b3SortData& b)
+{
+	return (a.m_key<b.m_key);
+}
+
+static bool b3ContactCmp(const b3Contact4& p, const b3Contact4& q)
+{
+	return ((p.m_bodyAPtrAndSignBit<q.m_bodyAPtrAndSignBit) ||
+		((p.m_bodyAPtrAndSignBit==q.m_bodyAPtrAndSignBit) && (p.m_bodyBPtrAndSignBit<q.m_bodyBPtrAndSignBit)) ||
+		((p.m_bodyAPtrAndSignBit==q.m_bodyAPtrAndSignBit) && (p.m_bodyBPtrAndSignBit==q.m_bodyBPtrAndSignBit)  &&     p.m_childIndexA<q.m_childIndexA ) ||
+		((p.m_bodyAPtrAndSignBit==q.m_bodyAPtrAndSignBit) && (p.m_bodyBPtrAndSignBit==q.m_bodyBPtrAndSignBit)  &&     p.m_childIndexA<q.m_childIndexA ) ||
+		((p.m_bodyAPtrAndSignBit==q.m_bodyAPtrAndSignBit) && (p.m_bodyBPtrAndSignBit==q.m_bodyBPtrAndSignBit)  &&     p.m_childIndexA==q.m_childIndexA  && p.m_childIndexB<q.m_childIndexB)
+		);
+}
+
+
+
+
+
+
+
+
+
+
+
+#define USE_SPATIAL_BATCHING 1
+#define USE_4x4_GRID 1
+
+#ifndef USE_SPATIAL_BATCHING
+static const int gridTable4x4[] = 
+{
+	0,1,17,16,
+	1,2,18,19,
+	17,18,32,3,
+	16,19,3,34
+};
+static const int gridTable8x8[] = 
+{
+	0,  2,  3, 16, 17, 18, 19,  1,
+	66, 64, 80, 67, 82, 81, 65, 83,
+	131,144,128,130,147,129,145,146,
+	208,195,194,192,193,211,210,209,
+	21, 22, 23,  5,  4,  6,  7, 20,
+	86, 85, 69, 87, 70, 68, 84, 71,
+	151,133,149,150,135,148,132,134,
+	197,27,214,213,212,199,198,196
+	
+};
+
+
+#endif
+
+
+void SetSortDataCPU(b3Contact4* gContact, b3RigidBodyData* gBodies, b3SortData* gSortDataOut, int nContacts,float scale,const b3Int4& nSplit,int staticIdx)
+{
+	for (int gIdx=0;gIdx<nContacts;gIdx++)
+	{
+		if( gIdx < nContacts )
+		{
+			int aPtrAndSignBit  = gContact[gIdx].m_bodyAPtrAndSignBit;
+			int bPtrAndSignBit  = gContact[gIdx].m_bodyBPtrAndSignBit;
+
+			int aIdx = abs(aPtrAndSignBit );
+			int bIdx = abs(bPtrAndSignBit);
+
+			bool aStatic = (aPtrAndSignBit<0) ||(aPtrAndSignBit==staticIdx);
+
+	#if USE_SPATIAL_BATCHING		
+			int idx = (aStatic)? bIdx: aIdx;
+			b3Vector3 p = gBodies[idx].m_pos;
+			int xIdx = (int)((p.x-((p.x<0.f)?1.f:0.f))*scale) & (nSplit.x-1);
+			int yIdx = (int)((p.y-((p.y<0.f)?1.f:0.f))*scale) & (nSplit.y-1);
+			int zIdx = (int)((p.z-((p.z<0.f)?1.f:0.f))*scale) & (nSplit.z-1);
+			
+			int newIndex = (xIdx+yIdx*nSplit.x+zIdx*nSplit.x*nSplit.y);
+		
+	#else//USE_SPATIAL_BATCHING
+			bool bStatic = (bPtrAndSignBit<0) ||(bPtrAndSignBit==staticIdx);
+
+		#if USE_4x4_GRID
+			int aa = aIdx&3;
+			int bb = bIdx&3;
+			if (aStatic)
+				aa = bb;
+			if (bStatic)
+				bb = aa;
+
+			int gridIndex = aa + bb*4;
+			int newIndex = gridTable4x4[gridIndex];
+		#else//USE_4x4_GRID
+			int aa = aIdx&7;
+			int bb = bIdx&7;
+			if (aStatic)
+				aa = bb;
+			if (bStatic)
+				bb = aa;
+
+			int gridIndex = aa + bb*8;
+			int newIndex = gridTable8x8[gridIndex];
+		#endif//USE_4x4_GRID
+	#endif//USE_SPATIAL_BATCHING
+
+
+			gSortDataOut[gIdx].x = newIndex;
+			gSortDataOut[gIdx].y = gIdx;
+		}
+		else
+		{
+			gSortDataOut[gIdx].x = 0xffffffff;
+		}
+	}
+}
+
+
+
+
+
+
+void b3GpuPgsContactSolver::solveContacts(int numBodies, cl_mem bodyBuf, cl_mem inertiaBuf, int numContacts, cl_mem contactBuf, const b3Config& config, int static0Index)
+{
+	B3_PROFILE("solveContacts");
+	m_data->m_bodyBufferGPU->setFromOpenCLBuffer(bodyBuf,numBodies);
+	m_data->m_inertiaBufferGPU->setFromOpenCLBuffer(inertiaBuf,numBodies);
+	m_data->m_pBufContactOutGPU->setFromOpenCLBuffer(contactBuf,numContacts);
+
+	if (optionalSortContactsDeterminism)
+	{
+		if (!gCpuSortContactsDeterminism)
+		{
+			B3_PROFILE("GPU Sort contact constraints (determinism)");
+
+			m_data->m_pBufContactOutGPUCopy->resize(numContacts);
+			m_data->m_contactKeyValues->resize(numContacts);
+
+			m_data->m_pBufContactOutGPU->copyToCL(m_data->m_pBufContactOutGPUCopy->getBufferCL(),numContacts,0,0);
+
+			{
+				b3LauncherCL launcher(m_data->m_queue, m_data->m_setDeterminismSortDataChildShapeBKernel,"m_setDeterminismSortDataChildShapeBKernel");
+				launcher.setBuffer(m_data->m_pBufContactOutGPUCopy->getBufferCL());
+				launcher.setBuffer(m_data->m_contactKeyValues->getBufferCL());
+				launcher.setConst(numContacts);
+				launcher.launch1D( numContacts, 64 );
+			}
+			m_data->m_solverGPU->m_sort32->execute(*m_data->m_contactKeyValues);
+			{
+				b3LauncherCL launcher(m_data->m_queue, m_data->m_setDeterminismSortDataChildShapeAKernel,"m_setDeterminismSortDataChildShapeAKernel");
+				launcher.setBuffer(m_data->m_pBufContactOutGPUCopy->getBufferCL());
+				launcher.setBuffer(m_data->m_contactKeyValues->getBufferCL());
+				launcher.setConst(numContacts);
+				launcher.launch1D( numContacts, 64 );
+			}
+			m_data->m_solverGPU->m_sort32->execute(*m_data->m_contactKeyValues);
+			{
+				b3LauncherCL launcher(m_data->m_queue, m_data->m_setDeterminismSortDataBodyBKernel,"m_setDeterminismSortDataBodyBKernel");
+				launcher.setBuffer(m_data->m_pBufContactOutGPUCopy->getBufferCL());
+				launcher.setBuffer(m_data->m_contactKeyValues->getBufferCL());
+				launcher.setConst(numContacts);
+				launcher.launch1D( numContacts, 64 );
+			}
+						
+			m_data->m_solverGPU->m_sort32->execute(*m_data->m_contactKeyValues);
+			
+			{
+				b3LauncherCL launcher(m_data->m_queue, m_data->m_setDeterminismSortDataBodyAKernel,"m_setDeterminismSortDataBodyAKernel");
+				launcher.setBuffer(m_data->m_pBufContactOutGPUCopy->getBufferCL());
+				launcher.setBuffer(m_data->m_contactKeyValues->getBufferCL());
+				launcher.setConst(numContacts);
+				launcher.launch1D( numContacts, 64 );
+			}
+
+			m_data->m_solverGPU->m_sort32->execute(*m_data->m_contactKeyValues);
+
+			{
+				B3_PROFILE("gpu reorderContactKernel (determinism)");
+                                
+				b3Int4 cdata;
+				cdata.x = numContacts;
+                                
+				//b3BufferInfoCL bInfo[] = { b3BufferInfoCL( m_data->m_pBufContactOutGPU->getBufferCL() ), b3BufferInfoCL( m_data->m_solverGPU->m_contactBuffer2->getBufferCL())
+				//	, b3BufferInfoCL( m_data->m_solverGPU->m_sortDataBuffer->getBufferCL()) };
+				b3LauncherCL launcher(m_data->m_queue,m_data->m_solverGPU->m_reorderContactKernel,"m_reorderContactKernel");
+				launcher.setBuffer(m_data->m_pBufContactOutGPUCopy->getBufferCL());
+				launcher.setBuffer(m_data->m_pBufContactOutGPU->getBufferCL());
+				launcher.setBuffer(m_data->m_contactKeyValues->getBufferCL());
+				launcher.setConst( cdata );
+				launcher.launch1D( numContacts, 64 );
+            }
+
+		} else
+		{
+			B3_PROFILE("CPU Sort contact constraints (determinism)");
+			b3AlignedObjectArray<b3Contact4> cpuConstraints;
+			m_data->m_pBufContactOutGPU->copyToHost(cpuConstraints);
+			bool sort = true;
+			if (sort)
+			{
+				cpuConstraints.quickSort(b3ContactCmp);
+
+				for (int i=0;i<cpuConstraints.size();i++)
+				{
+					cpuConstraints[i].m_batchIdx = i;
+				}
+			}
+			m_data->m_pBufContactOutGPU->copyFromHost(cpuConstraints);
+			if (m_debugOutput==100)
+			{
+				for (int i=0;i<cpuConstraints.size();i++)
+				{
+					printf("c[%d].m_bodyA = %d, m_bodyB = %d, batchId = %d\n",i,cpuConstraints[i].m_bodyAPtrAndSignBit,cpuConstraints[i].m_bodyBPtrAndSignBit, cpuConstraints[i].m_batchIdx);
+				}
+			}
+
+			m_debugOutput++;
+		}
+	}
+	
+
+
+
+	int nContactOut = m_data->m_pBufContactOutGPU->size();
+
+	bool useSolver = true;
+	
+
+    if (useSolver)
+    {
+        float dt=1./60.;
+        b3ConstraintCfg csCfg( dt );
+        csCfg.m_enableParallelSolve = true;
+        csCfg.m_batchCellSize = 6;
+        csCfg.m_staticIdx = static0Index;
+        
+        
+        b3OpenCLArray<b3RigidBodyData>* bodyBuf = m_data->m_bodyBufferGPU;
+
+        void* additionalData = 0;//m_data->m_frictionCGPU;
+        const b3OpenCLArray<b3InertiaData>* shapeBuf = m_data->m_inertiaBufferGPU;
+        b3OpenCLArray<b3GpuConstraint4>* contactConstraintOut = m_data->m_contactCGPU;
+        int nContacts = nContactOut;
+        
+        
+		int maxNumBatches = 0;
+ 
+		if (!gUseLargeBatches)
+        {
+            
+            if( m_data->m_solverGPU->m_contactBuffer2)
+            {
+                m_data->m_solverGPU->m_contactBuffer2->resize(nContacts);
+            }
+            
+            if( m_data->m_solverGPU->m_contactBuffer2 == 0 )
+            {
+				m_data->m_solverGPU->m_contactBuffer2 = new b3OpenCLArray<b3Contact4>(m_data->m_context,m_data->m_queue, nContacts );
+                m_data->m_solverGPU->m_contactBuffer2->resize(nContacts);
+            }
+			
+            //clFinish(m_data->m_queue);
+            
+            
+            
+			{
+				B3_PROFILE("batching");
+				//@todo: just reserve it, without copy of original contact (unless we use warmstarting)
+
+
+
+				//const b3OpenCLArray<b3RigidBodyData>* bodyNative = bodyBuf;
+
+
+				{
+
+					//b3OpenCLArray<b3RigidBodyData>* bodyNative = b3OpenCLArrayUtils::map<adl::TYPE_CL, true>( data->m_device, bodyBuf );
+					//b3OpenCLArray<b3Contact4>* contactNative = b3OpenCLArrayUtils::map<adl::TYPE_CL, true>( data->m_device, contactsIn );
+
+					const int sortAlignment = 512; // todo. get this out of sort
+					if( csCfg.m_enableParallelSolve )
+					{
+
+
+						int sortSize = B3NEXTMULTIPLEOF( nContacts, sortAlignment );
+
+						b3OpenCLArray<unsigned int>* countsNative = m_data->m_solverGPU->m_numConstraints;
+						b3OpenCLArray<unsigned int>* offsetsNative = m_data->m_solverGPU->m_offsets;
+
+
+						if (!gCpuSetSortData)
+						{	//	2. set cell idx
+							B3_PROFILE("GPU set cell idx");
+							struct CB
+							{
+								int m_nContacts;
+								int m_staticIdx;
+								float m_scale;
+								b3Int4 m_nSplit;
+							};
+
+							b3Assert( sortSize%64 == 0 );
+							CB cdata;
+							cdata.m_nContacts = nContacts;
+							cdata.m_staticIdx = csCfg.m_staticIdx;
+							cdata.m_scale = 1.f/csCfg.m_batchCellSize;
+							cdata.m_nSplit.x = B3_SOLVER_N_SPLIT_X;
+							cdata.m_nSplit.y = B3_SOLVER_N_SPLIT_Y;
+							cdata.m_nSplit.z = B3_SOLVER_N_SPLIT_Z;
+
+							m_data->m_solverGPU->m_sortDataBuffer->resize(nContacts);
+
+
+							b3BufferInfoCL bInfo[] = { b3BufferInfoCL( m_data->m_pBufContactOutGPU->getBufferCL() ), b3BufferInfoCL( bodyBuf->getBufferCL()), b3BufferInfoCL( m_data->m_solverGPU->m_sortDataBuffer->getBufferCL()) };
+							b3LauncherCL launcher(m_data->m_queue, m_data->m_solverGPU->m_setSortDataKernel,"m_setSortDataKernel" );
+							launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+							launcher.setConst( cdata.m_nContacts );
+							launcher.setConst( cdata.m_scale );
+							launcher.setConst(cdata.m_nSplit);
+							launcher.setConst(cdata.m_staticIdx);
+
+
+							launcher.launch1D( sortSize, 64 );
+						} else
+						{
+							m_data->m_solverGPU->m_sortDataBuffer->resize(nContacts);
+							b3AlignedObjectArray<b3SortData> sortDataCPU;
+							m_data->m_solverGPU->m_sortDataBuffer->copyToHost(sortDataCPU);
+
+							b3AlignedObjectArray<b3Contact4> contactCPU;
+							m_data->m_pBufContactOutGPU->copyToHost(contactCPU);
+							b3AlignedObjectArray<b3RigidBodyData> bodiesCPU;
+							bodyBuf->copyToHost(bodiesCPU);
+							float scale = 1.f/csCfg.m_batchCellSize;
+							b3Int4 nSplit;
+							nSplit.x = B3_SOLVER_N_SPLIT_X;
+							nSplit.y = B3_SOLVER_N_SPLIT_Y;
+							nSplit.z = B3_SOLVER_N_SPLIT_Z;
+
+							SetSortDataCPU(&contactCPU[0],  &bodiesCPU[0], &sortDataCPU[0], nContacts,scale,nSplit,csCfg.m_staticIdx);
+
+
+							m_data->m_solverGPU->m_sortDataBuffer->copyFromHost(sortDataCPU);
+						}
+
+
+
+						if (!gCpuRadixSort)
+						{	//	3. sort by cell idx
+							B3_PROFILE("gpuRadixSort");
+							//int n = B3_SOLVER_N_SPLIT*B3_SOLVER_N_SPLIT;
+							//int sortBit = 32;
+							//if( n <= 0xffff ) sortBit = 16;
+							//if( n <= 0xff ) sortBit = 8;
+							//adl::RadixSort<adl::TYPE_CL>::execute( data->m_sort, *data->m_sortDataBuffer, sortSize );
+							//adl::RadixSort32<adl::TYPE_CL>::execute( data->m_sort32, *data->m_sortDataBuffer, sortSize );
+							b3OpenCLArray<b3SortData>& keyValuesInOut = *(m_data->m_solverGPU->m_sortDataBuffer);
+							this->m_data->m_solverGPU->m_sort32->execute(keyValuesInOut);
+
+
+
+						} else
+						{
+							b3OpenCLArray<b3SortData>& keyValuesInOut = *(m_data->m_solverGPU->m_sortDataBuffer);
+							b3AlignedObjectArray<b3SortData> hostValues;
+							keyValuesInOut.copyToHost(hostValues);
+							hostValues.quickSort(sortfnc);
+							keyValuesInOut.copyFromHost(hostValues);
+						}
+
+
+						if (gUseScanHost)
+						{
+							//	4. find entries
+							B3_PROFILE("cpuBoundSearch");
+							b3AlignedObjectArray<unsigned int> countsHost;
+							countsNative->copyToHost(countsHost);
+
+							b3AlignedObjectArray<b3SortData> sortDataHost;
+							m_data->m_solverGPU->m_sortDataBuffer->copyToHost(sortDataHost);
+
+
+							//m_data->m_solverGPU->m_search->executeHost(*m_data->m_solverGPU->m_sortDataBuffer,nContacts,*countsNative,B3_SOLVER_N_CELLS,b3BoundSearchCL::COUNT);
+							m_data->m_solverGPU->m_search->executeHost(sortDataHost,nContacts,countsHost,B3_SOLVER_N_CELLS,b3BoundSearchCL::COUNT);
+
+							countsNative->copyFromHost(countsHost);
+
+
+							//adl::BoundSearch<adl::TYPE_CL>::execute( data->m_search, *data->m_sortDataBuffer, nContacts, *countsNative,
+							//	B3_SOLVER_N_SPLIT*B3_SOLVER_N_SPLIT, adl::BoundSearchBase::COUNT );
+
+							//unsigned int sum;
+							//m_data->m_solverGPU->m_scan->execute(*countsNative,*offsetsNative, B3_SOLVER_N_CELLS);//,&sum );
+							b3AlignedObjectArray<unsigned int> offsetsHost;
+							offsetsHost.resize(offsetsNative->size());
+
+
+							m_data->m_solverGPU->m_scan->executeHost(countsHost,offsetsHost, B3_SOLVER_N_CELLS);//,&sum );
+							offsetsNative->copyFromHost(offsetsHost);
+
+							//printf("sum = %d\n",sum);
+						}  else
+						{
+							//	4. find entries
+							B3_PROFILE("gpuBoundSearch");
+							m_data->m_solverGPU->m_search->execute(*m_data->m_solverGPU->m_sortDataBuffer,nContacts,*countsNative,B3_SOLVER_N_CELLS,b3BoundSearchCL::COUNT);
+							m_data->m_solverGPU->m_scan->execute(*countsNative,*offsetsNative, B3_SOLVER_N_CELLS);//,&sum );
+						} 
+
+
+
+
+						if (nContacts)
+						{	//	5. sort constraints by cellIdx
+							if (gReorderContactsOnCpu)
+							{
+								B3_PROFILE("cpu m_reorderContactKernel");
+								b3AlignedObjectArray<b3SortData> sortDataHost;
+								m_data->m_solverGPU->m_sortDataBuffer->copyToHost(sortDataHost);
+								b3AlignedObjectArray<b3Contact4> inContacts;
+								b3AlignedObjectArray<b3Contact4> outContacts;
+								m_data->m_pBufContactOutGPU->copyToHost(inContacts);
+								outContacts.resize(inContacts.size());
+								for (int i=0;i<nContacts;i++)
+								{
+									int srcIdx = sortDataHost[i].y;
+									outContacts[i] = inContacts[srcIdx];
+								}
+								m_data->m_solverGPU->m_contactBuffer2->copyFromHost(outContacts);
+
+								/*								"void ReorderContactKernel(__global struct b3Contact4Data* in, __global struct b3Contact4Data* out, __global int2* sortData, int4 cb )\n"
+								"{\n"
+								"	int nContacts = cb.x;\n"
+								"	int gIdx = GET_GLOBAL_IDX;\n"
+								"	if( gIdx < nContacts )\n"
+								"	{\n"
+								"		int srcIdx = sortData[gIdx].y;\n"
+								"		out[gIdx] = in[srcIdx];\n"
+								"	}\n"
+								"}\n"
+								*/
+							} else
+							{
+								B3_PROFILE("gpu m_reorderContactKernel");
+
+								b3Int4 cdata;
+								cdata.x = nContacts;
+
+								b3BufferInfoCL bInfo[] = { 
+									b3BufferInfoCL( m_data->m_pBufContactOutGPU->getBufferCL() ), 
+									b3BufferInfoCL( m_data->m_solverGPU->m_contactBuffer2->getBufferCL())
+									, b3BufferInfoCL( m_data->m_solverGPU->m_sortDataBuffer->getBufferCL()) };
+
+									b3LauncherCL launcher(m_data->m_queue,m_data->m_solverGPU->m_reorderContactKernel,"m_reorderContactKernel");
+									launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+									launcher.setConst( cdata );
+									launcher.launch1D( nContacts, 64 );
+							}
+						}
+
+
+
+
+					}
+
+				}
+
+				//clFinish(m_data->m_queue);
+
+				//				{
+				//				b3AlignedObjectArray<unsigned int> histogram;
+				//				m_data->m_solverGPU->m_numConstraints->copyToHost(histogram);
+				//				printf(",,,\n");
+				//				}
+
+
+				if (nContacts)
+				{
+
+					if (gUseCpuCopyConstraints)
+					{
+						for (int i=0;i<nContacts;i++)
+						{
+							m_data->m_pBufContactOutGPU->copyFromOpenCLArray(*m_data->m_solverGPU->m_contactBuffer2);
+							//							m_data->m_solverGPU->m_contactBuffer2->getBufferCL(); 
+							//						m_data->m_pBufContactOutGPU->getBufferCL() 
+						}
+
+					} else
+					{
+						B3_PROFILE("gpu m_copyConstraintKernel");
+						b3Int4 cdata; cdata.x = nContacts;
+						b3BufferInfoCL bInfo[] = { 
+							b3BufferInfoCL(  m_data->m_solverGPU->m_contactBuffer2->getBufferCL() ), 
+							b3BufferInfoCL( m_data->m_pBufContactOutGPU->getBufferCL() ) 
+						};
+
+						b3LauncherCL launcher(m_data->m_queue, m_data->m_solverGPU->m_copyConstraintKernel,"m_copyConstraintKernel" );
+						launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+						launcher.setConst(  cdata );
+						launcher.launch1D( nContacts, 64 );
+						//we use the clFinish for proper benchmark/profile
+						clFinish(m_data->m_queue);
+					}
+				}
+
+
+//				bool compareGPU = false;
+				if (nContacts)
+				{
+					if (!gCpuBatchContacts)
+					{
+						B3_PROFILE("gpu batchContacts");
+						maxNumBatches = 250;//250;
+						m_data->m_solverGPU->batchContacts( m_data->m_pBufContactOutGPU, nContacts, m_data->m_solverGPU->m_numConstraints, m_data->m_solverGPU->m_offsets, csCfg.m_staticIdx );
+						clFinish(m_data->m_queue);
+					} else
+					{
+						B3_PROFILE("cpu batchContacts");
+						static b3AlignedObjectArray<b3Contact4> cpuContacts;
+						b3OpenCLArray<b3Contact4>* contactsIn = m_data->m_solverGPU->m_contactBuffer2;
+						{
+							B3_PROFILE("copyToHost");
+							contactsIn->copyToHost(cpuContacts);
+						}
+						b3OpenCLArray<unsigned int>* countsNative = m_data->m_solverGPU->m_numConstraints;
+						b3OpenCLArray<unsigned int>* offsetsNative = m_data->m_solverGPU->m_offsets;
+
+						b3AlignedObjectArray<unsigned int> nNativeHost;
+						b3AlignedObjectArray<unsigned int> offsetsNativeHost;
+
+						{
+							B3_PROFILE("countsNative/offsetsNative copyToHost");
+							countsNative->copyToHost(nNativeHost);
+							offsetsNative->copyToHost(offsetsNativeHost);
+						}
+
+
+						int numNonzeroGrid=0;
+
+						if (gUseLargeBatches)
+						{
+							m_data->m_batchSizes.resize(B3_MAX_NUM_BATCHES);
+							int totalNumConstraints = cpuContacts.size();
+							//int simdWidth =numBodies+1;//-1;//64;//-1;//32;
+							int numBatches = sortConstraintByBatch3( &cpuContacts[0], totalNumConstraints, totalNumConstraints+1,csCfg.m_staticIdx ,numBodies,&m_data->m_batchSizes[0]);	//	on GPU
+							maxNumBatches = b3Max(numBatches,maxNumBatches);
+							static int globalMaxBatch = 0;
+							if (maxNumBatches>globalMaxBatch )
+							{
+								globalMaxBatch  = maxNumBatches;
+								b3Printf("maxNumBatches = %d\n",maxNumBatches);
+							}
+								
+						} else
+						{
+							m_data->m_batchSizes.resize(B3_SOLVER_N_CELLS*B3_MAX_NUM_BATCHES);
+							B3_PROFILE("cpu batch grid");
+							for(int i=0; i<B3_SOLVER_N_CELLS; i++)
+							{
+								int n = (nNativeHost)[i];
+								int offset = (offsetsNativeHost)[i];
+								if( n )
+								{
+									numNonzeroGrid++;
+									int simdWidth =numBodies+1;//-1;//64;//-1;//32;
+									int numBatches = sortConstraintByBatch3( &cpuContacts[0]+offset, n, simdWidth,csCfg.m_staticIdx ,numBodies,&m_data->m_batchSizes[i*B3_MAX_NUM_BATCHES]);	//	on GPU
+									maxNumBatches = b3Max(numBatches,maxNumBatches);
+									static int globalMaxBatch = 0;
+									if (maxNumBatches>globalMaxBatch )
+									{
+										globalMaxBatch  = maxNumBatches;
+										b3Printf("maxNumBatches = %d\n",maxNumBatches);
+									}
+									//we use the clFinish for proper benchmark/profile
+									
+								}
+							}
+							//clFinish(m_data->m_queue);
+						}
+						{
+							B3_PROFILE("m_contactBuffer->copyFromHost");
+							m_data->m_solverGPU->m_contactBuffer2->copyFromHost((b3AlignedObjectArray<b3Contact4>&)cpuContacts);
+						}
+
+					} 
+
+				}
+
+
+			
+
+
+			} 
+
+
+		} 
+
+
+			//printf("maxNumBatches = %d\n", maxNumBatches);
+
+		if (gUseLargeBatches)
+		{
+			if (nContacts)
+			{
+				B3_PROFILE("cpu batchContacts");
+				static b3AlignedObjectArray<b3Contact4> cpuContacts;
+//				b3OpenCLArray<b3Contact4>* contactsIn = m_data->m_solverGPU->m_contactBuffer2;
+				{
+					B3_PROFILE("copyToHost");
+					m_data->m_pBufContactOutGPU->copyToHost(cpuContacts);
+				}
+//				b3OpenCLArray<unsigned int>* countsNative = m_data->m_solverGPU->m_numConstraints;
+//				b3OpenCLArray<unsigned int>* offsetsNative = m_data->m_solverGPU->m_offsets;
+
+
+
+//				int numNonzeroGrid=0;
+
+				{
+					m_data->m_batchSizes.resize(B3_MAX_NUM_BATCHES);
+					int totalNumConstraints = cpuContacts.size();
+	//				int simdWidth =numBodies+1;//-1;//64;//-1;//32;
+					int numBatches = sortConstraintByBatch3( &cpuContacts[0], totalNumConstraints, totalNumConstraints+1,csCfg.m_staticIdx ,numBodies,&m_data->m_batchSizes[0]);	//	on GPU
+					maxNumBatches = b3Max(numBatches,maxNumBatches);
+					static int globalMaxBatch = 0;
+					if (maxNumBatches>globalMaxBatch )
+					{
+						globalMaxBatch  = maxNumBatches;
+						b3Printf("maxNumBatches = %d\n",maxNumBatches);
+					}
+								
+				}
+				{
+					B3_PROFILE("m_contactBuffer->copyFromHost");
+					m_data->m_solverGPU->m_contactBuffer2->copyFromHost((b3AlignedObjectArray<b3Contact4>&)cpuContacts);
+				}
+
+			} 
+
+		}
+
+		if (nContacts)
+		{
+			B3_PROFILE("gpu convertToConstraints");
+			m_data->m_solverGPU->convertToConstraints( bodyBuf, 
+				shapeBuf, m_data->m_solverGPU->m_contactBuffer2,
+				contactConstraintOut, 
+				additionalData, nContacts, 
+				(b3SolverBase::ConstraintCfg&) csCfg );
+			clFinish(m_data->m_queue);
+		}
+
+
+		if (1)
+		{
+			int numIter = 4;
+
+			m_data->m_solverGPU->m_nIterations = numIter;//10
+			if (!gCpuSolveConstraint)
+			{
+				B3_PROFILE("GPU solveContactConstraint");
+
+				/*m_data->m_solverGPU->solveContactConstraint(
+				m_data->m_bodyBufferGPU, 
+				m_data->m_inertiaBufferGPU,
+				m_data->m_contactCGPU,0,
+				nContactOut ,
+				maxNumBatches);
+				*/
+
+				//m_data->m_batchSizesGpu->copyFromHost(m_data->m_batchSizes);
+
+				if (gUseLargeBatches)
+				{
+					solveContactConstraintBatchSizes(m_data->m_bodyBufferGPU, 
+						m_data->m_inertiaBufferGPU,
+						m_data->m_contactCGPU,0,
+						nContactOut ,
+						maxNumBatches,numIter,&m_data->m_batchSizes);
+				} else
+				{
+					solveContactConstraint(
+						m_data->m_bodyBufferGPU, 
+						m_data->m_inertiaBufferGPU,
+						m_data->m_contactCGPU,0,
+						nContactOut ,
+						maxNumBatches,numIter,&m_data->m_batchSizes);//m_data->m_batchSizesGpu);
+				}
+			}
+			else
+			{
+				B3_PROFILE("Host solveContactConstraint");
+
+				m_data->m_solverGPU->solveContactConstraintHost(m_data->m_bodyBufferGPU, m_data->m_inertiaBufferGPU, m_data->m_contactCGPU,0, nContactOut ,maxNumBatches,&m_data->m_batchSizes);
+			}
+            
+            
+        }
+        
+        
+#if 0
+        if (0)
+        {
+            B3_PROFILE("read body velocities back to CPU");
+            //read body updated linear/angular velocities back to CPU
+            m_data->m_bodyBufferGPU->read(
+                                                  m_data->m_bodyBufferCPU->m_ptr,numOfConvexRBodies);
+            adl::DeviceUtils::waitForCompletion( m_data->m_deviceCL );
+        }
+#endif
+        
+    }
+
+}
+
+
+void b3GpuPgsContactSolver::batchContacts( b3OpenCLArray<b3Contact4>* contacts, int nContacts, b3OpenCLArray<unsigned int>* n, b3OpenCLArray<unsigned int>* offsets, int staticIdx )
+{
+}
+
+
+
+
+
+
+
+
+
+
+
+b3AlignedObjectArray<unsigned int> idxBuffer;
+b3AlignedObjectArray<b3SortData> sortData;
+b3AlignedObjectArray<b3Contact4> old;
+
+
+inline int b3GpuPgsContactSolver::sortConstraintByBatch( b3Contact4* cs, int n, int simdWidth , int staticIdx, int numBodies)
+{
+	
+	B3_PROFILE("sortConstraintByBatch");
+	int numIter = 0;
+    
+	sortData.resize(n);
+	idxBuffer.resize(n);
+	old.resize(n);
+	
+	unsigned int* idxSrc = &idxBuffer[0];
+	unsigned int* idxDst = &idxBuffer[0];
+	int nIdxSrc, nIdxDst;
+    
+	const int N_FLG = 256;
+	const int FLG_MASK = N_FLG-1;
+	unsigned int flg[N_FLG/32];
+#if defined(_DEBUG)
+	for(int i=0; i<n; i++)
+		cs[i].getBatchIdx() = -1;
+#endif
+	for(int i=0; i<n; i++) 
+		idxSrc[i] = i;
+	nIdxSrc = n;
+    
+	int batchIdx = 0;
+    
+	{
+		B3_PROFILE("cpu batch innerloop");
+		while( nIdxSrc )
+		{
+			numIter++;
+			nIdxDst = 0;
+			int nCurrentBatch = 0;
+            
+			//	clear flag
+			for(int i=0; i<N_FLG/32; i++) flg[i] = 0;
+            
+			for(int i=0; i<nIdxSrc; i++)
+			{
+				int idx = idxSrc[i];
+				
+
+				b3Assert( idx < n );
+				//	check if it can go
+				int bodyAS = cs[idx].m_bodyAPtrAndSignBit;
+				int bodyBS = cs[idx].m_bodyBPtrAndSignBit;
+                
+				
+                
+				int bodyA = abs(bodyAS);
+				int bodyB = abs(bodyBS);
+                
+				int aIdx = bodyA & FLG_MASK;
+				int bIdx = bodyB & FLG_MASK;
+                
+				unsigned int aUnavailable = flg[ aIdx/32 ] & (1<<(aIdx&31));
+				unsigned int bUnavailable = flg[ bIdx/32 ] & (1<<(bIdx&31));
+                
+				bool aIsStatic = (bodyAS<0) || bodyAS==staticIdx;
+				bool bIsStatic = (bodyBS<0) || bodyBS==staticIdx;
+
+                //use inv_mass!
+				aUnavailable = !aIsStatic? aUnavailable:0;//
+				bUnavailable = !bIsStatic? bUnavailable:0;
+                
+				if( aUnavailable==0 && bUnavailable==0 ) // ok
+				{
+					if (!aIsStatic)
+						flg[ aIdx/32 ] |= (1<<(aIdx&31));
+					if (!bIsStatic)
+						flg[ bIdx/32 ] |= (1<<(bIdx&31));
+
+					cs[idx].getBatchIdx() = batchIdx;
+					sortData[idx].m_key = batchIdx;
+					sortData[idx].m_value = idx;
+                    
+					{
+						nCurrentBatch++;
+						if( nCurrentBatch == simdWidth )
+						{
+							nCurrentBatch = 0;
+							for(int i=0; i<N_FLG/32; i++) flg[i] = 0;
+						}
+					}
+				}
+				else
+				{
+					idxDst[nIdxDst++] = idx;
+				}
+			}
+			b3Swap( idxSrc, idxDst );
+			b3Swap( nIdxSrc, nIdxDst );
+			batchIdx ++;
+		}
+	}
+	{
+		B3_PROFILE("quickSort");
+		sortData.quickSort(sortfnc);
+	}
+	
+	
+	{
+        B3_PROFILE("reorder");
+		//	reorder
+		
+		memcpy( &old[0], cs, sizeof(b3Contact4)*n);
+		for(int i=0; i<n; i++)
+		{
+			int idx = sortData[i].m_value;
+			cs[i] = old[idx];
+		}
+	}
+    
+	
+#if defined(_DEBUG)
+    //		debugPrintf( "nBatches: %d\n", batchIdx );
+	for(int i=0; i<n; i++)
+    {
+        b3Assert( cs[i].getBatchIdx() != -1 );
+    }
+#endif
+	return batchIdx;
+}
+
+
+b3AlignedObjectArray<int> bodyUsed2;
+
+inline int b3GpuPgsContactSolver::sortConstraintByBatch2( b3Contact4* cs, int numConstraints, int simdWidth , int staticIdx, int numBodies)
+{
+	
+	B3_PROFILE("sortConstraintByBatch2");
+	
+
+	
+	bodyUsed2.resize(2*simdWidth);
+
+	for (int q=0;q<2*simdWidth;q++)
+		bodyUsed2[q]=0;
+
+	int curBodyUsed = 0;
+
+	int numIter = 0;
+    
+	m_data->m_sortData.resize(numConstraints);
+	m_data->m_idxBuffer.resize(numConstraints);
+	m_data->m_old.resize(numConstraints);
+	
+	unsigned int* idxSrc = &m_data->m_idxBuffer[0];
+		
+#if defined(_DEBUG)
+	for(int i=0; i<numConstraints; i++)
+		cs[i].getBatchIdx() = -1;
+#endif
+	for(int i=0; i<numConstraints; i++) 
+		idxSrc[i] = i;
+    
+	int numValidConstraints = 0;
+//	int unprocessedConstraintIndex = 0;
+
+	int batchIdx = 0;
+    
+
+	{
+		B3_PROFILE("cpu batch innerloop");
+		
+		while( numValidConstraints < numConstraints)
+		{
+			numIter++;
+			int nCurrentBatch = 0;
+			//	clear flag
+			for(int i=0; i<curBodyUsed; i++) 
+				bodyUsed2[i] = 0;
+            curBodyUsed = 0;
+
+			for(int i=numValidConstraints; i<numConstraints; i++)
+			{
+				int idx = idxSrc[i];
+				b3Assert( idx < numConstraints );
+				//	check if it can go
+				int bodyAS = cs[idx].m_bodyAPtrAndSignBit;
+				int bodyBS = cs[idx].m_bodyBPtrAndSignBit;
+				int bodyA = abs(bodyAS);
+				int bodyB = abs(bodyBS);
+				bool aIsStatic = (bodyAS<0) || bodyAS==staticIdx;
+				bool bIsStatic = (bodyBS<0) || bodyBS==staticIdx;
+				int aUnavailable = 0;
+				int bUnavailable = 0;
+				if (!aIsStatic)
+				{
+					for (int j=0;j<curBodyUsed;j++)
+					{
+						if (bodyA == bodyUsed2[j])
+						{
+							aUnavailable=1;
+							break;
+						}
+					}
+				}
+				if (!aUnavailable)
+				if (!bIsStatic)
+				{
+					for (int j=0;j<curBodyUsed;j++)
+					{
+						if (bodyB == bodyUsed2[j])
+						{
+							bUnavailable=1;
+							break;
+						}
+					}
+				}
+                
+				if( aUnavailable==0 && bUnavailable==0 ) // ok
+				{
+					if (!aIsStatic)
+					{
+						bodyUsed2[curBodyUsed++] = bodyA;
+					}
+					if (!bIsStatic)
+					{
+						bodyUsed2[curBodyUsed++] = bodyB;
+					}
+
+					cs[idx].getBatchIdx() = batchIdx;
+					m_data->m_sortData[idx].m_key = batchIdx;
+					m_data->m_sortData[idx].m_value = idx;
+
+					if (i!=numValidConstraints)
+					{
+						b3Swap(idxSrc[i], idxSrc[numValidConstraints]);
+					}
+
+					numValidConstraints++;
+					{
+						nCurrentBatch++;
+						if( nCurrentBatch == simdWidth )
+						{
+							nCurrentBatch = 0;
+							for(int i=0; i<curBodyUsed; i++) 
+								bodyUsed2[i] = 0;
+
+							
+							curBodyUsed = 0;
+						}
+					}
+				}
+			}
+			
+			batchIdx ++;
+		}
+	}
+	{
+		B3_PROFILE("quickSort");
+		//m_data->m_sortData.quickSort(sortfnc);
+	}
+
+	{
+        B3_PROFILE("reorder");
+		//	reorder
+		
+		memcpy( &m_data->m_old[0], cs, sizeof(b3Contact4)*numConstraints);
+
+		for(int i=0; i<numConstraints; i++)
+		{
+			b3Assert(m_data->m_sortData[idxSrc[i]].m_value == idxSrc[i]);
+			int idx = m_data->m_sortData[idxSrc[i]].m_value;
+			cs[i] = m_data->m_old[idx];
+		}
+	}
+	
+#if defined(_DEBUG)
+    //		debugPrintf( "nBatches: %d\n", batchIdx );
+	for(int i=0; i<numConstraints; i++)
+    {
+        b3Assert( cs[i].getBatchIdx() != -1 );
+    }
+#endif
+
+	
+	return batchIdx;
+}
+
+
+b3AlignedObjectArray<int> bodyUsed;
+b3AlignedObjectArray<int> curUsed;
+
+
+inline int b3GpuPgsContactSolver::sortConstraintByBatch3( b3Contact4* cs, int numConstraints, int simdWidth , int staticIdx, int numBodies, int* batchSizes)
+{
+	
+	B3_PROFILE("sortConstraintByBatch3");
+	
+	static int maxSwaps = 0;
+	int numSwaps = 0;
+
+	curUsed.resize(2*simdWidth);
+
+	static int maxNumConstraints = 0;
+	if (maxNumConstraints<numConstraints)
+	{
+		maxNumConstraints = numConstraints;
+		//printf("maxNumConstraints  = %d\n",maxNumConstraints );
+	}
+
+	int numUsedArray = numBodies/32+1;
+	bodyUsed.resize(numUsedArray);
+
+	for (int q=0;q<numUsedArray;q++)
+		bodyUsed[q]=0;
+
+	
+	int curBodyUsed = 0;
+
+	int numIter = 0;
+    
+	m_data->m_sortData.resize(0);
+	m_data->m_idxBuffer.resize(0);
+	m_data->m_old.resize(0);
+	
+		
+#if defined(_DEBUG)
+	for(int i=0; i<numConstraints; i++)
+		cs[i].getBatchIdx() = -1;
+#endif
+	
+	int numValidConstraints = 0;
+//	int unprocessedConstraintIndex = 0;
+
+	int batchIdx = 0;
+    
+
+	{
+		B3_PROFILE("cpu batch innerloop");
+		
+		while( numValidConstraints < numConstraints)
+		{
+			numIter++;
+			int nCurrentBatch = 0;
+			batchSizes[batchIdx] = 0;
+
+			//	clear flag
+			for(int i=0; i<curBodyUsed; i++) 
+				bodyUsed[curUsed[i]/32] = 0;
+
+            curBodyUsed = 0;
+
+			for(int i=numValidConstraints; i<numConstraints; i++)
+			{
+				int idx = i;
+				b3Assert( idx < numConstraints );
+				//	check if it can go
+				int bodyAS = cs[idx].m_bodyAPtrAndSignBit;
+				int bodyBS = cs[idx].m_bodyBPtrAndSignBit;
+				int bodyA = abs(bodyAS);
+				int bodyB = abs(bodyBS);
+				bool aIsStatic = (bodyAS<0) || bodyAS==staticIdx;
+				bool bIsStatic = (bodyBS<0) || bodyBS==staticIdx;
+				int aUnavailable = 0;
+				int bUnavailable = 0;
+				if (!aIsStatic)
+				{
+					aUnavailable = bodyUsed[ bodyA/32 ] & (1<<(bodyA&31));
+				}
+				if (!aUnavailable)
+				if (!bIsStatic)
+				{
+					bUnavailable = bodyUsed[ bodyB/32 ] & (1<<(bodyB&31));
+				}
+                
+				if( aUnavailable==0 && bUnavailable==0 ) // ok
+				{
+					if (!aIsStatic)
+					{
+						bodyUsed[ bodyA/32 ] |= (1<<(bodyA&31));
+						curUsed[curBodyUsed++]=bodyA;
+					}
+					if (!bIsStatic)
+					{
+						bodyUsed[ bodyB/32 ] |= (1<<(bodyB&31));
+						curUsed[curBodyUsed++]=bodyB;
+					}
+
+					cs[idx].getBatchIdx() = batchIdx;
+
+					if (i!=numValidConstraints)
+					{
+						b3Swap(cs[i],cs[numValidConstraints]);
+						numSwaps++;
+					}
+
+					numValidConstraints++;
+					{
+						nCurrentBatch++;
+						if( nCurrentBatch == simdWidth )
+						{
+							batchSizes[batchIdx] += simdWidth;
+							nCurrentBatch = 0;
+							for(int i=0; i<curBodyUsed; i++) 
+								bodyUsed[curUsed[i]/32] = 0;
+							curBodyUsed = 0;
+						}
+					}
+				}
+			}
+
+			if (batchIdx>=B3_MAX_NUM_BATCHES)
+			{
+				b3Error("batchIdx>=B3_MAX_NUM_BATCHES");
+				b3Assert(0);
+				break;
+			}
+
+			batchSizes[batchIdx] += nCurrentBatch;
+
+			batchIdx ++;
+			
+		}
+	}
+	
+#if defined(_DEBUG)
+    //		debugPrintf( "nBatches: %d\n", batchIdx );
+	for(int i=0; i<numConstraints; i++)
+    {
+        b3Assert( cs[i].getBatchIdx() != -1 );
+    }
+#endif
+
+	batchSizes[batchIdx] =0;
+	
+	if (maxSwaps<numSwaps)
+	{
+		maxSwaps = numSwaps;
+		//printf("maxSwaps = %d\n", maxSwaps);
+	}
+	
+	return batchIdx;
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuPgsContactSolver.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuPgsContactSolver.h
new file mode 100644
index 0000000000..98e2a5b8c4
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuPgsContactSolver.h
@@ -0,0 +1,43 @@
+
+#ifndef B3_GPU_BATCHING_PGS_SOLVER_H
+#define B3_GPU_BATCHING_PGS_SOLVER_H
+
+#include "Bullet3OpenCL/Initialize/b3OpenCLInclude.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3RigidBodyData.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3Contact4.h"
+#include "b3GpuConstraint4.h"
+
+class b3GpuPgsContactSolver
+{
+protected:
+
+	int m_debugOutput;
+
+	struct b3GpuBatchingPgsSolverInternalData*		m_data;
+
+	void batchContacts( b3OpenCLArray<b3Contact4>* contacts, int nContacts, b3OpenCLArray<unsigned int>* n, b3OpenCLArray<unsigned int>* offsets, int staticIdx );
+	
+	inline int sortConstraintByBatch( b3Contact4* cs, int n, int simdWidth , int staticIdx, int numBodies);
+	inline int sortConstraintByBatch2( b3Contact4* cs, int n, int simdWidth , int staticIdx, int numBodies);
+	inline int sortConstraintByBatch3( b3Contact4* cs, int n, int simdWidth , int staticIdx, int numBodies, int* batchSizes);
+	
+
+	
+	void solveContactConstraintBatchSizes(  const b3OpenCLArray<b3RigidBodyData>* bodyBuf, const b3OpenCLArray<b3InertiaData>* shapeBuf, 
+			b3OpenCLArray<b3GpuConstraint4>* constraint, void* additionalData, int n ,int maxNumBatches, int numIterations, const b3AlignedObjectArray<int>* batchSizes);//const b3OpenCLArray<int>* gpuBatchSizes);
+
+		void solveContactConstraint(  const b3OpenCLArray<b3RigidBodyData>* bodyBuf, const b3OpenCLArray<b3InertiaData>* shapeBuf, 
+			b3OpenCLArray<b3GpuConstraint4>* constraint, void* additionalData, int n ,int maxNumBatches, int numIterations, const b3AlignedObjectArray<int>* batchSizes);//const b3OpenCLArray<int>* gpuBatchSizes);
+
+public:
+	
+	b3GpuPgsContactSolver(cl_context ctx,cl_device_id device, cl_command_queue  q,int pairCapacity);
+	virtual ~b3GpuPgsContactSolver();
+
+	void solveContacts(int numBodies, cl_mem bodyBuf, cl_mem inertiaBuf, int numContacts, cl_mem contactBuf, const struct b3Config& config, int static0Index);
+
+};
+
+#endif //B3_GPU_BATCHING_PGS_SOLVER_H
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuRigidBodyPipeline.cpp b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuRigidBodyPipeline.cpp
new file mode 100644
index 0000000000..783e443060
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuRigidBodyPipeline.cpp
@@ -0,0 +1,708 @@
+/*
+Copyright (c) 2013 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+#include "b3GpuRigidBodyPipeline.h"
+#include "b3GpuRigidBodyPipelineInternalData.h"
+#include "kernels/integrateKernel.h"
+#include "kernels/updateAabbsKernel.h"
+
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+#include "b3GpuNarrowPhase.h"
+#include "Bullet3Geometry/b3AabbUtil.h"
+#include "Bullet3OpenCL/BroadphaseCollision/b3SapAabb.h"
+#include "Bullet3OpenCL/BroadphaseCollision/b3GpuBroadphaseInterface.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h"
+#include "Bullet3Dynamics/ConstraintSolver/b3PgsJacobiSolver.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3UpdateAabbs.h"
+#include "Bullet3Collision/BroadPhaseCollision/b3DynamicBvhBroadphase.h"
+
+//#define TEST_OTHER_GPU_SOLVER
+
+#define B3_RIGIDBODY_INTEGRATE_PATH "src/Bullet3OpenCL/RigidBody/kernels/integrateKernel.cl"
+#define B3_RIGIDBODY_UPDATEAABB_PATH "src/Bullet3OpenCL/RigidBody/kernels/updateAabbsKernel.cl"
+
+bool useBullet2CpuSolver = true;
+
+//choice of contact solver
+bool gUseJacobi = false;
+bool gUseDbvt = false;
+bool gDumpContactStats = false;
+bool gCalcWorldSpaceAabbOnCpu = false;
+bool gUseCalculateOverlappingPairsHost = false;
+bool gIntegrateOnCpu = false;
+bool gClearPairsOnGpu = true;
+
+#define TEST_OTHER_GPU_SOLVER 1
+#ifdef TEST_OTHER_GPU_SOLVER
+#include "b3GpuJacobiContactSolver.h"
+#endif //TEST_OTHER_GPU_SOLVER
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3RigidBodyData.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3Contact4.h"
+#include "Bullet3OpenCL/RigidBody/b3GpuPgsConstraintSolver.h"
+
+#include "b3GpuPgsContactSolver.h"
+#include "b3Solver.h"
+
+#include "Bullet3Collision/NarrowPhaseCollision/b3Config.h"
+#include "Bullet3OpenCL/Raycast/b3GpuRaycast.h"
+
+	
+#include "Bullet3Dynamics/shared/b3IntegrateTransforms.h"
+#include "Bullet3OpenCL/RigidBody/b3GpuNarrowPhaseInternalData.h"
+
+b3GpuRigidBodyPipeline::b3GpuRigidBodyPipeline(cl_context ctx,cl_device_id device, cl_command_queue  q,class b3GpuNarrowPhase* narrowphase, class b3GpuBroadphaseInterface* broadphaseSap , struct b3DynamicBvhBroadphase* broadphaseDbvt, const b3Config& config)
+{
+	m_data = new b3GpuRigidBodyPipelineInternalData;
+	m_data->m_constraintUid=0;
+	m_data->m_config = config;
+	m_data->m_context = ctx;
+	m_data->m_device = device;
+	m_data->m_queue = q;
+
+	m_data->m_solver = new b3PgsJacobiSolver(true);//new b3PgsJacobiSolver(true);
+	m_data->m_gpuSolver = new b3GpuPgsConstraintSolver(ctx,device,q,true);//new b3PgsJacobiSolver(true);
+	
+	m_data->m_allAabbsGPU = new b3OpenCLArray<b3SapAabb>(ctx,q,config.m_maxConvexBodies);
+	m_data->m_overlappingPairsGPU = new b3OpenCLArray<b3BroadphasePair>(ctx,q,config.m_maxBroadphasePairs);
+
+	m_data->m_gpuConstraints = new b3OpenCLArray<b3GpuGenericConstraint>(ctx,q);
+#ifdef TEST_OTHER_GPU_SOLVER
+	m_data->m_solver3 = new b3GpuJacobiContactSolver(ctx,device,q,config.m_maxBroadphasePairs);	
+#endif //	TEST_OTHER_GPU_SOLVER
+	
+	m_data->m_solver2 = new b3GpuPgsContactSolver(ctx,device,q,config.m_maxBroadphasePairs);
+
+	m_data->m_raycaster = new b3GpuRaycast(ctx,device,q);
+
+	
+	m_data->m_broadphaseDbvt = broadphaseDbvt;
+	m_data->m_broadphaseSap = broadphaseSap;
+	m_data->m_narrowphase = narrowphase;
+	m_data->m_gravity.setValue(0.f,-9.8f,0.f);
+
+	cl_int errNum=0;
+
+	{
+		cl_program prog = b3OpenCLUtils::compileCLProgramFromString(m_data->m_context,m_data->m_device,integrateKernelCL,&errNum,"",B3_RIGIDBODY_INTEGRATE_PATH);
+		b3Assert(errNum==CL_SUCCESS);
+		m_data->m_integrateTransformsKernel = b3OpenCLUtils::compileCLKernelFromString(m_data->m_context, m_data->m_device,integrateKernelCL, "integrateTransformsKernel",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+		clReleaseProgram(prog);
+	}
+	{
+		cl_program prog = b3OpenCLUtils::compileCLProgramFromString(m_data->m_context,m_data->m_device,updateAabbsKernelCL,&errNum,"",B3_RIGIDBODY_UPDATEAABB_PATH);
+		b3Assert(errNum==CL_SUCCESS);
+		m_data->m_updateAabbsKernel = b3OpenCLUtils::compileCLKernelFromString(m_data->m_context, m_data->m_device,updateAabbsKernelCL, "initializeGpuAabbsFull",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+
+
+		m_data->m_clearOverlappingPairsKernel = b3OpenCLUtils::compileCLKernelFromString(m_data->m_context, m_data->m_device,updateAabbsKernelCL, "clearOverlappingPairsKernel",&errNum,prog);
+		b3Assert(errNum==CL_SUCCESS);
+
+		clReleaseProgram(prog);
+	}
+
+
+}
+
+b3GpuRigidBodyPipeline::~b3GpuRigidBodyPipeline()
+{
+	if (m_data->m_integrateTransformsKernel)
+		clReleaseKernel(m_data->m_integrateTransformsKernel);
+	
+	if (m_data->m_updateAabbsKernel)
+		clReleaseKernel(m_data->m_updateAabbsKernel);
+	
+	if (m_data->m_clearOverlappingPairsKernel)
+		clReleaseKernel(m_data->m_clearOverlappingPairsKernel);
+	delete m_data->m_raycaster;
+	delete m_data->m_solver;
+	delete m_data->m_allAabbsGPU;
+	delete m_data->m_gpuConstraints;
+	delete m_data->m_overlappingPairsGPU;
+
+#ifdef TEST_OTHER_GPU_SOLVER
+	delete m_data->m_solver3;
+#endif //TEST_OTHER_GPU_SOLVER
+	
+	delete m_data->m_solver2;
+	
+	
+	delete m_data;
+}
+
+void	b3GpuRigidBodyPipeline::reset()
+{
+	m_data->m_gpuConstraints->resize(0);
+	m_data->m_cpuConstraints.resize(0);
+	m_data->m_allAabbsGPU->resize(0);
+	m_data->m_allAabbsCPU.resize(0);
+}
+
+void	b3GpuRigidBodyPipeline::addConstraint(b3TypedConstraint* constraint)
+{
+	m_data->m_joints.push_back(constraint);
+}
+
+void	b3GpuRigidBodyPipeline::removeConstraint(b3TypedConstraint* constraint)
+{
+	m_data->m_joints.remove(constraint);
+}
+
+
+
+void  b3GpuRigidBodyPipeline::removeConstraintByUid(int uid)
+{
+	m_data->m_gpuSolver->recomputeBatches();
+	//slow linear search
+	m_data->m_gpuConstraints->copyToHost(m_data->m_cpuConstraints);
+	//remove
+	for (int i=0;i<m_data->m_cpuConstraints.size();i++)
+	{
+		if (m_data->m_cpuConstraints[i].m_uid == uid)
+		{
+			//m_data->m_cpuConstraints.remove(m_data->m_cpuConstraints[i]);
+			m_data->m_cpuConstraints.swap(i,m_data->m_cpuConstraints.size()-1);
+			m_data->m_cpuConstraints.pop_back();
+
+			break;
+		}
+	}
+
+	if (m_data->m_cpuConstraints.size())
+	{
+		m_data->m_gpuConstraints->copyFromHost(m_data->m_cpuConstraints);
+	} else
+	{
+		m_data->m_gpuConstraints->resize(0);
+	}
+
+}
+int b3GpuRigidBodyPipeline::createPoint2PointConstraint(int bodyA, int bodyB, const float* pivotInA, const float* pivotInB,float breakingThreshold)
+{
+	m_data->m_gpuSolver->recomputeBatches();
+	b3GpuGenericConstraint c;
+	c.m_uid = m_data->m_constraintUid;
+	m_data->m_constraintUid++;
+	c.m_flags = B3_CONSTRAINT_FLAG_ENABLED;
+	c.m_rbA = bodyA;
+	c.m_rbB = bodyB;
+	c.m_pivotInA.setValue(pivotInA[0],pivotInA[1],pivotInA[2]);
+	c.m_pivotInB.setValue(pivotInB[0],pivotInB[1],pivotInB[2]);
+	c.m_breakingImpulseThreshold = breakingThreshold;
+	c.m_constraintType = B3_GPU_POINT2POINT_CONSTRAINT_TYPE;
+	m_data->m_cpuConstraints.push_back(c);
+	return c.m_uid;
+}
+int b3GpuRigidBodyPipeline::createFixedConstraint(int bodyA, int bodyB, const float* pivotInA, const float* pivotInB, const float* relTargetAB,float breakingThreshold)
+{
+	m_data->m_gpuSolver->recomputeBatches();
+	b3GpuGenericConstraint c;
+	c.m_uid = m_data->m_constraintUid;
+	m_data->m_constraintUid++;
+	c.m_flags = B3_CONSTRAINT_FLAG_ENABLED;
+	c.m_rbA = bodyA;
+	c.m_rbB = bodyB;
+	c.m_pivotInA.setValue(pivotInA[0],pivotInA[1],pivotInA[2]);
+	c.m_pivotInB.setValue(pivotInB[0],pivotInB[1],pivotInB[2]);
+	c.m_relTargetAB.setValue(relTargetAB[0],relTargetAB[1],relTargetAB[2],relTargetAB[3]);
+	c.m_breakingImpulseThreshold = breakingThreshold;
+	c.m_constraintType = B3_GPU_FIXED_CONSTRAINT_TYPE;
+
+	m_data->m_cpuConstraints.push_back(c);
+	return c.m_uid;
+}
+
+
+void	b3GpuRigidBodyPipeline::stepSimulation(float deltaTime)
+{
+
+	//update worldspace AABBs from local AABB/worldtransform
+	{
+		B3_PROFILE("setupGpuAabbs");
+		setupGpuAabbsFull();
+	}
+
+	int numPairs =0;
+
+	//compute overlapping pairs
+	{
+
+		if (gUseDbvt)
+		{
+			{
+				B3_PROFILE("setAabb");
+				m_data->m_allAabbsGPU->copyToHost(m_data->m_allAabbsCPU);
+				for (int i=0;i<m_data->m_allAabbsCPU.size();i++)
+				{
+					b3Vector3 aabbMin=b3MakeVector3(m_data->m_allAabbsCPU[i].m_min[0],m_data->m_allAabbsCPU[i].m_min[1],m_data->m_allAabbsCPU[i].m_min[2]);
+					b3Vector3 aabbMax=b3MakeVector3(m_data->m_allAabbsCPU[i].m_max[0],m_data->m_allAabbsCPU[i].m_max[1],m_data->m_allAabbsCPU[i].m_max[2]);
+					m_data->m_broadphaseDbvt->setAabb(i,aabbMin,aabbMax,0);
+				}
+			}
+
+			{
+				B3_PROFILE("calculateOverlappingPairs");
+				m_data->m_broadphaseDbvt->calculateOverlappingPairs();
+			}
+			numPairs = m_data->m_broadphaseDbvt->getOverlappingPairCache()->getNumOverlappingPairs();
+
+		} else
+		{
+			if (gUseCalculateOverlappingPairsHost)
+			{
+				m_data->m_broadphaseSap->calculateOverlappingPairsHost(m_data->m_config.m_maxBroadphasePairs);
+			} else
+			{
+				m_data->m_broadphaseSap->calculateOverlappingPairs(m_data->m_config.m_maxBroadphasePairs);
+			}
+			numPairs = m_data->m_broadphaseSap->getNumOverlap();
+		}
+	}
+
+	//compute contact points
+//	printf("numPairs=%d\n",numPairs);
+	
+	int numContacts  = 0;
+
+
+	int numBodies = m_data->m_narrowphase->getNumRigidBodies();
+
+	if (numPairs)
+	{
+		cl_mem pairs =0;
+		cl_mem aabbsWS =0;
+		if (gUseDbvt)
+		{
+			B3_PROFILE("m_overlappingPairsGPU->copyFromHost");
+			m_data->m_overlappingPairsGPU->copyFromHost(m_data->m_broadphaseDbvt->getOverlappingPairCache()->getOverlappingPairArray());
+			pairs = m_data->m_overlappingPairsGPU->getBufferCL();
+			aabbsWS = m_data->m_allAabbsGPU->getBufferCL();
+		} else
+		{
+			pairs = m_data->m_broadphaseSap->getOverlappingPairBuffer();
+			aabbsWS = m_data->m_broadphaseSap->getAabbBufferWS();
+		}
+
+		m_data->m_overlappingPairsGPU->resize(numPairs);
+
+		//mark the contacts for each pair as 'unused'
+		if (numPairs)
+		{
+			b3OpenCLArray<b3BroadphasePair> gpuPairs(this->m_data->m_context,m_data->m_queue);
+			gpuPairs.setFromOpenCLBuffer(pairs,numPairs);
+
+			if (gClearPairsOnGpu)
+			{
+				
+
+				//b3AlignedObjectArray<b3BroadphasePair> hostPairs;//just for debugging
+				//gpuPairs.copyToHost(hostPairs);
+
+				b3LauncherCL launcher(m_data->m_queue,m_data->m_clearOverlappingPairsKernel,"clearOverlappingPairsKernel");
+				launcher.setBuffer(pairs);
+				launcher.setConst(numPairs);
+				launcher.launch1D(numPairs);
+
+
+				//gpuPairs.copyToHost(hostPairs);
+			
+
+			} else
+			{
+				b3AlignedObjectArray<b3BroadphasePair> hostPairs;
+				gpuPairs.copyToHost(hostPairs);
+
+				for (int i=0;i<hostPairs.size();i++)
+				{
+					hostPairs[i].z = 0xffffffff;
+				}
+
+				gpuPairs.copyFromHost(hostPairs);
+			}
+		}
+
+		m_data->m_narrowphase->computeContacts(pairs,numPairs,aabbsWS,numBodies);
+		numContacts = m_data->m_narrowphase->getNumContactsGpu();
+
+		if (gUseDbvt)
+		{
+			///store the cached information (contact locations in the 'z' component)
+			B3_PROFILE("m_overlappingPairsGPU->copyToHost");
+			m_data->m_overlappingPairsGPU->copyToHost(m_data->m_broadphaseDbvt->getOverlappingPairCache()->getOverlappingPairArray());
+		}
+		if (gDumpContactStats && numContacts)
+		{
+			m_data->m_narrowphase->getContactsGpu();
+			
+			printf("numContacts = %d\n", numContacts);
+
+			int totalPoints  = 0;
+			const b3Contact4* contacts = m_data->m_narrowphase->getContactsCPU();
+
+			for (int i=0;i<numContacts;i++)
+			{
+				totalPoints += contacts->getNPoints();
+			}
+			printf("totalPoints=%d\n",totalPoints);
+
+		}
+	}
+	
+
+	//convert contact points to contact constraints
+	
+	//solve constraints
+
+	b3OpenCLArray<b3RigidBodyData> gpuBodies(m_data->m_context,m_data->m_queue,0,true);
+	gpuBodies.setFromOpenCLBuffer(m_data->m_narrowphase->getBodiesGpu(),m_data->m_narrowphase->getNumRigidBodies());
+	b3OpenCLArray<b3InertiaData> gpuInertias(m_data->m_context,m_data->m_queue,0,true);
+	gpuInertias.setFromOpenCLBuffer(m_data->m_narrowphase->getBodyInertiasGpu(),m_data->m_narrowphase->getNumRigidBodies());
+	b3OpenCLArray<b3Contact4> gpuContacts(m_data->m_context,m_data->m_queue,0,true);
+	gpuContacts.setFromOpenCLBuffer(m_data->m_narrowphase->getContactsGpu(),m_data->m_narrowphase->getNumContactsGpu());
+
+	int numJoints =  m_data->m_joints.size() ?  m_data->m_joints.size() : m_data->m_cpuConstraints.size();
+	if (useBullet2CpuSolver && numJoints)
+	{
+
+	//	b3AlignedObjectArray<b3Contact4> hostContacts;
+		//gpuContacts.copyToHost(hostContacts);
+		{
+			bool useGpu = m_data->m_joints.size()==0;
+
+//			b3Contact4* contacts = numContacts? &hostContacts[0]: 0;
+			//m_data->m_solver->solveContacts(m_data->m_narrowphase->getNumBodiesGpu(),&hostBodies[0],&hostInertias[0],numContacts,contacts,numJoints, joints);
+			if (useGpu)
+			{
+				m_data->m_gpuSolver->solveJoints(m_data->m_narrowphase->getNumRigidBodies(),&gpuBodies,&gpuInertias,numJoints, m_data->m_gpuConstraints);
+			} else
+			{
+				b3AlignedObjectArray<b3RigidBodyData> hostBodies;
+				gpuBodies.copyToHost(hostBodies);
+				b3AlignedObjectArray<b3InertiaData> hostInertias;
+				gpuInertias.copyToHost(hostInertias);
+
+				b3TypedConstraint** joints = numJoints? &m_data->m_joints[0] : 0;
+				m_data->m_solver->solveContacts(m_data->m_narrowphase->getNumRigidBodies(),&hostBodies[0],&hostInertias[0],0,0,numJoints, joints);
+				gpuBodies.copyFromHost(hostBodies);
+			}
+		}
+	}
+
+	if (numContacts)
+	{
+
+#ifdef TEST_OTHER_GPU_SOLVER
+		
+		if (gUseJacobi)
+		{
+			bool useGpu = true;
+			if (useGpu)
+			{
+
+				bool forceHost = false;
+				if (forceHost)
+				{
+					b3AlignedObjectArray<b3RigidBodyData> hostBodies;
+					b3AlignedObjectArray<b3InertiaData> hostInertias;
+					b3AlignedObjectArray<b3Contact4> hostContacts;
+				
+					{
+						B3_PROFILE("copyToHost");
+						gpuBodies.copyToHost(hostBodies);
+						gpuInertias.copyToHost(hostInertias);
+						gpuContacts.copyToHost(hostContacts);
+					}
+
+					{
+						b3JacobiSolverInfo solverInfo;
+						m_data->m_solver3->solveGroupHost(&hostBodies[0], &hostInertias[0], hostBodies.size(),&hostContacts[0],hostContacts.size(),solverInfo);
+
+						
+					}
+					{
+						B3_PROFILE("copyFromHost");
+						gpuBodies.copyFromHost(hostBodies);
+					}
+				} else
+
+
+				{
+					int static0Index = m_data->m_narrowphase->getStatic0Index();
+					b3JacobiSolverInfo solverInfo;
+					//m_data->m_solver3->solveContacts(    >solveGroup(&gpuBodies, &gpuInertias, &gpuContacts,solverInfo);
+					//m_data->m_solver3->solveContacts(m_data->m_narrowphase->getNumBodiesGpu(),&hostBodies[0],&hostInertias[0],numContacts,&hostContacts[0]);
+					m_data->m_solver3->solveContacts(numBodies, gpuBodies.getBufferCL(),gpuInertias.getBufferCL(),numContacts, gpuContacts.getBufferCL(),m_data->m_config, static0Index);
+				}
+			} else
+			{
+				b3AlignedObjectArray<b3RigidBodyData> hostBodies;
+				gpuBodies.copyToHost(hostBodies);
+				b3AlignedObjectArray<b3InertiaData> hostInertias;
+				gpuInertias.copyToHost(hostInertias);
+				b3AlignedObjectArray<b3Contact4> hostContacts;
+				gpuContacts.copyToHost(hostContacts);
+				{
+					//m_data->m_solver->solveContacts(m_data->m_narrowphase->getNumBodiesGpu(),&hostBodies[0],&hostInertias[0],numContacts,&hostContacts[0]);
+				}
+				gpuBodies.copyFromHost(hostBodies);
+			}
+		
+		} else
+#endif //TEST_OTHER_GPU_SOLVER
+		{
+			
+			int static0Index = m_data->m_narrowphase->getStatic0Index();
+			m_data->m_solver2->solveContacts(numBodies, gpuBodies.getBufferCL(),gpuInertias.getBufferCL(),numContacts, gpuContacts.getBufferCL(),m_data->m_config, static0Index);
+			
+			//m_data->m_solver4->solveContacts(m_data->m_narrowphase->getNumBodiesGpu(), gpuBodies.getBufferCL(), gpuInertias.getBufferCL(), numContacts, gpuContacts.getBufferCL());
+			
+			
+			/*m_data->m_solver3->solveContactConstraintHost(
+			(b3OpenCLArray<RigidBodyBase::Body>*)&gpuBodies,
+			(b3OpenCLArray<RigidBodyBase::Inertia>*)&gpuInertias,
+			(b3OpenCLArray<Constraint4>*) &gpuContacts,
+			0,numContacts,256);
+			*/
+		}
+	}
+
+	integrate(deltaTime);
+
+}
+
+
+void	b3GpuRigidBodyPipeline::integrate(float timeStep)
+{
+	//integrate
+	int numBodies = m_data->m_narrowphase->getNumRigidBodies();
+	float angularDamp = 0.99f;
+
+	if (gIntegrateOnCpu)
+	{
+		if(numBodies)
+		{
+			b3GpuNarrowPhaseInternalData*	npData = m_data->m_narrowphase->getInternalData();
+			npData->m_bodyBufferGPU->copyToHost(*npData->m_bodyBufferCPU);
+
+			b3RigidBodyData_t* bodies = &npData->m_bodyBufferCPU->at(0);
+
+			for (int nodeID=0;nodeID<numBodies;nodeID++)
+			{
+				integrateSingleTransform( bodies,nodeID, timeStep, angularDamp, m_data->m_gravity);
+			}
+			npData->m_bodyBufferGPU->copyFromHost(*npData->m_bodyBufferCPU);
+		}
+	} else
+	{
+		b3LauncherCL launcher(m_data->m_queue,m_data->m_integrateTransformsKernel,"m_integrateTransformsKernel");
+		launcher.setBuffer(m_data->m_narrowphase->getBodiesGpu());
+		
+		launcher.setConst(numBodies);
+		launcher.setConst(timeStep);
+		launcher.setConst(angularDamp);
+		launcher.setConst(m_data->m_gravity);
+		launcher.launch1D(numBodies);
+	}
+}
+
+
+
+
+void	b3GpuRigidBodyPipeline::setupGpuAabbsFull()
+{
+	cl_int ciErrNum=0;
+
+	int numBodies = m_data->m_narrowphase->getNumRigidBodies();
+	if (!numBodies)
+		return;
+
+	if (gCalcWorldSpaceAabbOnCpu)
+	{
+		
+		if (numBodies)
+		{
+			if (gUseDbvt)
+			{
+				m_data->m_allAabbsCPU.resize(numBodies);
+				m_data->m_narrowphase->readbackAllBodiesToCpu();
+				for (int i=0;i<numBodies;i++)
+				{
+					b3ComputeWorldAabb(  i, m_data->m_narrowphase->getBodiesCpu(), m_data->m_narrowphase->getCollidablesCpu(), m_data->m_narrowphase->getLocalSpaceAabbsCpu(),&m_data->m_allAabbsCPU[0]);
+				}
+				m_data->m_allAabbsGPU->copyFromHost(m_data->m_allAabbsCPU);
+			} else
+			{
+				m_data->m_broadphaseSap->getAllAabbsCPU().resize(numBodies);
+				m_data->m_narrowphase->readbackAllBodiesToCpu();
+				for (int i=0;i<numBodies;i++)
+				{
+					b3ComputeWorldAabb(  i, m_data->m_narrowphase->getBodiesCpu(), m_data->m_narrowphase->getCollidablesCpu(), m_data->m_narrowphase->getLocalSpaceAabbsCpu(),&m_data->m_broadphaseSap->getAllAabbsCPU()[0]);
+				}
+				m_data->m_broadphaseSap->getAllAabbsGPU().copyFromHost(m_data->m_broadphaseSap->getAllAabbsCPU());
+				//m_data->m_broadphaseSap->writeAabbsToGpu();
+			}
+		}
+	} else
+	{
+		//__kernel void initializeGpuAabbsFull(  const int numNodes, __global Body* gBodies,__global Collidable* collidables, __global b3AABBCL* plocalShapeAABB, __global b3AABBCL* pAABB)
+		b3LauncherCL launcher(m_data->m_queue,m_data->m_updateAabbsKernel,"m_updateAabbsKernel");
+		launcher.setConst(numBodies);
+		cl_mem bodies = m_data->m_narrowphase->getBodiesGpu();
+		launcher.setBuffer(bodies);
+		cl_mem collidables = m_data->m_narrowphase->getCollidablesGpu();
+		launcher.setBuffer(collidables);
+		cl_mem localAabbs = m_data->m_narrowphase->getAabbLocalSpaceBufferGpu();
+		launcher.setBuffer(localAabbs);
+
+		cl_mem worldAabbs =0;
+		if (gUseDbvt)
+		{
+			worldAabbs = m_data->m_allAabbsGPU->getBufferCL();
+		} else
+		{
+			worldAabbs = m_data->m_broadphaseSap->getAabbBufferWS();
+		}
+		launcher.setBuffer(worldAabbs);
+		launcher.launch1D(numBodies);
+	
+		oclCHECKERROR(ciErrNum, CL_SUCCESS);
+	}
+
+	/*
+	b3AlignedObjectArray<b3SapAabb> aabbs;
+	m_data->m_broadphaseSap->m_allAabbsGPU.copyToHost(aabbs);
+
+	printf("numAabbs = %d\n",  aabbs.size());
+
+	for (int i=0;i<aabbs.size();i++)
+	{
+		printf("aabb[%d].m_min=%f,%f,%f,%d\n",i,aabbs[i].m_minVec[0],aabbs[i].m_minVec[1],aabbs[i].m_minVec[2],aabbs[i].m_minIndices[3]);
+		printf("aabb[%d].m_max=%f,%f,%f,%d\n",i,aabbs[i].m_maxVec[0],aabbs[i].m_maxVec[1],aabbs[i].m_maxVec[2],aabbs[i].m_signedMaxIndices[3]);
+
+	};
+	*/
+
+	
+
+
+
+}
+
+
+
+cl_mem	b3GpuRigidBodyPipeline::getBodyBuffer()
+{
+	return m_data->m_narrowphase->getBodiesGpu();
+}
+
+int	b3GpuRigidBodyPipeline::getNumBodies() const
+{
+	return m_data->m_narrowphase->getNumRigidBodies();
+}
+
+void	b3GpuRigidBodyPipeline::setGravity(const float* grav)
+{
+	m_data->m_gravity.setValue(grav[0],grav[1],grav[2]);
+}
+
+void 		b3GpuRigidBodyPipeline::copyConstraintsToHost()
+{
+	m_data->m_gpuConstraints->copyToHost(m_data->m_cpuConstraints);
+}
+
+void 		b3GpuRigidBodyPipeline::writeAllInstancesToGpu()
+{
+	m_data->m_allAabbsGPU->copyFromHost(m_data->m_allAabbsCPU);
+	m_data->m_gpuConstraints->copyFromHost(m_data->m_cpuConstraints);
+}
+
+
+int		b3GpuRigidBodyPipeline::registerPhysicsInstance(float mass, const float* position, const float* orientation, int collidableIndex, int userIndex, bool writeInstanceToGpu)
+{
+	
+	b3Vector3 aabbMin=b3MakeVector3(0,0,0),aabbMax=b3MakeVector3(0,0,0);
+
+	
+	if (collidableIndex>=0)
+	{
+		b3SapAabb localAabb = m_data->m_narrowphase->getLocalSpaceAabb(collidableIndex);
+		b3Vector3 localAabbMin=b3MakeVector3(localAabb.m_min[0],localAabb.m_min[1],localAabb.m_min[2]);
+		b3Vector3 localAabbMax=b3MakeVector3(localAabb.m_max[0],localAabb.m_max[1],localAabb.m_max[2]);
+		
+		b3Scalar margin = 0.01f;
+		b3Transform t;
+		t.setIdentity();
+		t.setOrigin(b3MakeVector3(position[0],position[1],position[2]));
+		t.setRotation(b3Quaternion(orientation[0],orientation[1],orientation[2],orientation[3]));
+		b3TransformAabb(localAabbMin,localAabbMax, margin,t,aabbMin,aabbMax);
+	} else
+	{
+		b3Error("registerPhysicsInstance using invalid collidableIndex\n");
+		return -1;
+	}
+			
+	
+	bool writeToGpu = false;
+	int bodyIndex = m_data->m_narrowphase->getNumRigidBodies();
+	bodyIndex = m_data->m_narrowphase->registerRigidBody(collidableIndex,mass,position,orientation,&aabbMin.getX(),&aabbMax.getX(),writeToGpu);
+
+	if (bodyIndex>=0)
+	{
+		if (gUseDbvt)
+		{
+			m_data->m_broadphaseDbvt->createProxy(aabbMin,aabbMax,bodyIndex,0,1,1);
+			b3SapAabb aabb;
+			for (int i=0;i<3;i++)
+			{
+				aabb.m_min[i] = aabbMin[i];
+				aabb.m_max[i] = aabbMax[i];
+				aabb.m_minIndices[3] = bodyIndex;
+			}
+			m_data->m_allAabbsCPU.push_back(aabb);
+			if (writeInstanceToGpu)
+			{
+				m_data->m_allAabbsGPU->copyFromHost(m_data->m_allAabbsCPU);
+			}
+		} else
+		{
+			if (mass)
+			{
+				m_data->m_broadphaseSap->createProxy(aabbMin,aabbMax,bodyIndex,1,1);//m_dispatcher);
+			} else
+			{
+				m_data->m_broadphaseSap->createLargeProxy(aabbMin,aabbMax,bodyIndex,1,1);//m_dispatcher);	
+			}
+		}
+	}
+
+	/*
+	if (mass>0.f)
+		m_numDynamicPhysicsInstances++;
+
+	m_numPhysicsInstances++;
+	*/
+
+	return bodyIndex;
+}
+
+void	b3GpuRigidBodyPipeline::castRays(const b3AlignedObjectArray<b3RayInfo>& rays,	b3AlignedObjectArray<b3RayHit>& hitResults)
+{
+	this->m_data->m_raycaster->castRays(rays,hitResults,
+		getNumBodies(),this->m_data->m_narrowphase->getBodiesCpu(),
+		m_data->m_narrowphase->getNumCollidablesGpu(), m_data->m_narrowphase->getCollidablesCpu(),
+		m_data->m_narrowphase->getInternalData(), m_data->m_broadphaseSap);
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuRigidBodyPipeline.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuRigidBodyPipeline.h
new file mode 100644
index 0000000000..b4eac6841a
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuRigidBodyPipeline.h
@@ -0,0 +1,74 @@
+/*
+Copyright (c) 2013 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+#ifndef B3_GPU_RIGIDBODY_PIPELINE_H
+#define B3_GPU_RIGIDBODY_PIPELINE_H
+
+#include "Bullet3OpenCL/Initialize/b3OpenCLInclude.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3Config.h"
+
+#include "Bullet3Common/b3AlignedObjectArray.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3RaycastInfo.h"
+
+class b3GpuRigidBodyPipeline
+{
+protected:
+	struct b3GpuRigidBodyPipelineInternalData*	m_data;
+
+	int allocateCollidable();
+
+public:
+
+
+	b3GpuRigidBodyPipeline(cl_context ctx,cl_device_id device, cl_command_queue  q , class b3GpuNarrowPhase* narrowphase, class b3GpuBroadphaseInterface* broadphaseSap, struct b3DynamicBvhBroadphase* broadphaseDbvt, const b3Config& config);
+	virtual ~b3GpuRigidBodyPipeline();
+
+	void	stepSimulation(float deltaTime);
+	void	integrate(float timeStep);
+	void	setupGpuAabbsFull();
+
+	int		registerConvexPolyhedron(class b3ConvexUtility* convex);
+
+	//int		registerConvexPolyhedron(const float* vertices, int strideInBytes, int numVertices, const float* scaling);
+	//int		registerSphereShape(float radius);
+	//int		registerPlaneShape(const b3Vector3& planeNormal, float planeConstant);
+	
+	//int		registerConcaveMesh(b3AlignedObjectArray<b3Vector3>* vertices, b3AlignedObjectArray<int>* indices, const float* scaling);
+	//int		registerCompoundShape(b3AlignedObjectArray<b3GpuChildShape>* childShapes);
+
+	
+	int		registerPhysicsInstance(float mass, const float* position, const float* orientation, int collisionShapeIndex, int userData, bool writeInstanceToGpu);
+	//if you passed "writeInstanceToGpu" false in the registerPhysicsInstance method (for performance) you need to call writeAllInstancesToGpu after all instances are registered
+	void	writeAllInstancesToGpu();
+	void	copyConstraintsToHost();
+	void	setGravity(const float* grav);
+	void reset();
+	
+	int createPoint2PointConstraint(int bodyA, int bodyB, const float* pivotInA, const float* pivotInB,float breakingThreshold);
+	int createFixedConstraint(int bodyA, int bodyB, const float* pivotInA, const float* pivotInB, const float* relTargetAB, float breakingThreshold);
+	void removeConstraintByUid(int uid);
+
+	void	addConstraint(class b3TypedConstraint* constraint);
+	void	removeConstraint(b3TypedConstraint* constraint);
+
+	void	castRays(const b3AlignedObjectArray<b3RayInfo>& rays,	b3AlignedObjectArray<b3RayHit>& hitResults);
+
+	cl_mem	getBodyBuffer();
+
+	int	getNumBodies() const;
+
+};
+
+#endif //B3_GPU_RIGIDBODY_PIPELINE_H
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuRigidBodyPipelineInternalData.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuRigidBodyPipelineInternalData.h
new file mode 100644
index 0000000000..5ac92f97d6
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuRigidBodyPipelineInternalData.h
@@ -0,0 +1,73 @@
+/*
+Copyright (c) 2013 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+#ifndef B3_GPU_RIGIDBODY_PIPELINE_INTERNAL_DATA_H
+#define B3_GPU_RIGIDBODY_PIPELINE_INTERNAL_DATA_H
+
+#include "Bullet3OpenCL/Initialize/b3OpenCLInclude.h"
+#include "Bullet3Common/b3AlignedObjectArray.h"
+
+#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Collidable.h"
+
+
+#include "Bullet3OpenCL/BroadphaseCollision/b3SapAabb.h"
+#include "Bullet3Dynamics/ConstraintSolver/b3TypedConstraint.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3Config.h"
+
+
+
+#include "Bullet3Collision/BroadPhaseCollision/b3OverlappingPair.h"
+#include "Bullet3OpenCL/RigidBody/b3GpuGenericConstraint.h"
+
+struct b3GpuRigidBodyPipelineInternalData
+{
+
+	cl_context			m_context;
+	cl_device_id		m_device;
+	cl_command_queue	m_queue;
+
+	cl_kernel	m_integrateTransformsKernel;
+	cl_kernel	m_updateAabbsKernel;
+	cl_kernel	m_clearOverlappingPairsKernel;
+	
+	class b3PgsJacobiSolver* m_solver;
+	
+	class b3GpuPgsConstraintSolver* m_gpuSolver;
+
+	class b3GpuPgsContactSolver* m_solver2;
+	class b3GpuJacobiContactSolver* m_solver3;
+	class b3GpuRaycast* m_raycaster;
+	
+	class b3GpuBroadphaseInterface* m_broadphaseSap;
+	
+	struct b3DynamicBvhBroadphase* m_broadphaseDbvt;
+	b3OpenCLArray<b3SapAabb>*	m_allAabbsGPU;
+	b3AlignedObjectArray<b3SapAabb>	m_allAabbsCPU;
+	b3OpenCLArray<b3BroadphasePair>*		m_overlappingPairsGPU;
+
+	b3OpenCLArray<b3GpuGenericConstraint>* m_gpuConstraints;
+	b3AlignedObjectArray<b3GpuGenericConstraint> m_cpuConstraints;
+
+	b3AlignedObjectArray<b3TypedConstraint*> m_joints;
+	int	m_constraintUid;
+	class b3GpuNarrowPhase*	m_narrowphase;
+	b3Vector3	m_gravity;
+
+	b3Config	m_config;
+};
+
+#endif //B3_GPU_RIGIDBODY_PIPELINE_INTERNAL_DATA_H
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuSolverBody.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuSolverBody.h
new file mode 100644
index 0000000000..f2a61801ac
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuSolverBody.h
@@ -0,0 +1,228 @@
+/*
+Copyright (c) 2013 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+
+#ifndef B3_GPU_SOLVER_BODY_H
+#define B3_GPU_SOLVER_BODY_H
+
+
+#include "Bullet3Common/b3Vector3.h"
+#include "Bullet3Common/b3Matrix3x3.h"
+
+#include "Bullet3Common/b3AlignedAllocator.h"
+#include "Bullet3Common/b3TransformUtil.h"
+
+///Until we get other contributions, only use SIMD on Windows, when using Visual Studio 2008 or later, and not double precision
+#ifdef B3_USE_SSE
+#define USE_SIMD 1
+#endif //
+
+
+
+///The b3SolverBody is an internal datastructure for the constraint solver. Only necessary data is packed to increase cache coherence/performance.
+B3_ATTRIBUTE_ALIGNED16 (struct)	b3GpuSolverBody
+{
+	B3_DECLARE_ALIGNED_ALLOCATOR();
+//	b3Transform		m_worldTransformUnused;
+	b3Vector3		m_deltaLinearVelocity;
+	b3Vector3		m_deltaAngularVelocity;
+	b3Vector3		m_angularFactor;
+	b3Vector3		m_linearFactor;
+	b3Vector3		m_invMass;
+	b3Vector3		m_pushVelocity;
+	b3Vector3		m_turnVelocity;
+	b3Vector3		m_linearVelocity;
+	b3Vector3		m_angularVelocity;
+
+	union 
+	{
+		void*	m_originalBody;
+		int		m_originalBodyIndex;
+	};
+
+	int padding[3];
+
+	/*
+	void	setWorldTransform(const b3Transform& worldTransform)
+	{
+		m_worldTransform = worldTransform;
+	}
+
+	const b3Transform& getWorldTransform() const
+	{
+		return m_worldTransform;
+	}
+	*/
+	B3_FORCE_INLINE void	getVelocityInLocalPointObsolete(const b3Vector3& rel_pos, b3Vector3& velocity ) const
+	{
+		if (m_originalBody)
+			velocity = m_linearVelocity+m_deltaLinearVelocity + (m_angularVelocity+m_deltaAngularVelocity).cross(rel_pos);
+		else
+			velocity.setValue(0,0,0);
+	}
+
+	B3_FORCE_INLINE void	getAngularVelocity(b3Vector3& angVel) const
+	{
+		if (m_originalBody)
+			angVel =m_angularVelocity+m_deltaAngularVelocity;
+		else
+			angVel.setValue(0,0,0);
+	}
+
+
+	//Optimization for the iterative solver: avoid calculating constant terms involving inertia, normal, relative position
+	B3_FORCE_INLINE void applyImpulse(const b3Vector3& linearComponent, const b3Vector3& angularComponent,const b3Scalar impulseMagnitude)
+	{
+		if (m_originalBody)
+		{
+			m_deltaLinearVelocity += linearComponent*impulseMagnitude*m_linearFactor;
+			m_deltaAngularVelocity += angularComponent*(impulseMagnitude*m_angularFactor);
+		}
+	}
+
+	B3_FORCE_INLINE void internalApplyPushImpulse(const b3Vector3& linearComponent, const b3Vector3& angularComponent,b3Scalar impulseMagnitude)
+	{
+		if (m_originalBody)
+		{
+			m_pushVelocity += linearComponent*impulseMagnitude*m_linearFactor;
+			m_turnVelocity += angularComponent*(impulseMagnitude*m_angularFactor);
+		}
+	}
+
+
+
+	const b3Vector3& getDeltaLinearVelocity() const
+	{
+		return m_deltaLinearVelocity;
+	}
+
+	const b3Vector3& getDeltaAngularVelocity() const
+	{
+		return m_deltaAngularVelocity;
+	}
+
+	const b3Vector3& getPushVelocity() const 
+	{
+		return m_pushVelocity;
+	}
+
+	const b3Vector3& getTurnVelocity() const 
+	{
+		return m_turnVelocity;
+	}
+
+
+	////////////////////////////////////////////////
+	///some internal methods, don't use them
+		
+	b3Vector3& internalGetDeltaLinearVelocity()
+	{
+		return m_deltaLinearVelocity;
+	}
+
+	b3Vector3& internalGetDeltaAngularVelocity()
+	{
+		return m_deltaAngularVelocity;
+	}
+
+	const b3Vector3& internalGetAngularFactor() const
+	{
+		return m_angularFactor;
+	}
+
+	const b3Vector3& internalGetInvMass() const
+	{
+		return m_invMass;
+	}
+
+	void internalSetInvMass(const b3Vector3& invMass)
+	{
+		m_invMass = invMass;
+	}
+	
+	b3Vector3& internalGetPushVelocity()
+	{
+		return m_pushVelocity;
+	}
+
+	b3Vector3& internalGetTurnVelocity()
+	{
+		return m_turnVelocity;
+	}
+
+	B3_FORCE_INLINE void	internalGetVelocityInLocalPointObsolete(const b3Vector3& rel_pos, b3Vector3& velocity ) const
+	{
+		velocity = m_linearVelocity+m_deltaLinearVelocity + (m_angularVelocity+m_deltaAngularVelocity).cross(rel_pos);
+	}
+
+	B3_FORCE_INLINE void	internalGetAngularVelocity(b3Vector3& angVel) const
+	{
+		angVel = m_angularVelocity+m_deltaAngularVelocity;
+	}
+
+
+	//Optimization for the iterative solver: avoid calculating constant terms involving inertia, normal, relative position
+	B3_FORCE_INLINE void internalApplyImpulse(const b3Vector3& linearComponent, const b3Vector3& angularComponent,const b3Scalar impulseMagnitude)
+	{
+		//if (m_originalBody)
+		{
+			m_deltaLinearVelocity += linearComponent*impulseMagnitude*m_linearFactor;
+			m_deltaAngularVelocity += angularComponent*(impulseMagnitude*m_angularFactor);
+		}
+	}
+		
+	
+	
+
+	void	writebackVelocity()
+	{
+		//if (m_originalBody>=0)
+		{
+			m_linearVelocity +=m_deltaLinearVelocity;
+			m_angularVelocity += m_deltaAngularVelocity;
+			
+			//m_originalBody->setCompanionId(-1);
+		}
+	}
+
+
+	void	writebackVelocityAndTransform(b3Scalar timeStep, b3Scalar splitImpulseTurnErp)
+	{
+        (void) timeStep;
+		if (m_originalBody)
+		{
+			m_linearVelocity += m_deltaLinearVelocity;
+			m_angularVelocity += m_deltaAngularVelocity;
+			
+			//correct the position/orientation based on push/turn recovery
+			b3Transform newTransform;
+			if (m_pushVelocity[0]!=0.f || m_pushVelocity[1]!=0 || m_pushVelocity[2]!=0 || m_turnVelocity[0]!=0.f || m_turnVelocity[1]!=0 || m_turnVelocity[2]!=0)
+			{
+			//	b3Quaternion orn = m_worldTransform.getRotation();
+//				b3TransformUtil::integrateTransform(m_worldTransform,m_pushVelocity,m_turnVelocity*splitImpulseTurnErp,timeStep,newTransform);
+//				m_worldTransform = newTransform;
+			}
+			//m_worldTransform.setRotation(orn);
+			//m_originalBody->setCompanionId(-1);
+		}
+	}
+	
+
+
+};
+
+#endif //B3_SOLVER_BODY_H
+
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuSolverConstraint.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuSolverConstraint.h
new file mode 100644
index 0000000000..60d235baab
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3GpuSolverConstraint.h
@@ -0,0 +1,82 @@
+/*
+Bullet Continuous Collision Detection and Physics Library
+Copyright (c) 2013 Erwin Coumans http://github.com/erwincoumans/bullet3
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+
+
+#ifndef B3_GPU_SOLVER_CONSTRAINT_H
+#define B3_GPU_SOLVER_CONSTRAINT_H
+
+
+#include "Bullet3Common/b3Vector3.h"
+#include "Bullet3Common/b3Matrix3x3.h"
+//#include "b3JacobianEntry.h"
+#include "Bullet3Common/b3AlignedObjectArray.h"
+
+//#define NO_FRICTION_TANGENTIALS 1
+
+
+
+///1D constraint along a normal axis between bodyA and bodyB. It can be combined to solve contact and friction constraints.
+B3_ATTRIBUTE_ALIGNED16 (struct)	b3GpuSolverConstraint
+{
+	B3_DECLARE_ALIGNED_ALLOCATOR();
+
+	b3Vector3		m_relpos1CrossNormal;
+	b3Vector3		m_contactNormal;
+
+	b3Vector3		m_relpos2CrossNormal;
+	//b3Vector3		m_contactNormal2;//usually m_contactNormal2 == -m_contactNormal
+
+	b3Vector3		m_angularComponentA;
+	b3Vector3		m_angularComponentB;
+	
+	mutable b3Scalar	m_appliedPushImpulse;
+	mutable b3Scalar	m_appliedImpulse;
+	int m_padding1;
+	int m_padding2;
+	b3Scalar	m_friction;
+	b3Scalar	m_jacDiagABInv;
+	b3Scalar		m_rhs;
+	b3Scalar		m_cfm;
+	
+    b3Scalar		m_lowerLimit;
+	b3Scalar		m_upperLimit;
+	b3Scalar		m_rhsPenetration;
+    union
+	{
+		void*		m_originalContactPoint;
+		int		m_originalConstraintIndex;
+		b3Scalar	m_unusedPadding4;
+	};
+
+	int	m_overrideNumSolverIterations;
+    int			m_frictionIndex;
+	int m_solverBodyIdA;
+	int m_solverBodyIdB;
+
+    
+	enum		b3SolverConstraintType
+	{
+		B3_SOLVER_CONTACT_1D = 0,
+		B3_SOLVER_FRICTION_1D
+	};
+};
+
+typedef b3AlignedObjectArray<b3GpuSolverConstraint>	b3GpuConstraintArray;
+
+
+#endif //B3_GPU_SOLVER_CONSTRAINT_H
+
+
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3Solver.cpp b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3Solver.cpp
new file mode 100644
index 0000000000..20bf6d47c5
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3Solver.cpp
@@ -0,0 +1,1225 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+#include "b3Solver.h"
+
+///useNewBatchingKernel  is a rewritten kernel using just a single thread of the warp, for experiments
+bool useNewBatchingKernel = true;
+bool gConvertConstraintOnCpu = false;
+
+#define B3_SOLVER_SETUP_KERNEL_PATH "src/Bullet3OpenCL/RigidBody/kernels/solverSetup.cl"
+#define B3_SOLVER_SETUP2_KERNEL_PATH "src/Bullet3OpenCL/RigidBody/kernels/solverSetup2.cl"
+#define B3_SOLVER_CONTACT_KERNEL_PATH "src/Bullet3OpenCL/RigidBody/kernels/solveContact.cl"
+#define B3_SOLVER_FRICTION_KERNEL_PATH "src/Bullet3OpenCL/RigidBody/kernels/solveFriction.cl"
+#define B3_BATCHING_PATH "src/Bullet3OpenCL/RigidBody/kernels/batchingKernels.cl"
+#define B3_BATCHING_NEW_PATH "src/Bullet3OpenCL/RigidBody/kernels/batchingKernelsNew.cl"
+
+#include "Bullet3Dynamics/shared/b3ConvertConstraint4.h"
+
+#include "kernels/solverSetup.h"
+#include "kernels/solverSetup2.h"
+
+#include "kernels/solveContact.h"
+#include "kernels/solveFriction.h"
+
+#include "kernels/batchingKernels.h"
+#include "kernels/batchingKernelsNew.h"
+
+
+#include "Bullet3OpenCL/ParallelPrimitives/b3LauncherCL.h"
+#include "Bullet3Common/b3Vector3.h"
+
+struct SolverDebugInfo
+{
+	int m_valInt0;
+	int m_valInt1;
+	int m_valInt2;
+	int m_valInt3;
+	
+	int m_valInt4;
+	int m_valInt5;
+	int m_valInt6;
+	int m_valInt7;
+
+	int m_valInt8;
+	int m_valInt9;
+	int m_valInt10;
+	int m_valInt11;
+
+	int	m_valInt12;
+	int	m_valInt13;
+	int	m_valInt14;
+	int	m_valInt15;
+
+
+	float m_val0;
+	float m_val1;
+	float m_val2;
+	float m_val3;
+};
+
+
+
+
+class SolverDeviceInl
+{
+public:
+	struct ParallelSolveData
+	{
+		b3OpenCLArray<unsigned int>* m_numConstraints;
+		b3OpenCLArray<unsigned int>* m_offsets;
+	};
+};
+
+
+
+b3Solver::b3Solver(cl_context ctx, cl_device_id device, cl_command_queue queue, int pairCapacity)
+			:
+			m_context(ctx),
+			m_device(device),
+			m_queue(queue),
+			m_batchSizes(ctx,queue),
+			m_nIterations(4)
+{
+	m_sort32 = new b3RadixSort32CL(ctx,device,queue);
+	m_scan = new b3PrefixScanCL(ctx,device,queue,B3_SOLVER_N_CELLS);
+	m_search = new b3BoundSearchCL(ctx,device,queue,B3_SOLVER_N_CELLS);
+
+	const int sortSize = B3NEXTMULTIPLEOF( pairCapacity, 512 );
+
+	m_sortDataBuffer = new b3OpenCLArray<b3SortData>(ctx,queue,sortSize);
+	m_contactBuffer2 = new b3OpenCLArray<b3Contact4>(ctx,queue);
+
+	m_numConstraints = new b3OpenCLArray<unsigned int>(ctx,queue,B3_SOLVER_N_CELLS );
+	m_numConstraints->resize(B3_SOLVER_N_CELLS);
+
+	m_offsets = new b3OpenCLArray<unsigned int>( ctx,queue,B3_SOLVER_N_CELLS);
+	m_offsets->resize(B3_SOLVER_N_CELLS);
+	const char* additionalMacros = "";
+//	const char* srcFileNameForCaching="";
+
+
+
+	cl_int pErrNum;
+	const char* batchKernelSource = batchingKernelsCL;
+	const char* batchKernelNewSource = batchingKernelsNewCL;
+	
+	const char* solverSetupSource = solverSetupCL;
+	const char* solverSetup2Source = solverSetup2CL;
+	const char* solveContactSource = solveContactCL;
+	const char* solveFrictionSource = solveFrictionCL;
+	
+	
+	
+	{
+		
+		cl_program solveContactProg= b3OpenCLUtils::compileCLProgramFromString( ctx, device, solveContactSource, &pErrNum,additionalMacros, B3_SOLVER_CONTACT_KERNEL_PATH);
+		b3Assert(solveContactProg);
+		
+		cl_program solveFrictionProg= b3OpenCLUtils::compileCLProgramFromString( ctx, device, solveFrictionSource, &pErrNum,additionalMacros, B3_SOLVER_FRICTION_KERNEL_PATH);
+		b3Assert(solveFrictionProg);
+
+		cl_program solverSetup2Prog= b3OpenCLUtils::compileCLProgramFromString( ctx, device, solverSetup2Source, &pErrNum,additionalMacros, B3_SOLVER_SETUP2_KERNEL_PATH);
+		b3Assert(solverSetup2Prog);
+
+		
+		cl_program solverSetupProg= b3OpenCLUtils::compileCLProgramFromString( ctx, device, solverSetupSource, &pErrNum,additionalMacros, B3_SOLVER_SETUP_KERNEL_PATH);
+		b3Assert(solverSetupProg);
+		
+		
+		m_solveFrictionKernel= b3OpenCLUtils::compileCLKernelFromString( ctx, device, solveFrictionSource, "BatchSolveKernelFriction", &pErrNum, solveFrictionProg,additionalMacros );
+		b3Assert(m_solveFrictionKernel);
+
+		m_solveContactKernel= b3OpenCLUtils::compileCLKernelFromString( ctx, device, solveContactSource, "BatchSolveKernelContact", &pErrNum, solveContactProg,additionalMacros );
+		b3Assert(m_solveContactKernel);
+		
+		m_contactToConstraintKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverSetupSource, "ContactToConstraintKernel", &pErrNum, solverSetupProg,additionalMacros );
+		b3Assert(m_contactToConstraintKernel);
+			
+		m_setSortDataKernel =  b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverSetup2Source, "SetSortDataKernel", &pErrNum, solverSetup2Prog,additionalMacros );
+		b3Assert(m_setSortDataKernel);
+				
+		m_reorderContactKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverSetup2Source, "ReorderContactKernel", &pErrNum, solverSetup2Prog,additionalMacros );
+		b3Assert(m_reorderContactKernel);
+		
+
+		m_copyConstraintKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, solverSetup2Source, "CopyConstraintKernel", &pErrNum, solverSetup2Prog,additionalMacros );
+		b3Assert(m_copyConstraintKernel);
+		
+	}
+
+	{
+		cl_program batchingProg = b3OpenCLUtils::compileCLProgramFromString( ctx, device, batchKernelSource, &pErrNum,additionalMacros, B3_BATCHING_PATH);
+		//cl_program batchingProg = b3OpenCLUtils::compileCLProgramFromString( ctx, device, 0, &pErrNum,additionalMacros, B3_BATCHING_PATH,true);
+		b3Assert(batchingProg);
+		
+		m_batchingKernel = b3OpenCLUtils::compileCLKernelFromString( ctx, device, batchKernelSource, "CreateBatches", &pErrNum, batchingProg,additionalMacros );
+		b3Assert(m_batchingKernel);
+	}
+	{
+		cl_program batchingNewProg = b3OpenCLUtils::compileCLProgramFromString( ctx, device, batchKernelNewSource, &pErrNum,additionalMacros, B3_BATCHING_NEW_PATH);
+		b3Assert(batchingNewProg);
+
+		m_batchingKernelNew = b3OpenCLUtils::compileCLKernelFromString( ctx, device, batchKernelNewSource, "CreateBatchesNew", &pErrNum, batchingNewProg,additionalMacros );
+		//m_batchingKernelNew = b3OpenCLUtils::compileCLKernelFromString( ctx, device, batchKernelNewSource, "CreateBatchesBruteForce", &pErrNum, batchingNewProg,additionalMacros );
+		b3Assert(m_batchingKernelNew);
+	}
+}
+		
+b3Solver::~b3Solver()
+{
+	delete m_offsets;
+	delete m_numConstraints;
+	delete m_sortDataBuffer;
+	delete m_contactBuffer2;
+
+	delete m_sort32;
+	delete m_scan;
+	delete m_search;
+
+
+	clReleaseKernel(m_batchingKernel);
+	clReleaseKernel(m_batchingKernelNew);
+	
+	clReleaseKernel( m_solveContactKernel);
+	clReleaseKernel( m_solveFrictionKernel);
+
+	clReleaseKernel( m_contactToConstraintKernel);
+	clReleaseKernel( m_setSortDataKernel);
+	clReleaseKernel( m_reorderContactKernel);
+	clReleaseKernel( m_copyConstraintKernel);
+			
+}
+
+
+ 
+
+template<bool JACOBI>
+static
+__inline
+void solveContact(b3GpuConstraint4& cs, 
+	const b3Vector3& posA, b3Vector3& linVelA, b3Vector3& angVelA, float invMassA, const b3Matrix3x3& invInertiaA,
+	const b3Vector3& posB, b3Vector3& linVelB, b3Vector3& angVelB, float invMassB, const b3Matrix3x3& invInertiaB, 
+	float maxRambdaDt[4], float minRambdaDt[4])
+{
+
+	b3Vector3 dLinVelA; dLinVelA.setZero();
+	b3Vector3 dAngVelA; dAngVelA.setZero();
+	b3Vector3 dLinVelB; dLinVelB.setZero();
+	b3Vector3 dAngVelB; dAngVelB.setZero();
+
+	for(int ic=0; ic<4; ic++)
+	{
+		//	dont necessary because this makes change to 0
+		if( cs.m_jacCoeffInv[ic] == 0.f ) continue;
+
+		{
+			b3Vector3 angular0, angular1, linear;
+			b3Vector3 r0 = cs.m_worldPos[ic] - (b3Vector3&)posA;
+			b3Vector3 r1 = cs.m_worldPos[ic] - (b3Vector3&)posB;
+			setLinearAndAngular( (const b3Vector3 &)cs.m_linear, (const b3Vector3 &)r0, (const b3Vector3 &)r1, &linear, &angular0, &angular1 );
+
+			float rambdaDt = calcRelVel((const b3Vector3 &)cs.m_linear,(const b3Vector3 &) -cs.m_linear, angular0, angular1,
+				linVelA, angVelA, linVelB, angVelB ) + cs.m_b[ic];
+			rambdaDt *= cs.m_jacCoeffInv[ic];
+
+			{
+				float prevSum = cs.m_appliedRambdaDt[ic];
+				float updated = prevSum;
+				updated += rambdaDt;
+				updated = b3Max( updated, minRambdaDt[ic] );
+				updated = b3Min( updated, maxRambdaDt[ic] );
+				rambdaDt = updated - prevSum;
+				cs.m_appliedRambdaDt[ic] = updated;
+			}
+
+			b3Vector3 linImp0 = invMassA*linear*rambdaDt;
+			b3Vector3 linImp1 = invMassB*(-linear)*rambdaDt;
+			b3Vector3 angImp0 = (invInertiaA* angular0)*rambdaDt;
+			b3Vector3 angImp1 = (invInertiaB* angular1)*rambdaDt;
+#ifdef _WIN32
+            b3Assert(_finite(linImp0.getX()));
+			b3Assert(_finite(linImp1.getX()));
+#endif
+			if( JACOBI )
+			{
+				dLinVelA += linImp0;
+				dAngVelA += angImp0;
+				dLinVelB += linImp1;
+				dAngVelB += angImp1;
+			}
+			else
+			{
+				linVelA += linImp0;
+				angVelA += angImp0;
+				linVelB += linImp1;
+				angVelB += angImp1;
+			}
+		}
+	}
+
+	if( JACOBI )
+	{
+		linVelA += dLinVelA;
+		angVelA += dAngVelA;
+		linVelB += dLinVelB;
+		angVelB += dAngVelB;
+	}
+
+}
+
+
+
+
+
+	static
+	__inline
+	void solveFriction(b3GpuConstraint4& cs, 
+		const b3Vector3& posA, b3Vector3& linVelA, b3Vector3& angVelA, float invMassA, const b3Matrix3x3& invInertiaA,
+		const b3Vector3& posB, b3Vector3& linVelB, b3Vector3& angVelB, float invMassB, const b3Matrix3x3& invInertiaB, 
+		float maxRambdaDt[4], float minRambdaDt[4])
+	{
+
+		if( cs.m_fJacCoeffInv[0] == 0 && cs.m_fJacCoeffInv[0] == 0 ) return;
+		const b3Vector3& center = (const b3Vector3&)cs.m_center;
+
+		b3Vector3 n = -(const b3Vector3&)cs.m_linear;
+
+		b3Vector3 tangent[2];
+#if 1		
+		b3PlaneSpace1 (n, tangent[0],tangent[1]);
+#else
+		b3Vector3 r = cs.m_worldPos[0]-center;
+		tangent[0] = cross3( n, r );
+		tangent[1] = cross3( tangent[0], n );
+		tangent[0] = normalize3( tangent[0] );
+		tangent[1] = normalize3( tangent[1] );
+#endif
+
+		b3Vector3 angular0, angular1, linear;
+		b3Vector3 r0 = center - posA;
+		b3Vector3 r1 = center - posB;
+		for(int i=0; i<2; i++)
+		{
+			setLinearAndAngular( tangent[i], r0, r1, &linear, &angular0, &angular1 );
+			float rambdaDt = calcRelVel(linear, -linear, angular0, angular1,
+				linVelA, angVelA, linVelB, angVelB );
+			rambdaDt *= cs.m_fJacCoeffInv[i];
+
+				{
+					float prevSum = cs.m_fAppliedRambdaDt[i];
+					float updated = prevSum;
+					updated += rambdaDt;
+					updated = b3Max( updated, minRambdaDt[i] );
+					updated = b3Min( updated, maxRambdaDt[i] );
+					rambdaDt = updated - prevSum;
+					cs.m_fAppliedRambdaDt[i] = updated;
+				}
+
+			b3Vector3 linImp0 = invMassA*linear*rambdaDt;
+			b3Vector3 linImp1 = invMassB*(-linear)*rambdaDt;
+			b3Vector3 angImp0 = (invInertiaA* angular0)*rambdaDt;
+			b3Vector3 angImp1 = (invInertiaB* angular1)*rambdaDt;
+#ifdef _WIN32
+			b3Assert(_finite(linImp0.getX()));
+			b3Assert(_finite(linImp1.getX()));
+#endif
+			linVelA += linImp0;
+			angVelA += angImp0;
+			linVelB += linImp1;
+			angVelB += angImp1;
+		}
+
+		{	//	angular damping for point constraint
+			b3Vector3 ab = ( posB - posA ).normalized();
+			b3Vector3 ac = ( center - posA ).normalized();
+			if( b3Dot( ab, ac ) > 0.95f || (invMassA == 0.f || invMassB == 0.f))
+			{
+				float angNA = b3Dot( n, angVelA );
+				float angNB = b3Dot( n, angVelB );
+
+				angVelA -= (angNA*0.1f)*n;
+				angVelB -= (angNB*0.1f)*n;
+			}
+		}
+
+	}
+/*
+ b3AlignedObjectArray<b3RigidBodyData>& m_bodies;
+	b3AlignedObjectArray<b3InertiaData>& m_shapes;
+	b3AlignedObjectArray<b3GpuConstraint4>& m_constraints;
+	b3AlignedObjectArray<int>* m_batchSizes;
+	int m_cellIndex;
+	int m_curWgidx;
+	int m_start;
+	int m_nConstraints;
+	bool m_solveFriction;
+	int m_maxNumBatches;
+ */
+
+struct SolveTask// : public ThreadPool::Task
+{
+	SolveTask(b3AlignedObjectArray<b3RigidBodyData>& bodies,  b3AlignedObjectArray<b3InertiaData>& shapes, b3AlignedObjectArray<b3GpuConstraint4>& constraints,
+		int start, int nConstraints,int maxNumBatches,b3AlignedObjectArray<int>* wgUsedBodies, int curWgidx, b3AlignedObjectArray<int>* batchSizes, int cellIndex)
+		: m_bodies( bodies ), m_shapes( shapes ), 
+		m_constraints( constraints ), 
+		m_batchSizes(batchSizes),
+		m_cellIndex(cellIndex),
+		m_curWgidx(curWgidx),
+		m_start( start ), 
+		m_nConstraints( nConstraints ),
+		m_solveFriction( true ),
+		m_maxNumBatches(maxNumBatches)
+	{}
+
+	unsigned short int getType(){ return 0; }
+
+	void run(int tIdx)
+	{
+		int offset = 0;
+		for (int ii=0;ii<B3_MAX_NUM_BATCHES;ii++)
+		{
+			int numInBatch = m_batchSizes->at(m_cellIndex*B3_MAX_NUM_BATCHES+ii);
+			if (!numInBatch)
+				break;
+
+			for (int jj=0;jj<numInBatch;jj++)
+			{
+				int i = m_start + offset+jj;
+				int batchId = m_constraints[i].m_batchIdx;
+				b3Assert(batchId==ii);
+				float frictionCoeff = m_constraints[i].getFrictionCoeff();
+				int aIdx = (int)m_constraints[i].m_bodyA;
+				int bIdx = (int)m_constraints[i].m_bodyB;
+//				int localBatch = m_constraints[i].m_batchIdx;
+				b3RigidBodyData& bodyA = m_bodies[aIdx];
+				b3RigidBodyData& bodyB = m_bodies[bIdx];
+
+				if( !m_solveFriction )
+				{
+					float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};
+					float minRambdaDt[4] = {0.f,0.f,0.f,0.f};
+
+					solveContact<false>( m_constraints[i], (b3Vector3&)bodyA.m_pos, (b3Vector3&)bodyA.m_linVel, (b3Vector3&)bodyA.m_angVel, bodyA.m_invMass, (const b3Matrix3x3 &)m_shapes[aIdx].m_invInertiaWorld, 
+							(b3Vector3&)bodyB.m_pos, (b3Vector3&)bodyB.m_linVel, (b3Vector3&)bodyB.m_angVel, bodyB.m_invMass, (const b3Matrix3x3 &)m_shapes[bIdx].m_invInertiaWorld,
+						maxRambdaDt, minRambdaDt );
+				}
+				else
+				{
+					float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};
+					float minRambdaDt[4] = {0.f,0.f,0.f,0.f};
+					float sum = 0;
+					for(int j=0; j<4; j++)
+					{
+						sum +=m_constraints[i].m_appliedRambdaDt[j];
+					}
+					frictionCoeff = 0.7f;
+					for(int j=0; j<4; j++)
+					{
+						maxRambdaDt[j] = frictionCoeff*sum;
+						minRambdaDt[j] = -maxRambdaDt[j];
+					}
+					solveFriction( m_constraints[i], (b3Vector3&)bodyA.m_pos, (b3Vector3&)bodyA.m_linVel, (b3Vector3&)bodyA.m_angVel, bodyA.m_invMass,(const b3Matrix3x3 &) m_shapes[aIdx].m_invInertiaWorld, 
+						(b3Vector3&)bodyB.m_pos, (b3Vector3&)bodyB.m_linVel, (b3Vector3&)bodyB.m_angVel, bodyB.m_invMass,(const b3Matrix3x3 &) m_shapes[bIdx].m_invInertiaWorld,
+						maxRambdaDt, minRambdaDt );
+			
+				}
+			}
+			offset+=numInBatch;
+
+
+		}
+/*		for (int bb=0;bb<m_maxNumBatches;bb++)
+		{
+			//for(int ic=m_nConstraints-1; ic>=0; ic--)
+			for(int ic=0; ic<m_nConstraints; ic++)
+			{
+				
+				int i = m_start + ic;
+				if (m_constraints[i].m_batchIdx != bb)
+					continue;
+
+				float frictionCoeff = m_constraints[i].getFrictionCoeff();
+				int aIdx = (int)m_constraints[i].m_bodyA;
+				int bIdx = (int)m_constraints[i].m_bodyB;
+				int localBatch = m_constraints[i].m_batchIdx;
+				b3RigidBodyData& bodyA = m_bodies[aIdx];
+				b3RigidBodyData& bodyB = m_bodies[bIdx];
+
+				if( !m_solveFriction )
+				{
+					float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};
+					float minRambdaDt[4] = {0.f,0.f,0.f,0.f};
+
+					solveContact<false>( m_constraints[i], (b3Vector3&)bodyA.m_pos, (b3Vector3&)bodyA.m_linVel, (b3Vector3&)bodyA.m_angVel, bodyA.m_invMass, (const b3Matrix3x3 &)m_shapes[aIdx].m_invInertiaWorld, 
+							(b3Vector3&)bodyB.m_pos, (b3Vector3&)bodyB.m_linVel, (b3Vector3&)bodyB.m_angVel, bodyB.m_invMass, (const b3Matrix3x3 &)m_shapes[bIdx].m_invInertiaWorld,
+						maxRambdaDt, minRambdaDt );
+				}
+				else
+				{
+					float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};
+					float minRambdaDt[4] = {0.f,0.f,0.f,0.f};
+					float sum = 0;
+					for(int j=0; j<4; j++)
+					{
+						sum +=m_constraints[i].m_appliedRambdaDt[j];
+					}
+					frictionCoeff = 0.7f;
+					for(int j=0; j<4; j++)
+					{
+						maxRambdaDt[j] = frictionCoeff*sum;
+						minRambdaDt[j] = -maxRambdaDt[j];
+					}
+					solveFriction( m_constraints[i], (b3Vector3&)bodyA.m_pos, (b3Vector3&)bodyA.m_linVel, (b3Vector3&)bodyA.m_angVel, bodyA.m_invMass,(const b3Matrix3x3 &) m_shapes[aIdx].m_invInertiaWorld, 
+						(b3Vector3&)bodyB.m_pos, (b3Vector3&)bodyB.m_linVel, (b3Vector3&)bodyB.m_angVel, bodyB.m_invMass,(const b3Matrix3x3 &) m_shapes[bIdx].m_invInertiaWorld,
+						maxRambdaDt, minRambdaDt );
+			
+				}
+			}
+		}
+		*/
+
+
+		
+	}
+
+	b3AlignedObjectArray<b3RigidBodyData>& m_bodies;
+	b3AlignedObjectArray<b3InertiaData>& m_shapes;
+	b3AlignedObjectArray<b3GpuConstraint4>& m_constraints;
+	b3AlignedObjectArray<int>* m_batchSizes;
+	int m_cellIndex;
+	int m_curWgidx;
+	int m_start;
+	int m_nConstraints;
+	bool m_solveFriction;
+	int m_maxNumBatches;
+};
+
+
+void b3Solver::solveContactConstraintHost(  b3OpenCLArray<b3RigidBodyData>* bodyBuf, b3OpenCLArray<b3InertiaData>* shapeBuf, 
+			b3OpenCLArray<b3GpuConstraint4>* constraint, void* additionalData, int n ,int maxNumBatches,b3AlignedObjectArray<int>* batchSizes)
+{
+
+#if 0
+	{	
+		int nSplitX = B3_SOLVER_N_SPLIT_X;
+		int nSplitY = B3_SOLVER_N_SPLIT_Y;
+		int numWorkgroups = B3_SOLVER_N_CELLS/B3_SOLVER_N_BATCHES;
+		for (int z=0;z<4;z++)
+		{
+			for (int y=0;y<4;y++)
+			{
+				for (int x=0;x<4;x++)
+				{
+					int newIndex = (x+y*nSplitX+z*nSplitX*nSplitY);
+				//	printf("newIndex=%d\n",newIndex);
+
+					int zIdx = newIndex/(nSplitX*nSplitY);
+					int remain = newIndex%(nSplitX*nSplitY);
+					int yIdx = remain/nSplitX;
+					int xIdx = remain%nSplitX;
+				//	printf("newIndex=%d\n",newIndex);
+				}
+			}
+		}
+
+		//for (int wgIdx=numWorkgroups-1;wgIdx>=0;wgIdx--)
+		for (int cellBatch=0;cellBatch<B3_SOLVER_N_BATCHES;cellBatch++)
+		{
+			for (int wgIdx=0;wgIdx<numWorkgroups;wgIdx++)
+			{
+				int zIdx = (wgIdx/((nSplitX*nSplitY)/4))*2+((cellBatch&4)>>2);
+				int remain= (wgIdx%((nSplitX*nSplitY)/4));
+				int yIdx = (remain/(nSplitX/2))*2 + ((cellBatch&2)>>1);
+				int xIdx = (remain%(nSplitX/2))*2 + (cellBatch&1);
+				
+				/*int zIdx = newIndex/(nSplitX*nSplitY);
+				int remain = newIndex%(nSplitX*nSplitY);
+				int yIdx = remain/nSplitX;
+				int xIdx = remain%nSplitX;
+				*/
+				int cellIdx = xIdx+yIdx*nSplitX+zIdx*(nSplitX*nSplitY);
+			//	printf("wgIdx %d: xIdx=%d, yIdx=%d, zIdx=%d, cellIdx=%d, cell Batch %d\n",wgIdx,xIdx,yIdx,zIdx,cellIdx,cellBatch);
+			}
+		}
+	}
+#endif
+
+	b3AlignedObjectArray<b3RigidBodyData> bodyNative;
+	bodyBuf->copyToHost(bodyNative);
+	b3AlignedObjectArray<b3InertiaData> shapeNative;
+	shapeBuf->copyToHost(shapeNative);
+	b3AlignedObjectArray<b3GpuConstraint4> constraintNative;
+	constraint->copyToHost(constraintNative);
+
+	b3AlignedObjectArray<unsigned int> numConstraintsHost;
+	m_numConstraints->copyToHost(numConstraintsHost);
+
+	//printf("------------------------\n");
+	b3AlignedObjectArray<unsigned int> offsetsHost;
+	m_offsets->copyToHost(offsetsHost);
+	static int frame=0;
+	bool useBatches=true;
+	if (useBatches)
+	{
+		for(int iter=0; iter<m_nIterations; iter++)
+		{
+			for (int cellBatch=0;cellBatch<B3_SOLVER_N_BATCHES;cellBatch++)
+			{
+				
+				int nSplitX = B3_SOLVER_N_SPLIT_X;
+				int nSplitY = B3_SOLVER_N_SPLIT_Y;
+				int numWorkgroups = B3_SOLVER_N_CELLS/B3_SOLVER_N_BATCHES;
+				//printf("cell Batch %d\n",cellBatch);
+				b3AlignedObjectArray<int> usedBodies[B3_SOLVER_N_CELLS];
+				for (int i=0;i<B3_SOLVER_N_CELLS;i++)
+				{
+					usedBodies[i].resize(0);
+				}
+
+				
+
+
+				//for (int wgIdx=numWorkgroups-1;wgIdx>=0;wgIdx--)
+				for (int wgIdx=0;wgIdx<numWorkgroups;wgIdx++)
+				{
+					int zIdx = (wgIdx/((nSplitX*nSplitY)/4))*2+((cellBatch&4)>>2);
+					int remain= (wgIdx%((nSplitX*nSplitY)/4));
+					int yIdx = (remain/(nSplitX/2))*2 + ((cellBatch&2)>>1);
+					int xIdx = (remain%(nSplitX/2))*2 + (cellBatch&1);
+					int cellIdx = xIdx+yIdx*nSplitX+zIdx*(nSplitX*nSplitY);
+					
+	
+					if( numConstraintsHost[cellIdx] == 0 ) 
+						continue;
+
+					//printf("wgIdx %d: xIdx=%d, yIdx=%d, zIdx=%d, cellIdx=%d, cell Batch %d\n",wgIdx,xIdx,yIdx,zIdx,cellIdx,cellBatch);
+					//printf("cell %d has %d constraints\n", cellIdx,numConstraintsHost[cellIdx]);
+					if (zIdx)
+					{
+					//printf("?\n");
+					}
+
+					if (iter==0)
+					{
+						//printf("frame=%d, Cell xIdx=%x, yIdx=%d ",frame, xIdx,yIdx);
+						//printf("cellBatch=%d, wgIdx=%d, #constraints in cell=%d\n",cellBatch,wgIdx,numConstraintsHost[cellIdx]);
+					}
+					const int start = offsetsHost[cellIdx];
+					int numConstraintsInCell = numConstraintsHost[cellIdx];
+	//				const int end = start + numConstraintsInCell;
+
+					SolveTask task( bodyNative, shapeNative, constraintNative, start, numConstraintsInCell ,maxNumBatches,usedBodies,wgIdx,batchSizes,cellIdx);
+					task.m_solveFriction = false;
+					task.run(0);
+				
+				}
+			}
+		}
+
+		for(int iter=0; iter<m_nIterations; iter++)
+		{
+			for (int cellBatch=0;cellBatch<B3_SOLVER_N_BATCHES;cellBatch++)
+			{
+				int nSplitX = B3_SOLVER_N_SPLIT_X;
+				int nSplitY = B3_SOLVER_N_SPLIT_Y;
+				
+
+				int numWorkgroups = B3_SOLVER_N_CELLS/B3_SOLVER_N_BATCHES;
+
+				for (int wgIdx=0;wgIdx<numWorkgroups;wgIdx++)
+				{
+					int zIdx = (wgIdx/((nSplitX*nSplitY)/4))*2+((cellBatch&4)>>2);
+					int remain= (wgIdx%((nSplitX*nSplitY)/4));
+					int yIdx = (remain/(nSplitX/2))*2 + ((cellBatch&2)>>1);
+					int xIdx = (remain%(nSplitX/2))*2 + (cellBatch&1);
+					
+					int cellIdx = xIdx+yIdx*nSplitX+zIdx*(nSplitX*nSplitY);
+	
+					if( numConstraintsHost[cellIdx] == 0 ) 
+						continue;
+	
+					//printf("yIdx=%d\n",yIdx);
+					
+					const int start = offsetsHost[cellIdx];
+					int numConstraintsInCell = numConstraintsHost[cellIdx];
+	//				const int end = start + numConstraintsInCell;
+
+					SolveTask task( bodyNative, shapeNative, constraintNative, start, numConstraintsInCell,maxNumBatches, 0,0,batchSizes,cellIdx);
+					task.m_solveFriction = true;
+					task.run(0);
+					
+				}
+			}
+		}
+
+
+	} else
+	{
+		for(int iter=0; iter<m_nIterations; iter++)
+		{
+			SolveTask task( bodyNative, shapeNative, constraintNative, 0, n ,maxNumBatches,0,0,0,0);
+			task.m_solveFriction = false;
+			task.run(0);
+		}
+
+		for(int iter=0; iter<m_nIterations; iter++)
+		{
+			SolveTask task( bodyNative, shapeNative, constraintNative, 0, n ,maxNumBatches,0,0,0,0);
+			task.m_solveFriction = true;
+			task.run(0);
+		}
+	}
+
+	bodyBuf->copyFromHost(bodyNative);
+	shapeBuf->copyFromHost(shapeNative);
+	constraint->copyFromHost(constraintNative);
+	frame++;
+	
+}
+
+void checkConstraintBatch(const b3OpenCLArray<b3RigidBodyData>* bodyBuf,
+					const b3OpenCLArray<b3InertiaData>* shapeBuf,
+					b3OpenCLArray<b3GpuConstraint4>* constraint, 
+					b3OpenCLArray<unsigned int>* m_numConstraints,
+					b3OpenCLArray<unsigned int>* m_offsets,
+					int batchId
+					)
+{
+//						b3BufferInfoCL( m_numConstraints->getBufferCL() ), 
+//						b3BufferInfoCL( m_offsets->getBufferCL() ) 
+	
+	int cellBatch = batchId;
+	const int nn = B3_SOLVER_N_CELLS;
+//	int numWorkItems = 64*nn/B3_SOLVER_N_BATCHES;
+
+	b3AlignedObjectArray<unsigned int> gN;
+	m_numConstraints->copyToHost(gN);
+	b3AlignedObjectArray<unsigned int> gOffsets;
+	m_offsets->copyToHost(gOffsets);
+	int nSplitX = B3_SOLVER_N_SPLIT_X;
+	int nSplitY = B3_SOLVER_N_SPLIT_Y;
+	
+//	int bIdx = batchId;
+
+	b3AlignedObjectArray<b3GpuConstraint4> cpuConstraints;
+	constraint->copyToHost(cpuConstraints);
+
+	printf("batch = %d\n", batchId);
+
+	int numWorkgroups = nn/B3_SOLVER_N_BATCHES;
+	b3AlignedObjectArray<int> usedBodies;
+
+
+	for (int wgIdx=0;wgIdx<numWorkgroups;wgIdx++)
+	{
+		printf("wgIdx = %d           ", wgIdx);
+
+		int zIdx = (wgIdx/((nSplitX*nSplitY))/2)*2+((cellBatch&4)>>2);					
+		int remain = wgIdx%((nSplitX*nSplitY));
+		int yIdx = (remain%(nSplitX/2))*2 + ((cellBatch&2)>>1);
+		int xIdx = (remain/(nSplitX/2))*2 + (cellBatch&1);
+
+		
+		int cellIdx = xIdx+yIdx*nSplitX+zIdx*(nSplitX*nSplitY);
+		printf("cellIdx=%d\n",cellIdx);
+		if( gN[cellIdx] == 0 ) 
+			continue;
+
+		const int start = gOffsets[cellIdx];
+		const int end = start + gN[cellIdx];
+
+		for (int c=start;c<end;c++)
+		{
+			b3GpuConstraint4& constraint = cpuConstraints[c];
+			//printf("constraint (%d,%d)\n", constraint.m_bodyA,constraint.m_bodyB);
+			if (usedBodies.findLinearSearch(constraint.m_bodyA)< usedBodies.size())
+			{
+				printf("error?\n");
+			}
+			if (usedBodies.findLinearSearch(constraint.m_bodyB)< usedBodies.size())
+			{
+				printf("error?\n");
+			}
+		}
+
+		for (int c=start;c<end;c++)
+		{
+			b3GpuConstraint4& constraint = cpuConstraints[c];
+			usedBodies.push_back(constraint.m_bodyA);
+			usedBodies.push_back(constraint.m_bodyB);
+		}
+
+	}
+}
+
+static bool verify=false;
+
+void b3Solver::solveContactConstraint(  const b3OpenCLArray<b3RigidBodyData>* bodyBuf, const b3OpenCLArray<b3InertiaData>* shapeBuf, 
+			b3OpenCLArray<b3GpuConstraint4>* constraint, void* additionalData, int n ,int maxNumBatches)
+{
+	
+	
+	b3Int4 cdata = b3MakeInt4( n, 0, 0, 0 );
+	{
+		
+		const int nn = B3_SOLVER_N_CELLS;
+
+		cdata.x = 0;
+		cdata.y = maxNumBatches;//250;
+
+
+		int numWorkItems = 64*nn/B3_SOLVER_N_BATCHES;
+#ifdef DEBUG_ME
+		SolverDebugInfo* debugInfo = new  SolverDebugInfo[numWorkItems];
+		adl::b3OpenCLArray<SolverDebugInfo> gpuDebugInfo(data->m_device,numWorkItems);
+#endif
+
+
+
+		{
+
+			B3_PROFILE("m_batchSolveKernel iterations");
+			for(int iter=0; iter<m_nIterations; iter++)
+			{
+				for(int ib=0; ib<B3_SOLVER_N_BATCHES; ib++)
+				{
+					
+					if (verify)
+					{
+						checkConstraintBatch(bodyBuf,shapeBuf,constraint,m_numConstraints,m_offsets,ib);
+					}
+
+#ifdef DEBUG_ME
+					memset(debugInfo,0,sizeof(SolverDebugInfo)*numWorkItems);
+					gpuDebugInfo.write(debugInfo,numWorkItems);
+#endif
+
+
+					cdata.z = ib;
+					
+
+				b3LauncherCL launcher( m_queue, m_solveContactKernel ,"m_solveContactKernel");
+#if 1
+                    
+					b3BufferInfoCL bInfo[] = { 
+
+						b3BufferInfoCL( bodyBuf->getBufferCL() ), 
+						b3BufferInfoCL( shapeBuf->getBufferCL() ), 
+						b3BufferInfoCL( constraint->getBufferCL() ),
+						b3BufferInfoCL( m_numConstraints->getBufferCL() ), 
+						b3BufferInfoCL( m_offsets->getBufferCL() ) 
+#ifdef DEBUG_ME
+						,	b3BufferInfoCL(&gpuDebugInfo)
+#endif
+						};
+
+					
+
+                    launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+					//launcher.setConst(  cdata.x );
+                    launcher.setConst(  cdata.y );
+                    launcher.setConst(  cdata.z );
+                    b3Int4 nSplit;
+					nSplit.x = B3_SOLVER_N_SPLIT_X;
+					nSplit.y = B3_SOLVER_N_SPLIT_Y;
+					nSplit.z = B3_SOLVER_N_SPLIT_Z;
+
+                    launcher.setConst(  nSplit );
+                    launcher.launch1D( numWorkItems, 64 );
+
+                    
+#else
+                    const char* fileName = "m_batchSolveKernel.bin";
+                    FILE* f = fopen(fileName,"rb");
+                    if (f)
+                    {
+                        int sizeInBytes=0;
+                        if (fseek(f, 0, SEEK_END) || (sizeInBytes = ftell(f)) == EOF || fseek(f, 0, SEEK_SET))
+                        {
+                            printf("error, cannot get file size\n");
+                            exit(0);
+                        }
+                        
+                        unsigned char* buf = (unsigned char*) malloc(sizeInBytes);
+                        fread(buf,sizeInBytes,1,f);
+                        int serializedBytes = launcher.deserializeArgs(buf, sizeInBytes,m_context);
+                        int num = *(int*)&buf[serializedBytes];
+                        
+                        launcher.launch1D( num);
+
+                        //this clFinish is for testing on errors
+                        clFinish(m_queue);
+                    }
+
+#endif
+					
+
+#ifdef DEBUG_ME
+					clFinish(m_queue);
+					gpuDebugInfo.read(debugInfo,numWorkItems);
+					clFinish(m_queue);
+					for (int i=0;i<numWorkItems;i++)
+					{
+						if (debugInfo[i].m_valInt2>0)
+						{
+							printf("debugInfo[i].m_valInt2 = %d\n",i,debugInfo[i].m_valInt2);
+						}
+
+						if (debugInfo[i].m_valInt3>0)
+						{
+							printf("debugInfo[i].m_valInt3 = %d\n",i,debugInfo[i].m_valInt3);
+						}
+					}
+#endif //DEBUG_ME
+
+
+				}
+			}
+		
+			clFinish(m_queue);
+
+
+		}
+
+		cdata.x = 1;
+		bool applyFriction=true;
+		if (applyFriction)
+    	{
+			B3_PROFILE("m_batchSolveKernel iterations2");
+			for(int iter=0; iter<m_nIterations; iter++)
+			{
+				for(int ib=0; ib<B3_SOLVER_N_BATCHES; ib++)
+				{
+					cdata.z = ib;
+					
+
+					b3BufferInfoCL bInfo[] = { 
+						b3BufferInfoCL( bodyBuf->getBufferCL() ), 
+						b3BufferInfoCL( shapeBuf->getBufferCL() ), 
+						b3BufferInfoCL( constraint->getBufferCL() ),
+						b3BufferInfoCL( m_numConstraints->getBufferCL() ), 
+						b3BufferInfoCL( m_offsets->getBufferCL() )
+#ifdef DEBUG_ME
+						,b3BufferInfoCL(&gpuDebugInfo)
+#endif //DEBUG_ME
+					};
+					b3LauncherCL launcher( m_queue, m_solveFrictionKernel,"m_solveFrictionKernel" );
+					launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+					//launcher.setConst(  cdata.x );
+                    launcher.setConst(  cdata.y );
+                    launcher.setConst(  cdata.z );
+                    b3Int4 nSplit;
+					nSplit.x = B3_SOLVER_N_SPLIT_X;
+					nSplit.y = B3_SOLVER_N_SPLIT_Y;
+					nSplit.z = B3_SOLVER_N_SPLIT_Z;
+
+                    launcher.setConst(  nSplit );
+                    
+					launcher.launch1D( 64*nn/B3_SOLVER_N_BATCHES, 64 );
+				}
+			}
+			clFinish(m_queue);
+			
+		}
+#ifdef DEBUG_ME
+		delete[] debugInfo;
+#endif //DEBUG_ME
+	}
+
+	
+}
+
+void b3Solver::convertToConstraints( const b3OpenCLArray<b3RigidBodyData>* bodyBuf, 
+	const b3OpenCLArray<b3InertiaData>* shapeBuf, 
+	b3OpenCLArray<b3Contact4>* contactsIn, b3OpenCLArray<b3GpuConstraint4>* contactCOut, void* additionalData, 
+	int nContacts, const ConstraintCfg& cfg )
+{
+//	b3OpenCLArray<b3GpuConstraint4>* constraintNative =0;
+	contactCOut->resize(nContacts);
+	struct CB
+	{
+		int m_nContacts;
+		float m_dt;
+		float m_positionDrift;
+		float m_positionConstraintCoeff;
+	};
+
+	{
+
+		CB cdata;
+		cdata.m_nContacts = nContacts;
+		cdata.m_dt = cfg.m_dt;
+		cdata.m_positionDrift = cfg.m_positionDrift;
+		cdata.m_positionConstraintCoeff = cfg.m_positionConstraintCoeff;
+
+		
+		if (gConvertConstraintOnCpu)
+		{
+			b3AlignedObjectArray<b3RigidBodyData> gBodies;
+		bodyBuf->copyToHost(gBodies);
+
+		b3AlignedObjectArray<b3Contact4> gContact;
+		contactsIn->copyToHost(gContact);
+
+		b3AlignedObjectArray<b3InertiaData> gShapes;
+		shapeBuf->copyToHost(gShapes);
+		
+		b3AlignedObjectArray<b3GpuConstraint4> gConstraintOut;
+		gConstraintOut.resize(nContacts);
+		
+			B3_PROFILE("cpu contactToConstraintKernel");
+			for (int gIdx=0;gIdx<nContacts;gIdx++)
+			{
+				int aIdx = abs(gContact[gIdx].m_bodyAPtrAndSignBit);
+				int bIdx = abs(gContact[gIdx].m_bodyBPtrAndSignBit);
+
+				b3Float4 posA = gBodies[aIdx].m_pos;
+				b3Float4 linVelA = gBodies[aIdx].m_linVel;
+				b3Float4 angVelA = gBodies[aIdx].m_angVel;
+				float invMassA = gBodies[aIdx].m_invMass;
+				b3Mat3x3 invInertiaA = gShapes[aIdx].m_initInvInertia;
+
+				b3Float4 posB = gBodies[bIdx].m_pos;
+				b3Float4 linVelB = gBodies[bIdx].m_linVel;
+				b3Float4 angVelB = gBodies[bIdx].m_angVel;
+				float invMassB = gBodies[bIdx].m_invMass;
+				b3Mat3x3 invInertiaB = gShapes[bIdx].m_initInvInertia;
+
+				b3ContactConstraint4_t cs;
+
+    			setConstraint4( posA, linVelA, angVelA, invMassA, invInertiaA, posB, linVelB, angVelB, invMassB, invInertiaB,
+					&gContact[gIdx], cdata.m_dt, cdata.m_positionDrift, cdata.m_positionConstraintCoeff,
+					&cs );
+		
+				cs.m_batchIdx = gContact[gIdx].m_batchIdx;
+
+				gConstraintOut[gIdx] = (b3GpuConstraint4&)cs;
+			}
+
+			contactCOut->copyFromHost(gConstraintOut);
+
+		} else
+		{
+			B3_PROFILE("gpu m_contactToConstraintKernel");
+
+		
+			b3BufferInfoCL bInfo[] = { b3BufferInfoCL( contactsIn->getBufferCL() ), b3BufferInfoCL( bodyBuf->getBufferCL() ), b3BufferInfoCL( shapeBuf->getBufferCL()),
+				b3BufferInfoCL( contactCOut->getBufferCL() )};
+			b3LauncherCL launcher( m_queue, m_contactToConstraintKernel,"m_contactToConstraintKernel" );
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			//launcher.setConst(  cdata );
+        
+			launcher.setConst(cdata.m_nContacts);
+			launcher.setConst(cdata.m_dt);
+			launcher.setConst(cdata.m_positionDrift);
+			launcher.setConst(cdata.m_positionConstraintCoeff);
+        
+			launcher.launch1D( nContacts, 64 );	
+			clFinish(m_queue);
+
+		}
+	}
+
+	
+}
+
+/*
+void b3Solver::sortContacts(  const b3OpenCLArray<b3RigidBodyData>* bodyBuf, 
+			b3OpenCLArray<b3Contact4>* contactsIn, void* additionalData, 
+			int nContacts, const b3Solver::ConstraintCfg& cfg )
+{
+	
+	
+
+	const int sortAlignment = 512; // todo. get this out of sort
+	if( cfg.m_enableParallelSolve )
+	{
+		
+
+		int sortSize = NEXTMULTIPLEOF( nContacts, sortAlignment );
+
+		b3OpenCLArray<unsigned int>* countsNative = m_numConstraints;//BufferUtils::map<TYPE_CL, false>( data->m_device, &countsHost );
+		b3OpenCLArray<unsigned int>* offsetsNative = m_offsets;//BufferUtils::map<TYPE_CL, false>( data->m_device, &offsetsHost );
+
+		{	//	2. set cell idx
+			struct CB
+			{
+				int m_nContacts;
+				int m_staticIdx;
+				float m_scale;
+				int m_nSplit;
+			};
+
+			b3Assert( sortSize%64 == 0 );
+			CB cdata;
+			cdata.m_nContacts = nContacts;
+			cdata.m_staticIdx = cfg.m_staticIdx;
+			cdata.m_scale = 1.f/(N_OBJ_PER_SPLIT*cfg.m_averageExtent);
+			cdata.m_nSplit = B3_SOLVER_N_SPLIT;
+
+			
+			b3BufferInfoCL bInfo[] = { b3BufferInfoCL( contactsIn->getBufferCL() ), b3BufferInfoCL( bodyBuf->getBufferCL() ), b3BufferInfoCL( m_sortDataBuffer->getBufferCL() ) };
+			b3LauncherCL launcher( m_queue, m_setSortDataKernel );
+			launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+			launcher.setConst(  cdata );
+			launcher.launch1D( sortSize, 64 );
+		}
+
+		{	//	3. sort by cell idx
+			int n = B3_SOLVER_N_SPLIT*B3_SOLVER_N_SPLIT;
+			int sortBit = 32;
+			//if( n <= 0xffff ) sortBit = 16;
+			//if( n <= 0xff ) sortBit = 8;
+			m_sort32->execute(*m_sortDataBuffer,sortSize);
+		}
+		{	//	4. find entries
+			m_search->execute( *m_sortDataBuffer, nContacts, *countsNative, B3_SOLVER_N_SPLIT*B3_SOLVER_N_SPLIT, b3BoundSearchCL::COUNT);
+
+			m_scan->execute( *countsNative, *offsetsNative, B3_SOLVER_N_SPLIT*B3_SOLVER_N_SPLIT );
+		}
+
+		{	//	5. sort constraints by cellIdx
+			//	todo. preallocate this
+//			b3Assert( contactsIn->getType() == TYPE_HOST );
+//			b3OpenCLArray<b3Contact4>* out = BufferUtils::map<TYPE_CL, false>( data->m_device, contactsIn );	//	copying contacts to this buffer
+
+			{
+				
+
+				b3Int4 cdata; cdata.x = nContacts;
+				b3BufferInfoCL bInfo[] = { b3BufferInfoCL( contactsIn->getBufferCL() ), b3BufferInfoCL( m_contactBuffer->getBufferCL() ), b3BufferInfoCL( m_sortDataBuffer->getBufferCL() ) };
+				b3LauncherCL launcher( m_queue, m_reorderContactKernel );
+				launcher.setBuffers( bInfo, sizeof(bInfo)/sizeof(b3BufferInfoCL) );
+				launcher.setConst(  cdata );
+				launcher.launch1D( nContacts, 64 );
+			}
+//			BufferUtils::unmap<true>( out, contactsIn, nContacts );
+		}
+	}
+
+	
+}
+
+*/
+void	b3Solver::batchContacts(  b3OpenCLArray<b3Contact4>* contacts, int nContacts, b3OpenCLArray<unsigned int>* nNative, b3OpenCLArray<unsigned int>* offsetsNative, int staticIdx )
+{
+	
+	int numWorkItems = 64*B3_SOLVER_N_CELLS;
+	{
+		B3_PROFILE("batch generation");
+		
+		b3Int4 cdata;
+		cdata.x = nContacts;
+		cdata.y = 0;
+		cdata.z = staticIdx;
+
+		
+#ifdef BATCH_DEBUG
+		SolverDebugInfo* debugInfo = new  SolverDebugInfo[numWorkItems];
+		adl::b3OpenCLArray<SolverDebugInfo> gpuDebugInfo(data->m_device,numWorkItems);
+		memset(debugInfo,0,sizeof(SolverDebugInfo)*numWorkItems);
+		gpuDebugInfo.write(debugInfo,numWorkItems);
+#endif
+
+		
+
+#if 0
+		b3BufferInfoCL bInfo[] = { 
+			b3BufferInfoCL( contacts->getBufferCL() ), 
+			b3BufferInfoCL(  m_contactBuffer2->getBufferCL()),
+			b3BufferInfoCL( nNative->getBufferCL() ), 
+			b3BufferInfoCL( offsetsNative->getBufferCL() ),
+#ifdef BATCH_DEBUG
+			,	b3BufferInfoCL(&gpuDebugInfo)
+#endif
+		};
+#endif
+		
+		
+
+		{
+			m_batchSizes.resize(nNative->size());
+			B3_PROFILE("batchingKernel");
+			//b3LauncherCL launcher( m_queue, m_batchingKernel);
+			cl_kernel k = useNewBatchingKernel ? m_batchingKernelNew : m_batchingKernel;
+
+			b3LauncherCL launcher( m_queue, k,"*batchingKernel");
+			if (!useNewBatchingKernel )
+			{
+				launcher.setBuffer( contacts->getBufferCL() );
+			}
+			launcher.setBuffer( m_contactBuffer2->getBufferCL() );
+			launcher.setBuffer( nNative->getBufferCL());
+			launcher.setBuffer( offsetsNative->getBufferCL());
+			
+			launcher.setBuffer(m_batchSizes.getBufferCL());
+			
+
+			//launcher.setConst(  cdata );
+            launcher.setConst(staticIdx);
+            
+			launcher.launch1D( numWorkItems, 64 );
+			//clFinish(m_queue);
+			//b3AlignedObjectArray<int> batchSizesCPU;
+			//m_batchSizes.copyToHost(batchSizesCPU);
+			//printf(".\n");
+		}
+
+#ifdef BATCH_DEBUG
+	aaaa
+		b3Contact4* hostContacts = new b3Contact4[nContacts];
+		m_contactBuffer->read(hostContacts,nContacts);
+		clFinish(m_queue);
+
+		gpuDebugInfo.read(debugInfo,numWorkItems);
+		clFinish(m_queue);
+
+		for (int i=0;i<numWorkItems;i++)
+		{
+			if (debugInfo[i].m_valInt1>0)
+			{
+				printf("catch\n");
+			}
+			if (debugInfo[i].m_valInt2>0)
+			{
+				printf("catch22\n");
+			}
+
+			if (debugInfo[i].m_valInt3>0)
+			{
+				printf("catch666\n");
+			}
+
+			if (debugInfo[i].m_valInt4>0)
+			{
+				printf("catch777\n");
+			}
+		}
+		delete[] debugInfo;
+#endif //BATCH_DEBUG
+
+	}
+
+//	copy buffer to buffer
+	//b3Assert(m_contactBuffer->size()==nContacts);
+	//contacts->copyFromOpenCLArray( *m_contactBuffer);
+	//clFinish(m_queue);//needed?
+	
+	
+	
+}
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3Solver.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3Solver.h
new file mode 100644
index 0000000000..b37f2f1bec
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/b3Solver.h
@@ -0,0 +1,126 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+#ifndef __ADL_SOLVER_H
+#define __ADL_SOLVER_H
+
+#include "Bullet3OpenCL/ParallelPrimitives/b3OpenCLArray.h"
+#include "b3GpuConstraint4.h"
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3RigidBodyData.h"
+#include "Bullet3Collision/NarrowPhaseCollision/b3Contact4.h"
+
+#include "Bullet3OpenCL/ParallelPrimitives/b3PrefixScanCL.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h"
+#include "Bullet3OpenCL/ParallelPrimitives/b3BoundSearchCL.h"
+
+#include "Bullet3OpenCL/Initialize/b3OpenCLUtils.h"
+
+
+#define B3NEXTMULTIPLEOF(num, alignment) (((num)/(alignment) + (((num)%(alignment)==0)?0:1))*(alignment))
+
+enum
+{
+	B3_SOLVER_N_SPLIT_X = 8,//16,//4,
+	B3_SOLVER_N_SPLIT_Y = 4,//16,//4,
+	B3_SOLVER_N_SPLIT_Z = 8,//,
+	B3_SOLVER_N_CELLS = B3_SOLVER_N_SPLIT_X*B3_SOLVER_N_SPLIT_Y*B3_SOLVER_N_SPLIT_Z,
+	B3_SOLVER_N_BATCHES = 8,//4,//8,//4,
+	B3_MAX_NUM_BATCHES = 128,
+};
+
+class b3SolverBase
+{
+	public:
+		
+
+		struct ConstraintCfg
+		{
+			ConstraintCfg( float dt = 0.f ): m_positionDrift( 0.005f ), m_positionConstraintCoeff( 0.2f ), m_dt(dt), m_staticIdx(-1) {}
+
+			float m_positionDrift;
+			float m_positionConstraintCoeff;
+			float m_dt;
+			bool m_enableParallelSolve;
+			float m_batchCellSize;
+			int m_staticIdx;
+		};
+		
+};
+
+class b3Solver : public b3SolverBase
+{
+	public:
+
+		cl_context m_context;
+		cl_device_id m_device;
+		cl_command_queue m_queue;
+				
+
+		b3OpenCLArray<unsigned int>* m_numConstraints;
+		b3OpenCLArray<unsigned int>* m_offsets;
+		b3OpenCLArray<int> m_batchSizes;
+		
+		
+		int m_nIterations;
+		cl_kernel m_batchingKernel;
+		cl_kernel m_batchingKernelNew;
+		cl_kernel m_solveContactKernel;
+		cl_kernel m_solveFrictionKernel;
+		cl_kernel m_contactToConstraintKernel;
+		cl_kernel m_setSortDataKernel;
+		cl_kernel m_reorderContactKernel;
+		cl_kernel m_copyConstraintKernel;
+
+		class b3RadixSort32CL*	m_sort32;
+		class b3BoundSearchCL*	m_search;
+		class b3PrefixScanCL*	m_scan;
+
+		b3OpenCLArray<b3SortData>* m_sortDataBuffer;
+		b3OpenCLArray<b3Contact4>* m_contactBuffer2;
+
+		enum
+		{
+			DYNAMIC_CONTACT_ALLOCATION_THRESHOLD = 2000000,
+		};
+
+		
+
+		
+		b3Solver(cl_context ctx, cl_device_id device, cl_command_queue queue, int pairCapacity);
+
+		virtual ~b3Solver();
+		
+		void solveContactConstraint( const b3OpenCLArray<b3RigidBodyData>* bodyBuf, const b3OpenCLArray<b3InertiaData>* inertiaBuf, 
+			b3OpenCLArray<b3GpuConstraint4>* constraint, void* additionalData, int n ,int maxNumBatches);
+
+		void solveContactConstraintHost(  b3OpenCLArray<b3RigidBodyData>* bodyBuf, b3OpenCLArray<b3InertiaData>* shapeBuf, 
+			b3OpenCLArray<b3GpuConstraint4>* constraint, void* additionalData, int n ,int maxNumBatches, b3AlignedObjectArray<int>* batchSizes);
+
+
+		void convertToConstraints( const b3OpenCLArray<b3RigidBodyData>* bodyBuf, 
+			const b3OpenCLArray<b3InertiaData>* shapeBuf, 
+			b3OpenCLArray<b3Contact4>* contactsIn, b3OpenCLArray<b3GpuConstraint4>* contactCOut, void* additionalData, 
+			int nContacts, const ConstraintCfg& cfg );
+
+		void	batchContacts( b3OpenCLArray<b3Contact4>* contacts, int nContacts, b3OpenCLArray<unsigned int>* n, b3OpenCLArray<unsigned int>* offsets, int staticIdx );
+
+};
+
+
+
+
+#endif //__ADL_SOLVER_H
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/batchingKernels.cl b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/batchingKernels.cl
new file mode 100644
index 0000000000..3b891b863d
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/batchingKernels.cl
@@ -0,0 +1,353 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Contact4Data.h"
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile __global int*
+#endif
+
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define make_float4 (float4)
+#define make_float2 (float2)
+#define make_uint4 (uint4)
+#define make_int4 (int4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+#define WG_SIZE 64
+
+
+
+
+
+typedef struct 
+{
+	int m_n;
+	int m_start;
+	int m_staticIdx;
+	int m_paddings[1];
+} ConstBuffer;
+
+typedef struct 
+{
+	int m_a;
+	int m_b;
+	u32 m_idx;
+}Elem;
+
+#define STACK_SIZE (WG_SIZE*10)
+//#define STACK_SIZE (WG_SIZE)
+#define RING_SIZE 1024
+#define RING_SIZE_MASK (RING_SIZE-1)
+#define CHECK_SIZE (WG_SIZE)
+
+
+#define GET_RING_CAPACITY (RING_SIZE - ldsRingEnd)
+#define RING_END ldsTmp
+
+u32 readBuf(__local u32* buff, int idx)
+{
+	idx = idx % (32*CHECK_SIZE);
+	int bitIdx = idx%32;
+	int bufIdx = idx/32;
+	return buff[bufIdx] & (1<<bitIdx);
+}
+
+void writeBuf(__local u32* buff, int idx)
+{
+	idx = idx % (32*CHECK_SIZE);
+	int bitIdx = idx%32;
+	int bufIdx = idx/32;
+//	buff[bufIdx] |= (1<<bitIdx);
+	atom_or( &buff[bufIdx], (1<<bitIdx) );
+}
+
+u32 tryWrite(__local u32* buff, int idx)
+{
+	idx = idx % (32*CHECK_SIZE);
+	int bitIdx = idx%32;
+	int bufIdx = idx/32;
+	u32 ans = (u32)atom_or( &buff[bufIdx], (1<<bitIdx) );
+	return ((ans >> bitIdx)&1) == 0;
+}
+
+//	batching on the GPU
+__kernel void CreateBatches( __global const struct b3Contact4Data* gConstraints, __global struct b3Contact4Data* gConstraintsOut,
+		__global const u32* gN, __global const u32* gStart, __global int* batchSizes, 
+		int m_staticIdx )
+{
+	__local u32 ldsStackIdx[STACK_SIZE];
+	__local u32 ldsStackEnd;
+	__local Elem ldsRingElem[RING_SIZE];
+	__local u32 ldsRingEnd;
+	__local u32 ldsTmp;
+	__local u32 ldsCheckBuffer[CHECK_SIZE];
+	__local u32 ldsFixedBuffer[CHECK_SIZE];
+	__local u32 ldsGEnd;
+	__local u32 ldsDstEnd;
+
+	int wgIdx = GET_GROUP_IDX;
+	int lIdx = GET_LOCAL_IDX;
+	
+	const int m_n = gN[wgIdx];
+	const int m_start = gStart[wgIdx];
+		
+	if( lIdx == 0 )
+	{
+		ldsRingEnd = 0;
+		ldsGEnd = 0;
+		ldsStackEnd = 0;
+		ldsDstEnd = m_start;
+	}
+	
+	
+	
+//	while(1)
+//was 250
+	int ie=0;
+	int maxBatch = 0;
+	for(ie=0; ie<50; ie++)
+	{
+		ldsFixedBuffer[lIdx] = 0;
+
+		for(int giter=0; giter<4; giter++)
+		{
+			int ringCap = GET_RING_CAPACITY;
+		
+			//	1. fill ring
+			if( ldsGEnd < m_n )
+			{
+				while( ringCap > WG_SIZE )
+				{
+					if( ldsGEnd >= m_n ) break;
+					if( lIdx < ringCap - WG_SIZE )
+					{
+						int srcIdx;
+						AtomInc1( ldsGEnd, srcIdx );
+						if( srcIdx < m_n )
+						{
+							int dstIdx;
+							AtomInc1( ldsRingEnd, dstIdx );
+							
+							int a = gConstraints[m_start+srcIdx].m_bodyAPtrAndSignBit;
+							int b = gConstraints[m_start+srcIdx].m_bodyBPtrAndSignBit;
+							ldsRingElem[dstIdx].m_a = (a>b)? b:a;
+							ldsRingElem[dstIdx].m_b = (a>b)? a:b;
+							ldsRingElem[dstIdx].m_idx = srcIdx;
+						}
+					}
+					ringCap = GET_RING_CAPACITY;
+				}
+			}
+
+			GROUP_LDS_BARRIER;
+	
+			//	2. fill stack
+			__local Elem* dst = ldsRingElem;
+			if( lIdx == 0 ) RING_END = 0;
+
+			int srcIdx=lIdx;
+			int end = ldsRingEnd;
+
+			{
+				for(int ii=0; ii<end; ii+=WG_SIZE, srcIdx+=WG_SIZE)
+				{
+					Elem e;
+					if(srcIdx<end) e = ldsRingElem[srcIdx];
+					bool done = (srcIdx<end)?false:true;
+
+					for(int i=lIdx; i<CHECK_SIZE; i+=WG_SIZE) ldsCheckBuffer[lIdx] = 0;
+					
+					if( !done )
+					{
+						int aUsed = readBuf( ldsFixedBuffer, abs(e.m_a));
+						int bUsed = readBuf( ldsFixedBuffer, abs(e.m_b));
+
+						if( aUsed==0 && bUsed==0 )
+						{
+							int aAvailable=1;
+							int bAvailable=1;
+							int ea = abs(e.m_a);
+							int eb = abs(e.m_b);
+
+							bool aStatic = (e.m_a<0) ||(ea==m_staticIdx);
+							bool bStatic = (e.m_b<0) ||(eb==m_staticIdx);
+							
+							if (!aStatic)
+								aAvailable = tryWrite( ldsCheckBuffer, ea );
+							if (!bStatic)
+								bAvailable = tryWrite( ldsCheckBuffer, eb );
+							
+							//aAvailable = aStatic? 1: aAvailable;
+							//bAvailable = bStatic? 1: bAvailable;
+
+							bool success = (aAvailable && bAvailable);
+							if(success)
+							{
+							
+								if (!aStatic)
+									writeBuf( ldsFixedBuffer, ea );
+								if (!bStatic)
+									writeBuf( ldsFixedBuffer, eb );
+							}
+							done = success;
+						}
+					}
+
+					//	put it aside
+					if(srcIdx<end)
+					{
+						if( done )
+						{
+							int dstIdx; AtomInc1( ldsStackEnd, dstIdx );
+							if( dstIdx < STACK_SIZE )
+								ldsStackIdx[dstIdx] = e.m_idx;
+							else{
+								done = false;
+								AtomAdd( ldsStackEnd, -1 );
+							}
+						}
+						if( !done )
+						{
+							int dstIdx; AtomInc1( RING_END, dstIdx );
+							dst[dstIdx] = e;
+						}
+					}
+
+					//	if filled, flush
+					if( ldsStackEnd == STACK_SIZE )
+					{
+						for(int i=lIdx; i<STACK_SIZE; i+=WG_SIZE)
+						{
+							int idx = m_start + ldsStackIdx[i];
+							int dstIdx; AtomInc1( ldsDstEnd, dstIdx );
+							gConstraintsOut[ dstIdx ] = gConstraints[ idx ];
+							gConstraintsOut[ dstIdx ].m_batchIdx = ie;
+						}
+						if( lIdx == 0 ) ldsStackEnd = 0;
+
+						//for(int i=lIdx; i<CHECK_SIZE; i+=WG_SIZE) 
+						ldsFixedBuffer[lIdx] = 0;
+					}
+				}
+			}
+
+			if( lIdx == 0 ) ldsRingEnd = RING_END;
+		}
+
+		GROUP_LDS_BARRIER;
+
+		for(int i=lIdx; i<ldsStackEnd; i+=WG_SIZE)
+		{
+			int idx = m_start + ldsStackIdx[i];
+			int dstIdx; AtomInc1( ldsDstEnd, dstIdx );
+			gConstraintsOut[ dstIdx ] = gConstraints[ idx ];
+			gConstraintsOut[ dstIdx ].m_batchIdx = ie;
+		}
+
+		//	in case it couldn't consume any pair. Flush them
+		//	todo. Serial batch worth while?
+		if( ldsStackEnd == 0 )
+		{
+			for(int i=lIdx; i<ldsRingEnd; i+=WG_SIZE)
+			{
+				int idx = m_start + ldsRingElem[i].m_idx;
+				int dstIdx; AtomInc1( ldsDstEnd, dstIdx );
+				gConstraintsOut[ dstIdx ] = gConstraints[ idx ];
+				int curBatch = 100+i;
+				if (maxBatch < curBatch)
+					maxBatch = curBatch;
+				
+				gConstraintsOut[ dstIdx ].m_batchIdx = curBatch;
+				
+			}
+			GROUP_LDS_BARRIER;
+			if( lIdx == 0 ) ldsRingEnd = 0;
+		}
+
+		if( lIdx == 0 ) ldsStackEnd = 0;
+
+		GROUP_LDS_BARRIER;
+
+		//	termination
+		if( ldsGEnd == m_n && ldsRingEnd == 0 )
+			break;
+	}
+
+	if( lIdx == 0 )
+	{
+		if (maxBatch < ie)
+			maxBatch=ie;
+		batchSizes[wgIdx]=maxBatch;
+	}
+
+}
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/batchingKernels.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/batchingKernels.h
new file mode 100644
index 0000000000..150eedc94b
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/batchingKernels.h
@@ -0,0 +1,388 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* batchingKernelsCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"#ifndef B3_CONTACT4DATA_H\n"
+"#define B3_CONTACT4DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#define B3_FLOAT4_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#define B3_PLATFORM_DEFINITIONS_H\n"
+"struct MyTest\n"
+"{\n"
+"	int bla;\n"
+"};\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"//keep B3_LARGE_FLOAT*B3_LARGE_FLOAT < FLT_MAX\n"
+"#define B3_LARGE_FLOAT 1e18f\n"
+"#define B3_INFINITY 1e18f\n"
+"#define b3Assert(a)\n"
+"#define b3ConstArray(a) __global const a*\n"
+"#define b3AtomicInc atomic_inc\n"
+"#define b3AtomicAdd atomic_add\n"
+"#define b3Fabs fabs\n"
+"#define b3Sqrt native_sqrt\n"
+"#define b3Sin native_sin\n"
+"#define b3Cos native_cos\n"
+"#define B3_STATIC\n"
+"#endif\n"
+"#endif\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Float4;\n"
+"	#define b3Float4ConstArg const b3Float4\n"
+"	#define b3MakeFloat4 (float4)\n"
+"	float b3Dot3F4(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return dot(a1, b1);\n"
+"	}\n"
+"	b3Float4 b3Cross3(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return cross(a1, b1);\n"
+"	}\n"
+"	#define b3MinFloat4 min\n"
+"	#define b3MaxFloat4 max\n"
+"	#define b3Normalized(a) normalize(a)\n"
+"#endif \n"
+"		\n"
+"inline bool b3IsAlmostZero(b3Float4ConstArg v)\n"
+"{\n"
+"	if(b3Fabs(v.x)>1e-6 || b3Fabs(v.y)>1e-6 || b3Fabs(v.z)>1e-6)	\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"inline int    b3MaxDot( b3Float4ConstArg vec, __global const b3Float4* vecArray, int vecLen, float* dotOut )\n"
+"{\n"
+"    float maxDot = -B3_INFINITY;\n"
+"    int i = 0;\n"
+"    int ptIndex = -1;\n"
+"    for( i = 0; i < vecLen; i++ )\n"
+"    {\n"
+"        float dot = b3Dot3F4(vecArray[i],vec);\n"
+"            \n"
+"        if( dot > maxDot )\n"
+"        {\n"
+"            maxDot = dot;\n"
+"            ptIndex = i;\n"
+"        }\n"
+"    }\n"
+"	b3Assert(ptIndex>=0);\n"
+"    if (ptIndex<0)\n"
+"	{\n"
+"		ptIndex = 0;\n"
+"	}\n"
+"    *dotOut = maxDot;\n"
+"    return ptIndex;\n"
+"}\n"
+"#endif //B3_FLOAT4_H\n"
+"typedef  struct b3Contact4Data b3Contact4Data_t;\n"
+"struct b3Contact4Data\n"
+"{\n"
+"	b3Float4	m_worldPosB[4];\n"
+"//	b3Float4	m_localPosA[4];\n"
+"//	b3Float4	m_localPosB[4];\n"
+"	b3Float4	m_worldNormalOnB;	//	w: m_nPoints\n"
+"	unsigned short  m_restituitionCoeffCmp;\n"
+"	unsigned short  m_frictionCoeffCmp;\n"
+"	int m_batchIdx;\n"
+"	int m_bodyAPtrAndSignBit;//x:m_bodyAPtr, y:m_bodyBPtr\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"	int	m_childIndexA;\n"
+"	int	m_childIndexB;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"};\n"
+"inline int b3Contact4Data_getNumPoints(const struct b3Contact4Data* contact)\n"
+"{\n"
+"	return (int)contact->m_worldNormalOnB.w;\n"
+"};\n"
+"inline void b3Contact4Data_setNumPoints(struct b3Contact4Data* contact, int numPoints)\n"
+"{\n"
+"	contact->m_worldNormalOnB.w = (float)numPoints;\n"
+"};\n"
+"#endif //B3_CONTACT4DATA_H\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile __global int*\n"
+"#endif\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"#define make_float4 (float4)\n"
+"#define make_float2 (float2)\n"
+"#define make_uint4 (uint4)\n"
+"#define make_int4 (int4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"#define WG_SIZE 64\n"
+"typedef struct \n"
+"{\n"
+"	int m_n;\n"
+"	int m_start;\n"
+"	int m_staticIdx;\n"
+"	int m_paddings[1];\n"
+"} ConstBuffer;\n"
+"typedef struct \n"
+"{\n"
+"	int m_a;\n"
+"	int m_b;\n"
+"	u32 m_idx;\n"
+"}Elem;\n"
+"#define STACK_SIZE (WG_SIZE*10)\n"
+"//#define STACK_SIZE (WG_SIZE)\n"
+"#define RING_SIZE 1024\n"
+"#define RING_SIZE_MASK (RING_SIZE-1)\n"
+"#define CHECK_SIZE (WG_SIZE)\n"
+"#define GET_RING_CAPACITY (RING_SIZE - ldsRingEnd)\n"
+"#define RING_END ldsTmp\n"
+"u32 readBuf(__local u32* buff, int idx)\n"
+"{\n"
+"	idx = idx % (32*CHECK_SIZE);\n"
+"	int bitIdx = idx%32;\n"
+"	int bufIdx = idx/32;\n"
+"	return buff[bufIdx] & (1<<bitIdx);\n"
+"}\n"
+"void writeBuf(__local u32* buff, int idx)\n"
+"{\n"
+"	idx = idx % (32*CHECK_SIZE);\n"
+"	int bitIdx = idx%32;\n"
+"	int bufIdx = idx/32;\n"
+"//	buff[bufIdx] |= (1<<bitIdx);\n"
+"	atom_or( &buff[bufIdx], (1<<bitIdx) );\n"
+"}\n"
+"u32 tryWrite(__local u32* buff, int idx)\n"
+"{\n"
+"	idx = idx % (32*CHECK_SIZE);\n"
+"	int bitIdx = idx%32;\n"
+"	int bufIdx = idx/32;\n"
+"	u32 ans = (u32)atom_or( &buff[bufIdx], (1<<bitIdx) );\n"
+"	return ((ans >> bitIdx)&1) == 0;\n"
+"}\n"
+"//	batching on the GPU\n"
+"__kernel void CreateBatches( __global const struct b3Contact4Data* gConstraints, __global struct b3Contact4Data* gConstraintsOut,\n"
+"		__global const u32* gN, __global const u32* gStart, __global int* batchSizes, \n"
+"		int m_staticIdx )\n"
+"{\n"
+"	__local u32 ldsStackIdx[STACK_SIZE];\n"
+"	__local u32 ldsStackEnd;\n"
+"	__local Elem ldsRingElem[RING_SIZE];\n"
+"	__local u32 ldsRingEnd;\n"
+"	__local u32 ldsTmp;\n"
+"	__local u32 ldsCheckBuffer[CHECK_SIZE];\n"
+"	__local u32 ldsFixedBuffer[CHECK_SIZE];\n"
+"	__local u32 ldsGEnd;\n"
+"	__local u32 ldsDstEnd;\n"
+"	int wgIdx = GET_GROUP_IDX;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	\n"
+"	const int m_n = gN[wgIdx];\n"
+"	const int m_start = gStart[wgIdx];\n"
+"		\n"
+"	if( lIdx == 0 )\n"
+"	{\n"
+"		ldsRingEnd = 0;\n"
+"		ldsGEnd = 0;\n"
+"		ldsStackEnd = 0;\n"
+"		ldsDstEnd = m_start;\n"
+"	}\n"
+"	\n"
+"	\n"
+"	\n"
+"//	while(1)\n"
+"//was 250\n"
+"	int ie=0;\n"
+"	int maxBatch = 0;\n"
+"	for(ie=0; ie<50; ie++)\n"
+"	{\n"
+"		ldsFixedBuffer[lIdx] = 0;\n"
+"		for(int giter=0; giter<4; giter++)\n"
+"		{\n"
+"			int ringCap = GET_RING_CAPACITY;\n"
+"		\n"
+"			//	1. fill ring\n"
+"			if( ldsGEnd < m_n )\n"
+"			{\n"
+"				while( ringCap > WG_SIZE )\n"
+"				{\n"
+"					if( ldsGEnd >= m_n ) break;\n"
+"					if( lIdx < ringCap - WG_SIZE )\n"
+"					{\n"
+"						int srcIdx;\n"
+"						AtomInc1( ldsGEnd, srcIdx );\n"
+"						if( srcIdx < m_n )\n"
+"						{\n"
+"							int dstIdx;\n"
+"							AtomInc1( ldsRingEnd, dstIdx );\n"
+"							\n"
+"							int a = gConstraints[m_start+srcIdx].m_bodyAPtrAndSignBit;\n"
+"							int b = gConstraints[m_start+srcIdx].m_bodyBPtrAndSignBit;\n"
+"							ldsRingElem[dstIdx].m_a = (a>b)? b:a;\n"
+"							ldsRingElem[dstIdx].m_b = (a>b)? a:b;\n"
+"							ldsRingElem[dstIdx].m_idx = srcIdx;\n"
+"						}\n"
+"					}\n"
+"					ringCap = GET_RING_CAPACITY;\n"
+"				}\n"
+"			}\n"
+"			GROUP_LDS_BARRIER;\n"
+"	\n"
+"			//	2. fill stack\n"
+"			__local Elem* dst = ldsRingElem;\n"
+"			if( lIdx == 0 ) RING_END = 0;\n"
+"			int srcIdx=lIdx;\n"
+"			int end = ldsRingEnd;\n"
+"			{\n"
+"				for(int ii=0; ii<end; ii+=WG_SIZE, srcIdx+=WG_SIZE)\n"
+"				{\n"
+"					Elem e;\n"
+"					if(srcIdx<end) e = ldsRingElem[srcIdx];\n"
+"					bool done = (srcIdx<end)?false:true;\n"
+"					for(int i=lIdx; i<CHECK_SIZE; i+=WG_SIZE) ldsCheckBuffer[lIdx] = 0;\n"
+"					\n"
+"					if( !done )\n"
+"					{\n"
+"						int aUsed = readBuf( ldsFixedBuffer, abs(e.m_a));\n"
+"						int bUsed = readBuf( ldsFixedBuffer, abs(e.m_b));\n"
+"						if( aUsed==0 && bUsed==0 )\n"
+"						{\n"
+"							int aAvailable=1;\n"
+"							int bAvailable=1;\n"
+"							int ea = abs(e.m_a);\n"
+"							int eb = abs(e.m_b);\n"
+"							bool aStatic = (e.m_a<0) ||(ea==m_staticIdx);\n"
+"							bool bStatic = (e.m_b<0) ||(eb==m_staticIdx);\n"
+"							\n"
+"							if (!aStatic)\n"
+"								aAvailable = tryWrite( ldsCheckBuffer, ea );\n"
+"							if (!bStatic)\n"
+"								bAvailable = tryWrite( ldsCheckBuffer, eb );\n"
+"							\n"
+"							//aAvailable = aStatic? 1: aAvailable;\n"
+"							//bAvailable = bStatic? 1: bAvailable;\n"
+"							bool success = (aAvailable && bAvailable);\n"
+"							if(success)\n"
+"							{\n"
+"							\n"
+"								if (!aStatic)\n"
+"									writeBuf( ldsFixedBuffer, ea );\n"
+"								if (!bStatic)\n"
+"									writeBuf( ldsFixedBuffer, eb );\n"
+"							}\n"
+"							done = success;\n"
+"						}\n"
+"					}\n"
+"					//	put it aside\n"
+"					if(srcIdx<end)\n"
+"					{\n"
+"						if( done )\n"
+"						{\n"
+"							int dstIdx; AtomInc1( ldsStackEnd, dstIdx );\n"
+"							if( dstIdx < STACK_SIZE )\n"
+"								ldsStackIdx[dstIdx] = e.m_idx;\n"
+"							else{\n"
+"								done = false;\n"
+"								AtomAdd( ldsStackEnd, -1 );\n"
+"							}\n"
+"						}\n"
+"						if( !done )\n"
+"						{\n"
+"							int dstIdx; AtomInc1( RING_END, dstIdx );\n"
+"							dst[dstIdx] = e;\n"
+"						}\n"
+"					}\n"
+"					//	if filled, flush\n"
+"					if( ldsStackEnd == STACK_SIZE )\n"
+"					{\n"
+"						for(int i=lIdx; i<STACK_SIZE; i+=WG_SIZE)\n"
+"						{\n"
+"							int idx = m_start + ldsStackIdx[i];\n"
+"							int dstIdx; AtomInc1( ldsDstEnd, dstIdx );\n"
+"							gConstraintsOut[ dstIdx ] = gConstraints[ idx ];\n"
+"							gConstraintsOut[ dstIdx ].m_batchIdx = ie;\n"
+"						}\n"
+"						if( lIdx == 0 ) ldsStackEnd = 0;\n"
+"						//for(int i=lIdx; i<CHECK_SIZE; i+=WG_SIZE) \n"
+"						ldsFixedBuffer[lIdx] = 0;\n"
+"					}\n"
+"				}\n"
+"			}\n"
+"			if( lIdx == 0 ) ldsRingEnd = RING_END;\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"		for(int i=lIdx; i<ldsStackEnd; i+=WG_SIZE)\n"
+"		{\n"
+"			int idx = m_start + ldsStackIdx[i];\n"
+"			int dstIdx; AtomInc1( ldsDstEnd, dstIdx );\n"
+"			gConstraintsOut[ dstIdx ] = gConstraints[ idx ];\n"
+"			gConstraintsOut[ dstIdx ].m_batchIdx = ie;\n"
+"		}\n"
+"		//	in case it couldn't consume any pair. Flush them\n"
+"		//	todo. Serial batch worth while?\n"
+"		if( ldsStackEnd == 0 )\n"
+"		{\n"
+"			for(int i=lIdx; i<ldsRingEnd; i+=WG_SIZE)\n"
+"			{\n"
+"				int idx = m_start + ldsRingElem[i].m_idx;\n"
+"				int dstIdx; AtomInc1( ldsDstEnd, dstIdx );\n"
+"				gConstraintsOut[ dstIdx ] = gConstraints[ idx ];\n"
+"				int curBatch = 100+i;\n"
+"				if (maxBatch < curBatch)\n"
+"					maxBatch = curBatch;\n"
+"				\n"
+"				gConstraintsOut[ dstIdx ].m_batchIdx = curBatch;\n"
+"				\n"
+"			}\n"
+"			GROUP_LDS_BARRIER;\n"
+"			if( lIdx == 0 ) ldsRingEnd = 0;\n"
+"		}\n"
+"		if( lIdx == 0 ) ldsStackEnd = 0;\n"
+"		GROUP_LDS_BARRIER;\n"
+"		//	termination\n"
+"		if( ldsGEnd == m_n && ldsRingEnd == 0 )\n"
+"			break;\n"
+"	}\n"
+"	if( lIdx == 0 )\n"
+"	{\n"
+"		if (maxBatch < ie)\n"
+"			maxBatch=ie;\n"
+"		batchSizes[wgIdx]=maxBatch;\n"
+"	}\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/batchingKernelsNew.cl b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/batchingKernelsNew.cl
new file mode 100644
index 0000000000..ba1b66d2c3
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/batchingKernelsNew.cl
@@ -0,0 +1,231 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Contact4Data.h"
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile __global int*
+#endif
+
+#define SIMD_WIDTH 64
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define make_float4 (float4)
+#define make_float2 (float2)
+#define make_uint4 (uint4)
+#define make_int4 (int4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+#define WG_SIZE 64
+
+
+
+
+
+typedef struct 
+{
+	int m_n;
+	int m_start;
+	int m_staticIdx;
+	int m_paddings[1];
+} ConstBuffer;
+
+typedef struct 
+{
+	int m_a;
+	int m_b;
+	u32 m_idx;
+}Elem;
+
+
+
+
+
+//	batching on the GPU
+__kernel void CreateBatchesBruteForce( __global struct b3Contact4Data* gConstraints, 	__global const u32* gN, __global const u32* gStart, int m_staticIdx )
+{
+	int wgIdx = GET_GROUP_IDX;
+	int lIdx = GET_LOCAL_IDX;
+	
+	const int m_n = gN[wgIdx];
+	const int m_start = gStart[wgIdx];
+		
+	if( lIdx == 0 )
+	{
+		for (int i=0;i<m_n;i++)
+		{
+			int srcIdx = i+m_start;
+			int batchIndex = i;
+			gConstraints[ srcIdx ].m_batchIdx = batchIndex;	
+		}
+	}
+}
+
+
+#define CHECK_SIZE (WG_SIZE)
+
+
+
+
+u32 readBuf(__local u32* buff, int idx)
+{
+	idx = idx % (32*CHECK_SIZE);
+	int bitIdx = idx%32;
+	int bufIdx = idx/32;
+	return buff[bufIdx] & (1<<bitIdx);
+}
+
+void writeBuf(__local u32* buff, int idx)
+{
+	idx = idx % (32*CHECK_SIZE);
+	int bitIdx = idx%32;
+	int bufIdx = idx/32;
+	buff[bufIdx] |= (1<<bitIdx);
+	//atom_or( &buff[bufIdx], (1<<bitIdx) );
+}
+
+u32 tryWrite(__local u32* buff, int idx)
+{
+	idx = idx % (32*CHECK_SIZE);
+	int bitIdx = idx%32;
+	int bufIdx = idx/32;
+	u32 ans = (u32)atom_or( &buff[bufIdx], (1<<bitIdx) );
+	return ((ans >> bitIdx)&1) == 0;
+}
+
+
+//	batching on the GPU
+__kernel void CreateBatchesNew( __global struct b3Contact4Data* gConstraints, __global const u32* gN, __global const u32* gStart, __global int* batchSizes, int staticIdx )
+{
+	int wgIdx = GET_GROUP_IDX;
+	int lIdx = GET_LOCAL_IDX;
+	const int numConstraints = gN[wgIdx];
+	const int m_start = gStart[wgIdx];
+	b3Contact4Data_t tmp;
+	
+	__local u32 ldsFixedBuffer[CHECK_SIZE];
+		
+	
+	
+	
+	
+	if( lIdx == 0 )
+	{
+	
+		
+		__global struct b3Contact4Data* cs = &gConstraints[m_start];	
+	
+		
+		int numValidConstraints = 0;
+		int batchIdx = 0;
+
+		while( numValidConstraints < numConstraints)
+		{
+			int nCurrentBatch = 0;
+			//	clear flag
+	
+			for(int i=0; i<CHECK_SIZE; i++) 
+				ldsFixedBuffer[i] = 0;		
+
+			for(int i=numValidConstraints; i<numConstraints; i++)
+			{
+
+				int bodyAS = cs[i].m_bodyAPtrAndSignBit;
+				int bodyBS = cs[i].m_bodyBPtrAndSignBit;
+				int bodyA = abs(bodyAS);
+				int bodyB = abs(bodyBS);
+				bool aIsStatic = (bodyAS<0) || bodyAS==staticIdx;
+				bool bIsStatic = (bodyBS<0) || bodyBS==staticIdx;
+				int aUnavailable = aIsStatic ? 0 : readBuf( ldsFixedBuffer, bodyA);
+				int bUnavailable = bIsStatic ? 0 : readBuf( ldsFixedBuffer, bodyB);
+				
+				if( aUnavailable==0 && bUnavailable==0 ) // ok
+				{
+					if (!aIsStatic)
+					{
+						writeBuf( ldsFixedBuffer, bodyA );
+					}
+					if (!bIsStatic)
+					{
+						writeBuf( ldsFixedBuffer, bodyB );
+					}
+
+					cs[i].m_batchIdx = batchIdx;
+
+					if (i!=numValidConstraints)
+					{
+
+						tmp = cs[i];
+						cs[i] = cs[numValidConstraints];
+						cs[numValidConstraints]  = tmp;
+
+
+					}
+
+					numValidConstraints++;
+					
+					nCurrentBatch++;
+					if( nCurrentBatch == SIMD_WIDTH)
+					{
+						nCurrentBatch = 0;
+						for(int i=0; i<CHECK_SIZE; i++) 
+							ldsFixedBuffer[i] = 0;
+						
+					}
+				}
+			}//for
+			batchIdx ++;
+		}//while
+		
+		batchSizes[wgIdx] = batchIdx;
+
+	}//if( lIdx == 0 )
+	
+	//return batchIdx;
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/batchingKernelsNew.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/batchingKernelsNew.h
new file mode 100644
index 0000000000..1e5957adae
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/batchingKernelsNew.h
@@ -0,0 +1,291 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* batchingKernelsNewCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Erwin Coumans\n"
+"#ifndef B3_CONTACT4DATA_H\n"
+"#define B3_CONTACT4DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#define B3_FLOAT4_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#define B3_PLATFORM_DEFINITIONS_H\n"
+"struct MyTest\n"
+"{\n"
+"	int bla;\n"
+"};\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"//keep B3_LARGE_FLOAT*B3_LARGE_FLOAT < FLT_MAX\n"
+"#define B3_LARGE_FLOAT 1e18f\n"
+"#define B3_INFINITY 1e18f\n"
+"#define b3Assert(a)\n"
+"#define b3ConstArray(a) __global const a*\n"
+"#define b3AtomicInc atomic_inc\n"
+"#define b3AtomicAdd atomic_add\n"
+"#define b3Fabs fabs\n"
+"#define b3Sqrt native_sqrt\n"
+"#define b3Sin native_sin\n"
+"#define b3Cos native_cos\n"
+"#define B3_STATIC\n"
+"#endif\n"
+"#endif\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Float4;\n"
+"	#define b3Float4ConstArg const b3Float4\n"
+"	#define b3MakeFloat4 (float4)\n"
+"	float b3Dot3F4(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return dot(a1, b1);\n"
+"	}\n"
+"	b3Float4 b3Cross3(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return cross(a1, b1);\n"
+"	}\n"
+"	#define b3MinFloat4 min\n"
+"	#define b3MaxFloat4 max\n"
+"	#define b3Normalized(a) normalize(a)\n"
+"#endif \n"
+"		\n"
+"inline bool b3IsAlmostZero(b3Float4ConstArg v)\n"
+"{\n"
+"	if(b3Fabs(v.x)>1e-6 || b3Fabs(v.y)>1e-6 || b3Fabs(v.z)>1e-6)	\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"inline int    b3MaxDot( b3Float4ConstArg vec, __global const b3Float4* vecArray, int vecLen, float* dotOut )\n"
+"{\n"
+"    float maxDot = -B3_INFINITY;\n"
+"    int i = 0;\n"
+"    int ptIndex = -1;\n"
+"    for( i = 0; i < vecLen; i++ )\n"
+"    {\n"
+"        float dot = b3Dot3F4(vecArray[i],vec);\n"
+"            \n"
+"        if( dot > maxDot )\n"
+"        {\n"
+"            maxDot = dot;\n"
+"            ptIndex = i;\n"
+"        }\n"
+"    }\n"
+"	b3Assert(ptIndex>=0);\n"
+"    if (ptIndex<0)\n"
+"	{\n"
+"		ptIndex = 0;\n"
+"	}\n"
+"    *dotOut = maxDot;\n"
+"    return ptIndex;\n"
+"}\n"
+"#endif //B3_FLOAT4_H\n"
+"typedef  struct b3Contact4Data b3Contact4Data_t;\n"
+"struct b3Contact4Data\n"
+"{\n"
+"	b3Float4	m_worldPosB[4];\n"
+"//	b3Float4	m_localPosA[4];\n"
+"//	b3Float4	m_localPosB[4];\n"
+"	b3Float4	m_worldNormalOnB;	//	w: m_nPoints\n"
+"	unsigned short  m_restituitionCoeffCmp;\n"
+"	unsigned short  m_frictionCoeffCmp;\n"
+"	int m_batchIdx;\n"
+"	int m_bodyAPtrAndSignBit;//x:m_bodyAPtr, y:m_bodyBPtr\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"	int	m_childIndexA;\n"
+"	int	m_childIndexB;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"};\n"
+"inline int b3Contact4Data_getNumPoints(const struct b3Contact4Data* contact)\n"
+"{\n"
+"	return (int)contact->m_worldNormalOnB.w;\n"
+"};\n"
+"inline void b3Contact4Data_setNumPoints(struct b3Contact4Data* contact, int numPoints)\n"
+"{\n"
+"	contact->m_worldNormalOnB.w = (float)numPoints;\n"
+"};\n"
+"#endif //B3_CONTACT4DATA_H\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile __global int*\n"
+"#endif\n"
+"#define SIMD_WIDTH 64\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"#define make_float4 (float4)\n"
+"#define make_float2 (float2)\n"
+"#define make_uint4 (uint4)\n"
+"#define make_int4 (int4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"#define WG_SIZE 64\n"
+"typedef struct \n"
+"{\n"
+"	int m_n;\n"
+"	int m_start;\n"
+"	int m_staticIdx;\n"
+"	int m_paddings[1];\n"
+"} ConstBuffer;\n"
+"typedef struct \n"
+"{\n"
+"	int m_a;\n"
+"	int m_b;\n"
+"	u32 m_idx;\n"
+"}Elem;\n"
+"//	batching on the GPU\n"
+"__kernel void CreateBatchesBruteForce( __global struct b3Contact4Data* gConstraints, 	__global const u32* gN, __global const u32* gStart, int m_staticIdx )\n"
+"{\n"
+"	int wgIdx = GET_GROUP_IDX;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	\n"
+"	const int m_n = gN[wgIdx];\n"
+"	const int m_start = gStart[wgIdx];\n"
+"		\n"
+"	if( lIdx == 0 )\n"
+"	{\n"
+"		for (int i=0;i<m_n;i++)\n"
+"		{\n"
+"			int srcIdx = i+m_start;\n"
+"			int batchIndex = i;\n"
+"			gConstraints[ srcIdx ].m_batchIdx = batchIndex;	\n"
+"		}\n"
+"	}\n"
+"}\n"
+"#define CHECK_SIZE (WG_SIZE)\n"
+"u32 readBuf(__local u32* buff, int idx)\n"
+"{\n"
+"	idx = idx % (32*CHECK_SIZE);\n"
+"	int bitIdx = idx%32;\n"
+"	int bufIdx = idx/32;\n"
+"	return buff[bufIdx] & (1<<bitIdx);\n"
+"}\n"
+"void writeBuf(__local u32* buff, int idx)\n"
+"{\n"
+"	idx = idx % (32*CHECK_SIZE);\n"
+"	int bitIdx = idx%32;\n"
+"	int bufIdx = idx/32;\n"
+"	buff[bufIdx] |= (1<<bitIdx);\n"
+"	//atom_or( &buff[bufIdx], (1<<bitIdx) );\n"
+"}\n"
+"u32 tryWrite(__local u32* buff, int idx)\n"
+"{\n"
+"	idx = idx % (32*CHECK_SIZE);\n"
+"	int bitIdx = idx%32;\n"
+"	int bufIdx = idx/32;\n"
+"	u32 ans = (u32)atom_or( &buff[bufIdx], (1<<bitIdx) );\n"
+"	return ((ans >> bitIdx)&1) == 0;\n"
+"}\n"
+"//	batching on the GPU\n"
+"__kernel void CreateBatchesNew( __global struct b3Contact4Data* gConstraints, __global const u32* gN, __global const u32* gStart, __global int* batchSizes, int staticIdx )\n"
+"{\n"
+"	int wgIdx = GET_GROUP_IDX;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	const int numConstraints = gN[wgIdx];\n"
+"	const int m_start = gStart[wgIdx];\n"
+"	b3Contact4Data_t tmp;\n"
+"	\n"
+"	__local u32 ldsFixedBuffer[CHECK_SIZE];\n"
+"		\n"
+"	\n"
+"	\n"
+"	\n"
+"	\n"
+"	if( lIdx == 0 )\n"
+"	{\n"
+"	\n"
+"		\n"
+"		__global struct b3Contact4Data* cs = &gConstraints[m_start];	\n"
+"	\n"
+"		\n"
+"		int numValidConstraints = 0;\n"
+"		int batchIdx = 0;\n"
+"		while( numValidConstraints < numConstraints)\n"
+"		{\n"
+"			int nCurrentBatch = 0;\n"
+"			//	clear flag\n"
+"	\n"
+"			for(int i=0; i<CHECK_SIZE; i++) \n"
+"				ldsFixedBuffer[i] = 0;		\n"
+"			for(int i=numValidConstraints; i<numConstraints; i++)\n"
+"			{\n"
+"				int bodyAS = cs[i].m_bodyAPtrAndSignBit;\n"
+"				int bodyBS = cs[i].m_bodyBPtrAndSignBit;\n"
+"				int bodyA = abs(bodyAS);\n"
+"				int bodyB = abs(bodyBS);\n"
+"				bool aIsStatic = (bodyAS<0) || bodyAS==staticIdx;\n"
+"				bool bIsStatic = (bodyBS<0) || bodyBS==staticIdx;\n"
+"				int aUnavailable = aIsStatic ? 0 : readBuf( ldsFixedBuffer, bodyA);\n"
+"				int bUnavailable = bIsStatic ? 0 : readBuf( ldsFixedBuffer, bodyB);\n"
+"				\n"
+"				if( aUnavailable==0 && bUnavailable==0 ) // ok\n"
+"				{\n"
+"					if (!aIsStatic)\n"
+"					{\n"
+"						writeBuf( ldsFixedBuffer, bodyA );\n"
+"					}\n"
+"					if (!bIsStatic)\n"
+"					{\n"
+"						writeBuf( ldsFixedBuffer, bodyB );\n"
+"					}\n"
+"					cs[i].m_batchIdx = batchIdx;\n"
+"					if (i!=numValidConstraints)\n"
+"					{\n"
+"						tmp = cs[i];\n"
+"						cs[i] = cs[numValidConstraints];\n"
+"						cs[numValidConstraints]  = tmp;\n"
+"					}\n"
+"					numValidConstraints++;\n"
+"					\n"
+"					nCurrentBatch++;\n"
+"					if( nCurrentBatch == SIMD_WIDTH)\n"
+"					{\n"
+"						nCurrentBatch = 0;\n"
+"						for(int i=0; i<CHECK_SIZE; i++) \n"
+"							ldsFixedBuffer[i] = 0;\n"
+"						\n"
+"					}\n"
+"				}\n"
+"			}//for\n"
+"			batchIdx ++;\n"
+"		}//while\n"
+"		\n"
+"		batchSizes[wgIdx] = batchIdx;\n"
+"	}//if( lIdx == 0 )\n"
+"	\n"
+"	//return batchIdx;\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/integrateKernel.cl b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/integrateKernel.cl
new file mode 100644
index 0000000000..e22bc9bc33
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/integrateKernel.cl
@@ -0,0 +1,32 @@
+/*
+Copyright (c) 2013 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3RigidBodyData.h"
+
+#include "Bullet3Dynamics/shared/b3IntegrateTransforms.h"
+
+
+
+__kernel void 
+  integrateTransformsKernel( __global b3RigidBodyData_t* bodies,const int numNodes, float timeStep, float angularDamping, float4 gravityAcceleration)
+{
+	int nodeID = get_global_id(0);
+	
+	if( nodeID < numNodes)
+	{
+		integrateSingleTransform(bodies,nodeID, timeStep, angularDamping,gravityAcceleration);
+	}
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/integrateKernel.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/integrateKernel.h
new file mode 100644
index 0000000000..a5a432947c
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/integrateKernel.h
@@ -0,0 +1,433 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* integrateKernelCL= \
+"/*\n"
+"Copyright (c) 2013 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Erwin Coumans\n"
+"#ifndef B3_RIGIDBODY_DATA_H\n"
+"#define B3_RIGIDBODY_DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#define B3_FLOAT4_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#define B3_PLATFORM_DEFINITIONS_H\n"
+"struct MyTest\n"
+"{\n"
+"	int bla;\n"
+"};\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"//keep B3_LARGE_FLOAT*B3_LARGE_FLOAT < FLT_MAX\n"
+"#define B3_LARGE_FLOAT 1e18f\n"
+"#define B3_INFINITY 1e18f\n"
+"#define b3Assert(a)\n"
+"#define b3ConstArray(a) __global const a*\n"
+"#define b3AtomicInc atomic_inc\n"
+"#define b3AtomicAdd atomic_add\n"
+"#define b3Fabs fabs\n"
+"#define b3Sqrt native_sqrt\n"
+"#define b3Sin native_sin\n"
+"#define b3Cos native_cos\n"
+"#define B3_STATIC\n"
+"#endif\n"
+"#endif\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Float4;\n"
+"	#define b3Float4ConstArg const b3Float4\n"
+"	#define b3MakeFloat4 (float4)\n"
+"	float b3Dot3F4(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return dot(a1, b1);\n"
+"	}\n"
+"	b3Float4 b3Cross3(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return cross(a1, b1);\n"
+"	}\n"
+"	#define b3MinFloat4 min\n"
+"	#define b3MaxFloat4 max\n"
+"	#define b3Normalized(a) normalize(a)\n"
+"#endif \n"
+"		\n"
+"inline bool b3IsAlmostZero(b3Float4ConstArg v)\n"
+"{\n"
+"	if(b3Fabs(v.x)>1e-6 || b3Fabs(v.y)>1e-6 || b3Fabs(v.z)>1e-6)	\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"inline int    b3MaxDot( b3Float4ConstArg vec, __global const b3Float4* vecArray, int vecLen, float* dotOut )\n"
+"{\n"
+"    float maxDot = -B3_INFINITY;\n"
+"    int i = 0;\n"
+"    int ptIndex = -1;\n"
+"    for( i = 0; i < vecLen; i++ )\n"
+"    {\n"
+"        float dot = b3Dot3F4(vecArray[i],vec);\n"
+"            \n"
+"        if( dot > maxDot )\n"
+"        {\n"
+"            maxDot = dot;\n"
+"            ptIndex = i;\n"
+"        }\n"
+"    }\n"
+"	b3Assert(ptIndex>=0);\n"
+"    if (ptIndex<0)\n"
+"	{\n"
+"		ptIndex = 0;\n"
+"	}\n"
+"    *dotOut = maxDot;\n"
+"    return ptIndex;\n"
+"}\n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_QUAT_H\n"
+"#define B3_QUAT_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif\n"
+"#endif\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Quat;\n"
+"	#define b3QuatConstArg const b3Quat\n"
+"	\n"
+"	\n"
+"inline float4 b3FastNormalize4(float4 v)\n"
+"{\n"
+"	v = (float4)(v.xyz,0.f);\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"	\n"
+"inline b3Quat b3QuatMul(b3Quat a, b3Quat b);\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in);\n"
+"inline b3Quat b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec);\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatMul(b3QuatConstArg a, b3QuatConstArg b)\n"
+"{\n"
+"	b3Quat ans;\n"
+"	ans = b3Cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - b3Dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in)\n"
+"{\n"
+"	b3Quat q;\n"
+"	q=in;\n"
+"	//return b3FastNormalize4(in);\n"
+"	float len = native_sqrt(dot(q, q));\n"
+"	if(len > 0.f)\n"
+"	{\n"
+"		q *= 1.f / len;\n"
+"	}\n"
+"	else\n"
+"	{\n"
+"		q.x = q.y = q.z = 0.f;\n"
+"		q.w = 1.f;\n"
+"	}\n"
+"	return q;\n"
+"}\n"
+"inline float4 b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	b3Quat qInv = b3QuatInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = b3QuatMul(b3QuatMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline float4 b3QuatInvRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	return b3QuatRotate( b3QuatInvert( q ), vec );\n"
+"}\n"
+"inline b3Float4 b3TransformPoint(b3Float4ConstArg point, b3Float4ConstArg translation, b3QuatConstArg  orientation)\n"
+"{\n"
+"	return b3QuatRotate( orientation, point ) + (translation);\n"
+"}\n"
+"	\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"#ifndef B3_MAT3x3_H\n"
+"#define B3_MAT3x3_H\n"
+"#ifndef B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"typedef struct\n"
+"{\n"
+"	b3Float4 m_row[3];\n"
+"}b3Mat3x3;\n"
+"#define b3Mat3x3ConstArg const b3Mat3x3\n"
+"#define b3GetRow(m,row) (m.m_row[row])\n"
+"inline b3Mat3x3 b3QuatGetRotationMatrix(b3Quat quat)\n"
+"{\n"
+"	b3Float4 quat2 = (b3Float4)(quat.x*quat.x, quat.y*quat.y, quat.z*quat.z, 0.f);\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0].x=1-2*quat2.y-2*quat2.z;\n"
+"	out.m_row[0].y=2*quat.x*quat.y-2*quat.w*quat.z;\n"
+"	out.m_row[0].z=2*quat.x*quat.z+2*quat.w*quat.y;\n"
+"	out.m_row[0].w = 0.f;\n"
+"	out.m_row[1].x=2*quat.x*quat.y+2*quat.w*quat.z;\n"
+"	out.m_row[1].y=1-2*quat2.x-2*quat2.z;\n"
+"	out.m_row[1].z=2*quat.y*quat.z-2*quat.w*quat.x;\n"
+"	out.m_row[1].w = 0.f;\n"
+"	out.m_row[2].x=2*quat.x*quat.z-2*quat.w*quat.y;\n"
+"	out.m_row[2].y=2*quat.y*quat.z+2*quat.w*quat.x;\n"
+"	out.m_row[2].z=1-2*quat2.x-2*quat2.y;\n"
+"	out.m_row[2].w = 0.f;\n"
+"	return out;\n"
+"}\n"
+"inline b3Mat3x3 b3AbsoluteMat3x3(b3Mat3x3ConstArg matIn)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = fabs(matIn.m_row[0]);\n"
+"	out.m_row[1] = fabs(matIn.m_row[1]);\n"
+"	out.m_row[2] = fabs(matIn.m_row[2]);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtZero();\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity();\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m);\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b);\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Mat3x3 mtZero()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(0.f);\n"
+"	m.m_row[1] = (b3Float4)(0.f);\n"
+"	m.m_row[2] = (b3Float4)(0.f);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(1,0,0,0);\n"
+"	m.m_row[1] = (b3Float4)(0,1,0,0);\n"
+"	m.m_row[2] = (b3Float4)(0,0,1,0);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = (b3Float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (b3Float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (b3Float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Mat3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	b3Mat3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = b3Dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = b3Dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = b3Dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b)\n"
+"{\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a.m_row[0], b );\n"
+"	ans.y = b3Dot3F4( a.m_row[1], b );\n"
+"	ans.z = b3Dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Float4 colx = b3MakeFloat4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	b3Float4 coly = b3MakeFloat4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	b3Float4 colz = b3MakeFloat4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a, colx );\n"
+"	ans.y = b3Dot3F4( a, coly );\n"
+"	ans.z = b3Dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"#endif\n"
+"#endif //B3_MAT3x3_H\n"
+"typedef struct b3RigidBodyData b3RigidBodyData_t;\n"
+"struct b3RigidBodyData\n"
+"{\n"
+"	b3Float4				m_pos;\n"
+"	b3Quat					m_quat;\n"
+"	b3Float4				m_linVel;\n"
+"	b3Float4				m_angVel;\n"
+"	int 					m_collidableIdx;\n"
+"	float 				m_invMass;\n"
+"	float 				m_restituitionCoeff;\n"
+"	float 				m_frictionCoeff;\n"
+"};\n"
+"typedef struct b3InertiaData b3InertiaData_t;\n"
+"struct b3InertiaData\n"
+"{\n"
+"	b3Mat3x3 m_invInertiaWorld;\n"
+"	b3Mat3x3 m_initInvInertia;\n"
+"};\n"
+"#endif //B3_RIGIDBODY_DATA_H\n"
+"	\n"
+"#ifndef B3_RIGIDBODY_DATA_H\n"
+"#endif //B3_RIGIDBODY_DATA_H\n"
+"	\n"
+"inline void integrateSingleTransform( __global b3RigidBodyData_t* bodies,int nodeID, float timeStep, float angularDamping, b3Float4ConstArg gravityAcceleration)\n"
+"{\n"
+"	\n"
+"	if (bodies[nodeID].m_invMass != 0.f)\n"
+"	{\n"
+"		float BT_GPU_ANGULAR_MOTION_THRESHOLD = (0.25f * 3.14159254f);\n"
+"		//angular velocity\n"
+"		{\n"
+"			b3Float4 axis;\n"
+"			//add some hardcoded angular damping\n"
+"			bodies[nodeID].m_angVel.x *= angularDamping;\n"
+"			bodies[nodeID].m_angVel.y *= angularDamping;\n"
+"			bodies[nodeID].m_angVel.z *= angularDamping;\n"
+"			\n"
+"			b3Float4 angvel = bodies[nodeID].m_angVel;\n"
+"			float fAngle = b3Sqrt(b3Dot3F4(angvel, angvel));\n"
+"			\n"
+"			//limit the angular motion\n"
+"			if(fAngle*timeStep > BT_GPU_ANGULAR_MOTION_THRESHOLD)\n"
+"			{\n"
+"				fAngle = BT_GPU_ANGULAR_MOTION_THRESHOLD / timeStep;\n"
+"			}\n"
+"			if(fAngle < 0.001f)\n"
+"			{\n"
+"				// use Taylor's expansions of sync function\n"
+"				axis = angvel * (0.5f*timeStep-(timeStep*timeStep*timeStep)*0.020833333333f * fAngle * fAngle);\n"
+"			}\n"
+"			else\n"
+"			{\n"
+"				// sync(fAngle) = sin(c*fAngle)/t\n"
+"				axis = angvel * ( b3Sin(0.5f * fAngle * timeStep) / fAngle);\n"
+"			}\n"
+"			\n"
+"			b3Quat dorn;\n"
+"			dorn.x = axis.x;\n"
+"			dorn.y = axis.y;\n"
+"			dorn.z = axis.z;\n"
+"			dorn.w = b3Cos(fAngle * timeStep * 0.5f);\n"
+"			b3Quat orn0 = bodies[nodeID].m_quat;\n"
+"			b3Quat predictedOrn = b3QuatMul(dorn, orn0);\n"
+"			predictedOrn = b3QuatNormalized(predictedOrn);\n"
+"			bodies[nodeID].m_quat=predictedOrn;\n"
+"		}\n"
+"		//linear velocity		\n"
+"		bodies[nodeID].m_pos +=  bodies[nodeID].m_linVel * timeStep;\n"
+"		\n"
+"		//apply gravity\n"
+"		bodies[nodeID].m_linVel += gravityAcceleration * timeStep;\n"
+"		\n"
+"	}\n"
+"	\n"
+"}\n"
+"inline void b3IntegrateTransform( __global b3RigidBodyData_t* body, float timeStep, float angularDamping, b3Float4ConstArg gravityAcceleration)\n"
+"{\n"
+"	float BT_GPU_ANGULAR_MOTION_THRESHOLD = (0.25f * 3.14159254f);\n"
+"	\n"
+"	if( (body->m_invMass != 0.f))\n"
+"	{\n"
+"		//angular velocity\n"
+"		{\n"
+"			b3Float4 axis;\n"
+"			//add some hardcoded angular damping\n"
+"			body->m_angVel.x *= angularDamping;\n"
+"			body->m_angVel.y *= angularDamping;\n"
+"			body->m_angVel.z *= angularDamping;\n"
+"			\n"
+"			b3Float4 angvel = body->m_angVel;\n"
+"			float fAngle = b3Sqrt(b3Dot3F4(angvel, angvel));\n"
+"			//limit the angular motion\n"
+"			if(fAngle*timeStep > BT_GPU_ANGULAR_MOTION_THRESHOLD)\n"
+"			{\n"
+"				fAngle = BT_GPU_ANGULAR_MOTION_THRESHOLD / timeStep;\n"
+"			}\n"
+"			if(fAngle < 0.001f)\n"
+"			{\n"
+"				// use Taylor's expansions of sync function\n"
+"				axis = angvel * (0.5f*timeStep-(timeStep*timeStep*timeStep)*0.020833333333f * fAngle * fAngle);\n"
+"			}\n"
+"			else\n"
+"			{\n"
+"				// sync(fAngle) = sin(c*fAngle)/t\n"
+"				axis = angvel * ( b3Sin(0.5f * fAngle * timeStep) / fAngle);\n"
+"			}\n"
+"			b3Quat dorn;\n"
+"			dorn.x = axis.x;\n"
+"			dorn.y = axis.y;\n"
+"			dorn.z = axis.z;\n"
+"			dorn.w = b3Cos(fAngle * timeStep * 0.5f);\n"
+"			b3Quat orn0 = body->m_quat;\n"
+"			b3Quat predictedOrn = b3QuatMul(dorn, orn0);\n"
+"			predictedOrn = b3QuatNormalized(predictedOrn);\n"
+"			body->m_quat=predictedOrn;\n"
+"		}\n"
+"		//apply gravity\n"
+"		body->m_linVel += gravityAcceleration * timeStep;\n"
+"		//linear velocity		\n"
+"		body->m_pos +=  body->m_linVel * timeStep;\n"
+"		\n"
+"	}\n"
+"	\n"
+"}\n"
+"__kernel void \n"
+"  integrateTransformsKernel( __global b3RigidBodyData_t* bodies,const int numNodes, float timeStep, float angularDamping, float4 gravityAcceleration)\n"
+"{\n"
+"	int nodeID = get_global_id(0);\n"
+"	\n"
+"	if( nodeID < numNodes)\n"
+"	{\n"
+"		integrateSingleTransform(bodies,nodeID, timeStep, angularDamping,gravityAcceleration);\n"
+"	}\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/jointSolver.cl b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/jointSolver.cl
new file mode 100644
index 0000000000..7f5dabe274
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/jointSolver.cl
@@ -0,0 +1,877 @@
+/*
+Copyright (c) 2013 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+#define B3_CONSTRAINT_FLAG_ENABLED 1
+
+#define B3_GPU_POINT2POINT_CONSTRAINT_TYPE 3
+#define B3_GPU_FIXED_CONSTRAINT_TYPE 4
+
+#define MOTIONCLAMP 100000 //unused, for debugging/safety in case constraint solver fails
+#define B3_INFINITY 1e30f
+
+#define mymake_float4 (float4)
+
+
+__inline float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = mymake_float4(a.xyz,0.f);
+	float4 b1 = mymake_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+
+typedef float4 Quaternion;
+
+
+typedef struct
+{
+	float4 m_row[3];
+}Matrix3x3;
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b);
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b);
+
+
+
+
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b)
+{
+	float4 ans;
+	ans.x = dot3F4( a.m_row[0], b );
+	ans.y = dot3F4( a.m_row[1], b );
+	ans.z = dot3F4( a.m_row[2], b );
+	ans.w = 0.f;
+	return ans;
+}
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b)
+{
+	float4 colx = mymake_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);
+	float4 coly = mymake_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);
+	float4 colz = mymake_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);
+
+	float4 ans;
+	ans.x = dot3F4( a, colx );
+	ans.y = dot3F4( a, coly );
+	ans.z = dot3F4( a, colz );
+	return ans;
+}
+
+
+
+typedef struct
+{
+	Matrix3x3 m_invInertiaWorld;
+	Matrix3x3 m_initInvInertia;
+} BodyInertia;
+
+
+typedef struct
+{
+	Matrix3x3 m_basis;//orientation
+	float4	m_origin;//transform
+}b3Transform;
+
+typedef struct
+{
+//	b3Transform		m_worldTransformUnused;
+	float4		m_deltaLinearVelocity;
+	float4		m_deltaAngularVelocity;
+	float4		m_angularFactor;
+	float4		m_linearFactor;
+	float4		m_invMass;
+	float4		m_pushVelocity;
+	float4		m_turnVelocity;
+	float4		m_linearVelocity;
+	float4		m_angularVelocity;
+
+	union 
+	{
+		void*	m_originalBody;
+		int		m_originalBodyIndex;
+	};
+	int padding[3];
+
+} b3GpuSolverBody;
+
+typedef struct
+{
+	float4 m_pos;
+	Quaternion m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	unsigned int m_shapeIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} b3RigidBodyCL;
+
+typedef struct
+{
+
+	float4		m_relpos1CrossNormal;
+	float4		m_contactNormal;
+
+	float4		m_relpos2CrossNormal;
+	//float4		m_contactNormal2;//usually m_contactNormal2 == -m_contactNormal
+
+	float4		m_angularComponentA;
+	float4		m_angularComponentB;
+	
+	float	m_appliedPushImpulse;
+	float	m_appliedImpulse;
+	int	m_padding1;
+	int	m_padding2;
+	float	m_friction;
+	float	m_jacDiagABInv;
+	float		m_rhs;
+	float		m_cfm;
+	
+    float		m_lowerLimit;
+	float		m_upperLimit;
+	float		m_rhsPenetration;
+	int			m_originalConstraint;
+
+
+	int	m_overrideNumSolverIterations;
+    int			m_frictionIndex;
+	int m_solverBodyIdA;
+	int m_solverBodyIdB;
+
+} b3SolverConstraint;
+
+typedef struct 
+{
+	int m_bodyAPtrAndSignBit;
+	int m_bodyBPtrAndSignBit;
+	int m_originalConstraintIndex;
+	int m_batchId;
+} b3BatchConstraint;
+
+
+
+
+
+
+typedef struct 
+{
+	int				m_constraintType;
+	int				m_rbA;
+	int				m_rbB;
+	float			m_breakingImpulseThreshold;
+
+	float4 m_pivotInA;
+	float4 m_pivotInB;
+	Quaternion m_relTargetAB;
+
+	int	m_flags;
+	int m_padding[3];
+} b3GpuGenericConstraint;
+
+
+/*b3Transform	getWorldTransform(b3RigidBodyCL* rb)
+{
+	b3Transform newTrans;
+	newTrans.setOrigin(rb->m_pos);
+	newTrans.setRotation(rb->m_quat);
+	return newTrans;
+}*/
+
+
+
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	v = mymake_float4(v.xyz,0.f);
+	return fast_normalize(v);
+}
+
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b);
+
+__inline
+Quaternion qtNormalize(Quaternion in);
+
+__inline
+float4 qtRotate(Quaternion q, float4 vec);
+
+__inline
+Quaternion qtInvert(Quaternion q);
+
+
+
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross3( a, b );
+	ans += a.w*b+b.w*a;
+//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+Quaternion qtNormalize(Quaternion in)
+{
+	return fastNormalize4(in);
+//	in /= length( in );
+//	return in;
+}
+__inline
+float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(qtMul(q,vcpy),qInv);
+	return out;
+}
+
+__inline
+Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+
+__inline void internalApplyImpulse(__global b3GpuSolverBody* body,  float4 linearComponent, float4 angularComponent,float impulseMagnitude)
+{
+	body->m_deltaLinearVelocity += linearComponent*impulseMagnitude*body->m_linearFactor;
+	body->m_deltaAngularVelocity += angularComponent*(impulseMagnitude*body->m_angularFactor);
+}
+
+
+void resolveSingleConstraintRowGeneric(__global b3GpuSolverBody* body1, __global b3GpuSolverBody* body2, __global b3SolverConstraint* c)
+{
+	float deltaImpulse = c->m_rhs-c->m_appliedImpulse*c->m_cfm;
+	float deltaVel1Dotn	=	dot3F4(c->m_contactNormal,body1->m_deltaLinearVelocity) 	+ dot3F4(c->m_relpos1CrossNormal,body1->m_deltaAngularVelocity);
+	float deltaVel2Dotn	=	-dot3F4(c->m_contactNormal,body2->m_deltaLinearVelocity) + dot3F4(c->m_relpos2CrossNormal,body2->m_deltaAngularVelocity);
+
+	deltaImpulse	-=	deltaVel1Dotn*c->m_jacDiagABInv;
+	deltaImpulse	-=	deltaVel2Dotn*c->m_jacDiagABInv;
+
+	float sum = c->m_appliedImpulse + deltaImpulse;
+	if (sum < c->m_lowerLimit)
+	{
+		deltaImpulse = c->m_lowerLimit-c->m_appliedImpulse;
+		c->m_appliedImpulse = c->m_lowerLimit;
+	}
+	else if (sum > c->m_upperLimit) 
+	{
+		deltaImpulse = c->m_upperLimit-c->m_appliedImpulse;
+		c->m_appliedImpulse = c->m_upperLimit;
+	}
+	else
+	{
+		c->m_appliedImpulse = sum;
+	}
+
+	internalApplyImpulse(body1,c->m_contactNormal*body1->m_invMass,c->m_angularComponentA,deltaImpulse);
+	internalApplyImpulse(body2,-c->m_contactNormal*body2->m_invMass,c->m_angularComponentB,deltaImpulse);
+
+}
+
+__kernel void solveJointConstraintRows(__global b3GpuSolverBody* solverBodies,
+					  __global b3BatchConstraint* batchConstraints,
+					  	__global b3SolverConstraint* rows,
+						__global unsigned int* numConstraintRowsInfo1, 
+						__global unsigned int* rowOffsets,
+						__global b3GpuGenericConstraint* constraints,
+						int batchOffset,
+						int numConstraintsInBatch
+                      )
+{
+	int b = get_global_id(0);
+	if (b>=numConstraintsInBatch)
+		return;
+
+	__global b3BatchConstraint* c = &batchConstraints[b+batchOffset];
+	int originalConstraintIndex = c->m_originalConstraintIndex;
+	if (constraints[originalConstraintIndex].m_flags&B3_CONSTRAINT_FLAG_ENABLED)
+	{
+		int numConstraintRows = numConstraintRowsInfo1[originalConstraintIndex];
+		int rowOffset = rowOffsets[originalConstraintIndex];
+		for (int jj=0;jj<numConstraintRows;jj++)
+		{
+			__global b3SolverConstraint* constraint = &rows[rowOffset+jj];
+			resolveSingleConstraintRowGeneric(&solverBodies[constraint->m_solverBodyIdA],&solverBodies[constraint->m_solverBodyIdB],constraint);
+		}
+	}
+};
+
+__kernel void initSolverBodies(__global b3GpuSolverBody* solverBodies,__global b3RigidBodyCL* bodiesCL, int numBodies)
+{
+	int i = get_global_id(0);
+	if (i>=numBodies)
+		return;
+
+	__global b3GpuSolverBody* solverBody = &solverBodies[i];
+	__global b3RigidBodyCL* bodyCL = &bodiesCL[i];
+
+	solverBody->m_deltaLinearVelocity = (float4)(0.f,0.f,0.f,0.f);
+	solverBody->m_deltaAngularVelocity  = (float4)(0.f,0.f,0.f,0.f);
+	solverBody->m_pushVelocity = (float4)(0.f,0.f,0.f,0.f);
+	solverBody->m_pushVelocity = (float4)(0.f,0.f,0.f,0.f);
+	solverBody->m_invMass = (float4)(bodyCL->m_invMass,bodyCL->m_invMass,bodyCL->m_invMass,0.f);
+	solverBody->m_originalBodyIndex = i;
+	solverBody->m_angularFactor = (float4)(1,1,1,0);
+	solverBody->m_linearFactor = (float4) (1,1,1,0);
+	solverBody->m_linearVelocity = bodyCL->m_linVel;
+	solverBody->m_angularVelocity = bodyCL->m_angVel;
+}
+
+__kernel void breakViolatedConstraintsKernel(__global b3GpuGenericConstraint* constraints, __global unsigned int* numConstraintRows, __global unsigned int* rowOffsets, __global b3SolverConstraint* rows, int numConstraints)
+{
+	int cid = get_global_id(0);
+	if (cid>=numConstraints)
+		return;
+	int numRows = numConstraintRows[cid];
+	if (numRows)
+	{
+		for (int i=0;i<numRows;i++)
+		{
+			int rowIndex = rowOffsets[cid]+i;
+			float breakingThreshold = constraints[cid].m_breakingImpulseThreshold;
+			if (fabs(rows[rowIndex].m_appliedImpulse) >= breakingThreshold)
+			{
+				constraints[cid].m_flags =0;//&= ~B3_CONSTRAINT_FLAG_ENABLED;
+			}
+		}
+	}
+}
+
+
+
+__kernel void getInfo1Kernel(__global unsigned int* infos, __global b3GpuGenericConstraint* constraints, int numConstraints)
+{
+	int i = get_global_id(0);
+	if (i>=numConstraints)
+		return;
+
+	__global b3GpuGenericConstraint* constraint = &constraints[i];
+
+	switch (constraint->m_constraintType)
+	{
+		case B3_GPU_POINT2POINT_CONSTRAINT_TYPE:
+		{
+			infos[i] = 3;
+			break;
+		}
+		case B3_GPU_FIXED_CONSTRAINT_TYPE:
+		{
+			infos[i] = 6;
+			break;
+		}
+		default:
+		{
+		}
+	}
+}
+
+__kernel void initBatchConstraintsKernel(__global unsigned int* numConstraintRows, __global unsigned int* rowOffsets, 
+										__global b3BatchConstraint* batchConstraints, 
+										__global b3GpuGenericConstraint* constraints,
+										__global b3RigidBodyCL* bodies,
+										int numConstraints)
+{
+	int i = get_global_id(0);
+	if (i>=numConstraints)
+		return;
+
+	int rbA = constraints[i].m_rbA;
+	int rbB = constraints[i].m_rbB;
+
+	batchConstraints[i].m_bodyAPtrAndSignBit = bodies[rbA].m_invMass != 0.f ? rbA : -rbA;
+	batchConstraints[i].m_bodyBPtrAndSignBit = bodies[rbB].m_invMass != 0.f ? rbB : -rbB;
+	batchConstraints[i].m_batchId = -1;
+	batchConstraints[i].m_originalConstraintIndex = i;
+
+}
+
+
+
+
+typedef struct
+{
+	// integrator parameters: frames per second (1/stepsize), default error
+	// reduction parameter (0..1).
+	float fps,erp;
+
+	// for the first and second body, pointers to two (linear and angular)
+	// n*3 jacobian sub matrices, stored by rows. these matrices will have
+	// been initialized to 0 on entry. if the second body is zero then the
+	// J2xx pointers may be 0.
+	union 
+	{
+		__global float4* m_J1linearAxisFloat4;
+		__global float* m_J1linearAxis;
+	};
+	union
+	{
+		__global float4* m_J1angularAxisFloat4;
+		__global float* m_J1angularAxis;
+
+	};
+	union
+	{
+	__global float4* m_J2linearAxisFloat4;
+	__global float* m_J2linearAxis;
+	};
+	union
+	{
+		__global float4* m_J2angularAxisFloat4;
+		__global float* m_J2angularAxis;
+	};
+	// elements to jump from one row to the next in J's
+	int rowskip;
+
+	// right hand sides of the equation J*v = c + cfm * lambda. cfm is the
+	// "constraint force mixing" vector. c is set to zero on entry, cfm is
+	// set to a constant value (typically very small or zero) value on entry.
+	__global float* m_constraintError;
+	__global float* cfm;
+
+	// lo and hi limits for variables (set to -/+ infinity on entry).
+	__global float* m_lowerLimit;
+	__global float* m_upperLimit;
+
+	// findex vector for variables. see the LCP solver interface for a
+	// description of what this does. this is set to -1 on entry.
+	// note that the returned indexes are relative to the first index of
+	// the constraint.
+	__global int *findex;
+	// number of solver iterations
+	int m_numIterations;
+
+	//damping of the velocity
+	float	m_damping;
+} b3GpuConstraintInfo2;
+
+
+void	getSkewSymmetricMatrix(float4 vecIn, __global float4* v0,__global float4* v1,__global float4* v2)
+{
+	*v0 = (float4)(0.		,-vecIn.z		,vecIn.y,0.f);
+	*v1 = (float4)(vecIn.z	,0.			,-vecIn.x,0.f);
+	*v2 = (float4)(-vecIn.y	,vecIn.x	,0.f,0.f);
+}
+
+
+void getInfo2Point2Point(__global b3GpuGenericConstraint* constraint,b3GpuConstraintInfo2* info,__global b3RigidBodyCL* bodies)
+{
+	float4 posA = bodies[constraint->m_rbA].m_pos;
+	Quaternion rotA = bodies[constraint->m_rbA].m_quat;
+
+	float4 posB = bodies[constraint->m_rbB].m_pos;
+	Quaternion rotB = bodies[constraint->m_rbB].m_quat;
+
+
+
+		// anchor points in global coordinates with respect to body PORs.
+   
+    // set jacobian
+    info->m_J1linearAxis[0] = 1;
+	info->m_J1linearAxis[info->rowskip+1] = 1;
+	info->m_J1linearAxis[2*info->rowskip+2] = 1;
+
+	float4 a1 = qtRotate(rotA,constraint->m_pivotInA);
+
+	{
+		__global float4* angular0 = (__global float4*)(info->m_J1angularAxis);
+		__global float4* angular1 = (__global float4*)(info->m_J1angularAxis+info->rowskip);
+		__global float4* angular2 = (__global float4*)(info->m_J1angularAxis+2*info->rowskip);
+		float4 a1neg = -a1;
+		getSkewSymmetricMatrix(a1neg,angular0,angular1,angular2);
+	}
+	if (info->m_J2linearAxis)
+	{
+		info->m_J2linearAxis[0] = -1;
+		info->m_J2linearAxis[info->rowskip+1] = -1;
+		info->m_J2linearAxis[2*info->rowskip+2] = -1;
+	}
+	
+	float4 a2 = qtRotate(rotB,constraint->m_pivotInB);
+   
+	{
+	//	float4 a2n = -a2;
+		__global float4* angular0 = (__global float4*)(info->m_J2angularAxis);
+		__global float4* angular1 = (__global float4*)(info->m_J2angularAxis+info->rowskip);
+		__global float4* angular2 = (__global float4*)(info->m_J2angularAxis+2*info->rowskip);
+		getSkewSymmetricMatrix(a2,angular0,angular1,angular2);
+	}
+    
+    // set right hand side
+//	float currERP = (m_flags & B3_P2P_FLAGS_ERP) ? m_erp : info->erp;
+	float currERP = info->erp;
+
+	float k = info->fps * currERP;
+    int j;
+	float4 result = a2 + posB - a1 - posA;
+	float* resultPtr = &result;
+
+	for (j=0; j<3; j++)
+    {
+        info->m_constraintError[j*info->rowskip] = k * (resultPtr[j]);
+    }
+}
+
+Quaternion nearest( Quaternion first, Quaternion qd)
+{
+	Quaternion diff,sum;
+	diff = first- qd;
+	sum = first + qd;
+	
+	if( dot(diff,diff) < dot(sum,sum) )
+		return qd;
+	return (-qd);
+}
+
+float b3Acos(float x) 
+{ 
+	if (x<-1)	
+		x=-1; 
+	if (x>1)	
+		x=1;
+	return acos(x); 
+}
+
+float getAngle(Quaternion orn)
+{
+	if (orn.w>=1.f)
+		orn.w=1.f;
+	float s = 2.f * b3Acos(orn.w);
+	return s;
+}
+
+void calculateDiffAxisAngleQuaternion( Quaternion orn0,Quaternion orn1a,float4* axis,float* angle)
+{
+	Quaternion orn1 = nearest(orn0,orn1a);
+	
+	Quaternion dorn = qtMul(orn1,qtInvert(orn0));
+	*angle = getAngle(dorn);
+	*axis = (float4)(dorn.x,dorn.y,dorn.z,0.f);
+	
+	//check for axis length
+	float len = dot3F4(*axis,*axis);
+	if (len < FLT_EPSILON*FLT_EPSILON)
+		*axis = (float4)(1,0,0,0);
+	else
+		*axis /= sqrt(len);
+}
+
+
+
+void getInfo2FixedOrientation(__global b3GpuGenericConstraint* constraint,b3GpuConstraintInfo2* info,__global b3RigidBodyCL* bodies, int start_row)
+{
+	Quaternion worldOrnA = bodies[constraint->m_rbA].m_quat;
+	Quaternion worldOrnB = bodies[constraint->m_rbB].m_quat;
+
+	int s = info->rowskip;
+	int start_index = start_row * s;
+
+	// 3 rows to make body rotations equal
+	info->m_J1angularAxis[start_index] = 1;
+	info->m_J1angularAxis[start_index + s + 1] = 1;
+	info->m_J1angularAxis[start_index + s*2+2] = 1;
+	if ( info->m_J2angularAxis)
+	{
+		info->m_J2angularAxis[start_index] = -1;
+		info->m_J2angularAxis[start_index + s+1] = -1;
+		info->m_J2angularAxis[start_index + s*2+2] = -1;
+	}
+	
+	float currERP = info->erp;
+	float k = info->fps * currERP;
+	float4 diff;
+	float angle;
+	float4 qrelCur = qtMul(worldOrnA,qtInvert(worldOrnB));
+	
+	calculateDiffAxisAngleQuaternion(constraint->m_relTargetAB,qrelCur,&diff,&angle);
+	diff*=-angle;
+		
+	float* resultPtr = &diff;
+	
+	for (int j=0; j<3; j++)
+    {
+        info->m_constraintError[(3+j)*info->rowskip] = k * resultPtr[j];
+    }
+	
+
+}
+
+
+__kernel void writeBackVelocitiesKernel(__global b3RigidBodyCL* bodies,__global b3GpuSolverBody* solverBodies,int numBodies)
+{
+	int i = get_global_id(0);
+	if (i>=numBodies)
+		return;
+
+	if (bodies[i].m_invMass)
+	{
+//		if (length(solverBodies[i].m_deltaLinearVelocity)<MOTIONCLAMP)
+		{
+			bodies[i].m_linVel += solverBodies[i].m_deltaLinearVelocity;
+		}
+//		if (length(solverBodies[i].m_deltaAngularVelocity)<MOTIONCLAMP)
+		{
+			bodies[i].m_angVel += solverBodies[i].m_deltaAngularVelocity;
+		} 
+	}
+}
+
+
+__kernel void getInfo2Kernel(__global b3SolverConstraint* solverConstraintRows, 
+							__global unsigned int* infos, 
+							__global unsigned int* constraintRowOffsets, 
+							__global b3GpuGenericConstraint* constraints, 
+							__global b3BatchConstraint* batchConstraints, 
+							__global b3RigidBodyCL* bodies,
+							__global BodyInertia* inertias,
+							__global b3GpuSolverBody* solverBodies,
+							float timeStep,
+							float globalErp,
+							float globalCfm,
+							float globalDamping,
+							int globalNumIterations,
+							int numConstraints)
+{
+
+	int i = get_global_id(0);
+	if (i>=numConstraints)
+		return;
+		
+	//for now, always initialize the batch info
+	int info1 = infos[i];
+			
+	__global b3SolverConstraint* currentConstraintRow = &solverConstraintRows[constraintRowOffsets[i]];
+	__global b3GpuGenericConstraint* constraint = &constraints[i];
+
+	__global b3RigidBodyCL* rbA = &bodies[ constraint->m_rbA];
+	__global b3RigidBodyCL* rbB = &bodies[ constraint->m_rbB];
+
+	int solverBodyIdA = constraint->m_rbA;
+	int solverBodyIdB = constraint->m_rbB;
+
+	__global b3GpuSolverBody* bodyAPtr = &solverBodies[solverBodyIdA];
+	__global b3GpuSolverBody* bodyBPtr = &solverBodies[solverBodyIdB];
+
+
+	if (rbA->m_invMass)
+	{
+		batchConstraints[i].m_bodyAPtrAndSignBit = solverBodyIdA;
+	} else
+	{
+//			if (!solverBodyIdA)
+//				m_staticIdx = 0;
+		batchConstraints[i].m_bodyAPtrAndSignBit = -solverBodyIdA;
+	}
+
+	if (rbB->m_invMass)
+	{
+		batchConstraints[i].m_bodyBPtrAndSignBit = solverBodyIdB;
+	} else
+	{
+//			if (!solverBodyIdB)
+//				m_staticIdx = 0;
+		batchConstraints[i].m_bodyBPtrAndSignBit = -solverBodyIdB;
+	}
+
+	if (info1)
+	{
+		int overrideNumSolverIterations = 0;//constraint->getOverrideNumSolverIterations() > 0 ? constraint->getOverrideNumSolverIterations() : infoGlobal.m_numIterations;
+//		if (overrideNumSolverIterations>m_maxOverrideNumSolverIterations)
+	//		m_maxOverrideNumSolverIterations = overrideNumSolverIterations;
+
+
+		int j;
+		for ( j=0;j<info1;j++)
+		{
+//			memset(&currentConstraintRow[j],0,sizeof(b3SolverConstraint));
+			currentConstraintRow[j].m_angularComponentA = (float4)(0,0,0,0);
+			currentConstraintRow[j].m_angularComponentB = (float4)(0,0,0,0);
+			currentConstraintRow[j].m_appliedImpulse = 0.f;
+			currentConstraintRow[j].m_appliedPushImpulse = 0.f;
+			currentConstraintRow[j].m_cfm = 0.f;
+			currentConstraintRow[j].m_contactNormal = (float4)(0,0,0,0);
+			currentConstraintRow[j].m_friction = 0.f;
+			currentConstraintRow[j].m_frictionIndex = 0;
+			currentConstraintRow[j].m_jacDiagABInv = 0.f;
+			currentConstraintRow[j].m_lowerLimit = 0.f;
+			currentConstraintRow[j].m_upperLimit = 0.f;
+
+			currentConstraintRow[j].m_originalConstraint = i;
+			currentConstraintRow[j].m_overrideNumSolverIterations = 0;
+			currentConstraintRow[j].m_relpos1CrossNormal = (float4)(0,0,0,0);
+			currentConstraintRow[j].m_relpos2CrossNormal = (float4)(0,0,0,0);
+			currentConstraintRow[j].m_rhs = 0.f;
+			currentConstraintRow[j].m_rhsPenetration = 0.f;
+			currentConstraintRow[j].m_solverBodyIdA = 0;
+			currentConstraintRow[j].m_solverBodyIdB = 0;
+							
+			currentConstraintRow[j].m_lowerLimit = -B3_INFINITY;
+			currentConstraintRow[j].m_upperLimit = B3_INFINITY;
+			currentConstraintRow[j].m_appliedImpulse = 0.f;
+			currentConstraintRow[j].m_appliedPushImpulse = 0.f;
+			currentConstraintRow[j].m_solverBodyIdA = solverBodyIdA;
+			currentConstraintRow[j].m_solverBodyIdB = solverBodyIdB;
+			currentConstraintRow[j].m_overrideNumSolverIterations = overrideNumSolverIterations;		
+		}
+
+		bodyAPtr->m_deltaLinearVelocity = (float4)(0,0,0,0);
+		bodyAPtr->m_deltaAngularVelocity = (float4)(0,0,0,0);
+		bodyAPtr->m_pushVelocity = (float4)(0,0,0,0);
+		bodyAPtr->m_turnVelocity = (float4)(0,0,0,0);
+		bodyBPtr->m_deltaLinearVelocity = (float4)(0,0,0,0);
+		bodyBPtr->m_deltaAngularVelocity = (float4)(0,0,0,0);
+		bodyBPtr->m_pushVelocity = (float4)(0,0,0,0);
+		bodyBPtr->m_turnVelocity  = (float4)(0,0,0,0);
+
+		int rowskip = sizeof(b3SolverConstraint)/sizeof(float);//check this
+
+		
+
+
+		b3GpuConstraintInfo2 info2;
+		info2.fps = 1.f/timeStep;
+		info2.erp = globalErp;
+		info2.m_J1linearAxisFloat4 = &currentConstraintRow->m_contactNormal;
+		info2.m_J1angularAxisFloat4 = &currentConstraintRow->m_relpos1CrossNormal;
+		info2.m_J2linearAxisFloat4 = 0;
+		info2.m_J2angularAxisFloat4 = &currentConstraintRow->m_relpos2CrossNormal;
+		info2.rowskip = sizeof(b3SolverConstraint)/sizeof(float);//check this
+
+		///the size of b3SolverConstraint needs be a multiple of float
+//		b3Assert(info2.rowskip*sizeof(float)== sizeof(b3SolverConstraint));
+		info2.m_constraintError = &currentConstraintRow->m_rhs;
+		currentConstraintRow->m_cfm = globalCfm;
+		info2.m_damping = globalDamping;
+		info2.cfm = &currentConstraintRow->m_cfm;
+		info2.m_lowerLimit = &currentConstraintRow->m_lowerLimit;
+		info2.m_upperLimit = &currentConstraintRow->m_upperLimit;
+		info2.m_numIterations = globalNumIterations;
+
+		switch (constraint->m_constraintType)
+		{
+			case B3_GPU_POINT2POINT_CONSTRAINT_TYPE:
+			{
+				getInfo2Point2Point(constraint,&info2,bodies);
+				break;
+			}
+			case B3_GPU_FIXED_CONSTRAINT_TYPE:
+			{
+				getInfo2Point2Point(constraint,&info2,bodies);
+
+				getInfo2FixedOrientation(constraint,&info2,bodies,3);
+
+				break;
+			}
+
+			default:
+			{
+			}
+		}
+
+		///finalize the constraint setup
+		for ( j=0;j<info1;j++)
+		{
+			__global b3SolverConstraint* solverConstraint = &currentConstraintRow[j];
+
+			if (solverConstraint->m_upperLimit>=constraint->m_breakingImpulseThreshold)
+			{
+				solverConstraint->m_upperLimit = constraint->m_breakingImpulseThreshold;
+			}
+
+			if (solverConstraint->m_lowerLimit<=-constraint->m_breakingImpulseThreshold)
+			{
+				solverConstraint->m_lowerLimit = -constraint->m_breakingImpulseThreshold;
+			}
+
+//						solverConstraint->m_originalContactPoint = constraint;
+							
+			Matrix3x3 invInertiaWorldA= inertias[constraint->m_rbA].m_invInertiaWorld;
+			{
+
+				//float4 angularFactorA(1,1,1);
+				float4 ftorqueAxis1 = solverConstraint->m_relpos1CrossNormal;
+				solverConstraint->m_angularComponentA = mtMul1(invInertiaWorldA,ftorqueAxis1);//*angularFactorA;
+			}
+						
+			Matrix3x3 invInertiaWorldB= inertias[constraint->m_rbB].m_invInertiaWorld;
+			{
+
+				float4 ftorqueAxis2 = solverConstraint->m_relpos2CrossNormal;
+				solverConstraint->m_angularComponentB = mtMul1(invInertiaWorldB,ftorqueAxis2);//*constraint->m_rbB.getAngularFactor();
+			}
+
+			{
+				//it is ok to use solverConstraint->m_contactNormal instead of -solverConstraint->m_contactNormal
+				//because it gets multiplied iMJlB
+				float4 iMJlA = solverConstraint->m_contactNormal*rbA->m_invMass;
+				float4 iMJaA = mtMul3(solverConstraint->m_relpos1CrossNormal,invInertiaWorldA);
+				float4 iMJlB = solverConstraint->m_contactNormal*rbB->m_invMass;//sign of normal?
+				float4 iMJaB = mtMul3(solverConstraint->m_relpos2CrossNormal,invInertiaWorldB);
+
+				float sum = dot3F4(iMJlA,solverConstraint->m_contactNormal);
+				sum += dot3F4(iMJaA,solverConstraint->m_relpos1CrossNormal);
+				sum += dot3F4(iMJlB,solverConstraint->m_contactNormal);
+				sum += dot3F4(iMJaB,solverConstraint->m_relpos2CrossNormal);
+				float fsum = fabs(sum);
+				if (fsum>FLT_EPSILON)
+				{
+					solverConstraint->m_jacDiagABInv = 1.f/sum;
+				} else
+				{
+					solverConstraint->m_jacDiagABInv = 0.f;
+				}
+			}
+
+
+			///fix rhs
+			///todo: add force/torque accelerators
+			{
+				float rel_vel;
+				float vel1Dotn = dot3F4(solverConstraint->m_contactNormal,rbA->m_linVel) + dot3F4(solverConstraint->m_relpos1CrossNormal,rbA->m_angVel);
+				float vel2Dotn = -dot3F4(solverConstraint->m_contactNormal,rbB->m_linVel) + dot3F4(solverConstraint->m_relpos2CrossNormal,rbB->m_angVel);
+
+				rel_vel = vel1Dotn+vel2Dotn;
+
+				float restitution = 0.f;
+				float positionalError = solverConstraint->m_rhs;//already filled in by getConstraintInfo2
+				float	velocityError = restitution - rel_vel * info2.m_damping;
+				float	penetrationImpulse = positionalError*solverConstraint->m_jacDiagABInv;
+				float	velocityImpulse = velocityError *solverConstraint->m_jacDiagABInv;
+				solverConstraint->m_rhs = penetrationImpulse+velocityImpulse;
+				solverConstraint->m_appliedImpulse = 0.f;
+
+			}
+		}
+	}
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/jointSolver.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/jointSolver.h
new file mode 100644
index 0000000000..d48ecf6ea6
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/jointSolver.h
@@ -0,0 +1,721 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* solveConstraintRowsCL= \
+"/*\n"
+"Copyright (c) 2013 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Erwin Coumans\n"
+"#define B3_CONSTRAINT_FLAG_ENABLED 1\n"
+"#define B3_GPU_POINT2POINT_CONSTRAINT_TYPE 3\n"
+"#define B3_GPU_FIXED_CONSTRAINT_TYPE 4\n"
+"#define MOTIONCLAMP 100000 //unused, for debugging/safety in case constraint solver fails\n"
+"#define B3_INFINITY 1e30f\n"
+"#define mymake_float4 (float4)\n"
+"__inline float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = mymake_float4(a.xyz,0.f);\n"
+"	float4 b1 = mymake_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"typedef float4 Quaternion;\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_row[3];\n"
+"}Matrix3x3;\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b);\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b);\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b)\n"
+"{\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a.m_row[0], b );\n"
+"	ans.y = dot3F4( a.m_row[1], b );\n"
+"	ans.z = dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b)\n"
+"{\n"
+"	float4 colx = mymake_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	float4 coly = mymake_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	float4 colz = mymake_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a, colx );\n"
+"	ans.y = dot3F4( a, coly );\n"
+"	ans.z = dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"typedef struct\n"
+"{\n"
+"	Matrix3x3 m_invInertiaWorld;\n"
+"	Matrix3x3 m_initInvInertia;\n"
+"} BodyInertia;\n"
+"typedef struct\n"
+"{\n"
+"	Matrix3x3 m_basis;//orientation\n"
+"	float4	m_origin;//transform\n"
+"}b3Transform;\n"
+"typedef struct\n"
+"{\n"
+"//	b3Transform		m_worldTransformUnused;\n"
+"	float4		m_deltaLinearVelocity;\n"
+"	float4		m_deltaAngularVelocity;\n"
+"	float4		m_angularFactor;\n"
+"	float4		m_linearFactor;\n"
+"	float4		m_invMass;\n"
+"	float4		m_pushVelocity;\n"
+"	float4		m_turnVelocity;\n"
+"	float4		m_linearVelocity;\n"
+"	float4		m_angularVelocity;\n"
+"	union \n"
+"	{\n"
+"		void*	m_originalBody;\n"
+"		int		m_originalBodyIndex;\n"
+"	};\n"
+"	int padding[3];\n"
+"} b3GpuSolverBody;\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	Quaternion m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"	unsigned int m_shapeIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} b3RigidBodyCL;\n"
+"typedef struct\n"
+"{\n"
+"	float4		m_relpos1CrossNormal;\n"
+"	float4		m_contactNormal;\n"
+"	float4		m_relpos2CrossNormal;\n"
+"	//float4		m_contactNormal2;//usually m_contactNormal2 == -m_contactNormal\n"
+"	float4		m_angularComponentA;\n"
+"	float4		m_angularComponentB;\n"
+"	\n"
+"	float	m_appliedPushImpulse;\n"
+"	float	m_appliedImpulse;\n"
+"	int	m_padding1;\n"
+"	int	m_padding2;\n"
+"	float	m_friction;\n"
+"	float	m_jacDiagABInv;\n"
+"	float		m_rhs;\n"
+"	float		m_cfm;\n"
+"	\n"
+"    float		m_lowerLimit;\n"
+"	float		m_upperLimit;\n"
+"	float		m_rhsPenetration;\n"
+"	int			m_originalConstraint;\n"
+"	int	m_overrideNumSolverIterations;\n"
+"    int			m_frictionIndex;\n"
+"	int m_solverBodyIdA;\n"
+"	int m_solverBodyIdB;\n"
+"} b3SolverConstraint;\n"
+"typedef struct \n"
+"{\n"
+"	int m_bodyAPtrAndSignBit;\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"	int m_originalConstraintIndex;\n"
+"	int m_batchId;\n"
+"} b3BatchConstraint;\n"
+"typedef struct \n"
+"{\n"
+"	int				m_constraintType;\n"
+"	int				m_rbA;\n"
+"	int				m_rbB;\n"
+"	float			m_breakingImpulseThreshold;\n"
+"	float4 m_pivotInA;\n"
+"	float4 m_pivotInB;\n"
+"	Quaternion m_relTargetAB;\n"
+"	int	m_flags;\n"
+"	int m_padding[3];\n"
+"} b3GpuGenericConstraint;\n"
+"/*b3Transform	getWorldTransform(b3RigidBodyCL* rb)\n"
+"{\n"
+"	b3Transform newTrans;\n"
+"	newTrans.setOrigin(rb->m_pos);\n"
+"	newTrans.setRotation(rb->m_quat);\n"
+"	return newTrans;\n"
+"}*/\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	v = mymake_float4(v.xyz,0.f);\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b);\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in);\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec);\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q);\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in)\n"
+"{\n"
+"	return fastNormalize4(in);\n"
+"//	in /= length( in );\n"
+"//	return in;\n"
+"}\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(qtMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"__inline void internalApplyImpulse(__global b3GpuSolverBody* body,  float4 linearComponent, float4 angularComponent,float impulseMagnitude)\n"
+"{\n"
+"	body->m_deltaLinearVelocity += linearComponent*impulseMagnitude*body->m_linearFactor;\n"
+"	body->m_deltaAngularVelocity += angularComponent*(impulseMagnitude*body->m_angularFactor);\n"
+"}\n"
+"void resolveSingleConstraintRowGeneric(__global b3GpuSolverBody* body1, __global b3GpuSolverBody* body2, __global b3SolverConstraint* c)\n"
+"{\n"
+"	float deltaImpulse = c->m_rhs-c->m_appliedImpulse*c->m_cfm;\n"
+"	float deltaVel1Dotn	=	dot3F4(c->m_contactNormal,body1->m_deltaLinearVelocity) 	+ dot3F4(c->m_relpos1CrossNormal,body1->m_deltaAngularVelocity);\n"
+"	float deltaVel2Dotn	=	-dot3F4(c->m_contactNormal,body2->m_deltaLinearVelocity) + dot3F4(c->m_relpos2CrossNormal,body2->m_deltaAngularVelocity);\n"
+"	deltaImpulse	-=	deltaVel1Dotn*c->m_jacDiagABInv;\n"
+"	deltaImpulse	-=	deltaVel2Dotn*c->m_jacDiagABInv;\n"
+"	float sum = c->m_appliedImpulse + deltaImpulse;\n"
+"	if (sum < c->m_lowerLimit)\n"
+"	{\n"
+"		deltaImpulse = c->m_lowerLimit-c->m_appliedImpulse;\n"
+"		c->m_appliedImpulse = c->m_lowerLimit;\n"
+"	}\n"
+"	else if (sum > c->m_upperLimit) \n"
+"	{\n"
+"		deltaImpulse = c->m_upperLimit-c->m_appliedImpulse;\n"
+"		c->m_appliedImpulse = c->m_upperLimit;\n"
+"	}\n"
+"	else\n"
+"	{\n"
+"		c->m_appliedImpulse = sum;\n"
+"	}\n"
+"	internalApplyImpulse(body1,c->m_contactNormal*body1->m_invMass,c->m_angularComponentA,deltaImpulse);\n"
+"	internalApplyImpulse(body2,-c->m_contactNormal*body2->m_invMass,c->m_angularComponentB,deltaImpulse);\n"
+"}\n"
+"__kernel void solveJointConstraintRows(__global b3GpuSolverBody* solverBodies,\n"
+"					  __global b3BatchConstraint* batchConstraints,\n"
+"					  	__global b3SolverConstraint* rows,\n"
+"						__global unsigned int* numConstraintRowsInfo1, \n"
+"						__global unsigned int* rowOffsets,\n"
+"						__global b3GpuGenericConstraint* constraints,\n"
+"						int batchOffset,\n"
+"						int numConstraintsInBatch\n"
+"                      )\n"
+"{\n"
+"	int b = get_global_id(0);\n"
+"	if (b>=numConstraintsInBatch)\n"
+"		return;\n"
+"	__global b3BatchConstraint* c = &batchConstraints[b+batchOffset];\n"
+"	int originalConstraintIndex = c->m_originalConstraintIndex;\n"
+"	if (constraints[originalConstraintIndex].m_flags&B3_CONSTRAINT_FLAG_ENABLED)\n"
+"	{\n"
+"		int numConstraintRows = numConstraintRowsInfo1[originalConstraintIndex];\n"
+"		int rowOffset = rowOffsets[originalConstraintIndex];\n"
+"		for (int jj=0;jj<numConstraintRows;jj++)\n"
+"		{\n"
+"			__global b3SolverConstraint* constraint = &rows[rowOffset+jj];\n"
+"			resolveSingleConstraintRowGeneric(&solverBodies[constraint->m_solverBodyIdA],&solverBodies[constraint->m_solverBodyIdB],constraint);\n"
+"		}\n"
+"	}\n"
+"};\n"
+"__kernel void initSolverBodies(__global b3GpuSolverBody* solverBodies,__global b3RigidBodyCL* bodiesCL, int numBodies)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numBodies)\n"
+"		return;\n"
+"	__global b3GpuSolverBody* solverBody = &solverBodies[i];\n"
+"	__global b3RigidBodyCL* bodyCL = &bodiesCL[i];\n"
+"	solverBody->m_deltaLinearVelocity = (float4)(0.f,0.f,0.f,0.f);\n"
+"	solverBody->m_deltaAngularVelocity  = (float4)(0.f,0.f,0.f,0.f);\n"
+"	solverBody->m_pushVelocity = (float4)(0.f,0.f,0.f,0.f);\n"
+"	solverBody->m_pushVelocity = (float4)(0.f,0.f,0.f,0.f);\n"
+"	solverBody->m_invMass = (float4)(bodyCL->m_invMass,bodyCL->m_invMass,bodyCL->m_invMass,0.f);\n"
+"	solverBody->m_originalBodyIndex = i;\n"
+"	solverBody->m_angularFactor = (float4)(1,1,1,0);\n"
+"	solverBody->m_linearFactor = (float4) (1,1,1,0);\n"
+"	solverBody->m_linearVelocity = bodyCL->m_linVel;\n"
+"	solverBody->m_angularVelocity = bodyCL->m_angVel;\n"
+"}\n"
+"__kernel void breakViolatedConstraintsKernel(__global b3GpuGenericConstraint* constraints, __global unsigned int* numConstraintRows, __global unsigned int* rowOffsets, __global b3SolverConstraint* rows, int numConstraints)\n"
+"{\n"
+"	int cid = get_global_id(0);\n"
+"	if (cid>=numConstraints)\n"
+"		return;\n"
+"	int numRows = numConstraintRows[cid];\n"
+"	if (numRows)\n"
+"	{\n"
+"		for (int i=0;i<numRows;i++)\n"
+"		{\n"
+"			int rowIndex = rowOffsets[cid]+i;\n"
+"			float breakingThreshold = constraints[cid].m_breakingImpulseThreshold;\n"
+"			if (fabs(rows[rowIndex].m_appliedImpulse) >= breakingThreshold)\n"
+"			{\n"
+"				constraints[cid].m_flags =0;//&= ~B3_CONSTRAINT_FLAG_ENABLED;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"}\n"
+"__kernel void getInfo1Kernel(__global unsigned int* infos, __global b3GpuGenericConstraint* constraints, int numConstraints)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numConstraints)\n"
+"		return;\n"
+"	__global b3GpuGenericConstraint* constraint = &constraints[i];\n"
+"	switch (constraint->m_constraintType)\n"
+"	{\n"
+"		case B3_GPU_POINT2POINT_CONSTRAINT_TYPE:\n"
+"		{\n"
+"			infos[i] = 3;\n"
+"			break;\n"
+"		}\n"
+"		case B3_GPU_FIXED_CONSTRAINT_TYPE:\n"
+"		{\n"
+"			infos[i] = 6;\n"
+"			break;\n"
+"		}\n"
+"		default:\n"
+"		{\n"
+"		}\n"
+"	}\n"
+"}\n"
+"__kernel void initBatchConstraintsKernel(__global unsigned int* numConstraintRows, __global unsigned int* rowOffsets, \n"
+"										__global b3BatchConstraint* batchConstraints, \n"
+"										__global b3GpuGenericConstraint* constraints,\n"
+"										__global b3RigidBodyCL* bodies,\n"
+"										int numConstraints)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numConstraints)\n"
+"		return;\n"
+"	int rbA = constraints[i].m_rbA;\n"
+"	int rbB = constraints[i].m_rbB;\n"
+"	batchConstraints[i].m_bodyAPtrAndSignBit = bodies[rbA].m_invMass != 0.f ? rbA : -rbA;\n"
+"	batchConstraints[i].m_bodyBPtrAndSignBit = bodies[rbB].m_invMass != 0.f ? rbB : -rbB;\n"
+"	batchConstraints[i].m_batchId = -1;\n"
+"	batchConstraints[i].m_originalConstraintIndex = i;\n"
+"}\n"
+"typedef struct\n"
+"{\n"
+"	// integrator parameters: frames per second (1/stepsize), default error\n"
+"	// reduction parameter (0..1).\n"
+"	float fps,erp;\n"
+"	// for the first and second body, pointers to two (linear and angular)\n"
+"	// n*3 jacobian sub matrices, stored by rows. these matrices will have\n"
+"	// been initialized to 0 on entry. if the second body is zero then the\n"
+"	// J2xx pointers may be 0.\n"
+"	union \n"
+"	{\n"
+"		__global float4* m_J1linearAxisFloat4;\n"
+"		__global float* m_J1linearAxis;\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		__global float4* m_J1angularAxisFloat4;\n"
+"		__global float* m_J1angularAxis;\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"	__global float4* m_J2linearAxisFloat4;\n"
+"	__global float* m_J2linearAxis;\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		__global float4* m_J2angularAxisFloat4;\n"
+"		__global float* m_J2angularAxis;\n"
+"	};\n"
+"	// elements to jump from one row to the next in J's\n"
+"	int rowskip;\n"
+"	// right hand sides of the equation J*v = c + cfm * lambda. cfm is the\n"
+"	// \"constraint force mixing\" vector. c is set to zero on entry, cfm is\n"
+"	// set to a constant value (typically very small or zero) value on entry.\n"
+"	__global float* m_constraintError;\n"
+"	__global float* cfm;\n"
+"	// lo and hi limits for variables (set to -/+ infinity on entry).\n"
+"	__global float* m_lowerLimit;\n"
+"	__global float* m_upperLimit;\n"
+"	// findex vector for variables. see the LCP solver interface for a\n"
+"	// description of what this does. this is set to -1 on entry.\n"
+"	// note that the returned indexes are relative to the first index of\n"
+"	// the constraint.\n"
+"	__global int *findex;\n"
+"	// number of solver iterations\n"
+"	int m_numIterations;\n"
+"	//damping of the velocity\n"
+"	float	m_damping;\n"
+"} b3GpuConstraintInfo2;\n"
+"void	getSkewSymmetricMatrix(float4 vecIn, __global float4* v0,__global float4* v1,__global float4* v2)\n"
+"{\n"
+"	*v0 = (float4)(0.		,-vecIn.z		,vecIn.y,0.f);\n"
+"	*v1 = (float4)(vecIn.z	,0.			,-vecIn.x,0.f);\n"
+"	*v2 = (float4)(-vecIn.y	,vecIn.x	,0.f,0.f);\n"
+"}\n"
+"void getInfo2Point2Point(__global b3GpuGenericConstraint* constraint,b3GpuConstraintInfo2* info,__global b3RigidBodyCL* bodies)\n"
+"{\n"
+"	float4 posA = bodies[constraint->m_rbA].m_pos;\n"
+"	Quaternion rotA = bodies[constraint->m_rbA].m_quat;\n"
+"	float4 posB = bodies[constraint->m_rbB].m_pos;\n"
+"	Quaternion rotB = bodies[constraint->m_rbB].m_quat;\n"
+"		// anchor points in global coordinates with respect to body PORs.\n"
+"   \n"
+"    // set jacobian\n"
+"    info->m_J1linearAxis[0] = 1;\n"
+"	info->m_J1linearAxis[info->rowskip+1] = 1;\n"
+"	info->m_J1linearAxis[2*info->rowskip+2] = 1;\n"
+"	float4 a1 = qtRotate(rotA,constraint->m_pivotInA);\n"
+"	{\n"
+"		__global float4* angular0 = (__global float4*)(info->m_J1angularAxis);\n"
+"		__global float4* angular1 = (__global float4*)(info->m_J1angularAxis+info->rowskip);\n"
+"		__global float4* angular2 = (__global float4*)(info->m_J1angularAxis+2*info->rowskip);\n"
+"		float4 a1neg = -a1;\n"
+"		getSkewSymmetricMatrix(a1neg,angular0,angular1,angular2);\n"
+"	}\n"
+"	if (info->m_J2linearAxis)\n"
+"	{\n"
+"		info->m_J2linearAxis[0] = -1;\n"
+"		info->m_J2linearAxis[info->rowskip+1] = -1;\n"
+"		info->m_J2linearAxis[2*info->rowskip+2] = -1;\n"
+"	}\n"
+"	\n"
+"	float4 a2 = qtRotate(rotB,constraint->m_pivotInB);\n"
+"   \n"
+"	{\n"
+"	//	float4 a2n = -a2;\n"
+"		__global float4* angular0 = (__global float4*)(info->m_J2angularAxis);\n"
+"		__global float4* angular1 = (__global float4*)(info->m_J2angularAxis+info->rowskip);\n"
+"		__global float4* angular2 = (__global float4*)(info->m_J2angularAxis+2*info->rowskip);\n"
+"		getSkewSymmetricMatrix(a2,angular0,angular1,angular2);\n"
+"	}\n"
+"    \n"
+"    // set right hand side\n"
+"//	float currERP = (m_flags & B3_P2P_FLAGS_ERP) ? m_erp : info->erp;\n"
+"	float currERP = info->erp;\n"
+"	float k = info->fps * currERP;\n"
+"    int j;\n"
+"	float4 result = a2 + posB - a1 - posA;\n"
+"	float* resultPtr = &result;\n"
+"	for (j=0; j<3; j++)\n"
+"    {\n"
+"        info->m_constraintError[j*info->rowskip] = k * (resultPtr[j]);\n"
+"    }\n"
+"}\n"
+"Quaternion nearest( Quaternion first, Quaternion qd)\n"
+"{\n"
+"	Quaternion diff,sum;\n"
+"	diff = first- qd;\n"
+"	sum = first + qd;\n"
+"	\n"
+"	if( dot(diff,diff) < dot(sum,sum) )\n"
+"		return qd;\n"
+"	return (-qd);\n"
+"}\n"
+"float b3Acos(float x) \n"
+"{ \n"
+"	if (x<-1)	\n"
+"		x=-1; \n"
+"	if (x>1)	\n"
+"		x=1;\n"
+"	return acos(x); \n"
+"}\n"
+"float getAngle(Quaternion orn)\n"
+"{\n"
+"	if (orn.w>=1.f)\n"
+"		orn.w=1.f;\n"
+"	float s = 2.f * b3Acos(orn.w);\n"
+"	return s;\n"
+"}\n"
+"void calculateDiffAxisAngleQuaternion( Quaternion orn0,Quaternion orn1a,float4* axis,float* angle)\n"
+"{\n"
+"	Quaternion orn1 = nearest(orn0,orn1a);\n"
+"	\n"
+"	Quaternion dorn = qtMul(orn1,qtInvert(orn0));\n"
+"	*angle = getAngle(dorn);\n"
+"	*axis = (float4)(dorn.x,dorn.y,dorn.z,0.f);\n"
+"	\n"
+"	//check for axis length\n"
+"	float len = dot3F4(*axis,*axis);\n"
+"	if (len < FLT_EPSILON*FLT_EPSILON)\n"
+"		*axis = (float4)(1,0,0,0);\n"
+"	else\n"
+"		*axis /= sqrt(len);\n"
+"}\n"
+"void getInfo2FixedOrientation(__global b3GpuGenericConstraint* constraint,b3GpuConstraintInfo2* info,__global b3RigidBodyCL* bodies, int start_row)\n"
+"{\n"
+"	Quaternion worldOrnA = bodies[constraint->m_rbA].m_quat;\n"
+"	Quaternion worldOrnB = bodies[constraint->m_rbB].m_quat;\n"
+"	int s = info->rowskip;\n"
+"	int start_index = start_row * s;\n"
+"	// 3 rows to make body rotations equal\n"
+"	info->m_J1angularAxis[start_index] = 1;\n"
+"	info->m_J1angularAxis[start_index + s + 1] = 1;\n"
+"	info->m_J1angularAxis[start_index + s*2+2] = 1;\n"
+"	if ( info->m_J2angularAxis)\n"
+"	{\n"
+"		info->m_J2angularAxis[start_index] = -1;\n"
+"		info->m_J2angularAxis[start_index + s+1] = -1;\n"
+"		info->m_J2angularAxis[start_index + s*2+2] = -1;\n"
+"	}\n"
+"	\n"
+"	float currERP = info->erp;\n"
+"	float k = info->fps * currERP;\n"
+"	float4 diff;\n"
+"	float angle;\n"
+"	float4 qrelCur = qtMul(worldOrnA,qtInvert(worldOrnB));\n"
+"	\n"
+"	calculateDiffAxisAngleQuaternion(constraint->m_relTargetAB,qrelCur,&diff,&angle);\n"
+"	diff*=-angle;\n"
+"		\n"
+"	float* resultPtr = &diff;\n"
+"	\n"
+"	for (int j=0; j<3; j++)\n"
+"    {\n"
+"        info->m_constraintError[(3+j)*info->rowskip] = k * resultPtr[j];\n"
+"    }\n"
+"	\n"
+"}\n"
+"__kernel void writeBackVelocitiesKernel(__global b3RigidBodyCL* bodies,__global b3GpuSolverBody* solverBodies,int numBodies)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numBodies)\n"
+"		return;\n"
+"	if (bodies[i].m_invMass)\n"
+"	{\n"
+"//		if (length(solverBodies[i].m_deltaLinearVelocity)<MOTIONCLAMP)\n"
+"		{\n"
+"			bodies[i].m_linVel += solverBodies[i].m_deltaLinearVelocity;\n"
+"		}\n"
+"//		if (length(solverBodies[i].m_deltaAngularVelocity)<MOTIONCLAMP)\n"
+"		{\n"
+"			bodies[i].m_angVel += solverBodies[i].m_deltaAngularVelocity;\n"
+"		} \n"
+"	}\n"
+"}\n"
+"__kernel void getInfo2Kernel(__global b3SolverConstraint* solverConstraintRows, \n"
+"							__global unsigned int* infos, \n"
+"							__global unsigned int* constraintRowOffsets, \n"
+"							__global b3GpuGenericConstraint* constraints, \n"
+"							__global b3BatchConstraint* batchConstraints, \n"
+"							__global b3RigidBodyCL* bodies,\n"
+"							__global BodyInertia* inertias,\n"
+"							__global b3GpuSolverBody* solverBodies,\n"
+"							float timeStep,\n"
+"							float globalErp,\n"
+"							float globalCfm,\n"
+"							float globalDamping,\n"
+"							int globalNumIterations,\n"
+"							int numConstraints)\n"
+"{\n"
+"	int i = get_global_id(0);\n"
+"	if (i>=numConstraints)\n"
+"		return;\n"
+"		\n"
+"	//for now, always initialize the batch info\n"
+"	int info1 = infos[i];\n"
+"			\n"
+"	__global b3SolverConstraint* currentConstraintRow = &solverConstraintRows[constraintRowOffsets[i]];\n"
+"	__global b3GpuGenericConstraint* constraint = &constraints[i];\n"
+"	__global b3RigidBodyCL* rbA = &bodies[ constraint->m_rbA];\n"
+"	__global b3RigidBodyCL* rbB = &bodies[ constraint->m_rbB];\n"
+"	int solverBodyIdA = constraint->m_rbA;\n"
+"	int solverBodyIdB = constraint->m_rbB;\n"
+"	__global b3GpuSolverBody* bodyAPtr = &solverBodies[solverBodyIdA];\n"
+"	__global b3GpuSolverBody* bodyBPtr = &solverBodies[solverBodyIdB];\n"
+"	if (rbA->m_invMass)\n"
+"	{\n"
+"		batchConstraints[i].m_bodyAPtrAndSignBit = solverBodyIdA;\n"
+"	} else\n"
+"	{\n"
+"//			if (!solverBodyIdA)\n"
+"//				m_staticIdx = 0;\n"
+"		batchConstraints[i].m_bodyAPtrAndSignBit = -solverBodyIdA;\n"
+"	}\n"
+"	if (rbB->m_invMass)\n"
+"	{\n"
+"		batchConstraints[i].m_bodyBPtrAndSignBit = solverBodyIdB;\n"
+"	} else\n"
+"	{\n"
+"//			if (!solverBodyIdB)\n"
+"//				m_staticIdx = 0;\n"
+"		batchConstraints[i].m_bodyBPtrAndSignBit = -solverBodyIdB;\n"
+"	}\n"
+"	if (info1)\n"
+"	{\n"
+"		int overrideNumSolverIterations = 0;//constraint->getOverrideNumSolverIterations() > 0 ? constraint->getOverrideNumSolverIterations() : infoGlobal.m_numIterations;\n"
+"//		if (overrideNumSolverIterations>m_maxOverrideNumSolverIterations)\n"
+"	//		m_maxOverrideNumSolverIterations = overrideNumSolverIterations;\n"
+"		int j;\n"
+"		for ( j=0;j<info1;j++)\n"
+"		{\n"
+"//			memset(&currentConstraintRow[j],0,sizeof(b3SolverConstraint));\n"
+"			currentConstraintRow[j].m_angularComponentA = (float4)(0,0,0,0);\n"
+"			currentConstraintRow[j].m_angularComponentB = (float4)(0,0,0,0);\n"
+"			currentConstraintRow[j].m_appliedImpulse = 0.f;\n"
+"			currentConstraintRow[j].m_appliedPushImpulse = 0.f;\n"
+"			currentConstraintRow[j].m_cfm = 0.f;\n"
+"			currentConstraintRow[j].m_contactNormal = (float4)(0,0,0,0);\n"
+"			currentConstraintRow[j].m_friction = 0.f;\n"
+"			currentConstraintRow[j].m_frictionIndex = 0;\n"
+"			currentConstraintRow[j].m_jacDiagABInv = 0.f;\n"
+"			currentConstraintRow[j].m_lowerLimit = 0.f;\n"
+"			currentConstraintRow[j].m_upperLimit = 0.f;\n"
+"			currentConstraintRow[j].m_originalConstraint = i;\n"
+"			currentConstraintRow[j].m_overrideNumSolverIterations = 0;\n"
+"			currentConstraintRow[j].m_relpos1CrossNormal = (float4)(0,0,0,0);\n"
+"			currentConstraintRow[j].m_relpos2CrossNormal = (float4)(0,0,0,0);\n"
+"			currentConstraintRow[j].m_rhs = 0.f;\n"
+"			currentConstraintRow[j].m_rhsPenetration = 0.f;\n"
+"			currentConstraintRow[j].m_solverBodyIdA = 0;\n"
+"			currentConstraintRow[j].m_solverBodyIdB = 0;\n"
+"							\n"
+"			currentConstraintRow[j].m_lowerLimit = -B3_INFINITY;\n"
+"			currentConstraintRow[j].m_upperLimit = B3_INFINITY;\n"
+"			currentConstraintRow[j].m_appliedImpulse = 0.f;\n"
+"			currentConstraintRow[j].m_appliedPushImpulse = 0.f;\n"
+"			currentConstraintRow[j].m_solverBodyIdA = solverBodyIdA;\n"
+"			currentConstraintRow[j].m_solverBodyIdB = solverBodyIdB;\n"
+"			currentConstraintRow[j].m_overrideNumSolverIterations = overrideNumSolverIterations;		\n"
+"		}\n"
+"		bodyAPtr->m_deltaLinearVelocity = (float4)(0,0,0,0);\n"
+"		bodyAPtr->m_deltaAngularVelocity = (float4)(0,0,0,0);\n"
+"		bodyAPtr->m_pushVelocity = (float4)(0,0,0,0);\n"
+"		bodyAPtr->m_turnVelocity = (float4)(0,0,0,0);\n"
+"		bodyBPtr->m_deltaLinearVelocity = (float4)(0,0,0,0);\n"
+"		bodyBPtr->m_deltaAngularVelocity = (float4)(0,0,0,0);\n"
+"		bodyBPtr->m_pushVelocity = (float4)(0,0,0,0);\n"
+"		bodyBPtr->m_turnVelocity  = (float4)(0,0,0,0);\n"
+"		int rowskip = sizeof(b3SolverConstraint)/sizeof(float);//check this\n"
+"		\n"
+"		b3GpuConstraintInfo2 info2;\n"
+"		info2.fps = 1.f/timeStep;\n"
+"		info2.erp = globalErp;\n"
+"		info2.m_J1linearAxisFloat4 = &currentConstraintRow->m_contactNormal;\n"
+"		info2.m_J1angularAxisFloat4 = &currentConstraintRow->m_relpos1CrossNormal;\n"
+"		info2.m_J2linearAxisFloat4 = 0;\n"
+"		info2.m_J2angularAxisFloat4 = &currentConstraintRow->m_relpos2CrossNormal;\n"
+"		info2.rowskip = sizeof(b3SolverConstraint)/sizeof(float);//check this\n"
+"		///the size of b3SolverConstraint needs be a multiple of float\n"
+"//		b3Assert(info2.rowskip*sizeof(float)== sizeof(b3SolverConstraint));\n"
+"		info2.m_constraintError = &currentConstraintRow->m_rhs;\n"
+"		currentConstraintRow->m_cfm = globalCfm;\n"
+"		info2.m_damping = globalDamping;\n"
+"		info2.cfm = &currentConstraintRow->m_cfm;\n"
+"		info2.m_lowerLimit = &currentConstraintRow->m_lowerLimit;\n"
+"		info2.m_upperLimit = &currentConstraintRow->m_upperLimit;\n"
+"		info2.m_numIterations = globalNumIterations;\n"
+"		switch (constraint->m_constraintType)\n"
+"		{\n"
+"			case B3_GPU_POINT2POINT_CONSTRAINT_TYPE:\n"
+"			{\n"
+"				getInfo2Point2Point(constraint,&info2,bodies);\n"
+"				break;\n"
+"			}\n"
+"			case B3_GPU_FIXED_CONSTRAINT_TYPE:\n"
+"			{\n"
+"				getInfo2Point2Point(constraint,&info2,bodies);\n"
+"				getInfo2FixedOrientation(constraint,&info2,bodies,3);\n"
+"				break;\n"
+"			}\n"
+"			default:\n"
+"			{\n"
+"			}\n"
+"		}\n"
+"		///finalize the constraint setup\n"
+"		for ( j=0;j<info1;j++)\n"
+"		{\n"
+"			__global b3SolverConstraint* solverConstraint = &currentConstraintRow[j];\n"
+"			if (solverConstraint->m_upperLimit>=constraint->m_breakingImpulseThreshold)\n"
+"			{\n"
+"				solverConstraint->m_upperLimit = constraint->m_breakingImpulseThreshold;\n"
+"			}\n"
+"			if (solverConstraint->m_lowerLimit<=-constraint->m_breakingImpulseThreshold)\n"
+"			{\n"
+"				solverConstraint->m_lowerLimit = -constraint->m_breakingImpulseThreshold;\n"
+"			}\n"
+"//						solverConstraint->m_originalContactPoint = constraint;\n"
+"							\n"
+"			Matrix3x3 invInertiaWorldA= inertias[constraint->m_rbA].m_invInertiaWorld;\n"
+"			{\n"
+"				//float4 angularFactorA(1,1,1);\n"
+"				float4 ftorqueAxis1 = solverConstraint->m_relpos1CrossNormal;\n"
+"				solverConstraint->m_angularComponentA = mtMul1(invInertiaWorldA,ftorqueAxis1);//*angularFactorA;\n"
+"			}\n"
+"						\n"
+"			Matrix3x3 invInertiaWorldB= inertias[constraint->m_rbB].m_invInertiaWorld;\n"
+"			{\n"
+"				float4 ftorqueAxis2 = solverConstraint->m_relpos2CrossNormal;\n"
+"				solverConstraint->m_angularComponentB = mtMul1(invInertiaWorldB,ftorqueAxis2);//*constraint->m_rbB.getAngularFactor();\n"
+"			}\n"
+"			{\n"
+"				//it is ok to use solverConstraint->m_contactNormal instead of -solverConstraint->m_contactNormal\n"
+"				//because it gets multiplied iMJlB\n"
+"				float4 iMJlA = solverConstraint->m_contactNormal*rbA->m_invMass;\n"
+"				float4 iMJaA = mtMul3(solverConstraint->m_relpos1CrossNormal,invInertiaWorldA);\n"
+"				float4 iMJlB = solverConstraint->m_contactNormal*rbB->m_invMass;//sign of normal?\n"
+"				float4 iMJaB = mtMul3(solverConstraint->m_relpos2CrossNormal,invInertiaWorldB);\n"
+"				float sum = dot3F4(iMJlA,solverConstraint->m_contactNormal);\n"
+"				sum += dot3F4(iMJaA,solverConstraint->m_relpos1CrossNormal);\n"
+"				sum += dot3F4(iMJlB,solverConstraint->m_contactNormal);\n"
+"				sum += dot3F4(iMJaB,solverConstraint->m_relpos2CrossNormal);\n"
+"				float fsum = fabs(sum);\n"
+"				if (fsum>FLT_EPSILON)\n"
+"				{\n"
+"					solverConstraint->m_jacDiagABInv = 1.f/sum;\n"
+"				} else\n"
+"				{\n"
+"					solverConstraint->m_jacDiagABInv = 0.f;\n"
+"				}\n"
+"			}\n"
+"			///fix rhs\n"
+"			///todo: add force/torque accelerators\n"
+"			{\n"
+"				float rel_vel;\n"
+"				float vel1Dotn = dot3F4(solverConstraint->m_contactNormal,rbA->m_linVel) + dot3F4(solverConstraint->m_relpos1CrossNormal,rbA->m_angVel);\n"
+"				float vel2Dotn = -dot3F4(solverConstraint->m_contactNormal,rbB->m_linVel) + dot3F4(solverConstraint->m_relpos2CrossNormal,rbB->m_angVel);\n"
+"				rel_vel = vel1Dotn+vel2Dotn;\n"
+"				float restitution = 0.f;\n"
+"				float positionalError = solverConstraint->m_rhs;//already filled in by getConstraintInfo2\n"
+"				float	velocityError = restitution - rel_vel * info2.m_damping;\n"
+"				float	penetrationImpulse = positionalError*solverConstraint->m_jacDiagABInv;\n"
+"				float	velocityImpulse = velocityError *solverConstraint->m_jacDiagABInv;\n"
+"				solverConstraint->m_rhs = penetrationImpulse+velocityImpulse;\n"
+"				solverConstraint->m_appliedImpulse = 0.f;\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solveContact.cl b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solveContact.cl
new file mode 100644
index 0000000000..5c4d62e4ec
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solveContact.cl
@@ -0,0 +1,501 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+//#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile global int*
+#endif
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define mymake_float4 (float4)
+//#define make_float2 (float2)
+//#define make_uint4 (uint4)
+//#define make_int4 (int4)
+//#define make_uint2 (uint2)
+//#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+///////////////////////////////////////
+//	Vector
+///////////////////////////////////////
+
+
+
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	return fast_normalize(v);
+}
+
+
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = mymake_float4(a.xyz,0.f);
+	float4 b1 = mymake_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+
+
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = mymake_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+//	float length = sqrtf(dot3F4(a, a));
+//	return 1.f/length * a;
+}
+
+
+
+
+///////////////////////////////////////
+//	Matrix3x3
+///////////////////////////////////////
+
+typedef struct
+{
+	float4 m_row[3];
+}Matrix3x3;
+
+
+
+
+
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b);
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b);
+
+
+
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b)
+{
+	float4 ans;
+	ans.x = dot3F4( a.m_row[0], b );
+	ans.y = dot3F4( a.m_row[1], b );
+	ans.z = dot3F4( a.m_row[2], b );
+	ans.w = 0.f;
+	return ans;
+}
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b)
+{
+	float4 colx = mymake_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);
+	float4 coly = mymake_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);
+	float4 colz = mymake_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);
+
+	float4 ans;
+	ans.x = dot3F4( a, colx );
+	ans.y = dot3F4( a, coly );
+	ans.z = dot3F4( a, colz );
+	return ans;
+}
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+
+
+
+
+
+
+#define WG_SIZE 64
+
+typedef struct
+{
+	float4 m_pos;
+	Quaternion m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_shapeIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} Body;
+
+typedef struct
+{
+	Matrix3x3 m_invInertia;
+	Matrix3x3 m_initInvInertia;
+} Shape;
+
+typedef struct
+{
+	float4 m_linear;
+	float4 m_worldPos[4];
+	float4 m_center;	
+	float m_jacCoeffInv[4];
+	float m_b[4];
+	float m_appliedRambdaDt[4];
+
+	float m_fJacCoeffInv[2];	
+	float m_fAppliedRambdaDt[2];	
+
+	u32 m_bodyA;
+	u32 m_bodyB;
+
+	int m_batchIdx;
+	u32 m_paddings[1];
+} Constraint4;
+
+
+
+typedef struct
+{
+	int m_nConstraints;
+	int m_start;
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBuffer;
+
+typedef struct
+{
+	int m_solveFriction;
+	int m_maxBatch;	//	long batch really kills the performance
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBufferBatchSolve;
+
+void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1);
+
+void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)
+{
+	*linear = mymake_float4(-n.xyz,0.f);
+	*angular0 = -cross3(r0, n);
+	*angular1 = cross3(r1, n);
+}
+
+float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 );
+
+float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )
+{
+	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);
+}
+
+
+float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,
+				   float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1);
+
+float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,
+					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1)
+{
+	//	linear0,1 are normlized
+	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;
+	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);
+	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;
+	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);
+	return -1.f/(jmj0+jmj1+jmj2+jmj3);
+}
+
+
+void solveContact(__global Constraint4* cs,
+				  float4 posA, float4* linVelA, float4* angVelA, float invMassA, Matrix3x3 invInertiaA,
+				  float4 posB, float4* linVelB, float4* angVelB, float invMassB, Matrix3x3 invInertiaB);
+
+void solveContact(__global Constraint4* cs,
+			float4 posA, float4* linVelA, float4* angVelA, float invMassA, Matrix3x3 invInertiaA,
+			float4 posB, float4* linVelB, float4* angVelB, float invMassB, Matrix3x3 invInertiaB)
+{
+	float minRambdaDt = 0;
+	float maxRambdaDt = FLT_MAX;
+
+	for(int ic=0; ic<4; ic++)
+	{
+		if( cs->m_jacCoeffInv[ic] == 0.f ) continue;
+
+		float4 angular0, angular1, linear;
+		float4 r0 = cs->m_worldPos[ic] - posA;
+		float4 r1 = cs->m_worldPos[ic] - posB;
+		setLinearAndAngular( -cs->m_linear, r0, r1, &linear, &angular0, &angular1 );
+
+		float rambdaDt = calcRelVel( cs->m_linear, -cs->m_linear, angular0, angular1, 
+			*linVelA, *angVelA, *linVelB, *angVelB ) + cs->m_b[ic];
+		rambdaDt *= cs->m_jacCoeffInv[ic];
+
+		{
+			float prevSum = cs->m_appliedRambdaDt[ic];
+			float updated = prevSum;
+			updated += rambdaDt;
+			updated = max2( updated, minRambdaDt );
+			updated = min2( updated, maxRambdaDt );
+			rambdaDt = updated - prevSum;
+			cs->m_appliedRambdaDt[ic] = updated;
+		}
+
+		float4 linImp0 = invMassA*linear*rambdaDt;
+		float4 linImp1 = invMassB*(-linear)*rambdaDt;
+		float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;
+		float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;
+
+		*linVelA += linImp0;
+		*angVelA += angImp0;
+		*linVelB += linImp1;
+		*angVelB += angImp1;
+	}
+}
+
+void btPlaneSpace1 (const float4* n, float4* p, float4* q);
+ void btPlaneSpace1 (const float4* n, float4* p, float4* q)
+{
+  if (fabs(n[0].z) > 0.70710678f) {
+    // choose p in y-z plane
+    float a = n[0].y*n[0].y + n[0].z*n[0].z;
+    float k = 1.f/sqrt(a);
+    p[0].x = 0;
+	p[0].y = -n[0].z*k;
+	p[0].z = n[0].y*k;
+    // set q = n x p
+    q[0].x = a*k;
+	q[0].y = -n[0].x*p[0].z;
+	q[0].z = n[0].x*p[0].y;
+  }
+  else {
+    // choose p in x-y plane
+    float a = n[0].x*n[0].x + n[0].y*n[0].y;
+    float k = 1.f/sqrt(a);
+    p[0].x = -n[0].y*k;
+	p[0].y = n[0].x*k;
+	p[0].z = 0;
+    // set q = n x p
+    q[0].x = -n[0].z*p[0].y;
+	q[0].y = n[0].z*p[0].x;
+	q[0].z = a*k;
+  }
+}
+
+void solveContactConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs);
+void solveContactConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs)
+{
+	//float frictionCoeff = ldsCs[0].m_linear.w;
+	int aIdx = ldsCs[0].m_bodyA;
+	int bIdx = ldsCs[0].m_bodyB;
+
+	float4 posA = gBodies[aIdx].m_pos;
+	float4 linVelA = gBodies[aIdx].m_linVel;
+	float4 angVelA = gBodies[aIdx].m_angVel;
+	float invMassA = gBodies[aIdx].m_invMass;
+	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;
+
+	float4 posB = gBodies[bIdx].m_pos;
+	float4 linVelB = gBodies[bIdx].m_linVel;
+	float4 angVelB = gBodies[bIdx].m_angVel;
+	float invMassB = gBodies[bIdx].m_invMass;
+	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;
+
+	solveContact( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,
+			posB, &linVelB, &angVelB, invMassB, invInertiaB );
+
+  if (gBodies[aIdx].m_invMass)
+  {
+		gBodies[aIdx].m_linVel = linVelA;
+		gBodies[aIdx].m_angVel = angVelA;
+	} else
+	{
+		gBodies[aIdx].m_linVel = mymake_float4(0,0,0,0);
+		gBodies[aIdx].m_angVel = mymake_float4(0,0,0,0);
+	
+	}
+	if (gBodies[bIdx].m_invMass)
+  {
+		gBodies[bIdx].m_linVel = linVelB;
+		gBodies[bIdx].m_angVel = angVelB;
+	} else
+	{
+		gBodies[bIdx].m_linVel = mymake_float4(0,0,0,0);
+		gBodies[bIdx].m_angVel = mymake_float4(0,0,0,0);
+	
+	}
+
+}
+
+
+
+typedef struct 
+{
+	int m_valInt0;
+	int m_valInt1;
+	int m_valInt2;
+	int m_valInt3;
+
+	float m_val0;
+	float m_val1;
+	float m_val2;
+	float m_val3;
+} SolverDebugInfo;
+
+
+
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void BatchSolveKernelContact(__global Body* gBodies,
+                      __global Shape* gShapes,
+                      __global Constraint4* gConstraints,
+                      __global int* gN,
+                      __global int* gOffsets,
+                      __global	int* batchSizes,
+                       int maxBatch1,
+                       int cellBatch,
+                       int4 nSplit
+                      )
+{
+	//__local int ldsBatchIdx[WG_SIZE+1];
+	__local int ldsCurBatch;
+	__local int ldsNextBatch;
+	__local int ldsStart;
+
+	int lIdx = GET_LOCAL_IDX;
+	int wgIdx = GET_GROUP_IDX;
+
+//	int gIdx = GET_GLOBAL_IDX;
+//	debugInfo[gIdx].m_valInt0 = gIdx;
+	//debugInfo[gIdx].m_valInt1 = GET_GROUP_SIZE;
+
+	
+	
+
+	int zIdx = (wgIdx/((nSplit.x*nSplit.y)/4))*2+((cellBatch&4)>>2);
+	int remain= (wgIdx%((nSplit.x*nSplit.y)/4));
+	int yIdx = (remain/(nSplit.x/2))*2 + ((cellBatch&2)>>1);
+	int xIdx = (remain%(nSplit.x/2))*2 + (cellBatch&1);
+	int cellIdx = xIdx+yIdx*nSplit.x+zIdx*(nSplit.x*nSplit.y);
+
+	//int xIdx = (wgIdx/(nSplit/2))*2 + (bIdx&1);
+	//int yIdx = (wgIdx%(nSplit/2))*2 + (bIdx>>1);
+	//int cellIdx = xIdx+yIdx*nSplit;
+	
+	if( gN[cellIdx] == 0 ) 
+		return;
+
+	int maxBatch = batchSizes[cellIdx];
+	
+	
+	const int start = gOffsets[cellIdx];
+	const int end = start + gN[cellIdx];
+
+	
+	
+	
+	if( lIdx == 0 )
+	{
+		ldsCurBatch = 0;
+		ldsNextBatch = 0;
+		ldsStart = start;
+	}
+
+
+	GROUP_LDS_BARRIER;
+
+	int idx=ldsStart+lIdx;
+	while (ldsCurBatch < maxBatch)
+	{
+		for(; idx<end; )
+		{
+			if (gConstraints[idx].m_batchIdx == ldsCurBatch)
+			{
+					solveContactConstraint( gBodies, gShapes, &gConstraints[idx] );
+
+				 idx+=64;
+			} else
+			{
+				break;
+			}
+		}
+		GROUP_LDS_BARRIER;
+	
+		if( lIdx == 0 )
+		{
+			ldsCurBatch++;
+		}
+		GROUP_LDS_BARRIER;
+	}
+	
+    
+}
+
+
+
+__kernel void solveSingleContactKernel(__global Body* gBodies,
+                      __global Shape* gShapes,
+                      __global Constraint4* gConstraints,
+                       int cellIdx,
+                       int batchOffset,
+                       int numConstraintsInBatch
+                      )
+{
+
+	int index = get_global_id(0);
+	if (index < numConstraintsInBatch)
+	{
+		int idx=batchOffset+index;
+		solveContactConstraint( gBodies, gShapes, &gConstraints[idx] );
+	}    
+}
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solveContact.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solveContact.h
new file mode 100644
index 0000000000..15a049992b
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solveContact.h
@@ -0,0 +1,393 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* solveContactCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"//#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile global int*\n"
+"#endif\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"#define mymake_float4 (float4)\n"
+"//#define make_float2 (float2)\n"
+"//#define make_uint4 (uint4)\n"
+"//#define make_int4 (int4)\n"
+"//#define make_uint2 (uint2)\n"
+"//#define make_int2 (int2)\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"///////////////////////////////////////\n"
+"//	Vector\n"
+"///////////////////////////////////////\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = mymake_float4(a.xyz,0.f);\n"
+"	float4 b1 = mymake_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = mymake_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"//	float length = sqrtf(dot3F4(a, a));\n"
+"//	return 1.f/length * a;\n"
+"}\n"
+"///////////////////////////////////////\n"
+"//	Matrix3x3\n"
+"///////////////////////////////////////\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_row[3];\n"
+"}Matrix3x3;\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b);\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b);\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b)\n"
+"{\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a.m_row[0], b );\n"
+"	ans.y = dot3F4( a.m_row[1], b );\n"
+"	ans.z = dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b)\n"
+"{\n"
+"	float4 colx = mymake_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	float4 coly = mymake_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	float4 colz = mymake_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a, colx );\n"
+"	ans.y = dot3F4( a, coly );\n"
+"	ans.z = dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"typedef float4 Quaternion;\n"
+"#define WG_SIZE 64\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	Quaternion m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"	u32 m_shapeIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} Body;\n"
+"typedef struct\n"
+"{\n"
+"	Matrix3x3 m_invInertia;\n"
+"	Matrix3x3 m_initInvInertia;\n"
+"} Shape;\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_linear;\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_center;	\n"
+"	float m_jacCoeffInv[4];\n"
+"	float m_b[4];\n"
+"	float m_appliedRambdaDt[4];\n"
+"	float m_fJacCoeffInv[2];	\n"
+"	float m_fAppliedRambdaDt[2];	\n"
+"	u32 m_bodyA;\n"
+"	u32 m_bodyB;\n"
+"	int m_batchIdx;\n"
+"	u32 m_paddings[1];\n"
+"} Constraint4;\n"
+"typedef struct\n"
+"{\n"
+"	int m_nConstraints;\n"
+"	int m_start;\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBuffer;\n"
+"typedef struct\n"
+"{\n"
+"	int m_solveFriction;\n"
+"	int m_maxBatch;	//	long batch really kills the performance\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBufferBatchSolve;\n"
+"void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1);\n"
+"void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)\n"
+"{\n"
+"	*linear = mymake_float4(-n.xyz,0.f);\n"
+"	*angular0 = -cross3(r0, n);\n"
+"	*angular1 = cross3(r1, n);\n"
+"}\n"
+"float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 );\n"
+"float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )\n"
+"{\n"
+"	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);\n"
+"}\n"
+"float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,\n"
+"				   float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1);\n"
+"float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,\n"
+"					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1)\n"
+"{\n"
+"	//	linear0,1 are normlized\n"
+"	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;\n"
+"	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);\n"
+"	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;\n"
+"	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);\n"
+"	return -1.f/(jmj0+jmj1+jmj2+jmj3);\n"
+"}\n"
+"void solveContact(__global Constraint4* cs,\n"
+"				  float4 posA, float4* linVelA, float4* angVelA, float invMassA, Matrix3x3 invInertiaA,\n"
+"				  float4 posB, float4* linVelB, float4* angVelB, float invMassB, Matrix3x3 invInertiaB);\n"
+"void solveContact(__global Constraint4* cs,\n"
+"			float4 posA, float4* linVelA, float4* angVelA, float invMassA, Matrix3x3 invInertiaA,\n"
+"			float4 posB, float4* linVelB, float4* angVelB, float invMassB, Matrix3x3 invInertiaB)\n"
+"{\n"
+"	float minRambdaDt = 0;\n"
+"	float maxRambdaDt = FLT_MAX;\n"
+"	for(int ic=0; ic<4; ic++)\n"
+"	{\n"
+"		if( cs->m_jacCoeffInv[ic] == 0.f ) continue;\n"
+"		float4 angular0, angular1, linear;\n"
+"		float4 r0 = cs->m_worldPos[ic] - posA;\n"
+"		float4 r1 = cs->m_worldPos[ic] - posB;\n"
+"		setLinearAndAngular( -cs->m_linear, r0, r1, &linear, &angular0, &angular1 );\n"
+"		float rambdaDt = calcRelVel( cs->m_linear, -cs->m_linear, angular0, angular1, \n"
+"			*linVelA, *angVelA, *linVelB, *angVelB ) + cs->m_b[ic];\n"
+"		rambdaDt *= cs->m_jacCoeffInv[ic];\n"
+"		{\n"
+"			float prevSum = cs->m_appliedRambdaDt[ic];\n"
+"			float updated = prevSum;\n"
+"			updated += rambdaDt;\n"
+"			updated = max2( updated, minRambdaDt );\n"
+"			updated = min2( updated, maxRambdaDt );\n"
+"			rambdaDt = updated - prevSum;\n"
+"			cs->m_appliedRambdaDt[ic] = updated;\n"
+"		}\n"
+"		float4 linImp0 = invMassA*linear*rambdaDt;\n"
+"		float4 linImp1 = invMassB*(-linear)*rambdaDt;\n"
+"		float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;\n"
+"		float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;\n"
+"		*linVelA += linImp0;\n"
+"		*angVelA += angImp0;\n"
+"		*linVelB += linImp1;\n"
+"		*angVelB += angImp1;\n"
+"	}\n"
+"}\n"
+"void btPlaneSpace1 (const float4* n, float4* p, float4* q);\n"
+" void btPlaneSpace1 (const float4* n, float4* p, float4* q)\n"
+"{\n"
+"  if (fabs(n[0].z) > 0.70710678f) {\n"
+"    // choose p in y-z plane\n"
+"    float a = n[0].y*n[0].y + n[0].z*n[0].z;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = 0;\n"
+"	p[0].y = -n[0].z*k;\n"
+"	p[0].z = n[0].y*k;\n"
+"    // set q = n x p\n"
+"    q[0].x = a*k;\n"
+"	q[0].y = -n[0].x*p[0].z;\n"
+"	q[0].z = n[0].x*p[0].y;\n"
+"  }\n"
+"  else {\n"
+"    // choose p in x-y plane\n"
+"    float a = n[0].x*n[0].x + n[0].y*n[0].y;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = -n[0].y*k;\n"
+"	p[0].y = n[0].x*k;\n"
+"	p[0].z = 0;\n"
+"    // set q = n x p\n"
+"    q[0].x = -n[0].z*p[0].y;\n"
+"	q[0].y = n[0].z*p[0].x;\n"
+"	q[0].z = a*k;\n"
+"  }\n"
+"}\n"
+"void solveContactConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs);\n"
+"void solveContactConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs)\n"
+"{\n"
+"	//float frictionCoeff = ldsCs[0].m_linear.w;\n"
+"	int aIdx = ldsCs[0].m_bodyA;\n"
+"	int bIdx = ldsCs[0].m_bodyB;\n"
+"	float4 posA = gBodies[aIdx].m_pos;\n"
+"	float4 linVelA = gBodies[aIdx].m_linVel;\n"
+"	float4 angVelA = gBodies[aIdx].m_angVel;\n"
+"	float invMassA = gBodies[aIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;\n"
+"	float4 posB = gBodies[bIdx].m_pos;\n"
+"	float4 linVelB = gBodies[bIdx].m_linVel;\n"
+"	float4 angVelB = gBodies[bIdx].m_angVel;\n"
+"	float invMassB = gBodies[bIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;\n"
+"	solveContact( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,\n"
+"			posB, &linVelB, &angVelB, invMassB, invInertiaB );\n"
+"  if (gBodies[aIdx].m_invMass)\n"
+"  {\n"
+"		gBodies[aIdx].m_linVel = linVelA;\n"
+"		gBodies[aIdx].m_angVel = angVelA;\n"
+"	} else\n"
+"	{\n"
+"		gBodies[aIdx].m_linVel = mymake_float4(0,0,0,0);\n"
+"		gBodies[aIdx].m_angVel = mymake_float4(0,0,0,0);\n"
+"	\n"
+"	}\n"
+"	if (gBodies[bIdx].m_invMass)\n"
+"  {\n"
+"		gBodies[bIdx].m_linVel = linVelB;\n"
+"		gBodies[bIdx].m_angVel = angVelB;\n"
+"	} else\n"
+"	{\n"
+"		gBodies[bIdx].m_linVel = mymake_float4(0,0,0,0);\n"
+"		gBodies[bIdx].m_angVel = mymake_float4(0,0,0,0);\n"
+"	\n"
+"	}\n"
+"}\n"
+"typedef struct \n"
+"{\n"
+"	int m_valInt0;\n"
+"	int m_valInt1;\n"
+"	int m_valInt2;\n"
+"	int m_valInt3;\n"
+"	float m_val0;\n"
+"	float m_val1;\n"
+"	float m_val2;\n"
+"	float m_val3;\n"
+"} SolverDebugInfo;\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void BatchSolveKernelContact(__global Body* gBodies,\n"
+"                      __global Shape* gShapes,\n"
+"                      __global Constraint4* gConstraints,\n"
+"                      __global int* gN,\n"
+"                      __global int* gOffsets,\n"
+"                      __global	int* batchSizes,\n"
+"                       int maxBatch1,\n"
+"                       int cellBatch,\n"
+"                       int4 nSplit\n"
+"                      )\n"
+"{\n"
+"	//__local int ldsBatchIdx[WG_SIZE+1];\n"
+"	__local int ldsCurBatch;\n"
+"	__local int ldsNextBatch;\n"
+"	__local int ldsStart;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	int wgIdx = GET_GROUP_IDX;\n"
+"//	int gIdx = GET_GLOBAL_IDX;\n"
+"//	debugInfo[gIdx].m_valInt0 = gIdx;\n"
+"	//debugInfo[gIdx].m_valInt1 = GET_GROUP_SIZE;\n"
+"	\n"
+"	\n"
+"	int zIdx = (wgIdx/((nSplit.x*nSplit.y)/4))*2+((cellBatch&4)>>2);\n"
+"	int remain= (wgIdx%((nSplit.x*nSplit.y)/4));\n"
+"	int yIdx = (remain/(nSplit.x/2))*2 + ((cellBatch&2)>>1);\n"
+"	int xIdx = (remain%(nSplit.x/2))*2 + (cellBatch&1);\n"
+"	int cellIdx = xIdx+yIdx*nSplit.x+zIdx*(nSplit.x*nSplit.y);\n"
+"	//int xIdx = (wgIdx/(nSplit/2))*2 + (bIdx&1);\n"
+"	//int yIdx = (wgIdx%(nSplit/2))*2 + (bIdx>>1);\n"
+"	//int cellIdx = xIdx+yIdx*nSplit;\n"
+"	\n"
+"	if( gN[cellIdx] == 0 ) \n"
+"		return;\n"
+"	int maxBatch = batchSizes[cellIdx];\n"
+"	\n"
+"	\n"
+"	const int start = gOffsets[cellIdx];\n"
+"	const int end = start + gN[cellIdx];\n"
+"	\n"
+"	\n"
+"	\n"
+"	if( lIdx == 0 )\n"
+"	{\n"
+"		ldsCurBatch = 0;\n"
+"		ldsNextBatch = 0;\n"
+"		ldsStart = start;\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	int idx=ldsStart+lIdx;\n"
+"	while (ldsCurBatch < maxBatch)\n"
+"	{\n"
+"		for(; idx<end; )\n"
+"		{\n"
+"			if (gConstraints[idx].m_batchIdx == ldsCurBatch)\n"
+"			{\n"
+"					solveContactConstraint( gBodies, gShapes, &gConstraints[idx] );\n"
+"				 idx+=64;\n"
+"			} else\n"
+"			{\n"
+"				break;\n"
+"			}\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"	\n"
+"		if( lIdx == 0 )\n"
+"		{\n"
+"			ldsCurBatch++;\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"	}\n"
+"	\n"
+"    \n"
+"}\n"
+"__kernel void solveSingleContactKernel(__global Body* gBodies,\n"
+"                      __global Shape* gShapes,\n"
+"                      __global Constraint4* gConstraints,\n"
+"                       int cellIdx,\n"
+"                       int batchOffset,\n"
+"                       int numConstraintsInBatch\n"
+"                      )\n"
+"{\n"
+"	int index = get_global_id(0);\n"
+"	if (index < numConstraintsInBatch)\n"
+"	{\n"
+"		int idx=batchOffset+index;\n"
+"		solveContactConstraint( gBodies, gShapes, &gConstraints[idx] );\n"
+"	}    \n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solveFriction.cl b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solveFriction.cl
new file mode 100644
index 0000000000..1d70fbbae3
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solveFriction.cl
@@ -0,0 +1,527 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+//#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile global int*
+#endif
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define mymake_float4 (float4)
+//#define make_float2 (float2)
+//#define make_uint4 (uint4)
+//#define make_int4 (int4)
+//#define make_uint2 (uint2)
+//#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+///////////////////////////////////////
+//	Vector
+///////////////////////////////////////
+
+
+
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	return fast_normalize(v);
+}
+
+
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = mymake_float4(a.xyz,0.f);
+	float4 b1 = mymake_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+
+
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = mymake_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+//	float length = sqrtf(dot3F4(a, a));
+//	return 1.f/length * a;
+}
+
+
+
+
+///////////////////////////////////////
+//	Matrix3x3
+///////////////////////////////////////
+
+typedef struct
+{
+	float4 m_row[3];
+}Matrix3x3;
+
+
+
+
+
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b);
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b);
+
+
+
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b)
+{
+	float4 ans;
+	ans.x = dot3F4( a.m_row[0], b );
+	ans.y = dot3F4( a.m_row[1], b );
+	ans.z = dot3F4( a.m_row[2], b );
+	ans.w = 0.f;
+	return ans;
+}
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b)
+{
+	float4 colx = mymake_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);
+	float4 coly = mymake_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);
+	float4 colz = mymake_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);
+
+	float4 ans;
+	ans.x = dot3F4( a, colx );
+	ans.y = dot3F4( a, coly );
+	ans.z = dot3F4( a, colz );
+	return ans;
+}
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+
+
+
+
+
+
+#define WG_SIZE 64
+
+typedef struct
+{
+	float4 m_pos;
+	Quaternion m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_shapeIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} Body;
+
+typedef struct
+{
+	Matrix3x3 m_invInertia;
+	Matrix3x3 m_initInvInertia;
+} Shape;
+
+typedef struct
+{
+	float4 m_linear;
+	float4 m_worldPos[4];
+	float4 m_center;	
+	float m_jacCoeffInv[4];
+	float m_b[4];
+	float m_appliedRambdaDt[4];
+
+	float m_fJacCoeffInv[2];	
+	float m_fAppliedRambdaDt[2];	
+
+	u32 m_bodyA;
+	u32 m_bodyB;
+
+	int m_batchIdx;
+	u32 m_paddings[1];
+} Constraint4;
+
+
+
+typedef struct
+{
+	int m_nConstraints;
+	int m_start;
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBuffer;
+
+typedef struct
+{
+	int m_solveFriction;
+	int m_maxBatch;	//	long batch really kills the performance
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBufferBatchSolve;
+
+void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1);
+
+void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)
+{
+	*linear = mymake_float4(-n.xyz,0.f);
+	*angular0 = -cross3(r0, n);
+	*angular1 = cross3(r1, n);
+}
+
+float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 );
+
+float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )
+{
+	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);
+}
+
+
+float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,
+				   float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1);
+
+float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,
+					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1)
+{
+	//	linear0,1 are normlized
+	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;
+	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);
+	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;
+	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);
+	return -1.f/(jmj0+jmj1+jmj2+jmj3);
+}
+void btPlaneSpace1 (const float4* n, float4* p, float4* q);
+ void btPlaneSpace1 (const float4* n, float4* p, float4* q)
+{
+  if (fabs(n[0].z) > 0.70710678f) {
+    // choose p in y-z plane
+    float a = n[0].y*n[0].y + n[0].z*n[0].z;
+    float k = 1.f/sqrt(a);
+    p[0].x = 0;
+	p[0].y = -n[0].z*k;
+	p[0].z = n[0].y*k;
+    // set q = n x p
+    q[0].x = a*k;
+	q[0].y = -n[0].x*p[0].z;
+	q[0].z = n[0].x*p[0].y;
+  }
+  else {
+    // choose p in x-y plane
+    float a = n[0].x*n[0].x + n[0].y*n[0].y;
+    float k = 1.f/sqrt(a);
+    p[0].x = -n[0].y*k;
+	p[0].y = n[0].x*k;
+	p[0].z = 0;
+    // set q = n x p
+    q[0].x = -n[0].z*p[0].y;
+	q[0].y = n[0].z*p[0].x;
+	q[0].z = a*k;
+  }
+}
+
+
+void solveFrictionConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs);
+void solveFrictionConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs)
+{
+	float frictionCoeff = ldsCs[0].m_linear.w;
+	int aIdx = ldsCs[0].m_bodyA;
+	int bIdx = ldsCs[0].m_bodyB;
+
+
+	float4 posA = gBodies[aIdx].m_pos;
+	float4 linVelA = gBodies[aIdx].m_linVel;
+	float4 angVelA = gBodies[aIdx].m_angVel;
+	float invMassA = gBodies[aIdx].m_invMass;
+	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;
+
+	float4 posB = gBodies[bIdx].m_pos;
+	float4 linVelB = gBodies[bIdx].m_linVel;
+	float4 angVelB = gBodies[bIdx].m_angVel;
+	float invMassB = gBodies[bIdx].m_invMass;
+	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;
+	
+
+	{
+		float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};
+		float minRambdaDt[4] = {0.f,0.f,0.f,0.f};
+
+		float sum = 0;
+		for(int j=0; j<4; j++)
+		{
+			sum +=ldsCs[0].m_appliedRambdaDt[j];
+		}
+		frictionCoeff = 0.7f;
+		for(int j=0; j<4; j++)
+		{
+			maxRambdaDt[j] = frictionCoeff*sum;
+			minRambdaDt[j] = -maxRambdaDt[j];
+		}
+
+		
+//		solveFriction( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,
+//			posB, &linVelB, &angVelB, invMassB, invInertiaB, maxRambdaDt, minRambdaDt );
+		
+		
+		{
+			
+			__global Constraint4* cs = ldsCs;
+			
+			if( cs->m_fJacCoeffInv[0] == 0 && cs->m_fJacCoeffInv[0] == 0 ) return;
+			const float4 center = cs->m_center;
+			
+			float4 n = -cs->m_linear;
+			
+			float4 tangent[2];
+			btPlaneSpace1(&n,&tangent[0],&tangent[1]);
+			float4 angular0, angular1, linear;
+			float4 r0 = center - posA;
+			float4 r1 = center - posB;
+			for(int i=0; i<2; i++)
+			{
+				setLinearAndAngular( tangent[i], r0, r1, &linear, &angular0, &angular1 );
+				float rambdaDt = calcRelVel(linear, -linear, angular0, angular1,
+											linVelA, angVelA, linVelB, angVelB );
+				rambdaDt *= cs->m_fJacCoeffInv[i];
+				
+				{
+					float prevSum = cs->m_fAppliedRambdaDt[i];
+					float updated = prevSum;
+					updated += rambdaDt;
+					updated = max2( updated, minRambdaDt[i] );
+					updated = min2( updated, maxRambdaDt[i] );
+					rambdaDt = updated - prevSum;
+					cs->m_fAppliedRambdaDt[i] = updated;
+				}
+				
+				float4 linImp0 = invMassA*linear*rambdaDt;
+				float4 linImp1 = invMassB*(-linear)*rambdaDt;
+				float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;
+				float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;
+				
+				linVelA += linImp0;
+				angVelA += angImp0;
+				linVelB += linImp1;
+				angVelB += angImp1;
+			}
+			{	//	angular damping for point constraint
+				float4 ab = normalize3( posB - posA );
+				float4 ac = normalize3( center - posA );
+				if( dot3F4( ab, ac ) > 0.95f  || (invMassA == 0.f || invMassB == 0.f))
+				{
+					float angNA = dot3F4( n, angVelA );
+					float angNB = dot3F4( n, angVelB );
+					
+					angVelA -= (angNA*0.1f)*n;
+					angVelB -= (angNB*0.1f)*n;
+				}
+			}
+		}
+
+		
+		
+	}
+
+	if (gBodies[aIdx].m_invMass)
+	{
+		gBodies[aIdx].m_linVel = linVelA;
+		gBodies[aIdx].m_angVel = angVelA;
+	} else
+	{
+		gBodies[aIdx].m_linVel = mymake_float4(0,0,0,0);
+		gBodies[aIdx].m_angVel = mymake_float4(0,0,0,0);
+	}
+	if (gBodies[bIdx].m_invMass)
+	{
+		gBodies[bIdx].m_linVel = linVelB;
+		gBodies[bIdx].m_angVel = angVelB;
+	} else
+	{
+		gBodies[bIdx].m_linVel = mymake_float4(0,0,0,0);
+		gBodies[bIdx].m_angVel = mymake_float4(0,0,0,0);
+	}
+ 
+
+}
+
+typedef struct 
+{
+	int m_valInt0;
+	int m_valInt1;
+	int m_valInt2;
+	int m_valInt3;
+
+	float m_val0;
+	float m_val1;
+	float m_val2;
+	float m_val3;
+} SolverDebugInfo;
+
+
+
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void BatchSolveKernelFriction(__global Body* gBodies,
+                      __global Shape* gShapes,
+                      __global Constraint4* gConstraints,
+                      __global int* gN,
+                      __global int* gOffsets,
+                      __global int* batchSizes,
+                       int maxBatch1,
+                       int cellBatch,
+                       int4 nSplit
+                      )
+{
+	//__local int ldsBatchIdx[WG_SIZE+1];
+	__local int ldsCurBatch;
+	__local int ldsNextBatch;
+	__local int ldsStart;
+
+	int lIdx = GET_LOCAL_IDX;
+	int wgIdx = GET_GROUP_IDX;
+
+//	int gIdx = GET_GLOBAL_IDX;
+//	debugInfo[gIdx].m_valInt0 = gIdx;
+	//debugInfo[gIdx].m_valInt1 = GET_GROUP_SIZE;
+
+
+	int zIdx = (wgIdx/((nSplit.x*nSplit.y)/4))*2+((cellBatch&4)>>2);
+	int remain= (wgIdx%((nSplit.x*nSplit.y)/4));
+	int yIdx = (remain/(nSplit.x/2))*2 + ((cellBatch&2)>>1);
+	int xIdx = (remain%(nSplit.x/2))*2 + (cellBatch&1);
+	int cellIdx = xIdx+yIdx*nSplit.x+zIdx*(nSplit.x*nSplit.y);
+
+	
+	if( gN[cellIdx] == 0 ) 
+		return;
+
+	int maxBatch = batchSizes[cellIdx];
+
+	const int start = gOffsets[cellIdx];
+	const int end = start + gN[cellIdx];
+
+	
+	if( lIdx == 0 )
+	{
+		ldsCurBatch = 0;
+		ldsNextBatch = 0;
+		ldsStart = start;
+	}
+
+
+	GROUP_LDS_BARRIER;
+
+	int idx=ldsStart+lIdx;
+	while (ldsCurBatch < maxBatch)
+	{
+		for(; idx<end; )
+		{
+			if (gConstraints[idx].m_batchIdx == ldsCurBatch)
+			{
+
+					solveFrictionConstraint( gBodies, gShapes, &gConstraints[idx] );
+
+				 idx+=64;
+			} else
+			{
+				break;
+			}
+		}
+		GROUP_LDS_BARRIER;
+		if( lIdx == 0 )
+		{
+			ldsCurBatch++;
+		}
+		GROUP_LDS_BARRIER;
+	}
+	
+    
+}
+
+
+
+
+
+
+__kernel void solveSingleFrictionKernel(__global Body* gBodies,
+                      __global Shape* gShapes,
+                      __global Constraint4* gConstraints,
+                       int cellIdx,
+                       int batchOffset,
+                       int numConstraintsInBatch
+                      )
+{
+
+	int index = get_global_id(0);
+	if (index < numConstraintsInBatch)
+	{
+		
+		int idx=batchOffset+index;
+	
+		solveFrictionConstraint( gBodies, gShapes, &gConstraints[idx] );
+	}    
+}
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solveFriction.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solveFriction.h
new file mode 100644
index 0000000000..eb58674f22
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solveFriction.h
@@ -0,0 +1,421 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* solveFrictionCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"//#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile global int*\n"
+"#endif\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"#define mymake_float4 (float4)\n"
+"//#define make_float2 (float2)\n"
+"//#define make_uint4 (uint4)\n"
+"//#define make_int4 (int4)\n"
+"//#define make_uint2 (uint2)\n"
+"//#define make_int2 (int2)\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"///////////////////////////////////////\n"
+"//	Vector\n"
+"///////////////////////////////////////\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = mymake_float4(a.xyz,0.f);\n"
+"	float4 b1 = mymake_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = mymake_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"//	float length = sqrtf(dot3F4(a, a));\n"
+"//	return 1.f/length * a;\n"
+"}\n"
+"///////////////////////////////////////\n"
+"//	Matrix3x3\n"
+"///////////////////////////////////////\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_row[3];\n"
+"}Matrix3x3;\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b);\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b);\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b)\n"
+"{\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a.m_row[0], b );\n"
+"	ans.y = dot3F4( a.m_row[1], b );\n"
+"	ans.z = dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b)\n"
+"{\n"
+"	float4 colx = mymake_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	float4 coly = mymake_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	float4 colz = mymake_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a, colx );\n"
+"	ans.y = dot3F4( a, coly );\n"
+"	ans.z = dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"typedef float4 Quaternion;\n"
+"#define WG_SIZE 64\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	Quaternion m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"	u32 m_shapeIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} Body;\n"
+"typedef struct\n"
+"{\n"
+"	Matrix3x3 m_invInertia;\n"
+"	Matrix3x3 m_initInvInertia;\n"
+"} Shape;\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_linear;\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_center;	\n"
+"	float m_jacCoeffInv[4];\n"
+"	float m_b[4];\n"
+"	float m_appliedRambdaDt[4];\n"
+"	float m_fJacCoeffInv[2];	\n"
+"	float m_fAppliedRambdaDt[2];	\n"
+"	u32 m_bodyA;\n"
+"	u32 m_bodyB;\n"
+"	int m_batchIdx;\n"
+"	u32 m_paddings[1];\n"
+"} Constraint4;\n"
+"typedef struct\n"
+"{\n"
+"	int m_nConstraints;\n"
+"	int m_start;\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBuffer;\n"
+"typedef struct\n"
+"{\n"
+"	int m_solveFriction;\n"
+"	int m_maxBatch;	//	long batch really kills the performance\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBufferBatchSolve;\n"
+"void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1);\n"
+"void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)\n"
+"{\n"
+"	*linear = mymake_float4(-n.xyz,0.f);\n"
+"	*angular0 = -cross3(r0, n);\n"
+"	*angular1 = cross3(r1, n);\n"
+"}\n"
+"float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 );\n"
+"float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )\n"
+"{\n"
+"	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);\n"
+"}\n"
+"float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,\n"
+"				   float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1);\n"
+"float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,\n"
+"					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1)\n"
+"{\n"
+"	//	linear0,1 are normlized\n"
+"	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;\n"
+"	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);\n"
+"	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;\n"
+"	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);\n"
+"	return -1.f/(jmj0+jmj1+jmj2+jmj3);\n"
+"}\n"
+"void btPlaneSpace1 (const float4* n, float4* p, float4* q);\n"
+" void btPlaneSpace1 (const float4* n, float4* p, float4* q)\n"
+"{\n"
+"  if (fabs(n[0].z) > 0.70710678f) {\n"
+"    // choose p in y-z plane\n"
+"    float a = n[0].y*n[0].y + n[0].z*n[0].z;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = 0;\n"
+"	p[0].y = -n[0].z*k;\n"
+"	p[0].z = n[0].y*k;\n"
+"    // set q = n x p\n"
+"    q[0].x = a*k;\n"
+"	q[0].y = -n[0].x*p[0].z;\n"
+"	q[0].z = n[0].x*p[0].y;\n"
+"  }\n"
+"  else {\n"
+"    // choose p in x-y plane\n"
+"    float a = n[0].x*n[0].x + n[0].y*n[0].y;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = -n[0].y*k;\n"
+"	p[0].y = n[0].x*k;\n"
+"	p[0].z = 0;\n"
+"    // set q = n x p\n"
+"    q[0].x = -n[0].z*p[0].y;\n"
+"	q[0].y = n[0].z*p[0].x;\n"
+"	q[0].z = a*k;\n"
+"  }\n"
+"}\n"
+"void solveFrictionConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs);\n"
+"void solveFrictionConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs)\n"
+"{\n"
+"	float frictionCoeff = ldsCs[0].m_linear.w;\n"
+"	int aIdx = ldsCs[0].m_bodyA;\n"
+"	int bIdx = ldsCs[0].m_bodyB;\n"
+"	float4 posA = gBodies[aIdx].m_pos;\n"
+"	float4 linVelA = gBodies[aIdx].m_linVel;\n"
+"	float4 angVelA = gBodies[aIdx].m_angVel;\n"
+"	float invMassA = gBodies[aIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;\n"
+"	float4 posB = gBodies[bIdx].m_pos;\n"
+"	float4 linVelB = gBodies[bIdx].m_linVel;\n"
+"	float4 angVelB = gBodies[bIdx].m_angVel;\n"
+"	float invMassB = gBodies[bIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;\n"
+"	\n"
+"	{\n"
+"		float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};\n"
+"		float minRambdaDt[4] = {0.f,0.f,0.f,0.f};\n"
+"		float sum = 0;\n"
+"		for(int j=0; j<4; j++)\n"
+"		{\n"
+"			sum +=ldsCs[0].m_appliedRambdaDt[j];\n"
+"		}\n"
+"		frictionCoeff = 0.7f;\n"
+"		for(int j=0; j<4; j++)\n"
+"		{\n"
+"			maxRambdaDt[j] = frictionCoeff*sum;\n"
+"			minRambdaDt[j] = -maxRambdaDt[j];\n"
+"		}\n"
+"		\n"
+"//		solveFriction( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,\n"
+"//			posB, &linVelB, &angVelB, invMassB, invInertiaB, maxRambdaDt, minRambdaDt );\n"
+"		\n"
+"		\n"
+"		{\n"
+"			\n"
+"			__global Constraint4* cs = ldsCs;\n"
+"			\n"
+"			if( cs->m_fJacCoeffInv[0] == 0 && cs->m_fJacCoeffInv[0] == 0 ) return;\n"
+"			const float4 center = cs->m_center;\n"
+"			\n"
+"			float4 n = -cs->m_linear;\n"
+"			\n"
+"			float4 tangent[2];\n"
+"			btPlaneSpace1(&n,&tangent[0],&tangent[1]);\n"
+"			float4 angular0, angular1, linear;\n"
+"			float4 r0 = center - posA;\n"
+"			float4 r1 = center - posB;\n"
+"			for(int i=0; i<2; i++)\n"
+"			{\n"
+"				setLinearAndAngular( tangent[i], r0, r1, &linear, &angular0, &angular1 );\n"
+"				float rambdaDt = calcRelVel(linear, -linear, angular0, angular1,\n"
+"											linVelA, angVelA, linVelB, angVelB );\n"
+"				rambdaDt *= cs->m_fJacCoeffInv[i];\n"
+"				\n"
+"				{\n"
+"					float prevSum = cs->m_fAppliedRambdaDt[i];\n"
+"					float updated = prevSum;\n"
+"					updated += rambdaDt;\n"
+"					updated = max2( updated, minRambdaDt[i] );\n"
+"					updated = min2( updated, maxRambdaDt[i] );\n"
+"					rambdaDt = updated - prevSum;\n"
+"					cs->m_fAppliedRambdaDt[i] = updated;\n"
+"				}\n"
+"				\n"
+"				float4 linImp0 = invMassA*linear*rambdaDt;\n"
+"				float4 linImp1 = invMassB*(-linear)*rambdaDt;\n"
+"				float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;\n"
+"				float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;\n"
+"				\n"
+"				linVelA += linImp0;\n"
+"				angVelA += angImp0;\n"
+"				linVelB += linImp1;\n"
+"				angVelB += angImp1;\n"
+"			}\n"
+"			{	//	angular damping for point constraint\n"
+"				float4 ab = normalize3( posB - posA );\n"
+"				float4 ac = normalize3( center - posA );\n"
+"				if( dot3F4( ab, ac ) > 0.95f  || (invMassA == 0.f || invMassB == 0.f))\n"
+"				{\n"
+"					float angNA = dot3F4( n, angVelA );\n"
+"					float angNB = dot3F4( n, angVelB );\n"
+"					\n"
+"					angVelA -= (angNA*0.1f)*n;\n"
+"					angVelB -= (angNB*0.1f)*n;\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"		\n"
+"		\n"
+"	}\n"
+"	if (gBodies[aIdx].m_invMass)\n"
+"	{\n"
+"		gBodies[aIdx].m_linVel = linVelA;\n"
+"		gBodies[aIdx].m_angVel = angVelA;\n"
+"	} else\n"
+"	{\n"
+"		gBodies[aIdx].m_linVel = mymake_float4(0,0,0,0);\n"
+"		gBodies[aIdx].m_angVel = mymake_float4(0,0,0,0);\n"
+"	}\n"
+"	if (gBodies[bIdx].m_invMass)\n"
+"	{\n"
+"		gBodies[bIdx].m_linVel = linVelB;\n"
+"		gBodies[bIdx].m_angVel = angVelB;\n"
+"	} else\n"
+"	{\n"
+"		gBodies[bIdx].m_linVel = mymake_float4(0,0,0,0);\n"
+"		gBodies[bIdx].m_angVel = mymake_float4(0,0,0,0);\n"
+"	}\n"
+" \n"
+"}\n"
+"typedef struct \n"
+"{\n"
+"	int m_valInt0;\n"
+"	int m_valInt1;\n"
+"	int m_valInt2;\n"
+"	int m_valInt3;\n"
+"	float m_val0;\n"
+"	float m_val1;\n"
+"	float m_val2;\n"
+"	float m_val3;\n"
+"} SolverDebugInfo;\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void BatchSolveKernelFriction(__global Body* gBodies,\n"
+"                      __global Shape* gShapes,\n"
+"                      __global Constraint4* gConstraints,\n"
+"                      __global int* gN,\n"
+"                      __global int* gOffsets,\n"
+"                      __global int* batchSizes,\n"
+"                       int maxBatch1,\n"
+"                       int cellBatch,\n"
+"                       int4 nSplit\n"
+"                      )\n"
+"{\n"
+"	//__local int ldsBatchIdx[WG_SIZE+1];\n"
+"	__local int ldsCurBatch;\n"
+"	__local int ldsNextBatch;\n"
+"	__local int ldsStart;\n"
+"	int lIdx = GET_LOCAL_IDX;\n"
+"	int wgIdx = GET_GROUP_IDX;\n"
+"//	int gIdx = GET_GLOBAL_IDX;\n"
+"//	debugInfo[gIdx].m_valInt0 = gIdx;\n"
+"	//debugInfo[gIdx].m_valInt1 = GET_GROUP_SIZE;\n"
+"	int zIdx = (wgIdx/((nSplit.x*nSplit.y)/4))*2+((cellBatch&4)>>2);\n"
+"	int remain= (wgIdx%((nSplit.x*nSplit.y)/4));\n"
+"	int yIdx = (remain/(nSplit.x/2))*2 + ((cellBatch&2)>>1);\n"
+"	int xIdx = (remain%(nSplit.x/2))*2 + (cellBatch&1);\n"
+"	int cellIdx = xIdx+yIdx*nSplit.x+zIdx*(nSplit.x*nSplit.y);\n"
+"	\n"
+"	if( gN[cellIdx] == 0 ) \n"
+"		return;\n"
+"	int maxBatch = batchSizes[cellIdx];\n"
+"	const int start = gOffsets[cellIdx];\n"
+"	const int end = start + gN[cellIdx];\n"
+"	\n"
+"	if( lIdx == 0 )\n"
+"	{\n"
+"		ldsCurBatch = 0;\n"
+"		ldsNextBatch = 0;\n"
+"		ldsStart = start;\n"
+"	}\n"
+"	GROUP_LDS_BARRIER;\n"
+"	int idx=ldsStart+lIdx;\n"
+"	while (ldsCurBatch < maxBatch)\n"
+"	{\n"
+"		for(; idx<end; )\n"
+"		{\n"
+"			if (gConstraints[idx].m_batchIdx == ldsCurBatch)\n"
+"			{\n"
+"					solveFrictionConstraint( gBodies, gShapes, &gConstraints[idx] );\n"
+"				 idx+=64;\n"
+"			} else\n"
+"			{\n"
+"				break;\n"
+"			}\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"		if( lIdx == 0 )\n"
+"		{\n"
+"			ldsCurBatch++;\n"
+"		}\n"
+"		GROUP_LDS_BARRIER;\n"
+"	}\n"
+"	\n"
+"    \n"
+"}\n"
+"__kernel void solveSingleFrictionKernel(__global Body* gBodies,\n"
+"                      __global Shape* gShapes,\n"
+"                      __global Constraint4* gConstraints,\n"
+"                       int cellIdx,\n"
+"                       int batchOffset,\n"
+"                       int numConstraintsInBatch\n"
+"                      )\n"
+"{\n"
+"	int index = get_global_id(0);\n"
+"	if (index < numConstraintsInBatch)\n"
+"	{\n"
+"		\n"
+"		int idx=batchOffset+index;\n"
+"	\n"
+"		solveFrictionConstraint( gBodies, gShapes, &gConstraints[idx] );\n"
+"	}    \n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverSetup.cl b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverSetup.cl
new file mode 100644
index 0000000000..8e2de7b5a6
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverSetup.cl
@@ -0,0 +1,277 @@
+
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+#include "Bullet3Dynamics/shared/b3ConvertConstraint4.h"
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile global int*
+#endif
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define make_float4 (float4)
+#define make_float2 (float2)
+#define make_uint4 (uint4)
+#define make_int4 (int4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+///////////////////////////////////////
+//	Vector
+///////////////////////////////////////
+__inline
+float fastDiv(float numerator, float denominator)
+{
+	return native_divide(numerator, denominator);	
+//	return numerator/denominator;	
+}
+
+__inline
+float4 fastDiv4(float4 numerator, float4 denominator)
+{
+	return native_divide(numerator, denominator);	
+}
+
+__inline
+float fastSqrtf(float f2)
+{
+	return native_sqrt(f2);
+//	return sqrt(f2);
+}
+
+__inline
+float fastRSqrt(float f2)
+{
+	return native_rsqrt(f2);
+}
+
+__inline
+float fastLength4(float4 v)
+{
+	return fast_length(v);
+}
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	return fast_normalize(v);
+}
+
+
+__inline
+float sqrtf(float a)
+{
+//	return sqrt(a);
+	return native_sqrt(a);
+}
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = make_float4(a.xyz,0.f);
+	float4 b1 = make_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+__inline
+float length3(const float4 a)
+{
+	return sqrtf(dot3F4(a,a));
+}
+
+__inline
+float dot4(const float4 a, const float4 b)
+{
+	return dot( a, b );
+}
+
+//	for height
+__inline
+float dot3w1(const float4 point, const float4 eqn)
+{
+	return dot3F4(point,eqn) + eqn.w;
+}
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = make_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+//	float length = sqrtf(dot3F4(a, a));
+//	return 1.f/length * a;
+}
+
+__inline
+float4 normalize4(const float4 a)
+{
+	float length = sqrtf(dot4(a, a));
+	return 1.f/length * a;
+}
+
+__inline
+float4 createEquation(const float4 a, const float4 b, const float4 c)
+{
+	float4 eqn;
+	float4 ab = b-a;
+	float4 ac = c-a;
+	eqn = normalize3( cross3(ab, ac) );
+	eqn.w = -dot3F4(eqn,a);
+	return eqn;
+}
+
+
+
+#define WG_SIZE 64
+
+
+
+
+
+
+
+typedef struct
+{
+	int m_nConstraints;
+	int m_start;
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBuffer;
+
+typedef struct
+{
+	int m_solveFriction;
+	int m_maxBatch;	//	long batch really kills the performance
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBufferBatchSolve;
+
+
+
+
+ 
+
+
+typedef struct 
+{
+	int m_valInt0;
+	int m_valInt1;
+	int m_valInt2;
+	int m_valInt3;
+
+	float m_val0;
+	float m_val1;
+	float m_val2;
+	float m_val3;
+} SolverDebugInfo;
+
+
+
+
+
+
+typedef struct
+{
+	int m_nContacts;
+	float m_dt;
+	float m_positionDrift;
+	float m_positionConstraintCoeff;
+} ConstBufferCTC;
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void ContactToConstraintKernel(__global struct b3Contact4Data* gContact, __global b3RigidBodyData_t* gBodies, __global b3InertiaData_t* gShapes, __global b3ContactConstraint4_t* gConstraintOut, 
+int nContacts,
+float dt,
+float positionDrift,
+float positionConstraintCoeff
+)
+{
+	int gIdx = GET_GLOBAL_IDX;
+	
+	if( gIdx < nContacts )
+	{
+		int aIdx = abs(gContact[gIdx].m_bodyAPtrAndSignBit);
+		int bIdx = abs(gContact[gIdx].m_bodyBPtrAndSignBit);
+
+		float4 posA = gBodies[aIdx].m_pos;
+		float4 linVelA = gBodies[aIdx].m_linVel;
+		float4 angVelA = gBodies[aIdx].m_angVel;
+		float invMassA = gBodies[aIdx].m_invMass;
+		b3Mat3x3 invInertiaA = gShapes[aIdx].m_initInvInertia;
+
+		float4 posB = gBodies[bIdx].m_pos;
+		float4 linVelB = gBodies[bIdx].m_linVel;
+		float4 angVelB = gBodies[bIdx].m_angVel;
+		float invMassB = gBodies[bIdx].m_invMass;
+		b3Mat3x3 invInertiaB = gShapes[bIdx].m_initInvInertia;
+
+		b3ContactConstraint4_t cs;
+
+    	setConstraint4( posA, linVelA, angVelA, invMassA, invInertiaA, posB, linVelB, angVelB, invMassB, invInertiaB,
+			&gContact[gIdx], dt, positionDrift, positionConstraintCoeff,
+			&cs );
+		
+		cs.m_batchIdx = gContact[gIdx].m_batchIdx;
+
+		gConstraintOut[gIdx] = cs;
+	}
+}
+
+
+
+
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverSetup.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverSetup.h
new file mode 100644
index 0000000000..eb1834ee00
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverSetup.h
@@ -0,0 +1,703 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* solverSetupCL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"#ifndef B3_CONTACT4DATA_H\n"
+"#define B3_CONTACT4DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#define B3_FLOAT4_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#define B3_PLATFORM_DEFINITIONS_H\n"
+"struct MyTest\n"
+"{\n"
+"	int bla;\n"
+"};\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"//keep B3_LARGE_FLOAT*B3_LARGE_FLOAT < FLT_MAX\n"
+"#define B3_LARGE_FLOAT 1e18f\n"
+"#define B3_INFINITY 1e18f\n"
+"#define b3Assert(a)\n"
+"#define b3ConstArray(a) __global const a*\n"
+"#define b3AtomicInc atomic_inc\n"
+"#define b3AtomicAdd atomic_add\n"
+"#define b3Fabs fabs\n"
+"#define b3Sqrt native_sqrt\n"
+"#define b3Sin native_sin\n"
+"#define b3Cos native_cos\n"
+"#define B3_STATIC\n"
+"#endif\n"
+"#endif\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Float4;\n"
+"	#define b3Float4ConstArg const b3Float4\n"
+"	#define b3MakeFloat4 (float4)\n"
+"	float b3Dot3F4(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return dot(a1, b1);\n"
+"	}\n"
+"	b3Float4 b3Cross3(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return cross(a1, b1);\n"
+"	}\n"
+"	#define b3MinFloat4 min\n"
+"	#define b3MaxFloat4 max\n"
+"	#define b3Normalized(a) normalize(a)\n"
+"#endif \n"
+"		\n"
+"inline bool b3IsAlmostZero(b3Float4ConstArg v)\n"
+"{\n"
+"	if(b3Fabs(v.x)>1e-6 || b3Fabs(v.y)>1e-6 || b3Fabs(v.z)>1e-6)	\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"inline int    b3MaxDot( b3Float4ConstArg vec, __global const b3Float4* vecArray, int vecLen, float* dotOut )\n"
+"{\n"
+"    float maxDot = -B3_INFINITY;\n"
+"    int i = 0;\n"
+"    int ptIndex = -1;\n"
+"    for( i = 0; i < vecLen; i++ )\n"
+"    {\n"
+"        float dot = b3Dot3F4(vecArray[i],vec);\n"
+"            \n"
+"        if( dot > maxDot )\n"
+"        {\n"
+"            maxDot = dot;\n"
+"            ptIndex = i;\n"
+"        }\n"
+"    }\n"
+"	b3Assert(ptIndex>=0);\n"
+"    if (ptIndex<0)\n"
+"	{\n"
+"		ptIndex = 0;\n"
+"	}\n"
+"    *dotOut = maxDot;\n"
+"    return ptIndex;\n"
+"}\n"
+"#endif //B3_FLOAT4_H\n"
+"typedef  struct b3Contact4Data b3Contact4Data_t;\n"
+"struct b3Contact4Data\n"
+"{\n"
+"	b3Float4	m_worldPosB[4];\n"
+"//	b3Float4	m_localPosA[4];\n"
+"//	b3Float4	m_localPosB[4];\n"
+"	b3Float4	m_worldNormalOnB;	//	w: m_nPoints\n"
+"	unsigned short  m_restituitionCoeffCmp;\n"
+"	unsigned short  m_frictionCoeffCmp;\n"
+"	int m_batchIdx;\n"
+"	int m_bodyAPtrAndSignBit;//x:m_bodyAPtr, y:m_bodyBPtr\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"	int	m_childIndexA;\n"
+"	int	m_childIndexB;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"};\n"
+"inline int b3Contact4Data_getNumPoints(const struct b3Contact4Data* contact)\n"
+"{\n"
+"	return (int)contact->m_worldNormalOnB.w;\n"
+"};\n"
+"inline void b3Contact4Data_setNumPoints(struct b3Contact4Data* contact, int numPoints)\n"
+"{\n"
+"	contact->m_worldNormalOnB.w = (float)numPoints;\n"
+"};\n"
+"#endif //B3_CONTACT4DATA_H\n"
+"#ifndef B3_CONTACT_CONSTRAINT5_H\n"
+"#define B3_CONTACT_CONSTRAINT5_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"typedef struct b3ContactConstraint4 b3ContactConstraint4_t;\n"
+"struct b3ContactConstraint4\n"
+"{\n"
+"	b3Float4 m_linear;//normal?\n"
+"	b3Float4 m_worldPos[4];\n"
+"	b3Float4 m_center;	//	friction\n"
+"	float m_jacCoeffInv[4];\n"
+"	float m_b[4];\n"
+"	float m_appliedRambdaDt[4];\n"
+"	float m_fJacCoeffInv[2];	//	friction\n"
+"	float m_fAppliedRambdaDt[2];	//	friction\n"
+"	unsigned int m_bodyA;\n"
+"	unsigned int m_bodyB;\n"
+"	int			m_batchIdx;\n"
+"	unsigned int m_paddings;\n"
+"};\n"
+"//inline	void setFrictionCoeff(float value) { m_linear[3] = value; }\n"
+"inline	float b3GetFrictionCoeff(b3ContactConstraint4_t* constraint) \n"
+"{\n"
+"	return constraint->m_linear.w; \n"
+"}\n"
+"#endif //B3_CONTACT_CONSTRAINT5_H\n"
+"#ifndef B3_RIGIDBODY_DATA_H\n"
+"#define B3_RIGIDBODY_DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_QUAT_H\n"
+"#define B3_QUAT_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif\n"
+"#endif\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Quat;\n"
+"	#define b3QuatConstArg const b3Quat\n"
+"	\n"
+"	\n"
+"inline float4 b3FastNormalize4(float4 v)\n"
+"{\n"
+"	v = (float4)(v.xyz,0.f);\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"	\n"
+"inline b3Quat b3QuatMul(b3Quat a, b3Quat b);\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in);\n"
+"inline b3Quat b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec);\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatMul(b3QuatConstArg a, b3QuatConstArg b)\n"
+"{\n"
+"	b3Quat ans;\n"
+"	ans = b3Cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - b3Dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in)\n"
+"{\n"
+"	b3Quat q;\n"
+"	q=in;\n"
+"	//return b3FastNormalize4(in);\n"
+"	float len = native_sqrt(dot(q, q));\n"
+"	if(len > 0.f)\n"
+"	{\n"
+"		q *= 1.f / len;\n"
+"	}\n"
+"	else\n"
+"	{\n"
+"		q.x = q.y = q.z = 0.f;\n"
+"		q.w = 1.f;\n"
+"	}\n"
+"	return q;\n"
+"}\n"
+"inline float4 b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	b3Quat qInv = b3QuatInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = b3QuatMul(b3QuatMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline float4 b3QuatInvRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	return b3QuatRotate( b3QuatInvert( q ), vec );\n"
+"}\n"
+"inline b3Float4 b3TransformPoint(b3Float4ConstArg point, b3Float4ConstArg translation, b3QuatConstArg  orientation)\n"
+"{\n"
+"	return b3QuatRotate( orientation, point ) + (translation);\n"
+"}\n"
+"	\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"#ifndef B3_MAT3x3_H\n"
+"#define B3_MAT3x3_H\n"
+"#ifndef B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"typedef struct\n"
+"{\n"
+"	b3Float4 m_row[3];\n"
+"}b3Mat3x3;\n"
+"#define b3Mat3x3ConstArg const b3Mat3x3\n"
+"#define b3GetRow(m,row) (m.m_row[row])\n"
+"inline b3Mat3x3 b3QuatGetRotationMatrix(b3Quat quat)\n"
+"{\n"
+"	b3Float4 quat2 = (b3Float4)(quat.x*quat.x, quat.y*quat.y, quat.z*quat.z, 0.f);\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0].x=1-2*quat2.y-2*quat2.z;\n"
+"	out.m_row[0].y=2*quat.x*quat.y-2*quat.w*quat.z;\n"
+"	out.m_row[0].z=2*quat.x*quat.z+2*quat.w*quat.y;\n"
+"	out.m_row[0].w = 0.f;\n"
+"	out.m_row[1].x=2*quat.x*quat.y+2*quat.w*quat.z;\n"
+"	out.m_row[1].y=1-2*quat2.x-2*quat2.z;\n"
+"	out.m_row[1].z=2*quat.y*quat.z-2*quat.w*quat.x;\n"
+"	out.m_row[1].w = 0.f;\n"
+"	out.m_row[2].x=2*quat.x*quat.z-2*quat.w*quat.y;\n"
+"	out.m_row[2].y=2*quat.y*quat.z+2*quat.w*quat.x;\n"
+"	out.m_row[2].z=1-2*quat2.x-2*quat2.y;\n"
+"	out.m_row[2].w = 0.f;\n"
+"	return out;\n"
+"}\n"
+"inline b3Mat3x3 b3AbsoluteMat3x3(b3Mat3x3ConstArg matIn)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = fabs(matIn.m_row[0]);\n"
+"	out.m_row[1] = fabs(matIn.m_row[1]);\n"
+"	out.m_row[2] = fabs(matIn.m_row[2]);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtZero();\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity();\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m);\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b);\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Mat3x3 mtZero()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(0.f);\n"
+"	m.m_row[1] = (b3Float4)(0.f);\n"
+"	m.m_row[2] = (b3Float4)(0.f);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(1,0,0,0);\n"
+"	m.m_row[1] = (b3Float4)(0,1,0,0);\n"
+"	m.m_row[2] = (b3Float4)(0,0,1,0);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = (b3Float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (b3Float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (b3Float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Mat3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	b3Mat3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = b3Dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = b3Dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = b3Dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b)\n"
+"{\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a.m_row[0], b );\n"
+"	ans.y = b3Dot3F4( a.m_row[1], b );\n"
+"	ans.z = b3Dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Float4 colx = b3MakeFloat4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	b3Float4 coly = b3MakeFloat4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	b3Float4 colz = b3MakeFloat4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a, colx );\n"
+"	ans.y = b3Dot3F4( a, coly );\n"
+"	ans.z = b3Dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"#endif\n"
+"#endif //B3_MAT3x3_H\n"
+"typedef struct b3RigidBodyData b3RigidBodyData_t;\n"
+"struct b3RigidBodyData\n"
+"{\n"
+"	b3Float4				m_pos;\n"
+"	b3Quat					m_quat;\n"
+"	b3Float4				m_linVel;\n"
+"	b3Float4				m_angVel;\n"
+"	int 					m_collidableIdx;\n"
+"	float 				m_invMass;\n"
+"	float 				m_restituitionCoeff;\n"
+"	float 				m_frictionCoeff;\n"
+"};\n"
+"typedef struct b3InertiaData b3InertiaData_t;\n"
+"struct b3InertiaData\n"
+"{\n"
+"	b3Mat3x3 m_invInertiaWorld;\n"
+"	b3Mat3x3 m_initInvInertia;\n"
+"};\n"
+"#endif //B3_RIGIDBODY_DATA_H\n"
+"	\n"
+"void b3PlaneSpace1 (b3Float4ConstArg n, b3Float4* p, b3Float4* q);\n"
+" void b3PlaneSpace1 (b3Float4ConstArg n, b3Float4* p, b3Float4* q)\n"
+"{\n"
+"  if (b3Fabs(n.z) > 0.70710678f) {\n"
+"    // choose p in y-z plane\n"
+"    float a = n.y*n.y + n.z*n.z;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = 0;\n"
+"	p[0].y = -n.z*k;\n"
+"	p[0].z = n.y*k;\n"
+"    // set q = n x p\n"
+"    q[0].x = a*k;\n"
+"	q[0].y = -n.x*p[0].z;\n"
+"	q[0].z = n.x*p[0].y;\n"
+"  }\n"
+"  else {\n"
+"    // choose p in x-y plane\n"
+"    float a = n.x*n.x + n.y*n.y;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = -n.y*k;\n"
+"	p[0].y = n.x*k;\n"
+"	p[0].z = 0;\n"
+"    // set q = n x p\n"
+"    q[0].x = -n.z*p[0].y;\n"
+"	q[0].y = n.z*p[0].x;\n"
+"	q[0].z = a*k;\n"
+"  }\n"
+"}\n"
+" \n"
+"void setLinearAndAngular( b3Float4ConstArg n, b3Float4ConstArg r0, b3Float4ConstArg r1, b3Float4* linear, b3Float4* angular0, b3Float4* angular1)\n"
+"{\n"
+"	*linear = b3MakeFloat4(n.x,n.y,n.z,0.f);\n"
+"	*angular0 = b3Cross3(r0, n);\n"
+"	*angular1 = -b3Cross3(r1, n);\n"
+"}\n"
+"float calcRelVel( b3Float4ConstArg l0, b3Float4ConstArg l1, b3Float4ConstArg a0, b3Float4ConstArg a1, b3Float4ConstArg linVel0,\n"
+"	b3Float4ConstArg angVel0, b3Float4ConstArg linVel1, b3Float4ConstArg angVel1 )\n"
+"{\n"
+"	return b3Dot3F4(l0, linVel0) + b3Dot3F4(a0, angVel0) + b3Dot3F4(l1, linVel1) + b3Dot3F4(a1, angVel1);\n"
+"}\n"
+"float calcJacCoeff(b3Float4ConstArg linear0, b3Float4ConstArg linear1, b3Float4ConstArg angular0, b3Float4ConstArg angular1,\n"
+"					float invMass0, const b3Mat3x3* invInertia0, float invMass1, const b3Mat3x3* invInertia1)\n"
+"{\n"
+"	//	linear0,1 are normlized\n"
+"	float jmj0 = invMass0;//b3Dot3F4(linear0, linear0)*invMass0;\n"
+"	float jmj1 = b3Dot3F4(mtMul3(angular0,*invInertia0), angular0);\n"
+"	float jmj2 = invMass1;//b3Dot3F4(linear1, linear1)*invMass1;\n"
+"	float jmj3 = b3Dot3F4(mtMul3(angular1,*invInertia1), angular1);\n"
+"	return -1.f/(jmj0+jmj1+jmj2+jmj3);\n"
+"}\n"
+"void setConstraint4( b3Float4ConstArg posA, b3Float4ConstArg linVelA, b3Float4ConstArg angVelA, float invMassA, b3Mat3x3ConstArg invInertiaA,\n"
+"	b3Float4ConstArg posB, b3Float4ConstArg linVelB, b3Float4ConstArg angVelB, float invMassB, b3Mat3x3ConstArg invInertiaB, \n"
+"	__global struct b3Contact4Data* src, float dt, float positionDrift, float positionConstraintCoeff,\n"
+"	b3ContactConstraint4_t* dstC )\n"
+"{\n"
+"	dstC->m_bodyA = abs(src->m_bodyAPtrAndSignBit);\n"
+"	dstC->m_bodyB = abs(src->m_bodyBPtrAndSignBit);\n"
+"	float dtInv = 1.f/dt;\n"
+"	for(int ic=0; ic<4; ic++)\n"
+"	{\n"
+"		dstC->m_appliedRambdaDt[ic] = 0.f;\n"
+"	}\n"
+"	dstC->m_fJacCoeffInv[0] = dstC->m_fJacCoeffInv[1] = 0.f;\n"
+"	dstC->m_linear = src->m_worldNormalOnB;\n"
+"	dstC->m_linear.w = 0.7f ;//src->getFrictionCoeff() );\n"
+"	for(int ic=0; ic<4; ic++)\n"
+"	{\n"
+"		b3Float4 r0 = src->m_worldPosB[ic] - posA;\n"
+"		b3Float4 r1 = src->m_worldPosB[ic] - posB;\n"
+"		if( ic >= src->m_worldNormalOnB.w )//npoints\n"
+"		{\n"
+"			dstC->m_jacCoeffInv[ic] = 0.f;\n"
+"			continue;\n"
+"		}\n"
+"		float relVelN;\n"
+"		{\n"
+"			b3Float4 linear, angular0, angular1;\n"
+"			setLinearAndAngular(src->m_worldNormalOnB, r0, r1, &linear, &angular0, &angular1);\n"
+"			dstC->m_jacCoeffInv[ic] = calcJacCoeff(linear, -linear, angular0, angular1,\n"
+"				invMassA, &invInertiaA, invMassB, &invInertiaB );\n"
+"			relVelN = calcRelVel(linear, -linear, angular0, angular1,\n"
+"				linVelA, angVelA, linVelB, angVelB);\n"
+"			float e = 0.f;//src->getRestituitionCoeff();\n"
+"			if( relVelN*relVelN < 0.004f ) e = 0.f;\n"
+"			dstC->m_b[ic] = e*relVelN;\n"
+"			//float penetration = src->m_worldPosB[ic].w;\n"
+"			dstC->m_b[ic] += (src->m_worldPosB[ic].w + positionDrift)*positionConstraintCoeff*dtInv;\n"
+"			dstC->m_appliedRambdaDt[ic] = 0.f;\n"
+"		}\n"
+"	}\n"
+"	if( src->m_worldNormalOnB.w > 0 )//npoints\n"
+"	{	//	prepare friction\n"
+"		b3Float4 center = b3MakeFloat4(0.f,0.f,0.f,0.f);\n"
+"		for(int i=0; i<src->m_worldNormalOnB.w; i++) \n"
+"			center += src->m_worldPosB[i];\n"
+"		center /= (float)src->m_worldNormalOnB.w;\n"
+"		b3Float4 tangent[2];\n"
+"		b3PlaneSpace1(src->m_worldNormalOnB,&tangent[0],&tangent[1]);\n"
+"		\n"
+"		b3Float4 r[2];\n"
+"		r[0] = center - posA;\n"
+"		r[1] = center - posB;\n"
+"		for(int i=0; i<2; i++)\n"
+"		{\n"
+"			b3Float4 linear, angular0, angular1;\n"
+"			setLinearAndAngular(tangent[i], r[0], r[1], &linear, &angular0, &angular1);\n"
+"			dstC->m_fJacCoeffInv[i] = calcJacCoeff(linear, -linear, angular0, angular1,\n"
+"				invMassA, &invInertiaA, invMassB, &invInertiaB );\n"
+"			dstC->m_fAppliedRambdaDt[i] = 0.f;\n"
+"		}\n"
+"		dstC->m_center = center;\n"
+"	}\n"
+"	for(int i=0; i<4; i++)\n"
+"	{\n"
+"		if( i<src->m_worldNormalOnB.w )\n"
+"		{\n"
+"			dstC->m_worldPos[i] = src->m_worldPosB[i];\n"
+"		}\n"
+"		else\n"
+"		{\n"
+"			dstC->m_worldPos[i] = b3MakeFloat4(0.f,0.f,0.f,0.f);\n"
+"		}\n"
+"	}\n"
+"}\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile global int*\n"
+"#endif\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"#define make_float4 (float4)\n"
+"#define make_float2 (float2)\n"
+"#define make_uint4 (uint4)\n"
+"#define make_int4 (int4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"///////////////////////////////////////\n"
+"//	Vector\n"
+"///////////////////////////////////////\n"
+"__inline\n"
+"float fastDiv(float numerator, float denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"//	return numerator/denominator;	\n"
+"}\n"
+"__inline\n"
+"float4 fastDiv4(float4 numerator, float4 denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"}\n"
+"__inline\n"
+"float fastSqrtf(float f2)\n"
+"{\n"
+"	return native_sqrt(f2);\n"
+"//	return sqrt(f2);\n"
+"}\n"
+"__inline\n"
+"float fastRSqrt(float f2)\n"
+"{\n"
+"	return native_rsqrt(f2);\n"
+"}\n"
+"__inline\n"
+"float fastLength4(float4 v)\n"
+"{\n"
+"	return fast_length(v);\n"
+"}\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"__inline\n"
+"float sqrtf(float a)\n"
+"{\n"
+"//	return sqrt(a);\n"
+"	return native_sqrt(a);\n"
+"}\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = make_float4(a.xyz,0.f);\n"
+"	float4 b1 = make_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"__inline\n"
+"float length3(const float4 a)\n"
+"{\n"
+"	return sqrtf(dot3F4(a,a));\n"
+"}\n"
+"__inline\n"
+"float dot4(const float4 a, const float4 b)\n"
+"{\n"
+"	return dot( a, b );\n"
+"}\n"
+"//	for height\n"
+"__inline\n"
+"float dot3w1(const float4 point, const float4 eqn)\n"
+"{\n"
+"	return dot3F4(point,eqn) + eqn.w;\n"
+"}\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = make_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"//	float length = sqrtf(dot3F4(a, a));\n"
+"//	return 1.f/length * a;\n"
+"}\n"
+"__inline\n"
+"float4 normalize4(const float4 a)\n"
+"{\n"
+"	float length = sqrtf(dot4(a, a));\n"
+"	return 1.f/length * a;\n"
+"}\n"
+"__inline\n"
+"float4 createEquation(const float4 a, const float4 b, const float4 c)\n"
+"{\n"
+"	float4 eqn;\n"
+"	float4 ab = b-a;\n"
+"	float4 ac = c-a;\n"
+"	eqn = normalize3( cross3(ab, ac) );\n"
+"	eqn.w = -dot3F4(eqn,a);\n"
+"	return eqn;\n"
+"}\n"
+"#define WG_SIZE 64\n"
+"typedef struct\n"
+"{\n"
+"	int m_nConstraints;\n"
+"	int m_start;\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBuffer;\n"
+"typedef struct\n"
+"{\n"
+"	int m_solveFriction;\n"
+"	int m_maxBatch;	//	long batch really kills the performance\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBufferBatchSolve;\n"
+" \n"
+"typedef struct \n"
+"{\n"
+"	int m_valInt0;\n"
+"	int m_valInt1;\n"
+"	int m_valInt2;\n"
+"	int m_valInt3;\n"
+"	float m_val0;\n"
+"	float m_val1;\n"
+"	float m_val2;\n"
+"	float m_val3;\n"
+"} SolverDebugInfo;\n"
+"typedef struct\n"
+"{\n"
+"	int m_nContacts;\n"
+"	float m_dt;\n"
+"	float m_positionDrift;\n"
+"	float m_positionConstraintCoeff;\n"
+"} ConstBufferCTC;\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void ContactToConstraintKernel(__global struct b3Contact4Data* gContact, __global b3RigidBodyData_t* gBodies, __global b3InertiaData_t* gShapes, __global b3ContactConstraint4_t* gConstraintOut, \n"
+"int nContacts,\n"
+"float dt,\n"
+"float positionDrift,\n"
+"float positionConstraintCoeff\n"
+")\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	\n"
+"	if( gIdx < nContacts )\n"
+"	{\n"
+"		int aIdx = abs(gContact[gIdx].m_bodyAPtrAndSignBit);\n"
+"		int bIdx = abs(gContact[gIdx].m_bodyBPtrAndSignBit);\n"
+"		float4 posA = gBodies[aIdx].m_pos;\n"
+"		float4 linVelA = gBodies[aIdx].m_linVel;\n"
+"		float4 angVelA = gBodies[aIdx].m_angVel;\n"
+"		float invMassA = gBodies[aIdx].m_invMass;\n"
+"		b3Mat3x3 invInertiaA = gShapes[aIdx].m_initInvInertia;\n"
+"		float4 posB = gBodies[bIdx].m_pos;\n"
+"		float4 linVelB = gBodies[bIdx].m_linVel;\n"
+"		float4 angVelB = gBodies[bIdx].m_angVel;\n"
+"		float invMassB = gBodies[bIdx].m_invMass;\n"
+"		b3Mat3x3 invInertiaB = gShapes[bIdx].m_initInvInertia;\n"
+"		b3ContactConstraint4_t cs;\n"
+"    	setConstraint4( posA, linVelA, angVelA, invMassA, invInertiaA, posB, linVelB, angVelB, invMassB, invInertiaB,\n"
+"			&gContact[gIdx], dt, positionDrift, positionConstraintCoeff,\n"
+"			&cs );\n"
+"		\n"
+"		cs.m_batchIdx = gContact[gIdx].m_batchIdx;\n"
+"		gConstraintOut[gIdx] = cs;\n"
+"	}\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverSetup2.cl b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverSetup2.cl
new file mode 100644
index 0000000000..3dc48d4350
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverSetup2.cl
@@ -0,0 +1,613 @@
+/*
+Copyright (c) 2012 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Takahiro Harada
+
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Contact4Data.h"
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile global int*
+#endif
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define make_float4 (float4)
+#define make_float2 (float2)
+#define make_uint4 (uint4)
+#define make_int4 (int4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+///////////////////////////////////////
+//	Vector
+///////////////////////////////////////
+__inline
+float fastDiv(float numerator, float denominator)
+{
+	return native_divide(numerator, denominator);	
+//	return numerator/denominator;	
+}
+
+__inline
+float4 fastDiv4(float4 numerator, float4 denominator)
+{
+	return native_divide(numerator, denominator);	
+}
+
+__inline
+float fastSqrtf(float f2)
+{
+	return native_sqrt(f2);
+//	return sqrt(f2);
+}
+
+__inline
+float fastRSqrt(float f2)
+{
+	return native_rsqrt(f2);
+}
+
+__inline
+float fastLength4(float4 v)
+{
+	return fast_length(v);
+}
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	return fast_normalize(v);
+}
+
+
+__inline
+float sqrtf(float a)
+{
+//	return sqrt(a);
+	return native_sqrt(a);
+}
+
+__inline
+float4 cross3(float4 a, float4 b)
+{
+	return cross(a,b);
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = make_float4(a.xyz,0.f);
+	float4 b1 = make_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+__inline
+float length3(const float4 a)
+{
+	return sqrtf(dot3F4(a,a));
+}
+
+__inline
+float dot4(const float4 a, const float4 b)
+{
+	return dot( a, b );
+}
+
+//	for height
+__inline
+float dot3w1(const float4 point, const float4 eqn)
+{
+	return dot3F4(point,eqn) + eqn.w;
+}
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = make_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+//	float length = sqrtf(dot3F4(a, a));
+//	return 1.f/length * a;
+}
+
+__inline
+float4 normalize4(const float4 a)
+{
+	float length = sqrtf(dot4(a, a));
+	return 1.f/length * a;
+}
+
+__inline
+float4 createEquation(const float4 a, const float4 b, const float4 c)
+{
+	float4 eqn;
+	float4 ab = b-a;
+	float4 ac = c-a;
+	eqn = normalize3( cross3(ab, ac) );
+	eqn.w = -dot3F4(eqn,a);
+	return eqn;
+}
+
+///////////////////////////////////////
+//	Matrix3x3
+///////////////////////////////////////
+
+typedef struct
+{
+	float4 m_row[3];
+}Matrix3x3;
+
+__inline
+Matrix3x3 mtZero();
+
+__inline
+Matrix3x3 mtIdentity();
+
+__inline
+Matrix3x3 mtTranspose(Matrix3x3 m);
+
+__inline
+Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b);
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b);
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b);
+
+__inline
+Matrix3x3 mtZero()
+{
+	Matrix3x3 m;
+	m.m_row[0] = (float4)(0.f);
+	m.m_row[1] = (float4)(0.f);
+	m.m_row[2] = (float4)(0.f);
+	return m;
+}
+
+__inline
+Matrix3x3 mtIdentity()
+{
+	Matrix3x3 m;
+	m.m_row[0] = (float4)(1,0,0,0);
+	m.m_row[1] = (float4)(0,1,0,0);
+	m.m_row[2] = (float4)(0,0,1,0);
+	return m;
+}
+
+__inline
+Matrix3x3 mtTranspose(Matrix3x3 m)
+{
+	Matrix3x3 out;
+	out.m_row[0] = (float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);
+	out.m_row[1] = (float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);
+	out.m_row[2] = (float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);
+	return out;
+}
+
+__inline
+Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b)
+{
+	Matrix3x3 transB;
+	transB = mtTranspose( b );
+	Matrix3x3 ans;
+	//	why this doesn't run when 0ing in the for{}
+	a.m_row[0].w = 0.f;
+	a.m_row[1].w = 0.f;
+	a.m_row[2].w = 0.f;
+	for(int i=0; i<3; i++)
+	{
+//	a.m_row[i].w = 0.f;
+		ans.m_row[i].x = dot3F4(a.m_row[i],transB.m_row[0]);
+		ans.m_row[i].y = dot3F4(a.m_row[i],transB.m_row[1]);
+		ans.m_row[i].z = dot3F4(a.m_row[i],transB.m_row[2]);
+		ans.m_row[i].w = 0.f;
+	}
+	return ans;
+}
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b)
+{
+	float4 ans;
+	ans.x = dot3F4( a.m_row[0], b );
+	ans.y = dot3F4( a.m_row[1], b );
+	ans.z = dot3F4( a.m_row[2], b );
+	ans.w = 0.f;
+	return ans;
+}
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b)
+{
+	float4 colx = make_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);
+	float4 coly = make_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);
+	float4 colz = make_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);
+
+	float4 ans;
+	ans.x = dot3F4( a, colx );
+	ans.y = dot3F4( a, coly );
+	ans.z = dot3F4( a, colz );
+	return ans;
+}
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b);
+
+__inline
+Quaternion qtNormalize(Quaternion in);
+
+__inline
+float4 qtRotate(Quaternion q, float4 vec);
+
+__inline
+Quaternion qtInvert(Quaternion q);
+
+
+
+
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross3( a, b );
+	ans += a.w*b+b.w*a;
+//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+Quaternion qtNormalize(Quaternion in)
+{
+	return fastNormalize4(in);
+//	in /= length( in );
+//	return in;
+}
+__inline
+float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(qtMul(q,vcpy),qInv);
+	return out;
+}
+
+__inline
+Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+__inline
+float4 qtInvRotate(const Quaternion q, float4 vec)
+{
+	return qtRotate( qtInvert( q ), vec );
+}
+
+
+
+
+#define WG_SIZE 64
+
+typedef struct
+{
+	float4 m_pos;
+	Quaternion m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_shapeIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} Body;
+
+typedef struct
+{
+	Matrix3x3 m_invInertia;
+	Matrix3x3 m_initInvInertia;
+} Shape;
+
+typedef struct
+{
+	float4 m_linear;
+	float4 m_worldPos[4];
+	float4 m_center;	
+	float m_jacCoeffInv[4];
+	float m_b[4];
+	float m_appliedRambdaDt[4];
+
+	float m_fJacCoeffInv[2];	
+	float m_fAppliedRambdaDt[2];	
+
+	u32 m_bodyA;
+	u32 m_bodyB;
+
+	int m_batchIdx;
+	u32 m_paddings[1];
+} Constraint4;
+
+
+
+typedef struct
+{
+	int m_nConstraints;
+	int m_start;
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBuffer;
+
+typedef struct
+{
+	int m_solveFriction;
+	int m_maxBatch;	//	long batch really kills the performance
+	int m_batchIdx;
+	int m_nSplit;
+//	int m_paddings[1];
+} ConstBufferBatchSolve;
+
+
+ 
+
+
+typedef struct 
+{
+	int m_valInt0;
+	int m_valInt1;
+	int m_valInt2;
+	int m_valInt3;
+
+	float m_val0;
+	float m_val1;
+	float m_val2;
+	float m_val3;
+} SolverDebugInfo;
+
+
+
+
+//	others
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void ReorderContactKernel(__global struct b3Contact4Data* in, __global struct b3Contact4Data* out, __global int2* sortData, int4 cb )
+{
+	int nContacts = cb.x;
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < nContacts )
+	{
+		int srcIdx = sortData[gIdx].y;
+		out[gIdx] = in[srcIdx];
+	}
+}
+
+__kernel __attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void SetDeterminismSortDataChildShapeB(__global struct b3Contact4Data* contactsIn, __global int2* sortDataOut, int nContacts)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < nContacts )
+	{
+		int2 sd;
+		sd.x = contactsIn[gIdx].m_childIndexB;
+		sd.y = gIdx;
+		sortDataOut[gIdx] = sd;
+	}
+}
+
+__kernel __attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void SetDeterminismSortDataChildShapeA(__global struct b3Contact4Data* contactsIn, __global int2* sortDataInOut, int nContacts)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < nContacts )
+	{
+		int2 sdIn;
+		sdIn = sortDataInOut[gIdx];
+		int2 sdOut;
+		sdOut.x = contactsIn[sdIn.y].m_childIndexA;
+		sdOut.y = sdIn.y;
+		sortDataInOut[gIdx] = sdOut;
+	}
+}
+
+__kernel __attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void SetDeterminismSortDataBodyA(__global struct b3Contact4Data* contactsIn, __global int2* sortDataInOut, int nContacts)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < nContacts )
+	{
+		int2 sdIn;
+		sdIn = sortDataInOut[gIdx];
+		int2 sdOut;
+		sdOut.x = contactsIn[sdIn.y].m_bodyAPtrAndSignBit;
+		sdOut.y = sdIn.y;
+		sortDataInOut[gIdx] = sdOut;
+	}
+}
+
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void SetDeterminismSortDataBodyB(__global struct b3Contact4Data* contactsIn, __global int2* sortDataInOut, int nContacts)
+{
+	int gIdx = GET_GLOBAL_IDX;
+
+	if( gIdx < nContacts )
+	{
+		int2 sdIn;
+		sdIn = sortDataInOut[gIdx];
+		int2 sdOut;
+		sdOut.x = contactsIn[sdIn.y].m_bodyBPtrAndSignBit;
+		sdOut.y = sdIn.y;
+		sortDataInOut[gIdx] = sdOut;
+	}
+}
+
+
+
+
+typedef struct
+{
+	int m_nContacts;
+	int m_staticIdx;
+	float m_scale;
+	int m_nSplit;
+} ConstBufferSSD;
+
+
+__constant const int gridTable4x4[] = 
+{
+    0,1,17,16,
+	1,2,18,19,
+	17,18,32,3,
+	16,19,3,34
+};
+
+__constant const int gridTable8x8[] = 
+{
+	  0,  2,  3, 16, 17, 18, 19,  1,
+	 66, 64, 80, 67, 82, 81, 65, 83,
+	131,144,128,130,147,129,145,146,
+	208,195,194,192,193,211,210,209,
+	 21, 22, 23,  5,  4,  6,  7, 20,
+	 86, 85, 69, 87, 70, 68, 84, 71,
+	151,133,149,150,135,148,132,134,
+	197,27,214,213,212,199,198,196
+	
+};
+
+
+
+
+#define USE_SPATIAL_BATCHING 1
+#define USE_4x4_GRID 1
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void SetSortDataKernel(__global struct b3Contact4Data* gContact, __global Body* gBodies, __global int2* gSortDataOut, 
+int nContacts,float scale,int4 nSplit,int staticIdx)
+
+{
+	int gIdx = GET_GLOBAL_IDX;
+	
+	if( gIdx < nContacts )
+	{
+		int aPtrAndSignBit  = gContact[gIdx].m_bodyAPtrAndSignBit;
+		int bPtrAndSignBit  = gContact[gIdx].m_bodyBPtrAndSignBit;
+
+		int aIdx = abs(aPtrAndSignBit );
+		int bIdx = abs(bPtrAndSignBit);
+
+		bool aStatic = (aPtrAndSignBit<0) ||(aPtrAndSignBit==staticIdx);
+		bool bStatic = (bPtrAndSignBit<0) ||(bPtrAndSignBit==staticIdx);
+
+#if USE_SPATIAL_BATCHING		
+		int idx = (aStatic)? bIdx: aIdx;
+		float4 p = gBodies[idx].m_pos;
+		int xIdx = (int)((p.x-((p.x<0.f)?1.f:0.f))*scale) & (nSplit.x-1);
+		int yIdx = (int)((p.y-((p.y<0.f)?1.f:0.f))*scale) & (nSplit.y-1);
+		int zIdx = (int)((p.z-((p.z<0.f)?1.f:0.f))*scale) & (nSplit.z-1);
+		int newIndex = (xIdx+yIdx*nSplit.x+zIdx*nSplit.x*nSplit.y);
+		
+#else//USE_SPATIAL_BATCHING
+	#if USE_4x4_GRID
+		int aa = aIdx&3;
+		int bb = bIdx&3;
+		if (aStatic)
+			aa = bb;
+		if (bStatic)
+			bb = aa;
+
+		int gridIndex = aa + bb*4;
+		int newIndex = gridTable4x4[gridIndex];
+	#else//USE_4x4_GRID
+		int aa = aIdx&7;
+		int bb = bIdx&7;
+		if (aStatic)
+			aa = bb;
+		if (bStatic)
+			bb = aa;
+
+		int gridIndex = aa + bb*8;
+		int newIndex = gridTable8x8[gridIndex];
+	#endif//USE_4x4_GRID
+#endif//USE_SPATIAL_BATCHING
+
+
+		gSortDataOut[gIdx].x = newIndex;
+		gSortDataOut[gIdx].y = gIdx;
+	}
+	else
+	{
+		gSortDataOut[gIdx].x = 0xffffffff;
+	}
+}
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void CopyConstraintKernel(__global struct b3Contact4Data* gIn, __global struct b3Contact4Data* gOut, int4 cb )
+{
+	int gIdx = GET_GLOBAL_IDX;
+	if( gIdx < cb.x )
+	{
+		gOut[gIdx] = gIn[gIdx];
+	}
+}
+
+
+
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverSetup2.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverSetup2.h
new file mode 100644
index 0000000000..1b5819f6cf
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverSetup2.h
@@ -0,0 +1,601 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* solverSetup2CL= \
+"/*\n"
+"Copyright (c) 2012 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Takahiro Harada\n"
+"#ifndef B3_CONTACT4DATA_H\n"
+"#define B3_CONTACT4DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#define B3_FLOAT4_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#define B3_PLATFORM_DEFINITIONS_H\n"
+"struct MyTest\n"
+"{\n"
+"	int bla;\n"
+"};\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"//keep B3_LARGE_FLOAT*B3_LARGE_FLOAT < FLT_MAX\n"
+"#define B3_LARGE_FLOAT 1e18f\n"
+"#define B3_INFINITY 1e18f\n"
+"#define b3Assert(a)\n"
+"#define b3ConstArray(a) __global const a*\n"
+"#define b3AtomicInc atomic_inc\n"
+"#define b3AtomicAdd atomic_add\n"
+"#define b3Fabs fabs\n"
+"#define b3Sqrt native_sqrt\n"
+"#define b3Sin native_sin\n"
+"#define b3Cos native_cos\n"
+"#define B3_STATIC\n"
+"#endif\n"
+"#endif\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Float4;\n"
+"	#define b3Float4ConstArg const b3Float4\n"
+"	#define b3MakeFloat4 (float4)\n"
+"	float b3Dot3F4(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return dot(a1, b1);\n"
+"	}\n"
+"	b3Float4 b3Cross3(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return cross(a1, b1);\n"
+"	}\n"
+"	#define b3MinFloat4 min\n"
+"	#define b3MaxFloat4 max\n"
+"	#define b3Normalized(a) normalize(a)\n"
+"#endif \n"
+"		\n"
+"inline bool b3IsAlmostZero(b3Float4ConstArg v)\n"
+"{\n"
+"	if(b3Fabs(v.x)>1e-6 || b3Fabs(v.y)>1e-6 || b3Fabs(v.z)>1e-6)	\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"inline int    b3MaxDot( b3Float4ConstArg vec, __global const b3Float4* vecArray, int vecLen, float* dotOut )\n"
+"{\n"
+"    float maxDot = -B3_INFINITY;\n"
+"    int i = 0;\n"
+"    int ptIndex = -1;\n"
+"    for( i = 0; i < vecLen; i++ )\n"
+"    {\n"
+"        float dot = b3Dot3F4(vecArray[i],vec);\n"
+"            \n"
+"        if( dot > maxDot )\n"
+"        {\n"
+"            maxDot = dot;\n"
+"            ptIndex = i;\n"
+"        }\n"
+"    }\n"
+"	b3Assert(ptIndex>=0);\n"
+"    if (ptIndex<0)\n"
+"	{\n"
+"		ptIndex = 0;\n"
+"	}\n"
+"    *dotOut = maxDot;\n"
+"    return ptIndex;\n"
+"}\n"
+"#endif //B3_FLOAT4_H\n"
+"typedef  struct b3Contact4Data b3Contact4Data_t;\n"
+"struct b3Contact4Data\n"
+"{\n"
+"	b3Float4	m_worldPosB[4];\n"
+"//	b3Float4	m_localPosA[4];\n"
+"//	b3Float4	m_localPosB[4];\n"
+"	b3Float4	m_worldNormalOnB;	//	w: m_nPoints\n"
+"	unsigned short  m_restituitionCoeffCmp;\n"
+"	unsigned short  m_frictionCoeffCmp;\n"
+"	int m_batchIdx;\n"
+"	int m_bodyAPtrAndSignBit;//x:m_bodyAPtr, y:m_bodyBPtr\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"	int	m_childIndexA;\n"
+"	int	m_childIndexB;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"};\n"
+"inline int b3Contact4Data_getNumPoints(const struct b3Contact4Data* contact)\n"
+"{\n"
+"	return (int)contact->m_worldNormalOnB.w;\n"
+"};\n"
+"inline void b3Contact4Data_setNumPoints(struct b3Contact4Data* contact, int numPoints)\n"
+"{\n"
+"	contact->m_worldNormalOnB.w = (float)numPoints;\n"
+"};\n"
+"#endif //B3_CONTACT4DATA_H\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile global int*\n"
+"#endif\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"#define make_float4 (float4)\n"
+"#define make_float2 (float2)\n"
+"#define make_uint4 (uint4)\n"
+"#define make_int4 (int4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"///////////////////////////////////////\n"
+"//	Vector\n"
+"///////////////////////////////////////\n"
+"__inline\n"
+"float fastDiv(float numerator, float denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"//	return numerator/denominator;	\n"
+"}\n"
+"__inline\n"
+"float4 fastDiv4(float4 numerator, float4 denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"}\n"
+"__inline\n"
+"float fastSqrtf(float f2)\n"
+"{\n"
+"	return native_sqrt(f2);\n"
+"//	return sqrt(f2);\n"
+"}\n"
+"__inline\n"
+"float fastRSqrt(float f2)\n"
+"{\n"
+"	return native_rsqrt(f2);\n"
+"}\n"
+"__inline\n"
+"float fastLength4(float4 v)\n"
+"{\n"
+"	return fast_length(v);\n"
+"}\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"__inline\n"
+"float sqrtf(float a)\n"
+"{\n"
+"//	return sqrt(a);\n"
+"	return native_sqrt(a);\n"
+"}\n"
+"__inline\n"
+"float4 cross3(float4 a, float4 b)\n"
+"{\n"
+"	return cross(a,b);\n"
+"}\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = make_float4(a.xyz,0.f);\n"
+"	float4 b1 = make_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"__inline\n"
+"float length3(const float4 a)\n"
+"{\n"
+"	return sqrtf(dot3F4(a,a));\n"
+"}\n"
+"__inline\n"
+"float dot4(const float4 a, const float4 b)\n"
+"{\n"
+"	return dot( a, b );\n"
+"}\n"
+"//	for height\n"
+"__inline\n"
+"float dot3w1(const float4 point, const float4 eqn)\n"
+"{\n"
+"	return dot3F4(point,eqn) + eqn.w;\n"
+"}\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = make_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"//	float length = sqrtf(dot3F4(a, a));\n"
+"//	return 1.f/length * a;\n"
+"}\n"
+"__inline\n"
+"float4 normalize4(const float4 a)\n"
+"{\n"
+"	float length = sqrtf(dot4(a, a));\n"
+"	return 1.f/length * a;\n"
+"}\n"
+"__inline\n"
+"float4 createEquation(const float4 a, const float4 b, const float4 c)\n"
+"{\n"
+"	float4 eqn;\n"
+"	float4 ab = b-a;\n"
+"	float4 ac = c-a;\n"
+"	eqn = normalize3( cross3(ab, ac) );\n"
+"	eqn.w = -dot3F4(eqn,a);\n"
+"	return eqn;\n"
+"}\n"
+"///////////////////////////////////////\n"
+"//	Matrix3x3\n"
+"///////////////////////////////////////\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_row[3];\n"
+"}Matrix3x3;\n"
+"__inline\n"
+"Matrix3x3 mtZero();\n"
+"__inline\n"
+"Matrix3x3 mtIdentity();\n"
+"__inline\n"
+"Matrix3x3 mtTranspose(Matrix3x3 m);\n"
+"__inline\n"
+"Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b);\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b);\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b);\n"
+"__inline\n"
+"Matrix3x3 mtZero()\n"
+"{\n"
+"	Matrix3x3 m;\n"
+"	m.m_row[0] = (float4)(0.f);\n"
+"	m.m_row[1] = (float4)(0.f);\n"
+"	m.m_row[2] = (float4)(0.f);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"Matrix3x3 mtIdentity()\n"
+"{\n"
+"	Matrix3x3 m;\n"
+"	m.m_row[0] = (float4)(1,0,0,0);\n"
+"	m.m_row[1] = (float4)(0,1,0,0);\n"
+"	m.m_row[2] = (float4)(0,0,1,0);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"Matrix3x3 mtTranspose(Matrix3x3 m)\n"
+"{\n"
+"	Matrix3x3 out;\n"
+"	out.m_row[0] = (float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b)\n"
+"{\n"
+"	Matrix3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	Matrix3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b)\n"
+"{\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a.m_row[0], b );\n"
+"	ans.y = dot3F4( a.m_row[1], b );\n"
+"	ans.z = dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b)\n"
+"{\n"
+"	float4 colx = make_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	float4 coly = make_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	float4 colz = make_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a, colx );\n"
+"	ans.y = dot3F4( a, coly );\n"
+"	ans.z = dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"typedef float4 Quaternion;\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b);\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in);\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec);\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q);\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in)\n"
+"{\n"
+"	return fastNormalize4(in);\n"
+"//	in /= length( in );\n"
+"//	return in;\n"
+"}\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(qtMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"__inline\n"
+"float4 qtInvRotate(const Quaternion q, float4 vec)\n"
+"{\n"
+"	return qtRotate( qtInvert( q ), vec );\n"
+"}\n"
+"#define WG_SIZE 64\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	Quaternion m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"	u32 m_shapeIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} Body;\n"
+"typedef struct\n"
+"{\n"
+"	Matrix3x3 m_invInertia;\n"
+"	Matrix3x3 m_initInvInertia;\n"
+"} Shape;\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_linear;\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_center;	\n"
+"	float m_jacCoeffInv[4];\n"
+"	float m_b[4];\n"
+"	float m_appliedRambdaDt[4];\n"
+"	float m_fJacCoeffInv[2];	\n"
+"	float m_fAppliedRambdaDt[2];	\n"
+"	u32 m_bodyA;\n"
+"	u32 m_bodyB;\n"
+"	int m_batchIdx;\n"
+"	u32 m_paddings[1];\n"
+"} Constraint4;\n"
+"typedef struct\n"
+"{\n"
+"	int m_nConstraints;\n"
+"	int m_start;\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBuffer;\n"
+"typedef struct\n"
+"{\n"
+"	int m_solveFriction;\n"
+"	int m_maxBatch;	//	long batch really kills the performance\n"
+"	int m_batchIdx;\n"
+"	int m_nSplit;\n"
+"//	int m_paddings[1];\n"
+"} ConstBufferBatchSolve;\n"
+" \n"
+"typedef struct \n"
+"{\n"
+"	int m_valInt0;\n"
+"	int m_valInt1;\n"
+"	int m_valInt2;\n"
+"	int m_valInt3;\n"
+"	float m_val0;\n"
+"	float m_val1;\n"
+"	float m_val2;\n"
+"	float m_val3;\n"
+"} SolverDebugInfo;\n"
+"//	others\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void ReorderContactKernel(__global struct b3Contact4Data* in, __global struct b3Contact4Data* out, __global int2* sortData, int4 cb )\n"
+"{\n"
+"	int nContacts = cb.x;\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < nContacts )\n"
+"	{\n"
+"		int srcIdx = sortData[gIdx].y;\n"
+"		out[gIdx] = in[srcIdx];\n"
+"	}\n"
+"}\n"
+"__kernel __attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void SetDeterminismSortDataChildShapeB(__global struct b3Contact4Data* contactsIn, __global int2* sortDataOut, int nContacts)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < nContacts )\n"
+"	{\n"
+"		int2 sd;\n"
+"		sd.x = contactsIn[gIdx].m_childIndexB;\n"
+"		sd.y = gIdx;\n"
+"		sortDataOut[gIdx] = sd;\n"
+"	}\n"
+"}\n"
+"__kernel __attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void SetDeterminismSortDataChildShapeA(__global struct b3Contact4Data* contactsIn, __global int2* sortDataInOut, int nContacts)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < nContacts )\n"
+"	{\n"
+"		int2 sdIn;\n"
+"		sdIn = sortDataInOut[gIdx];\n"
+"		int2 sdOut;\n"
+"		sdOut.x = contactsIn[sdIn.y].m_childIndexA;\n"
+"		sdOut.y = sdIn.y;\n"
+"		sortDataInOut[gIdx] = sdOut;\n"
+"	}\n"
+"}\n"
+"__kernel __attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void SetDeterminismSortDataBodyA(__global struct b3Contact4Data* contactsIn, __global int2* sortDataInOut, int nContacts)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < nContacts )\n"
+"	{\n"
+"		int2 sdIn;\n"
+"		sdIn = sortDataInOut[gIdx];\n"
+"		int2 sdOut;\n"
+"		sdOut.x = contactsIn[sdIn.y].m_bodyAPtrAndSignBit;\n"
+"		sdOut.y = sdIn.y;\n"
+"		sortDataInOut[gIdx] = sdOut;\n"
+"	}\n"
+"}\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void SetDeterminismSortDataBodyB(__global struct b3Contact4Data* contactsIn, __global int2* sortDataInOut, int nContacts)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < nContacts )\n"
+"	{\n"
+"		int2 sdIn;\n"
+"		sdIn = sortDataInOut[gIdx];\n"
+"		int2 sdOut;\n"
+"		sdOut.x = contactsIn[sdIn.y].m_bodyBPtrAndSignBit;\n"
+"		sdOut.y = sdIn.y;\n"
+"		sortDataInOut[gIdx] = sdOut;\n"
+"	}\n"
+"}\n"
+"typedef struct\n"
+"{\n"
+"	int m_nContacts;\n"
+"	int m_staticIdx;\n"
+"	float m_scale;\n"
+"	int m_nSplit;\n"
+"} ConstBufferSSD;\n"
+"__constant const int gridTable4x4[] = \n"
+"{\n"
+"    0,1,17,16,\n"
+"	1,2,18,19,\n"
+"	17,18,32,3,\n"
+"	16,19,3,34\n"
+"};\n"
+"__constant const int gridTable8x8[] = \n"
+"{\n"
+"	  0,  2,  3, 16, 17, 18, 19,  1,\n"
+"	 66, 64, 80, 67, 82, 81, 65, 83,\n"
+"	131,144,128,130,147,129,145,146,\n"
+"	208,195,194,192,193,211,210,209,\n"
+"	 21, 22, 23,  5,  4,  6,  7, 20,\n"
+"	 86, 85, 69, 87, 70, 68, 84, 71,\n"
+"	151,133,149,150,135,148,132,134,\n"
+"	197,27,214,213,212,199,198,196\n"
+"	\n"
+"};\n"
+"#define USE_SPATIAL_BATCHING 1\n"
+"#define USE_4x4_GRID 1\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void SetSortDataKernel(__global struct b3Contact4Data* gContact, __global Body* gBodies, __global int2* gSortDataOut, \n"
+"int nContacts,float scale,int4 nSplit,int staticIdx)\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	\n"
+"	if( gIdx < nContacts )\n"
+"	{\n"
+"		int aPtrAndSignBit  = gContact[gIdx].m_bodyAPtrAndSignBit;\n"
+"		int bPtrAndSignBit  = gContact[gIdx].m_bodyBPtrAndSignBit;\n"
+"		int aIdx = abs(aPtrAndSignBit );\n"
+"		int bIdx = abs(bPtrAndSignBit);\n"
+"		bool aStatic = (aPtrAndSignBit<0) ||(aPtrAndSignBit==staticIdx);\n"
+"		bool bStatic = (bPtrAndSignBit<0) ||(bPtrAndSignBit==staticIdx);\n"
+"#if USE_SPATIAL_BATCHING		\n"
+"		int idx = (aStatic)? bIdx: aIdx;\n"
+"		float4 p = gBodies[idx].m_pos;\n"
+"		int xIdx = (int)((p.x-((p.x<0.f)?1.f:0.f))*scale) & (nSplit.x-1);\n"
+"		int yIdx = (int)((p.y-((p.y<0.f)?1.f:0.f))*scale) & (nSplit.y-1);\n"
+"		int zIdx = (int)((p.z-((p.z<0.f)?1.f:0.f))*scale) & (nSplit.z-1);\n"
+"		int newIndex = (xIdx+yIdx*nSplit.x+zIdx*nSplit.x*nSplit.y);\n"
+"		\n"
+"#else//USE_SPATIAL_BATCHING\n"
+"	#if USE_4x4_GRID\n"
+"		int aa = aIdx&3;\n"
+"		int bb = bIdx&3;\n"
+"		if (aStatic)\n"
+"			aa = bb;\n"
+"		if (bStatic)\n"
+"			bb = aa;\n"
+"		int gridIndex = aa + bb*4;\n"
+"		int newIndex = gridTable4x4[gridIndex];\n"
+"	#else//USE_4x4_GRID\n"
+"		int aa = aIdx&7;\n"
+"		int bb = bIdx&7;\n"
+"		if (aStatic)\n"
+"			aa = bb;\n"
+"		if (bStatic)\n"
+"			bb = aa;\n"
+"		int gridIndex = aa + bb*8;\n"
+"		int newIndex = gridTable8x8[gridIndex];\n"
+"	#endif//USE_4x4_GRID\n"
+"#endif//USE_SPATIAL_BATCHING\n"
+"		gSortDataOut[gIdx].x = newIndex;\n"
+"		gSortDataOut[gIdx].y = gIdx;\n"
+"	}\n"
+"	else\n"
+"	{\n"
+"		gSortDataOut[gIdx].x = 0xffffffff;\n"
+"	}\n"
+"}\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void CopyConstraintKernel(__global struct b3Contact4Data* gIn, __global struct b3Contact4Data* gOut, int4 cb )\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	if( gIdx < cb.x )\n"
+"	{\n"
+"		gOut[gIdx] = gIn[gIdx];\n"
+"	}\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverUtils.cl b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverUtils.cl
new file mode 100644
index 0000000000..a21a08c3b4
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverUtils.cl
@@ -0,0 +1,968 @@
+/*
+Copyright (c) 2013 Advanced Micro Devices, Inc.  
+
+This software is provided 'as-is', without any express or implied warranty.
+In no event will the authors be held liable for any damages arising from the use of this software.
+Permission is granted to anyone to use this software for any purpose, 
+including commercial applications, and to alter it and redistribute it freely, 
+subject to the following restrictions:
+
+1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.
+2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.
+3. This notice may not be removed or altered from any source distribution.
+*/
+//Originally written by Erwin Coumans
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3Contact4Data.h"
+
+#pragma OPENCL EXTENSION cl_amd_printf : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable
+#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable
+
+
+#ifdef cl_ext_atomic_counters_32
+#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable
+#else
+#define counter32_t volatile global int*
+#endif
+
+typedef unsigned int u32;
+typedef unsigned short u16;
+typedef unsigned char u8;
+
+#define GET_GROUP_IDX get_group_id(0)
+#define GET_LOCAL_IDX get_local_id(0)
+#define GET_GLOBAL_IDX get_global_id(0)
+#define GET_GROUP_SIZE get_local_size(0)
+#define GET_NUM_GROUPS get_num_groups(0)
+#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)
+#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)
+#define AtomInc(x) atom_inc(&(x))
+#define AtomInc1(x, out) out = atom_inc(&(x))
+#define AppendInc(x, out) out = atomic_inc(x)
+#define AtomAdd(x, value) atom_add(&(x), value)
+#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )
+#define AtomXhg(x, value) atom_xchg ( &(x), value )
+
+
+#define SELECT_UINT4( b, a, condition ) select( b,a,condition )
+
+#define make_float4 (float4)
+#define make_float2 (float2)
+#define make_uint4 (uint4)
+#define make_int4 (int4)
+#define make_uint2 (uint2)
+#define make_int2 (int2)
+
+
+#define max2 max
+#define min2 min
+
+
+///////////////////////////////////////
+//	Vector
+///////////////////////////////////////
+__inline
+float fastDiv(float numerator, float denominator)
+{
+	return native_divide(numerator, denominator);	
+//	return numerator/denominator;	
+}
+
+__inline
+float4 fastDiv4(float4 numerator, float4 denominator)
+{
+	return native_divide(numerator, denominator);	
+}
+
+__inline
+float fastSqrtf(float f2)
+{
+	return native_sqrt(f2);
+//	return sqrt(f2);
+}
+
+__inline
+float fastRSqrt(float f2)
+{
+	return native_rsqrt(f2);
+}
+
+__inline
+float fastLength4(float4 v)
+{
+	return fast_length(v);
+}
+
+__inline
+float4 fastNormalize4(float4 v)
+{
+	return fast_normalize(v);
+}
+
+
+__inline
+float sqrtf(float a)
+{
+//	return sqrt(a);
+	return native_sqrt(a);
+}
+
+__inline
+float4 cross3(float4 a1, float4 b1)
+{
+
+	float4 	a=make_float4(a1.xyz,0.f);
+	float4 	b=make_float4(b1.xyz,0.f);
+	//float4 	a=a1;
+	//float4 	b=b1;
+	return cross(a,b);
+}
+
+__inline
+float dot3F4(float4 a, float4 b)
+{
+	float4 a1 = make_float4(a.xyz,0.f);
+	float4 b1 = make_float4(b.xyz,0.f);
+	return dot(a1, b1);
+}
+
+__inline
+float length3(const float4 a)
+{
+	return sqrtf(dot3F4(a,a));
+}
+
+__inline
+float dot4(const float4 a, const float4 b)
+{
+	return dot( a, b );
+}
+
+//	for height
+__inline
+float dot3w1(const float4 point, const float4 eqn)
+{
+	return dot3F4(point,eqn) + eqn.w;
+}
+
+__inline
+float4 normalize3(const float4 a)
+{
+	float4 n = make_float4(a.x, a.y, a.z, 0.f);
+	return fastNormalize4( n );
+//	float length = sqrtf(dot3F4(a, a));
+//	return 1.f/length * a;
+}
+
+__inline
+float4 normalize4(const float4 a)
+{
+	float length = sqrtf(dot4(a, a));
+	return 1.f/length * a;
+}
+
+__inline
+float4 createEquation(const float4 a, const float4 b, const float4 c)
+{
+	float4 eqn;
+	float4 ab = b-a;
+	float4 ac = c-a;
+	eqn = normalize3( cross3(ab, ac) );
+	eqn.w = -dot3F4(eqn,a);
+	return eqn;
+}
+
+///////////////////////////////////////
+//	Matrix3x3
+///////////////////////////////////////
+
+typedef struct
+{
+	float4 m_row[3];
+}Matrix3x3;
+
+__inline
+Matrix3x3 mtZero();
+
+__inline
+Matrix3x3 mtIdentity();
+
+__inline
+Matrix3x3 mtTranspose(Matrix3x3 m);
+
+__inline
+Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b);
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b);
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b);
+
+__inline
+Matrix3x3 mtZero()
+{
+	Matrix3x3 m;
+	m.m_row[0] = (float4)(0.f);
+	m.m_row[1] = (float4)(0.f);
+	m.m_row[2] = (float4)(0.f);
+	return m;
+}
+
+__inline
+Matrix3x3 mtIdentity()
+{
+	Matrix3x3 m;
+	m.m_row[0] = (float4)(1,0,0,0);
+	m.m_row[1] = (float4)(0,1,0,0);
+	m.m_row[2] = (float4)(0,0,1,0);
+	return m;
+}
+
+__inline
+Matrix3x3 mtTranspose(Matrix3x3 m)
+{
+	Matrix3x3 out;
+	out.m_row[0] = (float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);
+	out.m_row[1] = (float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);
+	out.m_row[2] = (float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);
+	return out;
+}
+
+__inline
+Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b)
+{
+	Matrix3x3 transB;
+	transB = mtTranspose( b );
+	Matrix3x3 ans;
+	//	why this doesn't run when 0ing in the for{}
+	a.m_row[0].w = 0.f;
+	a.m_row[1].w = 0.f;
+	a.m_row[2].w = 0.f;
+	for(int i=0; i<3; i++)
+	{
+//	a.m_row[i].w = 0.f;
+		ans.m_row[i].x = dot3F4(a.m_row[i],transB.m_row[0]);
+		ans.m_row[i].y = dot3F4(a.m_row[i],transB.m_row[1]);
+		ans.m_row[i].z = dot3F4(a.m_row[i],transB.m_row[2]);
+		ans.m_row[i].w = 0.f;
+	}
+	return ans;
+}
+
+__inline
+float4 mtMul1(Matrix3x3 a, float4 b)
+{
+	float4 ans;
+	ans.x = dot3F4( a.m_row[0], b );
+	ans.y = dot3F4( a.m_row[1], b );
+	ans.z = dot3F4( a.m_row[2], b );
+	ans.w = 0.f;
+	return ans;
+}
+
+__inline
+float4 mtMul3(float4 a, Matrix3x3 b)
+{
+	float4 colx = make_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);
+	float4 coly = make_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);
+	float4 colz = make_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);
+
+	float4 ans;
+	ans.x = dot3F4( a, colx );
+	ans.y = dot3F4( a, coly );
+	ans.z = dot3F4( a, colz );
+	return ans;
+}
+
+///////////////////////////////////////
+//	Quaternion
+///////////////////////////////////////
+
+typedef float4 Quaternion;
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b);
+
+__inline
+Quaternion qtNormalize(Quaternion in);
+
+__inline
+float4 qtRotate(Quaternion q, float4 vec);
+
+__inline
+Quaternion qtInvert(Quaternion q);
+
+
+
+
+
+__inline
+Quaternion qtMul(Quaternion a, Quaternion b)
+{
+	Quaternion ans;
+	ans = cross3( a, b );
+	ans += a.w*b+b.w*a;
+//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);
+	ans.w = a.w*b.w - dot3F4(a, b);
+	return ans;
+}
+
+__inline
+Quaternion qtNormalize(Quaternion in)
+{
+	return fastNormalize4(in);
+//	in /= length( in );
+//	return in;
+}
+__inline
+float4 qtRotate(Quaternion q, float4 vec)
+{
+	Quaternion qInv = qtInvert( q );
+	float4 vcpy = vec;
+	vcpy.w = 0.f;
+	float4 out = qtMul(qtMul(q,vcpy),qInv);
+	return out;
+}
+
+__inline
+Quaternion qtInvert(Quaternion q)
+{
+	return (Quaternion)(-q.xyz, q.w);
+}
+
+__inline
+float4 qtInvRotate(const Quaternion q, float4 vec)
+{
+	return qtRotate( qtInvert( q ), vec );
+}
+
+
+
+
+#define WG_SIZE 64
+
+typedef struct
+{
+	float4 m_pos;
+	Quaternion m_quat;
+	float4 m_linVel;
+	float4 m_angVel;
+
+	u32 m_shapeIdx;
+	float m_invMass;
+	float m_restituitionCoeff;
+	float m_frictionCoeff;
+} Body;
+
+
+
+typedef struct
+{
+	Matrix3x3 m_invInertia;
+	Matrix3x3 m_initInvInertia;
+} Shape;
+
+typedef struct
+{
+	float4 m_linear;
+	float4 m_worldPos[4];
+	float4 m_center;	
+	float m_jacCoeffInv[4];
+	float m_b[4];
+	float m_appliedRambdaDt[4];
+
+	float m_fJacCoeffInv[2];	
+	float m_fAppliedRambdaDt[2];	
+
+	u32 m_bodyA;
+	u32 m_bodyB;
+	int m_batchIdx;
+	u32 m_paddings;
+} Constraint4;
+
+
+
+
+
+
+__kernel void CountBodiesKernel(__global struct b3Contact4Data* manifoldPtr, __global unsigned int* bodyCount, __global int2* contactConstraintOffsets, int numContactManifolds, int fixedBodyIndex)
+{
+	int i = GET_GLOBAL_IDX;
+	
+	if( i < numContactManifolds)
+	{
+		int pa = manifoldPtr[i].m_bodyAPtrAndSignBit;
+		bool isFixedA = (pa <0) || (pa == fixedBodyIndex);
+		int bodyIndexA = abs(pa);
+		if (!isFixedA)
+		{
+			 AtomInc1(bodyCount[bodyIndexA],contactConstraintOffsets[i].x);
+		}
+		barrier(CLK_GLOBAL_MEM_FENCE);
+		int pb = manifoldPtr[i].m_bodyBPtrAndSignBit;
+		bool isFixedB = (pb <0) || (pb == fixedBodyIndex);
+		int bodyIndexB = abs(pb);
+		if (!isFixedB)
+		{
+			AtomInc1(bodyCount[bodyIndexB],contactConstraintOffsets[i].y);
+		} 
+	}
+}
+
+__kernel void ClearVelocitiesKernel(__global float4* linearVelocities,__global float4* angularVelocities, int numSplitBodies)
+{
+	int i = GET_GLOBAL_IDX;
+	
+	if( i < numSplitBodies)
+	{
+		linearVelocities[i] = make_float4(0);
+		angularVelocities[i] = make_float4(0);
+	}
+}
+
+
+__kernel void AverageVelocitiesKernel(__global Body* gBodies,__global int* offsetSplitBodies,__global const unsigned int* bodyCount,
+__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities, int numBodies)
+{
+	int i = GET_GLOBAL_IDX;
+	if (i<numBodies)
+	{
+		if (gBodies[i].m_invMass)
+		{
+			int bodyOffset = offsetSplitBodies[i];
+			int count = bodyCount[i];
+			float factor = 1.f/((float)count);
+			float4 averageLinVel = make_float4(0.f);
+			float4 averageAngVel = make_float4(0.f);
+			
+			for (int j=0;j<count;j++)
+			{
+				averageLinVel += deltaLinearVelocities[bodyOffset+j]*factor;
+				averageAngVel += deltaAngularVelocities[bodyOffset+j]*factor;
+			}
+			
+			for (int j=0;j<count;j++)
+			{
+				deltaLinearVelocities[bodyOffset+j] = averageLinVel;
+				deltaAngularVelocities[bodyOffset+j] = averageAngVel;
+			}
+			
+		}//bodies[i].m_invMass
+	}//i<numBodies
+}
+
+
+
+void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)
+{
+	*linear = make_float4(n.xyz,0.f);
+	*angular0 = cross3(r0, n);
+	*angular1 = -cross3(r1, n);
+}
+
+
+float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )
+{
+	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);
+}
+
+
+float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,
+					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1, float countA, float countB)
+{
+	//	linear0,1 are normlized
+	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;
+	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);
+	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;
+	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);
+	return -1.f/((jmj0+jmj1)*countA+(jmj2+jmj3)*countB);
+}
+
+
+void btPlaneSpace1 (float4 n, float4* p, float4* q);
+ void btPlaneSpace1 (float4 n, float4* p, float4* q)
+{
+  if (fabs(n.z) > 0.70710678f) {
+    // choose p in y-z plane
+    float a = n.y*n.y + n.z*n.z;
+    float k = 1.f/sqrt(a);
+    p[0].x = 0;
+	p[0].y = -n.z*k;
+	p[0].z = n.y*k;
+    // set q = n x p
+    q[0].x = a*k;
+	q[0].y = -n.x*p[0].z;
+	q[0].z = n.x*p[0].y;
+  }
+  else {
+    // choose p in x-y plane
+    float a = n.x*n.x + n.y*n.y;
+    float k = 1.f/sqrt(a);
+    p[0].x = -n.y*k;
+	p[0].y = n.x*k;
+	p[0].z = 0;
+    // set q = n x p
+    q[0].x = -n.z*p[0].y;
+	q[0].y = n.z*p[0].x;
+	q[0].z = a*k;
+  }
+}
+
+
+
+
+
+void solveContact(__global Constraint4* cs,
+			float4 posA, float4* linVelA, float4* angVelA, float invMassA, Matrix3x3 invInertiaA,
+			float4 posB, float4* linVelB, float4* angVelB, float invMassB, Matrix3x3 invInertiaB,
+			float4* dLinVelA, float4* dAngVelA, float4* dLinVelB, float4* dAngVelB)
+{
+	float minRambdaDt = 0;
+	float maxRambdaDt = FLT_MAX;
+
+	for(int ic=0; ic<4; ic++)
+	{
+		if( cs->m_jacCoeffInv[ic] == 0.f ) continue;
+
+		float4 angular0, angular1, linear;
+		float4 r0 = cs->m_worldPos[ic] - posA;
+		float4 r1 = cs->m_worldPos[ic] - posB;
+		setLinearAndAngular( cs->m_linear, r0, r1, &linear, &angular0, &angular1 );
+	
+
+
+		float rambdaDt = calcRelVel( cs->m_linear, -cs->m_linear, angular0, angular1, 
+			*linVelA+*dLinVelA, *angVelA+*dAngVelA, *linVelB+*dLinVelB, *angVelB+*dAngVelB ) + cs->m_b[ic];
+		rambdaDt *= cs->m_jacCoeffInv[ic];
+
+		
+		{
+			float prevSum = cs->m_appliedRambdaDt[ic];
+			float updated = prevSum;
+			updated += rambdaDt;
+			updated = max2( updated, minRambdaDt );
+			updated = min2( updated, maxRambdaDt );
+			rambdaDt = updated - prevSum;
+			cs->m_appliedRambdaDt[ic] = updated;
+		}
+
+			
+		float4 linImp0 = invMassA*linear*rambdaDt;
+		float4 linImp1 = invMassB*(-linear)*rambdaDt;
+		float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;
+		float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;
+
+		
+		if (invMassA)
+		{
+			*dLinVelA += linImp0;
+			*dAngVelA += angImp0;
+		}
+		if (invMassB)
+		{
+			*dLinVelB += linImp1;
+			*dAngVelB += angImp1;
+		}
+	}
+}
+
+
+//	solveContactConstraint( gBodies, gShapes, &gConstraints[i] ,contactConstraintOffsets,offsetSplitBodies, deltaLinearVelocities, deltaAngularVelocities);
+
+
+void solveContactConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs, 
+__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,
+__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities)
+{
+
+	//float frictionCoeff = ldsCs[0].m_linear.w;
+	int aIdx = ldsCs[0].m_bodyA;
+	int bIdx = ldsCs[0].m_bodyB;
+
+	float4 posA = gBodies[aIdx].m_pos;
+	float4 linVelA = gBodies[aIdx].m_linVel;
+	float4 angVelA = gBodies[aIdx].m_angVel;
+	float invMassA = gBodies[aIdx].m_invMass;
+	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;
+
+	float4 posB = gBodies[bIdx].m_pos;
+	float4 linVelB = gBodies[bIdx].m_linVel;
+	float4 angVelB = gBodies[bIdx].m_angVel;
+	float invMassB = gBodies[bIdx].m_invMass;
+	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;
+
+			
+	float4 dLinVelA = make_float4(0,0,0,0);
+	float4 dAngVelA = make_float4(0,0,0,0);
+	float4 dLinVelB = make_float4(0,0,0,0);
+	float4 dAngVelB = make_float4(0,0,0,0);
+			
+	int bodyOffsetA = offsetSplitBodies[aIdx];
+	int constraintOffsetA = contactConstraintOffsets[0].x;
+	int splitIndexA = bodyOffsetA+constraintOffsetA;
+	
+	if (invMassA)
+	{
+		dLinVelA = deltaLinearVelocities[splitIndexA];
+		dAngVelA = deltaAngularVelocities[splitIndexA];
+	}
+
+	int bodyOffsetB = offsetSplitBodies[bIdx];
+	int constraintOffsetB = contactConstraintOffsets[0].y;
+	int splitIndexB= bodyOffsetB+constraintOffsetB;
+
+	if (invMassB)
+	{
+		dLinVelB = deltaLinearVelocities[splitIndexB];
+		dAngVelB = deltaAngularVelocities[splitIndexB];
+	}
+
+	solveContact( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,
+			posB, &linVelB, &angVelB, invMassB, invInertiaB ,&dLinVelA, &dAngVelA, &dLinVelB, &dAngVelB);
+
+	if (invMassA)
+	{
+		deltaLinearVelocities[splitIndexA] = dLinVelA;
+		deltaAngularVelocities[splitIndexA] = dAngVelA;
+	} 
+	if (invMassB)
+	{
+		deltaLinearVelocities[splitIndexB] = dLinVelB;
+		deltaAngularVelocities[splitIndexB] = dAngVelB;
+	}
+
+}
+
+
+__kernel void SolveContactJacobiKernel(__global Constraint4* gConstraints, __global Body* gBodies, __global Shape* gShapes ,
+__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities,
+float deltaTime, float positionDrift, float positionConstraintCoeff, int fixedBodyIndex, int numManifolds
+)
+{
+	int i = GET_GLOBAL_IDX;
+	if (i<numManifolds)
+	{
+		solveContactConstraint( gBodies, gShapes, &gConstraints[i] ,&contactConstraintOffsets[i],offsetSplitBodies, deltaLinearVelocities, deltaAngularVelocities);
+	}
+}
+
+
+
+
+void solveFrictionConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs,
+							__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,
+							__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities)
+{
+	float frictionCoeff = 0.7f;//ldsCs[0].m_linear.w;
+	int aIdx = ldsCs[0].m_bodyA;
+	int bIdx = ldsCs[0].m_bodyB;
+
+
+	float4 posA = gBodies[aIdx].m_pos;
+	float4 linVelA = gBodies[aIdx].m_linVel;
+	float4 angVelA = gBodies[aIdx].m_angVel;
+	float invMassA = gBodies[aIdx].m_invMass;
+	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;
+
+	float4 posB = gBodies[bIdx].m_pos;
+	float4 linVelB = gBodies[bIdx].m_linVel;
+	float4 angVelB = gBodies[bIdx].m_angVel;
+	float invMassB = gBodies[bIdx].m_invMass;
+	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;
+	
+
+	float4 dLinVelA = make_float4(0,0,0,0);
+	float4 dAngVelA = make_float4(0,0,0,0);
+	float4 dLinVelB = make_float4(0,0,0,0);
+	float4 dAngVelB = make_float4(0,0,0,0);
+			
+	int bodyOffsetA = offsetSplitBodies[aIdx];
+	int constraintOffsetA = contactConstraintOffsets[0].x;
+	int splitIndexA = bodyOffsetA+constraintOffsetA;
+	
+	if (invMassA)
+	{
+		dLinVelA = deltaLinearVelocities[splitIndexA];
+		dAngVelA = deltaAngularVelocities[splitIndexA];
+	}
+
+	int bodyOffsetB = offsetSplitBodies[bIdx];
+	int constraintOffsetB = contactConstraintOffsets[0].y;
+	int splitIndexB= bodyOffsetB+constraintOffsetB;
+
+	if (invMassB)
+	{
+		dLinVelB = deltaLinearVelocities[splitIndexB];
+		dAngVelB = deltaAngularVelocities[splitIndexB];
+	}
+
+
+
+
+	{
+		float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};
+		float minRambdaDt[4] = {0.f,0.f,0.f,0.f};
+
+		float sum = 0;
+		for(int j=0; j<4; j++)
+		{
+			sum +=ldsCs[0].m_appliedRambdaDt[j];
+		}
+		frictionCoeff = 0.7f;
+		for(int j=0; j<4; j++)
+		{
+			maxRambdaDt[j] = frictionCoeff*sum;
+			minRambdaDt[j] = -maxRambdaDt[j];
+		}
+
+		
+//		solveFriction( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,
+//			posB, &linVelB, &angVelB, invMassB, invInertiaB, maxRambdaDt, minRambdaDt );
+		
+		
+		{
+			
+			__global Constraint4* cs = ldsCs;
+			
+			if( cs->m_fJacCoeffInv[0] == 0 && cs->m_fJacCoeffInv[0] == 0 ) return;
+			const float4 center = cs->m_center;
+			
+			float4 n = -cs->m_linear;
+			
+			float4 tangent[2];
+			btPlaneSpace1(n,&tangent[0],&tangent[1]);
+			float4 angular0, angular1, linear;
+			float4 r0 = center - posA;
+			float4 r1 = center - posB;
+			for(int i=0; i<2; i++)
+			{
+				setLinearAndAngular( tangent[i], r0, r1, &linear, &angular0, &angular1 );
+				float rambdaDt = calcRelVel(linear, -linear, angular0, angular1,
+											linVelA+dLinVelA, angVelA+dAngVelA, linVelB+dLinVelB, angVelB+dAngVelB );
+				rambdaDt *= cs->m_fJacCoeffInv[i];
+				
+				{
+					float prevSum = cs->m_fAppliedRambdaDt[i];
+					float updated = prevSum;
+					updated += rambdaDt;
+					updated = max2( updated, minRambdaDt[i] );
+					updated = min2( updated, maxRambdaDt[i] );
+					rambdaDt = updated - prevSum;
+					cs->m_fAppliedRambdaDt[i] = updated;
+				}
+				
+				float4 linImp0 = invMassA*linear*rambdaDt;
+				float4 linImp1 = invMassB*(-linear)*rambdaDt;
+				float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;
+				float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;
+				
+				dLinVelA += linImp0;
+				dAngVelA += angImp0;
+				dLinVelB += linImp1;
+				dAngVelB += angImp1;
+			}
+			{	//	angular damping for point constraint
+				float4 ab = normalize3( posB - posA );
+				float4 ac = normalize3( center - posA );
+				if( dot3F4( ab, ac ) > 0.95f  || (invMassA == 0.f || invMassB == 0.f))
+				{
+					float angNA = dot3F4( n, angVelA );
+					float angNB = dot3F4( n, angVelB );
+					
+					dAngVelA -= (angNA*0.1f)*n;
+					dAngVelB -= (angNB*0.1f)*n;
+				}
+			}
+		}
+
+		
+		
+	}
+
+	if (invMassA)
+	{
+		deltaLinearVelocities[splitIndexA] = dLinVelA;
+		deltaAngularVelocities[splitIndexA] = dAngVelA;
+	} 
+	if (invMassB)
+	{
+		deltaLinearVelocities[splitIndexB] = dLinVelB;
+		deltaAngularVelocities[splitIndexB] = dAngVelB;
+	}
+ 
+
+}
+
+
+__kernel void SolveFrictionJacobiKernel(__global Constraint4* gConstraints, __global Body* gBodies, __global Shape* gShapes ,
+										__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,
+										__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities,
+										float deltaTime, float positionDrift, float positionConstraintCoeff, int fixedBodyIndex, int numManifolds
+)
+{
+	int i = GET_GLOBAL_IDX;
+	if (i<numManifolds)
+	{
+		solveFrictionConstraint( gBodies, gShapes, &gConstraints[i] ,&contactConstraintOffsets[i],offsetSplitBodies, deltaLinearVelocities, deltaAngularVelocities);
+	}
+}
+
+
+__kernel void UpdateBodyVelocitiesKernel(__global Body* gBodies,__global int* offsetSplitBodies,__global const unsigned int* bodyCount,
+									__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities, int numBodies)
+{
+	int i = GET_GLOBAL_IDX;
+	if (i<numBodies)
+	{
+		if (gBodies[i].m_invMass)
+		{
+			int bodyOffset = offsetSplitBodies[i];
+			int count = bodyCount[i];
+			if (count)
+			{
+				gBodies[i].m_linVel += deltaLinearVelocities[bodyOffset];
+				gBodies[i].m_angVel += deltaAngularVelocities[bodyOffset];
+			}
+		}
+	}
+}
+
+
+
+void setConstraint4( const float4 posA, const float4 linVelA, const float4 angVelA, float invMassA, const Matrix3x3 invInertiaA,
+	const float4 posB, const float4 linVelB, const float4 angVelB, float invMassB, const Matrix3x3 invInertiaB, 
+	__global struct b3Contact4Data* src, float dt, float positionDrift, float positionConstraintCoeff,float countA, float countB,
+	Constraint4* dstC )
+{
+	dstC->m_bodyA = abs(src->m_bodyAPtrAndSignBit);
+	dstC->m_bodyB = abs(src->m_bodyBPtrAndSignBit);
+
+	float dtInv = 1.f/dt;
+	for(int ic=0; ic<4; ic++)
+	{
+		dstC->m_appliedRambdaDt[ic] = 0.f;
+	}
+	dstC->m_fJacCoeffInv[0] = dstC->m_fJacCoeffInv[1] = 0.f;
+
+
+	dstC->m_linear = src->m_worldNormalOnB;
+	dstC->m_linear.w = 0.7f ;//src->getFrictionCoeff() );
+	for(int ic=0; ic<4; ic++)
+	{
+		float4 r0 = src->m_worldPosB[ic] - posA;
+		float4 r1 = src->m_worldPosB[ic] - posB;
+
+		if( ic >= src->m_worldNormalOnB.w )//npoints
+		{
+			dstC->m_jacCoeffInv[ic] = 0.f;
+			continue;
+		}
+
+		float relVelN;
+		{
+			float4 linear, angular0, angular1;
+			setLinearAndAngular(src->m_worldNormalOnB, r0, r1, &linear, &angular0, &angular1);
+
+			dstC->m_jacCoeffInv[ic] = calcJacCoeff(linear, -linear, angular0, angular1,
+				invMassA, &invInertiaA, invMassB, &invInertiaB , countA, countB);
+
+			relVelN = calcRelVel(linear, -linear, angular0, angular1,
+				linVelA, angVelA, linVelB, angVelB);
+
+			float e = 0.f;//src->getRestituitionCoeff();
+			if( relVelN*relVelN < 0.004f ) e = 0.f;
+
+			dstC->m_b[ic] = e*relVelN;
+			//float penetration = src->m_worldPosB[ic].w;
+			dstC->m_b[ic] += (src->m_worldPosB[ic].w + positionDrift)*positionConstraintCoeff*dtInv;
+			dstC->m_appliedRambdaDt[ic] = 0.f;
+		}
+	}
+
+	if( src->m_worldNormalOnB.w > 0 )//npoints
+	{	//	prepare friction
+		float4 center = make_float4(0.f);
+		for(int i=0; i<src->m_worldNormalOnB.w; i++) 
+			center += src->m_worldPosB[i];
+		center /= (float)src->m_worldNormalOnB.w;
+
+		float4 tangent[2];
+		btPlaneSpace1(-src->m_worldNormalOnB,&tangent[0],&tangent[1]);
+		
+		float4 r[2];
+		r[0] = center - posA;
+		r[1] = center - posB;
+
+		for(int i=0; i<2; i++)
+		{
+			float4 linear, angular0, angular1;
+			setLinearAndAngular(tangent[i], r[0], r[1], &linear, &angular0, &angular1);
+
+			dstC->m_fJacCoeffInv[i] = calcJacCoeff(linear, -linear, angular0, angular1,
+				invMassA, &invInertiaA, invMassB, &invInertiaB ,countA, countB);
+			dstC->m_fAppliedRambdaDt[i] = 0.f;
+		}
+		dstC->m_center = center;
+	}
+
+	for(int i=0; i<4; i++)
+	{
+		if( i<src->m_worldNormalOnB.w )
+		{
+			dstC->m_worldPos[i] = src->m_worldPosB[i];
+		}
+		else
+		{
+			dstC->m_worldPos[i] = make_float4(0.f);
+		}
+	}
+}
+
+
+__kernel
+__attribute__((reqd_work_group_size(WG_SIZE,1,1)))
+void ContactToConstraintSplitKernel(__global const struct b3Contact4Data* gContact, __global const Body* gBodies, __global const Shape* gShapes, __global Constraint4* gConstraintOut, 
+__global const unsigned int* bodyCount,
+int nContacts,
+float dt,
+float positionDrift,
+float positionConstraintCoeff
+)
+{
+	int gIdx = GET_GLOBAL_IDX;
+	
+	if( gIdx < nContacts )
+	{
+		int aIdx = abs(gContact[gIdx].m_bodyAPtrAndSignBit);
+		int bIdx = abs(gContact[gIdx].m_bodyBPtrAndSignBit);
+
+		float4 posA = gBodies[aIdx].m_pos;
+		float4 linVelA = gBodies[aIdx].m_linVel;
+		float4 angVelA = gBodies[aIdx].m_angVel;
+		float invMassA = gBodies[aIdx].m_invMass;
+		Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;
+
+		float4 posB = gBodies[bIdx].m_pos;
+		float4 linVelB = gBodies[bIdx].m_linVel;
+		float4 angVelB = gBodies[bIdx].m_angVel;
+		float invMassB = gBodies[bIdx].m_invMass;
+		Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;
+
+		Constraint4 cs;
+
+		float countA = invMassA != 0.f ? (float)bodyCount[aIdx] : 1;
+		float countB = invMassB != 0.f ? (float)bodyCount[bIdx] : 1;
+
+    	setConstraint4( posA, linVelA, angVelA, invMassA, invInertiaA, posB, linVelB, angVelB, invMassB, invInertiaB,
+			&gContact[gIdx], dt, positionDrift, positionConstraintCoeff,countA,countB,
+			&cs  );
+		
+		cs.m_batchIdx = gContact[gIdx].m_batchIdx;
+
+		gConstraintOut[gIdx] = cs;
+	}
+}
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverUtils.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverUtils.h
new file mode 100644
index 0000000000..c0173ad9f4
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/solverUtils.h
@@ -0,0 +1,909 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* solverUtilsCL= \
+"/*\n"
+"Copyright (c) 2013 Advanced Micro Devices, Inc.  \n"
+"This software is provided 'as-is', without any express or implied warranty.\n"
+"In no event will the authors be held liable for any damages arising from the use of this software.\n"
+"Permission is granted to anyone to use this software for any purpose, \n"
+"including commercial applications, and to alter it and redistribute it freely, \n"
+"subject to the following restrictions:\n"
+"1. The origin of this software must not be misrepresented; you must not claim that you wrote the original software. If you use this software in a product, an acknowledgment in the product documentation would be appreciated but is not required.\n"
+"2. Altered source versions must be plainly marked as such, and must not be misrepresented as being the original software.\n"
+"3. This notice may not be removed or altered from any source distribution.\n"
+"*/\n"
+"//Originally written by Erwin Coumans\n"
+"#ifndef B3_CONTACT4DATA_H\n"
+"#define B3_CONTACT4DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#define B3_FLOAT4_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#define B3_PLATFORM_DEFINITIONS_H\n"
+"struct MyTest\n"
+"{\n"
+"	int bla;\n"
+"};\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"//keep B3_LARGE_FLOAT*B3_LARGE_FLOAT < FLT_MAX\n"
+"#define B3_LARGE_FLOAT 1e18f\n"
+"#define B3_INFINITY 1e18f\n"
+"#define b3Assert(a)\n"
+"#define b3ConstArray(a) __global const a*\n"
+"#define b3AtomicInc atomic_inc\n"
+"#define b3AtomicAdd atomic_add\n"
+"#define b3Fabs fabs\n"
+"#define b3Sqrt native_sqrt\n"
+"#define b3Sin native_sin\n"
+"#define b3Cos native_cos\n"
+"#define B3_STATIC\n"
+"#endif\n"
+"#endif\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Float4;\n"
+"	#define b3Float4ConstArg const b3Float4\n"
+"	#define b3MakeFloat4 (float4)\n"
+"	float b3Dot3F4(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return dot(a1, b1);\n"
+"	}\n"
+"	b3Float4 b3Cross3(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return cross(a1, b1);\n"
+"	}\n"
+"	#define b3MinFloat4 min\n"
+"	#define b3MaxFloat4 max\n"
+"	#define b3Normalized(a) normalize(a)\n"
+"#endif \n"
+"		\n"
+"inline bool b3IsAlmostZero(b3Float4ConstArg v)\n"
+"{\n"
+"	if(b3Fabs(v.x)>1e-6 || b3Fabs(v.y)>1e-6 || b3Fabs(v.z)>1e-6)	\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"inline int    b3MaxDot( b3Float4ConstArg vec, __global const b3Float4* vecArray, int vecLen, float* dotOut )\n"
+"{\n"
+"    float maxDot = -B3_INFINITY;\n"
+"    int i = 0;\n"
+"    int ptIndex = -1;\n"
+"    for( i = 0; i < vecLen; i++ )\n"
+"    {\n"
+"        float dot = b3Dot3F4(vecArray[i],vec);\n"
+"            \n"
+"        if( dot > maxDot )\n"
+"        {\n"
+"            maxDot = dot;\n"
+"            ptIndex = i;\n"
+"        }\n"
+"    }\n"
+"	b3Assert(ptIndex>=0);\n"
+"    if (ptIndex<0)\n"
+"	{\n"
+"		ptIndex = 0;\n"
+"	}\n"
+"    *dotOut = maxDot;\n"
+"    return ptIndex;\n"
+"}\n"
+"#endif //B3_FLOAT4_H\n"
+"typedef  struct b3Contact4Data b3Contact4Data_t;\n"
+"struct b3Contact4Data\n"
+"{\n"
+"	b3Float4	m_worldPosB[4];\n"
+"//	b3Float4	m_localPosA[4];\n"
+"//	b3Float4	m_localPosB[4];\n"
+"	b3Float4	m_worldNormalOnB;	//	w: m_nPoints\n"
+"	unsigned short  m_restituitionCoeffCmp;\n"
+"	unsigned short  m_frictionCoeffCmp;\n"
+"	int m_batchIdx;\n"
+"	int m_bodyAPtrAndSignBit;//x:m_bodyAPtr, y:m_bodyBPtr\n"
+"	int m_bodyBPtrAndSignBit;\n"
+"	int	m_childIndexA;\n"
+"	int	m_childIndexB;\n"
+"	int m_unused1;\n"
+"	int m_unused2;\n"
+"};\n"
+"inline int b3Contact4Data_getNumPoints(const struct b3Contact4Data* contact)\n"
+"{\n"
+"	return (int)contact->m_worldNormalOnB.w;\n"
+"};\n"
+"inline void b3Contact4Data_setNumPoints(struct b3Contact4Data* contact, int numPoints)\n"
+"{\n"
+"	contact->m_worldNormalOnB.w = (float)numPoints;\n"
+"};\n"
+"#endif //B3_CONTACT4DATA_H\n"
+"#pragma OPENCL EXTENSION cl_amd_printf : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_local_int32_extended_atomics : enable\n"
+"#pragma OPENCL EXTENSION cl_khr_global_int32_extended_atomics : enable\n"
+"#ifdef cl_ext_atomic_counters_32\n"
+"#pragma OPENCL EXTENSION cl_ext_atomic_counters_32 : enable\n"
+"#else\n"
+"#define counter32_t volatile global int*\n"
+"#endif\n"
+"typedef unsigned int u32;\n"
+"typedef unsigned short u16;\n"
+"typedef unsigned char u8;\n"
+"#define GET_GROUP_IDX get_group_id(0)\n"
+"#define GET_LOCAL_IDX get_local_id(0)\n"
+"#define GET_GLOBAL_IDX get_global_id(0)\n"
+"#define GET_GROUP_SIZE get_local_size(0)\n"
+"#define GET_NUM_GROUPS get_num_groups(0)\n"
+"#define GROUP_LDS_BARRIER barrier(CLK_LOCAL_MEM_FENCE)\n"
+"#define GROUP_MEM_FENCE mem_fence(CLK_LOCAL_MEM_FENCE)\n"
+"#define AtomInc(x) atom_inc(&(x))\n"
+"#define AtomInc1(x, out) out = atom_inc(&(x))\n"
+"#define AppendInc(x, out) out = atomic_inc(x)\n"
+"#define AtomAdd(x, value) atom_add(&(x), value)\n"
+"#define AtomCmpxhg(x, cmp, value) atom_cmpxchg( &(x), cmp, value )\n"
+"#define AtomXhg(x, value) atom_xchg ( &(x), value )\n"
+"#define SELECT_UINT4( b, a, condition ) select( b,a,condition )\n"
+"#define make_float4 (float4)\n"
+"#define make_float2 (float2)\n"
+"#define make_uint4 (uint4)\n"
+"#define make_int4 (int4)\n"
+"#define make_uint2 (uint2)\n"
+"#define make_int2 (int2)\n"
+"#define max2 max\n"
+"#define min2 min\n"
+"///////////////////////////////////////\n"
+"//	Vector\n"
+"///////////////////////////////////////\n"
+"__inline\n"
+"float fastDiv(float numerator, float denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"//	return numerator/denominator;	\n"
+"}\n"
+"__inline\n"
+"float4 fastDiv4(float4 numerator, float4 denominator)\n"
+"{\n"
+"	return native_divide(numerator, denominator);	\n"
+"}\n"
+"__inline\n"
+"float fastSqrtf(float f2)\n"
+"{\n"
+"	return native_sqrt(f2);\n"
+"//	return sqrt(f2);\n"
+"}\n"
+"__inline\n"
+"float fastRSqrt(float f2)\n"
+"{\n"
+"	return native_rsqrt(f2);\n"
+"}\n"
+"__inline\n"
+"float fastLength4(float4 v)\n"
+"{\n"
+"	return fast_length(v);\n"
+"}\n"
+"__inline\n"
+"float4 fastNormalize4(float4 v)\n"
+"{\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"__inline\n"
+"float sqrtf(float a)\n"
+"{\n"
+"//	return sqrt(a);\n"
+"	return native_sqrt(a);\n"
+"}\n"
+"__inline\n"
+"float4 cross3(float4 a1, float4 b1)\n"
+"{\n"
+"	float4 	a=make_float4(a1.xyz,0.f);\n"
+"	float4 	b=make_float4(b1.xyz,0.f);\n"
+"	//float4 	a=a1;\n"
+"	//float4 	b=b1;\n"
+"	return cross(a,b);\n"
+"}\n"
+"__inline\n"
+"float dot3F4(float4 a, float4 b)\n"
+"{\n"
+"	float4 a1 = make_float4(a.xyz,0.f);\n"
+"	float4 b1 = make_float4(b.xyz,0.f);\n"
+"	return dot(a1, b1);\n"
+"}\n"
+"__inline\n"
+"float length3(const float4 a)\n"
+"{\n"
+"	return sqrtf(dot3F4(a,a));\n"
+"}\n"
+"__inline\n"
+"float dot4(const float4 a, const float4 b)\n"
+"{\n"
+"	return dot( a, b );\n"
+"}\n"
+"//	for height\n"
+"__inline\n"
+"float dot3w1(const float4 point, const float4 eqn)\n"
+"{\n"
+"	return dot3F4(point,eqn) + eqn.w;\n"
+"}\n"
+"__inline\n"
+"float4 normalize3(const float4 a)\n"
+"{\n"
+"	float4 n = make_float4(a.x, a.y, a.z, 0.f);\n"
+"	return fastNormalize4( n );\n"
+"//	float length = sqrtf(dot3F4(a, a));\n"
+"//	return 1.f/length * a;\n"
+"}\n"
+"__inline\n"
+"float4 normalize4(const float4 a)\n"
+"{\n"
+"	float length = sqrtf(dot4(a, a));\n"
+"	return 1.f/length * a;\n"
+"}\n"
+"__inline\n"
+"float4 createEquation(const float4 a, const float4 b, const float4 c)\n"
+"{\n"
+"	float4 eqn;\n"
+"	float4 ab = b-a;\n"
+"	float4 ac = c-a;\n"
+"	eqn = normalize3( cross3(ab, ac) );\n"
+"	eqn.w = -dot3F4(eqn,a);\n"
+"	return eqn;\n"
+"}\n"
+"///////////////////////////////////////\n"
+"//	Matrix3x3\n"
+"///////////////////////////////////////\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_row[3];\n"
+"}Matrix3x3;\n"
+"__inline\n"
+"Matrix3x3 mtZero();\n"
+"__inline\n"
+"Matrix3x3 mtIdentity();\n"
+"__inline\n"
+"Matrix3x3 mtTranspose(Matrix3x3 m);\n"
+"__inline\n"
+"Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b);\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b);\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b);\n"
+"__inline\n"
+"Matrix3x3 mtZero()\n"
+"{\n"
+"	Matrix3x3 m;\n"
+"	m.m_row[0] = (float4)(0.f);\n"
+"	m.m_row[1] = (float4)(0.f);\n"
+"	m.m_row[2] = (float4)(0.f);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"Matrix3x3 mtIdentity()\n"
+"{\n"
+"	Matrix3x3 m;\n"
+"	m.m_row[0] = (float4)(1,0,0,0);\n"
+"	m.m_row[1] = (float4)(0,1,0,0);\n"
+"	m.m_row[2] = (float4)(0,0,1,0);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"Matrix3x3 mtTranspose(Matrix3x3 m)\n"
+"{\n"
+"	Matrix3x3 out;\n"
+"	out.m_row[0] = (float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"Matrix3x3 mtMul(Matrix3x3 a, Matrix3x3 b)\n"
+"{\n"
+"	Matrix3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	Matrix3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"float4 mtMul1(Matrix3x3 a, float4 b)\n"
+"{\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a.m_row[0], b );\n"
+"	ans.y = dot3F4( a.m_row[1], b );\n"
+"	ans.z = dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"float4 mtMul3(float4 a, Matrix3x3 b)\n"
+"{\n"
+"	float4 colx = make_float4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	float4 coly = make_float4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	float4 colz = make_float4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"	float4 ans;\n"
+"	ans.x = dot3F4( a, colx );\n"
+"	ans.y = dot3F4( a, coly );\n"
+"	ans.z = dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"///////////////////////////////////////\n"
+"//	Quaternion\n"
+"///////////////////////////////////////\n"
+"typedef float4 Quaternion;\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b);\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in);\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec);\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q);\n"
+"__inline\n"
+"Quaternion qtMul(Quaternion a, Quaternion b)\n"
+"{\n"
+"	Quaternion ans;\n"
+"	ans = cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtNormalize(Quaternion in)\n"
+"{\n"
+"	return fastNormalize4(in);\n"
+"//	in /= length( in );\n"
+"//	return in;\n"
+"}\n"
+"__inline\n"
+"float4 qtRotate(Quaternion q, float4 vec)\n"
+"{\n"
+"	Quaternion qInv = qtInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = qtMul(qtMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"Quaternion qtInvert(Quaternion q)\n"
+"{\n"
+"	return (Quaternion)(-q.xyz, q.w);\n"
+"}\n"
+"__inline\n"
+"float4 qtInvRotate(const Quaternion q, float4 vec)\n"
+"{\n"
+"	return qtRotate( qtInvert( q ), vec );\n"
+"}\n"
+"#define WG_SIZE 64\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_pos;\n"
+"	Quaternion m_quat;\n"
+"	float4 m_linVel;\n"
+"	float4 m_angVel;\n"
+"	u32 m_shapeIdx;\n"
+"	float m_invMass;\n"
+"	float m_restituitionCoeff;\n"
+"	float m_frictionCoeff;\n"
+"} Body;\n"
+"typedef struct\n"
+"{\n"
+"	Matrix3x3 m_invInertia;\n"
+"	Matrix3x3 m_initInvInertia;\n"
+"} Shape;\n"
+"typedef struct\n"
+"{\n"
+"	float4 m_linear;\n"
+"	float4 m_worldPos[4];\n"
+"	float4 m_center;	\n"
+"	float m_jacCoeffInv[4];\n"
+"	float m_b[4];\n"
+"	float m_appliedRambdaDt[4];\n"
+"	float m_fJacCoeffInv[2];	\n"
+"	float m_fAppliedRambdaDt[2];	\n"
+"	u32 m_bodyA;\n"
+"	u32 m_bodyB;\n"
+"	int m_batchIdx;\n"
+"	u32 m_paddings;\n"
+"} Constraint4;\n"
+"__kernel void CountBodiesKernel(__global struct b3Contact4Data* manifoldPtr, __global unsigned int* bodyCount, __global int2* contactConstraintOffsets, int numContactManifolds, int fixedBodyIndex)\n"
+"{\n"
+"	int i = GET_GLOBAL_IDX;\n"
+"	\n"
+"	if( i < numContactManifolds)\n"
+"	{\n"
+"		int pa = manifoldPtr[i].m_bodyAPtrAndSignBit;\n"
+"		bool isFixedA = (pa <0) || (pa == fixedBodyIndex);\n"
+"		int bodyIndexA = abs(pa);\n"
+"		if (!isFixedA)\n"
+"		{\n"
+"			 AtomInc1(bodyCount[bodyIndexA],contactConstraintOffsets[i].x);\n"
+"		}\n"
+"		barrier(CLK_GLOBAL_MEM_FENCE);\n"
+"		int pb = manifoldPtr[i].m_bodyBPtrAndSignBit;\n"
+"		bool isFixedB = (pb <0) || (pb == fixedBodyIndex);\n"
+"		int bodyIndexB = abs(pb);\n"
+"		if (!isFixedB)\n"
+"		{\n"
+"			AtomInc1(bodyCount[bodyIndexB],contactConstraintOffsets[i].y);\n"
+"		} \n"
+"	}\n"
+"}\n"
+"__kernel void ClearVelocitiesKernel(__global float4* linearVelocities,__global float4* angularVelocities, int numSplitBodies)\n"
+"{\n"
+"	int i = GET_GLOBAL_IDX;\n"
+"	\n"
+"	if( i < numSplitBodies)\n"
+"	{\n"
+"		linearVelocities[i] = make_float4(0);\n"
+"		angularVelocities[i] = make_float4(0);\n"
+"	}\n"
+"}\n"
+"__kernel void AverageVelocitiesKernel(__global Body* gBodies,__global int* offsetSplitBodies,__global const unsigned int* bodyCount,\n"
+"__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities, int numBodies)\n"
+"{\n"
+"	int i = GET_GLOBAL_IDX;\n"
+"	if (i<numBodies)\n"
+"	{\n"
+"		if (gBodies[i].m_invMass)\n"
+"		{\n"
+"			int bodyOffset = offsetSplitBodies[i];\n"
+"			int count = bodyCount[i];\n"
+"			float factor = 1.f/((float)count);\n"
+"			float4 averageLinVel = make_float4(0.f);\n"
+"			float4 averageAngVel = make_float4(0.f);\n"
+"			\n"
+"			for (int j=0;j<count;j++)\n"
+"			{\n"
+"				averageLinVel += deltaLinearVelocities[bodyOffset+j]*factor;\n"
+"				averageAngVel += deltaAngularVelocities[bodyOffset+j]*factor;\n"
+"			}\n"
+"			\n"
+"			for (int j=0;j<count;j++)\n"
+"			{\n"
+"				deltaLinearVelocities[bodyOffset+j] = averageLinVel;\n"
+"				deltaAngularVelocities[bodyOffset+j] = averageAngVel;\n"
+"			}\n"
+"			\n"
+"		}//bodies[i].m_invMass\n"
+"	}//i<numBodies\n"
+"}\n"
+"void setLinearAndAngular( float4 n, float4 r0, float4 r1, float4* linear, float4* angular0, float4* angular1)\n"
+"{\n"
+"	*linear = make_float4(n.xyz,0.f);\n"
+"	*angular0 = cross3(r0, n);\n"
+"	*angular1 = -cross3(r1, n);\n"
+"}\n"
+"float calcRelVel( float4 l0, float4 l1, float4 a0, float4 a1, float4 linVel0, float4 angVel0, float4 linVel1, float4 angVel1 )\n"
+"{\n"
+"	return dot3F4(l0, linVel0) + dot3F4(a0, angVel0) + dot3F4(l1, linVel1) + dot3F4(a1, angVel1);\n"
+"}\n"
+"float calcJacCoeff(const float4 linear0, const float4 linear1, const float4 angular0, const float4 angular1,\n"
+"					float invMass0, const Matrix3x3* invInertia0, float invMass1, const Matrix3x3* invInertia1, float countA, float countB)\n"
+"{\n"
+"	//	linear0,1 are normlized\n"
+"	float jmj0 = invMass0;//dot3F4(linear0, linear0)*invMass0;\n"
+"	float jmj1 = dot3F4(mtMul3(angular0,*invInertia0), angular0);\n"
+"	float jmj2 = invMass1;//dot3F4(linear1, linear1)*invMass1;\n"
+"	float jmj3 = dot3F4(mtMul3(angular1,*invInertia1), angular1);\n"
+"	return -1.f/((jmj0+jmj1)*countA+(jmj2+jmj3)*countB);\n"
+"}\n"
+"void btPlaneSpace1 (float4 n, float4* p, float4* q);\n"
+" void btPlaneSpace1 (float4 n, float4* p, float4* q)\n"
+"{\n"
+"  if (fabs(n.z) > 0.70710678f) {\n"
+"    // choose p in y-z plane\n"
+"    float a = n.y*n.y + n.z*n.z;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = 0;\n"
+"	p[0].y = -n.z*k;\n"
+"	p[0].z = n.y*k;\n"
+"    // set q = n x p\n"
+"    q[0].x = a*k;\n"
+"	q[0].y = -n.x*p[0].z;\n"
+"	q[0].z = n.x*p[0].y;\n"
+"  }\n"
+"  else {\n"
+"    // choose p in x-y plane\n"
+"    float a = n.x*n.x + n.y*n.y;\n"
+"    float k = 1.f/sqrt(a);\n"
+"    p[0].x = -n.y*k;\n"
+"	p[0].y = n.x*k;\n"
+"	p[0].z = 0;\n"
+"    // set q = n x p\n"
+"    q[0].x = -n.z*p[0].y;\n"
+"	q[0].y = n.z*p[0].x;\n"
+"	q[0].z = a*k;\n"
+"  }\n"
+"}\n"
+"void solveContact(__global Constraint4* cs,\n"
+"			float4 posA, float4* linVelA, float4* angVelA, float invMassA, Matrix3x3 invInertiaA,\n"
+"			float4 posB, float4* linVelB, float4* angVelB, float invMassB, Matrix3x3 invInertiaB,\n"
+"			float4* dLinVelA, float4* dAngVelA, float4* dLinVelB, float4* dAngVelB)\n"
+"{\n"
+"	float minRambdaDt = 0;\n"
+"	float maxRambdaDt = FLT_MAX;\n"
+"	for(int ic=0; ic<4; ic++)\n"
+"	{\n"
+"		if( cs->m_jacCoeffInv[ic] == 0.f ) continue;\n"
+"		float4 angular0, angular1, linear;\n"
+"		float4 r0 = cs->m_worldPos[ic] - posA;\n"
+"		float4 r1 = cs->m_worldPos[ic] - posB;\n"
+"		setLinearAndAngular( cs->m_linear, r0, r1, &linear, &angular0, &angular1 );\n"
+"	\n"
+"		float rambdaDt = calcRelVel( cs->m_linear, -cs->m_linear, angular0, angular1, \n"
+"			*linVelA+*dLinVelA, *angVelA+*dAngVelA, *linVelB+*dLinVelB, *angVelB+*dAngVelB ) + cs->m_b[ic];\n"
+"		rambdaDt *= cs->m_jacCoeffInv[ic];\n"
+"		\n"
+"		{\n"
+"			float prevSum = cs->m_appliedRambdaDt[ic];\n"
+"			float updated = prevSum;\n"
+"			updated += rambdaDt;\n"
+"			updated = max2( updated, minRambdaDt );\n"
+"			updated = min2( updated, maxRambdaDt );\n"
+"			rambdaDt = updated - prevSum;\n"
+"			cs->m_appliedRambdaDt[ic] = updated;\n"
+"		}\n"
+"			\n"
+"		float4 linImp0 = invMassA*linear*rambdaDt;\n"
+"		float4 linImp1 = invMassB*(-linear)*rambdaDt;\n"
+"		float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;\n"
+"		float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;\n"
+"		\n"
+"		if (invMassA)\n"
+"		{\n"
+"			*dLinVelA += linImp0;\n"
+"			*dAngVelA += angImp0;\n"
+"		}\n"
+"		if (invMassB)\n"
+"		{\n"
+"			*dLinVelB += linImp1;\n"
+"			*dAngVelB += angImp1;\n"
+"		}\n"
+"	}\n"
+"}\n"
+"//	solveContactConstraint( gBodies, gShapes, &gConstraints[i] ,contactConstraintOffsets,offsetSplitBodies, deltaLinearVelocities, deltaAngularVelocities);\n"
+"void solveContactConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs, \n"
+"__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,\n"
+"__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities)\n"
+"{\n"
+"	//float frictionCoeff = ldsCs[0].m_linear.w;\n"
+"	int aIdx = ldsCs[0].m_bodyA;\n"
+"	int bIdx = ldsCs[0].m_bodyB;\n"
+"	float4 posA = gBodies[aIdx].m_pos;\n"
+"	float4 linVelA = gBodies[aIdx].m_linVel;\n"
+"	float4 angVelA = gBodies[aIdx].m_angVel;\n"
+"	float invMassA = gBodies[aIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;\n"
+"	float4 posB = gBodies[bIdx].m_pos;\n"
+"	float4 linVelB = gBodies[bIdx].m_linVel;\n"
+"	float4 angVelB = gBodies[bIdx].m_angVel;\n"
+"	float invMassB = gBodies[bIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;\n"
+"			\n"
+"	float4 dLinVelA = make_float4(0,0,0,0);\n"
+"	float4 dAngVelA = make_float4(0,0,0,0);\n"
+"	float4 dLinVelB = make_float4(0,0,0,0);\n"
+"	float4 dAngVelB = make_float4(0,0,0,0);\n"
+"			\n"
+"	int bodyOffsetA = offsetSplitBodies[aIdx];\n"
+"	int constraintOffsetA = contactConstraintOffsets[0].x;\n"
+"	int splitIndexA = bodyOffsetA+constraintOffsetA;\n"
+"	\n"
+"	if (invMassA)\n"
+"	{\n"
+"		dLinVelA = deltaLinearVelocities[splitIndexA];\n"
+"		dAngVelA = deltaAngularVelocities[splitIndexA];\n"
+"	}\n"
+"	int bodyOffsetB = offsetSplitBodies[bIdx];\n"
+"	int constraintOffsetB = contactConstraintOffsets[0].y;\n"
+"	int splitIndexB= bodyOffsetB+constraintOffsetB;\n"
+"	if (invMassB)\n"
+"	{\n"
+"		dLinVelB = deltaLinearVelocities[splitIndexB];\n"
+"		dAngVelB = deltaAngularVelocities[splitIndexB];\n"
+"	}\n"
+"	solveContact( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,\n"
+"			posB, &linVelB, &angVelB, invMassB, invInertiaB ,&dLinVelA, &dAngVelA, &dLinVelB, &dAngVelB);\n"
+"	if (invMassA)\n"
+"	{\n"
+"		deltaLinearVelocities[splitIndexA] = dLinVelA;\n"
+"		deltaAngularVelocities[splitIndexA] = dAngVelA;\n"
+"	} \n"
+"	if (invMassB)\n"
+"	{\n"
+"		deltaLinearVelocities[splitIndexB] = dLinVelB;\n"
+"		deltaAngularVelocities[splitIndexB] = dAngVelB;\n"
+"	}\n"
+"}\n"
+"__kernel void SolveContactJacobiKernel(__global Constraint4* gConstraints, __global Body* gBodies, __global Shape* gShapes ,\n"
+"__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities,\n"
+"float deltaTime, float positionDrift, float positionConstraintCoeff, int fixedBodyIndex, int numManifolds\n"
+")\n"
+"{\n"
+"	int i = GET_GLOBAL_IDX;\n"
+"	if (i<numManifolds)\n"
+"	{\n"
+"		solveContactConstraint( gBodies, gShapes, &gConstraints[i] ,&contactConstraintOffsets[i],offsetSplitBodies, deltaLinearVelocities, deltaAngularVelocities);\n"
+"	}\n"
+"}\n"
+"void solveFrictionConstraint(__global Body* gBodies, __global Shape* gShapes, __global Constraint4* ldsCs,\n"
+"							__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,\n"
+"							__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities)\n"
+"{\n"
+"	float frictionCoeff = 0.7f;//ldsCs[0].m_linear.w;\n"
+"	int aIdx = ldsCs[0].m_bodyA;\n"
+"	int bIdx = ldsCs[0].m_bodyB;\n"
+"	float4 posA = gBodies[aIdx].m_pos;\n"
+"	float4 linVelA = gBodies[aIdx].m_linVel;\n"
+"	float4 angVelA = gBodies[aIdx].m_angVel;\n"
+"	float invMassA = gBodies[aIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;\n"
+"	float4 posB = gBodies[bIdx].m_pos;\n"
+"	float4 linVelB = gBodies[bIdx].m_linVel;\n"
+"	float4 angVelB = gBodies[bIdx].m_angVel;\n"
+"	float invMassB = gBodies[bIdx].m_invMass;\n"
+"	Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;\n"
+"	\n"
+"	float4 dLinVelA = make_float4(0,0,0,0);\n"
+"	float4 dAngVelA = make_float4(0,0,0,0);\n"
+"	float4 dLinVelB = make_float4(0,0,0,0);\n"
+"	float4 dAngVelB = make_float4(0,0,0,0);\n"
+"			\n"
+"	int bodyOffsetA = offsetSplitBodies[aIdx];\n"
+"	int constraintOffsetA = contactConstraintOffsets[0].x;\n"
+"	int splitIndexA = bodyOffsetA+constraintOffsetA;\n"
+"	\n"
+"	if (invMassA)\n"
+"	{\n"
+"		dLinVelA = deltaLinearVelocities[splitIndexA];\n"
+"		dAngVelA = deltaAngularVelocities[splitIndexA];\n"
+"	}\n"
+"	int bodyOffsetB = offsetSplitBodies[bIdx];\n"
+"	int constraintOffsetB = contactConstraintOffsets[0].y;\n"
+"	int splitIndexB= bodyOffsetB+constraintOffsetB;\n"
+"	if (invMassB)\n"
+"	{\n"
+"		dLinVelB = deltaLinearVelocities[splitIndexB];\n"
+"		dAngVelB = deltaAngularVelocities[splitIndexB];\n"
+"	}\n"
+"	{\n"
+"		float maxRambdaDt[4] = {FLT_MAX,FLT_MAX,FLT_MAX,FLT_MAX};\n"
+"		float minRambdaDt[4] = {0.f,0.f,0.f,0.f};\n"
+"		float sum = 0;\n"
+"		for(int j=0; j<4; j++)\n"
+"		{\n"
+"			sum +=ldsCs[0].m_appliedRambdaDt[j];\n"
+"		}\n"
+"		frictionCoeff = 0.7f;\n"
+"		for(int j=0; j<4; j++)\n"
+"		{\n"
+"			maxRambdaDt[j] = frictionCoeff*sum;\n"
+"			minRambdaDt[j] = -maxRambdaDt[j];\n"
+"		}\n"
+"		\n"
+"//		solveFriction( ldsCs, posA, &linVelA, &angVelA, invMassA, invInertiaA,\n"
+"//			posB, &linVelB, &angVelB, invMassB, invInertiaB, maxRambdaDt, minRambdaDt );\n"
+"		\n"
+"		\n"
+"		{\n"
+"			\n"
+"			__global Constraint4* cs = ldsCs;\n"
+"			\n"
+"			if( cs->m_fJacCoeffInv[0] == 0 && cs->m_fJacCoeffInv[0] == 0 ) return;\n"
+"			const float4 center = cs->m_center;\n"
+"			\n"
+"			float4 n = -cs->m_linear;\n"
+"			\n"
+"			float4 tangent[2];\n"
+"			btPlaneSpace1(n,&tangent[0],&tangent[1]);\n"
+"			float4 angular0, angular1, linear;\n"
+"			float4 r0 = center - posA;\n"
+"			float4 r1 = center - posB;\n"
+"			for(int i=0; i<2; i++)\n"
+"			{\n"
+"				setLinearAndAngular( tangent[i], r0, r1, &linear, &angular0, &angular1 );\n"
+"				float rambdaDt = calcRelVel(linear, -linear, angular0, angular1,\n"
+"											linVelA+dLinVelA, angVelA+dAngVelA, linVelB+dLinVelB, angVelB+dAngVelB );\n"
+"				rambdaDt *= cs->m_fJacCoeffInv[i];\n"
+"				\n"
+"				{\n"
+"					float prevSum = cs->m_fAppliedRambdaDt[i];\n"
+"					float updated = prevSum;\n"
+"					updated += rambdaDt;\n"
+"					updated = max2( updated, minRambdaDt[i] );\n"
+"					updated = min2( updated, maxRambdaDt[i] );\n"
+"					rambdaDt = updated - prevSum;\n"
+"					cs->m_fAppliedRambdaDt[i] = updated;\n"
+"				}\n"
+"				\n"
+"				float4 linImp0 = invMassA*linear*rambdaDt;\n"
+"				float4 linImp1 = invMassB*(-linear)*rambdaDt;\n"
+"				float4 angImp0 = mtMul1(invInertiaA, angular0)*rambdaDt;\n"
+"				float4 angImp1 = mtMul1(invInertiaB, angular1)*rambdaDt;\n"
+"				\n"
+"				dLinVelA += linImp0;\n"
+"				dAngVelA += angImp0;\n"
+"				dLinVelB += linImp1;\n"
+"				dAngVelB += angImp1;\n"
+"			}\n"
+"			{	//	angular damping for point constraint\n"
+"				float4 ab = normalize3( posB - posA );\n"
+"				float4 ac = normalize3( center - posA );\n"
+"				if( dot3F4( ab, ac ) > 0.95f  || (invMassA == 0.f || invMassB == 0.f))\n"
+"				{\n"
+"					float angNA = dot3F4( n, angVelA );\n"
+"					float angNB = dot3F4( n, angVelB );\n"
+"					\n"
+"					dAngVelA -= (angNA*0.1f)*n;\n"
+"					dAngVelB -= (angNB*0.1f)*n;\n"
+"				}\n"
+"			}\n"
+"		}\n"
+"		\n"
+"		\n"
+"	}\n"
+"	if (invMassA)\n"
+"	{\n"
+"		deltaLinearVelocities[splitIndexA] = dLinVelA;\n"
+"		deltaAngularVelocities[splitIndexA] = dAngVelA;\n"
+"	} \n"
+"	if (invMassB)\n"
+"	{\n"
+"		deltaLinearVelocities[splitIndexB] = dLinVelB;\n"
+"		deltaAngularVelocities[splitIndexB] = dAngVelB;\n"
+"	}\n"
+" \n"
+"}\n"
+"__kernel void SolveFrictionJacobiKernel(__global Constraint4* gConstraints, __global Body* gBodies, __global Shape* gShapes ,\n"
+"										__global int2* contactConstraintOffsets,__global unsigned int* offsetSplitBodies,\n"
+"										__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities,\n"
+"										float deltaTime, float positionDrift, float positionConstraintCoeff, int fixedBodyIndex, int numManifolds\n"
+")\n"
+"{\n"
+"	int i = GET_GLOBAL_IDX;\n"
+"	if (i<numManifolds)\n"
+"	{\n"
+"		solveFrictionConstraint( gBodies, gShapes, &gConstraints[i] ,&contactConstraintOffsets[i],offsetSplitBodies, deltaLinearVelocities, deltaAngularVelocities);\n"
+"	}\n"
+"}\n"
+"__kernel void UpdateBodyVelocitiesKernel(__global Body* gBodies,__global int* offsetSplitBodies,__global const unsigned int* bodyCount,\n"
+"									__global float4* deltaLinearVelocities, __global float4* deltaAngularVelocities, int numBodies)\n"
+"{\n"
+"	int i = GET_GLOBAL_IDX;\n"
+"	if (i<numBodies)\n"
+"	{\n"
+"		if (gBodies[i].m_invMass)\n"
+"		{\n"
+"			int bodyOffset = offsetSplitBodies[i];\n"
+"			int count = bodyCount[i];\n"
+"			if (count)\n"
+"			{\n"
+"				gBodies[i].m_linVel += deltaLinearVelocities[bodyOffset];\n"
+"				gBodies[i].m_angVel += deltaAngularVelocities[bodyOffset];\n"
+"			}\n"
+"		}\n"
+"	}\n"
+"}\n"
+"void setConstraint4( const float4 posA, const float4 linVelA, const float4 angVelA, float invMassA, const Matrix3x3 invInertiaA,\n"
+"	const float4 posB, const float4 linVelB, const float4 angVelB, float invMassB, const Matrix3x3 invInertiaB, \n"
+"	__global struct b3Contact4Data* src, float dt, float positionDrift, float positionConstraintCoeff,float countA, float countB,\n"
+"	Constraint4* dstC )\n"
+"{\n"
+"	dstC->m_bodyA = abs(src->m_bodyAPtrAndSignBit);\n"
+"	dstC->m_bodyB = abs(src->m_bodyBPtrAndSignBit);\n"
+"	float dtInv = 1.f/dt;\n"
+"	for(int ic=0; ic<4; ic++)\n"
+"	{\n"
+"		dstC->m_appliedRambdaDt[ic] = 0.f;\n"
+"	}\n"
+"	dstC->m_fJacCoeffInv[0] = dstC->m_fJacCoeffInv[1] = 0.f;\n"
+"	dstC->m_linear = src->m_worldNormalOnB;\n"
+"	dstC->m_linear.w = 0.7f ;//src->getFrictionCoeff() );\n"
+"	for(int ic=0; ic<4; ic++)\n"
+"	{\n"
+"		float4 r0 = src->m_worldPosB[ic] - posA;\n"
+"		float4 r1 = src->m_worldPosB[ic] - posB;\n"
+"		if( ic >= src->m_worldNormalOnB.w )//npoints\n"
+"		{\n"
+"			dstC->m_jacCoeffInv[ic] = 0.f;\n"
+"			continue;\n"
+"		}\n"
+"		float relVelN;\n"
+"		{\n"
+"			float4 linear, angular0, angular1;\n"
+"			setLinearAndAngular(src->m_worldNormalOnB, r0, r1, &linear, &angular0, &angular1);\n"
+"			dstC->m_jacCoeffInv[ic] = calcJacCoeff(linear, -linear, angular0, angular1,\n"
+"				invMassA, &invInertiaA, invMassB, &invInertiaB , countA, countB);\n"
+"			relVelN = calcRelVel(linear, -linear, angular0, angular1,\n"
+"				linVelA, angVelA, linVelB, angVelB);\n"
+"			float e = 0.f;//src->getRestituitionCoeff();\n"
+"			if( relVelN*relVelN < 0.004f ) e = 0.f;\n"
+"			dstC->m_b[ic] = e*relVelN;\n"
+"			//float penetration = src->m_worldPosB[ic].w;\n"
+"			dstC->m_b[ic] += (src->m_worldPosB[ic].w + positionDrift)*positionConstraintCoeff*dtInv;\n"
+"			dstC->m_appliedRambdaDt[ic] = 0.f;\n"
+"		}\n"
+"	}\n"
+"	if( src->m_worldNormalOnB.w > 0 )//npoints\n"
+"	{	//	prepare friction\n"
+"		float4 center = make_float4(0.f);\n"
+"		for(int i=0; i<src->m_worldNormalOnB.w; i++) \n"
+"			center += src->m_worldPosB[i];\n"
+"		center /= (float)src->m_worldNormalOnB.w;\n"
+"		float4 tangent[2];\n"
+"		btPlaneSpace1(-src->m_worldNormalOnB,&tangent[0],&tangent[1]);\n"
+"		\n"
+"		float4 r[2];\n"
+"		r[0] = center - posA;\n"
+"		r[1] = center - posB;\n"
+"		for(int i=0; i<2; i++)\n"
+"		{\n"
+"			float4 linear, angular0, angular1;\n"
+"			setLinearAndAngular(tangent[i], r[0], r[1], &linear, &angular0, &angular1);\n"
+"			dstC->m_fJacCoeffInv[i] = calcJacCoeff(linear, -linear, angular0, angular1,\n"
+"				invMassA, &invInertiaA, invMassB, &invInertiaB ,countA, countB);\n"
+"			dstC->m_fAppliedRambdaDt[i] = 0.f;\n"
+"		}\n"
+"		dstC->m_center = center;\n"
+"	}\n"
+"	for(int i=0; i<4; i++)\n"
+"	{\n"
+"		if( i<src->m_worldNormalOnB.w )\n"
+"		{\n"
+"			dstC->m_worldPos[i] = src->m_worldPosB[i];\n"
+"		}\n"
+"		else\n"
+"		{\n"
+"			dstC->m_worldPos[i] = make_float4(0.f);\n"
+"		}\n"
+"	}\n"
+"}\n"
+"__kernel\n"
+"__attribute__((reqd_work_group_size(WG_SIZE,1,1)))\n"
+"void ContactToConstraintSplitKernel(__global const struct b3Contact4Data* gContact, __global const Body* gBodies, __global const Shape* gShapes, __global Constraint4* gConstraintOut, \n"
+"__global const unsigned int* bodyCount,\n"
+"int nContacts,\n"
+"float dt,\n"
+"float positionDrift,\n"
+"float positionConstraintCoeff\n"
+")\n"
+"{\n"
+"	int gIdx = GET_GLOBAL_IDX;\n"
+"	\n"
+"	if( gIdx < nContacts )\n"
+"	{\n"
+"		int aIdx = abs(gContact[gIdx].m_bodyAPtrAndSignBit);\n"
+"		int bIdx = abs(gContact[gIdx].m_bodyBPtrAndSignBit);\n"
+"		float4 posA = gBodies[aIdx].m_pos;\n"
+"		float4 linVelA = gBodies[aIdx].m_linVel;\n"
+"		float4 angVelA = gBodies[aIdx].m_angVel;\n"
+"		float invMassA = gBodies[aIdx].m_invMass;\n"
+"		Matrix3x3 invInertiaA = gShapes[aIdx].m_invInertia;\n"
+"		float4 posB = gBodies[bIdx].m_pos;\n"
+"		float4 linVelB = gBodies[bIdx].m_linVel;\n"
+"		float4 angVelB = gBodies[bIdx].m_angVel;\n"
+"		float invMassB = gBodies[bIdx].m_invMass;\n"
+"		Matrix3x3 invInertiaB = gShapes[bIdx].m_invInertia;\n"
+"		Constraint4 cs;\n"
+"		float countA = invMassA != 0.f ? (float)bodyCount[aIdx] : 1;\n"
+"		float countB = invMassB != 0.f ? (float)bodyCount[bIdx] : 1;\n"
+"    	setConstraint4( posA, linVelA, angVelA, invMassA, invInertiaA, posB, linVelB, angVelB, invMassB, invInertiaB,\n"
+"			&gContact[gIdx], dt, positionDrift, positionConstraintCoeff,countA,countB,\n"
+"			&cs  );\n"
+"		\n"
+"		cs.m_batchIdx = gContact[gIdx].m_batchIdx;\n"
+"		gConstraintOut[gIdx] = cs;\n"
+"	}\n"
+"}\n"
+;
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/updateAabbsKernel.cl b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/updateAabbsKernel.cl
new file mode 100644
index 0000000000..ba8ba735d0
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/updateAabbsKernel.cl
@@ -0,0 +1,22 @@
+
+
+#include "Bullet3Collision/NarrowPhaseCollision/shared/b3UpdateAabbs.h"
+
+
+__kernel void initializeGpuAabbsFull(  const int numNodes, __global b3RigidBodyData_t* gBodies,__global b3Collidable_t* collidables, __global b3Aabb_t* plocalShapeAABB, __global b3Aabb_t* pAABB)
+{
+	int nodeID = get_global_id(0);
+	if( nodeID < numNodes )
+	{
+		b3ComputeWorldAabb(nodeID, gBodies, collidables, plocalShapeAABB,pAABB);
+	}
+}
+
+__kernel void clearOverlappingPairsKernel(  __global int4* pairs, int numPairs)
+{
+	int pairId = get_global_id(0);
+	if( pairId< numPairs )
+	{
+		pairs[pairId].z = 0xffffffff;
+	}
+}
\ No newline at end of file
diff --git a/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/updateAabbsKernel.h b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/updateAabbsKernel.h
new file mode 100644
index 0000000000..d70e74017a
--- /dev/null
+++ b/thirdparty/bullet/Bullet3OpenCL/RigidBody/kernels/updateAabbsKernel.h
@@ -0,0 +1,483 @@
+//this file is autogenerated using stringify.bat (premake --stringify) in the build folder of this project
+static const char* updateAabbsKernelCL= \
+"#ifndef B3_UPDATE_AABBS_H\n"
+"#define B3_UPDATE_AABBS_H\n"
+"#ifndef B3_AABB_H\n"
+"#define B3_AABB_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#define B3_FLOAT4_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#define B3_PLATFORM_DEFINITIONS_H\n"
+"struct MyTest\n"
+"{\n"
+"	int bla;\n"
+"};\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"//keep B3_LARGE_FLOAT*B3_LARGE_FLOAT < FLT_MAX\n"
+"#define B3_LARGE_FLOAT 1e18f\n"
+"#define B3_INFINITY 1e18f\n"
+"#define b3Assert(a)\n"
+"#define b3ConstArray(a) __global const a*\n"
+"#define b3AtomicInc atomic_inc\n"
+"#define b3AtomicAdd atomic_add\n"
+"#define b3Fabs fabs\n"
+"#define b3Sqrt native_sqrt\n"
+"#define b3Sin native_sin\n"
+"#define b3Cos native_cos\n"
+"#define B3_STATIC\n"
+"#endif\n"
+"#endif\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Float4;\n"
+"	#define b3Float4ConstArg const b3Float4\n"
+"	#define b3MakeFloat4 (float4)\n"
+"	float b3Dot3F4(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return dot(a1, b1);\n"
+"	}\n"
+"	b3Float4 b3Cross3(b3Float4ConstArg v0,b3Float4ConstArg v1)\n"
+"	{\n"
+"		float4 a1 = b3MakeFloat4(v0.xyz,0.f);\n"
+"		float4 b1 = b3MakeFloat4(v1.xyz,0.f);\n"
+"		return cross(a1, b1);\n"
+"	}\n"
+"	#define b3MinFloat4 min\n"
+"	#define b3MaxFloat4 max\n"
+"	#define b3Normalized(a) normalize(a)\n"
+"#endif \n"
+"		\n"
+"inline bool b3IsAlmostZero(b3Float4ConstArg v)\n"
+"{\n"
+"	if(b3Fabs(v.x)>1e-6 || b3Fabs(v.y)>1e-6 || b3Fabs(v.z)>1e-6)	\n"
+"		return false;\n"
+"	return true;\n"
+"}\n"
+"inline int    b3MaxDot( b3Float4ConstArg vec, __global const b3Float4* vecArray, int vecLen, float* dotOut )\n"
+"{\n"
+"    float maxDot = -B3_INFINITY;\n"
+"    int i = 0;\n"
+"    int ptIndex = -1;\n"
+"    for( i = 0; i < vecLen; i++ )\n"
+"    {\n"
+"        float dot = b3Dot3F4(vecArray[i],vec);\n"
+"            \n"
+"        if( dot > maxDot )\n"
+"        {\n"
+"            maxDot = dot;\n"
+"            ptIndex = i;\n"
+"        }\n"
+"    }\n"
+"	b3Assert(ptIndex>=0);\n"
+"    if (ptIndex<0)\n"
+"	{\n"
+"		ptIndex = 0;\n"
+"	}\n"
+"    *dotOut = maxDot;\n"
+"    return ptIndex;\n"
+"}\n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_MAT3x3_H\n"
+"#define B3_MAT3x3_H\n"
+"#ifndef B3_QUAT_H\n"
+"#define B3_QUAT_H\n"
+"#ifndef B3_PLATFORM_DEFINITIONS_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif\n"
+"#endif\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"	typedef float4	b3Quat;\n"
+"	#define b3QuatConstArg const b3Quat\n"
+"	\n"
+"	\n"
+"inline float4 b3FastNormalize4(float4 v)\n"
+"{\n"
+"	v = (float4)(v.xyz,0.f);\n"
+"	return fast_normalize(v);\n"
+"}\n"
+"	\n"
+"inline b3Quat b3QuatMul(b3Quat a, b3Quat b);\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in);\n"
+"inline b3Quat b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec);\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q);\n"
+"inline b3Quat b3QuatMul(b3QuatConstArg a, b3QuatConstArg b)\n"
+"{\n"
+"	b3Quat ans;\n"
+"	ans = b3Cross3( a, b );\n"
+"	ans += a.w*b+b.w*a;\n"
+"//	ans.w = a.w*b.w - (a.x*b.x+a.y*b.y+a.z*b.z);\n"
+"	ans.w = a.w*b.w - b3Dot3F4(a, b);\n"
+"	return ans;\n"
+"}\n"
+"inline b3Quat b3QuatNormalized(b3QuatConstArg in)\n"
+"{\n"
+"	b3Quat q;\n"
+"	q=in;\n"
+"	//return b3FastNormalize4(in);\n"
+"	float len = native_sqrt(dot(q, q));\n"
+"	if(len > 0.f)\n"
+"	{\n"
+"		q *= 1.f / len;\n"
+"	}\n"
+"	else\n"
+"	{\n"
+"		q.x = q.y = q.z = 0.f;\n"
+"		q.w = 1.f;\n"
+"	}\n"
+"	return q;\n"
+"}\n"
+"inline float4 b3QuatRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	b3Quat qInv = b3QuatInvert( q );\n"
+"	float4 vcpy = vec;\n"
+"	vcpy.w = 0.f;\n"
+"	float4 out = b3QuatMul(b3QuatMul(q,vcpy),qInv);\n"
+"	return out;\n"
+"}\n"
+"inline b3Quat b3QuatInverse(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline b3Quat b3QuatInvert(b3QuatConstArg q)\n"
+"{\n"
+"	return (b3Quat)(-q.xyz, q.w);\n"
+"}\n"
+"inline float4 b3QuatInvRotate(b3QuatConstArg q, b3QuatConstArg vec)\n"
+"{\n"
+"	return b3QuatRotate( b3QuatInvert( q ), vec );\n"
+"}\n"
+"inline b3Float4 b3TransformPoint(b3Float4ConstArg point, b3Float4ConstArg translation, b3QuatConstArg  orientation)\n"
+"{\n"
+"	return b3QuatRotate( orientation, point ) + (translation);\n"
+"}\n"
+"	\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"typedef struct\n"
+"{\n"
+"	b3Float4 m_row[3];\n"
+"}b3Mat3x3;\n"
+"#define b3Mat3x3ConstArg const b3Mat3x3\n"
+"#define b3GetRow(m,row) (m.m_row[row])\n"
+"inline b3Mat3x3 b3QuatGetRotationMatrix(b3Quat quat)\n"
+"{\n"
+"	b3Float4 quat2 = (b3Float4)(quat.x*quat.x, quat.y*quat.y, quat.z*quat.z, 0.f);\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0].x=1-2*quat2.y-2*quat2.z;\n"
+"	out.m_row[0].y=2*quat.x*quat.y-2*quat.w*quat.z;\n"
+"	out.m_row[0].z=2*quat.x*quat.z+2*quat.w*quat.y;\n"
+"	out.m_row[0].w = 0.f;\n"
+"	out.m_row[1].x=2*quat.x*quat.y+2*quat.w*quat.z;\n"
+"	out.m_row[1].y=1-2*quat2.x-2*quat2.z;\n"
+"	out.m_row[1].z=2*quat.y*quat.z-2*quat.w*quat.x;\n"
+"	out.m_row[1].w = 0.f;\n"
+"	out.m_row[2].x=2*quat.x*quat.z-2*quat.w*quat.y;\n"
+"	out.m_row[2].y=2*quat.y*quat.z+2*quat.w*quat.x;\n"
+"	out.m_row[2].z=1-2*quat2.x-2*quat2.y;\n"
+"	out.m_row[2].w = 0.f;\n"
+"	return out;\n"
+"}\n"
+"inline b3Mat3x3 b3AbsoluteMat3x3(b3Mat3x3ConstArg matIn)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = fabs(matIn.m_row[0]);\n"
+"	out.m_row[1] = fabs(matIn.m_row[1]);\n"
+"	out.m_row[2] = fabs(matIn.m_row[2]);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtZero();\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity();\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m);\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b);\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b);\n"
+"__inline\n"
+"b3Mat3x3 mtZero()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(0.f);\n"
+"	m.m_row[1] = (b3Float4)(0.f);\n"
+"	m.m_row[2] = (b3Float4)(0.f);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtIdentity()\n"
+"{\n"
+"	b3Mat3x3 m;\n"
+"	m.m_row[0] = (b3Float4)(1,0,0,0);\n"
+"	m.m_row[1] = (b3Float4)(0,1,0,0);\n"
+"	m.m_row[2] = (b3Float4)(0,0,1,0);\n"
+"	return m;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtTranspose(b3Mat3x3 m)\n"
+"{\n"
+"	b3Mat3x3 out;\n"
+"	out.m_row[0] = (b3Float4)(m.m_row[0].x, m.m_row[1].x, m.m_row[2].x, 0.f);\n"
+"	out.m_row[1] = (b3Float4)(m.m_row[0].y, m.m_row[1].y, m.m_row[2].y, 0.f);\n"
+"	out.m_row[2] = (b3Float4)(m.m_row[0].z, m.m_row[1].z, m.m_row[2].z, 0.f);\n"
+"	return out;\n"
+"}\n"
+"__inline\n"
+"b3Mat3x3 mtMul(b3Mat3x3 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Mat3x3 transB;\n"
+"	transB = mtTranspose( b );\n"
+"	b3Mat3x3 ans;\n"
+"	//	why this doesn't run when 0ing in the for{}\n"
+"	a.m_row[0].w = 0.f;\n"
+"	a.m_row[1].w = 0.f;\n"
+"	a.m_row[2].w = 0.f;\n"
+"	for(int i=0; i<3; i++)\n"
+"	{\n"
+"//	a.m_row[i].w = 0.f;\n"
+"		ans.m_row[i].x = b3Dot3F4(a.m_row[i],transB.m_row[0]);\n"
+"		ans.m_row[i].y = b3Dot3F4(a.m_row[i],transB.m_row[1]);\n"
+"		ans.m_row[i].z = b3Dot3F4(a.m_row[i],transB.m_row[2]);\n"
+"		ans.m_row[i].w = 0.f;\n"
+"	}\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul1(b3Mat3x3 a, b3Float4 b)\n"
+"{\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a.m_row[0], b );\n"
+"	ans.y = b3Dot3F4( a.m_row[1], b );\n"
+"	ans.z = b3Dot3F4( a.m_row[2], b );\n"
+"	ans.w = 0.f;\n"
+"	return ans;\n"
+"}\n"
+"__inline\n"
+"b3Float4 mtMul3(b3Float4 a, b3Mat3x3 b)\n"
+"{\n"
+"	b3Float4 colx = b3MakeFloat4(b.m_row[0].x, b.m_row[1].x, b.m_row[2].x, 0);\n"
+"	b3Float4 coly = b3MakeFloat4(b.m_row[0].y, b.m_row[1].y, b.m_row[2].y, 0);\n"
+"	b3Float4 colz = b3MakeFloat4(b.m_row[0].z, b.m_row[1].z, b.m_row[2].z, 0);\n"
+"	b3Float4 ans;\n"
+"	ans.x = b3Dot3F4( a, colx );\n"
+"	ans.y = b3Dot3F4( a, coly );\n"
+"	ans.z = b3Dot3F4( a, colz );\n"
+"	return ans;\n"
+"}\n"
+"#endif\n"
+"#endif //B3_MAT3x3_H\n"
+"typedef struct b3Aabb b3Aabb_t;\n"
+"struct b3Aabb\n"
+"{\n"
+"	union\n"
+"	{\n"
+"		float m_min[4];\n"
+"		b3Float4 m_minVec;\n"
+"		int m_minIndices[4];\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float	m_max[4];\n"
+"		b3Float4 m_maxVec;\n"
+"		int m_signedMaxIndices[4];\n"
+"	};\n"
+"};\n"
+"inline void b3TransformAabb2(b3Float4ConstArg localAabbMin,b3Float4ConstArg localAabbMax, float margin,\n"
+"						b3Float4ConstArg pos,\n"
+"						b3QuatConstArg orn,\n"
+"						b3Float4* aabbMinOut,b3Float4* aabbMaxOut)\n"
+"{\n"
+"		b3Float4 localHalfExtents = 0.5f*(localAabbMax-localAabbMin);\n"
+"		localHalfExtents+=b3MakeFloat4(margin,margin,margin,0.f);\n"
+"		b3Float4 localCenter = 0.5f*(localAabbMax+localAabbMin);\n"
+"		b3Mat3x3 m;\n"
+"		m = b3QuatGetRotationMatrix(orn);\n"
+"		b3Mat3x3 abs_b = b3AbsoluteMat3x3(m);\n"
+"		b3Float4 center = b3TransformPoint(localCenter,pos,orn);\n"
+"		\n"
+"		b3Float4 extent = b3MakeFloat4(b3Dot3F4(localHalfExtents,b3GetRow(abs_b,0)),\n"
+"										 b3Dot3F4(localHalfExtents,b3GetRow(abs_b,1)),\n"
+"										 b3Dot3F4(localHalfExtents,b3GetRow(abs_b,2)),\n"
+"										 0.f);\n"
+"		*aabbMinOut = center-extent;\n"
+"		*aabbMaxOut = center+extent;\n"
+"}\n"
+"/// conservative test for overlap between two aabbs\n"
+"inline bool b3TestAabbAgainstAabb(b3Float4ConstArg aabbMin1,b3Float4ConstArg aabbMax1,\n"
+"								b3Float4ConstArg aabbMin2, b3Float4ConstArg aabbMax2)\n"
+"{\n"
+"	bool overlap = true;\n"
+"	overlap = (aabbMin1.x > aabbMax2.x || aabbMax1.x < aabbMin2.x) ? false : overlap;\n"
+"	overlap = (aabbMin1.z > aabbMax2.z || aabbMax1.z < aabbMin2.z) ? false : overlap;\n"
+"	overlap = (aabbMin1.y > aabbMax2.y || aabbMax1.y < aabbMin2.y) ? false : overlap;\n"
+"	return overlap;\n"
+"}\n"
+"#endif //B3_AABB_H\n"
+"#ifndef B3_COLLIDABLE_H\n"
+"#define B3_COLLIDABLE_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"enum b3ShapeTypes\n"
+"{\n"
+"	SHAPE_HEIGHT_FIELD=1,\n"
+"	SHAPE_CONVEX_HULL=3,\n"
+"	SHAPE_PLANE=4,\n"
+"	SHAPE_CONCAVE_TRIMESH=5,\n"
+"	SHAPE_COMPOUND_OF_CONVEX_HULLS=6,\n"
+"	SHAPE_SPHERE=7,\n"
+"	MAX_NUM_SHAPE_TYPES,\n"
+"};\n"
+"typedef struct b3Collidable b3Collidable_t;\n"
+"struct b3Collidable\n"
+"{\n"
+"	union {\n"
+"		int m_numChildShapes;\n"
+"		int m_bvhIndex;\n"
+"	};\n"
+"	union\n"
+"	{\n"
+"		float m_radius;\n"
+"		int	m_compoundBvhIndex;\n"
+"	};\n"
+"	int m_shapeType;\n"
+"	union\n"
+"	{\n"
+"		int m_shapeIndex;\n"
+"		float m_height;\n"
+"	};\n"
+"};\n"
+"typedef struct b3GpuChildShape b3GpuChildShape_t;\n"
+"struct b3GpuChildShape\n"
+"{\n"
+"	b3Float4	m_childPosition;\n"
+"	b3Quat		m_childOrientation;\n"
+"	union\n"
+"	{\n"
+"		int			m_shapeIndex;//used for SHAPE_COMPOUND_OF_CONVEX_HULLS\n"
+"		int			m_capsuleAxis;\n"
+"	};\n"
+"	union \n"
+"	{\n"
+"		float		m_radius;//used for childshape of SHAPE_COMPOUND_OF_SPHERES or SHAPE_COMPOUND_OF_CAPSULES\n"
+"		int			m_numChildShapes;//used for compound shape\n"
+"	};\n"
+"	union \n"
+"	{\n"
+"		float		m_height;//used for childshape of SHAPE_COMPOUND_OF_CAPSULES\n"
+"		int	m_collidableShapeIndex;\n"
+"	};\n"
+"	int			m_shapeType;\n"
+"};\n"
+"struct b3CompoundOverlappingPair\n"
+"{\n"
+"	int m_bodyIndexA;\n"
+"	int m_bodyIndexB;\n"
+"//	int	m_pairType;\n"
+"	int m_childShapeIndexA;\n"
+"	int m_childShapeIndexB;\n"
+"};\n"
+"#endif //B3_COLLIDABLE_H\n"
+"#ifndef B3_RIGIDBODY_DATA_H\n"
+"#define B3_RIGIDBODY_DATA_H\n"
+"#ifndef B3_FLOAT4_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_FLOAT4_H\n"
+"#ifndef B3_QUAT_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif \n"
+"#endif //B3_QUAT_H\n"
+"#ifndef B3_MAT3x3_H\n"
+"#ifdef __cplusplus\n"
+"#else\n"
+"#endif\n"
+"#endif //B3_MAT3x3_H\n"
+"typedef struct b3RigidBodyData b3RigidBodyData_t;\n"
+"struct b3RigidBodyData\n"
+"{\n"
+"	b3Float4				m_pos;\n"
+"	b3Quat					m_quat;\n"
+"	b3Float4				m_linVel;\n"
+"	b3Float4				m_angVel;\n"
+"	int 					m_collidableIdx;\n"
+"	float 				m_invMass;\n"
+"	float 				m_restituitionCoeff;\n"
+"	float 				m_frictionCoeff;\n"
+"};\n"
+"typedef struct b3InertiaData b3InertiaData_t;\n"
+"struct b3InertiaData\n"
+"{\n"
+"	b3Mat3x3 m_invInertiaWorld;\n"
+"	b3Mat3x3 m_initInvInertia;\n"
+"};\n"
+"#endif //B3_RIGIDBODY_DATA_H\n"
+"	\n"
+"void b3ComputeWorldAabb(  int bodyId, __global const b3RigidBodyData_t* bodies, __global const  b3Collidable_t* collidables, __global const  b3Aabb_t* localShapeAABB, __global b3Aabb_t* worldAabbs)\n"
+"{\n"
+"	__global const b3RigidBodyData_t* body = &bodies[bodyId];\n"
+"	b3Float4 position = body->m_pos;\n"
+"	b3Quat	orientation = body->m_quat;\n"
+"	\n"
+"	int collidableIndex = body->m_collidableIdx;\n"
+"	int shapeIndex = collidables[collidableIndex].m_shapeIndex;\n"
+"		\n"
+"	if (shapeIndex>=0)\n"
+"	{\n"
+"				\n"
+"		b3Aabb_t localAabb = localShapeAABB[collidableIndex];\n"
+"		b3Aabb_t worldAabb;\n"
+"		\n"
+"		b3Float4 aabbAMinOut,aabbAMaxOut;	\n"
+"		float margin = 0.f;\n"
+"		b3TransformAabb2(localAabb.m_minVec,localAabb.m_maxVec,margin,position,orientation,&aabbAMinOut,&aabbAMaxOut);\n"
+"		\n"
+"		worldAabb.m_minVec =aabbAMinOut;\n"
+"		worldAabb.m_minIndices[3] = bodyId;\n"
+"		worldAabb.m_maxVec = aabbAMaxOut;\n"
+"		worldAabb.m_signedMaxIndices[3] = body[bodyId].m_invMass==0.f? 0 : 1;\n"
+"		worldAabbs[bodyId] = worldAabb;\n"
+"	}\n"
+"}\n"
+"#endif //B3_UPDATE_AABBS_H\n"
+"__kernel void initializeGpuAabbsFull(  const int numNodes, __global b3RigidBodyData_t* gBodies,__global b3Collidable_t* collidables, __global b3Aabb_t* plocalShapeAABB, __global b3Aabb_t* pAABB)\n"
+"{\n"
+"	int nodeID = get_global_id(0);\n"
+"	if( nodeID < numNodes )\n"
+"	{\n"
+"		b3ComputeWorldAabb(nodeID, gBodies, collidables, plocalShapeAABB,pAABB);\n"
+"	}\n"
+"}\n"
+"__kernel void clearOverlappingPairsKernel(  __global int4* pairs, int numPairs)\n"
+"{\n"
+"	int pairId = get_global_id(0);\n"
+"	if( pairId< numPairs )\n"
+"	{\n"
+"		pairs[pairId].z = 0xffffffff;\n"
+"	}\n"
+"}\n"
+;