diff options
author | Oussama <o.boukhelf@gmail.com> | 2019-01-03 14:26:51 +0100 |
---|---|---|
committer | RĂ©mi Verschelde <rverschelde@gmail.com> | 2019-01-07 12:30:35 +0100 |
commit | 22b7c9dfa80d0f7abca40f061865c2ab3c136a74 (patch) | |
tree | 311cd3f22b012329160f9d43810aea429994af48 /thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h | |
parent | a6722cf36251ddcb538e6ebed9fa4950342b68ba (diff) |
Update Bullet to the latest commit 126b676
Diffstat (limited to 'thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h')
-rw-r--r-- | thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h | 103 |
1 files changed, 46 insertions, 57 deletions
diff --git a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h index 975bd80e53..69caf182d7 100644 --- a/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h +++ b/thirdparty/bullet/Bullet3OpenCL/ParallelPrimitives/b3RadixSort32CL.h @@ -6,90 +6,79 @@ struct b3SortData { - union - { + union { unsigned int m_key; unsigned int x; }; - union - { + union { unsigned int m_value; unsigned int y; - }; }; #include "b3BufferInfoCL.h" -class b3RadixSort32CL +class b3RadixSort32CL { + b3OpenCLArray<unsigned int>* m_workBuffer1; + b3OpenCLArray<unsigned int>* m_workBuffer2; - 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; + b3OpenCLArray<b3SortData>* m_workBuffer3; + b3OpenCLArray<b3SortData>* m_workBuffer4; - cl_command_queue m_commandQueue; + b3OpenCLArray<unsigned int>* m_workBuffer3a; + b3OpenCLArray<unsigned int>* m_workBuffer4a; - cl_kernel m_streamCountSortDataKernel; - cl_kernel m_streamCountKernel; + cl_command_queue m_commandQueue; - cl_kernel m_prefixScanKernel; - cl_kernel m_sortAndScatterSortDataKernel; - cl_kernel m_sortAndScatterKernel; + cl_kernel m_streamCountSortDataKernel; + cl_kernel m_streamCountKernel; + cl_kernel m_prefixScanKernel; + cl_kernel m_sortAndScatterSortDataKernel; + cl_kernel m_sortAndScatterKernel; - bool m_deviceCPU; + bool m_deviceCPU; - class b3PrefixScanCL* m_scan; - class b3FillCL* m_fill; + class b3PrefixScanCL* m_scan; + class b3FillCL* m_fill; public: struct b3ConstData - { - int m_n; - int m_nWGs; - int m_startBit; - int m_nBlocksPerWG; - }; + { + 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 - }; - + { + 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); - b3RadixSort32CL(cl_context ctx, cl_device_id device, cl_command_queue queue, int initialCapacity =0); + virtual ~b3RadixSort32CL(); - 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); - 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); + ///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 - +#endif //B3_RADIXSORT32_H |