diff options
Diffstat (limited to 'thirdparty/embree/kernels/bvh')
12 files changed, 249 insertions, 28 deletions
diff --git a/thirdparty/embree/kernels/bvh/bvh.cpp b/thirdparty/embree/kernels/bvh/bvh.cpp index a84295f0da..f6cf626465 100644 --- a/thirdparty/embree/kernels/bvh/bvh.cpp +++ b/thirdparty/embree/kernels/bvh/bvh.cpp @@ -183,7 +183,7 @@ namespace embree template class BVHN<8>; #endif -#if !defined(__AVX__) || !defined(EMBREE_TARGET_SSE2) && !defined(EMBREE_TARGET_SSE42) +#if !defined(__AVX__) || !defined(EMBREE_TARGET_SSE2) && !defined(EMBREE_TARGET_SSE42) || defined(__aarch64__) template class BVHN<4>; #endif } diff --git a/thirdparty/embree/kernels/bvh/bvh_intersector_hybrid.cpp b/thirdparty/embree/kernels/bvh/bvh_intersector_hybrid.cpp index 6e9a5a538e..1d393fd06b 100644 --- a/thirdparty/embree/kernels/bvh/bvh_intersector_hybrid.cpp +++ b/thirdparty/embree/kernels/bvh/bvh_intersector_hybrid.cpp @@ -230,7 +230,7 @@ namespace embree continue; /* switch to single ray traversal */ -#if (!defined(__WIN32__) || defined(__X86_64__)) && defined(__SSE4_2__) +#if (!defined(__WIN32__) || defined(__X86_64__)) && ((defined(__aarch64__)) || defined(__SSE4_2__)) #if FORCE_SINGLE_MODE == 0 if (single) #endif @@ -676,7 +676,7 @@ namespace embree continue; /* switch to single ray traversal */ -#if (!defined(__WIN32__) || defined(__X86_64__)) && defined(__SSE4_2__) +#if (!defined(__WIN32__) || defined(__X86_64__)) && ((defined(__aarch64__)) || defined(__SSE4_2__)) #if FORCE_SINGLE_MODE == 0 if (single) #endif diff --git a/thirdparty/embree/kernels/bvh/bvh_intersector_stream.h b/thirdparty/embree/kernels/bvh/bvh_intersector_stream.h index 717f559677..c7e040fadb 100644 --- a/thirdparty/embree/kernels/bvh/bvh_intersector_stream.h +++ b/thirdparty/embree/kernels/bvh/bvh_intersector_stream.h @@ -170,12 +170,23 @@ namespace embree TravRayKStream<K,robust> &p = packets[rayID / K]; const size_t i = rayID % K; const vint<N> bitmask(shiftTable[rayID]); + +#if defined (__aarch64__) + const vfloat<N> tNearX = madd(bminX, p.rdir.x[i], p.neg_org_rdir.x[i]); + const vfloat<N> tNearY = madd(bminY, p.rdir.y[i], p.neg_org_rdir.y[i]); + const vfloat<N> tNearZ = madd(bminZ, p.rdir.z[i], p.neg_org_rdir.z[i]); + const vfloat<N> tFarX = madd(bmaxX, p.rdir.x[i], p.neg_org_rdir.x[i]); + const vfloat<N> tFarY = madd(bmaxY, p.rdir.y[i], p.neg_org_rdir.y[i]); + const vfloat<N> tFarZ = madd(bmaxZ, p.rdir.z[i], p.neg_org_rdir.z[i]); +#else const vfloat<N> tNearX = msub(bminX, p.rdir.x[i], p.org_rdir.x[i]); const vfloat<N> tNearY = msub(bminY, p.rdir.y[i], p.org_rdir.y[i]); const vfloat<N> tNearZ = msub(bminZ, p.rdir.z[i], p.org_rdir.z[i]); const vfloat<N> tFarX = msub(bmaxX, p.rdir.x[i], p.org_rdir.x[i]); const vfloat<N> tFarY = msub(bmaxY, p.rdir.y[i], p.org_rdir.y[i]); const vfloat<N> tFarZ = msub(bmaxZ, p.rdir.z[i], p.org_rdir.z[i]); +#endif + const vfloat<N> tNear = maxi(tNearX, tNearY, tNearZ, vfloat<N>(p.tnear[i])); const vfloat<N> tFar = mini(tFarX , tFarY , tFarZ, vfloat<N>(p.tfar[i])); diff --git a/thirdparty/embree/kernels/bvh/bvh_node_aabb.h b/thirdparty/embree/kernels/bvh/bvh_node_aabb.h index 57530692bc..3fd9fc7d18 100644 --- a/thirdparty/embree/kernels/bvh/bvh_node_aabb.h +++ b/thirdparty/embree/kernels/bvh/bvh_node_aabb.h @@ -46,6 +46,14 @@ namespace embree template<typename BuildRecord> __forceinline NodeRef operator() (const BuildRecord& precord, const BuildRecord* crecords, NodeRef ref, NodeRef* children, const size_t num) const { +#if defined(DEBUG) + // check that empty children are only at the end of the child list + bool emptyChild = false; + for (size_t i=0; i<num; i++) { + emptyChild |= (children[i] == NodeRef::emptyNode); + assert(emptyChild == (children[i] == NodeRef::emptyNode)); + } +#endif AABBNode_t* node = ref.getAABBNode(); for (size_t i=0; i<num; i++) node->setRef(i,children[i]); return ref; @@ -60,6 +68,14 @@ namespace embree template<typename BuildRecord> __forceinline NodeRef operator() (const BuildRecord& precord, const BuildRecord* crecords, NodeRef ref, NodeRef* children, const size_t num) const { +#if defined(DEBUG) + // check that empty children are only at the end of the child list + bool emptyChild = false; + for (size_t i=0; i<num; i++) { + emptyChild |= (children[i] == NodeRef::emptyNode); + assert(emptyChild == (children[i] == NodeRef::emptyNode)); + } +#endif AABBNode_t* node = ref.getAABBNode(); for (size_t i=0; i<num; i++) node->setRef(i,children[i]); diff --git a/thirdparty/embree/kernels/bvh/bvh_node_aabb_mb.h b/thirdparty/embree/kernels/bvh/bvh_node_aabb_mb.h index c4cea7d8ba..001f526c25 100644 --- a/thirdparty/embree/kernels/bvh/bvh_node_aabb_mb.h +++ b/thirdparty/embree/kernels/bvh/bvh_node_aabb_mb.h @@ -31,6 +31,14 @@ namespace embree template<typename BuildRecord> __forceinline NodeRecordMB operator() (const BuildRecord& precord, const BuildRecord* crecords, NodeRef ref, NodeRecordMB* children, const size_t num) const { +#if defined(DEBUG) + // check that empty children are only at the end of the child list + bool emptyChild = false; + for (size_t i=0; i<num; i++) { + emptyChild |= (children[i].ref == NodeRef::emptyNode); + assert(emptyChild == (children[i].ref == NodeRef::emptyNode)); + } +#endif AABBNodeMB_t* node = ref.getAABBNodeMB(); LBBox3fa bounds = empty; diff --git a/thirdparty/embree/kernels/bvh/bvh_node_aabb_mb4d.h b/thirdparty/embree/kernels/bvh/bvh_node_aabb_mb4d.h index 46a81d7581..3b966fd054 100644 --- a/thirdparty/embree/kernels/bvh/bvh_node_aabb_mb4d.h +++ b/thirdparty/embree/kernels/bvh/bvh_node_aabb_mb4d.h @@ -41,6 +41,14 @@ namespace embree template<typename BuildRecord> __forceinline void operator() (const BuildRecord&, const BuildRecord*, NodeRef ref, NodeRecordMB4D* children, const size_t num) const { +#if defined(DEBUG) + // check that empty children are only at the end of the child list + bool emptyChild = false; + for (size_t i=0; i<num; i++) { + emptyChild |= (children[i].ref == NodeRef::emptyNode); + assert(emptyChild == (children[i].ref == NodeRef::emptyNode)); + } +#endif if (likely(ref.isAABBNodeMB())) { for (size_t i=0; i<num; i++) ref.getAABBNodeMB()->set(i, children[i]); diff --git a/thirdparty/embree/kernels/bvh/bvh_node_qaabb.h b/thirdparty/embree/kernels/bvh/bvh_node_qaabb.h index 2afc8c98e7..99671ddc5a 100644 --- a/thirdparty/embree/kernels/bvh/bvh_node_qaabb.h +++ b/thirdparty/embree/kernels/bvh/bvh_node_qaabb.h @@ -190,6 +190,14 @@ namespace embree template<typename BuildRecord> __forceinline NodeRef operator() (const BuildRecord& precord, const BuildRecord* crecords, NodeRef ref, NodeRef* children, const size_t num) const { +#if defined(DEBUG) + // check that empty children are only at the end of the child list + bool emptyChild = false; + for (size_t i=0; i<num; i++) { + emptyChild |= (children[i] == NodeRef::emptyNode); + assert(emptyChild == (children[i] == NodeRef::emptyNode)); + } +#endif QuantizedNode_t* node = ref.quantizedNode(); for (size_t i=0; i<num; i++) node->setRef(i,children[i]); return ref; diff --git a/thirdparty/embree/kernels/bvh/bvh_statistics.cpp b/thirdparty/embree/kernels/bvh/bvh_statistics.cpp index d857ff7d95..57f75bfd7e 100644 --- a/thirdparty/embree/kernels/bvh/bvh_statistics.cpp +++ b/thirdparty/embree/kernels/bvh/bvh_statistics.cpp @@ -162,7 +162,7 @@ namespace embree template class BVHNStatistics<8>; #endif -#if !defined(__AVX__) || !defined(EMBREE_TARGET_SSE2) && !defined(EMBREE_TARGET_SSE42) +#if !defined(__AVX__) || (!defined(EMBREE_TARGET_SSE2) && !defined(EMBREE_TARGET_SSE42)) || defined(__aarch64__) template class BVHNStatistics<4>; #endif } diff --git a/thirdparty/embree/kernels/bvh/node_intersector1.h b/thirdparty/embree/kernels/bvh/node_intersector1.h index 1ec4fc63fc..17641fa888 100644 --- a/thirdparty/embree/kernels/bvh/node_intersector1.h +++ b/thirdparty/embree/kernels/bvh/node_intersector1.h @@ -5,6 +5,15 @@ #include "node_intersector.h" +#if defined(__AVX2__) +#define __FMA_X4__ +#endif + +#if defined(__aarch64__) +#define __FMA_X4__ +#endif + + namespace embree { namespace isa @@ -29,9 +38,15 @@ namespace embree org = Vec3vf<N>(ray_org.x,ray_org.y,ray_org.z); dir = Vec3vf<N>(ray_dir.x,ray_dir.y,ray_dir.z); rdir = Vec3vf<N>(ray_rdir.x,ray_rdir.y,ray_rdir.z); -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__FMA_X4__) const Vec3fa ray_org_rdir = ray_org*ray_rdir; +#if !defined(__aarch64__) org_rdir = Vec3vf<N>(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z); +#else + //for aarch64, we do not have msub equal instruction, so we negeate orig and use madd + //x86 will use msub + neg_org_rdir = Vec3vf<N>(-ray_org_rdir.x,-ray_org_rdir.y,-ray_org_rdir.z); +#endif #endif nearX = ray_rdir.x >= 0.0f ? 0*sizeof(vfloat<N>) : 1*sizeof(vfloat<N>); nearY = ray_rdir.y >= 0.0f ? 2*sizeof(vfloat<N>) : 3*sizeof(vfloat<N>); @@ -49,8 +64,12 @@ namespace embree org = Vec3vf<N>(ray_org.x[k], ray_org.y[k], ray_org.z[k]); dir = Vec3vf<N>(ray_dir.x[k], ray_dir.y[k], ray_dir.z[k]); rdir = Vec3vf<N>(ray_rdir.x[k], ray_rdir.y[k], ray_rdir.z[k]); -#if defined(__AVX2__) || defined(__ARM_NEON) - org_rdir = org*rdir; +#if defined(__FMA_X4__) +#if !defined(__aarch64__) + org_rdir = org*rdir; +#else + neg_org_rdir = -(org*rdir); +#endif #endif nearX = nearXYZ.x[k]; nearY = nearXYZ.y[k]; @@ -62,8 +81,14 @@ namespace embree Vec3fa org_xyz, dir_xyz; Vec3vf<N> org, dir, rdir; -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__FMA_X4__) +#if !defined(__aarch64__) Vec3vf<N> org_rdir; +#else + //aarch64 version are keeping negation of the org_rdir and use madd + //x86 uses msub + Vec3vf<N> neg_org_rdir; +#endif #endif size_t nearX, nearY, nearZ; size_t farX, farY, farZ; @@ -404,13 +429,22 @@ namespace embree template<> __forceinline size_t intersectNode<4>(const typename BVH4::AABBNode* node, const TravRay<4,false>& ray, vfloat4& dist) { -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__FMA_X4__) +#if defined(__aarch64__) + const vfloat4 tNearX = madd(vfloat4::load((float*)((const char*)&node->lower_x+ray.nearX)), ray.rdir.x, ray.neg_org_rdir.x); + const vfloat4 tNearY = madd(vfloat4::load((float*)((const char*)&node->lower_x+ray.nearY)), ray.rdir.y, ray.neg_org_rdir.y); + const vfloat4 tNearZ = madd(vfloat4::load((float*)((const char*)&node->lower_x+ray.nearZ)), ray.rdir.z, ray.neg_org_rdir.z); + const vfloat4 tFarX = madd(vfloat4::load((float*)((const char*)&node->lower_x+ray.farX )), ray.rdir.x, ray.neg_org_rdir.x); + const vfloat4 tFarY = madd(vfloat4::load((float*)((const char*)&node->lower_x+ray.farY )), ray.rdir.y, ray.neg_org_rdir.y); + const vfloat4 tFarZ = madd(vfloat4::load((float*)((const char*)&node->lower_x+ray.farZ )), ray.rdir.z, ray.neg_org_rdir.z); +#else const vfloat4 tNearX = msub(vfloat4::load((float*)((const char*)&node->lower_x+ray.nearX)), ray.rdir.x, ray.org_rdir.x); const vfloat4 tNearY = msub(vfloat4::load((float*)((const char*)&node->lower_x+ray.nearY)), ray.rdir.y, ray.org_rdir.y); const vfloat4 tNearZ = msub(vfloat4::load((float*)((const char*)&node->lower_x+ray.nearZ)), ray.rdir.z, ray.org_rdir.z); const vfloat4 tFarX = msub(vfloat4::load((float*)((const char*)&node->lower_x+ray.farX )), ray.rdir.x, ray.org_rdir.x); const vfloat4 tFarY = msub(vfloat4::load((float*)((const char*)&node->lower_x+ray.farY )), ray.rdir.y, ray.org_rdir.y); const vfloat4 tFarZ = msub(vfloat4::load((float*)((const char*)&node->lower_x+ray.farZ )), ray.rdir.z, ray.org_rdir.z); +#endif #else const vfloat4 tNearX = (vfloat4::load((float*)((const char*)&node->lower_x+ray.nearX)) - ray.org.x) * ray.rdir.x; const vfloat4 tNearY = (vfloat4::load((float*)((const char*)&node->lower_x+ray.nearY)) - ray.org.y) * ray.rdir.y; @@ -450,13 +484,23 @@ namespace embree template<> __forceinline size_t intersectNode<8>(const typename BVH8::AABBNode* node, const TravRay<8,false>& ray, vfloat8& dist) { -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__AVX2__) +#if defined(__aarch64__) + const vfloat8 tNearX = madd(vfloat8::load((float*)((const char*)&node->lower_x+ray.nearX)), ray.rdir.x, ray.neg_org_rdir.x); + const vfloat8 tNearY = madd(vfloat8::load((float*)((const char*)&node->lower_x+ray.nearY)), ray.rdir.y, ray.neg_org_rdir.y); + const vfloat8 tNearZ = madd(vfloat8::load((float*)((const char*)&node->lower_x+ray.nearZ)), ray.rdir.z, ray.neg_org_rdir.z); + const vfloat8 tFarX = madd(vfloat8::load((float*)((const char*)&node->lower_x+ray.farX )), ray.rdir.x, ray.neg_org_rdir.x); + const vfloat8 tFarY = madd(vfloat8::load((float*)((const char*)&node->lower_x+ray.farY )), ray.rdir.y, ray.neg_org_rdir.y); + const vfloat8 tFarZ = madd(vfloat8::load((float*)((const char*)&node->lower_x+ray.farZ )), ray.rdir.z, ray.neg_org_rdir.z); +#else const vfloat8 tNearX = msub(vfloat8::load((float*)((const char*)&node->lower_x+ray.nearX)), ray.rdir.x, ray.org_rdir.x); const vfloat8 tNearY = msub(vfloat8::load((float*)((const char*)&node->lower_x+ray.nearY)), ray.rdir.y, ray.org_rdir.y); const vfloat8 tNearZ = msub(vfloat8::load((float*)((const char*)&node->lower_x+ray.nearZ)), ray.rdir.z, ray.org_rdir.z); const vfloat8 tFarX = msub(vfloat8::load((float*)((const char*)&node->lower_x+ray.farX )), ray.rdir.x, ray.org_rdir.x); const vfloat8 tFarY = msub(vfloat8::load((float*)((const char*)&node->lower_x+ray.farY )), ray.rdir.y, ray.org_rdir.y); const vfloat8 tFarZ = msub(vfloat8::load((float*)((const char*)&node->lower_x+ray.farZ )), ray.rdir.z, ray.org_rdir.z); +#endif + #else const vfloat8 tNearX = (vfloat8::load((float*)((const char*)&node->lower_x+ray.nearX)) - ray.org.x) * ray.rdir.x; const vfloat8 tNearY = (vfloat8::load((float*)((const char*)&node->lower_x+ray.nearY)) - ray.org.y) * ray.rdir.y; @@ -522,13 +566,22 @@ namespace embree const vfloat<N>* pFarX = (const vfloat<N>*)((const char*)&node->lower_x+ray.farX); const vfloat<N>* pFarY = (const vfloat<N>*)((const char*)&node->lower_x+ray.farY); const vfloat<N>* pFarZ = (const vfloat<N>*)((const char*)&node->lower_x+ray.farZ); -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__FMA_X4__) +#if defined(__aarch64__) + const vfloat<N> tNearX = madd(madd(time,pNearX[6],vfloat<N>(pNearX[0])), ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<N> tNearY = madd(madd(time,pNearY[6],vfloat<N>(pNearY[0])), ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<N> tNearZ = madd(madd(time,pNearZ[6],vfloat<N>(pNearZ[0])), ray.rdir.z, ray.neg_org_rdir.z); + const vfloat<N> tFarX = madd(madd(time,pFarX [6],vfloat<N>(pFarX [0])), ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<N> tFarY = madd(madd(time,pFarY [6],vfloat<N>(pFarY [0])), ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<N> tFarZ = madd(madd(time,pFarZ [6],vfloat<N>(pFarZ [0])), ray.rdir.z, ray.neg_org_rdir.z); +#else const vfloat<N> tNearX = msub(madd(time,pNearX[6],vfloat<N>(pNearX[0])), ray.rdir.x, ray.org_rdir.x); const vfloat<N> tNearY = msub(madd(time,pNearY[6],vfloat<N>(pNearY[0])), ray.rdir.y, ray.org_rdir.y); const vfloat<N> tNearZ = msub(madd(time,pNearZ[6],vfloat<N>(pNearZ[0])), ray.rdir.z, ray.org_rdir.z); const vfloat<N> tFarX = msub(madd(time,pFarX [6],vfloat<N>(pFarX [0])), ray.rdir.x, ray.org_rdir.x); const vfloat<N> tFarY = msub(madd(time,pFarY [6],vfloat<N>(pFarY [0])), ray.rdir.y, ray.org_rdir.y); const vfloat<N> tFarZ = msub(madd(time,pFarZ [6],vfloat<N>(pFarZ [0])), ray.rdir.z, ray.org_rdir.z); +#endif #else const vfloat<N> tNearX = (madd(time,pNearX[6],vfloat<N>(pNearX[0])) - ray.org.x) * ray.rdir.x; const vfloat<N> tNearY = (madd(time,pNearY[6],vfloat<N>(pNearY[0])) - ray.org.y) * ray.rdir.y; @@ -537,7 +590,7 @@ namespace embree const vfloat<N> tFarY = (madd(time,pFarY [6],vfloat<N>(pFarY [0])) - ray.org.y) * ray.rdir.y; const vfloat<N> tFarZ = (madd(time,pFarZ [6],vfloat<N>(pFarZ [0])) - ray.org.z) * ray.rdir.z; #endif -#if defined(__AVX2__) && !defined(__AVX512F__) // HSW +#if defined(__FMA_X4__) && !defined(__AVX512F__) // HSW const vfloat<N> tNear = maxi(tNearX,tNearY,tNearZ,ray.tnear); const vfloat<N> tFar = mini(tFarX ,tFarY ,tFarZ ,ray.tfar); const vbool<N> vmask = asInt(tNear) > asInt(tFar); @@ -598,13 +651,22 @@ namespace embree const vfloat<N>* pFarX = (const vfloat<N>*)((const char*)&node->lower_x+ray.farX); const vfloat<N>* pFarY = (const vfloat<N>*)((const char*)&node->lower_x+ray.farY); const vfloat<N>* pFarZ = (const vfloat<N>*)((const char*)&node->lower_x+ray.farZ); -#if defined (__AVX2__) || defined(__ARM_NEON) +#if defined (__FMA_X4__) +#if defined(__aarch64__) + const vfloat<N> tNearX = madd(madd(time,pNearX[6],vfloat<N>(pNearX[0])), ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<N> tNearY = madd(madd(time,pNearY[6],vfloat<N>(pNearY[0])), ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<N> tNearZ = madd(madd(time,pNearZ[6],vfloat<N>(pNearZ[0])), ray.rdir.z, ray.neg_org_rdir.z); + const vfloat<N> tFarX = madd(madd(time,pFarX [6],vfloat<N>(pFarX [0])), ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<N> tFarY = madd(madd(time,pFarY [6],vfloat<N>(pFarY [0])), ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<N> tFarZ = madd(madd(time,pFarZ [6],vfloat<N>(pFarZ [0])), ray.rdir.z, ray.neg_org_rdir.z); +#else const vfloat<N> tNearX = msub(madd(time,pNearX[6],vfloat<N>(pNearX[0])), ray.rdir.x, ray.org_rdir.x); const vfloat<N> tNearY = msub(madd(time,pNearY[6],vfloat<N>(pNearY[0])), ray.rdir.y, ray.org_rdir.y); const vfloat<N> tNearZ = msub(madd(time,pNearZ[6],vfloat<N>(pNearZ[0])), ray.rdir.z, ray.org_rdir.z); const vfloat<N> tFarX = msub(madd(time,pFarX [6],vfloat<N>(pFarX [0])), ray.rdir.x, ray.org_rdir.x); const vfloat<N> tFarY = msub(madd(time,pFarY [6],vfloat<N>(pFarY [0])), ray.rdir.y, ray.org_rdir.y); const vfloat<N> tFarZ = msub(madd(time,pFarZ [6],vfloat<N>(pFarZ [0])), ray.rdir.z, ray.org_rdir.z); +#endif #else const vfloat<N> tNearX = (madd(time,pNearX[6],vfloat<N>(pNearX[0])) - ray.org.x) * ray.rdir.x; const vfloat<N> tNearY = (madd(time,pNearY[6],vfloat<N>(pNearY[0])) - ray.org.y) * ray.rdir.y; @@ -613,7 +675,7 @@ namespace embree const vfloat<N> tFarY = (madd(time,pFarY [6],vfloat<N>(pFarY [0])) - ray.org.y) * ray.rdir.y; const vfloat<N> tFarZ = (madd(time,pFarZ [6],vfloat<N>(pFarZ [0])) - ray.org.z) * ray.rdir.z; #endif -#if defined(__AVX2__) && !defined(__AVX512F__) +#if defined(__FMA_X4__) && !defined(__AVX512F__) const vfloat<N> tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,ray.tnear)); const vfloat<N> tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,ray.tfar )); #else @@ -687,13 +749,22 @@ namespace embree const vfloat4 lower_z = madd(node->dequantize<4>(ray.nearZ >> 2),scale_z,start_z); const vfloat4 upper_z = madd(node->dequantize<4>(ray.farZ >> 2),scale_z,start_z); -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__FMA_X4__) +#if defined(__aarch64__) + const vfloat4 tNearX = madd(lower_x, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat4 tNearY = madd(lower_y, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat4 tNearZ = madd(lower_z, ray.rdir.z, ray.neg_org_rdir.z); + const vfloat4 tFarX = madd(upper_x, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat4 tFarY = madd(upper_y, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat4 tFarZ = madd(upper_z, ray.rdir.z, ray.neg_org_rdir.z); +#else const vfloat4 tNearX = msub(lower_x, ray.rdir.x, ray.org_rdir.x); const vfloat4 tNearY = msub(lower_y, ray.rdir.y, ray.org_rdir.y); const vfloat4 tNearZ = msub(lower_z, ray.rdir.z, ray.org_rdir.z); const vfloat4 tFarX = msub(upper_x, ray.rdir.x, ray.org_rdir.x); const vfloat4 tFarY = msub(upper_y, ray.rdir.y, ray.org_rdir.y); const vfloat4 tFarZ = msub(upper_z, ray.rdir.z, ray.org_rdir.z); +#endif #else const vfloat4 tNearX = (lower_x - ray.org.x) * ray.rdir.x; const vfloat4 tNearY = (lower_y - ray.org.y) * ray.rdir.y; @@ -703,7 +774,7 @@ namespace embree const vfloat4 tFarZ = (upper_z - ray.org.z) * ray.rdir.z; #endif -#if defined(__SSE4_1__) && !defined(__AVX512F__) // up to HSW +#if defined(__aarch64__) || defined(__SSE4_1__) && !defined(__AVX512F__) // up to HSW const vfloat4 tNear = maxi(tNearX,tNearY,tNearZ,ray.tnear); const vfloat4 tFar = mini(tFarX ,tFarY ,tFarZ ,ray.tfar); const vbool4 vmask = asInt(tNear) > asInt(tFar); @@ -775,13 +846,22 @@ namespace embree const vfloat8 lower_z = madd(node->dequantize<8>(ray.nearZ >> 2),scale_z,start_z); const vfloat8 upper_z = madd(node->dequantize<8>(ray.farZ >> 2),scale_z,start_z); -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__AVX2__) +#if defined(__aarch64__) + const vfloat8 tNearX = madd(lower_x, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat8 tNearY = madd(lower_y, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat8 tNearZ = madd(lower_z, ray.rdir.z, ray.neg_org_rdir.z); + const vfloat8 tFarX = madd(upper_x, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat8 tFarY = madd(upper_y, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat8 tFarZ = madd(upper_z, ray.rdir.z, ray.neg_org_rdir.z); +#else const vfloat8 tNearX = msub(lower_x, ray.rdir.x, ray.org_rdir.x); const vfloat8 tNearY = msub(lower_y, ray.rdir.y, ray.org_rdir.y); const vfloat8 tNearZ = msub(lower_z, ray.rdir.z, ray.org_rdir.z); const vfloat8 tFarX = msub(upper_x, ray.rdir.x, ray.org_rdir.x); const vfloat8 tFarY = msub(upper_y, ray.rdir.y, ray.org_rdir.y); const vfloat8 tFarZ = msub(upper_z, ray.rdir.z, ray.org_rdir.z); +#endif #else const vfloat8 tNearX = (lower_x - ray.org.x) * ray.rdir.x; const vfloat8 tNearY = (lower_y - ray.org.y) * ray.rdir.y; @@ -857,13 +937,22 @@ namespace embree const vfloat<N> upper_y = node->dequantizeUpperY(time); const vfloat<N> lower_z = node->dequantizeLowerZ(time); const vfloat<N> upper_z = node->dequantizeUpperZ(time); -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__FMA_X4__) +#if defined(__aarch64__) + const vfloat<N> tNearX = madd(lower_x, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<N> tNearY = madd(lower_y, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<N> tNearZ = madd(lower_z, ray.rdir.z, ray.neg_org_rdir.z); + const vfloat<N> tFarX = madd(upper_x, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<N> tFarY = madd(upper_y, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<N> tFarZ = madd(upper_z, ray.rdir.z, ray.neg_org_rdir.z); +#else const vfloat<N> tNearX = msub(lower_x, ray.rdir.x, ray.org_rdir.x); const vfloat<N> tNearY = msub(lower_y, ray.rdir.y, ray.org_rdir.y); const vfloat<N> tNearZ = msub(lower_z, ray.rdir.z, ray.org_rdir.z); const vfloat<N> tFarX = msub(upper_x, ray.rdir.x, ray.org_rdir.x); const vfloat<N> tFarY = msub(upper_y, ray.rdir.y, ray.org_rdir.y); const vfloat<N> tFarZ = msub(upper_z, ray.rdir.z, ray.org_rdir.z); +#endif #else const vfloat<N> tNearX = (lower_x - ray.org.x) * ray.rdir.x; const vfloat<N> tNearY = (lower_y - ray.org.y) * ray.rdir.y; diff --git a/thirdparty/embree/kernels/bvh/node_intersector_frustum.h b/thirdparty/embree/kernels/bvh/node_intersector_frustum.h index 1f7215e5df..cad4e6de2d 100644 --- a/thirdparty/embree/kernels/bvh/node_intersector_frustum.h +++ b/thirdparty/embree/kernels/bvh/node_intersector_frustum.h @@ -75,9 +75,13 @@ namespace embree min_rdir = select(pos_rdir, reduced_min_rdir, reduced_max_rdir); max_rdir = select(pos_rdir, reduced_max_rdir, reduced_min_rdir); +#if defined (__aarch64__) + neg_min_org_rdir = -(min_rdir * select(pos_rdir, reduced_max_org, reduced_min_org)); + neg_max_org_rdir = -(max_rdir * select(pos_rdir, reduced_min_org, reduced_max_org)); +#else min_org_rdir = min_rdir * select(pos_rdir, reduced_max_org, reduced_min_org); max_org_rdir = max_rdir * select(pos_rdir, reduced_min_org, reduced_max_org); - +#endif min_dist = reduced_min_dist; max_dist = reduced_max_dist; @@ -95,9 +99,13 @@ namespace embree Vec3fa min_rdir; Vec3fa max_rdir; +#if defined (__aarch64__) + Vec3fa neg_min_org_rdir; + Vec3fa neg_max_org_rdir; +#else Vec3fa min_org_rdir; Vec3fa max_org_rdir; - +#endif float min_dist; float max_dist; }; @@ -191,13 +199,21 @@ namespace embree const vfloat<N> bmaxY = *(const vfloat<N>*)((const char*)&node->lower_x + frustum.nf.farY); const vfloat<N> bmaxZ = *(const vfloat<N>*)((const char*)&node->lower_x + frustum.nf.farZ); +#if defined (__aarch64__) + const vfloat<N> fminX = madd(bminX, vfloat<N>(frustum.min_rdir.x), vfloat<N>(frustum.neg_min_org_rdir.x)); + const vfloat<N> fminY = madd(bminY, vfloat<N>(frustum.min_rdir.y), vfloat<N>(frustum.neg_min_org_rdir.y)); + const vfloat<N> fminZ = madd(bminZ, vfloat<N>(frustum.min_rdir.z), vfloat<N>(frustum.neg_min_org_rdir.z)); + const vfloat<N> fmaxX = madd(bmaxX, vfloat<N>(frustum.max_rdir.x), vfloat<N>(frustum.neg_max_org_rdir.x)); + const vfloat<N> fmaxY = madd(bmaxY, vfloat<N>(frustum.max_rdir.y), vfloat<N>(frustum.neg_max_org_rdir.y)); + const vfloat<N> fmaxZ = madd(bmaxZ, vfloat<N>(frustum.max_rdir.z), vfloat<N>(frustum.neg_max_org_rdir.z)); +#else const vfloat<N> fminX = msub(bminX, vfloat<N>(frustum.min_rdir.x), vfloat<N>(frustum.min_org_rdir.x)); const vfloat<N> fminY = msub(bminY, vfloat<N>(frustum.min_rdir.y), vfloat<N>(frustum.min_org_rdir.y)); const vfloat<N> fminZ = msub(bminZ, vfloat<N>(frustum.min_rdir.z), vfloat<N>(frustum.min_org_rdir.z)); const vfloat<N> fmaxX = msub(bmaxX, vfloat<N>(frustum.max_rdir.x), vfloat<N>(frustum.max_org_rdir.x)); const vfloat<N> fmaxY = msub(bmaxY, vfloat<N>(frustum.max_rdir.y), vfloat<N>(frustum.max_org_rdir.y)); const vfloat<N> fmaxZ = msub(bmaxZ, vfloat<N>(frustum.max_rdir.z), vfloat<N>(frustum.max_org_rdir.z)); - +#endif const vfloat<N> fmin = maxi(fminX, fminY, fminZ, vfloat<N>(frustum.min_dist)); dist = fmin; const vfloat<N> fmax = mini(fmaxX, fmaxY, fmaxZ, vfloat<N>(frustum.max_dist)); diff --git a/thirdparty/embree/kernels/bvh/node_intersector_packet.h b/thirdparty/embree/kernels/bvh/node_intersector_packet.h index d5498fc5db..4deacd620d 100644 --- a/thirdparty/embree/kernels/bvh/node_intersector_packet.h +++ b/thirdparty/embree/kernels/bvh/node_intersector_packet.h @@ -39,7 +39,9 @@ namespace embree org = ray_org; dir = ray_dir; rdir = rcp_safe(ray_dir); -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__aarch64__) + neg_org_rdir = -(org * rdir); +#elif defined(__AVX2__) org_rdir = org * rdir; #endif @@ -55,7 +57,9 @@ namespace embree Vec3vf<K> org; Vec3vf<K> dir; Vec3vf<K> rdir; -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__aarch64__) + Vec3vf<K> neg_org_rdir; +#elif defined(__AVX2__) Vec3vf<K> org_rdir; #endif Vec3vi<K> nearXYZ; @@ -119,7 +123,14 @@ namespace embree const TravRayKFast<K>& ray, vfloat<K>& dist) { - #if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__aarch64__) + const vfloat<K> lclipMinX = madd(node->lower_x[i], ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<K> lclipMinY = madd(node->lower_y[i], ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<K> lclipMinZ = madd(node->lower_z[i], ray.rdir.z, ray.neg_org_rdir.z); + const vfloat<K> lclipMaxX = madd(node->upper_x[i], ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<K> lclipMaxY = madd(node->upper_y[i], ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<K> lclipMaxZ = madd(node->upper_z[i], ray.rdir.z, ray.neg_org_rdir.z); +#elif defined(__AVX2__) const vfloat<K> lclipMinX = msub(node->lower_x[i], ray.rdir.x, ray.org_rdir.x); const vfloat<K> lclipMinY = msub(node->lower_y[i], ray.rdir.y, ray.org_rdir.y); const vfloat<K> lclipMinZ = msub(node->lower_z[i], ray.rdir.z, ray.org_rdir.z); @@ -199,7 +210,14 @@ namespace embree const vfloat<K> vupper_y = madd(time, vfloat<K>(node->upper_dy[i]), vfloat<K>(node->upper_y[i])); const vfloat<K> vupper_z = madd(time, vfloat<K>(node->upper_dz[i]), vfloat<K>(node->upper_z[i])); -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__aarch64__) + const vfloat<K> lclipMinX = madd(vlower_x, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<K> lclipMinY = madd(vlower_y, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<K> lclipMinZ = madd(vlower_z, ray.rdir.z, ray.neg_org_rdir.z); + const vfloat<K> lclipMaxX = madd(vupper_x, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<K> lclipMaxY = madd(vupper_y, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<K> lclipMaxZ = madd(vupper_z, ray.rdir.z, ray.neg_org_rdir.z); +#elif defined(__AVX2__) const vfloat<K> lclipMinX = msub(vlower_x, ray.rdir.x, ray.org_rdir.x); const vfloat<K> lclipMinY = msub(vlower_y, ray.rdir.y, ray.org_rdir.y); const vfloat<K> lclipMinZ = msub(vlower_z, ray.rdir.z, ray.org_rdir.z); @@ -302,7 +320,14 @@ namespace embree const vfloat<K> vupper_y = madd(time, vfloat<K>(node->upper_dy[i]), vfloat<K>(node->upper_y[i])); const vfloat<K> vupper_z = madd(time, vfloat<K>(node->upper_dz[i]), vfloat<K>(node->upper_z[i])); -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__aarch64__) + const vfloat<K> lclipMinX = madd(vlower_x, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<K> lclipMinY = madd(vlower_y, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<K> lclipMinZ = madd(vlower_z, ray.rdir.z, ray.neg_org_rdir.z); + const vfloat<K> lclipMaxX = madd(vupper_x, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<K> lclipMaxY = madd(vupper_y, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<K> lclipMaxZ = madd(vupper_z, ray.rdir.z, ray.neg_org_rdir.z); +#elif defined(__AVX2__) const vfloat<K> lclipMinX = msub(vlower_x, ray.rdir.x, ray.org_rdir.x); const vfloat<K> lclipMinY = msub(vlower_y, ray.rdir.y, ray.org_rdir.y); const vfloat<K> lclipMinZ = msub(vlower_z, ray.rdir.z, ray.org_rdir.z); @@ -464,7 +489,14 @@ namespace embree const vfloat<N> lower_z = node->dequantizeLowerZ(); const vfloat<N> upper_z = node->dequantizeUpperZ(); - #if defined(__AVX2__) || defined(__ARM_NEON) + #if defined(__aarch64__) + const vfloat<K> lclipMinX = madd(lower_x[i], ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<K> lclipMinY = madd(lower_y[i], ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<K> lclipMinZ = madd(lower_z[i], ray.rdir.z, ray.neg_org_rdir.z); + const vfloat<K> lclipMaxX = madd(upper_x[i], ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<K> lclipMaxY = madd(upper_y[i], ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<K> lclipMaxZ = madd(upper_z[i], ray.rdir.z, ray.neg_org_rdir.z); + #elif defined(__AVX2__) const vfloat<K> lclipMinX = msub(lower_x[i], ray.rdir.x, ray.org_rdir.x); const vfloat<K> lclipMinY = msub(lower_y[i], ray.rdir.y, ray.org_rdir.y); const vfloat<K> lclipMinZ = msub(lower_z[i], ray.rdir.z, ray.org_rdir.z); @@ -549,7 +581,14 @@ namespace embree const vfloat<K> lower_z = node->template dequantizeLowerZ<K>(i,time); const vfloat<K> upper_z = node->template dequantizeUpperZ<K>(i,time); -#if defined(__AVX2__) || defined(__ARM_NEON) +#if defined(__aarch64__) + const vfloat<K> lclipMinX = madd(lower_x, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<K> lclipMinY = madd(lower_y, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<K> lclipMinZ = madd(lower_z, ray.rdir.z, ray.neg_org_rdir.z); + const vfloat<K> lclipMaxX = madd(upper_x, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<K> lclipMaxY = madd(upper_y, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<K> lclipMaxZ = madd(upper_z, ray.rdir.z, ray.neg_org_rdir.z); +#elif defined(__AVX2__) const vfloat<K> lclipMinX = msub(lower_x, ray.rdir.x, ray.org_rdir.x); const vfloat<K> lclipMinY = msub(lower_y, ray.rdir.y, ray.org_rdir.y); const vfloat<K> lclipMinZ = msub(lower_z, ray.rdir.z, ray.org_rdir.z); diff --git a/thirdparty/embree/kernels/bvh/node_intersector_packet_stream.h b/thirdparty/embree/kernels/bvh/node_intersector_packet_stream.h index 55b2c27231..943fd7043f 100644 --- a/thirdparty/embree/kernels/bvh/node_intersector_packet_stream.h +++ b/thirdparty/embree/kernels/bvh/node_intersector_packet_stream.h @@ -32,11 +32,19 @@ namespace embree __forceinline void init(const Vec3vf<K>& ray_org, const Vec3vf<K>& ray_dir) { rdir = rcp_safe(ray_dir); +#if defined(__aarch64__) + neg_org_rdir = -(ray_org * rdir); +#else org_rdir = ray_org * rdir; +#endif } Vec3vf<K> rdir; +#if defined(__aarch64__) + Vec3vf<K> neg_org_rdir; +#else Vec3vf<K> org_rdir; +#endif vfloat<K> tnear; vfloat<K> tfar; }; @@ -87,12 +95,21 @@ namespace embree const vfloat<N> bmaxY = vfloat<N>(*(const vfloat<N>*)((const char*)&node->lower_x + nf.farY)); const vfloat<N> bmaxZ = vfloat<N>(*(const vfloat<N>*)((const char*)&node->lower_x + nf.farZ)); +#if defined (__aarch64__) + const vfloat<N> rminX = madd(bminX, vfloat<N>(ray.rdir.x[k]), vfloat<N>(ray.neg_org_rdir.x[k])); + const vfloat<N> rminY = madd(bminY, vfloat<N>(ray.rdir.y[k]), vfloat<N>(ray.neg_org_rdir.y[k])); + const vfloat<N> rminZ = madd(bminZ, vfloat<N>(ray.rdir.z[k]), vfloat<N>(ray.neg_org_rdir.z[k])); + const vfloat<N> rmaxX = madd(bmaxX, vfloat<N>(ray.rdir.x[k]), vfloat<N>(ray.neg_org_rdir.x[k])); + const vfloat<N> rmaxY = madd(bmaxY, vfloat<N>(ray.rdir.y[k]), vfloat<N>(ray.neg_org_rdir.y[k])); + const vfloat<N> rmaxZ = madd(bmaxZ, vfloat<N>(ray.rdir.z[k]), vfloat<N>(ray.neg_org_rdir.z[k])); +#else const vfloat<N> rminX = msub(bminX, vfloat<N>(ray.rdir.x[k]), vfloat<N>(ray.org_rdir.x[k])); const vfloat<N> rminY = msub(bminY, vfloat<N>(ray.rdir.y[k]), vfloat<N>(ray.org_rdir.y[k])); const vfloat<N> rminZ = msub(bminZ, vfloat<N>(ray.rdir.z[k]), vfloat<N>(ray.org_rdir.z[k])); const vfloat<N> rmaxX = msub(bmaxX, vfloat<N>(ray.rdir.x[k]), vfloat<N>(ray.org_rdir.x[k])); const vfloat<N> rmaxY = msub(bmaxY, vfloat<N>(ray.rdir.y[k]), vfloat<N>(ray.org_rdir.y[k])); const vfloat<N> rmaxZ = msub(bmaxZ, vfloat<N>(ray.rdir.z[k]), vfloat<N>(ray.org_rdir.z[k])); +#endif const vfloat<N> rmin = maxi(rminX, rminY, rminZ, vfloat<N>(ray.tnear[k])); const vfloat<N> rmax = mini(rmaxX, rmaxY, rmaxZ, vfloat<N>(ray.tfar[k])); @@ -113,12 +130,21 @@ namespace embree const vfloat<K> bmaxY = *(const float*)(ptr + nf.farY); const vfloat<K> bmaxZ = *(const float*)(ptr + nf.farZ); +#if defined (__aarch64__) + const vfloat<K> rminX = madd(bminX, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<K> rminY = madd(bminY, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<K> rminZ = madd(bminZ, ray.rdir.z, ray.neg_org_rdir.z); + const vfloat<K> rmaxX = madd(bmaxX, ray.rdir.x, ray.neg_org_rdir.x); + const vfloat<K> rmaxY = madd(bmaxY, ray.rdir.y, ray.neg_org_rdir.y); + const vfloat<K> rmaxZ = madd(bmaxZ, ray.rdir.z, ray.neg_org_rdir.z); +#else const vfloat<K> rminX = msub(bminX, ray.rdir.x, ray.org_rdir.x); const vfloat<K> rminY = msub(bminY, ray.rdir.y, ray.org_rdir.y); const vfloat<K> rminZ = msub(bminZ, ray.rdir.z, ray.org_rdir.z); const vfloat<K> rmaxX = msub(bmaxX, ray.rdir.x, ray.org_rdir.x); const vfloat<K> rmaxY = msub(bmaxY, ray.rdir.y, ray.org_rdir.y); const vfloat<K> rmaxZ = msub(bmaxZ, ray.rdir.z, ray.org_rdir.z); +#endif const vfloat<K> rmin = maxi(rminX, rminY, rminZ, ray.tnear); const vfloat<K> rmax = mini(rmaxX, rmaxY, rmaxZ, ray.tfar); |