// Copyright 2009-2020 Intel Corporation // SPDX-License-Identifier: Apache-2.0 #pragma once #include "quadv.h" #include "triangle_intersector_moeller.h" namespace embree { namespace isa { template struct QuadHitM { __forceinline QuadHitM() {} __forceinline QuadHitM(const vbool& valid, const vfloat& U, const vfloat& V, const vfloat& T, const vfloat& absDen, const Vec3vf& Ng, const vbool& flags) : U(U), V(V), T(T), absDen(absDen), tri_Ng(Ng), valid(valid), flags(flags) {} __forceinline void finalize() { const vfloat rcpAbsDen = rcp(absDen); vt = T * rcpAbsDen; const vfloat u = min(U * rcpAbsDen,1.0f); const vfloat v = min(V * rcpAbsDen,1.0f); const vfloat u1 = vfloat(1.0f) - u; const vfloat v1 = vfloat(1.0f) - v; #if !defined(__AVX__) || defined(EMBREE_BACKFACE_CULLING) vu = select(flags,u1,u); vv = select(flags,v1,v); vNg = Vec3vf(tri_Ng.x,tri_Ng.y,tri_Ng.z); #else const vfloat flip = select(flags,vfloat(-1.0f),vfloat(1.0f)); vv = select(flags,u1,v); vu = select(flags,v1,u); vNg = Vec3vf(flip*tri_Ng.x,flip*tri_Ng.y,flip*tri_Ng.z); #endif } __forceinline Vec2f uv(const size_t i) { const float u = vu[i]; const float v = vv[i]; return Vec2f(u,v); } __forceinline float t(const size_t i) { return vt[i]; } __forceinline Vec3fa Ng(const size_t i) { return Vec3fa(vNg.x[i],vNg.y[i],vNg.z[i]); } private: vfloat U; vfloat V; vfloat T; vfloat absDen; Vec3vf tri_Ng; public: vbool valid; vfloat vu; vfloat vv; vfloat vt; Vec3vf vNg; public: const vbool flags; }; template struct QuadHitK { __forceinline QuadHitK(const vfloat& U, const vfloat& V, const vfloat& T, const vfloat& absDen, const Vec3vf& Ng, const vbool& flags) : U(U), V(V), T(T), absDen(absDen), flags(flags), tri_Ng(Ng) {} __forceinline std::tuple,vfloat,vfloat,Vec3vf> operator() () const { const vfloat rcpAbsDen = rcp(absDen); const vfloat t = T * rcpAbsDen; const vfloat u0 = min(U * rcpAbsDen,1.0f); const vfloat v0 = min(V * rcpAbsDen,1.0f); const vfloat u1 = vfloat(1.0f) - u0; const vfloat v1 = vfloat(1.0f) - v0; const vfloat u = select(flags,u1,u0); const vfloat v = select(flags,v1,v0); const Vec3vf Ng(tri_Ng.x,tri_Ng.y,tri_Ng.z); return std::make_tuple(u,v,t,Ng); } private: const vfloat U; const vfloat V; const vfloat T; const vfloat absDen; const vbool flags; const Vec3vf tri_Ng; }; /* ----------------------------- */ /* -- single ray intersectors -- */ /* ----------------------------- */ template struct QuadMIntersector1MoellerTrumbore; /*! Intersects M quads with 1 ray */ template struct QuadMIntersector1MoellerTrumbore { __forceinline QuadMIntersector1MoellerTrumbore() {} __forceinline QuadMIntersector1MoellerTrumbore(const Ray& ray, const void* ptr) {} __forceinline void intersect(RayHit& ray, IntersectContext* context, const Vec3vf& v0, const Vec3vf& v1, const Vec3vf& v2, const Vec3vf& v3, const vuint& geomID, const vuint& primID) const { MoellerTrumboreHitM hit; MoellerTrumboreIntersector1 intersector(ray,nullptr); Intersect1EpilogM epilog(ray,context,geomID,primID); /* intersect first triangle */ if (intersector.intersect(ray,v0,v1,v3,hit)) epilog(hit.valid,hit); /* intersect second triangle */ if (intersector.intersect(ray,v2,v3,v1,hit)) { hit.U = hit.absDen - hit.U; hit.V = hit.absDen - hit.V; epilog(hit.valid,hit); } } __forceinline bool occluded(Ray& ray, IntersectContext* context, const Vec3vf& v0, const Vec3vf& v1, const Vec3vf& v2, const Vec3vf& v3, const vuint& geomID, const vuint& primID) const { MoellerTrumboreHitM hit; MoellerTrumboreIntersector1 intersector(ray,nullptr); Occluded1EpilogM epilog(ray,context,geomID,primID); /* intersect first triangle */ if (intersector.intersect(ray,v0,v1,v3,hit)) { if (epilog(hit.valid,hit)) return true; } /* intersect second triangle */ if (intersector.intersect(ray,v2,v3,v1,hit)) { hit.U = hit.absDen - hit.U; hit.V = hit.absDen - hit.V; if (epilog(hit.valid,hit)) return true; } return false; } }; #if defined(__AVX512ER__) // KNL /*! Intersects 4 quads with 1 ray using AVX512 */ template struct QuadMIntersector1MoellerTrumbore<4,filter> { __forceinline QuadMIntersector1MoellerTrumbore() {} __forceinline QuadMIntersector1MoellerTrumbore(const Ray& ray, const void* ptr) {} template __forceinline bool intersect(Ray& ray, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const Epilog& epilog) const { const Vec3vf16 vtx0(select(0x0f0f,vfloat16(v0.x),vfloat16(v2.x)), select(0x0f0f,vfloat16(v0.y),vfloat16(v2.y)), select(0x0f0f,vfloat16(v0.z),vfloat16(v2.z))); #if !defined(EMBREE_BACKFACE_CULLING) const Vec3vf16 vtx1(vfloat16(v1.x),vfloat16(v1.y),vfloat16(v1.z)); const Vec3vf16 vtx2(vfloat16(v3.x),vfloat16(v3.y),vfloat16(v3.z)); #else const Vec3vf16 vtx1(select(0x0f0f,vfloat16(v1.x),vfloat16(v3.x)), select(0x0f0f,vfloat16(v1.y),vfloat16(v3.y)), select(0x0f0f,vfloat16(v1.z),vfloat16(v3.z))); const Vec3vf16 vtx2(select(0x0f0f,vfloat16(v3.x),vfloat16(v1.x)), select(0x0f0f,vfloat16(v3.y),vfloat16(v1.y)), select(0x0f0f,vfloat16(v3.z),vfloat16(v1.z))); #endif const vbool16 flags(0xf0f0); MoellerTrumboreHitM<16> hit; MoellerTrumboreIntersector1<16> intersector(ray,nullptr); if (unlikely(intersector.intersect(ray,vtx0,vtx1,vtx2,hit))) { vfloat16 U = hit.U, V = hit.V, absDen = hit.absDen; #if !defined(EMBREE_BACKFACE_CULLING) hit.U = select(flags,absDen-V,U); hit.V = select(flags,absDen-U,V); hit.vNg *= select(flags,vfloat16(-1.0f),vfloat16(1.0f)); // FIXME: use XOR #else hit.U = select(flags,absDen-U,U); hit.V = select(flags,absDen-V,V); #endif if (likely(epilog(hit.valid,hit))) return true; } return false; } __forceinline bool intersect(RayHit& ray, IntersectContext* context, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const vuint4& geomID, const vuint4& primID) const { return intersect(ray,v0,v1,v2,v3,Intersect1EpilogM<8,16,filter>(ray,context,vuint8(geomID),vuint8(primID))); } __forceinline bool occluded(Ray& ray, IntersectContext* context, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const vuint4& geomID, const vuint4& primID) const { return intersect(ray,v0,v1,v2,v3,Occluded1EpilogM<8,16,filter>(ray,context,vuint8(geomID),vuint8(primID))); } }; #elif defined(__AVX__) /*! Intersects 4 quads with 1 ray using AVX */ template struct QuadMIntersector1MoellerTrumbore<4,filter> { __forceinline QuadMIntersector1MoellerTrumbore() {} __forceinline QuadMIntersector1MoellerTrumbore(const Ray& ray, const void* ptr) {} template __forceinline bool intersect(Ray& ray, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const Epilog& epilog) const { const Vec3vf8 vtx0(vfloat8(v0.x,v2.x),vfloat8(v0.y,v2.y),vfloat8(v0.z,v2.z)); #if !defined(EMBREE_BACKFACE_CULLING) const Vec3vf8 vtx1(vfloat8(v1.x),vfloat8(v1.y),vfloat8(v1.z)); const Vec3vf8 vtx2(vfloat8(v3.x),vfloat8(v3.y),vfloat8(v3.z)); #else const Vec3vf8 vtx1(vfloat8(v1.x,v3.x),vfloat8(v1.y,v3.y),vfloat8(v1.z,v3.z)); const Vec3vf8 vtx2(vfloat8(v3.x,v1.x),vfloat8(v3.y,v1.y),vfloat8(v3.z,v1.z)); #endif MoellerTrumboreHitM<8> hit; MoellerTrumboreIntersector1<8> intersector(ray,nullptr); const vbool8 flags(0,0,0,0,1,1,1,1); if (unlikely(intersector.intersect(ray,vtx0,vtx1,vtx2,hit))) { vfloat8 U = hit.U, V = hit.V, absDen = hit.absDen; #if !defined(EMBREE_BACKFACE_CULLING) hit.U = select(flags,absDen-V,U); hit.V = select(flags,absDen-U,V); hit.vNg *= select(flags,vfloat8(-1.0f),vfloat8(1.0f)); // FIXME: use XOR #else hit.U = select(flags,absDen-U,U); hit.V = select(flags,absDen-V,V); #endif if (unlikely(epilog(hit.valid,hit))) return true; } return false; } __forceinline bool intersect(RayHit& ray, IntersectContext* context, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const vuint4& geomID, const vuint4& primID) const { return intersect(ray,v0,v1,v2,v3,Intersect1EpilogM<8,8,filter>(ray,context,vuint8(geomID),vuint8(primID))); } __forceinline bool occluded(Ray& ray, IntersectContext* context, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const vuint4& geomID, const vuint4& primID) const { return intersect(ray,v0,v1,v2,v3,Occluded1EpilogM<8,8,filter>(ray,context,vuint8(geomID),vuint8(primID))); } }; #endif /* ----------------------------- */ /* -- ray packet intersectors -- */ /* ----------------------------- */ struct MoellerTrumboreIntersector1KTriangleM { /*! Intersect k'th ray from ray packet of size K with M triangles. */ template static __forceinline bool intersect(RayK& ray, size_t k, const Vec3vf& tri_v0, const Vec3vf& tri_e1, const Vec3vf& tri_e2, const Vec3vf& tri_Ng, const vbool& flags, const Epilog& epilog) { /* calculate denominator */ const Vec3vf O = broadcast>(ray.org,k); const Vec3vf D = broadcast>(ray.dir,k); const Vec3vf C = Vec3vf(tri_v0) - O; const Vec3vf R = cross(C,D); const vfloat den = dot(Vec3vf(tri_Ng),D); const vfloat absDen = abs(den); const vfloat sgnDen = signmsk(den); /* perform edge tests */ const vfloat U = dot(R,Vec3vf(tri_e2)) ^ sgnDen; const vfloat V = dot(R,Vec3vf(tri_e1)) ^ sgnDen; /* perform backface culling */ #if defined(EMBREE_BACKFACE_CULLING) vbool valid = (den < vfloat(zero)) & (U >= 0.0f) & (V >= 0.0f) & (U+V<=absDen); #else vbool valid = (den != vfloat(zero)) & (U >= 0.0f) & (V >= 0.0f) & (U+V<=absDen); #endif if (likely(none(valid))) return false; /* perform depth test */ const vfloat T = dot(Vec3vf(tri_Ng),C) ^ sgnDen; valid &= (absDen*vfloat(ray.tnear()[k]) < T) & (T <= absDen*vfloat(ray.tfar[k])); if (likely(none(valid))) return false; /* calculate hit information */ QuadHitM hit(valid,U,V,T,absDen,tri_Ng,flags); return epilog(valid,hit); } template static __forceinline bool intersect1(RayK& ray, size_t k, const Vec3vf& v0, const Vec3vf& v1, const Vec3vf& v2, const vbool& flags, const Epilog& epilog) { const Vec3vf e1 = v0-v1; const Vec3vf e2 = v2-v0; const Vec3vf Ng = cross(e2,e1); return intersect(ray,k,v0,e1,e2,Ng,flags,epilog); } }; template struct QuadMIntersectorKMoellerTrumboreBase { __forceinline QuadMIntersectorKMoellerTrumboreBase(const vbool& valid, const RayK& ray) {} /*! Intersects K rays with one of M triangles. */ template __forceinline vbool intersectK(const vbool& valid0, RayK& ray, const Vec3vf& tri_v0, const Vec3vf& tri_e1, const Vec3vf& tri_e2, const Vec3vf& tri_Ng, const vbool& flags, const Epilog& epilog) const { /* calculate denominator */ vbool valid = valid0; const Vec3vf C = tri_v0 - ray.org; const Vec3vf R = cross(C,ray.dir); const vfloat den = dot(tri_Ng,ray.dir); const vfloat absDen = abs(den); const vfloat sgnDen = signmsk(den); /* test against edge p2 p0 */ const vfloat U = dot(R,tri_e2) ^ sgnDen; valid &= U >= 0.0f; if (likely(none(valid))) return false; /* test against edge p0 p1 */ const vfloat V = dot(R,tri_e1) ^ sgnDen; valid &= V >= 0.0f; if (likely(none(valid))) return false; /* test against edge p1 p2 */ const vfloat W = absDen-U-V; valid &= W >= 0.0f; if (likely(none(valid))) return false; /* perform depth test */ const vfloat T = dot(tri_Ng,C) ^ sgnDen; valid &= (absDen*ray.tnear() < T) & (T <= absDen*ray.tfar); if (unlikely(none(valid))) return false; /* perform backface culling */ #if defined(EMBREE_BACKFACE_CULLING) valid &= den < vfloat(zero); if (unlikely(none(valid))) return false; #else valid &= den != vfloat(zero); if (unlikely(none(valid))) return false; #endif /* calculate hit information */ QuadHitK hit(U,V,T,absDen,tri_Ng,flags); return epilog(valid,hit); } /*! Intersects K rays with one of M quads. */ template __forceinline vbool intersectK(const vbool& valid0, RayK& ray, const Vec3vf& tri_v0, const Vec3vf& tri_v1, const Vec3vf& tri_v2, const vbool& flags, const Epilog& epilog) const { const Vec3vf e1 = tri_v0-tri_v1; const Vec3vf e2 = tri_v2-tri_v0; const Vec3vf Ng = cross(e2,e1); return intersectK(valid0,ray,tri_v0,e1,e2,Ng,flags,epilog); } /*! Intersects K rays with one of M quads. */ template __forceinline bool intersectK(const vbool& valid0, RayK& ray, const Vec3vf& v0, const Vec3vf& v1, const Vec3vf& v2, const Vec3vf& v3, const Epilog& epilog) const { intersectK(valid0,ray,v0,v1,v3,vbool(false),epilog); if (none(valid0)) return true; intersectK(valid0,ray,v2,v3,v1,vbool(true ),epilog); return none(valid0); } }; template struct QuadMIntersectorKMoellerTrumbore : public QuadMIntersectorKMoellerTrumboreBase { __forceinline QuadMIntersectorKMoellerTrumbore(const vbool& valid, const RayK& ray) : QuadMIntersectorKMoellerTrumboreBase(valid,ray) {} __forceinline void intersect1(RayHitK& ray, size_t k, IntersectContext* context, const Vec3vf& v0, const Vec3vf& v1, const Vec3vf& v2, const Vec3vf& v3, const vuint& geomID, const vuint& primID) const { Intersect1KEpilogM epilog(ray,k,context,geomID,primID); MoellerTrumboreIntersector1KTriangleM::intersect1(ray,k,v0,v1,v3,vbool(false),epilog); MoellerTrumboreIntersector1KTriangleM::intersect1(ray,k,v2,v3,v1,vbool(true ),epilog); } __forceinline bool occluded1(RayK& ray, size_t k, IntersectContext* context, const Vec3vf& v0, const Vec3vf& v1, const Vec3vf& v2, const Vec3vf& v3, const vuint& geomID, const vuint& primID) const { Occluded1KEpilogM epilog(ray,k,context,geomID,primID); if (MoellerTrumboreIntersector1KTriangleM::intersect1(ray,k,v0,v1,v3,vbool(false),epilog)) return true; if (MoellerTrumboreIntersector1KTriangleM::intersect1(ray,k,v2,v3,v1,vbool(true ),epilog)) return true; return false; } }; #if defined(__AVX512ER__) // KNL /*! Intersects 4 quads with 1 ray using AVX512 */ template struct QuadMIntersectorKMoellerTrumbore<4,K,filter> : public QuadMIntersectorKMoellerTrumboreBase<4,K,filter> { __forceinline QuadMIntersectorKMoellerTrumbore(const vbool& valid, const RayK& ray) : QuadMIntersectorKMoellerTrumboreBase<4,K,filter>(valid,ray) {} template __forceinline bool intersect1(RayK& ray, size_t k, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const Epilog& epilog) const { const Vec3vf16 vtx0(select(0x0f0f,vfloat16(v0.x),vfloat16(v2.x)), select(0x0f0f,vfloat16(v0.y),vfloat16(v2.y)), select(0x0f0f,vfloat16(v0.z),vfloat16(v2.z))); #if !defined(EMBREE_BACKFACE_CULLING) const Vec3vf16 vtx1(vfloat16(v1.x),vfloat16(v1.y),vfloat16(v1.z)); const Vec3vf16 vtx2(vfloat16(v3.x),vfloat16(v3.y),vfloat16(v3.z)); #else const Vec3vf16 vtx1(select(0x0f0f,vfloat16(v1.x),vfloat16(v3.x)), select(0x0f0f,vfloat16(v1.y),vfloat16(v3.y)), select(0x0f0f,vfloat16(v1.z),vfloat16(v3.z))); const Vec3vf16 vtx2(select(0x0f0f,vfloat16(v3.x),vfloat16(v1.x)), select(0x0f0f,vfloat16(v3.y),vfloat16(v1.y)), select(0x0f0f,vfloat16(v3.z),vfloat16(v1.z))); #endif const vbool16 flags(0xf0f0); return MoellerTrumboreIntersector1KTriangleM::intersect1(ray,k,vtx0,vtx1,vtx2,flags,epilog); } __forceinline bool intersect1(RayHitK& ray, size_t k, IntersectContext* context, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const vuint4& geomID, const vuint4& primID) const { return intersect1(ray,k,v0,v1,v2,v3,Intersect1KEpilogM<8,16,K,filter>(ray,k,context,vuint8(geomID),vuint8(primID))); } __forceinline bool occluded1(RayK& ray, size_t k, IntersectContext* context, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const vuint4& geomID, const vuint4& primID) const { return intersect1(ray,k,v0,v1,v2,v3,Occluded1KEpilogM<8,16,K,filter>(ray,k,context,vuint8(geomID),vuint8(primID))); } }; #elif defined(__AVX__) /*! Intersects 4 quads with 1 ray using AVX */ template struct QuadMIntersectorKMoellerTrumbore<4,K,filter> : public QuadMIntersectorKMoellerTrumboreBase<4,K,filter> { __forceinline QuadMIntersectorKMoellerTrumbore(const vbool& valid, const RayK& ray) : QuadMIntersectorKMoellerTrumboreBase<4,K,filter>(valid,ray) {} template __forceinline bool intersect1(RayK& ray, size_t k, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const Epilog& epilog) const { const Vec3vf8 vtx0(vfloat8(v0.x,v2.x),vfloat8(v0.y,v2.y),vfloat8(v0.z,v2.z)); #if !defined(EMBREE_BACKFACE_CULLING) const Vec3vf8 vtx1(vfloat8(v1.x),vfloat8(v1.y),vfloat8(v1.z)); const Vec3vf8 vtx2(vfloat8(v3.x),vfloat8(v3.y),vfloat8(v3.z)); #else const Vec3vf8 vtx1(vfloat8(v1.x,v3.x),vfloat8(v1.y,v3.y),vfloat8(v1.z,v3.z)); const Vec3vf8 vtx2(vfloat8(v3.x,v1.x),vfloat8(v3.y,v1.y),vfloat8(v3.z,v1.z)); #endif const vbool8 flags(0,0,0,0,1,1,1,1); return MoellerTrumboreIntersector1KTriangleM::intersect1(ray,k,vtx0,vtx1,vtx2,flags,epilog); } __forceinline bool intersect1(RayHitK& ray, size_t k, IntersectContext* context, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const vuint4& geomID, const vuint4& primID) const { return intersect1(ray,k,v0,v1,v2,v3,Intersect1KEpilogM<8,8,K,filter>(ray,k,context,vuint8(geomID),vuint8(primID))); } __forceinline bool occluded1(RayK& ray, size_t k, IntersectContext* context, const Vec3vf4& v0, const Vec3vf4& v1, const Vec3vf4& v2, const Vec3vf4& v3, const vuint4& geomID, const vuint4& primID) const { return intersect1(ray,k,v0,v1,v2,v3,Occluded1KEpilogM<8,8,K,filter>(ray,k,context,vuint8(geomID),vuint8(primID))); } }; #endif } }