Function intersectNode

kernels/bvh/node_intersector1.h:561–611

    template<int N>
      __forceinline size_t intersectNode(const typename BVHN<N>::AABBNodeMB* node, const TravRay<N,false>& ray, const float time, vfloat<N>& dist)
    {
      const vfloat<N>* pNearX = (const vfloat<N>*)((const char*)&node->lower_x+ray.nearX);
      const vfloat<N>* pNearY = (const vfloat<N>*)((const char*)&node->lower_x+ray.nearY);
      const vfloat<N>* pNearZ = (const vfloat<N>*)((const char*)&node->lower_x+ray.nearZ);
      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(__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;
      const vfloat<N> tNearZ = (madd(time,pNearZ[6],vfloat<N>(pNearZ[0])) - ray.org.z) * ray.rdir.z;
      const vfloat<N> tFarX  = (madd(time,pFarX [6],vfloat<N>(pFarX [0])) - ray.org.x) * ray.rdir.x;
      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(__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);
      const size_t mask = movemask(vmask) ^ ((1<<N)-1);
#elif defined(__AVX512F__) // SKX
      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);
      const size_t mask = movemask(vmask);
#else
      const vfloat<N> tNear = max(ray.tnear,tNearX,tNearY,tNearZ);
      const vfloat<N> tFar  = min(ray.tfar, tFarX ,tFarY ,tFarZ );
      const vbool<N> vmask = tNear <= tFar;
      const size_t mask = movemask(vmask);
#endif
      dist = tNear;
      return mask;
    }

    //////////////////////////////////////////////////////////////////////////////////////
    // Robust AABBNodeMB intersection