void BVH4Intersector4Single<types,robust,PrimitiveIntersector4>::occluded(sseb* valid_i, BVH4* bvh, Ray4& ray)
{
/* load ray */
const sseb valid = *valid_i;
sseb terminated = !valid;
sse3f ray_org = ray.org, ray_dir = ray.dir;
ssef ray_tnear = ray.tnear, ray_tfar = ray.tfar;
const sse3f rdir = rcp_safe(ray_dir);
const sse3f org(ray_org), org_rdir = org * rdir;
ray_tnear = select(valid,ray_tnear,ssef(pos_inf));
ray_tfar = select(valid,ray_tfar ,ssef(neg_inf));
const ssef inf = ssef(pos_inf);
Precalculations pre(valid,ray);
/* compute near/far per ray */
sse3i nearXYZ;
nearXYZ.x = select(rdir.x >= 0.0f,ssei(0*(int)sizeof(ssef)),ssei(1*(int)sizeof(ssef)));
nearXYZ.y = select(rdir.y >= 0.0f,ssei(2*(int)sizeof(ssef)),ssei(3*(int)sizeof(ssef)));
nearXYZ.z = select(rdir.z >= 0.0f,ssei(4*(int)sizeof(ssef)),ssei(5*(int)sizeof(ssef)));
/* we have no packet implementation for OBB nodes yet */
size_t bits = movemask(valid);
for (size_t i=__bsf(bits); bits!=0; bits=__btc(bits,i), i=__bsf(bits)) {
if (occluded1(bvh,bvh->root,i,pre,ray,ray_org,ray_dir,rdir,ray_tnear,ray_tfar,nearXYZ))
terminated[i] = -1;
}
store4i(valid & terminated,&ray.geomID,0);
AVX_ZERO_UPPER();
}
void BVH4Intersector8Single<types,robust,PrimitiveIntersector8>::intersect(avxb* valid_i, BVH4* bvh, Ray8& ray)
{
/* load ray */
const avxb valid0 = *valid_i;
avx3f ray_org = ray.org;
avx3f ray_dir = ray.dir;
avxf ray_tnear = ray.tnear, ray_tfar = ray.tfar;
const avx3f rdir = rcp_safe(ray_dir);
const avx3f org(ray_org), org_rdir = org * rdir;
ray_tnear = select(valid0,ray_tnear,avxf(pos_inf));
ray_tfar = select(valid0,ray_tfar ,avxf(neg_inf));
const avxf inf = avxf(pos_inf);
Precalculations pre(valid0,ray);
/* compute near/far per ray */
avx3i nearXYZ;
nearXYZ.x = select(rdir.x >= 0.0f,avxi(0*(int)sizeof(ssef)),avxi(1*(int)sizeof(ssef)));
nearXYZ.y = select(rdir.y >= 0.0f,avxi(2*(int)sizeof(ssef)),avxi(3*(int)sizeof(ssef)));
nearXYZ.z = select(rdir.z >= 0.0f,avxi(4*(int)sizeof(ssef)),avxi(5*(int)sizeof(ssef)));
/* we have no packet implementation for OBB nodes yet */
size_t bits = movemask(valid0);
for (size_t i=__bsf(bits); bits!=0; bits=__btc(bits,i), i=__bsf(bits)) {
intersect1(bvh, bvh->root, i, pre, ray, ray_org, ray_dir, rdir, ray_tnear, ray_tfar, nearXYZ);
}
AVX_ZERO_UPPER();
}
void BVH4Intersector4FromIntersector1<Intersector1>::intersect(sseb* valid_i, BVH4* bvh, Ray4& ray)
{
Ray rays[4];
ray.get(rays);
size_t bits = movemask(*valid_i);
for (size_t i=__bsf(bits); bits!=0; bits=__btc(bits,i), i=__bsf(bits)) {
Intersector1::intersect(bvh,rays[i]);
}
ray.set(rays);
AVX_ZERO_UPPER();
}
INLINE bool RTCameraPacketGen::generateIA(RayPacket &pckt, int x, int y) const
{
const ssef fw = (float) pckt.width;
const ssef fh = (float) pckt.height;
const ssef fx = (float) x;
const ssef fy = (float) y;
const ssef bottomLeft = aImagePlaneOrg + fx * axAxis + fy * azAxis;
const ssef bottomRight = bottomLeft + fw * axAxis;
const ssef topRight = bottomRight + fh * azAxis;
const ssef topLeft = bottomLeft + fh * azAxis;
const ssef dmin = fixup(min(min(bottomLeft, bottomRight), min(topLeft, topRight)));
const ssef dmax = fixup(max(max(bottomLeft, bottomRight), max(topLeft, topRight)));
const ssef rcpMin = rcp(dmax).xyzz(); // avoid issues with unused w
const ssef rcpMax = rcp(dmin).xyzz(); // avoid issues with unused w
const ssef minusMin = -rcpMin;
const ssef minusMax = -rcpMax;
const size_t mask = movemask(rcpMin);
const sseb maskv = unmovemask(mask);
pckt.iasign = maskv;
pckt.iaMinrDir = select(maskv, minusMax, rcpMin);
pckt.iaMaxrDir = select(maskv, minusMin, rcpMax);
return movemask(dmin ^ dmax) == 0;
}
/* ray/box intersection */
__forceinline size_t intersectBox(const avx3f& org, const avx3f& rdir, const avxf& tnear, const avxf& tfar, const BVH2::Node* node, const int i)
{
const avxf dminx = (avxf(node->lower_upper_x[i+0]) - org.x) * rdir.x;
const avxf dminy = (avxf(node->lower_upper_y[i+0]) - org.y) * rdir.y;
const avxf dminz = (avxf(node->lower_upper_z[i+0]) - org.z) * rdir.z;
const avxf dmaxx = (avxf(node->lower_upper_x[i+2]) - org.x) * rdir.x;
const avxf dmaxy = (avxf(node->lower_upper_y[i+2]) - org.y) * rdir.y;
const avxf dmaxz = (avxf(node->lower_upper_z[i+2]) - org.z) * rdir.z;
const avxf dlowerx = min(dminx,dmaxx);
const avxf dlowery = min(dminy,dmaxy);
const avxf dlowerz = min(dminz,dmaxz);
const avxf dupperx = max(dminx,dmaxx);
const avxf duppery = max(dminy,dmaxy);
const avxf dupperz = max(dminz,dmaxz);
const avxf near = max(dlowerx,dlowery,dlowerz,tnear);
const avxf far = min(dupperx,duppery,dupperz,tfar );
return movemask(near <= far);
}
void BVH4MBIntersector1<TriangleIntersector>::intersect(const BVH4MB* bvh, Ray& ray)
{
AVX_ZERO_UPPER();
STAT3(normal.travs,1,1,1);
/*! stack state */
Base* popCur = bvh->root; //!< pre-popped top node from the stack
float popDist = neg_inf; //!< pre-popped distance of top node from the stack
StackItem stack[1+3*BVH4MB::maxDepth]; //!< stack of nodes that still need to get traversed
StackItem* stackPtr = stack+1; //!< current stack pointer
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray.dir.x >= 0 ? 0*2*sizeof(ssef) : 1*2*sizeof(ssef);
const size_t nearY = ray.dir.y >= 0 ? 2*2*sizeof(ssef) : 3*2*sizeof(ssef);
const size_t nearZ = ray.dir.z >= 0 ? 4*2*sizeof(ssef) : 5*2*sizeof(ssef);
const size_t farX = nearX ^ 32;
const size_t farY = nearY ^ 32;
const size_t farZ = nearZ ^ 32;
/*! load the ray into SIMD registers */
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const ssef rayNear(ray.tnear);
ssef rayFar(ray.tfar);
while (true)
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
Base* cur = popCur;
/*! if popped node is too far, pop next one */
if (unlikely(popDist > ray.tfar)) {
popCur = (Base*)stackPtr[-1].ptr;
popDist = stackPtr[-1].dist;
continue;
}
next:
/*! we mostly go into the inner node case */
if (likely(cur->isNode()))
{
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur->node();
const ssef* pNearX = (const ssef*)((const char*)node+nearX);
const ssef* pNearY = (const ssef*)((const char*)node+nearY);
const ssef* pNearZ = (const ssef*)((const char*)node+nearZ);
const ssef tNearX = (norg.x + ssef(pNearX[0]) + ray.time*pNearX[1]) * rdir.x;
const ssef tNearY = (norg.y + ssef(pNearY[0]) + ray.time*pNearY[1]) * rdir.y;
const ssef tNearZ = (norg.z + ssef(pNearZ[0]) + ray.time*pNearZ[1]) * rdir.z;
const ssef tNear = max(tNearX,tNearY,tNearZ,rayNear);
const ssef* pFarX = (const ssef*)((const char*)node+farX);
const ssef* pFarY = (const ssef*)((const char*)node+farY);
const ssef* pFarZ = (const ssef*)((const char*)node+farZ);
const ssef tFarX = (norg.x + ssef(pFarX[0]) + ray.time*pFarX[1]) * rdir.x;
const ssef tFarY = (norg.y + ssef(pFarY[0]) + ray.time*pFarY[1]) * rdir.y;
const ssef tFarZ = (norg.z + ssef(pFarZ[0]) + ray.time*pFarZ[1]) * rdir.z;
popCur = (Base*) stackPtr[-1].ptr; //!< pre-pop of topmost stack item
popDist = stackPtr[-1].dist; //!< pre-pop of distance of topmost stack item
const ssef tFar = min(tFarX,tFarY,tFarZ,rayFar);
size_t _hit = movemask(tNear <= tFar);
/*! if no child is hit, pop next node */
if (unlikely(_hit == 0))
continue;
/*! one child is hit, continue with that child */
size_t r = __bsf(_hit); _hit = __btc(_hit,r);
if (likely(_hit == 0)) {
cur = node->child[r];
goto next;
}
/*! two children are hit, push far child, and continue with closer child */
Base* c0 = node->child[r]; const float d0 = tNear[r];
r = __bsf(_hit); _hit = __btc(_hit,r);
Base* c1 = node->child[r]; const float d1 = tNear[r];
if (likely(_hit == 0)) {
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; goto next; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; goto next; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
r = __bsf(_hit); _hit = __btc(_hit,r);
Base* c = node->child[r]; float d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
if (likely(_hit == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (Base*) stackPtr[-1].ptr; stackPtr--;
goto next;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
r = __bsf(_hit); _hit = __btc(_hit,r);
c = node->child[r]; d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (Base*) stackPtr[-1].ptr; stackPtr--;
goto next;
}
/*! this is a leaf node */
else
{
STAT3(normal.trav_leaves,1,1,1);
size_t num; Triangle* tri = (Triangle*) cur->leaf(num);
for (size_t i=0; i<num; i++)
TriangleIntersector::intersect(ray,tri[i],bvh->geometry);
popCur = (Base*) stackPtr[-1].ptr; //!< pre-pop of topmost stack item
popDist = stackPtr[-1].dist; //!< pre-pop of distance of topmost stack item
rayFar = ray.tfar;
}
}
AVX_ZERO_UPPER();
}
__forceinline bool BVH8iIntersector8Hybrid<TriangleIntersector8>::occluded1(const BVH8i* bvh, NodeRef root, const size_t k, Ray8& ray,const avx3f &ray_org, const avx3f &ray_dir, const avx3f &ray_rdir, const avxf &ray_tnear, const avxf &ray_tfar, const avx3i& nearXYZ)
{
/*! stack state */
NodeRef stack[stackSizeSingle]; //!< stack of nodes that still need to get traversed
NodeRef* stackPtr = stack+1; //!< current stack pointer
NodeRef* stackEnd = stack+stackSizeSingle;
stack[0] = root;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = nearXYZ.x[k];
const size_t nearY = nearXYZ.y[k];
const size_t nearZ = nearXYZ.z[k];
/*! load the ray into SIMD registers */
const avx3f org (ray_org .x[k],ray_org .y[k],ray_org .z[k]);
const avx3f rdir(ray_rdir.x[k],ray_rdir.y[k],ray_rdir.z[k]);
const avx3f norg = -org, org_rdir(org*rdir);
const avxf rayNear(ray_tnear[k]), rayFar(ray_tfar[k]);
const Node * __restrict__ nodes = (Node *)bvh->nodePtr();
const Triangle * __restrict__ accel = (Triangle*)bvh->triPtr();
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = (Node*)cur.node(nodes);
const size_t farX = nearX ^ sizeof(avxf), farY = nearY ^ sizeof(avxf), farZ = nearZ ^ sizeof(avxf);
#if defined (__AVX2__)
const avxf tNearX = msub(load8f((const char*)node+nearX), rdir.x, org_rdir.x);
const avxf tNearY = msub(load8f((const char*)node+nearY), rdir.y, org_rdir.y);
const avxf tNearZ = msub(load8f((const char*)node+nearZ), rdir.z, org_rdir.z);
const avxf tFarX = msub(load8f((const char*)node+farX ), rdir.x, org_rdir.x);
const avxf tFarY = msub(load8f((const char*)node+farY ), rdir.y, org_rdir.y);
const avxf tFarZ = msub(load8f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const avxf tNearX = (norg.x + load8f((const char*)node+nearX)) * rdir.x;
const avxf tNearY = (norg.y + load8f((const char*)node+nearY)) * rdir.y;
const avxf tNearZ = (norg.z + load8f((const char*)node+nearZ)) * rdir.z;
const avxf tFarX = (norg.x + load8f((const char*)node+farX )) * rdir.x;
const avxf tFarY = (norg.y + load8f((const char*)node+farY )) * rdir.y;
const avxf tFarZ = (norg.z + load8f((const char*)node+farZ )) * rdir.z;
#endif
#if defined(__AVX2__)
const avxf tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,rayNear));
const avxf tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,rayFar ));
const avxb vmask = cast(tNear) > cast(tFar);
unsigned int mask = movemask(vmask)^0xff;
#else
const avxf tNear = max(tNearX,tNearY,tNearZ,rayNear);
const avxf tFar = min(tFarX ,tFarY ,tFarZ ,rayFar);
const avxb vmask = tNear <= tFar;
unsigned int mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
assert(cur != BVH4i::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const unsigned int d1 = ((unsigned int*)&tNear)[r];
assert(c0 != BVH4i::emptyNode);
assert(c1 != BVH4i::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { *stackPtr = c1; stackPtr++; cur = c0; continue; }
else { *stackPtr = c0; stackPtr++; cur = c1; continue; }
}
assert(stackPtr < stackEnd);
*stackPtr = c0; stackPtr++;
assert(stackPtr < stackEnd);
*stackPtr = c1; stackPtr++;
/*! three children are hit */
r = __bscf(mask);
cur = node->child(r);
assert(cur != BVH4i::emptyNode);
if (likely(mask == 0)) continue;
while(1)
{
r = __bscf(mask);
NodeRef c = node->child(r); *stackPtr = c; stackPtr++;
if (unlikely(mask == 0)) break;
}
cur = (NodeRef) stackPtr[-1]; stackPtr--;
// assert(stackPtr < stackEnd);
// *stackPtr = cur; stackPtr++;
// /*! four children are hit */
// cur = node->child(3);
// assert(cur != BVH4i::emptyNode);
}
/*! this is a leaf node */
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Triangle* prim = (Triangle*) cur.leaf(accel,num);
if (TriangleIntersector8::occluded(ray,k,prim,num,bvh->geometry)) {
ray.geomID[k] = 0;
return true;
}
}
return false;
}
void BVH4Intersector4Hybrid<types,robust,PrimitiveIntersector4>::intersect(bool4* valid_i, BVH4* bvh, Ray4& ray)
{
/* verify correct input */
bool4 valid0 = *valid_i;
#if defined(RTCORE_IGNORE_INVALID_RAYS)
valid0 &= ray.valid();
#endif
assert(all(valid0,ray.tnear > -FLT_MIN));
assert(!(types & BVH4::FLAG_NODE_MB) || all(valid0,ray.time >= 0.0f & ray.time <= 1.0f));
/* load ray */
Vec3f4 ray_org = ray.org;
Vec3f4 ray_dir = ray.dir;
float4 ray_tnear = ray.tnear, ray_tfar = ray.tfar;
const Vec3f4 rdir = rcp_safe(ray_dir);
const Vec3f4 org(ray_org), org_rdir = org * rdir;
ray_tnear = select(valid0,ray_tnear,float4(pos_inf));
ray_tfar = select(valid0,ray_tfar ,float4(neg_inf));
const float4 inf = float4(pos_inf);
Precalculations pre(valid0,ray);
/* compute near/far per ray */
Vec3i4 nearXYZ;
nearXYZ.x = select(rdir.x >= 0.0f,int4(0*(int)sizeof(float4)),int4(1*(int)sizeof(float4)));
nearXYZ.y = select(rdir.y >= 0.0f,int4(2*(int)sizeof(float4)),int4(3*(int)sizeof(float4)));
nearXYZ.z = select(rdir.z >= 0.0f,int4(4*(int)sizeof(float4)),int4(5*(int)sizeof(float4)));
/* allocate stack and push root node */
float4 stack_near[stackSizeChunk];
NodeRef stack_node[stackSizeChunk];
stack_node[0] = BVH4::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* stackEnd = stack_node+stackSizeChunk;
NodeRef* __restrict__ sptr_node = stack_node + 2;
float4* __restrict__ sptr_near = stack_near + 2;
while (1) pop:
{
/* pop next node from stack */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
NodeRef cur = *sptr_node;
if (unlikely(cur == BVH4::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* cull node if behind closest hit point */
float4 curDist = *sptr_near;
const bool4 active = curDist < ray_tfar;
if (unlikely(none(active)))
continue;
/* switch to single ray traversal */
#if !defined(__WIN32__) || defined(__X86_64__)
size_t bits = movemask(active);
if (unlikely(__popcnt(bits) <= SWITCH_THRESHOLD)) {
for (size_t i=__bsf(bits); bits!=0; bits=__btc(bits,i), i=__bsf(bits)) {
BVH4Intersector4Single<types,robust,PrimitiveIntersector4>::intersect1(bvh, cur, i, pre, ray, ray_org, ray_dir, rdir, ray_tnear, ray_tfar, nearXYZ);
}
ray_tfar = min(ray_tfar,ray.tfar);
continue;
}
#endif
while (1)
{
/* process normal nodes */
if (likely((types & 0x1) && cur.isNode()))
{
const bool4 valid_node = ray_tfar > curDist;
STAT3(normal.trav_nodes,1,popcnt(valid_node),4);
const Node* __restrict__ const node = cur.node();
/* pop of next node */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
cur = *sptr_node;
curDist = *sptr_near;
#pragma unroll(4)
for (unsigned i=0; i<BVH4::N; i++)
{
const NodeRef child = node->children[i];
if (unlikely(child == BVH4::emptyNode)) break;
float4 lnearP; const bool4 lhit = intersect_node<robust>(node,i,org,rdir,org_rdir,ray_tnear,ray_tfar,lnearP);
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
assert(sptr_node < stackEnd);
assert(child != BVH4::emptyNode);
const float4 childDist = select(lhit,lnearP,inf);
sptr_node++;
sptr_near++;
/* push cur node onto stack and continue with hit child */
if (any(childDist < curDist))
{
*(sptr_node-1) = cur;
*(sptr_near-1) = curDist;
curDist = childDist;
cur = child;
}
/* push hit child onto stack */
else {
*(sptr_node-1) = child;
*(sptr_near-1) = childDist;
}
}
}
#if SWITCH_DURING_DOWN_TRAVERSAL == 1
// seems to be the best place for testing utilization
if (unlikely(popcnt(ray_tfar > curDist) <= SWITCH_THRESHOLD))
{
*sptr_node++ = cur;
*sptr_near++ = curDist;
goto pop;
}
#endif
}
/* process motion blur nodes */
else if (likely((types & 0x10) && cur.isNodeMB()))
{
const bool4 valid_node = ray_tfar > curDist;
STAT3(normal.trav_nodes,1,popcnt(valid_node),4);
const BVH4::NodeMB* __restrict__ const node = cur.nodeMB();
/* pop of next node */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
cur = *sptr_node;
curDist = *sptr_near;
#pragma unroll(4)
for (unsigned i=0; i<BVH4::N; i++)
{
const NodeRef child = node->child(i);
if (unlikely(child == BVH4::emptyNode)) break;
float4 lnearP; const bool4 lhit = intersect_node(node,i,org,rdir,org_rdir,ray_tnear,ray_tfar,ray.time,lnearP);
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
assert(sptr_node < stackEnd);
assert(child != BVH4::emptyNode);
const float4 childDist = select(lhit,lnearP,inf);
sptr_node++;
sptr_near++;
/* push cur node onto stack and continue with hit child */
if (any(childDist < curDist))
{
*(sptr_node-1) = cur;
*(sptr_near-1) = curDist;
curDist = childDist;
cur = child;
}
/* push hit child onto stack */
else {
*(sptr_node-1) = child;
*(sptr_near-1) = childDist;
}
}
}
#if SWITCH_DURING_DOWN_TRAVERSAL == 1
// seems to be the best place for testing utilization
if (unlikely(popcnt(ray_tfar > curDist) <= SWITCH_THRESHOLD))
{
*sptr_node++ = cur;
*sptr_near++ = curDist;
goto pop;
}
#endif
}
else
break;
}
bool BVH4iIntersector1<TriangleIntersector>::occluded(const BVH4iIntersector1* This, Ray& ray)
{
AVX_ZERO_UPPER();
STAT3(shadow.travs,1,1,1);
/*! stack state */
const BVH4i* bvh = This->bvh;
NodeRef stack[1+3*BVH4i::maxDepth]; //!< stack of nodes that still need to get traversed
NodeRef* stackPtr = stack+1; //!< current stack pointer
stack[0] = bvh->root;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray.dir.x >= 0 ? 0*sizeof(ssef_m) : 1*sizeof(ssef_m);
const size_t nearY = ray.dir.y >= 0 ? 2*sizeof(ssef_m) : 3*sizeof(ssef_m);
const size_t nearZ = ray.dir.z >= 0 ? 4*sizeof(ssef_m) : 5*sizeof(ssef_m);
/*! load the ray into SIMD registers */
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vector3f ray_rdir = rcp_safe(ray.dir);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vector3f ray_org_rdir = ray.org*ray_rdir;
const sse3f org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const ssef rayNear(ray.tnear);
const ssef rayFar(ray.tfar);
const void* nodePtr = bvh->nodePtr();
const void* triPtr = bvh->triPtr();
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node(nodePtr);
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
#if defined (__AVX2__)
const ssef tNearX = msub(ssef((const char*)nodePtr+(size_t)cur+nearX), rdir.x, org_rdir.x);
const ssef tNearY = msub(ssef((const char*)nodePtr+(size_t)cur+nearY), rdir.y, org_rdir.y);
const ssef tNearZ = msub(ssef((const char*)nodePtr+(size_t)cur+nearZ), rdir.z, org_rdir.z);
const ssef tFarX = msub(ssef((const char*)nodePtr+(size_t)cur+farX ), rdir.x, org_rdir.x);
const ssef tFarY = msub(ssef((const char*)nodePtr+(size_t)cur+farY ), rdir.y, org_rdir.y);
const ssef tFarZ = msub(ssef((const char*)nodePtr+(size_t)cur+farZ ), rdir.z, org_rdir.z);
#else
const ssef tNearX = (norg.x + ssef((const char*)nodePtr+(size_t)cur+nearX)) * rdir.x;
const ssef tNearY = (norg.y + ssef((const char*)nodePtr+(size_t)cur+nearY)) * rdir.y;
const ssef tNearZ = (norg.z + ssef((const char*)nodePtr+(size_t)cur+nearZ)) * rdir.z;
const ssef tFarX = (norg.x + ssef((const char*)nodePtr+(size_t)cur+farX )) * rdir.x;
const ssef tFarY = (norg.y + ssef((const char*)nodePtr+(size_t)cur+farY )) * rdir.y;
const ssef tFarZ = (norg.z + ssef((const char*)nodePtr+(size_t)cur+farZ )) * rdir.z;
#endif
const ssef tNear = max(tNearX,tNearY,tNearZ,rayNear);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,rayFar);
size_t mask = movemask(tNear <= tFar);
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bsf(mask); mask = __btc(mask,r);
if (likely(mask == 0)) {
cur = node->child(r);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const float d0 = tNear[r];
r = __bsf(mask); mask = __btc(mask,r);
NodeRef c1 = node->child(r); const float d1 = tNear[r];
if (likely(mask == 0)) {
if (d0 < d1) { *stackPtr = c1; stackPtr++; cur = c0; continue; }
else { *stackPtr = c0; stackPtr++; cur = c1; continue; }
}
*stackPtr = c0; stackPtr++;
*stackPtr = c1; stackPtr++;
/*! three children are hit */
r = __bsf(mask); mask = __btc(mask,r);
cur = node->child(r); *stackPtr = cur; stackPtr++;
if (likely(mask == 0)) {
stackPtr--;
continue;
}
/*! four children are hit */
cur = node->child(3);
}
/*! this is a leaf node */
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Triangle* tri = (Triangle*) cur.leaf(triPtr,num);
for (size_t i=0; i<num; i++) {
if (TriangleIntersector::occluded(ray,tri[i],bvh->vertices)) {
AVX_ZERO_UPPER();
return true;
}
}
}
AVX_ZERO_UPPER();
return false;
}
void BVH8Intersector1<robust,PrimitiveIntersector>::intersect(const BVH8* bvh, Ray& ray)
{
/*! perform per ray precalculations required by the primitive intersector */
Precalculations pre(ray,bvh);
/*! stack state */
StackItemT<NodeRef> stack[stackSize]; //!< stack of nodes
StackItemT<NodeRef>* stackPtr = stack+1; //!< current stack pointer
StackItemT<NodeRef>* stackEnd = stack+stackSize;
stack[0].ptr = bvh->root;
stack[0].dist = neg_inf;
/* filter out invalid rays */
#if defined(RTCORE_IGNORE_INVALID_RAYS)
if (!ray.valid()) return;
#endif
/* verify correct input */
assert(ray.tnear > -FLT_MIN);
//assert(!(types & BVH4::FLAG_NODE_MB) || (ray.time >= 0.0f && ray.time <= 1.0f));
/*! load the ray into SIMD registers */
const Vec3f8 norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const Vec3f8 rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vec3fa ray_org_rdir = ray.org*ray_rdir;
const Vec3f8 org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const float8 ray_near(ray.tnear);
float8 ray_far(ray.tfar);
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray_rdir.x >= 0.0f ? 0*sizeof(float8) : 1*sizeof(float8);
const size_t nearY = ray_rdir.y >= 0.0f ? 2*sizeof(float8) : 3*sizeof(float8);
const size_t nearZ = ray_rdir.z >= 0.0f ? 4*sizeof(float8) : 5*sizeof(float8);
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(*(float*)&stackPtr->dist > ray.tfar))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ sizeof(float8), farY = nearY ^ sizeof(float8), farZ = nearZ ^ sizeof(float8);
#if defined (__AVX2__)
const float8 tNearX = msub(load8f((const char*)node+nearX), rdir.x, org_rdir.x);
const float8 tNearY = msub(load8f((const char*)node+nearY), rdir.y, org_rdir.y);
const float8 tNearZ = msub(load8f((const char*)node+nearZ), rdir.z, org_rdir.z);
const float8 tFarX = msub(load8f((const char*)node+farX ), rdir.x, org_rdir.x);
const float8 tFarY = msub(load8f((const char*)node+farY ), rdir.y, org_rdir.y);
const float8 tFarZ = msub(load8f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const float8 tNearX = (norg.x + load8f((const char*)node+nearX)) * rdir.x;
const float8 tNearY = (norg.y + load8f((const char*)node+nearY)) * rdir.y;
const float8 tNearZ = (norg.z + load8f((const char*)node+nearZ)) * rdir.z;
const float8 tFarX = (norg.x + load8f((const char*)node+farX )) * rdir.x;
const float8 tFarY = (norg.y + load8f((const char*)node+farY )) * rdir.y;
const float8 tFarZ = (norg.z + load8f((const char*)node+farZ )) * rdir.z;
#endif
const float round_down = 1.0f-2.0f*float(ulp);
const float round_up = 1.0f+2.0f*float(ulp);
#if defined(__AVX2__)
const float8 tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,ray_near));
const float8 tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,ray_far ));
const bool8 vmask = robust ? (round_down*tNear > round_up*tFar) : cast(tNear) > cast(tFar);
size_t mask = movemask(vmask)^0xff;
#else
const float8 tNear = max(tNearX,tNearY,tNearZ,ray_near);
const float8 tFar = min(tFarX ,tFarY ,tFarZ ,ray_far);
const bool8 vmask = robust ? (round_down*tNear > round_up*tFar) : tNear <= tFar;
size_t mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r); cur.prefetch();
assert(cur != BVH8::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); c0.prefetch(); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); c1.prefetch(); const unsigned int d1 = ((unsigned int*)&tNear)[r];
assert(c0 != BVH8::emptyNode);
assert(c1 != BVH8::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
assert(stackPtr < stackEnd);
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
assert(stackPtr < stackEnd);
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
NodeRef c = node->child(r); c.prefetch(); unsigned int d = ((unsigned int*)&tNear)[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c != BVH8::emptyNode);
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
r = __bscf(mask);
c = node->child(r); c.prefetch(); d = *(unsigned int*)&tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! fallback case if more than 4 children are hit */
while (1)
{
r = __bscf(mask);
assert(stackPtr < stackEnd);
c = node->child(r); c.prefetch(); d = *(unsigned int*)&tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
if (unlikely(mask == 0)) break;
}
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
assert(cur != BVH8::emptyNode);
STAT3(normal.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
size_t lazy_node = 0;
PrimitiveIntersector::intersect(pre,ray,prim,num,bvh->scene,lazy_node);
ray_far = ray.tfar;
if (unlikely(lazy_node)) {
stackPtr->ptr = lazy_node;
stackPtr->dist = inf;
stackPtr++;
}
}
AVX_ZERO_UPPER();
}
void BVH8Intersector1<robust,PrimitiveIntersector>::occluded(const BVH8* bvh, Ray& ray)
{
/*! perform per ray precalculations required by the primitive intersector */
Precalculations pre(ray,bvh);
/*! stack state */
NodeRef stack[stackSize]; //!< stack of nodes that still need to get traversed
NodeRef* stackPtr = stack+1; //!< current stack pointer
NodeRef* stackEnd = stack+stackSize;
stack[0] = bvh->root;
/* filter out invalid rays */
#if defined(RTCORE_IGNORE_INVALID_RAYS)
if (!ray.valid()) return;
#endif
/* verify correct input */
assert(ray.tnear > -FLT_MIN);
//assert(!(types & BVH4::FLAG_NODE_MB) || (ray.time >= 0.0f && ray.time <= 1.0f));
/*! load the ray into SIMD registers */
const Vec3f8 norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const Vec3f8 rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vec3fa ray_org_rdir = ray.org*ray_rdir;
const Vec3f8 org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const float8 ray_near(ray.tnear);
float8 ray_far(ray.tfar);
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray_rdir.x >= 0 ? 0*sizeof(float8) : 1*sizeof(float8);
const size_t nearY = ray_rdir.y >= 0 ? 2*sizeof(float8) : 3*sizeof(float8);
const size_t nearZ = ray_rdir.z >= 0 ? 4*sizeof(float8) : 5*sizeof(float8);
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ sizeof(float8), farY = nearY ^ sizeof(float8), farZ = nearZ ^ sizeof(float8);
#if defined (__AVX2__)
const float8 tNearX = msub(load8f((const char*)node+nearX), rdir.x, org_rdir.x);
const float8 tNearY = msub(load8f((const char*)node+nearY), rdir.y, org_rdir.y);
const float8 tNearZ = msub(load8f((const char*)node+nearZ), rdir.z, org_rdir.z);
const float8 tFarX = msub(load8f((const char*)node+farX ), rdir.x, org_rdir.x);
const float8 tFarY = msub(load8f((const char*)node+farY ), rdir.y, org_rdir.y);
const float8 tFarZ = msub(load8f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const float8 tNearX = (norg.x + load8f((const char*)node+nearX)) * rdir.x;
const float8 tNearY = (norg.y + load8f((const char*)node+nearY)) * rdir.y;
const float8 tNearZ = (norg.z + load8f((const char*)node+nearZ)) * rdir.z;
const float8 tFarX = (norg.x + load8f((const char*)node+farX )) * rdir.x;
const float8 tFarY = (norg.y + load8f((const char*)node+farY )) * rdir.y;
const float8 tFarZ = (norg.z + load8f((const char*)node+farZ )) * rdir.z;
#endif
#if defined(__AVX2__)
const float8 tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,ray_near));
const float8 tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,ray_far ));
const bool8 vmask = cast(tNear) > cast(tFar);
size_t mask = movemask(vmask)^0xff;
#else
const float8 tNear = max(tNearX,tNearY,tNearZ,ray_near);
const float8 tFar = min(tFarX ,tFarY ,tFarZ ,ray_far);
const bool8 vmask = tNear <= tFar;
size_t mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r); cur.prefetch();
assert(cur != BVH8::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); c0.prefetch(); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); c1.prefetch(); const unsigned int d1 = ((unsigned int*)&tNear)[r];
assert(c0 != BVH8::emptyNode);
assert(c1 != BVH8::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { *stackPtr = c1; stackPtr++; cur = c0; continue; }
else { *stackPtr = c0; stackPtr++; cur = c1; continue; }
}
assert(stackPtr < stackEnd);
*stackPtr = c0; stackPtr++;
assert(stackPtr < stackEnd);
*stackPtr = c1; stackPtr++;
/*! three children are hit */
r = __bscf(mask);
cur = node->child(r); cur.prefetch(); *stackPtr = cur; stackPtr++;
if (likely(mask == 0)) {
stackPtr--;
continue;
}
/*! process more than three children */
while(1)
{
r = __bscf(mask);
NodeRef c = node->child(r); c.prefetch(); *stackPtr = c; stackPtr++;
if (unlikely(mask == 0)) break;
}
cur = (NodeRef) stackPtr[-1]; stackPtr--;
}
/*! this is a leaf node */
assert(cur != BVH8::emptyNode);
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
size_t lazy_node = 0;
if (PrimitiveIntersector::occluded(pre,ray,prim,num,bvh->scene,lazy_node)) {
ray.geomID = 0;
break;
}
if (unlikely(lazy_node)) {
*stackPtr = (NodeRef)lazy_node;
stackPtr++;
}
}
AVX_ZERO_UPPER();
}
size_t BVH4MB::rotate(Base* nodeID, size_t depth)
{
/*! nothing to rotate if we reached a leaf node. */
if (nodeID->isLeaf()) return 0;
Node* parent = nodeID->node();
/*! rotate all children first */
ssei cdepth;
for (size_t c=0; c<4; c++)
cdepth[c] = (int)rotate(parent->child[c],depth+1);
/* compute current area of all children */
ssef sizeX = parent->upper_x-parent->lower_x;
ssef sizeY = parent->upper_y-parent->lower_y;
ssef sizeZ = parent->upper_z-parent->lower_z;
ssef childArea = sizeX*(sizeY + sizeZ) + sizeY*sizeZ;
/*! transpose node bounds */
ssef plower0,plower1,plower2,plower3; transpose(parent->lower_x,parent->lower_y,parent->lower_z,ssef(zero),plower0,plower1,plower2,plower3);
ssef pupper0,pupper1,pupper2,pupper3; transpose(parent->upper_x,parent->upper_y,parent->upper_z,ssef(zero),pupper0,pupper1,pupper2,pupper3);
BBox<ssef> other0(plower0,pupper0), other1(plower1,pupper1), other2(plower2,pupper2), other3(plower3,pupper3);
/*! Find best rotation. We pick a target child of a first child,
and swap this with an other child. We perform the best such
swap. */
float bestCost = pos_inf;
int bestChild = -1, bestTarget = -1, bestOther = -1;
for (size_t c=0; c<4; c++)
{
/*! ignore leaf nodes as we cannot descent into */
if (parent->child[c]->isLeaf()) continue;
Node* child = parent->child[c]->node();
/*! transpose child bounds */
ssef clower0,clower1,clower2,clower3; transpose(child->lower_x,child->lower_y,child->lower_z,ssef(zero),clower0,clower1,clower2,clower3);
ssef cupper0,cupper1,cupper2,cupper3; transpose(child->upper_x,child->upper_y,child->upper_z,ssef(zero),cupper0,cupper1,cupper2,cupper3);
BBox<ssef> target0(clower0,cupper0), target1(clower1,cupper1), target2(clower2,cupper2), target3(clower3,cupper3);
/*! put other0 at each target position */
float cost00 = halfArea3f(merge(other0 ,target1,target2,target3));
float cost01 = halfArea3f(merge(target0,other0 ,target2,target3));
float cost02 = halfArea3f(merge(target0,target1,other0 ,target3));
float cost03 = halfArea3f(merge(target0,target1,target2,other0 ));
ssef cost0 = ssef(cost00,cost01,cost02,cost03);
ssef min0 = vreduce_min(cost0);
int pos0 = (int)__bsf(movemask(min0 == cost0));
/*! put other1 at each target position */
float cost10 = halfArea3f(merge(other1 ,target1,target2,target3));
float cost11 = halfArea3f(merge(target0,other1 ,target2,target3));
float cost12 = halfArea3f(merge(target0,target1,other1 ,target3));
float cost13 = halfArea3f(merge(target0,target1,target2,other1 ));
ssef cost1 = ssef(cost10,cost11,cost12,cost13);
ssef min1 = vreduce_min(cost1);
int pos1 = (int)__bsf(movemask(min1 == cost1));
/*! put other2 at each target position */
float cost20 = halfArea3f(merge(other2 ,target1,target2,target3));
float cost21 = halfArea3f(merge(target0,other2 ,target2,target3));
float cost22 = halfArea3f(merge(target0,target1,other2 ,target3));
float cost23 = halfArea3f(merge(target0,target1,target2,other2 ));
ssef cost2 = ssef(cost20,cost21,cost22,cost23);
ssef min2 = vreduce_min(cost2);
int pos2 = (int)__bsf(movemask(min2 == cost2));
/*! put other3 at each target position */
float cost30 = halfArea3f(merge(other3 ,target1,target2,target3));
float cost31 = halfArea3f(merge(target0,other3 ,target2,target3));
float cost32 = halfArea3f(merge(target0,target1,other3 ,target3));
float cost33 = halfArea3f(merge(target0,target1,target2,other3 ));
ssef cost3 = ssef(cost30,cost31,cost32,cost33);
ssef min3 = vreduce_min(cost3);
int pos3 = (int)__bsf(movemask(min3 == cost3));
/*! find best other child */
ssef otherCost = ssef(extract<0>(min0),extract<0>(min1),extract<0>(min2),extract<0>(min3));
int pos[4] = { pos0,pos1,pos2,pos3 };
sseb valid = ssei(int(depth+1))+cdepth <= ssei(maxDepth); // only select swaps that fulfill depth constraints
if (none(valid)) continue;
size_t n = select_min(valid,otherCost);
float cost = otherCost[n]-childArea[c]; //< increasing the original child bound is bad, decreasing good
/*! accept a swap when it reduces cost and is not swapping a node with itself */
if (cost < bestCost && n != c) {
bestCost = cost;
bestChild = (int)c;
bestOther = (int)n;
bestTarget = pos[n];
}
}
/*! if we did not find a swap that improves the SAH then do nothing */
if (bestCost >= 0) return 1+reduce_max(cdepth);
/*! perform the best found tree rotation */
Node* child = parent->child[bestChild]->node();
swap(parent,bestOther,child,bestTarget);
parent->lower_x[bestChild] = reduce_min(child->lower_x);
parent->lower_y[bestChild] = reduce_min(child->lower_y);
parent->lower_z[bestChild] = reduce_min(child->lower_z);
parent->upper_x[bestChild] = reduce_max(child->upper_x);
parent->upper_y[bestChild] = reduce_max(child->upper_y);
parent->upper_z[bestChild] = reduce_max(child->upper_z);
parent->lower_dx[bestChild] = reduce_min(child->lower_dx);
parent->lower_dy[bestChild] = reduce_min(child->lower_dy);
parent->lower_dz[bestChild] = reduce_min(child->lower_dz);
parent->upper_dx[bestChild] = reduce_max(child->upper_dx);
parent->upper_dy[bestChild] = reduce_max(child->upper_dy);
parent->upper_dz[bestChild] = reduce_max(child->upper_dz);
/*! This returned depth is conservative as the child that was
* pulled up in the tree could have been on the critical path. */
cdepth[bestOther]++; // bestOther was pushed down one level
return 1+reduce_max(cdepth);
}
void BVH4Intersector4Hybrid<TriangleIntersector4>::intersect(const BVH4Intersector4Hybrid* This, Ray4& ray, const __m128 valid_i)
{
sseb valid = valid_i;
const BVH4* bvh = This->bvh;
STAT3(normal.travs,1,popcnt(valid),4);
NodeRef invalid = (NodeRef)1;
/* load ray into registers */
ssef ray_near = select(valid,ray.tnear,pos_inf);
ssef ray_far = select(valid,ray.tfar ,neg_inf);
sse3f rdir = rcp_safe(ray.dir);
ray.tfar = ray_far;
/* allocate stack and push root node */
NodeRef stack_node[3*BVH4::maxDepth+1];
ssef stack_near[3*BVH4::maxDepth+1];
stack_node[0] = invalid;
stack_near[0] = ssef(inf);
stack_node[1] = bvh->root;
stack_near[1] = ray_near;
NodeRef* sptr_node = stack_node+2;
ssef * sptr_near = stack_near+2;
while (1)
{
/* pop next node from stack */
sptr_node--;
sptr_near--;
ssef curDist = *sptr_near;
NodeRef curNode = *sptr_node;
if (unlikely(curNode == invalid))
break;
/* cull node if behind closest hit point */
const sseb active = curDist < ray_far;
if (unlikely(none(active)))
continue;
/* switch to single ray traversal */
size_t bits = movemask(active);
if (unlikely(__popcnt(bits) <= SWITCH_THRESHOLD)) {
for (size_t i=__bsf(bits); bits!=0; bits=__btc(bits,i), i=__bsf(bits)) {
BVH4Intersector1<TriangleIntersector1>::intersect1(bvh,curNode,i,ray,rdir);
}
ray_far = ray.tfar;
continue;
}
while (1)
{
/* test if this is a leaf node */
if (unlikely(curNode.isLeaf()))
break;
const Node* const node = curNode.node(bvh->nodePtr()); //NodeRef(curNode).node(nodes);
/* pop of next node */
sptr_node--;
sptr_near--;
curNode = *sptr_node;
curDist = *sptr_near;
for (unsigned i=0; i<4; i++)
{
const ssef dminx = (ssef(node->lower_x[i]) - ray.org.x) * rdir.x;
const ssef dmaxx = (ssef(node->upper_x[i]) - ray.org.x) * rdir.x;
const ssef dminy = (ssef(node->lower_y[i]) - ray.org.y) * rdir.y;
const ssef dmaxy = (ssef(node->upper_y[i]) - ray.org.y) * rdir.y;
const ssef dminz = (ssef(node->lower_z[i]) - ray.org.z) * rdir.z;
const ssef dmaxz = (ssef(node->upper_z[i]) - ray.org.z) * rdir.z;
const NodeRef child = node->child(i);
const ssef dlowerx = min(dminx,dmaxx);
const ssef dupperx = max(dminx,dmaxx);
const ssef dlowery = min(dminy,dmaxy);
const ssef duppery = max(dminy,dmaxy);
const ssef dlowerz = min(dminz,dmaxz);
const ssef dupperz = max(dminz,dmaxz);
const ssef near = max(dlowerx,dlowery,dlowerz,ray_near);
const ssef far = min(dupperx,duppery,dupperz,ray_far );
const sseb mhit = near <= far;
const ssef childDist = select(mhit,near,inf);
const sseb closer = childDist < curDist;
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(mhit)))
{
//if (child != invalid)
{
sptr_node++;
sptr_near++;
/* push cur node onto stack and continue with hit child */
if (any(closer)) {
*(sptr_node-1) = curNode;
*(sptr_near-1) = curDist;
curDist = childDist;
curNode = child;
}
/* push hit child onto stack*/
else {
*(sptr_node-1) = child;
*(sptr_near-1) = childDist;
}
}
}
}
}
/* return if stack is empty */
if (unlikely(curNode == invalid))
break;
/* decode leaf node */
size_t num;
Triangle* tri = (Triangle*) curNode.leaf(bvh->triPtr(),num);
/* intersect triangles */
for (size_t i=0; i<num; i++)
TriangleIntersector4::intersect(valid,ray,tri[i],bvh->vertices);
ray_far = ray.tfar;
}
}
size_t BVHNRotate<4>::rotate(NodeRef parentRef, size_t depth)
{
/*! nothing to rotate if we reached a leaf node. */
if (parentRef.isBarrier()) return 0;
if (parentRef.isLeaf()) return 0;
Node* parent = parentRef.node();
/*! rotate all children first */
vint4 cdepth;
for (size_t c=0; c<4; c++)
cdepth[c] = (int)rotate(parent->child(c),depth+1);
/* compute current areas of all children */
vfloat4 sizeX = parent->upper_x-parent->lower_x;
vfloat4 sizeY = parent->upper_y-parent->lower_y;
vfloat4 sizeZ = parent->upper_z-parent->lower_z;
vfloat4 childArea = sizeX*(sizeY + sizeZ) + sizeY*sizeZ;
/*! get node bounds */
BBox<vfloat4> child1_0,child1_1,child1_2,child1_3;
parent->bounds(child1_0,child1_1,child1_2,child1_3);
/*! Find best rotation. We pick a first child (child1) and a sub-child
(child2child) of a different second child (child2), and swap child1
and child2child. We perform the best such swap. */
float bestArea = 0;
size_t bestChild1 = -1, bestChild2 = -1, bestChild2Child = -1;
for (size_t c2=0; c2<4; c2++)
{
/*! ignore leaf nodes as we cannot descent into them */
if (parent->child(c2).isBarrier()) continue;
if (parent->child(c2).isLeaf()) continue;
Node* child2 = parent->child(c2).node();
/*! transpose child bounds */
BBox<vfloat4> child2c0,child2c1,child2c2,child2c3;
child2->bounds(child2c0,child2c1,child2c2,child2c3);
/*! put child1_0 at each child2 position */
float cost00 = halfArea3f(merge(child1_0,child2c1,child2c2,child2c3));
float cost01 = halfArea3f(merge(child2c0,child1_0,child2c2,child2c3));
float cost02 = halfArea3f(merge(child2c0,child2c1,child1_0,child2c3));
float cost03 = halfArea3f(merge(child2c0,child2c1,child2c2,child1_0));
vfloat4 cost0 = vfloat4(cost00,cost01,cost02,cost03);
vfloat4 min0 = vreduce_min(cost0);
int pos0 = (int)__bsf(movemask(min0 == cost0));
/*! put child1_1 at each child2 position */
float cost10 = halfArea3f(merge(child1_1,child2c1,child2c2,child2c3));
float cost11 = halfArea3f(merge(child2c0,child1_1,child2c2,child2c3));
float cost12 = halfArea3f(merge(child2c0,child2c1,child1_1,child2c3));
float cost13 = halfArea3f(merge(child2c0,child2c1,child2c2,child1_1));
vfloat4 cost1 = vfloat4(cost10,cost11,cost12,cost13);
vfloat4 min1 = vreduce_min(cost1);
int pos1 = (int)__bsf(movemask(min1 == cost1));
/*! put child1_2 at each child2 position */
float cost20 = halfArea3f(merge(child1_2,child2c1,child2c2,child2c3));
float cost21 = halfArea3f(merge(child2c0,child1_2,child2c2,child2c3));
float cost22 = halfArea3f(merge(child2c0,child2c1,child1_2,child2c3));
float cost23 = halfArea3f(merge(child2c0,child2c1,child2c2,child1_2));
vfloat4 cost2 = vfloat4(cost20,cost21,cost22,cost23);
vfloat4 min2 = vreduce_min(cost2);
int pos2 = (int)__bsf(movemask(min2 == cost2));
/*! put child1_3 at each child2 position */
float cost30 = halfArea3f(merge(child1_3,child2c1,child2c2,child2c3));
float cost31 = halfArea3f(merge(child2c0,child1_3,child2c2,child2c3));
float cost32 = halfArea3f(merge(child2c0,child2c1,child1_3,child2c3));
float cost33 = halfArea3f(merge(child2c0,child2c1,child2c2,child1_3));
vfloat4 cost3 = vfloat4(cost30,cost31,cost32,cost33);
vfloat4 min3 = vreduce_min(cost3);
int pos3 = (int)__bsf(movemask(min3 == cost3));
/*! find best other child */
vfloat4 area0123 = vfloat4(extract<0>(min0),extract<0>(min1),extract<0>(min2),extract<0>(min3)) - vfloat4(childArea[c2]);
int pos[4] = { pos0,pos1,pos2,pos3 };
const size_t mbd = BVH4::maxBuildDepth;
vbool4 valid = vint4(int(depth+1))+cdepth <= vint4(mbd); // only select swaps that fulfill depth constraints
valid &= vint4(c2) != vint4(step);
if (none(valid)) continue;
size_t c1 = select_min(valid,area0123);
float area = area0123[c1];
if (c1 == c2) continue; // can happen if bounds are NANs
/*! accept a swap when it reduces cost and is not swapping a node with itself */
if (area < bestArea) {
bestArea = area;
bestChild1 = c1;
bestChild2 = c2;
bestChild2Child = pos[c1];
}
}
/*! if we did not find a swap that improves the SAH then do nothing */
if (bestChild1 == size_t(-1)) return 1+reduce_max(cdepth);
/*! perform the best found tree rotation */
Node* child2 = parent->child(bestChild2).node();
BVH4::swap(parent,bestChild1,child2,bestChild2Child);
parent->set(bestChild2,child2->bounds());
BVH4::compact(parent);
BVH4::compact(child2);
/*! This returned depth is conservative as the child that was
* pulled up in the tree could have been on the critical path. */
cdepth[bestChild1]++; // bestChild1 was pushed down one level
return 1+reduce_max(cdepth);
}
void BVH4Intersector1<PrimitiveIntersector>::intersect(const BVH4* bvh, Ray& ray)
{
/*! stack state */
StackItemInt32<NodeRef> stack[stackSize]; //!< stack of nodes
StackItemInt32<NodeRef>* stackPtr = stack+1; //!< current stack pointer
StackItemInt32<NodeRef>* stackEnd = stack+stackSize;
stack[0].ptr = bvh->root;
stack[0].dist = neg_inf;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray.dir.x >= 0.0f ? 0*sizeof(ssef) : 1*sizeof(ssef);
const size_t nearY = ray.dir.y >= 0.0f ? 2*sizeof(ssef) : 3*sizeof(ssef);
const size_t nearZ = ray.dir.z >= 0.0f ? 4*sizeof(ssef) : 5*sizeof(ssef);
#if 0 // FIXME: why is this slower
/*! load the ray */
Vec3fa ray_org = ray.org;
Vec3fa ray_dir = ray.dir;
ssef ray_near = max(ray.tnear,FLT_MIN); // we do not support negative tnear values in this kernel due to integer optimizations
ssef ray_far = ray.tfar;
#if defined(__FIX_RAYS__)
const float float_range = 0.1f*FLT_MAX;
ray_org = clamp(ray_org,Vec3fa(-float_range),Vec3fa(+float_range));
ray_dir = clamp(ray_dir,Vec3fa(-float_range),Vec3fa(+float_range));
ray_far = min(ray_far,float(inf));
#endif
const Vec3fa ray_rdir = rcp_safe(ray_dir);
const sse3f org(ray_org), dir(ray_dir);
const sse3f norg(-ray_org), rdir(ray_rdir), org_rdir(ray_org*ray_rdir);
#else
/*! load the ray into SIMD registers */
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vec3fa ray_org_rdir = ray.org*ray_rdir;
const sse3f org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const ssef ray_near(ray.tnear);
ssef ray_far(ray.tfar);
#endif
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(*(float*)&stackPtr->dist > ray.tfar))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
#if defined (__AVX2__)
const ssef tNearX = msub(load4f((const char*)node+nearX), rdir.x, org_rdir.x);
const ssef tNearY = msub(load4f((const char*)node+nearY), rdir.y, org_rdir.y);
const ssef tNearZ = msub(load4f((const char*)node+nearZ), rdir.z, org_rdir.z);
const ssef tFarX = msub(load4f((const char*)node+farX ), rdir.x, org_rdir.x);
const ssef tFarY = msub(load4f((const char*)node+farY ), rdir.y, org_rdir.y);
const ssef tFarZ = msub(load4f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const ssef tNearX = (norg.x + load4f((const char*)node+nearX)) * rdir.x;
const ssef tNearY = (norg.y + load4f((const char*)node+nearY)) * rdir.y;
const ssef tNearZ = (norg.z + load4f((const char*)node+nearZ)) * rdir.z;
const ssef tFarX = (norg.x + load4f((const char*)node+farX )) * rdir.x;
const ssef tFarY = (norg.y + load4f((const char*)node+farY )) * rdir.y;
const ssef tFarZ = (norg.z + load4f((const char*)node+farZ )) * rdir.z;
#endif
#if defined(__SSE4_1__)
const ssef tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,ray_near));
const ssef tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,ray_far ));
const sseb vmask = cast(tNear) > cast(tFar);
size_t mask = movemask(vmask)^0xf;
#else
const ssef tNear = max(tNearX,tNearY,tNearZ,ray_near);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,ray_far);
const sseb vmask = tNear <= tFar;
size_t mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
assert(cur != BVH4::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const unsigned int d1 = ((unsigned int*)&tNear)[r];
assert(c0 != BVH4::emptyNode);
assert(c1 != BVH4::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
assert(stackPtr < stackEnd);
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
assert(stackPtr < stackEnd);
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
NodeRef c = node->child(r); unsigned int d = ((unsigned int*)&tNear)[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c != BVH4::emptyNode);
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
c = node->child(r); d = *(unsigned int*)&tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c != BVH4::emptyNode);
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
STAT3(normal.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
PrimitiveIntersector::intersect(ray,prim,num,bvh->geometry);
ray_far = ray.tfar;
}
}
void BVH8Intersector8Hybrid<PrimitiveIntersector8>::occluded(bool8* valid_i, BVH8* bvh, Ray8& ray)
{
/* load ray */
const bool8 valid = *valid_i;
bool8 terminated = !valid;
Vec3f8 ray_org = ray.org, ray_dir = ray.dir;
float8 ray_tnear = ray.tnear, ray_tfar = ray.tfar;
const Vec3f8 rdir = rcp_safe(ray_dir);
const Vec3f8 org(ray_org), org_rdir = org * rdir;
ray_tnear = select(valid,ray_tnear,float8(pos_inf));
ray_tfar = select(valid,ray_tfar ,float8(neg_inf));
const float8 inf = float8(pos_inf);
Precalculations pre(valid,ray);
/* compute near/far per ray */
Vec3i8 nearXYZ;
nearXYZ.x = select(rdir.x >= 0.0f,int8(0*(int)sizeof(float8)),int8(1*(int)sizeof(float8)));
nearXYZ.y = select(rdir.y >= 0.0f,int8(2*(int)sizeof(float8)),int8(3*(int)sizeof(float8)));
nearXYZ.z = select(rdir.z >= 0.0f,int8(4*(int)sizeof(float8)),int8(5*(int)sizeof(float8)));
/* allocate stack and push root node */
float8 stack_near[stackSizeChunk];
NodeRef stack_node[stackSizeChunk];
stack_node[0] = BVH8::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* stackEnd = stack_node+stackSizeChunk;
NodeRef* __restrict__ sptr_node = stack_node + 2;
float8* __restrict__ sptr_near = stack_near + 2;
while (1)
{
/* pop next node from stack */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
NodeRef cur = *sptr_node;
if (unlikely(cur == BVH8::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* cull node if behind closest hit point */
float8 curDist = *sptr_near;
const bool8 active = curDist < ray_tfar;
if (unlikely(none(active)))
continue;
/* switch to single ray traversal */
#if !defined(__WIN32__) || defined(__X86_64__)
size_t bits = movemask(active);
if (unlikely(__popcnt(bits) <= SWITCH_THRESHOLD)) {
for (size_t i=__bsf(bits); bits!=0; bits=__btc(bits,i), i=__bsf(bits)) {
if (occluded1(bvh,cur,i,pre,ray,ray_org,ray_dir,rdir,ray_tnear,ray_tfar,nearXYZ))
terminated[i] = -1;
}
if (all(terminated)) break;
ray_tfar = select(terminated,float8(neg_inf),ray_tfar);
continue;
}
#endif
while (1)
{
/* test if this is a leaf node */
if (unlikely(cur.isLeaf()))
break;
const bool8 valid_node = ray_tfar > curDist;
STAT3(shadow.trav_nodes,1,popcnt(valid_node),8);
const Node* __restrict__ const node = (Node*)cur.node();
/* pop of next node */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
cur = *sptr_node;
curDist = *sptr_near;
for (unsigned i=0; i<BVH8::N; i++)
{
const NodeRef child = node->children[i];
if (unlikely(child == BVH8::emptyNode)) break;
#if defined(__AVX2__)
const float8 lclipMinX = msub(node->lower_x[i],rdir.x,org_rdir.x);
const float8 lclipMinY = msub(node->lower_y[i],rdir.y,org_rdir.y);
const float8 lclipMinZ = msub(node->lower_z[i],rdir.z,org_rdir.z);
const float8 lclipMaxX = msub(node->upper_x[i],rdir.x,org_rdir.x);
const float8 lclipMaxY = msub(node->upper_y[i],rdir.y,org_rdir.y);
const float8 lclipMaxZ = msub(node->upper_z[i],rdir.z,org_rdir.z);
const float8 lnearP = maxi(maxi(mini(lclipMinX, lclipMaxX), mini(lclipMinY, lclipMaxY)), mini(lclipMinZ, lclipMaxZ));
const float8 lfarP = mini(mini(maxi(lclipMinX, lclipMaxX), maxi(lclipMinY, lclipMaxY)), maxi(lclipMinZ, lclipMaxZ));
const bool8 lhit = maxi(lnearP,ray_tnear) <= mini(lfarP,ray_tfar);
#else
const float8 lclipMinX = (node->lower_x[i] - org.x) * rdir.x;
const float8 lclipMinY = (node->lower_y[i] - org.y) * rdir.y;
const float8 lclipMinZ = (node->lower_z[i] - org.z) * rdir.z;
const float8 lclipMaxX = (node->upper_x[i] - org.x) * rdir.x;
const float8 lclipMaxY = (node->upper_y[i] - org.y) * rdir.y;
const float8 lclipMaxZ = (node->upper_z[i] - org.z) * rdir.z;
const float8 lnearP = max(max(min(lclipMinX, lclipMaxX), min(lclipMinY, lclipMaxY)), min(lclipMinZ, lclipMaxZ));
const float8 lfarP = min(min(max(lclipMinX, lclipMaxX), max(lclipMinY, lclipMaxY)), max(lclipMinZ, lclipMaxZ));
const bool8 lhit = max(lnearP,ray_tnear) <= min(lfarP,ray_tfar);
#endif
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
assert(sptr_node < stackEnd);
assert(child != BVH8::emptyNode);
const float8 childDist = select(lhit,lnearP,inf);
sptr_node++;
sptr_near++;
/* push cur node onto stack and continue with hit child */
if (any(childDist < curDist))
{
*(sptr_node-1) = cur;
*(sptr_near-1) = curDist;
curDist = childDist;
cur = child;
}
/* push hit child onto stack */
else {
*(sptr_node-1) = child;
*(sptr_near-1) = childDist;
}
}
}
}
/* return if stack is empty */
if (unlikely(cur == BVH8::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* intersect leaf */
assert(cur != BVH8::emptyNode);
const bool8 valid_leaf = ray_tfar > curDist;
STAT3(shadow.trav_leaves,1,popcnt(valid_leaf),8);
size_t items; const Triangle* prim = (Triangle*) cur.leaf(items);
terminated |= PrimitiveIntersector8::occluded(!terminated,pre,ray,prim,items,bvh->scene);
if (all(terminated)) break;
ray_tfar = select(terminated,float8(neg_inf),ray_tfar);
}
store8i(valid & terminated,&ray.geomID,0);
AVX_ZERO_UPPER();
}
void BVH4iIntersector1<TriangleIntersector>::intersect(const BVH4iIntersector1* This, Ray& ray)
{
AVX_ZERO_UPPER();
STAT3(normal.travs,1,1,1);
/*! stack state */
const BVH4i* bvh = This->bvh;
StackItem stack[1+3*BVH4i::maxDepth]; //!< stack of nodes
StackItem* stackPtr = stack+1; //!< current stack pointer
stack[0].ptr = bvh->root;
stack[0].dist = neg_inf;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray.dir.x >= 0.0f ? 0*sizeof(ssef_m) : 1*sizeof(ssef_m);
const size_t nearY = ray.dir.y >= 0.0f ? 2*sizeof(ssef_m) : 3*sizeof(ssef_m);
const size_t nearZ = ray.dir.z >= 0.0f ? 4*sizeof(ssef_m) : 5*sizeof(ssef_m);
/*! load the ray into SIMD registers */
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vector3f ray_rdir = rcp_safe(ray.dir);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vector3f ray_org_rdir = ray.org*ray_rdir;
const sse3f org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const ssef rayNear(ray.tnear);
ssef rayFar(ray.tfar);
const void* nodePtr = bvh->nodePtr();
const void* triPtr = bvh->triPtr();
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(stackPtr->dist > ray.tfar))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node(nodePtr);
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
#if defined (__AVX2__)
const ssef tNearX = msub(ssef((const char*)nodePtr+(size_t)cur+nearX), rdir.x, org_rdir.x);
const ssef tNearY = msub(ssef((const char*)nodePtr+(size_t)cur+nearY), rdir.y, org_rdir.y);
const ssef tNearZ = msub(ssef((const char*)nodePtr+(size_t)cur+nearZ), rdir.z, org_rdir.z);
const ssef tFarX = msub(ssef((const char*)nodePtr+(size_t)cur+farX ), rdir.x, org_rdir.x);
const ssef tFarY = msub(ssef((const char*)nodePtr+(size_t)cur+farY ), rdir.y, org_rdir.y);
const ssef tFarZ = msub(ssef((const char*)nodePtr+(size_t)cur+farZ ), rdir.z, org_rdir.z);
#else
const ssef tNearX = (norg.x + ssef((const char*)nodePtr+(size_t)cur+nearX)) * rdir.x;
const ssef tNearY = (norg.y + ssef((const char*)nodePtr+(size_t)cur+nearY)) * rdir.y;
const ssef tNearZ = (norg.z + ssef((const char*)nodePtr+(size_t)cur+nearZ)) * rdir.z;
const ssef tFarX = (norg.x + ssef((const char*)nodePtr+(size_t)cur+farX )) * rdir.x;
const ssef tFarY = (norg.y + ssef((const char*)nodePtr+(size_t)cur+farY )) * rdir.y;
const ssef tFarZ = (norg.z + ssef((const char*)nodePtr+(size_t)cur+farZ )) * rdir.z;
#endif
const ssef tNear = max(tNearX,tNearY,tNearZ,rayNear);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,rayFar);
size_t mask = movemask(tNear <= tFar);
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bsf(mask); mask = __btc(mask,r);
if (likely(mask == 0)) {
cur = node->child(r);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const float d0 = tNear[r];
r = __bsf(mask); mask = __btc(mask,r);
NodeRef c1 = node->child(r); const float d1 = tNear[r];
if (likely(mask == 0)) {
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
r = __bsf(mask); mask = __btc(mask,r);
NodeRef c = node->child(r); float d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
r = __bsf(mask); mask = __btc(mask,r);
c = node->child(r); d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
STAT3(normal.trav_leaves,1,1,1);
size_t num; Triangle* tri = (Triangle*) cur.leaf(triPtr,num);
for (size_t i=0; i<num; i++)
TriangleIntersector::intersect(ray,tri[i],bvh->vertices);
rayFar = ray.tfar;
}
AVX_ZERO_UPPER();
}
__forceinline void BVH8iIntersector8Hybrid<TriangleIntersector8>::intersect1(const BVH8i* bvh, NodeRef root, const size_t k, Ray8& ray,const avx3f &ray_org, const avx3f &ray_dir, const avx3f &ray_rdir, const avxf &ray_tnear, const avxf &ray_tfar, const avx3i& nearXYZ)
{
/*! stack state */
StackItemInt64 stack[stackSizeSingle]; //!< stack of nodes
StackItemInt64* stackPtr = stack+1; //!< current stack pointer
StackItemInt64* stackEnd = stack+stackSizeSingle;
stack[0].ptr = root;
stack[0].dist = neg_inf;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = nearXYZ.x[k];
const size_t nearY = nearXYZ.y[k];
const size_t nearZ = nearXYZ.z[k];
/*! load the ray into SIMD registers */
const avx3f org (ray_org .x[k],ray_org .y[k],ray_org .z[k]);
const avx3f rdir(ray_rdir.x[k],ray_rdir.y[k],ray_rdir.z[k]);
const avx3f org_rdir(org*rdir);
avxf rayNear(ray_tnear[k]), rayFar(ray_tfar[k]);
const Node * __restrict__ nodes = (Node *)bvh->nodePtr();
const Triangle * __restrict__ accel = (Triangle*)bvh->triPtr();
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(*(float*)&stackPtr->dist > ray.tfar[k]))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = (Node*)cur.node(nodes);
const size_t farX = nearX ^ sizeof(avxf), farY = nearY ^ sizeof(avxf), farZ = nearZ ^ sizeof(avxf);
#if defined (__AVX2__)
const avxf tNearX = msub(load8f((const char*)node+nearX), rdir.x, org_rdir.x);
const avxf tNearY = msub(load8f((const char*)node+nearY), rdir.y, org_rdir.y);
const avxf tNearZ = msub(load8f((const char*)node+nearZ), rdir.z, org_rdir.z);
const avxf tFarX = msub(load8f((const char*)node+farX ), rdir.x, org_rdir.x);
const avxf tFarY = msub(load8f((const char*)node+farY ), rdir.y, org_rdir.y);
const avxf tFarZ = msub(load8f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const avxf tNearX = (load8f((const char*)node+nearX) - org.x) * rdir.x;
const avxf tNearY = (load8f((const char*)node+nearY) - org.y) * rdir.y;
const avxf tNearZ = (load8f((const char*)node+nearZ) - org.z) * rdir.z;
const avxf tFarX = (load8f((const char*)node+farX ) - org.x) * rdir.x;
const avxf tFarY = (load8f((const char*)node+farY ) - org.y) * rdir.y;
const avxf tFarZ = (load8f((const char*)node+farZ ) - org.z) * rdir.z;
#endif
#if defined(__AVX2__)
const avxf tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,rayNear));
const avxf tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,rayFar ));
const avxb vmask = cast(tNear) > cast(tFar);
unsigned int mask = movemask(vmask)^0xff;
#else
const avxf tNear = max(tNearX,tNearY,tNearZ,rayNear);
const avxf tFar = min(tFarX ,tFarY ,tFarZ ,rayFar);
const avxb vmask = tNear <= tFar;
unsigned int mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
assert(cur != BVH4i::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const unsigned int d1 = ((unsigned int*)&tNear)[r];
assert(c0 != BVH4i::emptyNode);
assert(c1 != BVH4i::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
assert(stackPtr < stackEnd);
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
assert(stackPtr < stackEnd);
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
NodeRef c = node->child(r); unsigned int d = ((unsigned int*)&tNear)[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c0 != BVH4i::emptyNode);
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
c = node->child(r); d = ((unsigned int*)&tNear)[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c != BVH4i::emptyNode);
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
while(1)
{
r = __bscf(mask);
c = node->child(r); d = ((unsigned int*)&tNear)[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
if (unlikely(mask == 0)) break;
}
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
STAT3(normal.trav_leaves,1,1,1);
size_t num; Triangle* prim = (Triangle*) cur.leaf(accel,num);
TriangleIntersector8::intersect(ray,k,prim,num,bvh->geometry);
rayFar = ray.tfar[k];
}
}
__forceinline bool occludedT(const BVH4* bvh, Ray& ray)
{
typedef typename TriangleIntersector::Triangle Triangle;
typedef StackItemT<size_t> StackItem;
typedef typename BVH4::NodeRef NodeRef;
typedef typename BVH4::Node Node;
/*! stack state */
NodeRef stack[1+3*BVH4::maxDepth]; //!< stack of nodes that still need to get traversed
NodeRef* stackPtr = stack+1; //!< current stack pointer
stack[0] = bvh->root;
/*! load the ray into SIMD registers */
const avxf pos_neg = avxf(ssef(+0.0f),ssef(-0.0f));
const avxf neg_pos = avxf(ssef(-0.0f),ssef(+0.0f));
const avxf flipSignX = swapX ? neg_pos : pos_neg;
const avxf flipSignY = swapY ? neg_pos : pos_neg;
const avxf flipSignZ = swapZ ? neg_pos : pos_neg;
const avx3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vector3f ray_rdir = rcp_safe(ray.dir);
const avx3f rdir(ray_rdir.x^flipSignX,ray_rdir.y^flipSignY,ray_rdir.z^flipSignZ);
const avx3f org_rdir(avx3f(ray.org.x,ray.org.y,ray.org.z)*rdir);
const avxf rayNearFar(ssef(ray.tnear),-ssef(ray.tfar));
const void* nodePtr = bvh->nodePtr();
const void* triPtr = bvh->triPtr();
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node(nodePtr);
#if defined (__AVX2__) || defined(__MIC__)
const avxf tLowerUpperX = msub(avxf::load(&node->lower_x), rdir.x, org_rdir.x);
const avxf tLowerUpperY = msub(avxf::load(&node->lower_y), rdir.y, org_rdir.y);
const avxf tLowerUpperZ = msub(avxf::load(&node->lower_z), rdir.z, org_rdir.z);
#else
const avxf tLowerUpperX = (norg.x + avxf::load(&node->lower_x)) * rdir.x;
const avxf tLowerUpperY = (norg.y + avxf::load(&node->lower_y)) * rdir.y;
const avxf tLowerUpperZ = (norg.z + avxf::load(&node->lower_z)) * rdir.z;
#endif
const avxf tNearFarX = swapX ? shuffle<1,0>(tLowerUpperX) : tLowerUpperX;
const avxf tNearFarY = swapY ? shuffle<1,0>(tLowerUpperY) : tLowerUpperY;
const avxf tNearFarZ = swapZ ? shuffle<1,0>(tLowerUpperZ) : tLowerUpperZ;
const avxf tNearFar = max(tNearFarX,tNearFarY,tNearFarZ,rayNearFar);
const ssef tNear = extract<0>(tNearFar);
const ssef tFar = extract<1>(tNearFar);
size_t mask = movemask(-tNear >= tFar);
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bsf(mask); mask = __btc(mask,r);
if (likely(mask == 0)) {
cur = node->child(r);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const float d0 = tNear[r];
r = __bsf(mask); mask = __btc(mask,r);
NodeRef c1 = node->child(r); const float d1 = tNear[r];
if (likely(mask == 0)) {
if (d0 < d1) { *stackPtr = c1; stackPtr++; cur = c0; continue; }
else { *stackPtr = c0; stackPtr++; cur = c1; continue; }
}
*stackPtr = c0; stackPtr++;
*stackPtr = c1; stackPtr++;
/*! three children are hit */
r = __bsf(mask); mask = __btc(mask,r);
cur = node->child(r); *stackPtr = cur; stackPtr++;
if (likely(mask == 0)) {
stackPtr--;
continue;
}
/*! four children are hit */
cur = node->child(3);
}
/*! this is a leaf node */
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Triangle* tri = (Triangle*) cur.leaf(triPtr,num);
for (size_t i=0; i<num; i++) {
if (TriangleIntersector::occluded(ray,tri[i],bvh->vertices)) {
AVX_ZERO_UPPER();
return true;
}
}
}
AVX_ZERO_UPPER();
return false;
}
void BVH8iIntersector8Hybrid<TriangleIntersector8>::occluded(avxb* valid_i, BVH8i* bvh, Ray8& ray)
{
/* load ray */
const avxb valid = *valid_i;
avxb terminated = !valid;
avx3f ray_org = ray.org, ray_dir = ray.dir;
avxf ray_tnear = ray.tnear, ray_tfar = ray.tfar;
#if defined(__FIX_RAYS__)
const avxf float_range = 0.1f*FLT_MAX;
ray_org = clamp(ray_org,avx3f(-float_range),avx3f(+float_range));
ray_dir = clamp(ray_dir,avx3f(-float_range),avx3f(+float_range));
ray_tnear = max(ray_tnear,FLT_MIN);
ray_tfar = min(ray_tfar,float(inf));
#endif
const avx3f rdir = rcp_safe(ray_dir);
const avx3f org(ray_org), org_rdir = org * rdir;
ray_tnear = select(valid,ray_tnear,avxf(pos_inf));
ray_tfar = select(valid,ray_tfar ,avxf(neg_inf));
const avxf inf = avxf(pos_inf);
/* compute near/far per ray */
avx3i nearXYZ;
nearXYZ.x = select(rdir.x >= 0.0f,avxi(0*(int)sizeof(avxf)),avxi(1*(int)sizeof(avxf)));
nearXYZ.y = select(rdir.y >= 0.0f,avxi(2*(int)sizeof(avxf)),avxi(3*(int)sizeof(avxf)));
nearXYZ.z = select(rdir.z >= 0.0f,avxi(4*(int)sizeof(avxf)),avxi(5*(int)sizeof(avxf)));
/* allocate stack and push root node */
avxf stack_near[stackSizeChunk];
NodeRef stack_node[stackSizeChunk];
stack_node[0] = BVH4i::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* stackEnd = stack_node+stackSizeChunk;
NodeRef* __restrict__ sptr_node = stack_node + 2;
avxf* __restrict__ sptr_near = stack_near + 2;
const Node * __restrict__ nodes = (Node *)bvh->nodePtr();
const Triangle * __restrict__ accel = (Triangle*)bvh->triPtr();
while (1)
{
/* pop next node from stack */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
NodeRef curNode = *sptr_node;
if (unlikely(curNode == BVH4i::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* cull node if behind closest hit point */
avxf curDist = *sptr_near;
const avxb active = curDist < ray_tfar;
if (unlikely(none(active)))
continue;
/* switch to single ray traversal */
#if !defined(__WIN32__) || defined(__X86_64__)
size_t bits = movemask(active);
if (unlikely(__popcnt(bits) <= SWITCH_THRESHOLD)) {
for (size_t i=__bsf(bits); bits!=0; bits=__btc(bits,i), i=__bsf(bits)) {
if (occluded1(bvh,curNode,i,ray,ray_org,ray_dir,rdir,ray_tnear,ray_tfar,nearXYZ))
terminated[i] = -1;
}
if (all(terminated)) break;
ray_tfar = select(terminated,avxf(neg_inf),ray_tfar);
continue;
}
#endif
while (1)
{
/* test if this is a leaf node */
if (unlikely(curNode.isLeaf()))
break;
const avxb valid_node = ray_tfar > curDist;
STAT3(shadow.trav_nodes,1,popcnt(valid_node),8);
const Node* __restrict__ const node = (Node*)curNode.node(nodes);
/* pop of next node */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
curNode = *sptr_node;
curDist = *sptr_near;
for (unsigned i=0; i<8; i++)
{
const NodeRef child = node->children[i];
if (unlikely(child == BVH4i::emptyNode)) break;
#if defined(__AVX2__)
const avxf lclipMinX = msub(node->lower_x[i],rdir.x,org_rdir.x);
const avxf lclipMinY = msub(node->lower_y[i],rdir.y,org_rdir.y);
const avxf lclipMinZ = msub(node->lower_z[i],rdir.z,org_rdir.z);
const avxf lclipMaxX = msub(node->upper_x[i],rdir.x,org_rdir.x);
const avxf lclipMaxY = msub(node->upper_y[i],rdir.y,org_rdir.y);
const avxf lclipMaxZ = msub(node->upper_z[i],rdir.z,org_rdir.z);
const avxf lnearP = maxi(maxi(mini(lclipMinX, lclipMaxX), mini(lclipMinY, lclipMaxY)), mini(lclipMinZ, lclipMaxZ));
const avxf lfarP = mini(mini(maxi(lclipMinX, lclipMaxX), maxi(lclipMinY, lclipMaxY)), maxi(lclipMinZ, lclipMaxZ));
const avxb lhit = maxi(lnearP,ray_tnear) <= mini(lfarP,ray_tfar);
#else
const avxf lclipMinX = (node->lower_x[i] - org.x) * rdir.x;
const avxf lclipMinY = (node->lower_y[i] - org.y) * rdir.y;
const avxf lclipMinZ = (node->lower_z[i] - org.z) * rdir.z;
const avxf lclipMaxX = (node->upper_x[i] - org.x) * rdir.x;
const avxf lclipMaxY = (node->upper_y[i] - org.y) * rdir.y;
const avxf lclipMaxZ = (node->upper_z[i] - org.z) * rdir.z;
const avxf lnearP = max(max(min(lclipMinX, lclipMaxX), min(lclipMinY, lclipMaxY)), min(lclipMinZ, lclipMaxZ));
const avxf lfarP = min(min(max(lclipMinX, lclipMaxX), max(lclipMinY, lclipMaxY)), max(lclipMinZ, lclipMaxZ));
const avxb lhit = max(lnearP,ray_tnear) <= min(lfarP,ray_tfar);
#endif
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
assert(sptr_node < stackEnd);
assert(child != BVH4i::emptyNode);
const avxf childDist = select(lhit,lnearP,inf);
sptr_node++;
sptr_near++;
/* push cur node onto stack and continue with hit child */
if (any(childDist < curDist))
{
*(sptr_node-1) = curNode;
*(sptr_near-1) = curDist;
curDist = childDist;
curNode = child;
}
/* push hit child onto stack */
else {
*(sptr_node-1) = child;
*(sptr_near-1) = childDist;
}
}
}
}
/* return if stack is empty */
if (unlikely(curNode == BVH4i::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* intersect leaf */
const avxb valid_leaf = ray_tfar > curDist;
STAT3(shadow.trav_leaves,1,popcnt(valid_leaf),8);
size_t items; const Triangle* prim = (Triangle*) curNode.leaf(accel,items);
terminated |= TriangleIntersector8::occluded(!terminated,ray,prim,items,bvh->geometry);
if (all(terminated)) break;
ray_tfar = select(terminated,avxf(neg_inf),ray_tfar);
}
store8i(valid & terminated,&ray.geomID,0);
AVX_ZERO_UPPER();
}
__forceinline void intersectT(const BVH4* bvh, Ray& ray)
{
typedef typename TriangleIntersector::Triangle Triangle;
typedef StackItemT<size_t> StackItem;
typedef typename BVH4::NodeRef NodeRef;
typedef typename BVH4::Node Node;
/*! stack state */
StackItem stack[1+3*BVH4::maxDepth]; //!< stack of nodes
StackItem* stackPtr = stack+1; //!< current stack pointer
stack[0].ptr = bvh->root;
stack[0].dist = neg_inf;
/*! load the ray into SIMD registers */
const avxf pos_neg = avxf(ssef(+0.0f),ssef(-0.0f));
const avxf neg_pos = avxf(ssef(-0.0f),ssef(+0.0f));
const avxf flipSignX = swapX ? neg_pos : pos_neg;
const avxf flipSignY = swapY ? neg_pos : pos_neg;
const avxf flipSignZ = swapZ ? neg_pos : pos_neg;
const Vector3f ray_rdir = rcp_safe(ray.dir);
const avx3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const avx3f rdir(ray_rdir.x^flipSignX,ray_rdir.y^flipSignY,ray_rdir.z^flipSignZ);
const avx3f org_rdir(avx3f(ray.org.x,ray.org.y,ray.org.z)*rdir);
avxf rayNearFar(ssef(ray.tnear),-ssef(ray.tfar));
const void* nodePtr = bvh->nodePtr();
const void* triPtr = bvh->triPtr();
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(stackPtr->dist > ray.tfar))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node(nodePtr);
#if defined (__AVX2__) || defined(__MIC__)
const avxf tLowerUpperX = msub(avxf::load(&node->lower_x), rdir.x, org_rdir.x);
const avxf tLowerUpperY = msub(avxf::load(&node->lower_y), rdir.y, org_rdir.y);
const avxf tLowerUpperZ = msub(avxf::load(&node->lower_z), rdir.z, org_rdir.z);
#else
const avxf tLowerUpperX = (norg.x + avxf::load(&node->lower_x)) * rdir.x;
const avxf tLowerUpperY = (norg.y + avxf::load(&node->lower_y)) * rdir.y;
const avxf tLowerUpperZ = (norg.z + avxf::load(&node->lower_z)) * rdir.z;
#endif
const avxf tNearFarX = swapX ? shuffle<1,0>(tLowerUpperX) : tLowerUpperX;
const avxf tNearFarY = swapY ? shuffle<1,0>(tLowerUpperY) : tLowerUpperY;
const avxf tNearFarZ = swapZ ? shuffle<1,0>(tLowerUpperZ) : tLowerUpperZ;
const avxf tNearFar = max(tNearFarX,tNearFarY,tNearFarZ,rayNearFar);
const ssef tNear = extract<0>(tNearFar);
const ssef tFar = extract<1>(tNearFar);
size_t mask = movemask(-tNear >= tFar);
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bsf(mask); mask = __btc(mask,r);
if (likely(mask == 0)) {
cur = node->child(r);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const float d0 = tNear[r];
r = __bsf(mask); mask = __btc(mask,r);
NodeRef c1 = node->child(r); const float d1 = tNear[r];
if (likely(mask == 0)) {
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
r = __bsf(mask); mask = __btc(mask,r);
NodeRef c = node->child(r); float d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
r = __bsf(mask); mask = __btc(mask,r);
c = node->child(r); d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
STAT3(normal.trav_leaves,1,1,1);
size_t num; Triangle* tri = (Triangle*) cur.leaf(triPtr,num);
for (size_t i=0; i<num; i++)
TriangleIntersector::intersect(ray,tri[i],bvh->vertices);
rayNearFar = insert<1>(rayNearFar,-ssef(ray.tfar));
}
}
void BVH4Intersector4Hybrid<PrimitiveIntersector4>::intersect(sseb* valid_i, BVH4* bvh, Ray4& ray)
{
/* load ray */
const sseb valid0 = *valid_i;
sse3f ray_org = ray.org, ray_dir = ray.dir;
ssef ray_tnear = ray.tnear, ray_tfar = ray.tfar;
#if defined(__FIX_RAYS__)
const ssef float_range = 0.1f*FLT_MAX;
ray_org = clamp(ray_org,sse3f(-float_range),sse3f(+float_range));
ray_dir = clamp(ray_dir,sse3f(-float_range),sse3f(+float_range));
ray_tnear = max(ray_tnear,FLT_MIN);
ray_tfar = min(ray_tfar,float(inf));
#endif
const sse3f rdir = rcp_safe(ray_dir);
const sse3f org(ray_org), org_rdir = org * rdir;
ray_tnear = select(valid0,ray_tnear,ssef(pos_inf));
ray_tfar = select(valid0,ray_tfar ,ssef(neg_inf));
const ssef inf = ssef(pos_inf);
/* allocate stack and push root node */
ssef stack_near[stackSizeChunk];
NodeRef stack_node[stackSizeChunk];
stack_node[0] = BVH4::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* stackEnd = stack_node+stackSizeChunk;
NodeRef* __restrict__ sptr_node = stack_node + 2;
ssef* __restrict__ sptr_near = stack_near + 2;
while (1)
{
/* pop next node from stack */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
NodeRef curNode = *sptr_node;
if (unlikely(curNode == BVH4::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* cull node if behind closest hit point */
ssef curDist = *sptr_near;
const sseb active = curDist < ray_tfar;
if (unlikely(none(active)))
continue;
/* switch to single ray traversal */
#if !defined(__WIN32__) || defined(__X86_64__)
size_t bits = movemask(active);
if (unlikely(__popcnt(bits) <= SWITCH_THRESHOLD)) {
for (size_t i=__bsf(bits); bits!=0; bits=__btc(bits,i), i=__bsf(bits)) {
intersect1(bvh,curNode,i,ray,ray_org,ray_dir,rdir,ray_tnear,ray_tfar);
}
ray_tfar = ray.tfar;
continue;
}
#endif
while (1)
{
/* test if this is a leaf node */
if (unlikely(curNode.isLeaf()))
break;
const sseb valid_node = ray_tfar > curDist;
STAT3(normal.trav_nodes,1,popcnt(valid_node),4);
const Node* __restrict__ const node = curNode.node();
/* pop of next node */
assert(sptr_node > stack_node);
sptr_node--;
sptr_near--;
curNode = *sptr_node;
curDist = *sptr_near;
#pragma unroll(4)
for (unsigned i=0; i<4; i++)
{
const NodeRef child = node->children[i];
if (unlikely(child == BVH4::emptyNode)) break;
#if defined(__AVX2__)
const ssef lclipMinX = msub(node->lower_x[i],rdir.x,org_rdir.x);
const ssef lclipMinY = msub(node->lower_y[i],rdir.y,org_rdir.y);
const ssef lclipMinZ = msub(node->lower_z[i],rdir.z,org_rdir.z);
const ssef lclipMaxX = msub(node->upper_x[i],rdir.x,org_rdir.x);
const ssef lclipMaxY = msub(node->upper_y[i],rdir.y,org_rdir.y);
const ssef lclipMaxZ = msub(node->upper_z[i],rdir.z,org_rdir.z);
#else
const ssef lclipMinX = (node->lower_x[i] - org.x) * rdir.x;
const ssef lclipMinY = (node->lower_y[i] - org.y) * rdir.y;
const ssef lclipMinZ = (node->lower_z[i] - org.z) * rdir.z;
const ssef lclipMaxX = (node->upper_x[i] - org.x) * rdir.x;
const ssef lclipMaxY = (node->upper_y[i] - org.y) * rdir.y;
const ssef lclipMaxZ = (node->upper_z[i] - org.z) * rdir.z;
#endif
#if defined(__SSE4_1__)
const ssef lnearP = maxi(maxi(mini(lclipMinX, lclipMaxX), mini(lclipMinY, lclipMaxY)), mini(lclipMinZ, lclipMaxZ));
const ssef lfarP = mini(mini(maxi(lclipMinX, lclipMaxX), maxi(lclipMinY, lclipMaxY)), maxi(lclipMinZ, lclipMaxZ));
const sseb lhit = maxi(lnearP,ray_tnear) <= mini(lfarP,ray_tfar);
#else
const ssef lnearP = max(max(min(lclipMinX, lclipMaxX), min(lclipMinY, lclipMaxY)), min(lclipMinZ, lclipMaxZ));
const ssef lfarP = min(min(max(lclipMinX, lclipMaxX), max(lclipMinY, lclipMaxY)), max(lclipMinZ, lclipMaxZ));
const sseb lhit = max(lnearP,ray_tnear) <= min(lfarP,ray_tfar);
#endif
/* if we hit the child we choose to continue with that child if it
is closer than the current next child, or we push it onto the stack */
if (likely(any(lhit)))
{
assert(sptr_node < stackEnd);
const ssef childDist = select(lhit,lnearP,inf);
const NodeRef child = node->children[i];
assert(child != BVH4::emptyNode);
sptr_node++;
sptr_near++;
/* push cur node onto stack and continue with hit child */
if (any(childDist < curDist))
{
*(sptr_node-1) = curNode;
*(sptr_near-1) = curDist;
curDist = childDist;
curNode = child;
}
/* push hit child onto stack */
else {
*(sptr_node-1) = child;
*(sptr_near-1) = childDist;
}
}
}
}
/* return if stack is empty */
if (unlikely(curNode == BVH4::invalidNode)) {
assert(sptr_node == stack_node);
break;
}
/* intersect leaf */
const sseb valid_leaf = ray_tfar > curDist;
STAT3(normal.trav_leaves,1,popcnt(valid_leaf),4);
size_t items; const Primitive* prim = (Primitive*) curNode.leaf(items);
PrimitiveIntersector4::intersect(valid_leaf,ray,prim,items,bvh->geometry);
ray_tfar = select(valid_leaf,ray.tfar,ray_tfar);
}
AVX_ZERO_UPPER();
}
__forceinline void BVH4Intersector4Hybrid<PrimitiveIntersector4>::intersect1(const BVH4* bvh, NodeRef root, size_t k, Ray4& ray,
const sse3f& ray_org, const sse3f& ray_dir, const sse3f& ray_rdir,
const ssef& ray_tnear, const ssef& ray_tfar)
{
/*! stack state */
StackItem stack[stackSizeSingle]; //!< stack of nodes
StackItem* stackPtr = stack+1; //!< current stack pointer
StackItem* stackEnd = stack+stackSizeSingle;
stack[0].ptr = root;
stack[0].dist = neg_inf;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray_dir.x[k] >= 0.0f ? 0*sizeof(ssef) : 1*sizeof(ssef);
const size_t nearY = ray_dir.y[k] >= 0.0f ? 2*sizeof(ssef) : 3*sizeof(ssef);
const size_t nearZ = ray_dir.z[k] >= 0.0f ? 4*sizeof(ssef) : 5*sizeof(ssef);
/*! load the ray into SIMD registers */
const sse3f org (ray_org .x[k],ray_org .y[k],ray_org .z[k]);
const sse3f rdir(ray_rdir.x[k],ray_rdir.y[k],ray_rdir.z[k]);
const sse3f norg = -org, org_rdir(org*rdir);
ssef rayNear(ray_tnear[k]), rayFar(ray_tfar[k]);
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(stackPtr->dist > ray.tfar[k]))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
#if defined (__AVX2__)
const ssef tNearX = msub(load4f((const char*)node+nearX), rdir.x, org_rdir.x);
const ssef tNearY = msub(load4f((const char*)node+nearY), rdir.y, org_rdir.y);
const ssef tNearZ = msub(load4f((const char*)node+nearZ), rdir.z, org_rdir.z);
const ssef tFarX = msub(load4f((const char*)node+farX ), rdir.x, org_rdir.x);
const ssef tFarY = msub(load4f((const char*)node+farY ), rdir.y, org_rdir.y);
const ssef tFarZ = msub(load4f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const ssef tNearX = (norg.x + load4f((const char*)node+nearX)) * rdir.x;
const ssef tNearY = (norg.y + load4f((const char*)node+nearY)) * rdir.y;
const ssef tNearZ = (norg.z + load4f((const char*)node+nearZ)) * rdir.z;
const ssef tFarX = (norg.x + load4f((const char*)node+farX )) * rdir.x;
const ssef tFarY = (norg.y + load4f((const char*)node+farY )) * rdir.y;
const ssef tFarZ = (norg.z + load4f((const char*)node+farZ )) * rdir.z;
#endif
#if defined(__SSE4_1__)
const ssef tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,rayNear));
const ssef tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,rayFar ));
const sseb vmask = cast(tNear) > cast(tFar);
size_t mask = movemask(vmask)^0xf;
#else
const ssef tNear = max(tNearX,tNearY,tNearZ,rayNear);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,rayFar);
const sseb vmask = tNear <= tFar;
size_t mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
assert(cur != BVH4::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const float d0 = tNear[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const float d1 = tNear[r];
assert(c0 != BVH4::emptyNode);
assert(c1 != BVH4::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
assert(stackPtr < stackEnd);
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
assert(stackPtr < stackEnd);
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
NodeRef c = node->child(r); float d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c != BVH4::emptyNode);
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
c = node->child(r); d = tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
assert(c != BVH4::emptyNode);
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
STAT3(normal.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
PrimitiveIntersector4::intersect(ray,k,prim,num,bvh->geometry);
rayFar = ray.tfar[k];
}
}
void BVH4MBIntersector1<TriangleIntersector>::occluded(const BVH4MB* bvh, Ray& ray)
{
AVX_ZERO_UPPER();
STAT3(shadow.travs,1,1,1);
/*! stack state */
Base* stack[1+3*BVH4MB::maxDepth]; //!< stack of nodes that still need to get traversed
Base** stackPtr = stack+1; //!< current stack pointer
stack[0] = bvh->root; //!< push first node onto stack
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = (ray.dir.x >= 0) ? 0*2*sizeof(ssef) : 1*2*sizeof(ssef);
const size_t nearY = (ray.dir.y >= 0) ? 2*2*sizeof(ssef) : 3*2*sizeof(ssef);
const size_t nearZ = (ray.dir.z >= 0) ? 4*2*sizeof(ssef) : 5*2*sizeof(ssef);
const size_t farX = nearX ^ 32;
const size_t farY = nearY ^ 32;
const size_t farZ = nearZ ^ 32;
/*! load the ray into SIMD registers */
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const ssef rayNear(ray.tnear);
const ssef rayFar (ray.tfar);
/*! pop node from stack */
while (true)
{
/* finish when the stack is empty */
if (unlikely(stackPtr == stack)) break;
Base* cur = *(--stackPtr);
/*! this is an inner node */
if (likely(cur->isNode()))
{
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur->node();
const ssef* pNearX = (const ssef*)((const char*)node+nearX);
const ssef* pNearY = (const ssef*)((const char*)node+nearY);
const ssef* pNearZ = (const ssef*)((const char*)node+nearZ);
const ssef tNearX = (norg.x + ssef(pNearX[0]) + ray.time*pNearX[1]) * rdir.x;
const ssef tNearY = (norg.y + ssef(pNearY[0]) + ray.time*pNearY[1]) * rdir.y;
const ssef tNearZ = (norg.z + ssef(pNearZ[0]) + ray.time*pNearZ[1]) * rdir.z;
const ssef tNear = max(tNearX,tNearY,tNearZ,rayNear);
const ssef* pFarX = (const ssef*)((const char*)node+farX);
const ssef* pFarY = (const ssef*)((const char*)node+farY);
const ssef* pFarZ = (const ssef*)((const char*)node+farZ);
const ssef tFarX = (norg.x + ssef(pFarX[0]) + ray.time*pFarX[1]) * rdir.x;
const ssef tFarY = (norg.y + ssef(pFarY[0]) + ray.time*pFarY[1]) * rdir.y;
const ssef tFarZ = (norg.z + ssef(pFarZ[0]) + ray.time*pFarZ[1]) * rdir.z;
const ssef tFar = min(tFarX,tFarY,tFarZ,rayFar);
size_t _hit = movemask(tNear <= tFar);
/*! push hit nodes onto stack */
if (likely(_hit == 0)) continue;
size_t r = __bsf(_hit); _hit = __btc(_hit,r);
*stackPtr = node->child[r]; stackPtr++;
if (likely(_hit == 0)) continue;
r = __bsf(_hit); _hit = __btc(_hit,r);
*stackPtr = node->child[r]; stackPtr++;
if (likely(_hit == 0)) continue;
r = __bsf(_hit); _hit = __btc(_hit,r);
*stackPtr = node->child[r]; stackPtr++;
if (likely(_hit == 0)) continue;
r = __bsf(_hit); _hit = __btc(_hit,r);
*stackPtr = node->child[r]; stackPtr++;
}
/*! this is a leaf node */
else
{
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Triangle* tri = (Triangle*) cur->leaf(num);
for (size_t i=0; i<num; i++)
if (TriangleIntersector::occluded(ray,tri[i],bvh->geometry)) {
ray.geomID = 0;
break;
}
}
}
AVX_ZERO_UPPER();
}
__forceinline bool BVH4Intersector4Hybrid<PrimitiveIntersector4>::occluded1(const BVH4* bvh, NodeRef root, size_t k, Ray4& ray,
const sse3f& ray_org, const sse3f& ray_dir, const sse3f& ray_rdir,
const ssef& ray_tnear, const ssef& ray_tfar)
{
/*! stack state */
NodeRef stack[stackSizeSingle]; //!< stack of nodes that still need to get traversed
NodeRef* stackPtr = stack+1; //!< current stack pointer
NodeRef* stackEnd = stack+stackSizeSingle;
stack[0] = root;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray_dir.x[k] >= 0.0f ? 0*sizeof(ssef) : 1*sizeof(ssef);
const size_t nearY = ray_dir.y[k] >= 0.0f ? 2*sizeof(ssef) : 3*sizeof(ssef);
const size_t nearZ = ray_dir.z[k] >= 0.0f ? 4*sizeof(ssef) : 5*sizeof(ssef);
/*! load the ray into SIMD registers */
const sse3f org (ray_org .x[k],ray_org .y[k],ray_org .z[k]);
const sse3f rdir(ray_rdir.x[k],ray_rdir.y[k],ray_rdir.z[k]);
const sse3f norg = -org, org_rdir(org*rdir);
const ssef rayNear(ray_tnear[k]), rayFar(ray_tfar[k]);
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
#if defined (__AVX2__)
const ssef tNearX = msub(load4f((const char*)node+nearX), rdir.x, org_rdir.x);
const ssef tNearY = msub(load4f((const char*)node+nearY), rdir.y, org_rdir.y);
const ssef tNearZ = msub(load4f((const char*)node+nearZ), rdir.z, org_rdir.z);
const ssef tFarX = msub(load4f((const char*)node+farX ), rdir.x, org_rdir.x);
const ssef tFarY = msub(load4f((const char*)node+farY ), rdir.y, org_rdir.y);
const ssef tFarZ = msub(load4f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const ssef tNearX = (norg.x + load4f((const char*)node+nearX)) * rdir.x;
const ssef tNearY = (norg.y + load4f((const char*)node+nearY)) * rdir.y;
const ssef tNearZ = (norg.z + load4f((const char*)node+nearZ)) * rdir.z;
const ssef tFarX = (norg.x + load4f((const char*)node+farX )) * rdir.x;
const ssef tFarY = (norg.y + load4f((const char*)node+farY )) * rdir.y;
const ssef tFarZ = (norg.z + load4f((const char*)node+farZ )) * rdir.z;
#endif
#if defined(__SSE4_1__)
const ssef tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,rayNear));
const ssef tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,rayFar ));
const sseb vmask = cast(tNear) > cast(tFar);
size_t mask = movemask(vmask)^0xf;
#else
const ssef tNear = max(tNearX,tNearY,tNearZ,rayNear);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,rayFar);
const sseb vmask = tNear <= tFar;
size_t mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
assert(cur != BVH4::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const float d0 = tNear[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const float d1 = tNear[r];
assert(c0 != BVH4::emptyNode);
assert(c1 != BVH4::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { *stackPtr = c1; stackPtr++; cur = c0; continue; }
else { *stackPtr = c0; stackPtr++; cur = c1; continue; }
}
assert(stackPtr < stackEnd);
*stackPtr = c0; stackPtr++;
assert(stackPtr < stackEnd);
*stackPtr = c1; stackPtr++;
/*! three children are hit */
r = __bscf(mask);
cur = node->child(r);
assert(cur != BVH4::emptyNode);
if (likely(mask == 0)) continue;
assert(stackPtr < stackEnd);
*stackPtr = cur; stackPtr++;
/*! four children are hit */
cur = node->child(3);
assert(cur != BVH4::emptyNode);
}
/*! this is a leaf node */
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
if (PrimitiveIntersector4::occluded(ray,k,prim,num,bvh->geometry)) {
ray.geomID[k] = 0;
return true;
}
}
return false;
}
void kdtreebenthin::draw<1>(scene& scene, ray4* r, hit4* hit4)
{
unsigned int signx = movemask(r->D().x());
unsigned int signy = movemask(r->D().y());
unsigned int signz = movemask(r->D().z());
//If the traversal direction is not same for all rays we
// do a single ray traversal
if (((signx - 1) < 14) // sign of x is 0xF or 0
|| ((signy - 1) < 14) // sign of y is 0xF or 0
|| ((signz - 1) < 14)) { // sign of z is 0xF or 0
hit hit[4];
for (int i = 0; i < 4; ++i) {
vec3f d(r->D().x()[i], r->D().y()[i], r->D().z()[i]);
ray ray(r->O(), d);
hit[i].prim = -1;
draw(scene, ray, hit[i]);
}
hit4->prim = ssei(hit[0].prim, hit[1].prim, hit[2].prim, hit[3].prim);
hit4->u = ssef(hit[0].u, hit[1].u, hit[2].u, hit[3].u);
hit4->v = ssef(hit[0].v, hit[1].v, hit[2].v, hit[3].v);
return;
}
ssef tnear, tfar;
_boundingBox.clip(*r, tnear, tfar);
if (movemask(tnear >= tfar) == 0xF)
return;
const unsigned int dir[3][2] = {
{ signx & 1 , 1 - (signx & 1) },
{ signy & 1 , 1 - (signy & 1) },
{ signz & 1 , 1 - (signz & 1) } };
ssef far[MAX_STACK_SIZE];
ssef near[MAX_STACK_SIZE];
int nodes[MAX_STACK_SIZE];
//push dummyNode onto stack which will cause us to exit
nodes[0] = 0;
far[0] = BPRAY_INF;
uint32_t stackptr = 1;
kdnode* currNode = _nodes + 1;
int activemask = 0xF;
#if MAILBOX
static uint64_t rayid = 0;
__sync_add_and_fetch(&rayid, 1);
#endif
while (true) {
if (!currNode->isLeaf()) {
const int axis = currNode->getAxis();
const int front = currNode->getLeft() + dir[axis][0];
const int back = currNode->getLeft() + dir[axis][1];
const ssef dist = currNode->getSplit() - r->O()[axis];
const ssef t = dist * r->rcpD()[axis];
currNode = _nodes + back;
if (!(movemask(tnear <= t) & activemask)) continue;
currNode = _nodes + front;
if (!(movemask(tfar >= t) & activemask)) continue;
nodes[stackptr] = back;
near[stackptr] = max(tnear, t);
far[stackptr] = tfar;
tfar = min(tfar, t);
activemask &= movemask(tnear <= tfar);
++stackptr;
} else {
int primidx = currNode->getPrimitiveOffset();
int primcount = currNode->getNumPrims();
for (int i = 0; i != primcount; ++i) {
int t = _prims[primidx + i];
//prefetch
int t2 = _prims[primidx + i + 1];
_mm_prefetch((char*)&scene._accels[t2], _MM_HINT_T0);
#if MAILBOX
//mailboxing
if (mbox.find(scene, rayid, t)) continue;
#endif
scene.intersect(t, *r, *hit4);
#if MAILBOX
mbox.add(scene, rayid, t);
#endif
}
if (movemask(tfar < r->tfar) == 0) return;
--stackptr;
currNode = nodes[stackptr] + _nodes;
tfar = far[stackptr];
tnear = near[stackptr];
activemask = movemask(tnear <= tfar);
}
}
}
void BVH4Intersector1<PrimitiveIntersector>::occluded(const BVH4* bvh, Ray& ray)
{
/*! stack state */
NodeRef stack[stackSize]; //!< stack of nodes that still need to get traversed
NodeRef* stackPtr = stack+1; //!< current stack pointer
NodeRef* stackEnd = stack+stackSize;
stack[0] = bvh->root;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray.dir.x >= 0 ? 0*sizeof(ssef) : 1*sizeof(ssef);
const size_t nearY = ray.dir.y >= 0 ? 2*sizeof(ssef) : 3*sizeof(ssef);
const size_t nearZ = ray.dir.z >= 0 ? 4*sizeof(ssef) : 5*sizeof(ssef);
#if 0 // FIXME: why is this slower
/*! load the ray */
Vec3fa ray_org = ray.org;
Vec3fa ray_dir = ray.dir;
ssef ray_near = max(ray.tnear,FLT_MIN); // we do not support negative tnear values in this kernel due to integer optimizations
ssef ray_far = ray.tfar;
#if defined(__FIX_RAYS__)
const float float_range = 0.1f*FLT_MAX;
ray_org = clamp(ray_org,Vec3fa(-float_range),Vec3fa(+float_range));
ray_dir = clamp(ray_dir,Vec3fa(-float_range),Vec3fa(+float_range));
ray_far = min(ray_far,float(inf));
#endif
const Vec3fa ray_rdir = rcp_safe(ray_dir);
const sse3f org(ray_org), dir(ray_dir);
const sse3f norg(-ray_org), rdir(ray_rdir), org_rdir(ray_org*ray_rdir);
#else
/*! load the ray into SIMD registers */
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vec3fa ray_org_rdir = ray.org*ray_rdir;
const sse3f org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const ssef ray_near(ray.tnear);
ssef ray_far(ray.tfar);
#endif
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(shadow.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
#if defined (__AVX2__)
const ssef tNearX = msub(load4f((const char*)node+nearX), rdir.x, org_rdir.x);
const ssef tNearY = msub(load4f((const char*)node+nearY), rdir.y, org_rdir.y);
const ssef tNearZ = msub(load4f((const char*)node+nearZ), rdir.z, org_rdir.z);
const ssef tFarX = msub(load4f((const char*)node+farX ), rdir.x, org_rdir.x);
const ssef tFarY = msub(load4f((const char*)node+farY ), rdir.y, org_rdir.y);
const ssef tFarZ = msub(load4f((const char*)node+farZ ), rdir.z, org_rdir.z);
#else
const ssef tNearX = (norg.x + load4f((const char*)node+nearX)) * rdir.x;
const ssef tNearY = (norg.y + load4f((const char*)node+nearY)) * rdir.y;
const ssef tNearZ = (norg.z + load4f((const char*)node+nearZ)) * rdir.z;
const ssef tFarX = (norg.x + load4f((const char*)node+farX )) * rdir.x;
const ssef tFarY = (norg.y + load4f((const char*)node+farY )) * rdir.y;
const ssef tFarZ = (norg.z + load4f((const char*)node+farZ )) * rdir.z;
#endif
#if defined(__SSE4_1__)
const ssef tNear = maxi(maxi(tNearX,tNearY),maxi(tNearZ,ray_near));
const ssef tFar = mini(mini(tFarX ,tFarY ),mini(tFarZ ,ray_far ));
const sseb vmask = cast(tNear) > cast(tFar);
size_t mask = movemask(vmask)^0xf;
#else
const ssef tNear = max(tNearX,tNearY,tNearZ,ray_near);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,ray_far);
const sseb vmask = tNear <= tFar;
size_t mask = movemask(vmask);
#endif
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
assert(cur != BVH4::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const unsigned int d1 = ((unsigned int*)&tNear)[r];
assert(c0 != BVH4::emptyNode);
assert(c1 != BVH4::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { *stackPtr = c1; stackPtr++; cur = c0; continue; }
else { *stackPtr = c0; stackPtr++; cur = c1; continue; }
}
assert(stackPtr < stackEnd);
*stackPtr = c0; stackPtr++;
assert(stackPtr < stackEnd);
*stackPtr = c1; stackPtr++;
/*! three children are hit */
r = __bscf(mask);
cur = node->child(r);
assert(cur != BVH4::emptyNode);
if (likely(mask == 0)) continue;
assert(stackPtr < stackEnd);
*stackPtr = cur; stackPtr++;
/*! four children are hit */
cur = node->child(3);
assert(cur != BVH4::emptyNode);
}
/*! this is a leaf node */
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
if (PrimitiveIntersector::occluded(ray,prim,num,bvh->geometry)) {
ray.geomID = 0;
break;
}
}
AVX_ZERO_UPPER();
}
void BVHNIntersectorKHybrid<N,K,types,robust,PrimitiveIntersectorK,single>::intersect(vint<K>* __restrict__ valid_i, BVH* __restrict__ bvh, RayK<K>& __restrict__ ray, IntersectContext* context)
{
/* filter out invalid rays */
vbool<K> valid = *valid_i == -1;
#if defined(EMBREE_IGNORE_INVALID_RAYS)
valid &= ray.valid();
#endif
/* verify correct input */
assert(all(valid,ray.valid()));
assert(all(valid,ray.tnear >= 0.0f));
assert(!(types & BVH_MB) || all(valid,(ray.time >= 0.0f) & (ray.time <= 1.0f)));
/* if the rays belong to different time segments, immediately switch to single ray traversal */
Precalculations pre(valid,ray,bvh->numTimeSteps);
size_t valid_bits = movemask(valid);
const size_t valid_first = __bsf(valid_bits);
if (unlikely((types & BVH_MB) && valid_bits && (movemask(pre.itime() == pre.itime(valid_first)) != valid_bits)))
{
intersectSingle(valid, bvh, pre, ray, context);
AVX_ZERO_UPPER();
return;
}
/* load ray */
Vec3vfK ray_org = ray.org;
Vec3vfK ray_dir = ray.dir;
vfloat<K> ray_tnear = max(ray.tnear,0.0f);
vfloat<K> ray_tfar = max(ray.tfar ,0.0f);
const Vec3vfK rdir = rcp_safe(ray_dir);
const Vec3vfK org(ray_org), org_rdir = org * rdir;
void BVH4Intersector1Bezier<PrimitiveIntersector>::intersect(const BVH4* bvh, Ray& ray)
{
/*! perform per ray precalculations required by the primitive intersector */
Precalculations pre(ray);
/*! stack state */
StackItemInt32<NodeRef> stack[stackSize]; //!< stack of nodes
StackItemInt32<NodeRef>* stackPtr = stack+1; //!< current stack pointer
StackItemInt32<NodeRef>* stackEnd = stack+stackSize;
stack[0].ptr = bvh->root;
stack[0].dist = neg_inf;
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray.dir.x >= 0.0f ? 0*sizeof(ssef) : 1*sizeof(ssef);
const size_t nearY = ray.dir.y >= 0.0f ? 2*sizeof(ssef) : 3*sizeof(ssef);
const size_t nearZ = ray.dir.z >= 0.0f ? 4*sizeof(ssef) : 5*sizeof(ssef);
/*! load the ray into SIMD registers */
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
const Vec3fa ray_org_rdir = ray.org*ray_rdir;
const sse3f org_rdir(ray_org_rdir.x,ray_org_rdir.y,ray_org_rdir.z);
const ssef ray_near(ray.tnear);
ssef ray_far(ray.tfar);
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/*! if popped node is too far, pop next one */
if (unlikely(*(float*)&stackPtr->dist > ray.tfar))
continue;
/* downtraversal loop */
while (true)
{
/*! stop if we found a leaf */
if (unlikely(cur.isLeaf())) break;
STAT3(normal.trav_nodes,1,1,1);
/*! single ray intersection with 4 boxes */
const Node* node = cur.node();
const size_t farX = nearX ^ 16, farY = nearY ^ 16, farZ = nearZ ^ 16;
const ssef tFarX0 = abs((norg.x + load4f((const char*)node+farX )) * rdir.x);
const ssef tFarY0 = abs((norg.y + load4f((const char*)node+farY )) * rdir.y);
const ssef tFarZ0 = abs((norg.z + load4f((const char*)node+farZ )) * rdir.z);
const ssef tFar0 = min(tFarX0 ,tFarY0 ,tFarZ0);
const ssef radius = abs(ssef(ray.org.w) + tFar0 * ssef(ray.dir.w));
//const ssef radius = zero;
//PRINT2(tFar0,radius);
const ssef tLowerX = (norg.x + node->lower_x - radius) * rdir.x;
const ssef tLowerY = (norg.y + node->lower_y - radius) * rdir.y;
const ssef tLowerZ = (norg.z + node->lower_z - radius) * rdir.z;
const ssef tUpperX = (norg.x + node->upper_x + radius) * rdir.x;
const ssef tUpperY = (norg.y + node->upper_y + radius) * rdir.y;
const ssef tUpperZ = (norg.z + node->upper_z + radius) * rdir.z;
const ssef tNearX = min(tLowerX,tUpperX);
const ssef tNearY = min(tLowerY,tUpperY);
const ssef tNearZ = min(tLowerZ,tUpperZ);
const ssef tFarX = max(tLowerX,tUpperX);
const ssef tFarY = max(tLowerY,tUpperY);
const ssef tFarZ = max(tLowerZ,tUpperZ);
const ssef tNear = max(tNearX,tNearY,tNearZ,ray_near);
const ssef tFar = min(tFarX ,tFarY ,tFarZ ,ray_far);
const sseb vmask = tNear <= tFar;
size_t mask = movemask(vmask);
/*! if no child is hit, pop next node */
if (unlikely(mask == 0))
goto pop;
/*! one child is hit, continue with that child */
size_t r = __bscf(mask);
if (likely(mask == 0)) {
cur = node->child(r);
//assert(cur != BVH4::emptyNode); // FIXME: enable these assertions again, currently traversing empty children
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); const unsigned int d1 = ((unsigned int*)&tNear)[r];
//assert(c0 != BVH4::emptyNode);
//assert(c1 != BVH4::emptyNode);
if (likely(mask == 0)) {
assert(stackPtr < stackEnd);
if (d0 < d1) { stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++; cur = c0; continue; }
else { stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++; cur = c1; continue; }
}
/*! Here starts the slow path for 3 or 4 hit children. We push
* all nodes onto the stack to sort them there. */
assert(stackPtr < stackEnd);
stackPtr->ptr = c0; stackPtr->dist = d0; stackPtr++;
assert(stackPtr < stackEnd);
stackPtr->ptr = c1; stackPtr->dist = d1; stackPtr++;
/*! three children are hit, push all onto stack and sort 3 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
NodeRef c = node->child(r); unsigned int d = ((unsigned int*)&tNear)[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
//assert(c != BVH4::emptyNode);
if (likely(mask == 0)) {
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
continue;
}
/*! four children are hit, push all onto stack and sort 4 stack items, continue with closest child */
assert(stackPtr < stackEnd);
r = __bscf(mask);
c = node->child(r); d = *(unsigned int*)&tNear[r]; stackPtr->ptr = c; stackPtr->dist = d; stackPtr++;
//assert(c != BVH4::emptyNode);
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr; stackPtr--;
}
/*! this is a leaf node */
STAT3(normal.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
PrimitiveIntersector::intersect(pre,ray,prim,num,bvh->geometry);
ray_far = ray.tfar;
}
}
