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();
}
void BVH4Intersector1AVX<TriangleIntersector>::intersect(const BVH4Intersector1AVX* This, Ray& ray)
{
AVX_ZERO_UPPER();
STAT3(normal.travs,1,1,1);
const BVH4* bvh = This->bvh;
int swapX = ray.dir.x < 0.0f;
int swapY = ray.dir.y < 0.0f;
int swapZ = ray.dir.z < 0.0f;
int swap = 4*swapX+2*swapY+swapZ;
switch (swap) {
case 0: intersectT<TriangleIntersector,false,false,false>(bvh,ray); break;
case 1: intersectT<TriangleIntersector,false,false,true >(bvh,ray); break;
case 2: intersectT<TriangleIntersector,false,true ,false>(bvh,ray); break;
case 3: intersectT<TriangleIntersector,false,true ,true >(bvh,ray); break;
case 4: intersectT<TriangleIntersector,true ,false,false>(bvh,ray); break;
case 5: intersectT<TriangleIntersector,true ,false,true >(bvh,ray); break;
case 6: intersectT<TriangleIntersector,true ,true ,false>(bvh,ray); break;
case 7: intersectT<TriangleIntersector,true ,true ,true >(bvh,ray); break;
}
AVX_ZERO_UPPER();
}
void SubdivMeshAVX::interpolateN(const void* valid_i, const unsigned* primIDs, const float* u, const float* v, size_t numUVs,
RTCBufferType buffer, float* P, float* dPdu, float* dPdv, float* ddPdudu, float* ddPdvdv, float* ddPdudv, size_t numFloats)
{
#if defined(DEBUG)
if ((parent->aflags & RTC_INTERPOLATE) == 0)
throw_RTCError(RTC_INVALID_OPERATION,"rtcInterpolate can only get called when RTC_INTERPOLATE is enabled for the scene");
#endif
const int* valid = (const int*) valid_i;
for (size_t i=0; i<numUVs;)
{
if (i+4 >= numUVs)
{
vbool4 valid1 = vint4(int(i))+vint4(step) < vint4(numUVs);
if (valid) valid1 &= vint4::loadu(&valid[i]) == vint4(-1);
if (none(valid1)) { i+=4; continue; }
interpolateHelper(valid1,vint4::loadu(&primIDs[i]),vfloat4::loadu(&u[i]),vfloat4::loadu(&v[i]),numUVs,buffer,
P ? P+i : nullptr,
dPdu ? dPdu+i : nullptr,
dPdv ? dPdv+i : nullptr,
ddPdudu ? ddPdudu+i : nullptr,
ddPdvdv ? ddPdvdv+i : nullptr,
ddPdudv ? ddPdudv+i : nullptr,
numFloats);
i+=4;
}
else
{
vbool8 valid1 = vint8(int(i))+vint8(step) < vint8(int(numUVs));
if (valid) valid1 &= vint8::loadu(&valid[i]) == vint8(-1);
if (none(valid1)) { i+=8; continue; }
interpolateHelper(valid1,vint8::loadu(&primIDs[i]),vfloat8::loadu(&u[i]),vfloat8::loadu(&v[i]),numUVs,buffer,
P ? P+i : nullptr,
dPdu ? dPdu+i : nullptr,
dPdv ? dPdv+i : nullptr,
ddPdudu ? ddPdudu+i : nullptr,
ddPdvdv ? ddPdvdv+i : nullptr,
ddPdudv ? ddPdudv+i : nullptr,
numFloats);
i+=8;
}
}
AVX_ZERO_UPPER();
}
bool BVH4Intersector1AVX<TriangleIntersector>::occluded(const BVH4Intersector1AVX* This, Ray& ray)
{
AVX_ZERO_UPPER();
STAT3(shadow.travs,1,1,1);
const BVH4* bvh = This->bvh;
int swapX = ray.dir.x < 0.0f;
int swapY = ray.dir.y < 0.0f;
int swapZ = ray.dir.z < 0.0f;
int swap = 4*swapX+2*swapY+swapZ;
switch (swap) {
case 0: return occludedT<TriangleIntersector,false,false,false>(bvh,ray); break;
case 1: return occludedT<TriangleIntersector,false,false,true >(bvh,ray); break;
case 2: return occludedT<TriangleIntersector,false,true ,false>(bvh,ray); break;
case 3: return occludedT<TriangleIntersector,false,true ,true >(bvh,ray); break;
case 4: return occludedT<TriangleIntersector,true ,false,false>(bvh,ray); break;
case 5: return occludedT<TriangleIntersector,true ,false,true >(bvh,ray); break;
case 6: return occludedT<TriangleIntersector,true ,true ,false>(bvh,ray); break;
case 7: return occludedT<TriangleIntersector,true ,true ,true >(bvh,ray); break;
default: return false;
}
}
void BVH4Intersector1<types,robust,PrimitiveIntersector>::intersect(const BVH4* bvh, Ray& ray)
{
/*! perform per ray precalculations required by the primitive intersector */
Precalculations pre(ray);
BVH4::UnalignedNodeMB::Precalculations pre1(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;
/*! load the ray into SIMD registers */
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const Vec3fa ray_org_rdir = ray.org*ray_rdir;
const sse3f org(ray.org.x,ray.org.y,ray.org.z);
const sse3f dir(ray.dir.x,ray.dir.y,ray.dir.z);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
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);
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray_rdir.x >= 0.0f ? 0*sizeof(ssef) : 1*sizeof(ssef);
const size_t nearY = ray_rdir.y >= 0.0f ? 2*sizeof(ssef) : 3*sizeof(ssef);
const size_t nearZ = ray_rdir.z >= 0.0f ? 4*sizeof(ssef) : 5*sizeof(ssef);
/* 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)
{
size_t mask;
ssef tNear;
/*! stop if we found a leaf node */
if (unlikely(cur.isLeaf(types))) break;
STAT3(normal.trav_nodes,1,1,1);
/* process standard nodes */
if (likely(cur.isNode(types)))
mask = cur.node()->intersect<robust>(nearX,nearY,nearZ,org,rdir,org_rdir,ray_near,ray_far,tNear);
/* process motion blur nodes */
else if (likely(cur.isNodeMB(types)))
mask = cur.nodeMB()->intersect(nearX,nearY,nearZ,org,rdir,org_rdir,ray_near,ray_far,ray.time,tNear);
/*! process nodes with unaligned bounds */
else if (unlikely(cur.isUnalignedNode(types)))
mask = cur.unalignedNode()->intersect(org,dir,ray_near,ray_far,tNear);
/*! process nodes with unaligned bounds and motion blur */
else if (unlikely(cur.isUnalignedNodeMB(types)))
mask = cur.unalignedNodeMB()->intersect(pre1,org,dir,ray_near,ray_far,ray.time,tNear);
/*! if no child is hit, pop next node */
const BVH4::BaseNode* node = cur.baseNode(types);
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(types);
assert(cur != BVH4::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); c0.prefetch(types); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); c1.prefetch(types); 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); c.prefetch(types); 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); c.prefetch(types); 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 */
assert(cur != BVH4::emptyNode);
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;
}
AVX_ZERO_UPPER();
}
void BVH4Intersector1<types,robust,PrimitiveIntersector>::occluded(const BVH4* bvh, Ray& ray)
{
/*! perform per ray precalculations required by the primitive intersector */
Precalculations pre(ray);
BVH4::UnalignedNodeMB::Precalculations pre1(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;
/*! load the ray into SIMD registers */
const Vec3fa ray_rdir = rcp_safe(ray.dir);
const Vec3fa ray_org_rdir = ray.org*ray_rdir;
const sse3f org(ray.org.x,ray.org.y,ray.org.z);
const sse3f dir(ray.dir.x,ray.dir.y,ray.dir.z);
const sse3f rdir(ray_rdir.x,ray_rdir.y,ray_rdir.z);
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);
/*! offsets to select the side that becomes the lower or upper bound */
const size_t nearX = ray_rdir.x >= 0 ? 0*sizeof(ssef) : 1*sizeof(ssef);
const size_t nearY = ray_rdir.y >= 0 ? 2*sizeof(ssef) : 3*sizeof(ssef);
const size_t nearZ = ray_rdir.z >= 0 ? 4*sizeof(ssef) : 5*sizeof(ssef);
/* pop loop */
while (true) pop:
{
/*! pop next node */
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = (NodeRef) *stackPtr;
/* downtraversal loop */
while (true)
{
size_t mask;
ssef tNear;
/*! stop if we found a leaf node */
if (unlikely(cur.isLeaf(types))) break;
STAT3(shadow.trav_nodes,1,1,1);
/* process standard nodes */
if (likely(cur.isNode(types)))
mask = cur.node()->intersect<robust>(nearX,nearY,nearZ,org,rdir,org_rdir,ray_near,ray_far,tNear);
/* process motion blur nodes */
else if (likely(cur.isNodeMB(types)))
mask = cur.nodeMB()->intersect(nearX,nearY,nearZ,org,rdir,org_rdir,ray_near,ray_far,ray.time,tNear);
/*! process nodes with unaligned bounds */
else if (unlikely(cur.isUnalignedNode(types)))
mask = cur.unalignedNode()->intersect(org,dir,ray_near,ray_far,tNear);
/*! process nodes with unaligned bounds and motion blur */
else if (unlikely(cur.isUnalignedNodeMB(types)))
mask = cur.unalignedNodeMB()->intersect(pre1,org,dir,ray_near,ray_far,ray.time,tNear);
/*! if no child is hit, pop next node */
const BVH4::BaseNode* node = cur.baseNode(types);
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(types);
assert(cur != BVH4::emptyNode);
continue;
}
/*! two children are hit, push far child, and continue with closer child */
NodeRef c0 = node->child(r); c0.prefetch(types); const unsigned int d0 = ((unsigned int*)&tNear)[r];
r = __bscf(mask);
NodeRef c1 = node->child(r); c1.prefetch(types); 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); cur.prefetch(types);
assert(cur != BVH4::emptyNode);
if (likely(mask == 0)) continue;
assert(stackPtr < stackEnd);
*stackPtr = cur; stackPtr++;
/*! four children are hit */
cur = node->child(3); cur.prefetch(types);
assert(cur != BVH4::emptyNode);
}
/*! this is a leaf node */
assert(cur != BVH4::emptyNode);
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Primitive* prim = (Primitive*) cur.leaf(num);
if (PrimitiveIntersector::occluded(pre,ray,prim,num,bvh->geometry)) {
ray.geomID = 0;
break;
}
}
AVX_ZERO_UPPER();
}
void BVH4Intersector4Chunk<PrimitiveIntersector4>::intersect(sseb* valid_i, BVH4* bvh, Ray4& ray)
{
/* load ray */
const sseb valid0 = *valid_i;
const sse3f rdir = rcp_safe(ray.dir);
const sse3f org(ray.org), org_rdir = org * rdir;
ssef ray_tnear = select(valid0,ray.tnear,ssef(pos_inf));
ssef ray_tfar = select(valid0,ray.tfar ,ssef(neg_inf));
const ssef inf = ssef(pos_inf);
Precalculations pre(valid0,ray);
/* allocate stack and push root node */
ssef stack_near[stackSize];
NodeRef stack_node[stackSize];
stack_node[0] = BVH4::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* stackEnd = stack_node+stackSize;
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;
if (unlikely(none(ray_tfar > curDist)))
continue;
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<BVH4::N; 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,pre,ray,prim,items,bvh->geometry);
ray_tfar = select(valid_leaf,ray.tfar,ray_tfar);
}
AVX_ZERO_UPPER();
}
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();
}
void BVH2Intersector<TriangleIntersector>::intersect(const Ray& ray, Hit& hit) const
{
AVX_ZERO_UPPER();
STAT3(normal.travs,1,1,1);
struct StackItem {
Base* ptr; //!< node pointer
float dist; //!< distance of node
};
/*! stack state */
StackItem stack[1+BVH2::maxDepth]; //!< stack of nodes that still need to get traversed
StackItem* stackPtr = stack; //!< current stack pointer
Base* cur = bvh->root; //!< in cur we track the ID of the current node
/*! precomputed shuffles, to switch lower and upper bounds depending on ray direction */
const ssei identity = _mm_set_epi8(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
const ssei swap = _mm_set_epi8( 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8);
const ssei shuffleX = ray.dir.x >= 0 ? identity : swap;
const ssei shuffleY = ray.dir.y >= 0 ? identity : swap;
const ssei shuffleZ = ray.dir.z >= 0 ? identity : swap;
/*! load the ray into SIMD registers */
const ssei pn = ssei(0x00000000,0x00000000,0x80000000,0x80000000);
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const sse3f rdir = sse3f(ssef(ray.rdir.x) ^ pn, ssef(ray.rdir.y) ^ pn, ssef(ray.rdir.z) ^ pn);
ssef nearFar(ray.near, ray.near, -ray.far, -ray.far);
hit.t = min(hit.t,ray.far);
while (true)
{
/*! downtraversal loop */
while (likely(cur->isNode()))
{
/*! single ray intersection with box of both children. */
const Node* node = cur->node();
const ssef tNearFarX = (shuffle8(node->lower_upper_x,shuffleX) + norg.x) * rdir.x;
const ssef tNearFarY = (shuffle8(node->lower_upper_y,shuffleY) + norg.y) * rdir.y;
const ssef tNearFarZ = (shuffle8(node->lower_upper_z,shuffleZ) + norg.z) * rdir.z;
const ssef tNearFar = max(tNearFarX,tNearFarY,tNearFarZ,nearFar) ^ pn;
const sseb lrhit = tNearFar <= shuffle8(tNearFar,swap);
/*! if two children hit, push far node onto stack and continue with closer node */
if (likely(lrhit[0] != 0 && lrhit[1] != 0)) {
if (likely(tNearFar[0] < tNearFar[1])) {
stackPtr->ptr = node->child[1];
stackPtr->dist = tNearFar[1];
cur = node->child[0];
stackPtr++;
}
else {
stackPtr->ptr = node->child[0];
stackPtr->dist = tNearFar[0];
cur = node->child[1];
stackPtr++;
}
}
/*! if one child hit, continue with that child */
else {
if (likely(lrhit[0] != 0)) cur = node->child[0];
else if (likely(lrhit[1] != 0)) cur = node->child[1];
else goto pop_node;
}
}
/*! leaf node, intersect all triangles */
{
STAT3(shadow.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,hit,tri[i],bvh->vertices);
nearFar = shuffle<0,1,2,3>(nearFar,-hit.t);
}
/*! pop next node from stack */
pop_node:
if (unlikely(stackPtr == stack)) break;
--stackPtr;
cur = stackPtr->ptr;
if (unlikely(stackPtr->dist > hit.t)) goto pop_node;
}
AVX_ZERO_UPPER();
}
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();
}
bool BVH2Intersector<TriangleIntersector>::occluded(const Ray& ray) const
{
AVX_ZERO_UPPER();
/*! stack state */
Base* stack[1+BVH2::maxDepth]; //!< stack of nodes that still need to get traversed
Base** stackPtr = stack; //!< current stack pointer
Base* cur = bvh->root; //!< in cur we track the ID of the current node
/*! precomputed shuffles, to switch lower and upper bounds depending on ray direction */
const ssei identity = _mm_set_epi8(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
const ssei swap = _mm_set_epi8( 7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8);
const ssei shuffleX = ray.dir.x >= 0 ? identity : swap;
const ssei shuffleY = ray.dir.y >= 0 ? identity : swap;
const ssei shuffleZ = ray.dir.z >= 0 ? identity : swap;
/*! load the ray into SIMD registers */
const ssei pn = ssei(0x00000000,0x00000000,0x80000000,0x80000000);
const sse3f norg(-ray.org.x,-ray.org.y,-ray.org.z);
const sse3f rdir = sse3f(ssef(ray.rdir.x) ^ pn, ssef(ray.rdir.y) ^ pn, ssef(ray.rdir.z) ^ pn);
ssef nearFar(ray.near, ray.near, -ray.far, -ray.far);
while (true)
{
/*! this is an inner node */
while (likely(cur->isNode()))
{
/*! Single ray intersection with box of both children. See bvh2i.h for node layout. */
const Node* node = cur->node();
const ssef tNearFarX = (shuffle8(node->lower_upper_x,shuffleX) + norg.x) * rdir.x;
const ssef tNearFarY = (shuffle8(node->lower_upper_y,shuffleY) + norg.y) * rdir.y;
const ssef tNearFarZ = (shuffle8(node->lower_upper_z,shuffleZ) + norg.z) * rdir.z;
const ssef tNearFar = max(tNearFarX,tNearFarY,tNearFarZ,nearFar) ^ pn;
const sseb lrhit = tNearFar <= shuffle8(tNearFar,swap);
/*! if two children hit, push far node onto stack and continue with closer node */
if (likely(lrhit[0] != 0 && lrhit[1] != 0)) {
*stackPtr++ = node->child[0]; cur = node->child[1];
}
/*! if one child hit, continue with that child */
else {
if (lrhit[0] != 0) cur = node->child[0];
else if (lrhit[1] != 0) cur = node->child[1];
else goto pop_node;
}
}
/*! leaf node, intersect all triangles */
{
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->vertices)) {
AVX_ZERO_UPPER();
return true;
}
}
/*! pop next node from stack */
pop_node:
if (unlikely(stackPtr == stack)) break;
cur = *(--stackPtr);
}
AVX_ZERO_UPPER();
return false;
}
void BVH4iIntersector4Chunk<TriangleIntersector4>::occluded(sseb* valid_i, BVH4i* bvh, Ray4& ray)
{
/* load node and primitive array */
const Node * __restrict__ nodes = (Node *)bvh->nodePtr();
const Triangle * __restrict__ accel = (Triangle*)bvh->triPtr();
/* load ray */
const sseb valid = *valid_i;
sseb terminated = !valid;
const sse3f rdir = rcp_safe(ray.dir);
const sse3f org_rdir = ray.org * rdir;
ssef ray_tnear = select(valid,ray.tnear,pos_inf);
ssef ray_tfar = select(valid,ray.tfar ,neg_inf);
const ssef inf = ssef(pos_inf);
/* allocate stack and push root node */
ssef stack_near[3*BVH4i::maxDepth+1];
NodeRef stack_node[3*BVH4i::maxDepth+1];
stack_node[0] = BVH4i::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* __restrict__ sptr_node = stack_node + 2;
ssef* __restrict__ sptr_near = stack_near + 2;
while (1)
{
/* pop next node from stack */
sptr_node--;
sptr_near--;
NodeRef curNode = *sptr_node;
if (unlikely(curNode == BVH4i::invalidNode))
break;
/* cull node if behind closest hit point */
ssef curDist = *sptr_near;
if (unlikely(none(ray_tfar > curDist)))
continue;
while (1)
{
/* test if this is a leaf node */
if (unlikely(curNode.isLeaf()))
break;
const sseb valid_node = ray_tfar > curDist;
STAT3(shadow.trav_nodes,1,popcnt(valid_node),4);
const Node* __restrict__ const node = curNode.node(nodes);
/* pop of next node */
sptr_node--;
sptr_near--;
curNode = *sptr_node; // FIXME: this trick creates issues with stack depth
curDist = *sptr_near;
#pragma unroll(4)
for (unsigned i=0; i<4; i++)
{
const NodeRef child = node->children[i];
if (unlikely(child == BVH4i::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);
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 lclipMinX = node->lower_x[i] * rdir.x - org_rdir.x;
const ssef lclipMinY = node->lower_y[i] * rdir.y - org_rdir.y;
const ssef lclipMinZ = node->lower_z[i] * rdir.z - org_rdir.z;
const ssef lclipMaxX = node->upper_x[i] * rdir.x - org_rdir.x;
const ssef lclipMaxY = node->upper_y[i] * rdir.y - org_rdir.y;
const ssef lclipMaxZ = node->upper_z[i] * rdir.z - org_rdir.z;
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)))
{
const ssef 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;
}
assert(sptr_node - stack_node < BVH4i::maxDepth);
}
}
}
/* return if stack is empty */
if (unlikely(curNode == BVH4i::invalidNode))
break;
/* intersect leaf */
const sseb valid_leaf = ray_tfar > curDist;
STAT3(shadow.trav_leaves,1,popcnt(valid_leaf),4);
size_t items; const Triangle* tri = (Triangle*) curNode.leaf(accel, items);
terminated |= TriangleIntersector4::occluded(!terminated,ray,tri,items,bvh->geometry);
if (all(terminated)) break;
ray_tfar = select(terminated,neg_inf,ray_tfar);
}
store4i(valid & terminated,&ray.geomID,0);
AVX_ZERO_UPPER();
}
void BVH8Intersector16Chunk<PrimitiveIntersector16>::occluded(bool16* valid_i, BVH8* bvh, Ray16& ray)
{
#if defined(__AVX512__)
/* load ray */
const bool16 valid = *valid_i;
bool16 terminated = !valid;
const Vec3f16 rdir = rcp_safe(ray.dir);
const Vec3f16 org_rdir = ray.org * rdir;
float16 ray_tnear = select(valid,ray.tnear,pos_inf);
float16 ray_tfar = select(valid,ray.tfar ,neg_inf);
const float16 inf = float16(pos_inf);
Precalculations pre(valid,ray);
/* allocate stack and push root node */
float16 stack_near[3*BVH8::maxDepth+1];
NodeRef stack_node[3*BVH8::maxDepth+1];
stack_node[0] = BVH8::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* __restrict__ sptr_node = stack_node + 2;
float16* __restrict__ sptr_near = stack_near + 2;
while (1)
{
/* pop next node from stack */
sptr_node--;
sptr_near--;
NodeRef cur = *sptr_node;
if (unlikely(cur == BVH8::invalidNode))
break;
/* cull node if behind closest hit point */
float16 curDist = *sptr_near;
if (unlikely(none(ray_tfar > curDist)))
continue;
while (1)
{
/* test if this is a leaf node */
if (unlikely(cur.isLeaf()))
break;
const bool16 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 */
sptr_node--;
sptr_near--;
cur = *sptr_node; // FIXME: this trick creates issues with stack depth
curDist = *sptr_near;
for (unsigned i=0; i<BVH8::N; i++)
{
const NodeRef child = node->children[i];
if (unlikely(child == BVH8::emptyNode)) break;
const float16 lclipMinX = msub(node->lower_x[i],rdir.x,org_rdir.x);
const float16 lclipMinY = msub(node->lower_y[i],rdir.y,org_rdir.y);
const float16 lclipMinZ = msub(node->lower_z[i],rdir.z,org_rdir.z);
const float16 lclipMaxX = msub(node->upper_x[i],rdir.x,org_rdir.x);
const float16 lclipMaxY = msub(node->upper_y[i],rdir.y,org_rdir.y);
const float16 lclipMaxZ = msub(node->upper_z[i],rdir.z,org_rdir.z);
const float16 lnearP = max(max(min(lclipMinX, lclipMaxX), min(lclipMinY, lclipMaxY)), min(lclipMinZ, lclipMaxZ));
const float16 lfarP = min(min(max(lclipMinX, lclipMaxX), max(lclipMinY, lclipMaxY)), max(lclipMinZ, lclipMaxZ));
const bool16 lhit = max(lnearP,ray_tnear) <= min(lfarP,ray_tfar);
/* 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)))
{
const float16 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;
}
assert(sptr_node - stack_node < BVH8::maxDepth);
}
}
}
/* return if stack is empty */
if (unlikely(cur == BVH8::invalidNode))
break;
/* intersect leaf */
assert(cur != BVH8::emptyNode);
const bool16 valid_leaf = ray_tfar > curDist;
STAT3(shadow.trav_leaves,1,popcnt(valid_leaf),8);
size_t items; const Triangle* tri = (Triangle*) cur.leaf(items);
terminated |= PrimitiveIntersector16::occluded(!terminated,pre,ray,tri,items,bvh->scene);
if (all(terminated)) break;
ray_tfar = select(terminated,neg_inf,ray_tfar);
}
store16i(valid & terminated,&ray.geomID,0);
AVX_ZERO_UPPER();
#endif
}
__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 BVH4iIntersector1Scalar<TriangleIntersector>::occluded(const BVH4i* bvh, Ray& ray)
{
/*! stack state */
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;
/*! load the ray into SIMD registers */
const Vec3f rdir = rcp_safe(ray.dir);
const Vec3f org_rdir = ray.org*rdir;
const void* nodePtr = bvh->nodePtr();
const void* triPtr = bvh->triPtr();
/* pop loop */
while (true) pop:
{
if (unlikely(stackPtr == stack)) break;
stackPtr--;
NodeRef cur = NodeRef(stackPtr->ptr);
/* 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);
size_t pushed = 0;
for (size_t i=0;i<4;i++)
{
const float nearX = node->lower_x[i] * rdir.x - org_rdir.x;
const float farX = node->upper_x[i] * rdir.x - org_rdir.x;
const float nearY = node->lower_y[i] * rdir.y - org_rdir.y;
const float farY = node->upper_y[i] * rdir.y - org_rdir.y;
const float nearZ = node->lower_z[i] * rdir.z - org_rdir.z;
const float farZ = node->upper_z[i] * rdir.z - org_rdir.z;
const float tNearX = min(nearX,farX);
const float tFarX = max(nearX,farX);
const float tNearY = min(nearY,farY);
const float tFarY = max(nearY,farY);
const float tNearZ = min(nearZ,farZ);
const float tFarZ = max(nearZ,farZ);
const float tNear = max(tNearX,tNearY,tNearZ,ray.tnear);
const float tFar = min(tFarX ,tFarY ,tFarZ ,ray.tfar);
if (tNear <= tFar)
{
stackPtr->ptr = node->child(i);
stackPtr->dist = tNear;
stackPtr++;
pushed++;
}
}
if (pushed == 0)
{
goto pop;
}
else if (pushed == 1)
{
cur = (NodeRef) stackPtr[-1].ptr;
stackPtr--;
continue;
}
else if (pushed == 2)
{
sort(stackPtr[-1],stackPtr[-2]);
cur = (NodeRef) stackPtr[-1].ptr;
stackPtr--;
continue;
}
else if (pushed == 3)
{
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3]);
cur = (NodeRef) stackPtr[-1].ptr;
stackPtr--;
continue;
}
else
{
sort(stackPtr[-1],stackPtr[-2],stackPtr[-3],stackPtr[-4]);
cur = (NodeRef) stackPtr[-1].ptr;
stackPtr--;
}
}
/*! this is a leaf node */
STAT3(shadow.trav_leaves,1,1,1);
size_t num; Triangle1* tri = (Triangle1*) cur.leaf(triPtr,num);
for (size_t i=0;i<num;i++)
if (occluded_vec3f(ray,tri[i],bvh->geometry))
{
ray.geomID = 0;
break;
}
if (ray.geomID == 0) break;
}
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();
}
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 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();
}
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();
}
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();
}
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 BVH8Intersector8Chunk<PrimitiveIntersector8>::intersect(avxb* valid_i, BVH8* bvh, Ray8& ray)
{
#if defined(__AVX__)
/* load ray */
const avxb valid0 = *valid_i;
const avx3f rdir = rcp_safe(ray.dir);
const avx3f org_rdir = ray.org * rdir;
avxf ray_tnear = select(valid0,ray.tnear,pos_inf);
avxf ray_tfar = select(valid0,ray.tfar ,neg_inf);
const avxf inf = avxf(pos_inf);
Precalculations pre(valid0,ray);
/* allocate stack and push root node */
avxf stack_near[3*BVH8::maxDepth+1];
NodeRef stack_node[3*BVH8::maxDepth+1];
stack_node[0] = BVH8::invalidNode;
stack_near[0] = inf;
stack_node[1] = bvh->root;
stack_near[1] = ray_tnear;
NodeRef* __restrict__ sptr_node = stack_node + 2;
avxf* __restrict__ sptr_near = stack_near + 2;
while (1)
{
/* pop next node from stack */
sptr_node--;
sptr_near--;
NodeRef cur = *sptr_node;
if (unlikely(cur == BVH8::invalidNode))
break;
/* cull node if behind closest hit point */
avxf curDist = *sptr_near;
if (unlikely(none(ray_tfar > curDist)))
continue;
while (1)
{
/* test if this is a leaf node */
if (unlikely(cur.isLeaf()))
break;
const avxb valid_node = ray_tfar > curDist;
STAT3(normal.trav_nodes,1,popcnt(valid_node),8);
const Node* __restrict__ const node = (BVH8::Node*)cur.node();
/* pop of next node */
sptr_node--;
sptr_near--;
cur = *sptr_node; // FIXME: this trick creates issues with stack depth
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 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] * rdir.x - org_rdir.x;
const avxf lclipMinY = node->lower_y[i] * rdir.y - org_rdir.y;
const avxf lclipMinZ = node->lower_z[i] * rdir.z - org_rdir.z;
const avxf lclipMaxX = node->upper_x[i] * rdir.x - org_rdir.x;
const avxf lclipMaxY = node->upper_y[i] * rdir.y - org_rdir.y;
const avxf lclipMaxZ = node->upper_z[i] * rdir.z - org_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)))
{
const avxf childDist = select(lhit,lnearP,inf);
const NodeRef child = node->children[i];
/* push cur node onto stack and continue with hit child */
if (any(childDist < curDist))
{
*sptr_node = cur;
*sptr_near = curDist;
sptr_node++;
sptr_near++;
curDist = childDist;
cur = child;
}
/* push hit child onto stack*/
else {
*sptr_node = child;
*sptr_near = childDist;
sptr_node++;
sptr_near++;
}
assert(sptr_node - stack_node < BVH8::maxDepth);
}
}
}
/* return if stack is empty */
if (unlikely(cur == BVH8::invalidNode))
break;
/* intersect leaf */
assert(cur != BVH8::emptyNode);
const avxb valid_leaf = ray_tfar > curDist;
STAT3(normal.trav_leaves,1,popcnt(valid_leaf),8);
size_t items; const Triangle* tri = (Triangle*) cur.leaf(items);
PrimitiveIntersector8::intersect(valid_leaf,pre,ray,tri,items,bvh->geometry);
ray_tfar = select(valid_leaf,ray.tfar,ray_tfar);
}
AVX_ZERO_UPPER();
#endif
}
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 SubdivMeshAVX::interpolate(unsigned primID, float u, float v, RTCBufferType buffer, float* P, float* dPdu, float* dPdv, float* ddPdudu, float* ddPdvdv, float* ddPdudv, size_t numFloats)
{
#if defined(DEBUG)
if ((parent->aflags & RTC_INTERPOLATE) == 0)
throw_RTCError(RTC_INVALID_OPERATION,"rtcInterpolate can only get called when RTC_INTERPOLATE is enabled for the scene");
#endif
/* calculate base pointer and stride */
assert((buffer >= RTC_VERTEX_BUFFER0 && buffer <= RTC_VERTEX_BUFFER1) ||
(buffer >= RTC_USER_VERTEX_BUFFER0 && buffer <= RTC_USER_VERTEX_BUFFER1));
const char* src = nullptr;
size_t stride = 0;
size_t bufID = buffer&0xFFFF;
std::vector<SharedLazyTessellationCache::CacheEntry>* baseEntry = nullptr;
if (buffer >= RTC_USER_VERTEX_BUFFER0) {
src = userbuffers[buffer&0xFFFF]->getPtr();
stride = userbuffers[buffer&0xFFFF]->getStride();
baseEntry = &user_buffer_tags[bufID];
} else {
src = vertices[buffer&0xFFFF].getPtr();
stride = vertices[buffer&0xFFFF].getStride();
baseEntry = &vertex_buffer_tags[bufID];
}
for (size_t i=0,slot=0; i<numFloats; slot++)
{
if (i+4 >= numFloats)
{
vfloat4 Pt, dPdut, dPdvt, ddPdudut, ddPdvdvt, ddPdudvt;;
isa::PatchEval<vfloat4>(baseEntry->at(interpolationSlot(primID,slot,stride)),parent->commitCounterSubdiv,
getHalfEdge(primID),src+i*sizeof(float),stride,u,v,
P ? &Pt : nullptr,
dPdu ? &dPdut : nullptr,
dPdv ? &dPdvt : nullptr,
ddPdudu ? &ddPdudut : nullptr,
ddPdvdv ? &ddPdvdvt : nullptr,
ddPdudv ? &ddPdudvt : nullptr);
if (P) {
for (size_t j=i; j<min(i+4,numFloats); j++)
P[j] = Pt[j-i];
}
if (dPdu)
{
for (size_t j=i; j<min(i+4,numFloats); j++) {
dPdu[j] = dPdut[j-i];
dPdv[j] = dPdvt[j-i];
}
}
if (ddPdudu)
{
for (size_t j=i; j<min(i+4,numFloats); j++) {
ddPdudu[j] = ddPdudut[j-i];
ddPdvdv[j] = ddPdvdvt[j-i];
ddPdudv[j] = ddPdudvt[j-i];
}
}
i+=4;
}
else
{
vfloat8 Pt, dPdut, dPdvt, ddPdudut, ddPdvdvt, ddPdudvt;
isa::PatchEval<vfloat8>(baseEntry->at(interpolationSlot(primID,slot,stride)),parent->commitCounterSubdiv,
getHalfEdge(primID),src+i*sizeof(float),stride,u,v,
P ? &Pt : nullptr,
dPdu ? &dPdut : nullptr,
dPdv ? &dPdvt : nullptr,
ddPdudu ? &ddPdudut : nullptr,
ddPdvdv ? &ddPdvdvt : nullptr,
ddPdudv ? &ddPdudvt : nullptr);
if (P) {
for (size_t j=i; j<i+8; j++)
P[j] = Pt[j-i];
}
if (dPdu)
{
for (size_t j=i; j<i+8; j++) {
dPdu[j] = dPdut[j-i];
dPdv[j] = dPdvt[j-i];
}
}
if (ddPdudu)
{
for (size_t j=i; j<i+8; j++) {
ddPdudu[j] = ddPdudut[j-i];
ddPdvdv[j] = ddPdvdvt[j-i];
ddPdudv[j] = ddPdudvt[j-i];
}
}
i+=8;
}
}
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();
}
