FallBackSplit FallBackSplit::find(size_t threadIndex, PrimRefBlockAlloc<PrimRef>& alloc, PrimRefList& prims,
PrimRefList& lprims_o, PrimInfo& linfo_o,
PrimRefList& rprims_o, PrimInfo& rinfo_o)
{
size_t num = 0;
BBox3fa lbounds = empty, rbounds = empty;
PrimRefList::item* lblock = lprims_o.insert(alloc.malloc(threadIndex));
PrimRefList::item* rblock = rprims_o.insert(alloc.malloc(threadIndex));
linfo_o.reset();
rinfo_o.reset();
while (PrimRefList::item* block = prims.take())
{
for (size_t i=0; i<block->size(); i++)
{
const PrimRef& prim = block->at(i);
const BBox3fa bounds = prim.bounds();
if ((num++)%2)
{
linfo_o.add(bounds,prim.center2());
if (likely(lblock->insert(prim))) continue;
lblock = lprims_o.insert(alloc.malloc(threadIndex));
lblock->insert(prim);
}
else
{
rinfo_o.add(bounds,prim.center2());
if (likely(rblock->insert(prim))) continue;
rblock = rprims_o.insert(alloc.malloc(threadIndex));
rblock->insert(prim);
}
}
alloc.free(threadIndex,block);
}
return FallBackSplit(linfo_o.geomBounds,linfo_o.size(),rinfo_o.geomBounds,rinfo_o.size());
}
void StrandSplit::Split::split<false>(size_t threadIndex, size_t threadCount, LockStepTaskScheduler* scheduler, PrimRefBlockAlloc<BezierPrim>& alloc,
BezierRefList& prims,
BezierRefList& lprims_o, PrimInfo& linfo_o,
BezierRefList& rprims_o, PrimInfo& rinfo_o) const
{
BezierRefList::item* lblock = lprims_o.insert(alloc.malloc(threadIndex));
BezierRefList::item* rblock = rprims_o.insert(alloc.malloc(threadIndex));
linfo_o.reset();
rinfo_o.reset();
while (BezierRefList::item* block = prims.take())
{
for (size_t i=0; i<block->size(); i++)
{
const BezierPrim& prim = block->at(i);
const Vec3fa axisi = normalize(prim.p3-prim.p0);
const float cos0 = abs(dot(axisi,axis0));
const float cos1 = abs(dot(axisi,axis1));
if (cos0 > cos1)
{
linfo_o.add(prim.bounds(),prim.center());
if (likely(lblock->insert(prim))) continue;
lblock = lprims_o.insert(alloc.malloc(threadIndex));
lblock->insert(prim);
}
else
{
rinfo_o.add(prim.bounds(),prim.center());
if (likely(rblock->insert(prim))) continue;
rblock = rprims_o.insert(alloc.malloc(threadIndex));
rblock->insert(prim);
}
}
alloc.free(threadIndex,block);
}
}
void BVH4BuilderTwoLevel::build(size_t threadIndex, size_t threadCount)
{
/* delete some objects */
size_t N = scene->size();
if (N < objects.size()) {
parallel_for(N, objects.size(), [&] (const range<size_t>& r) {
for (size_t i=r.begin(); i<r.end(); i++) {
delete builders[i]; builders[i] = nullptr;
delete objects[i]; objects[i] = nullptr;
}
});
}
/* reset memory allocator */
bvh->alloc.reset();
/* skip build for empty scene */
const size_t numPrimitives = scene->getNumPrimitives<TriangleMesh,1>();
if (numPrimitives == 0) {
prims.resize(0);
bvh->set(BVH4::emptyNode,empty,0);
return;
}
double t0 = bvh->preBuild(TOSTRING(isa) "::BVH4BuilderTwoLevel");
#if PROFILE
profile(2,20,numPrimitives,[&] (ProfileTimer& timer)
{
#endif
/* resize object array if scene got larger */
if (objects.size() < N) objects.resize(N);
if (builders.size() < N) builders.resize(N);
if (refs.size() < N) refs.resize(N);
nextRef = 0;
/* create of acceleration structures */
parallel_for(size_t(0), N, [&] (const range<size_t>& r)
{
for (size_t objectID=r.begin(); objectID<r.end(); objectID++)
{
TriangleMesh* mesh = scene->getTriangleMeshSafe(objectID);
/* verify meshes got deleted properly */
if (mesh == nullptr || mesh->numTimeSteps != 1) {
assert(objectID < objects.size () && objects[objectID] == nullptr);
assert(objectID < builders.size() && builders[objectID] == nullptr);
continue;
}
/* create BVH and builder for new meshes */
if (objects[objectID] == nullptr)
createTriangleMeshAccel(mesh,(AccelData*&)objects[objectID],builders[objectID]);
}
});
/* parallel build of acceleration structures */
parallel_for(size_t(0), N, [&] (const range<size_t>& r)
{
for (size_t objectID=r.begin(); objectID<r.end(); objectID++)
{
/* ignore if no triangle mesh or not enabled */
TriangleMesh* mesh = scene->getTriangleMeshSafe(objectID);
if (mesh == nullptr || !mesh->isEnabled() || mesh->numTimeSteps != 1)
continue;
BVH4* object = objects [objectID]; assert(object);
Builder* builder = builders[objectID]; assert(builder);
/* build object if it got modified */
#if !PROFILE
if (mesh->isModified())
#endif
builder->build(0,0);
/* create build primitive */
if (!object->bounds.empty())
refs[nextRef++] = BVH4BuilderTwoLevel::BuildRef(object->bounds,object->root);
}
});
/* fast path for single geometry scenes */
if (nextRef == 1) {
bvh->set(refs[0].node,refs[0].bounds(),numPrimitives);
return;
}
/* open all large nodes */
refs.resize(nextRef);
open_sequential(numPrimitives);
/* fast path for small geometries */
if (refs.size() == 1) {
bvh->set(refs[0].node,refs[0].bounds(),numPrimitives);
return;
}
/* compute PrimRefs */
prims.resize(refs.size());
const PrimInfo pinfo = parallel_reduce(size_t(0), refs.size(), size_t(1024), PrimInfo(empty), [&] (const range<size_t>& r) -> PrimInfo
{
PrimInfo pinfo(empty);
for (size_t i=r.begin(); i<r.end(); i++) {
pinfo.add(refs[i].bounds());
prims[i] = PrimRef(refs[i].bounds(),(size_t)refs[i].node);
}
return pinfo;
}, [] (const PrimInfo& a, const PrimInfo& b) { return PrimInfo::merge(a,b); });
/* skip if all objects where empty */
if (pinfo.size() == 0)
bvh->set(BVH4::emptyNode,empty,0);
/* otherwise build toplevel hierarchy */
else
{
BVH4::NodeRef root;
BVHBuilderBinnedSAH::build<BVH4::NodeRef>
(root,
[&] { return bvh->alloc.threadLocal2(); },
[&] (const isa::BVHBuilderBinnedSAH::BuildRecord& current, BVHBuilderBinnedSAH::BuildRecord* children, const size_t N, FastAllocator::ThreadLocal2* alloc) -> int
{
BVH4::Node* node = (BVH4::Node*) alloc->alloc0.malloc(sizeof(BVH4::Node)); node->clear();
for (size_t i=0; i<N; i++) {
node->set(i,children[i].pinfo.geomBounds);
children[i].parent = (size_t*)&node->child(i);
}
*current.parent = bvh->encodeNode(node);
return 0;
},
[&] (const BVHBuilderBinnedSAH::BuildRecord& current, FastAllocator::ThreadLocal2* alloc) -> int
{
assert(current.prims.size() == 1);
*current.parent = (BVH4::NodeRef) prims[current.prims.begin()].ID();
return 1;
},
[&] (size_t dn) { bvh->scene->progressMonitor(0); },
prims.data(),pinfo,BVH4::N,BVH4::maxBuildDepthLeaf,4,1,1,1.0f,1.0f);
bvh->set(root,pinfo.geomBounds,numPrimitives);
}
#if PROFILE
});
#endif
bvh->alloc.cleanup();
bvh->postBuild(t0);
}
void SpatialSplit::Split::split<false>(size_t threadIndex, size_t threadCount, LockStepTaskScheduler* scheduler, PrimRefBlockAlloc<PrimRef>& alloc,
Scene* scene, TriRefList& prims,
TriRefList& lprims_o, PrimInfo& linfo_o,
TriRefList& rprims_o, PrimInfo& rinfo_o) const
{
assert(valid());
TriRefList::item* lblock = lprims_o.insert(alloc.malloc(threadIndex));
TriRefList::item* rblock = rprims_o.insert(alloc.malloc(threadIndex));
linfo_o.reset();
rinfo_o.reset();
/* sort each primitive to left, right, or left and right */
while (atomic_set<PrimRefBlock>::item* block = prims.take())
{
for (size_t i=0; i<block->size(); i++)
{
const PrimRef& prim = block->at(i);
const BBox3fa bounds = prim.bounds();
const int bin0 = mapping.bin(bounds.lower)[dim];
const int bin1 = mapping.bin(bounds.upper)[dim];
/* sort to the left side */
if (bin1 < pos)
{
linfo_o.add(bounds,center2(bounds));
if (likely(lblock->insert(prim))) continue;
lblock = lprims_o.insert(alloc.malloc(threadIndex));
lblock->insert(prim);
continue;
}
/* sort to the right side */
if (bin0 >= pos)
{
rinfo_o.add(bounds,center2(bounds));
if (likely(rblock->insert(prim))) continue;
rblock = rprims_o.insert(alloc.malloc(threadIndex));
rblock->insert(prim);
continue;
}
/* split and sort to left and right */
TriangleMesh* mesh = (TriangleMesh*) scene->get(prim.geomID());
TriangleMesh::Triangle tri = mesh->triangle(prim.primID());
const Vec3fa v0 = mesh->vertex(tri.v[0]);
const Vec3fa v1 = mesh->vertex(tri.v[1]);
const Vec3fa v2 = mesh->vertex(tri.v[2]);
PrimRef left,right;
float fpos = mapping.pos(pos,dim);
splitTriangle(prim,dim,fpos,v0,v1,v2,left,right);
if (!left.bounds().empty()) {
linfo_o.add(left.bounds(),center2(left.bounds()));
if (!lblock->insert(left)) {
lblock = lprims_o.insert(alloc.malloc(threadIndex));
lblock->insert(left);
}
}
if (!right.bounds().empty()) {
rinfo_o.add(right.bounds(),center2(right.bounds()));
if (!rblock->insert(right)) {
rblock = rprims_o.insert(alloc.malloc(threadIndex));
rblock->insert(right);
}
}
}
alloc.free(threadIndex,block);
}
}
void SpatialSplit::Split::split<false>(size_t threadIndex, size_t threadCount, LockStepTaskScheduler* scheduler, PrimRefBlockAlloc<BezierPrim>& alloc,
Scene* scene, BezierRefList& prims,
BezierRefList& lprims_o, PrimInfo& linfo_o,
BezierRefList& rprims_o, PrimInfo& rinfo_o) const
{
assert(valid());
BezierRefList::item* lblock = lprims_o.insert(alloc.malloc(threadIndex));
BezierRefList::item* rblock = rprims_o.insert(alloc.malloc(threadIndex));
linfo_o.reset();
rinfo_o.reset();
while (BezierRefList::item* block = prims.take())
{
for (size_t i=0; i<block->size(); i++)
{
const BezierPrim& prim = block->at(i);
const int bin0 = mapping.bin(min(prim.p0,prim.p3))[dim];
const int bin1 = mapping.bin(max(prim.p0,prim.p3))[dim];
/* sort to the left side */
if (bin0 < pos && bin1 < pos) // FIXME: optimize
{
linfo_o.add(prim.bounds(),prim.center());
if (likely(lblock->insert(prim))) continue;
lblock = lprims_o.insert(alloc.malloc(threadIndex));
lblock->insert(prim);
continue;
}
/* sort to the right side */
if (bin0 >= pos && bin1 >= pos)// FIXME: optimize
{
rinfo_o.add(prim.bounds(),prim.center());
if (likely(rblock->insert(prim))) continue;
rblock = rprims_o.insert(alloc.malloc(threadIndex));
rblock->insert(prim);
continue;
}
/* split and sort to left and right */
BezierPrim left,right;
float fpos = mapping.pos(pos,dim);
if (prim.split(dim,fpos,left,right))
{
if (!left.bounds().empty()) {
linfo_o.add(left.bounds(),left.center());
if (!lblock->insert(left)) {
lblock = lprims_o.insert(alloc.malloc(threadIndex));
lblock->insert(left);
}
}
if (!right.bounds().empty()) {
rinfo_o.add(right.bounds(),right.center());
if (!rblock->insert(right)) {
rblock = rprims_o.insert(alloc.malloc(threadIndex));
rblock->insert(right);
}
}
continue;
}
/* insert to left side as fallback */
linfo_o.add(prim.bounds(),prim.center());
if (!lblock->insert(prim)) {
lblock = lprims_o.insert(alloc.malloc(threadIndex));
lblock->insert(prim);
}
}
alloc.free(threadIndex,block);
}
}
