void BVH4BuilderTopLevel::recurseSAH(size_t depth, BuildRecord& task, const size_t mode, const size_t threadID, const size_t numThreads)
{
/* return leaf node */
assert(task.end-task.begin > 0);
if (unlikely(task.end-task.begin == 1)) {
*(NodeRef*)task.parentNode = refs[task.begin].node;
return;
}
/* create leaf node */
if (unlikely(task.depth >= BVH4::maxBuildDepth)) {
createLeaf(task,threadID,numThreads);
return;
}
/*! initialize task list */
BuildRecord childTasks[4];
childTasks[0] = task;
size_t numChildren = 1;
/*! split until node is full */
do {
/*! find best child to split */
float bestArea = inf;
ssize_t bestChild = -1;
for (size_t i=0; i<numChildren; i++)
{
float A = childTasks[i].sceneArea();
size_t items = childTasks[i].items();
if (items > 1 && A <= bestArea) {
bestChild = i;
bestArea = A;
}
}
if (bestChild == -1) break;
/*! split best child into left and right child */
__align(64) BuildRecord left, right;
split(childTasks[bestChild],left,right,mode,threadID,numThreads);
/* add new children left and right */
left.depth = right.depth = task.depth+1;
childTasks[bestChild] = childTasks[numChildren-1];
childTasks[numChildren-1] = left;
childTasks[numChildren+0] = right;
numChildren++;
} while (numChildren < 4);
/* recurse */
BVH4::Node* node = bvh->allocNode(threadID);
for (ssize_t i=numChildren-1; i>=0; i--) {
childTasks[i].parentNode = (size_t)&node->child(i);
recurse(depth+1,childTasks[i],mode,threadID,numThreads);
node->set(i,childTasks[i].bounds.geometry);
}
*(NodeRef*)task.parentNode = bvh->encodeNode(node);
}
/* prints the bvh4.triangle4v data structure */
void print_bvh4_triangle4v(BVH4::NodeRef node, size_t depth)
{
if (node.isNode())
{
BVH4::Node* n = node.node();
std::cout << "Node {" << std::endl;
for (size_t i=0; i<BVH4::N; i++)
{
for (size_t k=0; k<depth; k++) std::cout << " ";
std::cout << " bounds" << i << " = " << n->bounds(i) << std::endl;
}
for (size_t i=0; i<BVH4::N; i++)
{
if (n->child(i) == BVH4::emptyNode)
continue;
for (size_t k=0; k<depth; k++) std::cout << " ";
std::cout << " child" << i << " = ";
print_bvh4_triangle4v(n->child(i),depth+1);
}
for (size_t k=0; k<depth; k++) std::cout << " ";
std::cout << "}" << std::endl;
}
else
{
size_t num;
const Triangle4v* tri = (const Triangle4v*) node.leaf(num);
std::cout << "Leaf {" << std::endl;
for (size_t i=0; i<num; i++) {
for (size_t j=0; j<tri[i].size(); j++) {
for (size_t k=0; k<depth; k++) std::cout << " ";
std::cout << " Triangle { v0 = (" << tri[i].v0.x[j] << ", " << tri[i].v0.y[j] << ", " << tri[i].v0.z[j] << "), "
"v1 = (" << tri[i].v1.x[j] << ", " << tri[i].v1.y[j] << ", " << tri[i].v1.z[j] << "), "
"v2 = (" << tri[i].v2.x[j] << ", " << tri[i].v2.y[j] << ", " << tri[i].v2.z[j] << "), "
"geomID = " << tri[i].geomID(j) << ", primID = " << tri[i].primID(j) << " }" << std::endl;
}
}
for (size_t k=0; k<depth; k++) std::cout << " ";
std::cout << "}" << std::endl;
}
}
void BVH4BuilderTopLevel::createLeaf(BuildRecord& current, size_t threadIndex, size_t threadCount)
{
#if defined(DEBUG)
if (current.depth > BVH4::maxBuildDepthLeaf)
throw std::runtime_error("ERROR: depth limit reached");
#endif
/* return empty node */
if (current.end-current.begin == 0) {
*(NodeRef*)current.parentNode = BVH4::emptyNode;
return;
}
/* return leaf node */
if (current.end-current.begin == 1) {
*(NodeRef*)current.parentNode = refs[current.begin].node;
return;
}
/* first split level */
BuildRecord record0, record1;
split_fallback2(&refs[0],current,record0,record1);
/* second split level */
BuildRecord children[4];
split_fallback2(&refs[0],record0,children[0],children[1]);
split_fallback2(&refs[0],record1,children[2],children[3]);
/* allocate next four nodes */
BVH4::Node* node = bvh->allocNode(threadIndex);
*(NodeRef*)current.parentNode = bvh->encodeNode(node);
/* recurse into each child */
for (size_t i=0; i<4; i++)
{
children[i].parentNode = (size_t)&node->child(i);
children[i].depth = current.depth+1;
createLeaf(children[i],threadIndex,threadCount);
node->set(i,children[i].bounds.geometry);
}
BVH4::compact(node); // move empty nodes to the end
}
void BVH4BuilderTopLevel::open_sequential()
{
size_t N = max(2*refs.size(),size_t(MIN_OPEN_SIZE));
refs.reserve(N);
std::make_heap(refs.begin(),refs.end());
while (refs.size()+3 <= N)
{
std::pop_heap (refs.begin(),refs.end());
BVH4::NodeRef ref = refs.back().node;
if (ref.isLeaf()) break;
refs.pop_back();
BVH4::Node* node = ref.node();
for (size_t i=0; i<4; i++) {
if (node->child(i) == BVH4::emptyNode) continue;
refs.push_back(BuildRef(node->bounds(i),node->child(i)));
std::push_heap (refs.begin(),refs.end());
}
}
}
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 build(size_t, size_t)
{
/* progress monitor */
auto progress = [&] (size_t dn) { bvh->scene->progressMonitor(dn); };
auto virtualprogress = BuildProgressMonitorFromClosure(progress);
/* fast path for empty BVH */
const size_t numPrimitives = scene->getNumPrimitives<BezierCurves,1>();
if (numPrimitives == 0) {
prims.clear();
bvh->set(BVH4::emptyNode,empty,0);
return;
}
double t0 = bvh->preBuild(TOSTRING(isa) "::BVH4BuilderHairSAH");
//profile(1,5,numPrimitives,[&] (ProfileTimer& timer) {
/* create primref array */
bvh->alloc.init_estimate(numPrimitives*sizeof(Primitive));
prims.resize(numPrimitives);
const PrimInfo pinfo = createBezierRefArray<1>(scene,prims,virtualprogress);
/* build hierarchy */
BVH4::NodeRef root = bvh_obb_builder_binned_sah
(
[&] () { return bvh->alloc.threadLocal2(); },
[&] (const PrimInfo* children, const size_t numChildren,
HeuristicArrayBinningSAH<BezierPrim> alignedHeuristic,
FastAllocator::ThreadLocal2* alloc) -> BVH4::Node*
{
BVH4::Node* node = (BVH4::Node*) alloc->alloc0.malloc(sizeof(BVH4::Node),16); node->clear();
for (size_t i=0; i<numChildren; i++)
node->set(i,children[i].geomBounds);
return node;
},
[&] (const PrimInfo* children, const size_t numChildren,
UnalignedHeuristicArrayBinningSAH<BezierPrim> unalignedHeuristic,
FastAllocator::ThreadLocal2* alloc) -> BVH4::UnalignedNode*
{
BVH4::UnalignedNode* node = (BVH4::UnalignedNode*) alloc->alloc0.malloc(sizeof(BVH4::UnalignedNode),16); node->clear();
for (size_t i=0; i<numChildren; i++)
{
const LinearSpace3fa space = unalignedHeuristic.computeAlignedSpace(children[i]);
const PrimInfo sinfo = unalignedHeuristic.computePrimInfo(children[i],space);
node->set(i,OBBox3fa(space,sinfo.geomBounds));
}
return node;
},
[&] (size_t depth, const PrimInfo& pinfo, FastAllocator::ThreadLocal2* alloc) -> BVH4::NodeRef
{
size_t items = pinfo.size();
size_t start = pinfo.begin;
Primitive* accel = (Primitive*) alloc->alloc1.malloc(items*sizeof(Primitive));
BVH4::NodeRef node = bvh->encodeLeaf((char*)accel,items);
for (size_t i=0; i<items; i++) {
accel[i].fill(prims.data(),start,pinfo.end,bvh->scene,false);
}
return node;
},
progress,
prims.data(),pinfo,BVH4::N,BVH4::maxBuildDepthLeaf,1,1,BVH4::maxLeafBlocks);
bvh->set(root,pinfo.geomBounds,pinfo.size());
//});
/* clear temporary data for static geometry */
if (scene->isStatic()) prims.clear();
bvh->alloc.cleanup();
bvh->postBuild(t0);
}
void BVH4BuilderTopLevel::task_open_parallel(size_t threadIndex, size_t threadCount, size_t taskIndex, size_t taskCount, TaskScheduler::Event* event)
{
size_t N = global_dest;
size_t M = refs1.size();
const size_t start0 = (threadIndex+0)*N/threadCount;
const size_t end0 = (threadIndex+1)*N/threadCount;
const size_t start1 = (threadIndex+0)*M/threadCount;
const size_t end1 = (threadIndex+1)*M/threadCount;
assert(end1-start1 >= end0-start0);
BuildRef* prefs1 = &refs1[0];
/* copy from refs buffer to refs1 buffer */
for (size_t i=start0, j=start1; i<end0; i++, j++)
refs1[j] = refs[i];
/* create max heap in our set of items */
size_t start = start1;
size_t end = start1+end0-start0;
std::make_heap(&prefs1[start],&prefs1[end]);
float max_volume = 0.0f;
while (true)
{
barrier.wait(threadIndex,threadCount);
if (threadIndex == 0) global_max_volume = 0.0f;
barrier.wait(threadIndex,threadCount);
/* parallel calculation of maximal volume */
max_volume = 0.0f;
if (end+3 <= end1)
for (size_t i=start; i<end; i++)
max_volume = max(max_volume,prefs1[i].lower.w);
atomic_max_f32(&global_max_volume,max_volume);
barrier.wait(threadIndex,threadCount);
max_volume = global_max_volume;
barrier.wait(threadIndex,threadCount);
/* if maximal volume is 0, all threads are finished */
if (max_volume == 0.0f) break;
/* open all nodes that are considered large in this iteration */
while (end+3 <= end1)
{
if (end-start == 0) break;
std::pop_heap(&prefs1[start],&prefs1[end]);
BVH4::NodeRef ref = prefs1[end-1].node;
float vol = prefs1[end-1].lower.w;
if (ref.isLeaf() || vol < 0.5f*max_volume) {
std::push_heap(&prefs1[start],&prefs1[end]);
break;
}
end--;
BVH4::Node* node = ref.node();
for (size_t i=0; i<4; i++) {
if (node->child(i) == BVH4::emptyNode) continue;
prefs1[end++] = BuildRef(node->bounds(i),node->child(i));
std::push_heap(&prefs1[start],&prefs1[end]);
}
}
}
if (threadIndex == 0) global_dest = 0;
barrier.wait(threadIndex,threadCount);
/* copy again back to refs array */
size_t dest = atomic_add(&global_dest,end-start);
for (size_t i=start, j=dest; i<end; i++, j++)
refs[j] = refs1[i];
}