typename BVHN<N>::NodeRef BVHN<N>::layoutLargeNodesRecursion(NodeRef& node, FastAllocator::ThreadLocal& allocator)
{
if (node.isBarrier()) {
node.clearBarrier();
return node;
}
else if (node.isAlignedNode())
{
AlignedNode* oldnode = node.alignedNode();
AlignedNode* newnode = (BVHN::AlignedNode*) allocator.malloc(sizeof(BVHN::AlignedNode),byteNodeAlignment);
*newnode = *oldnode;
for (size_t c=0; c<N; c++)
newnode->child(c) = layoutLargeNodesRecursion(oldnode->child(c),allocator);
return encodeNode(newnode);
}
else return node;
}
BBox3fa BVHNRefitter<N>::refit_toplevel(NodeRef& ref,
size_t &subtrees,
const BBox3fa *const subTreeBounds,
const size_t depth)
{
if (depth >= MAX_SUB_TREE_EXTRACTION_DEPTH)
{
assert(subtrees < MAX_NUM_SUB_TREES);
assert(subTrees[subtrees] == ref);
return subTreeBounds[subtrees++];
}
if (ref.isAlignedNode())
{
AlignedNode* node = ref.alignedNode();
BBox3fa bounds[N];
for (size_t i=0; i<N; i++)
{
NodeRef& child = node->child(i);
if (unlikely(child == BVH::emptyNode))
bounds[i] = BBox3fa(empty);
else
bounds[i] = refit_toplevel(child,subtrees,subTreeBounds,depth+1);
}
BBox3vf<N> boundsT = transpose<N>(bounds);
/* set new bounds */
node->lower_x = boundsT.lower.x;
node->lower_y = boundsT.lower.y;
node->lower_z = boundsT.lower.z;
node->upper_x = boundsT.upper.x;
node->upper_y = boundsT.upper.y;
node->upper_z = boundsT.upper.z;
return merge<N>(bounds);
}
else
return leafBounds.leafBounds(ref);
}
void BVHNRefitter<N>::gather_subtree_refs(NodeRef& ref,
size_t &subtrees,
const size_t depth)
{
if (depth >= MAX_SUB_TREE_EXTRACTION_DEPTH)
{
assert(subtrees < MAX_NUM_SUB_TREES);
subTrees[subtrees++] = ref;
return;
}
if (ref.isAlignedNode())
{
AlignedNode* node = ref.alignedNode();
for (size_t i=0; i<N; i++) {
NodeRef& child = node->child(i);
if (unlikely(child == BVH::emptyNode)) continue;
gather_subtree_refs(child,subtrees,depth+1);
}
}
}
void BVHN<N>::layoutLargeNodes(size_t num)
{
struct NodeArea
{
__forceinline NodeArea() {}
__forceinline NodeArea(NodeRef& node, const BBox3fa& bounds)
: node(&node), A(node.isLeaf() ? float(neg_inf) : area(bounds)) {}
__forceinline bool operator< (const NodeArea& other) const {
return this->A < other.A;
}
NodeRef* node;
float A;
};
std::vector<NodeArea> lst;
lst.reserve(num);
lst.push_back(NodeArea(root,empty));
while (lst.size() < num)
{
std::pop_heap(lst.begin(), lst.end());
NodeArea n = lst.back(); lst.pop_back();
if (!n.node->isAlignedNode()) break;
AlignedNode* node = n.node->alignedNode();
for (size_t i=0; i<N; i++) {
if (node->child(i) == BVHN::emptyNode) continue;
lst.push_back(NodeArea(node->child(i),node->bounds(i)));
std::push_heap(lst.begin(), lst.end());
}
}
for (size_t i=0; i<lst.size(); i++)
lst[i].node->setBarrier();
root = layoutLargeNodesRecursion(root,*alloc.threadLocal2()->alloc0);
}
BBox3fa BVHNRefitter<N>::recurse_bottom(NodeRef& ref)
{
/* this is a leaf node */
if (unlikely(ref.isLeaf()))
return leafBounds.leafBounds(ref);
/* recurse if this is an internal node */
AlignedNode* node = ref.alignedNode();
/* enable exclusive prefetch for >= AVX platforms */
#if defined(__AVX__)
ref.prefetchW();
#endif
BBox3fa bounds[N];
for (size_t i=0; i<N; i++)
if (unlikely(node->child(i) == BVH::emptyNode))
{
bounds[i] = BBox3fa(empty);
}
else
bounds[i] = recurse_bottom(node->child(i));
/* AOS to SOA transform */
BBox3vf<N> boundsT = transpose<N>(bounds);
/* set new bounds */
node->lower_x = boundsT.lower.x;
node->lower_y = boundsT.lower.y;
node->lower_z = boundsT.lower.z;
node->upper_x = boundsT.upper.x;
node->upper_y = boundsT.upper.y;
node->upper_z = boundsT.upper.z;
return merge<N>(bounds);
}
typename BVHNStatistics<N>::Statistics BVHNStatistics<N>::statistics(NodeRef node, const double A, const BBox1f t0t1)
{
Statistics s;
double dt = max(0.0f,t0t1.size());
if (node.isAlignedNode())
{
AlignedNode* n = node.alignedNode();
for (size_t i=0; i<N; i++) {
if (n->child(i) == BVH::emptyNode) continue;
s.statAlignedNodes.numChildren++;
const double Ai = max(0.0f,halfArea(n->extend(i)));
s = s + statistics(n->child(i),Ai,t0t1);
}
s.statAlignedNodes.numNodes++;
s.statAlignedNodes.nodeSAH += dt*A;
s.depth++;
}
else if (node.isUnalignedNode())
{
UnalignedNode* n = node.unalignedNode();
for (size_t i=0; i<N; i++) {
if (n->child(i) == BVH::emptyNode) continue;
s.statUnalignedNodes.numChildren++;
const double Ai = max(0.0f,halfArea(n->extend(i)));
s = s + statistics(n->child(i),Ai,t0t1);
}
s.statUnalignedNodes.numNodes++;
s.statUnalignedNodes.nodeSAH += dt*A;
s.depth++;
}
else if (node.isAlignedNodeMB())
{
AlignedNodeMB* n = node.alignedNodeMB();
for (size_t i=0; i<N; i++) {
if (n->child(i) == BVH::emptyNode) continue;
s.statAlignedNodesMB.numChildren++;
const double Ai = max(0.0f,halfArea(n->extend0(i)));
s = s + statistics(n->child(i),Ai,t0t1);
}
s.statAlignedNodesMB.numNodes++;
s.statAlignedNodesMB.nodeSAH += dt*A;
s.depth++;
}
else if (node.isUnalignedNodeMB())
{
UnalignedNodeMB* n = node.unalignedNodeMB();
for (size_t i=0; i<N; i++) {
if (n->child(i) == BVH::emptyNode) continue;
s.statUnalignedNodesMB.numChildren++;
const double Ai = max(0.0f,halfArea(n->extend0(i)));
s = s + statistics(n->child(i),Ai,t0t1);
}
s.statUnalignedNodesMB.numNodes++;
s.statUnalignedNodesMB.nodeSAH += dt*A;
s.depth++;
}
else if (node.isTransformNode())
{
s.statTransformNodes.numNodes++;
s.statTransformNodes.nodeSAH += dt*A;
s.depth++;
}
else if (node.isQuantizedNode())
{
QuantizedNode* n = node.quantizedNode();
for (size_t i=0; i<N; i++) {
if (n->child(i) == BVH::emptyNode) continue;
s.statQuantizedNodes.numChildren++;
const double Ai = max(0.0f,halfArea(n->extend(i)));
s = s + statistics(n->child(i),Ai,t0t1);
}
s.statQuantizedNodes.numNodes++;
s.statQuantizedNodes.nodeSAH += dt*A;
s.depth++;
}
else if (node.isLeaf())
{
size_t num; const char* tri = node.leaf(num);
if (num)
{
for (size_t i=0; i<num; i++) {
s.statLeaf.numPrims += bvh->primTy.size(tri+i*bvh->primTy.bytes);
}
s.statLeaf.numLeaves++;
s.statLeaf.numPrimBlocks += num;
s.statLeaf.leafSAH += dt*A*num;
if (num-1 < Statistics::LeafStat::NHIST) {
s.statLeaf.numPrimBlocksHistogram[num-1]++;
}
}
}
else {
throw std::runtime_error("not supported node type in bvh_statistics");
}
return s;
}
void BVHNStatistics<N>::statistics(NodeRef node, const float A, size_t& depth)
{
if (node.isNode())
{
numAlignedNodes++;
AlignedNode* n = node.node();
bvhSAH += A*travCostAligned;
depth = 0;
for (size_t i=0; i<N; i++) {
if (n->child(i) == BVH::emptyNode) continue;
childrenAlignedNodes++;
const float Ai = max(0.0f,halfArea(n->extend(i)));
size_t cdepth; statistics(n->child(i),Ai,cdepth);
depth=max(depth,cdepth);
}
depth++;
}
else if (node.isUnalignedNode())
{
numUnalignedNodes++;
UnalignedNode* n = node.unalignedNode();
bvhSAH += A*travCostUnaligned;
depth = 0;
for (size_t i=0; i<N; i++) {
if (n->child(i) == BVH::emptyNode) continue;
childrenUnalignedNodes++;
const float Ai = max(0.0f,halfArea(n->extend(i)));
size_t cdepth; statistics(n->child(i),Ai,cdepth);
depth=max(depth,cdepth);
}
depth++;
}
else if (node.isNodeMB())
{
numAlignedNodesMB++;
AlignedNodeMB* n = node.nodeMB();
bvhSAH += A*travCostAligned;
depth = 0;
for (size_t i=0; i<N; i++) {
if (n->child(i) == BVH::emptyNode) continue;
childrenAlignedNodesMB++;
const float Ai = max(0.0f,halfArea(n->extend0(i)));
size_t cdepth; statistics(n->child(i),Ai,cdepth);
depth=max(depth,cdepth);
}
depth++;
}
else if (node.isUnalignedNodeMB())
{
numUnalignedNodesMB++;
UnalignedNodeMB* n = node.unalignedNodeMB();
bvhSAH += A*travCostUnaligned;
depth = 0;
for (size_t i=0; i<N; i++) {
if (n->child(i) == BVH::emptyNode) continue;
childrenUnalignedNodesMB++;
const float Ai = max(0.0f,halfArea(n->extend0(i)));
size_t cdepth; statistics(n->child(i),Ai,cdepth);
depth=max(depth,cdepth);
}
depth++;
}
else if (node.isTransformNode())
{
numTransformNodes++;
TransformNode* n = node.transformNode();
bvhSAH += A*travCostTransform;
depth = 0;
const BBox3fa worldBounds = xfmBounds(n->local2world,n->localBounds);
const float Ai = max(0.0f,halfArea(worldBounds));
//size_t cdepth; statistics(n->child,Ai,cdepth);
//depth=max(depth,cdepth)+1;
}
else
{
depth = 0;
size_t num; const char* tri = node.leaf(num);
if (!num) return;
numLeaves++;
numPrimBlocks += num;
for (size_t i=0; i<num; i++)
numPrims += bvh->primTy.size(tri+i*bvh->primTy.bytes);
float sah = A * intCost * num;
leafSAH += sah;
}
}
void BVH4Statistics::statistics(NodeRef node, const float A, size_t& depth)
{
if (node.isNode())
{
hash += 0x1234;
numAlignedNodes++;
AlignedNode* n = node.node();
bvhSAH += A*BVH4::travCostAligned;
depth = 0;
for (size_t i=0; i<BVH4::N; i++) {
if (n->child(i) == BVH4::emptyNode) continue;
childrenAlignedNodes++;
const float Ai = max(0.0f,halfArea(n->extend(i)));
size_t cdepth; statistics(n->child(i),Ai,cdepth);
depth=max(depth,cdepth);
}
depth++;
hash += 0x76767*depth;
}
else if (node.isUnalignedNode())
{
hash += 0x1232344;
numUnalignedNodes++;
UnalignedNode* n = node.unalignedNode();
bvhSAH += A*BVH4::travCostUnaligned;
depth = 0;
for (size_t i=0; i<BVH4::N; i++) {
if (n->child(i) == BVH4::emptyNode) continue;
childrenUnalignedNodes++;
const float Ai = max(0.0f,halfArea(n->extend(i)));
size_t cdepth; statistics(n->child(i),Ai,cdepth);
depth=max(depth,cdepth);
}
depth++;
hash += 0x76767*depth;
}
else if (node.isNodeMB())
{
hash += 0xEF343;
numAlignedNodesMB++;
BVH4::NodeMB* n = node.nodeMB();
bvhSAH += A*BVH4::travCostAligned;
depth = 0;
for (size_t i=0; i<BVH4::N; i++) {
if (n->child(i) == BVH4::emptyNode) continue;
childrenAlignedNodesMB++;
const float Ai = max(0.0f,halfArea(n->extend0(i)));
size_t cdepth; statistics(n->child(i),Ai,cdepth);
depth=max(depth,cdepth);
}
depth++;
hash += 0x76767*depth;
}
else if (node.isUnalignedNodeMB())
{
hash += 0x1EEF4;
numUnalignedNodesMB++;
BVH4::UnalignedNodeMB* n = node.unalignedNodeMB();
bvhSAH += A*BVH4::travCostUnaligned;
depth = 0;
for (size_t i=0; i<BVH4::N; i++) {
if (n->child(i) == BVH4::emptyNode) continue;
childrenUnalignedNodesMB++;
const float Ai = max(0.0f,halfArea(n->extend0(i)));
size_t cdepth; statistics(n->child(i),Ai,cdepth);
depth=max(depth,cdepth);
}
depth++;
hash += 0x76767*depth;
}
else
{
depth = 0;
size_t num; const char* tri = node.leaf(num);
hash += 0xDD776*num+0x878;
if (!num) return;
hash += bvh->primTy.hash(tri,num);
numLeaves++;
numPrims += num;
float sah = A * BVH4::intCost * num;
bvhSAH += sah;
}
}