Skip to content

Free, open-source 3D sculpting for Linux, macOS and Windows: dense meshes, remeshing, painting, built-in AI (image/text to 3D), Send to Unreal, game-ready LODs, Gaussian splats and OpenUSD.

Small dabs stay on one thread: the vertex query by the triangles under its tasks, set listing by chunks

islandmed2a7a47dbcfd5
diff --git a/src/core/BVH.cpp b/src/core/BVH.cpp
index ee507cc6..9ecfac34 100644
--- a/src/core/BVH.cpp
+++ b/src/core/BVH.cpp
@@ -259,11 +259,12 @@ void BVH::queryVertices(const Mesh& mesh, const glm::vec3& center, float radius,
// The units of parallel work: the subtrees the sphere touches, split breadth first until there are enough
// of them (the same ones for any thread count).
- constexpr size_t kTasks = 128;
+ constexpr size_t kTasks = 128, kSerialTriangles = 16384;
std::vector<uint32_t>& tasks = queryTasks_;
tasks.clear();
if (sphereBox(nodes_[0].box, center, r2)) tasks.push_back(0);
- for (bool split = true; split && tasks.size() < kTasks;) {
+ int depth = 0; // of every inner node in `tasks` (median splits: each holds about nt >> depth triangles)
+ for (bool split = true; split && tasks.size() < kTasks; ++depth) {
split = false;
queryNext_.clear();
for (uint32_t ni : tasks) {
@@ -323,6 +324,13 @@ void BVH::queryVertices(const Mesh& mesh, const glm::vec3& center, float radius,
};
// The set doesn't depend on which thread found what, take() lists it in index order, and the leaves are
// joined in task order.
+ size_t work = 0; // triangles under the tasks, roughly
+ for (uint32_t ni : tasks) work += nodes_[ni].count > 0 ? nodes_[ni].count : mesh.triangleCount() >> depth;
+ if (work < kSerialTriangles) { // waking threads would cost more
+ for (uint32_t root : tasks) walk(root, stack_, queryLeaves_, vertBits_);
+ vertBits_.take(out);
+ return;
+ }
if (queryParts_.size() < tasks.size()) queryParts_.resize(tasks.size());
vertBits_.insertParallel(tasks.size(), 1, [&](size_t b, size_t e, IndexBits& found) {
std::vector<uint32_t> nodeStack;
diff --git a/src/core/IndexBits.cpp b/src/core/IndexBits.cpp
index ae1035a3..25bc278a 100644
--- a/src/core/IndexBits.cpp
+++ b/src/core/IndexBits.cpp
@@ -78,7 +78,7 @@ void IndexBits::absorb(std::span<IndexBits> parts) {
}
}
}
- }, 8);
+ }, 32);
for (IndexBits& p : parts) std::fill(p.summary_.begin(), p.summary_.end(), 0);
}
@@ -87,7 +87,7 @@ void IndexBits::take(std::vector<uint32_t>& out, const std::function<void(size_t
const size_t nc = chunks_.size();
if (nc == 0) return;
// Count each chunk's members, then write them at their offsets (blocks of chunks, so small sets run inline).
- constexpr size_t kBlock = 8;
+ constexpr size_t kBlock = 32; // chunks per job: small sets run inline
counts_.resize(nc + 1);
poolFor(nc, [&](size_t b, size_t e) {
for (size_t k = b; k < e; ++k) {