griefith/test2
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Mesh: Tune the parallelism of normals_calc_corners

Browse Source 
Tune the grain size used for the parallel_for to alleviate excessive
mutex contention inside `handle_fan_result_and_custom_normals`.

I happened to notice that the 4004 Moore Lane USD scene[1] experienced a
load time regression compared to the prior release. It looks due to the
grain size used and here are some 3-run averages for the import:
```
Grain      | Time in seconds
256 (main) | (14.6+14.6+14.8)/3 = 14.6667
1024       | (13+12.8+12.9)/3 = 12.9
4096       | (13.3+13.1+13.1)/3 = 13.1667
16384      | (12.2+12+ 12.5)/3 = 12.2333
65536      | (9.4+9.2+9.6)/3 = 9.4
131072     | (7.9+7.7+8)/3 = 7.8667
262144     | (7.3+7.1+7.2)/3 = 7.2
max(16384, #verts/2) (PR) | (7.1+6.9+6.8)/3 = 6.9333
```

This PR gets the scenario loading in just under 7 seconds now compared
to over 14 originally.
[1] https://dpel.aswf.io/4004-moore-lane/

Pull Request: https://projects.blender.org/blender/blender/pulls/141249
This commit is contained in:
Jesse Yurkovich
2025-07-03 20:43:43 +02:00
committed by Jesse Yurkovich
parent d92523d0b4
commit c8a4026984
1 changed files with 8 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
source/blender/blenkernel/intern/mesh_normals.cc
View File

@@ -1243,7 +1243,14 @@ void normals_calc_corners(const Span<float3> vert_positions,
r_fan_spaces->corners_by_space.reserve(corner_verts.size());
}
}
threading::parallel_for(vert_positions.index_range(), 256, [&](const IndexRange range) {
int64_t grain_size = 256;
/* Decrease parallelism in case where lock is used to avoid contention. */
if (!custom_normals.is_empty() || r_fan_spaces) {
grain_size = std::max(int64_t(16384), vert_positions.size() / 2);
}
threading::parallel_for(vert_positions.index_range(), grain_size, [&](const IndexRange range) {
Vector<VertCornerInfo, 16> corner_infos;
LocalEdgeVectorSet local_edge_by_vert;
Vector<VertEdgeInfo, 16> edge_infos;