Mikktspace tangent calculator optimizations #112256

Eugene-Kuznetsov · 2023-09-11T22:11:19+02:00

Eugene-Kuznetsov commented

2023-09-11 22:11:19 +02:00

The logic in intern/mikktspace is sometimes a major performance bottleneck. On my dual Xeon E5-2699 v4, I get as low as 10 fps in "animation player" in viewport shading mode with only a single 20k face mesh without any fancy modifiers. The bottleneck occurs because, under some circumstances, the method getTangSpace() is executed once per frame for every mesh, even if the mesh stays exactly the same. As far as I can tell, the sufficient conditions are (1) the mesh should be rigged to an armature with keypoints; and (2) at least one material on the mesh should contain a 'normal map' node.

This PR is a rework of tangent calculation logic. Individual steps are the same (and numbers it produces match the old module with reasonable accuracy), but their implementations were tweaked to various degrees. I've benchmarked it on two systems and I observe a 20% to 40% reduction in getTangSpace() execution time.

The logic in intern/mikktspace is sometimes a major performance bottleneck. On my dual Xeon E5-2699 v4, I get as low as 10 fps in "animation player" in viewport shading mode with only a single 20k face mesh without any fancy modifiers. The bottleneck occurs because, under some circumstances, the method getTangSpace() is executed once per frame for every mesh, even if the mesh stays exactly the same. As far as I can tell, the sufficient conditions are (1) the mesh should be rigged to an armature with keypoints; and (2) at least one material on the mesh should contain a 'normal map' node. This PR is a rework of tangent calculation logic. Individual steps are the same (and numbers it produces match the old module with reasonable accuracy), but their implementations were tweaked to various degrees. I've benchmarked it on two systems and I observe a 20% to 40% reduction in getTangSpace() execution time.

❤️ 3 🚀 1

Eugene-Kuznetsov force-pushed ek_mikktspace_rework from d536e37748 to 1d3fb1b5a8

2023-09-11 23:09:30 +02:00

Compare

Eugene-Kuznetsov force-pushed ek_mikktspace_rework from 1d3fb1b5a8 to 09be0b062f

2023-09-11 23:17:56 +02:00

Compare

Eugene-Kuznetsov commented

2023-09-17 04:03:30 +02:00

Quick update. PR still in development. I made a standalone version that lets me test performance and accuracy using previously dumped meshes and without launching the whole graphical UI. Found a couple of bugs. Added parallelization of build4RuleGroups() (the last major piece that was not parallelized, and sometimes taking over a quarter of total time) for well-behaved meshes only (it seems almost impossible to parallelize it and to have it match the behavior of reference implementation for all meshes).

The biggest obstacle right now is that I can reliably match reference implementation for "good" meshes, but things get complicated for non-manifold meshes and meshes with messed-up UV maps. For example, there's a piece of logic in there that marks triangles as "neighbors" pairwise if they share edges. Problems start when some edge is shared by three or more triangles, because each triangle can only have one neighbor along that edge, and the choice of which triangle is paired with which has an effect on generated tangents down the road.

Out of 20 meshes I'm testing on, I see some mismatches in 7, fewer than 1% of normals affected in each case but they are still mismatches and I need to find out what's causing them.

Quick update. PR still in development. I made a standalone version that lets me test performance and accuracy using previously dumped meshes and without launching the whole graphical UI. Found a couple of bugs. Added parallelization of build4RuleGroups() (the last major piece that was not parallelized, and sometimes taking over a quarter of total time) for well-behaved meshes only (it seems almost impossible to parallelize it and to have it match the behavior of reference implementation for all meshes). The biggest obstacle right now is that I can reliably match reference implementation for "good" meshes, but things get complicated for non-manifold meshes and meshes with messed-up UV maps. For example, there's a piece of logic in there that marks triangles as "neighbors" pairwise if they share edges. Problems start when some edge is shared by three or more triangles, because each triangle can only have one neighbor along that edge, and the choice of which triangle is paired with which has an effect on generated tangents down the road. Out of 20 meshes I'm testing on, I see some mismatches in 7, fewer than 1% of normals affected in each case but they are still mismatches and I need to find out what's causing them.

Eugene-Kuznetsov force-pushed ek_mikktspace_rework from 09be0b062f to 8df23fd1d9

2023-10-05 07:39:57 +02:00

Compare

Eugene-Kuznetsov force-pushed ek_mikktspace_rework from 8df23fd1d9 to 075bd919fa

2023-10-05 17:56:33 +02:00

Compare

Eugene-Kuznetsov force-pushed ek_mikktspace_rework from 075bd919fa to 245fe2c34c

2023-11-21 02:06:09 +01:00

Compare

Eugene-Kuznetsov force-pushed ek_mikktspace_rework from bcc11845be to 09a0c1feed

2023-12-06 18:34:07 +01:00

Compare

Eugene-Kuznetsov changed title from ~~WIP: Mikktspace tangent calculator optimizations~~ to Mikktspace tangent calculator optimizations

2023-12-07 16:05:22 +01:00

Iliya Katushenock added this to the Modeling project 2023-12-07 16:07:15 +01:00

Iliya Katushenock added the

 @ -871,3 +871,3 @@
   # different results. This will lead to automated test failures. So disable
   # this until we support it.
   set(PLATFORM_CFLAGS "-pipe -fPIC -funsigned-char -fno-strict-aliasing -ffp-contract=off")
   set(PLATFORM_CFLAGS "-pipe -fPIC -funsigned-char -fno-strict-aliasing -ffp-contract=off -march=native")

 @ -0,0 +12,4 @@
   /* we have to be compiled with -mavx2 or with -march=native on an
    * avx2-capable chip (this has not been the case for Blender previously) */
   #if  __has_builtin(__builtin_ia32_permvarsf256)
     #pragma message ("Using AVX2")

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Mikktspace tangent calculator optimizations #112256

Checkout