Use threading for simple gpu buffer mesh extractions. #88350

New Issue

Jeroen Bakker · 2021-05-17T14:46:58+02:00

Jeroen Bakker commented

2021-05-17 14:46:58 +02:00

When using smaller meshes threading adds overhead. For this reason simple extractions are run single threaded.
At a point a mesh can so large that it benefits using threading.

In file draw_cache_extract_mesh.c several batch extractions are defined. The extractions that don't use threading are:

extract_tris
extract_lines
extract_lines_with_lines_loose
extract_points
extract_fdots
extract_lines_paint_mask
extract_lines_adjacency
extract_edituv_tris
extract_edituv_lines
extract_edituv_points
extract_edituv_fdots
extract_uv
extract_tan/hq
extract_sculpt_data
extract_vcol
extract_edge_fac
extract_edituv_stretch_area
extract_edituv_stretch_angle
extract_mesh_analysis
extract_fdots_nor/hq
extract_skin_roots

This task should find out the reason why threading is disabled on them. Document the reason why threading is disabled.

Proof

Using D11255: Performance: GPU Batch Baseline TestCase. the baseline takes between 1111ms and 1180ms.
When changing extract_points to use threading between 1089ms and 1100ms.

Using similar tests when using small meshes (1000 faces) 622ms (not threaded vs 625ms using threading)

diff --git a/source/blender/draw/intern/draw_cache_extract_mesh.c b/source/blender/draw/intern/draw_cache_extract_mesh.c
index f167ea3d540..dcacbf104dc 100644
--- a/source/blender/draw/intern/draw_cache_extract_mesh.c
+++ b/source/blender/draw/intern/draw_cache_extract_mesh.c
@@ -1248,7 +1248,7 @@ static const MeshExtract extract_points = {
     .iter_lvert_mesh = extract_points_iter_lvert_mesh,
     .finish = extract_points_finish,
     .data_flag = 0,
-    .use_threading = false,
+    .use_threading = true,
 };
 
 /** \} */
@@ -1329,7 +1329,7 @@ static const MeshExtract extract_fdots = {
     .iter_poly_mesh = extract_fdots_iter_poly_mesh,
     .finish = extract_fdots_finish,
     .data_flag = 0,
-    .use_threading = false,
+    .use_threading = true,
 };
 
 /** \} */

Edit mode batches and their threading state:

vbo.pos_nor (threaded)
vbo.lnor (threaded)
vbo.edit_data (threaded)
vbo.fdots_pos (threaded)
vbo.fdots_nor/hq (not threaded): Whole batch creation is done during finish with the clarification that it is faster. Would like to see both implementations and compare both. I don't see any issues makeing this multithreaded.
ibo.tris (not threaded): Current implementation also sorts the faces on its material number. This might need more changes to be threaded. (thread per material, or special implementation when using one material. We could also consider a different tris when material sorting isn't needed.
ibo.points (not threaded): Don't see any reason why this one cannot be threaded.
ibo.fdots (not threaded): Don't see any reason why this one cannot be threaded. Didn't see that much of performance differences.

Conclusion

make points and fdots threaded. #88351
research threading for ibo.tris #88352
research threading for ibo.fdots_nor #88353

When using smaller meshes threading adds overhead. For this reason simple extractions are run single threaded. At a point a mesh can so large that it benefits using threading. In file `draw_cache_extract_mesh.c` several batch extractions are defined. The extractions that don't use threading are: * extract_tris * extract_lines * extract_lines_with_lines_loose * extract_points * extract_fdots * extract_lines_paint_mask * extract_lines_adjacency * extract_edituv_tris * extract_edituv_lines * extract_edituv_points * extract_edituv_fdots * extract_uv * extract_tan/hq * extract_sculpt_data * extract_vcol * extract_edge_fac * extract_edituv_stretch_area * extract_edituv_stretch_angle * extract_mesh_analysis * extract_fdots_nor/hq * extract_skin_roots This task should find out the reason why threading is disabled on them. Document the reason why threading is disabled. ## Proof Using [D11255: Performance: GPU Batch Baseline TestCase.](https://archive.blender.org/developer/D11255) the baseline takes between 1111ms and 1180ms. When changing extract_points to use threading between 1089ms and 1100ms. Using similar tests when using small meshes (1000 faces) 622ms (not threaded vs 625ms using threading) ``` diff --git a/source/blender/draw/intern/draw_cache_extract_mesh.c b/source/blender/draw/intern/draw_cache_extract_mesh.c index f167ea3d540..dcacbf104dc 100644 --- a/source/blender/draw/intern/draw_cache_extract_mesh.c +++ b/source/blender/draw/intern/draw_cache_extract_mesh.c @@ -1248,7 +1248,7 @@ static const MeshExtract extract_points = { .iter_lvert_mesh = extract_points_iter_lvert_mesh, .finish = extract_points_finish, .data_flag = 0, - .use_threading = false, + .use_threading = true, }; /** \} */ @@ -1329,7 +1329,7 @@ static const MeshExtract extract_fdots = { .iter_poly_mesh = extract_fdots_iter_poly_mesh, .finish = extract_fdots_finish, .data_flag = 0, - .use_threading = false, + .use_threading = true, }; /** \} */ ``` Edit mode batches and their threading state: * vbo.pos_nor (threaded) * vbo.lnor (threaded) * vbo.edit_data (threaded) * vbo.fdots_pos (threaded) * vbo.fdots_nor/hq (not threaded): Whole batch creation is done during `finish` with the clarification that it is faster. Would like to see both implementations and compare both. I don't see any issues makeing this multithreaded. * ibo.tris (not threaded): Current implementation also sorts the faces on its material number. This might need more changes to be threaded. (thread per material, or special implementation when using one material. We could also consider a different tris when material sorting isn't needed. * ibo.points (not threaded): Don't see any reason why this one cannot be threaded. * ibo.fdots (not threaded): Don't see any reason why this one cannot be threaded. Didn't see that much of performance differences. ## Conclusion * make points and fdots threaded. #88351 * research threading for ibo.tris #88352 * research threading for ibo.fdots_nor #88353

Jeroen Bakker self-assigned this 2021-05-17 14:46:58 +02:00

Jeroen Bakker commented

2021-05-17 14:46:58 +02:00

Added subscribers: @Jeroen-Bakker, @machieb, @warcanin, @fclem, @GeorgiaPacific, @kioku, @JorgeBernalMartinez, @elias.andersson92, @TheRedWaxPolice, @JacobMerrill-1, @ideasman42

Jeroen Bakker commented

2021-05-17 14:54:10 +02:00

Changed status from 'Needs Triage' to: 'Confirmed'

Jeroen Bakker removed their assignment 2021-05-17 15:02:49 +02:00

Philipp Oeser removed the

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Use threading for simple gpu buffer mesh extractions. #88350

Proof

Conclusion