Details

Reviewers

Commits

rB6b03621c018a: DrawManager: Use Compute Shader to Update Hair.
rB8f9599d17e80: DrawManager: Use Compute Shader to Update Hair.

Summary

This patch will use compute shaders to create the VBO for hair.
The previous implementation uses transform feedback.

Timings before: between 0.000069s and 0.000362s.
Timings after: between 0.000032s and 0.000092s.

Speedup isn't noticeable by end-users. The patch is used to test
the new compute shader pipeline and integrate it with the draw
manager. Allowing EEVEE, Workbench and other draw engines to
use compute shaders with the introduction of DRW_shgroup_call_compute
and DRW_shgroup_vertex_buffer.

Future improvements are possible by generating the index buffer
of hair directly on the GPU.

NOTE: that compute shaders aren't supported by Apple and still use the transform feedback workaround.

Diff Detail

Repository

rB Blender

Branch

temp-gpu-compute-shader-hair

Build Status

Buildable 14439
Build 14439: arc lint + arc unit

Event Timeline

Jeroen Bakker (jbakker) requested review of this revision.Apr 23 2021, 2:12 PM

Jeroen Bakker (jbakker) created this revision.

Harbormaster completed remote builds in B14193: Diff 36445.Apr 23 2021, 2:12 PM

Jeroen Bakker (jbakker) added a parent revision: D10913: GPU: Compute Pipeline..Apr 23 2021, 2:13 PM

Merge branch 'temp-gpu-compute-shader-hair' of git.blender.org:blender into temp-gpu-compute-shader-hair
Added memory barrier.

Harbormaster completed remote builds in B14198: Diff 36450.Apr 23 2021, 3:46 PM

Retry with correct base.

Harbormaster completed remote builds in B14199: Diff 36451.Apr 23 2021, 3:50 PM

Reupload as arcanist did select too many commits.

Harbormaster completed remote builds in B14200: Diff 36452.Apr 23 2021, 3:51 PM

Jeroen Bakker (jbakker) retitled this revision from DrawManager: Use Compute Shader to Update Hair. to [WIP] DrawManager: Use Compute Shader to Update Hair..Apr 23 2021, 3:53 PM

Jeroen Bakker (jbakker) edited the summary of this revision. (Show Details)

Jeroen Bakker (jbakker) added a project: EEVEE & Viewport.

Jeroen Bakker (jbakker) edited the summary of this revision. (Show Details)

Jeroen Bakker (jbakker) edited the summary of this revision. (Show Details)Apr 23 2021, 4:05 PM

Kenzie (kenziemac130) awarded a token.Apr 24 2021, 2:13 AM

Kenzie (kenziemac130) added a subscriber: Kenzie (kenziemac130).

Jeroen Bakker (jbakker) edited the summary of this revision. (Show Details)Apr 28 2021, 2:35 PM

Merge branch 'temp-gpu-compute-shaders' into temp-gpu-compute-shader-hair
Merge branch 'temp-gpu-compute-shaders' into temp-gpu-compute-shader-hair
Use GPU calls for hair.
Merge branch 'temp-gpu-compute-shaders' into temp-gpu-compute-shader-hair
Use attach vertex buffer.
Remove unused code.

Harbormaster completed remote builds in B14277: Diff 36600.Apr 28 2021, 2:39 PM

Jeroen Bakker (jbakker) retitled this revision from [WIP] DrawManager: Use Compute Shader to Update Hair. to DrawManager: Use Compute Shader to Update Hair..Apr 28 2021, 2:46 PM

Jeroen Bakker (jbakker) edited the summary of this revision. (Show Details)

Jeroen Bakker (jbakker) added a reviewer: Clément Foucault (fclem).

Jeroen Bakker (jbakker) added projects: BF Blender (3.0), Performance.

Jeroen Bakker (jbakker) edited the summary of this revision. (Show Details)Apr 28 2021, 2:49 PM

Jeroen Bakker (jbakker) edited the summary of this revision. (Show Details)Apr 28 2021, 3:36 PM

Jeroen Bakker (jbakker) edited the summary of this revision. (Show Details)Apr 28 2021, 3:49 PM

Reusing the uniform storage in DRWShadingGroup when not an uniform. Similar to Transform feedback

I don't see it as a weak point. The uniforms inside DRWShadingGroup are actually resources (which should be renamed one day). Binding a SSBO enters this category.

increase the number of hairs (655 and it fails. Related to 65536?

This might be linked to one of the limit of the implementation. See https://www.reddit.com/r/opengl/comments/6thlhy/how_big_can_a_single_compute_shader_work_group_get/ .
Check:

GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS
GL_MAX_COMPUTE_WORK_GROUP_COUNT
GL_MAX_COMPUTE_WORK_GROUP_SIZE

I don't know if these have a minimum size guaranteed by the standard. So you might want to make this exposed from the GPU module.

source/blender/draw/intern/DRW_render.h
583	Do not expose location. Expose name.
source/blender/draw/intern/draw_manager.h
301–302	I would still call it `DRW_UNIFORM_VERTEX_BUFFER` to make it clear it is a value of this enum. Also you could eventually bind these to other shader stages if that's needed in the future. So I would remove the `for compute shaders` part. If you want to make it extra clear, you could name it `DRW_UNIFORM_VERTEX_BUFFER_AS_STORAGE`. While the normal storage buffer would be `DRW_UNIFORM_STORAGE`. All things considered, the `UNIFORM` should be renamed to `RESOURCE`, but that's for a later cleanup.
source/blender/draw/intern/draw_manager_data.c
454	Get location from name here.
source/blender/draw/intern/shaders/common_hair_refine_vert.glsl
4	I would advise against this for new code. This is something that creates hard to read code with `#ifdefs` everywhere and it's hard to see what resources are needed. I would prefer to see the refine functions go into a shared `lib.glsl` file and have a dedicated `_comp.glsl` file.

Merge branch 'temp-gpu-compute-shaders' into temp-gpu-compute-shader-hair
Draw: Bind vertex buffers on name.
Separate vertex and compute shaders.
Rename to DRW_UNIFORM_VERTEX_BUFFER_AS_STORAGE
Batch hair when using compute shaders.

Harbormaster completed remote builds in B14438: Diff 36870.May 7 2021, 4:19 PM

Updating using parent patch as base.

Harbormaster completed remote builds in B14439: Diff 36871.May 7 2021, 4:20 PM

Jeroen Bakker (jbakker) edited the summary of this revision. (Show Details)May 7 2021, 4:23 PM

Can this same compute pipe be used for GPU armature skinning down the line?

Xavier Cho (mysticfall) added a subscriber: Xavier Cho (mysticfall).May 9 2021, 3:00 AM

Not related to this patch but I think we should enforce the vertex format is at least using 32bit floats/ints as expected by the buffer backed storage (I think). Better enforce this in the GPU module directly.

source/blender/draw/intern/shaders/common_hair_refine_comp.glsl
7	Use `writeonly` qualifier.
9	Not sure where this capitalization came from. :) Also unless blender became a spatial 4D app, the positions are only 3D ;). From the vertex format, I think it is `posTime` and would prefer to reuse it to keep consistency. In new code I tend to prefix all the outputs with `out_*` so shader stage outputs are easily identifiable . Alternatively you can name the instance of the buffer `out_vertbuf` like so. layout(std430, binding = 0) buffer hairPointOutputBuffer { vec4 posTime[]; } out_vertbuf;