Cycles-X: Spring demo scene shows large performance regression #91428

New Issue

Alaska · 2021-09-15T09:12:32+02:00

Alaska commented

2021-09-15 09:12:32 +02:00

System Information
Operating system: Linux-5.10.0-8-amd64-x86_64-with-glibc2.31 64 Bits
Graphics card: NVIDIA GeForce RTX 3090/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 470.63.01

Blender Version
Broken: version: 3.0.0 Alpha, branch: cycles-x, commit date: 2021-09-14 15:01, hash: 58ab616d4c
Worked: Prior to Cycles-X

Short description of error
The Spring demo scene shows large performance regressions on GPUs compared to normal Cycles. For example, normal Cycles may only take about a minute on my GPU, but Cycles-X takes 30 minutes. This was reported by Raimund Klink in blender chat:


Master with OptiX	Cycles-X with OptiX

Exact steps for others to reproduce the error

Open the Spring demo file and render it on a GPU, comparing performance to normal Cycles. https://www.blender.org/download/demo-files/

Upon investigation I have found what seems to be the cause. The hair on Spring seems to be the issue. It is made with a shader like this:

And there are a lot of intersecting hairs and a lot of "transparent bounces". Looking at just this culprit object, you can simplify the scene with something like this:
Transparent hair performance regression Cycles-X.blend

Here are some steps to recreate the file above:

Create an object with a lot of hair (a cube with 400,000 strands of hair works)
Get the hairs to overlap with each other (Using the "Brownian" value in the "Physics -> Forces" tab of the particle settings (with "Advanced" enabled at the top of the particle setting section) worked for me).
Give the cube a material with the culprit material (mix shader with a "Transparent BSDF")
Change the "Light Paths" count for "Transparent" to something high (64) for the biggest effect
Position the camera and objects in such a way that the object with the hair only takes up a small portion of the frame
Render

**System Information** Operating system: Linux-5.10.0-8-amd64-x86_64-with-glibc2.31 64 Bits Graphics card: NVIDIA GeForce RTX 3090/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 470.63.01 **Blender Version** Broken: version: 3.0.0 Alpha, branch: cycles-x, commit date: 2021-09-14 15:01, hash: `58ab616d4c` Worked: Prior to Cycles-X **Short description of error** The Spring demo scene shows large performance regressions on GPUs compared to normal Cycles. For example, normal Cycles may only take about a minute on my GPU, but Cycles-X takes 30 minutes. This was reported by Raimund Klink in blender chat: |![Master.png](https://archive.blender.org/developer/F10418747/Master.png)|![Cycles-X.png](https://archive.blender.org/developer/F10418750/Cycles-X.png)| | -- | -- | |Master with OptiX|Cycles-X with OptiX| **Exact steps for others to reproduce the error** 1. Open the Spring demo file and render it on a GPU, comparing performance to normal Cycles. https://www.blender.org/download/demo-files/ Upon investigation I have found what seems to be the cause. The hair on Spring seems to be the issue. It is made with a shader like this: ![image.png](https://archive.blender.org/developer/F10418757/image.png) And there are a lot of intersecting hairs and a lot of "transparent bounces". Looking at just this culprit object, you can simplify the scene with something like this: [Transparent hair performance regression Cycles-X.blend](https://archive.blender.org/developer/F10418762/Transparent_hair_performance_regression_Cycles-X.blend) Here are some steps to recreate the file above: 1. Create an object with a lot of hair (a cube with 400,000 strands of hair works) 2. Get the hairs to overlap with each other (Using the "Brownian" value in the "Physics -> Forces" tab of the particle settings (with "Advanced" enabled at the top of the particle setting section) worked for me). 3. Give the cube a material with the culprit material (mix shader with a "Transparent BSDF") 4. Change the "Light Paths" count for "Transparent" to something high (64) for the biggest effect 5. Position the camera and objects in such a way that the object with the hair only takes up a small portion of the frame 6. Render

Alaska commented

2021-09-15 09:12:32 +02:00

Added subscriber: @Alaska

Alaska commented

2021-09-15 09:13:28 +02:00

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

Alaska closed this issue

2021-09-15 09:13:28 +02:00

Alaska commented

2021-09-15 09:13:28 +02:00

Bug report for Cycles-X. Reported here for ease of access but archived because we do not accept bug reports for Cycles-X yet.

Raimund Klink commented

2021-09-15 09:24:51 +02:00

Added subscriber: @Raimund58

Brecht Van Lommel commented

2021-09-17 19:17:16 +02:00

Added subscriber: @brecht

Brecht Van Lommel commented

2021-09-17 19:17:16 +02:00

Changed status from 'Archived' to: 'Confirmed'

Brecht Van Lommel reopened this issue

2021-09-17 19:17:16 +02:00

Brecht Van Lommel commented

2021-09-17 19:17:16 +02:00

I'll open this bug since we won't fully solve it before the merge.

Part of the problem is that we are repeating some ray-curve intersection, to reduce memory usage. D12524 helps with this but does not fully solve it. Increasing the number of stored intersections helps a bit but not all that much. For the hair case, a good solution may be to add an option to bake transparency to curve vertices, and enable it by default. This would allow applying it to the throughput in the hit function, and would be a good optimization for transparent hair rendering in general.

But there is another problem, which is that full frame progressive rendering is in some way fundamentally bad at this. A subset of the pixels takes much longer to render than others, which leaves GPU occupancy low when rendering 1 sample at a time. What makes this worse is that we keep rendering 1 sample at a time because the render scheduler is not aware of this low occupancy. That could be addressed somehow. However to get the best performance we still need to render practically all samples at once, as rendering 320 samples is significantly faster than 32 samples in this scene still. So you'd have no interactive updates at all, unless you enable tiling.

There may be some other way to improve GPU occupancy, Allowing main paths to spawn multiple shadow paths if there is room available could help.

I'll open this bug since we won't fully solve it before the merge. Part of the problem is that we are repeating some ray-curve intersection, to reduce memory usage. [D12524](https://archive.blender.org/developer/D12524) helps with this but does not fully solve it. Increasing the number of stored intersections helps a bit but not all that much. For the hair case, a good solution may be to add an option to bake transparency to curve vertices, and enable it by default. This would allow applying it to the throughput in the hit function, and would be a good optimization for transparent hair rendering in general. But there is another problem, which is that full frame progressive rendering is in some way fundamentally bad at this. A subset of the pixels takes much longer to render than others, which leaves GPU occupancy low when rendering 1 sample at a time. What makes this worse is that we keep rendering 1 sample at a time because the render scheduler is not aware of this low occupancy. That could be addressed somehow. However to get the best performance we still need to render practically all samples at once, as rendering 320 samples is significantly faster than 32 samples in this scene still. So you'd have no interactive updates at all, unless you enable tiling. There may be some other way to improve GPU occupancy, Allowing main paths to spawn multiple shadow paths if there is room available could help.

blender-admin commented

2021-09-20 15:49:10 +02:00

This issue was referenced by a9ad15a88f129a0593432e1217610ac840687497