blender3.3 HIP performance down 30% on Windows #100891

New Issue

System Information
Operating system:windows10 21h2
Graphics card:rx6800xt

Blender Version：3.3 Long-term support version
Broken: Version ：3.2 has no problems
Worked: no
Short description of error3.2
3.2 version of the classroom scene only needs 32s, 3.3lts need 44s
Standard classroom, no need for me to provide it.

**System Information** Operating system:windows10 21h2 Graphics card:rx6800xt **Blender Version**：3.3 Long-term support version Broken: Version ：3.2 has no problems Worked: no **Short description of error**3.2 3.2 version of the classroom scene only needs 32s, 3.3lts need 44s Standard classroom, no need for me to provide it. ![3.3lts.png](https://archive.blender.org/developer/F13465926/3.3lts.png) ![3.2.jpeg](https://archive.blender.org/developer/F13465925/3.2.jpeg)

Added subscriber: @2905710881

#101151 was marked as duplicate of this issue

![3.3lts.png](https://archive.blender.org/developer/F13466009/3.3lts.png)

Added subscriber: @mano-wii

It must be a problem only with AMD GPUs as I didn't notice any difference using an RTX 3060.

Operating system: Windows-10-10.0.22000-SP0 64 Bits
Graphics card: NVIDIA GeForce RTX 3060 Laptop GPU/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 516.94

It must be a problem only with AMD GPUs as I didn't notice any difference using an RTX 3060. --- Operating system: Windows-10-10.0.22000-SP0 64 Bits Graphics card: NVIDIA GeForce RTX 3060 Laptop GPU/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 516.94

Added subscribers: @BrianSavery, @OmarEmaraDev

Changed status from 'Needs Triage' to: 'Needs Developer To Reproduce'

@BrianSavery Are you aware about any slow downs in 3.3? It seem Open Data queries seems to confirm some slow down, see:

@BrianSavery Are you aware about any slow downs in 3.3? It seem Open Data queries seems to confirm some slow down, see: - https://opendata.blender.org/benchmarks/query/?device_name=AMD%20Radeon%20RX%206800%20XT&compute_type=HIP&blender_version=3.3.0 - https://opendata.blender.org/benchmarks/query/?device_name=AMD%20Radeon%20RX%206800%20XT&compute_type=HIP&blender_version=3.2.0

Added subscriber: @Sayak-Biswas

@Sayak-Biswas is taking a look.

Added subscriber: @sayakbiswas

@2905710881 Are those two numbers on the same driver or different?

In #100891#1416621, @sayakbiswas wrote:
@2905710881 Are those two numbers on the same driver or different?

Similarly, I tested several drivers with the same result, including 7.1,8.1,8.2,5.2 22q2, each time using Display Driver Uninstaller to uninstall

> In #100891#1416621, @sayakbiswas wrote: > @2905710881 Are those two numbers on the same driver or different? Similarly, I tested several drivers with the same result, including 7.1,8.1,8.2,5.2 22q2, each time using Display Driver Uninstaller to uninstall

In #100891#1416621, @sayakbiswas wrote:
@2905710881 Are those two numbers on the same driver or different?

I purposely downloaded version 3.22 to compare the run, the current driver is 22Q2 (I tested multiple versions of the driver, each time using the duu tool to uninstall)

> In #100891#1416621, @sayakbiswas wrote: > @2905710881 Are those two numbers on the same driver or different? I purposely downloaded version 3.22 to compare the run, the current driver is 22Q2 (I tested multiple versions of the driver, each time using the duu tool to uninstall) ![T@ZY%KNTF)4QVC%~70~L_4N.png](https://archive.blender.org/developer/F13498556/T_ZY_KNTF_4QVC__70_L_4N.png)

Added subscriber: @Pionpouage

Added subscriber: @deadpin

Additional data from the merged report:

Benchmark using 3.2.1:
https://opendata.blender.org/benchmarks/4b575db1-e191-401a-898d-e4e33b3820a8/

Benchmark using 3.3.0:
https://opendata.blender.org/benchmarks/28b80b4a-fb76-4af8-8b8c-c4599b01c6e2/

Additional data from the merged report: Benchmark using 3.2.1: https://opendata.blender.org/benchmarks/4b575db1-e191-401a-898d-e4e33b3820a8/ Benchmark using 3.3.0: https://opendata.blender.org/benchmarks/28b80b4a-fb76-4af8-8b8c-c4599b01c6e2/

Okay it seems this commit is causing the performance differential: 023eb2ea7c
On my W6800, the classroom scene renders in about 59s on a build from this commit. However the previous commit (d567785658) the scene renders in ~44s. I'm not entirely sure yet which change is exactly causing the issue, maybe the new transform_inverse?

Okay it seems this commit is causing the performance differential: https://github.com/blender/blender/commit/023eb2ea7c16a00272f83d564145e28aeb9ed2b7 On my W6800, the classroom scene renders in about 59s on a build from this commit. However the previous commit (d567785658349504dc98c693c8c46c30e9a60c44) the scene renders in ~44s. I'm not entirely sure yet which change is exactly causing the issue, maybe the new transform_inverse?

In #100891#1419724, @Sayak-Biswas wrote:
Okay it seems this commit is causing the performance differential: 023eb2ea7c
On my W6800, the classroom scene renders in about 59s on a build from this commit. However the previous commit (d567785658) the scene renders in ~44s. I'm not entirely sure yet which change is exactly causing the issue, maybe the new transform_inverse?

https://devtalk.blender.org/t/cycles-amd-hip-device-feedback/21400/405
There is a tested linux without performance degradation problem

> In #100891#1419724, @Sayak-Biswas wrote: > Okay it seems this commit is causing the performance differential: https://github.com/blender/blender/commit/023eb2ea7c16a00272f83d564145e28aeb9ed2b7 > On my W6800, the classroom scene renders in about 59s on a build from this commit. However the previous commit (d567785658349504dc98c693c8c46c30e9a60c44) the scene renders in ~44s. I'm not entirely sure yet which change is exactly causing the issue, maybe the new transform_inverse? https://devtalk.blender.org/t/cycles-amd-hip-device-feedback/21400/405 There is a tested linux without performance degradation problem

In #100891#1419724, @Sayak-Biswas wrote:
Okay it seems this commit is causing the performance differential: 023eb2ea7c
On my W6800, the classroom scene renders in about 59s on a build from this commit. However the previous commit (d567785658) the scene renders in ~44s. I'm not entirely sure yet which change is exactly causing the issue, maybe the new transform_inverse?

I can also confirm this behavior with my RX 6800XT.

With version "3.3.0 Alpha, branch: master, commit date: 2022-07-25 22:17, hash: 4cf6524731", the scene takes 44s to render.
With version "3.3.0 Alpha, branch: master, commit date: 2022-07-24 19:33, hash: 6db059e3d7", the build just before the commit in question, the scene takes 33s to render.

> In #100891#1419724, @Sayak-Biswas wrote: > Okay it seems this commit is causing the performance differential: https://github.com/blender/blender/commit/023eb2ea7c16a00272f83d564145e28aeb9ed2b7 > On my W6800, the classroom scene renders in about 59s on a build from this commit. However the previous commit (d567785658349504dc98c693c8c46c30e9a60c44) the scene renders in ~44s. I'm not entirely sure yet which change is exactly causing the issue, maybe the new transform_inverse? I can also confirm this behavior with my RX 6800XT. With version "3.3.0 Alpha, branch: master, commit date: 2022-07-25 22:17, hash: `4cf6524731`", the scene takes 44s to render. With version "3.3.0 Alpha, branch: master, commit date: 2022-07-24 19:33, hash: `6db059e3d7`", the build just before the commit in question, the scene takes 33s to render.

This issue was referenced by d20be55c1a

This issue was referenced by d20be55c1a9fc9f503a507fe47ace6ef5d70573e

Added subscriber: @brecht

I committed a change that may help but I don't really expect it to.

@Sayak-Biswas are you looking into this? The way I suggest to approach it is to effectively bisect that commit, step by step undo changes by that commit until you hit the change(s) that caused the performance difference.

Like, revert changes in ray_triangle_intersect, revert changes to use cross(), revert division by det to be done once in advance, etc.

I committed a change that may help but I don't really expect it to. @Sayak-Biswas are you looking into this? The way I suggest to approach it is to effectively bisect that commit, step by step undo changes by that commit until you hit the change(s) that caused the performance difference. Like, revert changes in ray_triangle_intersect, revert changes to use cross(), revert division by det to be done once in advance, etc.

Changed status from 'Needs Developer To Reproduce' to: 'Confirmed'

@brecht thanks, will try the change. We have the compiler team looking into the issue as well.

In #100891#1427026, @Sayak-Biswas wrote:
@brecht thanks, will try the change. We have the compiler team looking into the issue as well.

Can you restore the performance when 3.4 is released?

> In #100891#1427026, @Sayak-Biswas wrote: > @brecht thanks, will try the change. We have the compiler team looking into the issue as well. Can you restore the performance when 3.4 is released?

The current status is that we expect to get a new HIP compiler version in the coming weeks that will fix this regression. This should then fix the issue in time for the 3.4 release, and also 3.3.2 or 3.3.3 LTS depending on the timing.

This issue was referenced by 3e41332798

This issue was referenced by 3e413327984e9d10613cfdb05e150353f0ed39a0

This issue was referenced by c94583cd64

This issue was referenced by c94583cd64b4effe489c6a0fb73c1c7fa69c9a17

Changed status from 'Confirmed' to: 'Resolved'

This comment was removed by @2905710881

*This comment was removed by @2905710881*

This comment was removed by @2905710881

*This comment was removed by @2905710881*

![1E.jpg](https://archive.blender.org/developer/F14049568/1E.jpg)

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

blender3.3 HIP performance down 30% on Windows #100891