Progressive refine performance regression with denoising data (Cycles) #75432

New Issue

Alaska · 2020-04-06T08:38:33+02:00

Alaska commented

2020-04-06 08:38:33 +02:00

System Information:
Operating system: Linux (POP!_OS 18.04)
Graphics card: GTX 1050ti

Blender Version:
Broken: Blender 2.83 (d0d20de183) 2020-04-05 22:52
Worked: Blender 2.82a

Short description of error:
Blender 2.83 shows a regression in the speed of progressive refine rendering in Cycles when denoisng data is enabled when compared to 2.82a. This happens with both CPU and GPU (Only tested CUDA) rendering.

Tests were conducted using a CPU, the Ryzen 9 3900X locked at 4.0Ghz with sufficient cooling. All tests were also conducted 2-3 times to ensure consistency:

Blender version	Progressive without denoising data	Progressive with denoising data	Tile rendering without denoising data	Tile rendering with denoising data
2.82a	30 seconds	32 seconds	30 seconds	33 seconds
2.83	26 seconds	194 seconds	25 seconds	29 seconds

I should note that I monitored CPU usage in all tests and nothing appeared too abnormal. All 24 threads were at 100% load most of the render, however with 2.83 + progressive refine + denoising data, CPU usuage would drop periodically (every 3 seconds or so - inline with the image preview of the progressive refine updating).

This graph shows CPU usage over a 60 second period including CPU usage during the 2.83 render tests.
Key:
1 = 2.83 progressive refine without denoising data
2 = 2.83 progressive refine with denoisng data

It should be noted that 2.82a does not render the denoising data when progressive refine is enabled. The bug I'm describing in this post is most likely related to the fix for that bug.

I should also note that the progress bar for progressive refine renders with denoising data in 2.83 reachs 100% at the half way mark of the render. So if the render has 1000 samples, the progress bar will reach 100% at 500 samples.

Exact steps for others to reproduce the error:
Take any scene and render it with progressive refine and denoising data enabled in 2.83 and compare it to 2.82a.

**System Information:** Operating system: Linux (POP!_OS 18.04) Graphics card: GTX 1050ti **Blender Version:** Broken: Blender 2.83 (d0d20de183f1) 2020-04-05 22:52 Worked: Blender 2.82a **Short description of error:** Blender 2.83 shows a regression in the speed of progressive refine rendering in Cycles when denoisng data is enabled when compared to 2.82a. This happens with both CPU and GPU (Only tested CUDA) rendering. Tests were conducted using a CPU, the Ryzen 9 3900X locked at 4.0Ghz with sufficient cooling. All tests were also conducted 2-3 times to ensure consistency: | **Blender version** |**Progressive without denoising data**|**Progressive with denoising data**|**Tile rendering without denoising data**|**Tile rendering with denoising data**| | -- | -- | -- | -- | -- | |2.82a|30 seconds|32 seconds|30 seconds|33 seconds| |2.83|26 seconds|**194 seconds**|25 seconds|29 seconds| I should note that I monitored CPU usage in all tests and nothing appeared too abnormal. All 24 threads were at 100% load most of the render, however with 2.83 + progressive refine + denoising data, CPU usuage would drop periodically (every 3 seconds or so - inline with the image preview of the progressive refine updating). ![CPU Usage.png](https://archive.blender.org/developer/F8452248/CPU_Usage.png) This graph shows CPU usage over a 60 second period including CPU usage during the 2.83 render tests. Key: 1 = 2.83 progressive refine without denoising data 2 = 2.83 progressive refine with denoisng data It should be noted that 2.82a does not render the denoising data when progressive refine is enabled. The bug I'm describing in this post is most likely related to the fix for that bug. I should also note that the progress bar for progressive refine renders with denoising data in 2.83 reachs 100% at the half way mark of the render. So if the render has 1000 samples, the progress bar will reach 100% at 500 samples. **Exact steps for others to reproduce the error:** Take any scene and render it with progressive refine and denoising data enabled in 2.83 and compare it to 2.82a.

Alaska commented

2020-04-06 08:38:33 +02:00

Added subscriber: @Alaska

Alaska changed title from ~~Progressive performance regression with denoising data~~ to Progressive refine performance regression with denoising data (Cycles)

2020-04-06 11:48:11 +02:00

Ankit Meel commented

2020-04-06 12:06:25 +02:00

Added subscriber: @ankitm

Evan Wilson commented

2020-04-13 00:14:23 +02:00

Added subscriber: @EAW

Evan Wilson commented

2020-04-13 00:14:23 +02:00

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

Evan Wilson commented

2020-04-13 00:14:23 +02:00

This is sitting on the top of my "Needs Triage" queue so I thought I would confirm it.

Blender version	Progressive without denoising data	Progressive with denoising data	Tile rendering without denoising data	Tile rendering with denoising data
2.83.13	30 seconds	8:39	27 seconds	44 seconds

I misread the task at first and rendered progressive refine while denoising. I stopped at 7 samples which took 7:13

I also noticed (am confirming) that the progressive refine time estimate is calculating only the time to 50% of the samples are done, which I think is already reported.

Intel i7 950 @3.07GHz

2.83 (sub 13), branch: master, commit date: 2020-04-11 19:01, hash: 0a747cd4e3

This is sitting on the top of my "Needs Triage" queue so I thought I would confirm it. | **Blender version** |**Progressive without denoising data**|**Progressive with denoising data**|**Tile rendering without denoising data**|**Tile rendering with denoising data**| | -- | -- | -- | -- | -- | |2.83.13|30 seconds|**8:39**|27 seconds|44 seconds| I misread the task at first and rendered progressive refine while denoising. I stopped at 7 samples which took 7:13 I also noticed (am confirming) that the progressive refine time estimate is calculating only the time to 50% of the samples are done, which I think is already reported. --- Intel i7 950 @3.07GHz 2.83 (sub 13), branch: master, commit date: 2020-04-11 19:01, hash: 0a747cd4e3

blender-admin commented

2020-04-28 20:27:03 +02:00

This issue was referenced by blender/cycles@935da86122

This issue was referenced by blender/cycles@935da8612254a1059dd6cb2ca7f307cbdda58aaf

blender-admin commented

2020-04-28 20:27:03 +02:00

This issue was referenced by 05274ca829

This issue was referenced by 05274ca82971332866d94bac07c843aa2ae97f80

Brecht Van Lommel commented

2020-04-28 22:41:57 +02:00

Changed status from 'Confirmed' to: 'Resolved'

Brecht Van Lommel closed this issue

2020-04-28 22:41:57 +02:00

Brecht Van Lommel self-assigned this 2020-04-28 22:41:57 +02:00

Alaska commented

2020-04-29 01:34:43 +02:00

@brecht Running tests with 2.83 05274ca829 resolves the issue, thank you, but causes other bugs to occur: #76205, #76206

@brecht Running tests with 2.83 05274ca82971 resolves the issue, thank you, but causes other bugs to occur: #76205, #76206

Thomas Dinges added this to the 2.83 LTS milestone 2023-02-08 16:38:06 +01:00

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Progressive refine performance regression with denoising data (Cycles) #75432