Cycles: adaptive sampling improvements #74581

New Issue

Brecht Van Lommel · 2020-03-09T18:23:01+01:00

Brecht Van Lommel commented

2020-03-09 18:23:01 +01:00

Optimize PMJ pattern generation, and generate more samples
Solve issues with combined CPU and GPU not using the same amount of samples. Possibly by using a higher adaptive step for both CPU and GPU.
Render multiple small times at the same time for GPU, to allow adaptive sampling to stop and match CPU without too much performance impact.
Denoising can detect patterns due to adaptive sampling only working per tile. (#79190)
Detect errors in larger windows or using some hierarchical methods, as the simple 3x3 window fails for noisy scenes
Analyze different error metrics, and consider adding some user control over this

- [x] Optimize PMJ pattern generation, and generate more samples - [x] Solve issues with combined CPU and GPU not using the same amount of samples. Possibly by using a higher adaptive step for both CPU and GPU. - [x] Render multiple small times at the same time for GPU, to allow adaptive sampling to stop and match CPU without too much performance impact. - [x] Denoising can detect patterns due to adaptive sampling only working per tile. (#79190) - [ ] Detect errors in larger windows or using some hierarchical methods, as the simple 3x3 window fails for noisy scenes - [ ] Analyze different error metrics, and consider adding some user control over this

Brecht Van Lommel commented

2020-03-09 18:23:01 +01:00

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

Brecht Van Lommel commented

2020-03-09 18:23:01 +01:00

Added subscriber: @brecht

blender-admin commented

2020-03-09 18:23:01 +01:00

#99135 was marked as duplicate of this issue

Brecht Van Lommel commented

2020-03-09 18:24:54 +01:00

Added subscriber: @Stefan_Werner

Steffen Dünner commented

2020-03-09 19:04:11 +01:00

Added subscriber: @SteffenD

Lord Odin commented

2020-03-09 19:13:11 +01:00

Added subscriber: @derekbarker

Lord Odin commented

2020-03-09 19:13:11 +01:00

Will this allow us to render GPU at smaller tiles even when not using adaptive sampling? Because that would be a dream come true!

Russell commented

2020-05-02 21:45:16 +02:00

Added subscriber: @Russ1642

Russell commented

2020-05-02 21:45:16 +02:00

Without digging into the code it seems that adaptive sampling counts up to your render samples setting but won't change pixels if they're within your noise threshhold. It doesn't stop rendering when each pixel in the tile reaches the noise threshhold, it must keep running noise checks or something on every pixel for every sample.
Try this: With the default cube set the noise threshold to 0.1, minimum samples to 200, and the render samples to 10,000,000. It'll render the noisy tile and then spend minutes not changing the image at all but counting up to the 10,000,000 samples for no reason. Even for tiles that are entirely background. It seems this algorithm doesn't actually stop when the noise threshhold is reached.

Edit: This is for GPU rendering. CPU rendering doesn't do this.

Without digging into the code it seems that adaptive sampling counts up to your render samples setting but won't change pixels if they're within your noise threshhold. It doesn't stop rendering when each pixel in the tile reaches the noise threshhold, it must keep running noise checks or something on every pixel for every sample. Try this: With the default cube set the noise threshold to 0.1, minimum samples to 200, and the render samples to 10,000,000. It'll render the noisy tile and then spend minutes not changing the image at all but counting up to the 10,000,000 samples for no reason. Even for tiles that are entirely background. It seems this algorithm doesn't actually stop when the noise threshhold is reached. Edit: This is for GPU rendering. CPU rendering doesn't do this.

Stefan Werner commented

2020-05-11 14:35:58 +02:00

Cooperative groups in CUDA 9+ should allow GPUs with SM 6.0 and newer to do adaptive sampling at the same granularity as the CPU.

Alaska commented

2020-05-18 12:36:02 +02:00

Added subscriber: @Alaska

Lsscpp commented

2020-05-28 02:25:40 +02:00

Added subscriber: @lsscpp

Christian Beckers commented

2020-05-31 21:08:01 +02:00

Added subscriber: @BeckersC

msalv commented

2020-08-11 11:23:32 +02:00

Added subscriber: @moisessalvador

msalv commented

2020-08-11 11:23:32 +02:00

In #74581#922698, @Russ1642 wrote:
Without digging into the code it seems that adaptive sampling counts up to your render samples setting but won't change pixels if they're within your noise threshhold. It doesn't stop rendering when each pixel in the tile reaches the noise threshhold, it must keep running noise checks or something on every pixel for every sample.
Try this: With the default cube set the noise threshold to 0.1, minimum samples to 200, and the render samples to 10,000,000. It'll render the noisy tile and then spend minutes not changing the image at all but counting up to the 10,000,000 samples for no reason. Even for tiles that are entirely background. It seems this algorithm doesn't actually stop when the noise threshhold is reached.

Edit: This is for GPU rendering. CPU rendering doesn't do this.

Yes, it's weird. I've been wondering the same. In other software, adaptative sampling is supposed to give you an uniform noise all around, by stopping sampling on areas that otherwise would come out clean. But blender's one does almost like the opposite? It's hard to understand how it works. Yes, it saves some rendertime with the default settings but I couldn't tell you why

> In #74581#922698, @Russ1642 wrote: > Without digging into the code it seems that adaptive sampling counts up to your render samples setting but won't change pixels if they're within your noise threshhold. It doesn't stop rendering when each pixel in the tile reaches the noise threshhold, it must keep running noise checks or something on every pixel for every sample. > Try this: With the default cube set the noise threshold to 0.1, minimum samples to 200, and the render samples to 10,000,000. It'll render the noisy tile and then spend minutes not changing the image at all but counting up to the 10,000,000 samples for no reason. Even for tiles that are entirely background. It seems this algorithm doesn't actually stop when the noise threshhold is reached. > > Edit: This is for GPU rendering. CPU rendering doesn't do this. Yes, it's weird. I've been wondering the same. In other software, adaptative sampling is supposed to give you an uniform noise all around, by stopping sampling on areas that otherwise would come out clean. But blender's one does almost like the opposite? It's hard to understand how it works. Yes, it saves some rendertime with the default settings but I couldn't tell you why

Riccardo Giovanetti commented

2020-10-26 16:52:42 +01:00

Added subscriber: @Harvester