Renderings vom 2.9 LTS produce different results than 3.0 #93710

New Issue

Daniel Thiele · 2021-12-05T17:10:52+01:00

Daniel Thiele commented

2021-12-05 17:10:52 +01:00

System Information
Operating system: Windows-10-10.0.19041-SP0 64 Bits
Graphics card: NVIDIA GeForce RTX 3070/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 472.12

Blender Version
Broken: version: 3.0
Worked: 2.93.5, branch: master, commit date: 2021-10-05 12:04, hash: a791bdabd0

Short description of error
Render looks broken in 3.0 with denoising

Exact steps for others to reproduce the error
Download https://drive.google.com/file/d/10VwtBGw_wD3nBIFMGLUW8qooYDEs8sta/view?usp=sharing

Render it with the settings in that exact same resolution with denoising enablebed. The denoised image hast some strange light shining on the humans in the very front row. the noisy image does not have that. It works as expected in 2.9 LTS

This image shows the artifact. It's the "yellow" arms in the bottom of the image that only shows up after denoising.

Here is a simplified file and set of steps provided by @Alaska

Create a new .blend file, add an object to it that fills the entire view of the camera, and change the render engine to Cycles.

Change the render resolution to 9464 x 4320 (I couldn't reproduce the issue with other render resoltuions)

Ensure Adaptive Sampling is on. You can reduce the Max Samples parameter to something like 64 to speed up renders.

Enabling Denoising for final renders and change the denoiser to OptiX with the passes set to Albedo and Normal

Ensure Use Tiling is enabled and the tile size is less than the render resolution.

Go to the Render Properties tab and enable ONLY these render passes.

Combined

Z

Diffuse Color

Emission

Ambient Occlusion

Render the image and notice the artifact.

Here is a file with steps 1-6 done for you.
#93710 Simplified.blend

The artifact is the horizontal line near the top of this image:

**System Information** Operating system: Windows-10-10.0.19041-SP0 64 Bits Graphics card: NVIDIA GeForce RTX 3070/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 472.12 **Blender Version** Broken: version: 3.0 Worked: 2.93.5, branch: master, commit date: 2021-10-05 12:04, hash: `a791bdabd0` **Short description of error** Render looks broken in 3.0 with denoising **Exact steps for others to reproduce the error** Download https://drive.google.com/file/d/10VwtBGw_wD3nBIFMGLUW8qooYDEs8sta/view?usp=sharing Render it with the settings in that exact same resolution with denoising enablebed. The denoised image hast some strange light shining on the humans in the very front row. the noisy image does not have that. It works as expected in 2.9 LTS This image shows the artifact. It's the "yellow" arms in the bottom of the image that only shows up after denoising. ![#93710 Denoising Artifacts.jpg](https://archive.blender.org/developer/F12693890/T93710_Denoising_Artifacts.jpg) --- Here is a simplified file and set of steps provided by @Alaska > 1. Create a new .blend file, add an object to it that fills the entire view of the camera, and change the render engine to Cycles. > 2. Change the render resolution to `9464 x 4320` (I couldn't reproduce the issue with other render resoltuions) > 3. Ensure `Adaptive Sampling` is on. You can reduce the `Max Samples` parameter to something like 64 to speed up renders. > 4. Enabling Denoising for final renders and change the denoiser to `OptiX` with the passes set to `Albedo and Normal` > 5. Ensure `Use Tiling` is enabled and the tile size is less than the render resolution. > 6. Go to the `Render Properties` tab and enable ONLY these render passes. > - Combined > - Z > - Diffuse Color > - Emission > - Ambient Occlusion > 7. Render the image and notice the artifact. > > Here is a file with steps 1-6 done for you. > [#93710 Simplified.blend](https://archive.blender.org/developer/F12694668/T93710_Simplified.blend) > > The artifact is the horizontal line near the top of this image: > ![#93710 - Simplified Artifact.jpg](https://archive.blender.org/developer/F12694636/T93710_-_Simplified_Artifact.jpg)

Daniel Thiele commented

2021-12-05 17:10:52 +01:00

Added subscriber: @DMan

blender-admin commented

2021-12-05 17:10:52 +01:00

#95814 was marked as duplicate of this issue

Alaska commented

2021-12-05 23:01:56 +01:00

Added subscribers: @pmoursnv, @Alaska

Alaska commented

2021-12-05 23:01:56 +01:00

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

Alaska commented

2021-12-05 23:01:56 +01:00

I am able to reproduce this issue on my setup. However I can't figure out what's wrong because if I enable Denoising Data so I can look at it and investigate what's wrong, the issue goes away.

Here is an image showing the artifacts (Notice that the people in the bottom portion of the image have yellow "glowing" arms):

CC @pmoursnv as it appears to be limited to the OptiX denoiser.

System Information
Operating system: Linux-5.15.0-2-amd64-x86_64-with-glibc2.32 64 Bits
Graphics card: NVIDIA GeForce RTX 3090/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 495.44
Blender version: 3.1.0 Alpha, branch: master, commit date: 2021-12-04 01:26, hash: 5ef5a9fc24

I am able to reproduce this issue on my setup. However I can't figure out what's wrong because if I enable `Denoising Data` so I can look at it and investigate what's wrong, the issue goes away. Here is an image showing the artifacts (Notice that the people in the bottom portion of the image have yellow "glowing" arms): ![#93710 Denoising Artifacts.jpg](https://archive.blender.org/developer/F12693890/T93710_Denoising_Artifacts.jpg) CC @pmoursnv as it appears to be limited to the OptiX denoiser. **System Information** Operating system: Linux-5.15.0-2-amd64-x86_64-with-glibc2.32 64 Bits Graphics card: NVIDIA GeForce RTX 3090/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 495.44 Blender version: 3.1.0 Alpha, branch: master, commit date: 2021-12-04 01:26, hash: `5ef5a9fc24`

Alaska commented

2021-12-05 23:42:02 +01:00

Added subscriber: @Sergey

Alaska commented

2021-12-05 23:42:02 +01:00

I will also CC @Sergey because upon further testing the issue mostly disappears if Use tiling is disabled or the tile size is increased to cover the entire image.

Note: I also tested with Blender 3.0 and the same thing I've noticed also occur there.

I will also CC @Sergey because upon further testing the issue mostly disappears if `Use tiling` is disabled or the tile size is increased to cover the entire image. Note: I also tested with Blender 3.0 and the same thing I've noticed also occur there.

Alaska commented

2021-12-06 01:21:27 +01:00

I have simplified the file and got some steps to reproduce it from scratch.

Create a new .blend file, add an object to it that fills the entire view of the camera, and change the render engine to Cycles.
Change the render resolution to 9464 x 4320 (I couldn't reproduce the issue with other render resoltuions)
Ensure Adaptive Sampling is on. You can reduce the Max Samples parameter to something like 64 to speed up renders.
Enabling Denoising for final renders and change the denoiser to OptiX with the passes set to Albedo and Normal
Ensure Use Tiling is enabled and the tile size is less than the render resolution.
Go to the Render Properties tab and enable ONLY these render passes.

Combined
Z
Diffuse Color
Emission
Ambient Occlusion

Render the image and notice the artifact.

Here is a file with steps 1-6 done for you.
#93710 Simplified.blend

The artifact is the horizontal line near the top of this image:

I have simplified the file and got some steps to reproduce it from scratch. 1. Create a new .blend file, add an object to it that fills the entire view of the camera, and change the render engine to Cycles. 2. Change the render resolution to `9464 x 4320` (I couldn't reproduce the issue with other render resoltuions) 3. Ensure `Adaptive Sampling` is on. You can reduce the `Max Samples` parameter to something like 64 to speed up renders. 4. Enabling Denoising for final renders and change the denoiser to `OptiX` with the passes set to `Albedo and Normal` 5. Ensure `Use Tiling` is enabled and the tile size is less than the render resolution. 6. Go to the `Render Properties` tab and enable ONLY these render passes. - Combined - Z - Diffuse Color - Emission - Ambient Occlusion 7. Render the image and notice the artifact. Here is a file with steps 1-6 done for you. [#93710 Simplified.blend](https://archive.blender.org/developer/F12694668/T93710_Simplified.blend) The artifact is the horizontal line near the top of this image: ![#93710 - Simplified Artifact.jpg](https://archive.blender.org/developer/F12694636/T93710_-_Simplified_Artifact.jpg)

Patrick Mours self-assigned this 2021-12-09 12:37:03 +01:00

Patrick Mours commented

2021-12-09 12:37:03 +01:00

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

Patrick Mours commented

2021-12-09 12:37:03 +01:00

So the problem here is a traditional case of integer overflow. What is happening is that for denoising, Cycles loads the entire huge image into a GPU buffer. Since 1766549418 this now theoretically works with the OptiX denoiser by splitting up this image buffer into smaller tiles that are denoised separately. But there is a catch: the width, height and rowStrideInBytes values passed to the denoiser are 32-bit integers and unfortunately there is code to calculate the size of the entire image by multiplying height (or a y coordinate close to that) and rowStrideInBytes. In this case the result is larger than what a 32-bit integer can hold and therefore overflows, causing the denoiser to read/write data at bogus locations, which shows up as artifacts. rowStrideInBytes is affected by the width of the image and how many render passes are active, hence why the problem goes away if you reduce the width or disable enough render passes (including hidden render passes that one does not select in the GUI, e.g. enabling auto tiling adds a hidden render pass, adaptive sampling adds two hidden render passes). To fix this the calculation will need to be done in 64-bit (by casting to 64-bit before multiplying).

Unfortunately this needs fixing in both Cycles and inside the driver, so it's not a quick thing to do ...

So the problem here is a traditional case of integer overflow. What is happening is that for denoising, Cycles loads the entire huge image into a GPU buffer. Since 1766549418 this now theoretically works with the OptiX denoiser by splitting up this image buffer into smaller tiles that are denoised separately. But there is a catch: the `width`, `height` and `rowStrideInBytes` values passed to the denoiser are 32-bit integers and unfortunately there is code to calculate the size of the entire image by multiplying `height` (or a y coordinate close to that) and `rowStrideInBytes`. In this case the result is larger than what a 32-bit integer can hold and therefore overflows, causing the denoiser to read/write data at bogus locations, which shows up as artifacts. `rowStrideInBytes` is affected by the width of the image and how many render passes are active, hence why the problem goes away if you reduce the width or disable enough render passes (including hidden render passes that one does not select in the GUI, e.g. enabling auto tiling adds a hidden render pass, adaptive sampling adds two hidden render passes). To fix this the calculation will need to be done in 64-bit (by casting to 64-bit before multiplying). Unfortunately this needs fixing in both Cycles and inside the driver, so it's not a quick thing to do ...

Raimund Klink commented

2021-12-10 10:14:39 +01:00

Added subscriber: @Raimund58

Vladimir commented

2021-12-10 14:02:01 +01:00

Added subscriber: @Wovchick

Sergey Sharybin commented

2022-02-18 16:48:50 +01:00

Added subscriber: @Renderbicks

Sergey Sharybin commented

2022-02-23 16:20:01 +01:00

@pmoursnv, Think is better (for the time being, at least) copy an implementation of the tiled denoiser to optix/device_impl.cpp so that we don't bump driver version requirement.

Additionally to the integer overflow in the optixUtilDenoiserInvokeTiled we might have an overflow in the denoiser related kernels. I've tried addressing those, but is only the first tile which is denoised correctly. There is something I am missing and wouldn't mind a second pair of eyes. Attached current state of the patch. It is a bit excessive on int64 usage in the optixUtilDenoiserInvokeTiled, in theory should be enough to cast before multiplication, I was just trying to eliminate any doubt of me missing some overflow.

optix_int_overflow.patch

@pmoursnv, Think is better (for the time being, at least) copy an implementation of the tiled denoiser to `optix/device_impl.cpp` so that we don't bump driver version requirement. Additionally to the integer overflow in the `optixUtilDenoiserInvokeTiled` we might have an overflow in the denoiser related kernels. I've tried addressing those, but is only the first tile which is denoised correctly. There is something I am missing and wouldn't mind a second pair of eyes. Attached current state of the patch. It is a bit excessive on int64 usage in the `optixUtilDenoiserInvokeTiled`, in theory should be enough to cast before multiplication, I was just trying to eliminate any doubt of me missing some overflow. [optix_int_overflow.patch](https://archive.blender.org/developer/F12884273/optix_int_overflow.patch)

Patrick Mours commented

2022-02-23 16:37:37 +01:00

I was able to get rid of the artifacts with just this patch and a driver that has the driver-side fixes (they made it into the last r510 release, so need at least Windows 511.79 or Linux 510.54):

--- "a/optix_denoiser_tiling.h"
+++ "b/optix_denoiser_tiling.h"
@@ -147,15 +147,15 @@ inline OptixResult optixUtilDenoiserSplitImage(
                                       std::min( tileWidth, input.width - copied_x );

             OptixUtilDenoiserImageTile tile;
-            tile.input.data               = input.data + ( inp_y - inputOffsetY ) * input.rowStrideInBytes
-                                            + ( inp_x - inputOffsetX ) * inPixelStride;
+            tile.input.data               = input.data + (size_t)( inp_y - inputOffsetY ) * input.rowStrideInBytes
+                                            + (size_t)( inp_x - inputOffsetX ) * inPixelStride;
             tile.input.width              = inp_w;
             tile.input.height             = inp_h;
             tile.input.rowStrideInBytes   = input.rowStrideInBytes;
             tile.input.pixelStrideInBytes = input.pixelStrideInBytes;
             tile.input.format             = input.format;

-            tile.output.data               = output.data + inp_y * output.rowStrideInBytes + inp_x * outPixelStride;
+            tile.output.data               = output.data + (size_t)inp_y * output.rowStrideInBytes + (size_t)inp_x * outPixelStride;
             tile.output.width              = copy_x;
             tile.output.height             = copy_y;
             tile.output.rowStrideInBytes   = output.rowStrideInBytes;

But had last tried this a while ago, and had since been waiting for the driver to release, so will need to recheck too.

I was able to get rid of the artifacts with just this patch and a driver that has the driver-side fixes (they made it into the last r510 release, so need at least Windows 511.79 or Linux 510.54): ``` --- "a/optix_denoiser_tiling.h" +++ "b/optix_denoiser_tiling.h" @@ -147,15 +147,15 @@ inline OptixResult optixUtilDenoiserSplitImage( std::min( tileWidth, input.width - copied_x ); OptixUtilDenoiserImageTile tile; - tile.input.data = input.data + ( inp_y - inputOffsetY ) * input.rowStrideInBytes - + ( inp_x - inputOffsetX ) * inPixelStride; + tile.input.data = input.data + (size_t)( inp_y - inputOffsetY ) * input.rowStrideInBytes + + (size_t)( inp_x - inputOffsetX ) * inPixelStride; tile.input.width = inp_w; tile.input.height = inp_h; tile.input.rowStrideInBytes = input.rowStrideInBytes; tile.input.pixelStrideInBytes = input.pixelStrideInBytes; tile.input.format = input.format; - tile.output.data = output.data + inp_y * output.rowStrideInBytes + inp_x * outPixelStride; + tile.output.data = output.data + (size_t)inp_y * output.rowStrideInBytes + (size_t)inp_x * outPixelStride; tile.output.width = copy_x; tile.output.height = copy_y; tile.output.rowStrideInBytes = output.rowStrideInBytes; ``` But had last tried this a while ago, and had since been waiting for the driver to release, so will need to recheck too.

Michael Klein commented

2022-02-24 08:31:49 +01:00

In #93710#1312853, @pmoursnv wrote:
I was able to get rid of the artifacts with just this patch and a driver that has the driver-side fixes (they made it into the last r510 release, so need at least Windows 511.79 or Linux 510.54):

Have you also tested the latest Studio driver 511.65?

> In #93710#1312853, @pmoursnv wrote: > I was able to get rid of the artifacts with just this patch and a driver that has the driver-side fixes (they made it into the last r510 release, so need at least Windows 511.79 or Linux 510.54): Have you also tested the latest Studio driver 511.65?

Patrick Mours commented

2022-02-24 14:37:32 +01:00

Like I said, you need at least 511.79 on Windows, 511.65 is too old.

blender-admin commented

2022-03-10 11:55:43 +01:00

This issue was referenced by 8cdee3a6d4

This issue was referenced by 8cdee3a6d4306fa7cae73b9f7bcd761e90f39937

Sergey Sharybin commented

2022-03-10 16:30:13 +01:00

Changed status from 'Confirmed' to: 'Resolved'

Sergey Sharybin closed this issue

2022-03-10 16:30:13 +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

Renderings vom 2.9 LTS produce different results than 3.0 #93710