OptiX denoise error on Blender 3.1 and nvidia 495.44: Misaligned address in cuMemFree_v2 #92985

New Issue

YAFU · 2021-11-10T10:28:41+01:00

YAFU commented

2021-11-10 10:28:41 +01:00

System Information
Operating system: Linux-5.11.0-38-generic-x86_64-with-glibc2.31 64 Bits
Graphics card: NVIDIA GeForce GTX 960/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 495.44

Blender Version
Broken: version: 3.1.0 Alpha, branch: master, commit date: 2021-11-09 23:23, hash: aa440923c8
Worked: Blender 2.93.5

Exact steps for others to reproduce the error
Open Car Demo (bmw27_gpu.blend) scene on Blender 3.1.0 alpha:
https://www.blender.org/download/demo-files/#cycles

(The problem also occurs with other scenes)

Enable viewport Rendered view. Enable OptiX Denoise for viewport. You will get the next error:

Misaligned address in cuMemFree_v2(mem.device_pointer) (intern/cycles/device/cuda/device_impl.cpp:801)

The error for final render image with OptiX denoise is:

Error: Misaligned address in CUDA queue copy_to_device ()

Edit:
The aforementioned was using CUDA as Cycles Render Device in preferences. If I choose OptiX as Cycles Render Device, the error message I get is:

Error: Misaligned address in CUDA queue synchronize (film_convert_float3 filter_guiding_preprocess)

**System Information** Operating system: Linux-5.11.0-38-generic-x86_64-with-glibc2.31 64 Bits Graphics card: NVIDIA GeForce GTX 960/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 495.44 **Blender Version** Broken: version: 3.1.0 Alpha, branch: master, commit date: 2021-11-09 23:23, hash: `aa440923c8` Worked: Blender 2.93.5 **Exact steps for others to reproduce the error** Open Car Demo (bmw27_gpu.blend) scene on Blender 3.1.0 alpha: https://www.blender.org/download/demo-files/#cycles (The problem also occurs with other scenes) Enable viewport Rendered view. Enable OptiX Denoise for viewport. You will get the next error: ``` Misaligned address in cuMemFree_v2(mem.device_pointer) (intern/cycles/device/cuda/device_impl.cpp:801) ``` The error for final render image with OptiX denoise is: ``` Error: Misaligned address in CUDA queue copy_to_device () ``` Edit: The aforementioned was using CUDA as Cycles Render Device in preferences. If I choose OptiX as Cycles Render Device, the error message I get is: ``` Error: Misaligned address in CUDA queue synchronize (film_convert_float3 filter_guiding_preprocess) ```

YAFU commented

2021-11-10 10:28:41 +01:00

Added subscriber: @YAFU

blender-admin commented

2021-11-10 10:28:41 +01:00

#92990 was marked as duplicate of this issue

Brecht Van Lommel commented

2021-11-10 12:46:36 +01:00

Added subscriber: @Wyk

Brecht Van Lommel commented

2021-11-10 12:46:44 +01:00

Added subscribers: @pmoursnv, @brecht

Brecht Van Lommel commented

2021-11-10 12:46:44 +01:00

CC @pmoursnv

Brecht Van Lommel commented

2021-11-10 12:47:03 +01:00

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

Steffen Dünner commented

2021-11-10 12:49:58 +01:00

Added subscriber: @SteffenD

Steffen Dünner commented

2021-11-10 12:49:58 +01:00

Just in case this helps: Doing a factory reset gets rid of the problem. In my case I had to completely erase my 3.1 user folder. Just using "File > Defaults > Load Factory Settings" didn't help for me.

YAFU commented

2021-11-10 12:53:53 +01:00

Hi. In my case renaming 3.1 config folder doesn't help. Perhaps this is a very new problem a little different from what others were having.

Brecht Van Lommel commented

2021-11-10 12:57:11 +01:00

There were changes to this code yesterday, it's just something that needs to be fixed there.

Patrick Mours commented

2021-11-10 13:05:49 +01:00

Is caused by the MSL changes from 3a4c8f406a (reverting that commit fixes it), hence why it's only happening in 3.1, but not 3.0. Not sure why yet.

But I'm already getting complains by the compiler with that commit in (that disappear with it reverted), so I'm suspecting the new macros don't do what they are supposed to do in some cases:

...
3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba is out of range. minnctapersm will be ignored
3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_shadow_catcher_matte_with_shadow is out of range. minnctapersm will be ignored
3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_shadow_catcher_half_rgba is out of range. minnctapersm will be ignored
3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_shadow_catcher is out of range. minnctapersm will be ignored
3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_cryptomatte_half_rgba is out of range. minnctapersm will be ignored
3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_cryptomatte is out of range. minnctapersm will be ignored
3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_motion_half_rgba is out of range. minnctapersm will be ignored
3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_motion is out of range. minnctapersm will be ignored
3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_float3_half_rgba is out of range. minnctapersm will be ignored
3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_float3 is out of range. minnctapersm will be ignored
...

Is caused by the MSL changes from 3a4c8f406a (reverting that commit fixes it), hence why it's only happening in 3.1, but not 3.0. Not sure why yet. But I'm already getting complains by the compiler with that commit in (that disappear with it reverted), so I'm suspecting the new macros don't do what they are supposed to do in some cases: ``` ... 3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_shadow_catcher_matte_with_shadow_half_rgba is out of range. minnctapersm will be ignored 3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_shadow_catcher_matte_with_shadow is out of range. minnctapersm will be ignored 3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_shadow_catcher_half_rgba is out of range. minnctapersm will be ignored 3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_shadow_catcher is out of range. minnctapersm will be ignored 3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_cryptomatte_half_rgba is out of range. minnctapersm will be ignored 3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_cryptomatte is out of range. minnctapersm will be ignored 3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_motion_half_rgba is out of range. minnctapersm will be ignored 3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_motion is out of range. minnctapersm will be ignored 3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_float3_half_rgba is out of range. minnctapersm will be ignored 3>CUSTOMBUILD : ptxas warning : Value of minnctapersm for entry kernel_gpu_film_convert_float3 is out of range. minnctapersm will be ignored ... ```

Steffen Dünner commented

2021-11-10 13:50:25 +01:00

In #92985#1251794, @SteffenD wrote:
Just in case this helps: Doing a factory reset gets rid of the problem. In my case I had to completely erase my 3.1 user folder. Just using "File > Defaults > Load Factory Settings" didn't help for me.

Sadly after a few hours the same problem started happening again. So resetting to factory defaults is only a temporary workaround :(

> In #92985#1251794, @SteffenD wrote: > Just in case this helps: Doing a factory reset gets rid of the problem. In my case I had to completely erase my 3.1 user folder. Just using "File > Defaults > Load Factory Settings" didn't help for me. Sadly after a few hours the same problem started happening again. So resetting to factory defaults is only a temporary workaround :(

blender-admin commented

2021-11-10 14:37:15 +01:00

This issue was referenced by blender/cycles@6ca201f56d

This issue was referenced by blender/cycles@6ca201f56d945f6029476af31858e7d010e86fb8

blender-admin commented

2021-11-10 14:37:15 +01:00

This issue was referenced by f565620435

This issue was referenced by f56562043521a5c160585aea3f28167b4d3bc77d

Patrick Mours commented

2021-11-10 15:52:16 +01:00

Changed status from 'Confirmed' to: 'Resolved'

Patrick Mours closed this issue

2021-11-10 15:52:16 +01:00

Patrick Mours self-assigned this 2021-11-10 15:52:16 +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

OptiX denoise error on Blender 3.1 and nvidia 495.44: Misaligned address in cuMemFree_v2 #92985