search

GPUOpen-LibrariesAndSDKs / RadeonImageFilter

https://gpuopen.com/radeon-prorender-suite/
Other
49 stars 15 forks source link

Out of video memory error when running AIDenoiser sample #6

Closed deranen closed 3 years ago

deranen commented 3 years ago

Hello,

This used to work, but after updating to latest SDK/binaries I get the following output:

D:\dev\RadeonImageFilter\samples-build\AIDenoiser\Debug>AIDenoiser.exe
[MSG  @RIF]: Loaded RIF API version: 1.6.0.0x42f99d51
Fatal error: Unable to create Direct3D12 device
INFO: MIOpen version: 2.0.1
[MSG  @RIF]: Loading model: ./models/denoise_c9_hdr_gamma_2.5.pb
[ERROR@RIF]: RIF_ERROR_OUT_OF_VIDEO_MEMORY at file C:\JN\WS\RadeonProImageProcessor_Build\src\AIFilters.cpp line 653

I'm running Windows 10 with a Radeon, GeForce and integrated Intel GPU. The AMD GPU is a Radeon R9 280X.

Let me know if you need more info.

bsavery commented 3 years ago

Hi there. The AI denoiser does use 4GB memory by minimum.

deranen commented 3 years ago

Hi @bsavery. Thanks for the answer. Is there a way to control max memory used? The thing that is disappointing is that it used to work before.

BenjaminCoquelle commented 3 years ago

@deranen as Brian explained we limited to 4GB in FP32 to run the denoiser. We had to introduce that limit as it does indeed require 4GB or more, depending on the input image size. If windows 10 allows to over commit memory, it is actually bad practice as the system will likely be unstable especially if others processes run. And that's the main reason why we introduced that limit. Moreover 4-5 years old mid range card all have more than 4GB

That said we now also support FP16 inference, which will reduce the memory usage by 2x. To that matter you need to use that API rifImageFilterSetComputeType https://github.com/GPUOpen-LibrariesAndSDKs/RadeonImageFilter/blob/master/include/RadeonImageFilters.h#L925 with the flag RIF_COMPUTE_TYPE_HALF.

However, AFAIR Tahiti(the chipset on your R9) doesn't support half. It is an old card not really designed for inference But it worth trying as I haven't worked on Tahiti for a while

We do have an environment variable to change the limit, and I can share it with you. But if you get issues because of it, we will unlikely give support.

Finally we have others ideas to reduce the memory usage of our models. But it will take some time to get them on production.

deranen commented 3 years ago

@BenjaminCoquelle Thank you for the answer. I will try your suggestions, but I also understand that you can't fully support older cards. This is not a bug, so I will close this issue.