slow denoise Ubuntu 22.04

[stevelcb]

Hi everyone
GraX 3.0.2 using AI denoise 3.0.1
Beelink Ryzen 7, 5800 8c/16t with Radeon graphics

A 6000x4000 pixel image takes over 30 minutes to denoise.
I've tried binning the image, inference batch sizes of 1, 2 and 4 and with/without hardware acceleration checked but without improvement in denoising speed.
AI 1.0.0 denoises the same image in around 5 minutes and background extraction using AI 1.0.1 is almost instantaneous.

Any ideas?
Cheers and thanks for a great project.

[schmelly]

Hi Steve,

hardware acceleration on Linux is available via onnxruntime and CUDA only for our pre-build package (afaik this requires an NVIDIA card). As onnxruntime does not provide pre-build rocm binaries at the moment unfortunately we cannot provide a pre-build GraXpert package for amd hardware either currently. Your only option would probably be to run GraXpert from source and build onnxruntime by yourself.

For completeness, the current Linux dependencies are:

• CUDA 12
• libcudnn8
• a CUDA 12 compatible kernel module

CS, David

[stevelcb]

Hi and thanks for the prompt reply.
If I build from source with onnxruntime, will that give me gpu acceleration on amd?
I've had a look but it doesn't look trivial. If anyone has a recipe/method for this, do please share.
Cheers

[schmelly]

Hi Steve,

according to the installation instructions of onnxruntime, there is no binary distribution of it for AMD. That is, in my understanding, you would have to build onnxruntime first (I don't have experience with this step) and use that for starting GraXpert from sources. If you can build onnxruntime successfully and all driver dependencies are met it should work for you.

CS, David