Can DirectML help run a CUDA based AI App on AMD GPU?

[SomeAB]

Hello I came across DirectML as I was looking for setting up the following app by facebookresearch on a local windows10 machine. As I don't have Nvidia card, but rather a AMD Vega64, I have not been able to run it so far.

I read in DirectML documentation that it can run cross-platform code, so I'm wondering if its possible to run the below mentioned app on a Windows 10 PC (Ubuntu subsystem?) ??

https://github.com/facebookresearch/pifuhd

[PatriceVignola]

@SomeAB

The app you linked is using PyTorch. We don't have a DirectML backend for PyTorch at the moment, but this is definitely something we could be interested in supporting in the future if there is a demand from the community.

[SomeAB]

Hi Patrice. Thats good to hear, but I briefly read at one point that I can somehow achieve it using onnxruntime. Onnx does support Directml, but a person needs to build it manually. Any idea how? Also, If there was a easy pip3 command for Onnx-Directml, it would be great.

Also are there any video/walkthrough tutorials for Directml yet? perhaps some showing the samples listed on microsoft website's DirectML docs section?

[PatriceVignola]

I'm not very familiar with this app, but as far as I can tell, it exclusively uses PyTorch. It's true that you can convert PyTorch models to ONNX models and use onnxruntime to run them, but you would need to swap the PyTorch calls that the app uses for onnxruntime calls.

To use DirectML with onnxruntime, you will need to follow the Building from source instructions.

[wchao1115]

@SomeAB I'm not super familiar with the app you refer to, but as for ONNX runtime, we work closely with them. If your question is how to get a version of ONNX runtime that works with DirectML, the easiest way currently is to install this nuget package on your Windows PC. This package comes with a version of DirectML that works well with it, but would work on Windows only. Good luck with your experiment. Let us know how it goes.

[alimoezzi]

@SomeAB I'm not super familiar with the app you refer to, but as for ONNX runtime, we work closely with them. If your question is how to get a version of ONNX runtime that works with DirectML, the easiest way currently is to install this nuget package on your Windows PC. This package comes with a version of DirectML that works well with it, but would work on Windows only. Good luck with your experiment. Let us know how it goes.

It's only useful when you have VS solution not working in with pytorch

[AGenchev]

I read of DirectML today and I must say I'm much impressed by Microsoft for this one. Might make me to boot more often in Windows if this works better than TF on CUDA in Linux. All AMD APU owners probably will do so, because of the non-existing support AMD provides for its APUs (If this works on APUs). Eh, if this was under Vulkan it'd be more portable and we could use on RaspberryPi...

[jstoecker]

There are preview builds of PyTorch-DirectML now: https://devblogs.microsoft.com/windowsai/introducing-pytorch-directml-train-your-machine-learning-models-on-any-gpu/

[6p5ra]

@SomeAB

The app you linked is using PyTorch. We don't have a DirectML backend for PyTorch at the moment, but this is definitely something we could be interested in supporting in the future if there is a demand from the community.

The future never came.

[AGenchev]

They made it. Nowadays pytorch just works on AMD GPU & APU. I use it on Ryzen 5700G APU.