pyg-lib and ROCm

[Aethor]

Hi,

Thank you for this repo, thanks to it I was able to run some pytorch_geometric operations on my AMD GPU.

However, one thing is missing for a complete pytorch_geometric install: pyg-lib. I tried to build it myself, but of course it fails by telling you you need a CUDA install. Honestly I'm not too sure where to start. I don't think pyg-lib currently supports ROCm out of the box. I would guess since it is lower level, there might be more work involved to adapt it. Do you have any information on that matter?

[Looong01]

Hi,

Thank you for this repo, thanks to it I was able to run some pytorch_geometric operations on my AMD GPU.

However, one thing is missing for a complete pytorch_geometric install: pyg-lib. I tried to build it myself, but of course it fails by telling you you need a CUDA install. Honestly I'm not too sure where to start. I don't think pyg-lib currently supports ROCm out of the box. I would guess since it is lower level, there might be more work involved to adapt it. Do you have any information on that matter?

Yes, u r right. I am trying to port pyg-lib from CUDA to ROCM but I meet some difficulties. pyg-lib uses some Custom Operators by cutlass but rocm doesn't have something called hiptlass instead it use composable kernel (ck). And all the things are different between them. I do not know much about ck now. But I am still trying that.

[Aethor]

Interesting. Is there anywhere we can help / follow your efforts? I'm not at all versed in cuda/rocm kernels, but I would be interested in learning to help advance the state of the ROCm ecosystem.

[Looong01]

Interesting. Is there anywhere we can help / follow your efforts? I'm not at all versed in cuda/rocm kernels, but I would be interested in learning to help advance the state of the ROCm ecosystem.

Thank u for your help. I open a repo here to do the porting. And the only one thing left is implement matmul_kernel operator by composable kernel. We can do this by learning the CUDA impl that how it could be and find the same/similar functions and APIs of composable kernel and implement it. You can find the references here.

Now this implement is by hipblaslt and I find that it doesn't work totally during the testing. So we must re-implement it by composable kernel. I know it is difficult but if u can provide any help, it will be great!

[Aethor]

Thank you for the pointers!! I will have an in-depth look at it after work.

[Aethor]

Sorry for the time off, I have been having difficulties building composable kernel (ROCm/composable_kernel#2759). Also, is it me or the composable kernel documentation is very incomplete? Am I missing something?

[Looong01]

Sorry for the time off, I have been having difficulties building composable kernel (ROCm/composable_kernel#2759). Also, is it me or the composable kernel documentation is very incomplete? Am I missing something?

Actually, U dont need to build composable kernel by yourself. If u install rocm stack by the official instructions, composable kernel should be install automatically. Check whether composable kernel is already installed in your rocm home path, like "/opt/rocm/". U can directly include the headers from "/opt/rocm/include/ck" or "/opt/rocm/include/ck_tile" if there it is.

[Aethor]

You are right, it's just that I installed rocm using the Fedora package manager, and sadly ck is not packaged yet