Custom rocm hip and c++ extensions

[salykova]

🚀 The feature, motivation and pitch

Dear PyTorch developers and community,

We have nice tutorial cpp_extension on custom cuda extensions written by Peter Goldsborough. I’m wondering if the same can be done but on AMD GPUs with kernels written using rocm HIP. I mean the following: call custom forward+backward hip kernel from pytorch and include it in deep learning pipeline. Is it currently supported and are there any limitations?

Does somebody have experience of writing custom hip/c++ kernels and using them in pytorch?

cc @sekyondaMeta @svekars @carljparker @NicolasHug @kit1980 @subramen @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport

[svekars]

@jeffdaily

[jithunnair-amd]

@jeffdaily We have some internal documentation that highlights some of the differences in enabling PyTorch extensions for ROCm. Shall I put that together into something we can publish on the pytorch documentation?

[jeffdaily]

/assigntome

[svekars]

This issue has been unassigned due to inactivity. If you are still planning to work on this, you can still send a PR referencing this issue.