Writing custom pytorch kernels with futhark

[nlgranger]

I could not find any resources about writing custom CUDA kernels for the pytorch framework. Is it something that has been investigated ?

[athas]

I have not heard of anyone who has done so. As long as you can extract the raw memory underlying the PyTorch arrays (which I assume must be possible for custom CUDA kernels to be possible), it should not be so difficult to use Futhark's C API to operate on them.