I was trying to add GPU support for Raspberry Pi 3 B . I did found a C++ Library for Raspberry pi GPU computation. There are two possible ways now: to customize my c++ extension, or to extend a new dispatcher. Now I do have cross compile tools, yet I was kind of confused what to do next. It would be really nice if anyone has any recommendations!
Are you trying to train or do inference? If you’re doing inference you can get quite far with model quantization + qnnpack. qnnpack is optimized for arm CPUs and gets pretty good performance on the RPi.
https://pytorch.org/tutorials/intermediate/realtime_rpi.html
Adding a new GPU backend to pytorch is possible though quite a bit of work to support and maintain