Conv on CUDA docs

I want to understand how convolution inference works on CUDA. I found ATen sources, but it’s a bit hard to understand just from code. If you could share some docs it would help me.
Especially interesting how parallelization is implemented.

I still need answer to this, thanks

It seems you’ve already found the code, which should show a native im2col followed by a matmul approach and other paths which dispatch to accelerating libraries, such as cuDNN. Which part exactly are you struggling with?