Years ago I started a collection of convolution optimization resources: https://github.com/mratsim/laser/wiki/Convolution-optimisati...

Also checked and apparently Nvidia Cutlass now supports generic convolutions: https://github.com/NVIDIA/cutlass