I am using cuDNN to perform convolutions of images with predefined 2D filters, taking advantage of its state-of-the-art and fine-tuned convolution implementations.
My filters can become quite big (up to 101x101, maybe more in the future), but feature a sparse format (only 1 to ~5 contiguous elements per row).
I was wondering if support for sparse input / output / filter was planned for cuDNN. I searched the documentation for mentions to sparse structures, without any luck.
If not, do you have any idea if cuSPARSE’s GEMM could be a better option, or a custom kernel ?
Thank you for any help.