What is "custom" "custom-back" size for SGEMM in cutlass?

Accelerated Computing GPU-Accelerated Libraries
1

Hi! I am learning the SGEMM implementation and find a “dispatch policy” in cutlass! (Although in official page disappear…only find it in huggingface’s cutlass) See here: pytorch_block_sparse/dispatch_policies.h at e71b54273a25ecebd2aa57181d87664e5de98dc3 · huggingface/pytorch_block_sparse · GitHub

So what does this “custom” “custom back” means? Thank you!!!