[Triton-MLIR] a detail questions related with mma_layout

[goostavz]

1, in the legacy code:
contig_per_thread of mma_layout sm<80 is always set to {1, 1}, which means that the convert_layout(mma->blocked) would never be able to do vectorized store to share memory.
However, in my understanding contig_per_thread should be 2 in row direction according to its definition mma_layout ver1
I guess there should be some story in the legacy code that set contig_per_thread to {1,1}, or am I missing something?

[ptillet]

Ah, I somewhat remember that. You are right that because contig_per_thread=1, shared memory ops aren't well vectorized on V100. I don't think there is any deep reason why; this happened at a time during which I was very busy with OpenAI stuff, and probably I just wanted to get good A100 perf ASAP without risking to break the V100 codegen. :D