Misaligned address when doing hgemm with odd M or K about cutlass

[herberteuler]

An early study shows that this is due to vectorized loading, odd dimensions may cause unaligned address accessing hence the error. Perhaps the best solution is to set ScalarsPerLdg to 1 to disable vectorized loading.

[kerrmudgeon]

@herberteuler 's comment is correct.

Note, we have substantially refactored CUTLASS with version 2.0 released in November 2019. The HGEMM kernel should not be performing vectorized loads and is not expected to have this issue.

There also exist kernels targeting Tensor Cores for mixed-precision matrix multiply with varying alignment restrictions (making performance tradeoffs). You could consider trying those to see if they address your use-case.

Closing this issue. Feel free to re-open if you have any issues with the current codebaase.

Thanks!