[BUG+FIX] sgemm_sm80: s2r_atom_* type is incomplete

[osayamenja]

Describe the bug sgemm_sm80 fails to compile when dispatched to non-fp16 gemm_* functions.

Steps/Code to reproduce bug Change TA = TB = float or any non-fp16 type.

Expected behavior Code should compile.

Environment details (please complete the following information): A100

Additional context For here and elsewhere, we need to use CopyAtom<AutoVectorizingCopy, TA>.

[thakkarV]

those are not expected to compile based on just changing the dtype. you will likely have to change every tiled copy and tiled mma. The goal of that example is not to be generic for all data types. That's what the CUTLASS API is for.

[osayamenja]

I see, will close this then.

[ccecka]

Reopen, that example does specialize for half_t and I believe the templated functions were intended to be specialized for float but somehow remained a template. These should be fixed.

It appears to me that gemm_nt should be functional for float, but gemm_tn may be broken? Either way, it appears that neither one should be templated.

[github-actions[bot]]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[jwu1980]

@osayamenja The fix is in the 4.2 tag release. Please reopen the issue if the fix doesn't work.