[QST] a redundant code

[zhoutianzi666]

What is your question?

cutlass/include/cutlass/gemm/kernel/gemm.h

Line 273 in b5d8a5d

if (!kSplitKSerial || gemm_k_iterations > 0) {

why add if (!kSplitKSerial || gemm_k_iterations > 0) {, I think if (gemm_k_iterations > 0) { is enough.

[JieRen98]

I guess keeping it is better, which will make things clear. It hints that, for the SplitK Serial case, we do not need to do mma, just do epilogue reduction only.

[zhoutianzi666]

I guess keeping it is better, which will make things clear. It hints that, for the SplitK Serial case, we do not need to do mma, just do epilogue reduction only.
ok，it means in SplitK Serial case when some thread blocks having gemm_k_iterations == 0， we do not do mma。
but there is any this case？

[JieRen98]

In addition, !kSplitKSerial will enable compile time optimization

[hwu36]

kernel/gemm.h is the old implementation. Now we use kernel/gemm_universal.h. It is called universal because it supports, batch gemm, gemm array, parallel splitk, serail splitk, standard gemm in one implementation. it is also used by cublas.

in the old days, these go through different kernel level code though they are very similar to each other.

as to this line, i think we should not check kSplitKSerial

[github-actions]

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.

[mnicely]

Closing due to inactivity. Feel free to reopen if needed.