[QST]Question About the Use of MMA In-Flight in SS_WarpSpecialized

[ziyuhuang123]

I don’t understand why MMA in-flight is used in SS_WarpSpecialized. As shown in the diagram below, I’ve illustrated my understanding with a pipeline diagram. If there is MMA in-flight, then after the computation of one stage is completed, the producer cannot immediately read the next set of values. Instead, it has to wait until the next MMA finishes before reading the current stage’s producer values. Does this waiting make sense?

[hwu36]

How did you get your figure? Is it what you profiled or your understanding. compute warp starts when its data is ready https://github.com/NVIDIA/cutlass/blob/main/include/cutlass/gemm/collective/sm90_mma_array_tma_gmma_ss_warpspecialized.hpp#L532

[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.

[github-actions]

This issue has been labeled inactive-90d due to no recent activity in the past 90 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.