You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In scenarios where both producer and consumer threads exist, how can we achieve synchronization using CUTLASS's barrier.sync/arrive? I understand that in barrier.arrive(a, b), a represents the number of threads required to arrive, and b is the barrier_ID. However, the number of producer and consumer threads is often different.
Here, the a parameter includes the consumer thread count (256) plus the active threads in the producer (32). However, I don't understand why it is written this way.
The text was updated successfully, but these errors were encountered:
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.
In scenarios where both producer and consumer threads exist, how can we achieve synchronization using CUTLASS's barrier.sync/arrive? I understand that in barrier.arrive(a, b), a represents the number of threads required to arrive, and b is the barrier_ID. However, the number of producer and consumer threads is often different.
In FlashAttention3, I saw this example:
https://github.com/Dao-AILab/flash-attention/blob/0dfb28174333d9eefb7c1dd4292690a8458d1e89/hopper/mainloop_fwd_sm90_tma_gmma_ws.hpp#L651
Here, the a parameter includes the consumer thread count (256) plus the active threads in the producer (32). However, I don't understand why it is written this way.
The text was updated successfully, but these errors were encountered: