Fix: barrier_wait → mbarrier_wait, barrier_arrive → mbarrier_arrive
This commit is contained in:
@@ -435,11 +435,11 @@ class FmhaV3StageCMulti:
|
||||
cute.arch.fence_view_async_tmem_load()
|
||||
# Async-proxy fence so the TMA store sees the SMEM writes.
|
||||
cute.arch.fence_proxy("async.shared", space="cta")
|
||||
cute.arch.barrier_arrive(
|
||||
cute.arch.mbarrier_arrive(
|
||||
barrier_id=self.epilog_sync_bar_id,
|
||||
number_of_threads=32 * len(self.epilogue_warp_id),
|
||||
)
|
||||
cute.arch.barrier_wait(
|
||||
cute.arch.mbarrier_wait(
|
||||
barrier_id=self.epilog_sync_bar_id,
|
||||
number_of_threads=32 * len(self.epilogue_warp_id),
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user