msccl: ensure memory coherence after data receive (#1062)

[ROCm/rccl commit: f7550d83b8]
이 커밋은 다음에 포함됨:
Wenkai Du
2024-01-30 08:22:50 -08:00
커밋한 사람 GitHub
부모 367ccf5688
커밋 3e66d42d95
+1 -1
파일 보기
@@ -372,7 +372,7 @@ __device__ __forceinline__ void mscclRunInterpreter(
return;
}
if (t->hasDependence && tid == nthreads-1)
__atomic_store_n(&mscclFlags[bid].flag, (uint64_t) COMPUTE_FLAG(workIndex, iter, step), t->type == MSCCL_REDUCE ? __ATOMIC_RELEASE : __ATOMIC_RELAXED);
__atomic_store_n(&mscclFlags[bid].flag, (uint64_t) COMPUTE_FLAG(workIndex, iter, step), (t->type == MSCCL_REDUCE || t->type == MSCCL_RECV) ? __ATOMIC_RELEASE : __ATOMIC_RELAXED);
step++;
}
}