msccl: use relaxed atomics on scratch buffer (#1075)

[ROCm/rccl commit: 5257c753c5]
This commit is contained in:
Wenkai Du
2024-02-08 12:09:56 -08:00
zatwierdzone przez GitHub
rodzic ff2be03272
commit 1538b908ac
@@ -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 || t->type == MSCCL_RECV) ? __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) && (t->dstBuffer != MSCCL_SCRATCH_BUFFER)) ? __ATOMIC_RELEASE : __ATOMIC_RELAXED);
step++;
}
}