From 48107b18c95b696022f3daa49aca0c0d5840b4bc Mon Sep 17 00:00:00 2001 From: Wenkai Du <43822138+wenkaidu@users.noreply.github.com> Date: Mon, 11 Dec 2023 20:33:15 -0800 Subject: [PATCH] msccl: fix data corruption with MTYPE_RW (#1014) [ROCm/rccl commit: 81602814a7486844a325bfc28b214ce17efaf85e] --- projects/rccl/src/collectives/device/msccl_kernel_impl.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/projects/rccl/src/collectives/device/msccl_kernel_impl.h b/projects/rccl/src/collectives/device/msccl_kernel_impl.h index c211c86921..897481e695 100644 --- a/projects/rccl/src/collectives/device/msccl_kernel_impl.h +++ b/projects/rccl/src/collectives/device/msccl_kernel_impl.h @@ -460,7 +460,7 @@ __device__ __forceinline__ void mscclRunInterpreter( return; } if (t->hasDependence && tid == nthreads-1) - __atomic_store_n(&mscclFlags[bid].flag, (uint64_t) COMPUTE_FLAG(workIndex, iter, step), __ATOMIC_RELAXED); + __atomic_store_n(&mscclFlags[bid].flag, (uint64_t) COMPUTE_FLAG(workIndex, iter, step), t->type == MSCCL_REDUCE ? __ATOMIC_RELEASE : __ATOMIC_RELAXED); step++; } }