Fix for GenericOp device primitive bug

[ROCm/rccl commit: 8ae1bce3bb]
这个提交包含在:
Gilbert Lee
2019-10-10 22:39:45 -07:00
父节点 fbcdfd8348
当前提交 d257970ad1
@@ -193,10 +193,11 @@ class ncclPrimitives {
if (SEND) __threadfence_system();
if (tid == 0) FOR_SEND(postSend);
if (tid == 0) FOR_RECV(postRecv);
for (int i=0; i<RECV*NRECV+SRC; i++) srcs[i] += sliceSize;
for (int i=0; i<SEND*NSEND+DST; i++) dsts[i] += sliceSize;
offset += sliceSize;
}
for (int i=0; i<RECV*NRECV+SRC; i++) srcs[i] += sliceSize;
for (int i=0; i<SEND*NSEND+DST; i++) dsts[i] += sliceSize;
offset += sliceSize;
}
__device__ void loadRecvConn(struct ncclConnInfo* conn, int i, T* directBuff) {