1.3.2 release
Broadcast tuning Better checking of inputs Copy/reduce code simplification
This commit is contained in:
+2
-2
@@ -34,7 +34,7 @@ ncclResult_t enqueue(const void* sendbuff,
|
||||
{
|
||||
if (stream != comm->prevStream) { // sync required for calls in different streams
|
||||
comm->prevStream = stream;
|
||||
CUDACHECK( cudaStreamWaitEvent(stream, comm->doneEvent, 0) );
|
||||
CUDACHECK(cudaStreamWaitEvent(stream, comm->doneEvent, 0), ncclUnhandledCudaError);
|
||||
}
|
||||
|
||||
ncclResult_t ret;
|
||||
@@ -42,7 +42,7 @@ ncclResult_t enqueue(const void* sendbuff,
|
||||
|
||||
// Always have to record done event because we don't know what stream next
|
||||
// collective will be in.
|
||||
CUDACHECK( cudaEventRecord(comm->doneEvent, stream) );
|
||||
CUDACHECK(cudaEventRecord(comm->doneEvent, stream), ncclUnhandledCudaError);
|
||||
comm->opSched += 1;
|
||||
return ret;
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user