How does torch.cuda.synchronize() behave?

4

Generally, torch.utils.benchmark is a great tool to profile code, as it’s adding warmup iterations, synchronizes, etc.
For information about using cudaEvents for profiling, take a look at this post, which shows an example as:

cudaEvent_t start, stop;
cudaEventCreate(&start);
cudaEventCreate(&stop);

cudaMemcpy(d_x, x, N*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_y, y, N*sizeof(float), cudaMemcpyHostToDevice);

cudaEventRecord(start);
saxpy<<<(N+255)/256, 256>>>(N, 2.0f, d_x, d_y);
cudaEventRecord(stop);

cudaMemcpy(y, d_y, N*sizeof(float), cudaMemcpyDeviceToHost);

cudaEventSynchronize(stop);
float milliseconds = 0;
cudaEventElapsedTime(&milliseconds, start, stop);

CUDA events are of type cudaEvent_t and are created and destroyed with cudaEventCreate() and cudaEventDestroy(). In the above code cudaEventRecord() places the start and stop events into the default stream, stream 0. The device will record a time stamp for the event when it reaches that event in the stream. The function cudaEventSynchronize() blocks CPU execution until the specified event is recorded. The cudaEventElapsedTime() function returns in the first argument the number of milliseconds time elapsed between the recording of start and stop. This value has a resolution of approximately one half microsecond.

The TL;DR: you have to synchronize the event either directly via the event object or globally via torch.cuda.synchronize().

1 Like