Synchronize CUDA calls in Libtorch

C++
1

Hi,
I’m trying to improve performance and in order to do so I want to measure the accurate running time of different functions calls.
Do anybody knows how to synchronize CUDA calls in Libtorch?
In python you can do:
torch.cuda.synchronize()

Thanks!

2

This should work:

CUDAStream stream = getCurrentCUDAStream();
AT_CUDA_CHECK(cudaStreamSynchronize(stream));
1 Like
3

I’m new to libtorch myself. But isn’t it possible to simply use

#include <torch/cuda.h>
torch::cuda::synchronize();

See https://pytorch.org/cppdocs/api/function_namespacetorch_1_1cuda_1a576cf6ec2b223bcabee2f80dfdf8cdc8.html.