Get data is slow

Hi everyone,

This is my part of my code but I have a problem that when I want to get data it takes ~90ms. where is the problem?

at::Tensor image_tensor = torch::from_blob(, { 1, _img_h, _img_w, _img_d }, at::kFloat);

image_tensor =*_device);

image_tensor = image_tensor.permute({ 0, 3, 1, 2 });

image_tensor[0][2] =
image_tensor[0][1] =
image_tensor[0][0] =

std::vectortorch::jit::IValue inputs;
torch::Tensor logits_tensor = _module->forward(inputs).toTensor();
torch::Tensor argmax_tensor = logits_tensor.argmax(1, false);

argmax_tensor = argmax_tensor.toType(at::kInt);

// it takes about ~90ms
auto begin = std::chrono::high_resolution_clock::now();
argmax_tensor =;
auto end = std::chrono::high_resolution_clock::now();

OS: Win 10
GPU: GTX 1080Ti
CUDA: 10.1


Keep in mind that the cuda api is asynchronous. And only wait for ops to actually run when the new value is needed outside of the GPU. So here the copy will have to wait for all the ops above to finish before running. You want to synchronize the cuda device before gettting timings if you want to see the runtime of a single op.

Thanks for the reply, Yes I know that but time on Ubuntu is less than 30. Should I change the code in Windows?

You should not rely on any of these timings if you don’t do proper sync. It is most likely due to a slower GPU or a faster CPU on your windows machine.