search

lizhihao6 / Forward-Warp

An optical flow forward warp's lib with backpropagation using pytorch.
MIT License
109 stars 17 forks source link

Solved some compilation warning when using PyTorch 1.7.1 with CUDA 10.2 #4

Closed ccj5351 closed 3 years ago

ccj5351 commented 3 years ago

Thanks for sharing the codes, which are good examples of how to implement and access 4D tensors (like, in size [N, C, H, W]) in a low-level CUDA kernel (e.g., idx=blockIdx.x * blockDim.x + threadIdx.x).

I would like to share the following:

  1. Uncomment the macro definitions CHECK_CUDA, CHECK_CONTIGUOUS, and CHECK_INPUT. Otherwise, you probably get incorrect results when you run the test/test.py. Since the tensors generated by e.g., im0 = torch.FloatTensor(im0).permute(0, 3, 1, 2) are not contiguous. Use im0 = torch.FloatTensor(im0).permute(0, 3, 1, 2).contiguous() instead.

  2. Compilation Deprecation Warning: the solution includes:

    • a) In cuda/forward_warp_cuda.cpp file: change "#define CHECK_CUDA(x) TORCH_CHECK(x.type().is_cuda(), ..." to "#define CHECK_CUDA(x) TORCH_CHECK(x.is_cuda(), ...";
    • b) In cuda/forward_warp_cuda_kernel.cu file: change "AT_DISPATCH_FLOATING_TYPES(im0.type(), ..." to "AT_DISPATCH_FLOATING_TYPES(im0.scalar_type(), ...", and im0.data<scalar_t>() to im0.data_ptr<scalar_t>();

Now I can compile the CUDA code and get the correct results when running test/test.py.

lizhihao6 commented 3 years ago

Thank you very much for your suggestion, all parts have been modified and committed.