CPP extensions Update

[UsernameN0tAvailable]

The cpp extensions do not compile anymore with newer version of torch, I've encountered the following two errors using versions >= 1.13.0

cpp_functions.cpp:85:28: error: ‘struct DLTensor’ has no member named ‘ctx’
   85 |       dlMTensor->dl_tensor.ctx.device_id = device_id;
      |                        

cpp_functions.cpp:24:14: error: ‘cudnn_convolution_backward_weight’ is not a member of ‘at’; did you mean ‘mps_convolution_backward_out’?
   24 |   return at::cudnn_convolution_backward_weight(
      |              ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
      |              mps_convolution_backward_out
cpp_functions.cpp:48:14: error: ‘cudnn_convolution_backward_input’ is not a member of ‘at’; did you mean ‘mps_convolution_backward_out’?
   48 |   return at::cudnn_convolution_backward_input(
      |              ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
      |              mps_convolution_backward_out

This is very likely due to versions mismatch, is there any fix for this or a newer version of the cpp extensions?

[hanspinckaers]

Hi! I don't have time to fix this now, but you can rewrite scnn to use these functions here, these use the CUDA kernels now too, so should be fast: https://github.com/pytorch/pytorch/blob/main/torch/nn/grad.py

it's relatively easy:

---

replace

StreamingCNN/scnn.py

Lines 128 to 130 in 350df03

	grad_in = cpp_functions.backward_input(inpt.shape, grad_output, weight.to(inpt.dtype), padding,
	stride, dilation, groups,
	torch.backends.cudnn.benchmark, torch.backends.cudnn.deterministic)

with:

grad_in = torch.nn.grad.conv2d_input(inpt.shape, weight, grad_output, stride,  padding, dilation, groups)  

---

replace

StreamingCNN/scnn.py

Lines 215 to 221 in 350df03

	grad_weight = cpp_functions.backward(weight.shape,
	relevant_grad.to(weight.dtype),
	relevant_input.to(weight.dtype),
	(0, 0), # padding
	stride[1:3], dilation, groups,
	torch.backends.cudnn.benchmark, # benchmark
	torch.backends.cudnn.deterministic) # deterministic

with:

grad_weight = torch.nn.grad.conv2d_weight(relevant_input.to(weight.dtype), 
                                      weight.shape, 
                                      relevant_grad.to(weight.dtype), 
                                      stride[1:3], 
                                      (0, 0),  # padding 
                                      dilation, 
                                      groups) 

Good luck!

[UsernameN0tAvailable]

yes, that is exactly it, I would have opened a PR if I had the permissions, however it's a very quick fix, thanks a lot!