Hi, recently I need to implement functional.conv_transpose2d() with pure C++ code from scratch, and use cudnn if needed. Where can I follow in detail how this function in PyTorch or ATen is actually implemented?
The general dispatching is done in ATen/native/Convolution.cpp, in particular the cudnn implementation is called in these lines of code.
This will then call into cudnn_convolution_transpose_forward, which is the backward operation of a vanilla convolution.
1 Like