Source Code: F.conv2d() Definition Location

ghee-air-moe · May 16, 2020, 11:05pm

I keep trying to find WHERE F.conv2d() is defined, like where all of it is ACTUALLY written out logically. Does anyone know where it hides?

Under torch/nn/modules/conv.py line 339 calls F.conv2d() but as I go to torch/nn/functional.py to where it supposed to be defined all i get are comments about it at line 48.

ptrblck · May 17, 2020, 7:57am

The call is eventually calling into Convolution.cpp, where some checks are applied to further dispatch the call to the proper backend (MKL, cudnn, native implementation etc.).

ghee-air-moe · May 20, 2020, 2:22pm

Thanks! It’s close, but only shows a call to some other function…sigh… I suppose it’s the secret sauce.

github.com

pytorch/pytorch/blob/1465970a343e61f2f2b104859ca7f5d7e03f5d02/aten/src/ATen/native/Convolution.cpp#L539


}


at::Tensor conv1d(
    const Tensor& input, const Tensor& weight, const Tensor& bias,
    IntArrayRef stride, IntArrayRef padding, IntArrayRef dilation, int64_t groups) {
  return at::convolution(input, weight, bias, stride, padding, dilation,
                         false, {0}, groups);
}

at::Tensor conv2d(
    const Tensor& input, const Tensor& weight, const Tensor& bias,
    IntArrayRef stride, IntArrayRef padding, IntArrayRef dilation, int64_t groups) {
  return at::convolution(input, weight, bias, stride, padding, dilation,
                         false, {{0, 0}}, groups);
}

at::Tensor conv3d(
    const Tensor& input, const Tensor& weight, const Tensor& bias,
    IntArrayRef stride, IntArrayRef padding, IntArrayRef dilation, int64_t groups) {
  return at::convolution(input, weight, bias, stride, padding, dilation,

ptrblck · May 21, 2020, 7:21am

The native implementation are not closed source and the CPU version can be found here, while the CUDA version is here.

ghee-air-moe · May 24, 2020, 8:42pm

Ptrblck, you’re awesome! Thank you for your dedication!

LearnedLately · January 30, 2022, 6:00pm

@ptrblck Is aten/src/THCUNN/generic/SpatialConvolutionMM.cu still the CUDA convolution implementation that pytorch uses? When I clone the pytorch repository with git clone --recursive https://github.com/pytorch/pytorch that file isn’t even included.

ptrblck · January 30, 2022, 9:18pm

You can find the current implementation here as it was moved to the ATen namespace.