[QST] Where is exactly the definiton code of fprop convolution?

What is your question?
I am studying example 16 and the hierarchy of fprop convolution within cutlass. I would like to know where is the low-level definition of the convolutions and their params.

Inside the example, the constructor used is linked with https://github.com/NVIDIA/cutlass/blob/main/include/cutlass/conv/device/implicit_gemm_convolution.h, but I cannot see there the source kernel or the structure of its parameters (activation, filters...). All I can see is a reference template constructor, but I don't know the code on which is based:

  /// Argument structure
  using Arguments = typename UnderlyingKernel::Arguments;

  /// Kernel parameters object
  typename UnderlyingKernel::Params params_;

I would like if it is used the constructor that the Convolution Readme links for the device (in my case GPU ampere), which is the following

Start with examples/16_ampere_tensorop_conv2dfprop and include/cutlass/conv/device/implicit_gemm_convolution.h, then follow UnderlyingKernel::Arguments and Params into the implementation. Compare that path with media/docs/implicit_gemm_convolution.md and tools/util/include/cutlass/util/reference/device/convolution.h for the Ampere case. Done means identifying the low-level convolution definition and parameter structures requested.