Illegal memory access in backward kernel

[5had3z]

I tried to adapt this and another variant (ARFlow https://github.com/lliuz/ARFlow), to pytorch 1.6.0 (I am also using CUDA 10.2).

When running the auxillary test, it works when I have small height and widths of a dummy input (e.g. 16x64x64x64), however if I try to have bigger HW dimensions (e.g. 16x3x96x96), I get "an illegal memory access was encountered".

I'm not sure what else I need to do to adapt this/or ARFlow's variant to pytorch 1.6.0 so I can use this. I have also tried to bring everything up to date by using torch::Tensors and AT Dispatching.

My adaptation can be found below
https://github.com/5had3z/stereo-to-all/tree/debug/correlation_pkg/nnet_training_framework/correlation_package