Channel wise convolution

Thanks for clarifying. @J_Johnson is right and the order will be the same. Shuffling the kernels would also break the training.
This post visualizes the processing and also links to another code snippet verifying the results.