Implementation of Flownet

[r0drigor]

Hello, I'm trying to implement FlowNet in PipeCNN. From what I can tell, I need to (maybe) implement the kernels that enable the preparation of both images and also a Deconvolution (Transposed Convolution) kernel.

I'm having a bit of a hard time implementing the Transposed Convolution kernel. I started to think about altering the MemRead so that it can load the input and weight values to the buffer so it can perform the correct operation but I'm having a hard time doing it and I'm not even sure if that's the correct strategy. Can someone, with more knowledge (@doonny) , give me some tips/hints at what I should be doing?