Backprop over kernels in a CNN

I dont know if it helps, but I wrote a post on it here: http://soumith.ch/ex/pages/2014/08/07/why-rotate-weights-convolution-gradient/

It goes into the calculations in a bit more detail.