What is the right buffer layout for performing cnn related operations - convolution, batch normalization etc. ? -c++

I have to implement CNN related operations in OpenCL and c++11. I generally keep my buffer layouts ( input, output, filter) as WHC but while looking at various dl based frameworks, I found that they tend to employ a different buffer layout even for operations like batch normalization i.e. NCHW or NHWC. I don't understand the reasons for doing so. Kindly shed some light on it . (especially in the context of GPU programming)