understanding the layout of data in PipeCNN

[saman-aghazadeh]

Hi,

I'm trying to apply some modifications on the PiepCNN. There are two specific questions I have with regard to the PipeCNN. First, I cannot fully understand the organization of the input and intermediate data in PipeCNN. In other words, how PipeCNN is storing the RGB (channel) data of each pixel and in which order? Further how the LANE_NUM and vector sizes do affect this organization? I can see that during the preparation of data and weights, it applies data reordering. I believe there should be a correlation between the data order and the implementation of the operations. I first need to understand this part to be able to understand how the implementation of operations is done. Second, I would like to know more about the implementation strategy of each operation, and also what are the potential directions for further performance improvements.

Any documentation on either of the above issues will be highly appreciated.

[r0drigor]

It is storing it with BGR. You can see it using the tool provided here https://github.com/pmgysel/alexnet-forwardpath which is named here https://github.com/doonny/PipeCNN/tree/master/data.