TODO: 'Much higher performance when HW-packed tensors are used for all of x, dy, dx' TODO: We should keep running tally of where we are in dataGPU and

TODO: 'Much higher performance for HW-packed tensors for both x and y' TODO: We should keep running tally of where we are in dataGPU and diffGPU TODO:

for testing or extract feature, have to call after Malloc