Comment by FL33TW00D

It takes a different code path for efficiency.

e.g

if (batch_size > 1024): kernel_x else: kernel_y