Binary Convolution about riptide HOT 2 CLOSED

Hi,

So, the computational behavior is the same although operations used are different, right?
Thanks for your reply!

from riptide.

Hello,

Thanks for your interest in Riptide! We only use TVM during inference. At training time all binary ops are simulated in tensorflow, so there is no xnor-popcount being applied.

You're right about xnor-popcount missing from that topi definition. The non-unipolar case there isn't actually quite bipolar, its when both weights and activations represent values in [0, 1], whereas the case listed unipolar is for when activations represent values in [0, 1] and weights represent values in [-1, 1]. Although we discuss the alternative case where both activations and weights represent [-1, 1] in our paper, it is computationally equivalent to the [0, 1] and would only require changing the and operation for an xnor.

from riptide.