binary-search thresholding RTL expects 2^N-1 thresholds, however, Brevitas implements signed activations using "narrow range" which drops one quantized values to keep the symmetry around 0. to get around this, we prepend a dummy threshold and offset the bias by -1.
binary-search thresholding RTL expects 2^N-1 thresholds, however, Brevitas implements signed activations using "narrow range" which drops one quantized values to keep the symmetry around 0. to get around this, we prepend a dummy threshold and offset the bias by -1.