Created by: wojtuss

PR types

New features

PR changes

OPs

Describe

This patch adds support for quantization with shift (aka asymmetric quantization). The support is required by the fusion_gru op quantization (https://github.com/PaddlePaddle/Paddle/issues/27330).

Quantization with shift is performed according to the formula:

out_u8 = in_f32 * scale + shift

When shift is nonzero, the output of quantization is always of unsigned int8 data type. Dequantization formula is:

out_f32 = in_u8/scale - shift/scale

Dequantization with shift expects unsigned int8 input.

Support in quantization passes will come in a separate PR.