MVP:Matrix-vector multiplication implemented on PYNQ board

Introduction

There are a huge number of Matrix-Vector Multiplication operations in deep learning.This work is aimed at building a general module that can support different forms of neural network.

Both activation and weight are quantified into 4 bit (the MSB is the signal bit) with Learned step-size qunatization(LSQ) algorithm.

A bit serial method is adopted to complete the MAC operation and it can greatly reduce power consumption compared to tradtional method.

Architecture

Note that DMA is used to move weigth and activation. Of course you can also use .coe to load the data into BRAM.

Block Design

Software Implementation

More at ./notebook/mvp.py

Acknowledgment

This work is based on the work of @zhuiyuanjiao and @zhutmost.