WinogradConvKernel

Introduction

This is a test code for fast Winograd convolution which accelerates the convolution inference by reducing the number of multiplications. This repo implementes two widely used output tile sizes, i.e., 2 and 4.

There are two versions of Winograd kernel, including a) original Winograd with element-wise matrix multiplication (EWMM) and without parallelization for transformations and; b) parallized Winograd with general matrix multiplication (GEMM) and parallelization for transformations.

Version a) can significantly accelerate the computation process compared with version b), especially when the dimension is large.

Command for Testing

python run_conv.py

The code will run with regular convolution, Winograd w/o parallelization, and Winograd w/ parallelization. We can set m = 2 or m=4 to change the output tile size.

Extension

Based on the Winograd kernel, we can easliy apply it to neural networks, e.g., ResNet by replacing the class nn.Conv2d with Winograd(). It is worth noting that for Winograd training, remember to modify the test code to make the weight be trainable using nn.Parameter() in __init__.