Awesome

Introduction

This code implements fast cuda kernels for DNN inference, especially for convolution layers / residule blocks in ResNet. Specifically, the kernels combine three parts into one piece:

• Convolution
• Batch Nomalization (BN + Scale)
• Activation (ReLU)

For implementation details, please refer to the technical report included in this repo. Winograd algorithm is used for 3 * 3 convolutional kernels.

Usage

mkdir data
python data_generator.py
make
./Test 0

• Set parameters in data_generator.py
• Run 6 test cases with changing numbers from 0 to 5 after ./Test

Results

3 * 3 Kernels

1 * 1 Kernels [BUGGY NUMBERS]