swDNNv1.0

A Deep Learning Library for Sunway TaihuLight

Features

Convolution layer

three convolutional layer implementations.

1. Explicit-GEMM
2. Implicit-GEMM
3. Winograd

Pooling

Batch Normalization

LSTM

Tensor Transformation

How to

Use

mkdir ./build/
cd build && cmake .. && make

Add new layers

1. mkdir a directory using your new layer name in ./slave/
   mkdir ./slave/conv
2. write your slave code with name sw_slave_XXX, i.e. sw_slave_conv.c
3. write your host code (interface to call slave code) in ./src/swXXX.c, i.e. swconv.c
4. write your header file in ./include/swXXX.h, i.e. swconv.h
5. write your test code in ./unitest/src/test_XXX.c
6. update test header file in ./unitest/include/test_XXX.h

Author

Jiarui Fang fang_jiarui@163.com