Archtecture-aware Neural Networks

A simple deep learning framework that optimizes task scheduling and memory usuage on different CPU/GPU architectures.

• I started this as a research project trying to explore optimization on task scheduling and memory usuage with DNN workloads on different architecutures.
• The ISSUE pages (some of those marked as "closed") contain partial preliminary results and interesting performance patterns.
• A classmate in GPU computing classes added GPU support too.

branch	build status
master
feng

Instructions

• For the most recent documentations online, see here
• There is a changelog in the link above.
• To test with caffe, see my forked caffe

prepare

git submodule update --init
mkdir build
cd build

Build

We use mkl for cpu gemm().

source /opt/intel/bin/compilervars.sh intel64
source /opt/intel/mkl/bin/mklvars.sh intel64

Then build with

cmake -DUSE_MKL=on -DAWNN_USE_FLT32=on ..

• BUILD in stampede2

in the builddir, run

../scripts/build_stampede2.sh

When mkl is not avaible install openblas and build with -DUSE_OPENBLAS=on

sudo ./install-apt.sh
cmake -DUSE_OPENBLAS=on -DAWNN_USE_FLT32=on ..