cgo-artifact-2020

Artifact repository for paper Automatic Generation of High-Performance Quantized Machine Learning Kernels

Hardware Requirements:

Host machine able to run Docker.
Raspberry Pi 3B+

Installation:

Host machine:

Download code and build docker image

$ git clone https://github.com/cowanmeg/cgo-artifact-2020.git

$ docker build -t cgo-art cgo-artifact-2020/docker

$ docker run -it-v <absolute/path/cgo-artifact-2020>:/artifact cgo-art

Build TVM

$ cd tvm

$ mkdir build && cd build

$ cmake .. && make -j4

Copy scripts to raspberry pi

$ cd /artifact$

$ export PI=<Raspberry Pi's IP address>

$ scp -r pi <pi-user@ip>:/home/pi-user

Raspberry Pi:

$ cd /home/pi-user/pi

$ ./download_build.sh

Experiments:

Raspberry Pi:

Benchmark handwritten quantized convolutions .

$ ./convolutions_pytorch.sh

Start TVM RPC server - leave running for following experiments

$ ./rpc.sh

Note: if port of server is not 9090 on the host machine run:

$ export PORT=<port>

Host:

Synthesize ARM micro-kernels

./synthesize.sh

Benchmark quantized convolutions

./convolutions.sh

Benchmark ResNet18

./end_to_end.sh

Generate graphs

./generate_graphs.sh