AGNA-FCCM2023

Model-Platform Optimized Deep Neural Network Accelerator Generation through Mixed-integer Geometric Programming

Link to the paper: https://ieeexplore.ieee.org/abstract/document/10171535

Citation:

@INPROCEEDINGS{10171535,
  author={Ding, Yuhao and Wu, Jiajun and Gao, Yizhao and Wang, Maolin and So, Hayden Kwok-Hay},
  booktitle={2023 IEEE 31st Annual International Symposium on Field-Programmable Custom Computing Machines (FCCM)}, 
  title={Model-Platform Optimized Deep Neural Network Accelerator Generation through Mixed-Integer Geometric Programming}, 
  year={2023},
  volume={},
  number={},
  pages={83-93},
  doi={10.1109/FCCM57271.2023.00018}}

Introduction

AGNA is an open-source hardware generator for Deep Neural Network (DNN). Given the specifications of target DNN model and FPGA platform, AGNA can produce FPGA accelerator that is optimized for target model-platform combination. AGNA can be generally separated into software and hardware parts:

• Software: AGNA analyzes the specifications of each layer and solves mixed-integer geometric programming. AGNA first generates a high-efficiency accelerator based on a general PE array architecture. Then based on the generated accelerator, AGNA generates the schedule and instruction of each layer.
• Hardware: AGNA provides the synthesizable source code of each hardware component. The target accelerator can be built by substituting the parameter with the generated one from the software. A template project for zcu102 is also provided.

Verified environment

• SCIP 8.0.3 with TPI=omp and Ipopt 3.14.10
• Python 3.9
• Vivado 2021.2

Usage

• Run in docker (Optional):

  We also provide a docker image that contains the necessary environment (except for Vivado). The image can be built by:

  docker build --build-arg HOST_UID=`id -u` --build-arg HOST_GID=`id -g` -t agna-local .

  Run docker image:

  docker run -it -v `pwd`:/home/user/workspace agna-local

  Current directory will be mounted to /home/user/workspace in the container.

• Run software:

  cd software
  export SCIPOPTDIR=<SCIPOPT_PATH>  # not required in docker
  conda env create --file environment.yml
  conda activate agna
  make all PLATFORM=<TARGET_PLATFORM> MODEL=<TARGET_MODEL>

  Make sure <SCIPOPT_PATH>/bin/scip is executable and specification files are available at software/spec/platforms/<TARGET_PLATFORM>.json and software/spec/models/<TARGET_MODEL>.json.

  Generated architecture and schedule are in software/results/<TARGET_PLATFORM>-<TARGET_MODEL>.

• Build hardware:

  cd hardware
  make all

  Generated project and bitstream are in hardware/prj.

Build software environment from scratch

1. Prerequisite:

   sudo apt update
   sudo apt install -y wget cmake g++ m4 xz-utils libgmp-dev unzip zlib1g-dev libboost-program-options-dev libboost-serialization-dev libboost-regex-dev libboost-iostreams-dev libtbb-dev libreadline-dev pkg-config git liblapack-dev libgsl-dev flex bison libcliquer-dev gfortran file dpkg-dev libopenblas-dev rpm
   sudo apt install -y libopenmpi-dev libomp-dev

2. Build Ipopt:

   mkdir coinbrew && cd coinbrew
   wget https://raw.githubusercontent.com/coin-or/coinbrew/master/coinbrew
   chmod +x coinbrew
   ./coinbrew fetch [email protected]
   export IPOPT_DIR=/tools/Ipopt  # install directory of Ipopt, could be other places
   mkdir -p ${IPOPT_DIR}
   ./coinbrew build Ipopt --prefix=${IPOPT_DIR} --test --no-prompt --verbosity=3
   sudo ./coinbrew install Ipopt --no-prompt

3. Build SCIPOpt:

   • Download scipoptsuite-8.0.3.tgz.

   • Install:

   tar xzf scipoptsuite-8.0.3.tgz
   cd scipoptsuite-8.0.3
   mkdir build && cd build
   export SCIPOPT_PATH=/tools/scipoptsuite-8.0.3 # install directory of SCIPOPT, could be other places
   cmake .. -DCMAKE_INSTALL_PREFIX=${SCIPOPT_PATH} -DIPOPT_DIR=${IPOPT_DIR} -DTPI=omp
   make
   make check
   make install