CGRAOmp: OpenMP Compiler for CGRAs

CGRAOmp is a front-end to compile OpenMP codes for Coarse-Grained Reconfigurable Architectures (CGRAs) based on LLVM.

The API document generated by doxgen is available.

To use this tool, you can either use docker images or build the LLVM libraries on your machine.

Publication

1. Takuya Kojima, Boma Adhi, Carlos Cortes, Yiyu Tan, Kentaro Sano, "An Architecture-Independent CGRA Compiler enabling OpenMP Applications", 2022 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), virtual, May 2022.

Using CGRAOmp with Docker

There are two options: 1) using prebuilt images and 2) building images with Dockerfile.

Prebuilt images

You can pull the prebuilt images from DockerHub.

CentOS based image

$ docker pull tkojima0107/cgraomp:centos

Ubuntu 20.04 based image

$ docker pull tkojima0107/cgraomp:ubuntu-20.04

They uses llvm_docker as base images.

Building these images

You can build the above images with Dockerfiles in docker

Building CGRAOmp

Prerequisites

Installation of LLVM 12.0.x

CGRAOmp consists of several out-of-tree Passes. In the beginning, you need to build and install LLVM 12.0.x according to the official guide (see llvm-project).

Python3 packages (Optional)

Compiler driver needs the following python packages:

• To visualize data flow graph
  1. graphviz
  2. watchdog
• To run backend mapping process
  1. watchdog
  2. rich (for rich console mode)

Build and Install

git clone https://github.com/hal-lab-u-tokyo/CGRAOmp.git
cd CGRAOmp
mkdir build
cmake .. [-DCMAKE_INSTALL_PREFIX=<path>] [-DCGRAOMP_ENABLE_DOXYGEN=On]
make -j`proc`
make install

If you want to change the path where the tools and libraries are installed, please specify the directory path by -DCMAKE_INSTALL_PREFIX=<path>.

If you want to build the API document with Doxgen, please use -DCGRAOMP_ENABLE_DOXYGEN=On.

Architecture types

The target architecture is defined by a JSON file. This repository includes the following types in share/presents

1. Decoupled CGRA
2. Time-Multiplexed CGRA (Experimental)

Usage

Sample code (vec_madd.c)

#include <cgraomp.h>
#define N 100
void kernel() {
    int X[N], Y[N], Z[N];
    const int a = 3;

    #pragma omp target parallel for map(to:X[0:N],Y[0:N]) map(from:Z[0:N])
    for (int64_t i = 0; i < N; i++) {
        Z[i] = X[i] * a + Y[i];
    }
}

Compilation commmad

$ cgraomp-cc vec_madd.c --cgra-config decoupled_affine_AG.json

If no error occurs, it will generate a dot file as a data flow graph.

How to connect your back-end mapper

This compiler only extracts a data flow graph of the OpenMP offloading kernel. To run mapping with the generated DFGs automatically, a script file is needed.

#!/bin/sh

your_mapping_command ${DOTFILE_NAME}

In the script, an environment variable ${DOTFILE_NAME} provides an actual file name of the generated DFG file.

Options

Necessary option

• --cgra-config (-cc): specify the path of CGRA configuration file

Generation options

• -o: specifies the output file name
• -v: enables verbose mode
• -save-temps: saves temporary files during the compilation
• --enable-cgraomp-debug: shows debug messsage in the CGRA OpenMP pass

Options for host code or pre-optimization

• -O(0|1|2|3|s|z): optimization level for host code
• -Xclang=<arg>: passes an argument for clang
• --preopt-pipeline: changes optimizaiton pipeline in pre-optimization stage

Options for CGRAOmp Pass

• --enable-custom-inst: enables custom instruction support
• --diagnostic-file=<path>: specifies the file path of a diagnostic file
• -Xcgraomp=<arg>: passes other options for CGRAOmp passes

Options for DFG generation

• --load-dfg-pass-plugin=<path>: loads DFG pass plugins
• --dfg-pass-pipeline: specifies the pass pipeline for DFG optimization (comma-separeted)
• --dfg-file-prefix: specifies the prefix used for data flow graph file name
• --visualize-dfg: generates image files of DFGs (graphviz is needed)
• --visualize-dfg-type: specifies file type of the visualized file (default: png)
• --simplify-dfg-name: uses simplified file name for generated DFG files
• --cgra-dfg-plain: uses plain node label

Options for backend process

• --backend-runner: specifies a runner script to drive a back-end mapping
• --backend-runner-command: specifies command to run the runner script (default: bash)
• --backend-proc-num: specifies the number of backend process to run simulatanceously (default: 1)
• --backend-panel-num: specifies the number of panel to view the backend process stdout (only in rich console mode)
• --backend-nowait: finishes immediately when all the backend processes are finished
• --no-rich-console: disables rich console mode

Related tools

• GenMap: a mapping algorith based on genetic algortihm

License

CGRAOmp is published under the MIT license (see LICENSE file).