GPU-accelerated compiler for a simple programming language, which outputs RISC-V machine code.
The project consists of several binaries, some of which serve as tools during the compilation process:
pareas
, the compiler itself.pareas-json
, a json parser implemented using similar techniques as the compiler.pareas-lpg
, a lexer and parser generator for parallel lexers and parsers.
The basic usage of the Pareas compiler is
$ pareas <input path> -o <output path>
See pareas --help
for additional options. See doc/language.md
for a description of the programming language.
Usage of the json parser is similar to the compiler itself. There is no output, however. It simply parses the supplied json file and optionally prints some statistics.
The lexer and parser generator is used to generate Futhark sources from a grammar definition, and its most basic invocation is
pareas-lpg --lexer <lexer grammar path> --parser <parser grammar path> -o <output basename> --namespace <namespace>
The output of this tools is a bunch of tables that are to be embedded in C++ and Futhark programs. The following files are generated:
<output basename>.hpp
contains C++ definitions of:- Token- and production constants.
- Utility functions.
- Structures.
- Extern variables representing the generated tables.
<output basename>.cpp
contains the implementation of the utility functions, the definitions of the extern variables representing the generated tables.<output basename>.dat
containing the actual generated tables.<output basename>.S
containing an incbin statement for the generated data files.<output basename>.fut
containing Futhark definitions for tokens and productions.
See doc/lpg.md
for a syntax description of both the lexical analyzer and parser generators. Also see src/json/json.lex
and src/json/json.g
for an example of how lexer and parser grammar files should look like.
Pareas is built using the help of several tools which are also located in this project and are built as part of the compilation process. The project is laid out as follows:
src/tools/compile_futhark.py
is a tool used during building that helps with compiling Futhark. Normally, the Futhark compiler is invoked on a single source root and finds other imports by relative paths. This projects generates some Futhark files during it's build process. To avoid polluting the source directory, we copy the source tree of Futhark files into the source directory, where the generated files are also placed in. Generated files appear under thegen
folder as if relative to the project root, so to import a generated file fromsrc/compiler/frontent.fut
one has to import../../gen/generated_file
.src/compiler/
contains the compiler itself. The Futhark files in this directory implement the meat of the compiler, while the c++ files implement some driving logic such as reading the input and writing the output.src/json/
contains an example json parser implemented using similar techniques used for the main compiler.src/lpg/
contains the lexer- and parser generator.src/profiler/
contains a very simple profiler used for measuring the performance of the compiler.
The frontend of the project consists of src/compiler/frontend.fut
and all the files it references, as well as src/json/
and src/lpg/
. The backend consists of src/compiler/backend.fut
and all the files it references.
Building Pareas requires the following dependencies:
- A C++20-capable compiler such as clang or gcc.
- The Meson build system.
- Ninja or Samurai to build.
- A Futhark compiler. The latest tested version is 20.6.
- Python, which is required for Meson as well as some build tools included in the project.
Some additional dependencies such as {fmt}
are automatically downloaded by Meson, and so this requires an active internet connection.
Dependending on the selected Futhark backend, some more dependencies might be requires:
- The OpenCL backend requires OpenCL development files.
- The CUDA backend requires CUDA development files.
- The multithreaded backend requires pthread.
To compile the project, please run:
$ mkdir build
$ cd build
$ meson .. -Dfuthark-backend=[opencl|cuda|c|multicore]
$ ninja
This repository contains the source code from the following Master Thesis projects:
- Robin Voetter, "Parallel Lexing, Parsing and Semantic Analysis on the GPU", 2021, https://theses.liacs.nl/2052
- Marcel Huijben, "Parallel Code Generation on the GPU", 2021, https://theses.liacs.nl/2053
And the following publication:
- Robin F. Voetter, Marcel Huijben, Kristian F. D. Rietveld, "Compilation on the GPU?: a feasibility study", 2022, https://dl.acm.org/doi/10.1145/3528416.3530249
If you use any of the material provided in the repository, please cite the above.
@mastersthesis{voetter2021,
author = {Robin Voetter},
title = {Parallel Lexing, Parsing and Semantic Analysis on the GPU},
school = {Leiden University},
year = {2021},
}
@mastersthesis{huijben2021,
author = {Marcel Huijben},
title = {Parallel Code Generation on the GPU},
school = {Leiden University},
year = {2021},
}
@inproceedings{10.1145/3528416.3530249,
author = {Voetter, Robin F. and Huijben, Marcel and Rietveld, Kristian F. D.},
title = {Compilation on the GPU? A Feasibility Study},
year = {2022},
isbn = {9781450393386},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3528416.3530249},
doi = {10.1145/3528416.3530249},
abstract = {The emergence of highly parallel architectures has led to a renewed interest in parallel compilation. In particular, the widespread availability of GPU architectures raises the question whether compilation on the GPU is feasible. In this paper, we describe the first design and implementation of a parallel compiler from a simple imperative programming language to RISC-V machine code, that is fully executed on a GPU. To accomplish this, all stages from parsing to machine code generation were redesigned to exploit fine-grained parallelism. Experimental evaluation of the implemented prototype demonstrates our proposed parallel techniques to be effective and implementation of compilation on the GPU to be feasible. Finally, we propose a number of avenues for future work and hope to revitalize research into parallel compilation conducted in the 1980s.},
booktitle = {Proceedings of the 19th ACM International Conference on Computing Frontiers},
pages = {230–236},
numpages = {7},
keywords = {GPUs, code generation, compiler construction, parallel compilation, parsing},
location = {Turin, Italy},
series = {CF '22}
}
Futhark requires OpenCL 1.2. Unfortunately, Mesa only implements OpenCL up to version 1.1. In order to work around this, one can install the AMDGPU-PRO version of OpenCL. Note that this library does not require the entire AMDGPU-PRO stack, just the OpenCL implementation is sufficient, and this works alongside Mesa.
To install the OpenCL implementation from AMDGPU-PRO, first download the latest version of the driver from AMD. Then run:
# Extract the package
$ tar xf amdgpu-pro-<version>.tar.xz
$ cd amdgpu-pro-<version>
# If you run a debian-based distro, this deb file can be installed directly.
$ ar x opencl-orca-amdgpu-pro-icd<another version>_amd64.deb
$ tar xf data.tar.xz
# Put the driver somewhere on the system
$ mv opt/amdgpu-pro/lib/x86_64-linux-gnu/libamdocl-orca64.so /usr/lib/libamdocl-orca64.so
# Make sure that the OpenCL ICD loader can find it
$ echo libamdocl-orca64.so > /etc/OpenCL/vendors/amdocl-orca64.icd
After this, clinfo
should print an additional platform with name AMD Accelerated Parallel PRocessing
. When invoking Pareas, make sure that the right device is selected by passing pareas -d DeviceName
. DeviceName
can be found in the output of clinfo
.