Human-oriented and high-performing transpiler for Python.
The main aim of transpyle is to let everyone who can code well enough in Python, benefit from modern high-performing computer hardware without need to reimplement their application in one of traditional efficient languages such as C or Fortran.
Framework consists of mainly the following kinds of modules:
- parser
- abstract syntax tree (AST) generalizer
- unparser
- compiler
- binder
At least some of the modules are expected to be implemented for each language supported by the framework.
The modules are responsible for transforming the data between the following states:
- language-specific code
- language-specific AST
- extended Python AST
- compiled binary
- Python interface for compiled binary
And thus:
- parser transforms language-specific code into language-specific AST
- AST generalizer transforms language-specific AST into extended Python AST
- unparser transforms extended Python AST into language-specific code
- compiler transforms language-specific code into compiled binary
- binder transforms compiled binary into Python interface for compiled binary
The intermediate meeting point which effectively allows code to actually be transpiled between languages, is the extended Python AST.
Using Python AST as the intermediate representation, enables the AST to be directly manipulated, and certain performance-oriented transformations can be applied. Current transpiler implementation aims at:
- inlining selected calls
- decorating selected loops with compiler-extension pragmas
More optimizations will be introduced in the future.
Some (if not all) of the above optimizations may have very limited (if not no) performance impact in Python, however when C, C++ or Fortran code is generated, the performance gains can be much greater.
The command-line interface (CLI) of transpyle allows one to translate source code files in supported languages.
The API of transpyle allows using it to make your Python code faster.
The most notable part of the API is the transpile
decorator, which in it's most basic form
is not very different from Numba's jit
decorator.
import transpyle
@transpyle.transpile('Fortran')
def my_function(a: int, b: int) -> int:
return a + b
Additionally, you can use each of the modules of the transpiler individually, therefore transpyle can support any transformation sequence you are able to express:
import pathlib
import transpyle
path = pathlib.Path('my_script.py')
code_reader = transpyle.CodeReader()
code = code_reader.read_file(path)
from_language = transpyle.Language.find('Python 3.6')
to_language = transpyle.Language.find('Fortran 95')
translator = transpyle.AutoTranslator(from_language, to_language)
fortran_code = translator.translate(code, path)
print(fortran_code)
As transpyle is under heavy development, the API might change significantly between versions.
Transpyle intends to support selected subsets of: C, C++, Cython, Fortran, OpenCL and Python.
For each language pair and direction of translation, the set of supported features may differ.
C-specific AST is created via pycparse, and some of elementary C syntax is transformed into Python AST.
Not implemented yet.
Parsing declarations, but not definitions (i.e. function signature, not body). And only selected subset of basic types and basic syntax is supported.
Only very basic syntax is supported currently.
Not implemented yet.
Not implemented yet.
Fortran-specific AST is created via Open Fortran Parser, then that AST is translated into Python AST.
Currently, the Fortran unparser uses special attribute fortran_metadata
attached
to selected Python AST nodes, and therefore unparsing raw Python AST created directly from ordinary
Python file might not work as expected.
The above behaviour will change in the future.
Not implemented yet.
Not implemented yet.
Python 3.6 with whole-line comments outside expressions is fully supported. Presence of end-of-line comments or comments in expressions might result in errors.
Python 3.6 with whole-line comments outside expressions is fully supported. Presence of end-of-line comments or comments in expressions might result in errors.
Python 3.5 or later.
Python libraries as specified in requirements.txt.
Building and running tests additionally requires packages listed in dev_requirements.txt.
Support for transpilation from/to specific language requires additional Python packages
specified in extras_requirements.json, which can be installed using the pip extras
installation formula pip3 install transpyle[extras]
where those extras
can be one or more of the following:
- All supported languages:
all
- C:
c
- C++:
cpp
- Cython:
cython
- Fortran:
fortran
- OpenCL:
opencl
Therefore to enable support for all languages, execute pip3 install transpyle[all]
.
Alternatively, to enable support for C++ and Fortran only, execute
pip3 install transpyle[cpp,fortran]
.
Additionally, full support for some languages requires the following software to be installed:
- C++:
- a modern C++ compiler -- fully tested with GNU's
g++
versions 7 and 8 and partially tested with LLVM'sclang++
version 7 - SWIG (Simplified Wrapper and Interface Generator) -- tested with version 3
- a modern C++ compiler -- fully tested with GNU's
- Fortran:
- a modern Fortran compiler -- fully tested with GNU's
gfortran
versions 7 and 8 and partially tested with PGI'spgfortran
version 2018
- a modern Fortran compiler -- fully tested with GNU's
The core functionality of transpyle is platform-independent. However, as support of some languages depends on presence of additional software, some functionality might be limited/unavailable on selected platforms.
Transpyle is fully tested on Linux, and partially tested on OS X and Windows.
pip3 install transpyle[all]
There is a docker image prepared so that you can easily try the transpiler.
First, download and run the docker container (migth require sudo):
docker pull "mbdevpl/transpyle"
docker run -h transmachine -it "mbdevpl/transpyle"
By default, this will download latest more or less stable development build,
if you wish to use a specific release, use "mbdevpl/transpyle:version"
instead.
Then, in the container:
python3 -m jupyter notebook --ip="$(hostname -i)" --port=8080
Open the shown link in your host's web browser, navigate to examples.ipynb, and start transpiling!
Below is the list of papers describing various aspects of transpyle and/or principles behind it. Further research is ongoing, so the list might be extended in the future.
M. Bysiek, A. Drozd and S. Matsuoka, Migrating Legacy Fortran to Python While Retaining Fortran-Level Performance Through Transpilation and Type Hints, PyHPC 2016: 6th Workshop on Python for High-Performance and Scientific Computing @ SC16, Salt Lake City, Utah, United States of America, 2016, pp. 9-18
Abstract:
We propose a method of accelerating Python code by just-in-time compilation leveraging type hints mechanism introduced in Python 3.5. In our approach performance-critical kernels are expected to be written as if Python was a strictly typed language, however without the need to extend Python syntax. This approach can be applied to any Python application, however we focus on a special case when legacy Fortran applications are automatically translated into Python for easier maintenance. We developed a framework implementing two-way transpilation and achieved performance equivalent to that of Python manually translated to Fortran, and better than using other currently available JIT alternatives (up to 5x times faster than Numba in some experiments).
M. Bysiek, M. Wahib, A. Drozd and S. Matsuoka, Towards Portable High Performance in Python: Transpilation, High-Level IR, Code Transformations and Compiler Directives (Unreferred Workshop Manuscript), 2018-HPC-165: 研究報告ハイパフォーマンスコンピューティング, Kumamoto, Kumamoto, Japan, 2018, pp. 1-7
Abstract:
We present a method for accelerating the execution of Python programs. We rely on just-in-time automatic code translation and compilation with Python itself being used as a high-level intermediate representation. We also employ performance-oriented code transformations and compiler directives to achieve high performance portability while enabling end users to keep their codebase in pure Python. To evaluate our method, we implement an open-source transpilation framework with an easy-to-use interface that achieves performance better than state-of-the-art methods for accelerating Python.