FluidPythran: easily speedup your Python code with Pythran

Documentation:

FluidPythran is a pure Python package (requiring Python >= 3.6 or Pypy3) to help to write Python code that can use Pythran if it is available.

Let's recall that "Pythran is an ahead-of-time (AOT) compiler for a subset of the Python language, with a focus on scientific computing. It takes a Python module annotated with a few interface description and turns it into a native Python module with the same interface, but (hopefully) faster."

Pythran is able to produce very efficient C++ code and binaries from high level Numpy code. If the algorithm is easier to express without loops, don't write loops!

Pythran always releases the GIL and can use SIMD instructions and OpenMP!

Pythran is not a hard dependency of FluidPythran: Python code using FluidPythran run fine without Pythran and without compilation (and of course without speedup)!

Overview

Python + Numpy + Pythran is a great combo to easily write highly efficient scientific programs and libraries.

To use Pythran, one needs to isolate the numerical kernels functions in modules that are compiled by Pythran. The C++ code produced by Pythran never uses the Python interpreter. It means that only a subset of what is doable in Python can be done in Pythran files. Some language features are not supported by Pythran (for example no classes) and most of the extension packages cannot be used in Pythran files (basically only Numpy and some Scipy functions).

Another cause of frustration for Python developers when using Pythran is related to manual writting of Pythran function signatures in comments, which can not be automated. Pythran uses C++ templates but Pythran users can not think with this concept. We would like to be able to express the templated nature of Pythran with modern Python syntax (in particular type annotations).

Finally, another limitation is that it is not possible to use Pythran for just-in-time (JIT) compilation so one needs to manually write all argument types.

With FluidPythran, we try to overcome these limitations. FluidPythran provides few supplementary Pythran commands and a small Python API to accelerate functions and methods with Pythran without writing the Pythran modules. The code of the numerical kernels can stay in the modules and in the classes where they were written. The Pythran files (i.e. the files compiled by Pythran), which are usually written by the user, are produced automatically by FluidPythran.

Bonus: There are FluidPythran syntaxes for both ahead-of-time and just-in-time compilations!

At run time, FluidPythran uses when possible the pythranized functions, but let's stress again that codes using FluidPythran work fine without Pythran (of course without speedup)!

To summarize, a strategy to quickly develop a very efficient scientific application/library with Python could be:

• Use modern Python coding, standard Numpy/Scipy for the computations and all the cool libraries you want.
• Profile your applications on real cases, detect the bottlenecks and apply standard optimizations with Numpy.
• Add few lines of FluidPythran to compile the hot spots.

Implementation details: Under the hood, FluidPythran creates Pythran files (one per module for AOT compilation and one per function for JIT compilation) that can be compiled at build, import or run times depending of the cases. Note that the developers can still read the Pythran files if needed.

Installation

The environment variable FLUIDPYTHRAN_DIR can be set to control where the cached files are saved.

A short tour of FluidPythran syntaxes

Decorator boost and command # pythran def

Most of this code looks familiar to Pythran users. The differences:

• One can use (for example) h5py and mpi4py (of course not in the Pythran functions).
• # pythran def instead of # pythran export (to stress that it is not the same command).
• A tiny bit of Python... The decorator @boost replaces the Python function by the pythranized function if FluidPythran has been used to produced the associated Pythran file.

Pythran using type annotations

The previous example can be rewritten without Pythran commands:

Nice (shorter and clearer than with the Pythran command) but very limited... So one can also elegantly define many Pythran signatures using in the annotations type variables and Pythran types in strings (see these examples). Moreover, it is possible to mix type hints and # pythran def commands.

Cached Just-In-Time compilation

With FluidPythran, one can use the Ahead-Of-Time compiler Pythran in a Just-In-Time mode. It is really the easiest way to speedup a function with Pythran, just by adding a decorator! And it also works in notebooks!

It is a "work in progress" so (i) it could be buggy and (ii) the API is not great, but it is a good start!

Note that the @cachedjit decorator takes into account type hints (see the example in the documentation).

Implementation details for just-in-time compilation: A Pythran file is produced for each "cachedjited" function (function decorated with @cachedjit). The file is compiled at the first call of the function and the compiled version is used as soon as it is ready. The warmup can be quite long but the compiled version is saved and can be reused (without warmup!) by another process.

Command # pythran block

FluidPythran blocks can be used with classes and more generally in functions with lines that cannot be compiled by Pythran.

For blocks, we need a little bit more of Python.

• At import time, we have fp = FluidPythran(), which detects which Pythran module should be used and imports it. This is done at import time since we want to be very fast at run time.
• In the function, we define a block with three lines of Python and special Pythran annotations (# pythran block). The 3 lines of Python are used (i) at run time to choose between the two branches (is_transpiled or not) and (ii) at compile time to detect the blocks.

Note that the annotations in the command # pythran block are different (and somehow easier to write) than in the standard command # pythran export.

Blocks can now also be defined with type hints!

Python classes: @boost and @cachedjit for methods

For simple methods only using attributes, we can write:

More examples of how to use FluidPythran for Object Oriented Programing are given here.

Make the Pythran files

There is a command-line tool fluidpythran which makes the associated Pythran files from Python files with annotations and fluidpythran code. By default and if Pythran is available, the Pythran files are compiled.

There is also a function make_pythran_files that can be used in a setup.py like this:

Note that the function make_pythran_files does not use Pythran. Compiling the associated Pythran file can be done if wanted (see for example how it is done in the example package example_package_fluidpythran or in fluidsim's setup.py).

License

FluidDyn is distributed under the CeCILL-B License, a BSD compatible french license.