Make your Python code fly at transonic speeds!

Documentation: https://transonic.readthedocs.io

Transonic is a pure Python package (requiring Python >= 3.6) to easily accelerate modern Python-Numpy code with different accelerators (currently Cython, Pythran and Numba, but potentially later Cupy, PyTorch, JAX, Weld, Pyccel, Uarray, etc...).

The accelerators are not hard dependencies of Transonic: Python codes using Transonic run fine without any accelerators installed (of course without speedup)!

You can try Transonic online by clicking this button: .

The long-term project

Transonic targets Python end-users and library developers.

It is based on the following principles:

• We'd like to write scientific / computing applications / libraries with pythonic, readable, modern code (Python >= 3.6).

• In some cases, Python-Numpy is too slow. However, there are tools to accelerate such Python-Numpy code which lead to very good performances!

• Let's try to write universal code which express what we want to compute and not the special hacks we want to use to make it fast. We just need nice ways to express that a function, a method or a block of code has to be accelerated (and how it has to be accelerated). We'd like to be able to do this in a pythonic way, with decorators and context managers.

• There are many tools to accelerate Python-Numpy code! Let's avoid writting code specialized for only one of these tools.

• Let's try to keep the code as it would be written without acceleration. For example, with Transonic, we are able to accelerate (simple) methods of classes even though some accelerators don't support classes.

• Let's accelerate/compile only what needs to be accelerated, i.e. only the bottlenecks. Python and its interpreters are good for the rest. In most cases, the benefice of writting big compiled extensions (with Cython or in other languages) is negligible.

• Adding types is sometimes necessary. In modern Python, we have nice syntaxes for type annotations! Let's use them.

• Ahead-of-time (AOT) and just-in-time (JIT) compilation modes are both useful. We'd like to have a nice, simple and unified API for these two modes.

  • AOT is useful to be able to distribute compiled packages and in some cases, more optimizations can be applied.
  • JIT is simpler to use (no need for type annotations) and optimizations can be more hardware specific.

  Note that with Transonic, AOT compilers (Pythran and Cython) can be used as JIT compilers (with a cache mechanism).

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

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

What we have now

We start to have a good API to accelerate Python-Numpy code (functions, methods and blocks of code). The default Transonic backend uses Pythran and works well. Here, we explain why Pythran is so great for Python users and why Transonic is great for Pythran users. There are also (more experimental) backends for Cython and Numba.

Installation and configuration

pip install transonic

Transonic is sensible to environment variables:

• TRANSONIC_DIR can be set to control where the cached files are saved.
• TRANSONIC_DEBUG triggers a verbose mode.
• TRANSONIC_COMPILE_AT_IMPORT can be set to enable a mode for which Transonic compiles at import time the Pythran file associated with the imported module. This behavior can also be triggered programmatically by using the function set_compile_at_import.
• TRANSONIC_NO_REPLACE can be set to disable all code replacements. This is useful to compare execution times and when measuring code coverage.
• TRANSONIC_COMPILE_JIT can be set to false to disable the compilation of jited functions. This can be useful for unittests.
• TRANSONIC_BACKEND to choose between the supported backends. The default backend "pythran" is quite robust. There are now 3 other backends: "cython", "numba" and "python" (prototypes).
• TRANSONIC_MPI_TIMEOUT sets the MPI timeout (default to 5 s).

A short tour of Transonic public API

Transonic supports both ahead-of-time and just-in-time compilations. When using the API for AOT compilation, the files need to be "compiled" to get speedup.

Decorator boost and command # transonic def

import h5py
import mpi4py

from transonic import boost

# transonic def myfunc(int, float)

@boost
def myfunc(a, b):
    return a * b

...

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).
• # transonic def instead of # pythran export.
• A tiny bit of Python... The decorator @boost replaces the Python function by the compiled function if Transonic has been used to produced the associated Pythran/Cython/Numba file.

With type annotations

The previous example can be rewritten without # transonic def. It is the recommended syntaxes for ahead-of-time compilation:

import numpy as np
import h5py

from transonic import boost

@boost
def myfunc(a: float, d: int):
    return a * np.ones(d * [10])

...

Nice (shorter and clearer than with the Pythran command) but very limited (only simple types and only one signature)... So one can also elegantly define many signatures using Transonic types and/or Pythran types in strings (see these examples and our API to define types (and fused types) in transonic.typing).

Moreover, it is possible to add more signatures with # transonic def commands.

Targetting Cython

Cython needs to know the types of local variables to really speedup the computations. Transonic is able to write fast Cython from such code:

from transonic import boost

@boost(boundscheck=False, wraparound=False)
def mysum(arr: "float[:]"):
    i: int
    n: int = arr.shape[0]
    result: float = 0.0
    for i in range(n):
        result += arr[i]
    return result

Just-In-Time compilation

With Transonic, one can use the Ahead-Of-Time compilers Pythran and Cython 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!

import numpy as np

from transonic import jit

def func0(a, b):
    return a + b

@jit
def func1(a, b):
    return np.exp(a) * b * func0(a, b)

Note that the @jit 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 "JITed" function (function decorated with @jit). 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.

Define accelerated blocks

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

from transonic import Transonic

ts = Transonic()

class MyClass:

    ...

    def func(self, n):
        a, b = self.something_that_cannot_be_pythranized()

        if ts.is_transpiled:
            result = ts.use_block("name_block")
        else:
            # transonic block (
            #     float a, b;
            #     int n
            # )

            # transonic block (
            #     complex a, b;
            #     int n
            # )

            result = a**n + b**n

        return self.another_func_that_cannot_be_pythranized(result)

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

• At import time, we have ts = Transonic(), 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 (# transonic 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 # transonic block are different (and somehow easier to write) than in the standard command # pythran export.

Blocks can also be defined with type hints!

Python classes: @boost and @jit for methods

For simple methods only using attributes, we can write:

import numpy as np

from transonic import boost

A = "float[:]"

@boost
class MyClass:

    arr0: A
    arr1: A

    def __init__(self, n):
        self.arr0 = np.zeros(n)
        self.arr1 = np.zeros(n)

    @boost
    def compute(self, alpha: float):
        return (self.arr0 + self.arr1).mean() ** alpha

More examples on how to use Transonic for Object Oriented Programing are given here.

Make the Pythran/Cython/Numba files and compile the extensions

There is a command-line tool transonic which makes the associated Pythran/Cython/Numba files from a Python file. For example one can run:

# Pythran is the default backend
transonic myfile.py -af "-march=native -DUSE_XSIMD -Ofast"
# Now using Cython
transonic myfile.py -b cython

By default and if the Python compiler is available, the produced files are compiled.

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

from pathlib import Path

from transonic.dist import make_backend_files

here = Path(__file__).parent.absolute()

paths = ["fluidsim/base/time_stepping/pseudo_spect.py"]
make_backend_files([here / path for path in paths])

Note that make_backend_files does not compile the backend files. The compilation has to be done after the call of this function (see for example how it is done in the example packages or in fluidsim's setup.py).

License

Transonic is distributed under the BSD License.