So you want to write algorithms that are easy to maintain as in Python and
you want performance as in FORTRAN or C++? Lets give a try to Pythran!
Pythran is a Python-to-c++ translator that turns Python modules into native
c++11 modules. From a user point of view, you still import your module, but
under the hood… There is much more happening!
Pythran is not a full Python-to-c++ converter, as is shedskin. Instead it takes a subset of the Python language and turns it into heavily templatized c++ code instantiated for your particular types.
Say hello to:
polymorphic functions (!)
lambdas
list comprehension
map, reduce and the like
dictionary, set, list
exceptions
file handling
(partial) numpy support
Say bye bye to:
classes
polymorphic variables [ but not all of them :-) ]
In a nutshell, Pythran makes it possible to write numerical algorithms in Python and to have them run faster. Nuff said.
Pythran depends on the following packages:
ply>=3.4
setuptools
gast~=0.5.0
numpy
beniget~=0.4.0
Pythran also depends on Boost and xsimd, however Pythran’s PyPI packages vendor these dependencies for convenience (note though that conda-forge and some Linux distros may unvendor one or both of these).
You also need a modern C++11 enabled compiler (e.g. g++>=5, clang>=3.5), that supports atomic operations (N3290) and variadic template (N2555).
The prefered way to install Pythran is using pip install pythran or
conda install pythran. Yet if you want to install from sources,
here is the procedure.
First get the sources:
$> git clone https://github.com/serge-sans-paille/pythran
From the source directory, run:
$> pip install .
pythran should now be on your PATH. If not, it’s possible pip
installed to .local (this happens if the default site-packages
location requires elevated permissions) - fix this by setting your path to:
$> export PATH=$PATH:$HOME/.local/bin
It makes the pythran command available to you.
The setup.py scripts automates this. The test target, as in:
$> python setup.py test
runs a whole (and long) validation suite (you will need to install the
pytest module first to use it).
If these tests fail, you are likely missing some of the requirements. You can
set site specific flags in your ~/.pythranrc, read the doc a bit further!
To begin with, you need… a Python function in a module. Something like:
<<dprod.py>>
def dprod(arr0, arr1):
return sum([x*y for x,y in zip(arr0, arr1)])
is perfect. But due to \_o< typing, arr0 and arr1 can be of any type,
so Pythran needs a small hint there. Add the following line somewhere in your
file, say at the top head, or right before the function definition:
#pythran export dprod(int list, int list)
This basically tells Pythran the type of the forthcoming arguments.
Afterwards, frenetically type:
$> pythran dprod.py
\o/ a dprod.so native module has been created and you can play with it
right now, as if it where a normal module:
>>> import dprod # this imports the native version if available
>>> dprod.dprod([1,2], [3,4])
11
The speedup will not be terrific because of the conversion cost from Python to C++.
So let’s try again with a well-known example. Let me introduce the almighty matrix multiply!:
<<mm.py>>
def zero(n,m): return [[0]*n for col in range(m)]
def matrix_multiply(m0, m1):
new_matrix = zero(len(m0),len(m1[0]))
for i in range(len(m0)):
for j in range(len(m1[0])):
for k in range(len(m1)):
new_matrix[i][j] += m0[i][k]*m1[k][j]
return new_matrix
This is a slightly more complex example, as a few intrinsics such as range and
len are used, with a function call and even nested list comprehension. But
Pythran can make its way through this. As you only want to export the
matrix_multiply function, you can safely ignore the zero function and
just add:
#pythran export matrix_multiply(float list list, float list list)
to the source file. Note how Pythran can combine different types and infer the resulting type. It also respects the nested list structure of Python, so you are not limited to matrices…
Enough talk, run:
$> pythran mm.py
One touch of magic wand and you have your native binary. Be amazed by the
generation of a mm.so native module that runs around 20x faster than the
original one. timeit approved!
But scientific computing in Python usually means Numpy. Here is a well-known Numpy snippet:
<<arc_distance.py>>
import numpy as np
def arc_distance(theta_1, phi_1, theta_2, phi_2):
"""
Calculates the pairwise arc distance
between all points in vector a and b.
"""
temp = (np.sin((theta_2-theta_1)/2)**2
+ np.cos(theta_1)*np.cos(theta_2) * np.sin((phi_2-phi_1)/2)**2)
distance_matrix = 2 * np.arctan2(np.sqrt(temp), np.sqrt(1-temp))
return distance_matrix
This example uses a lot of Numpy ufunc. Pythran is reasonably good at handling such expressions. As you already know, you need to export it, giving its argument types by adding:
#pythran export arc_distance(float[], float[], float[], float[])
To the input file. You can compile it as the previous code:
$> pythran arc_distance.py
and you’ll get a decent binary. But what you really want to do is:
$> pythran -DUSE_XSIMD -fopenmp -march=native arc_distance.py
which basically tells the compiler to parallelize and vectorize loops using whatever hardware available on your machine. Then you’ll get really fast code!
The #pythran export commands are critical to Pythran. In fact if they are
missing, Pythran will complain loudly (and fail miserably). So let us dive into
this complex language!
There is currently only one Pythran command, the export command. Its syntax is:
#pythran export function_name(argument_type* [, argument_type ? *])
where function_name is the name of a function defined in the module, and
argument_type* is a comma separated list of argument types, composed of any
combination of basic types and constructed types. argument_type ? * is a
comma separated list of optional argument types, similar to argument_type
but followed by a ?.
What is an argument_type? Anything that looks like a Python basic type!
Constructed types are either tuples, introduced by parenthesis, like (int,
(float, str)) or lists (resp. set), introduced by the list (resp.
set) keyword:
argument_type = basic_type
| (argument_type+) # this is a tuple
| argument_type list # this is a list
| argument_type set # this is a set
| argument_type []+ # this is a ndarray, C-style
| argument_type [::]+ # this is a strided ndarray
| argument_type [:,...,:]+ # this is a ndarray, Cython style
| argument_type [:,...,3]+ # this is a ndarray, some dimension fixed
| argument_type:argument_type dict # this is a dictionary
basic_type = bool | byte | int | float | str | None | slice
| uint8 | uint16 | uint32 | uint64 | uintp
| int8 | int16 | int32 | int64 | intp
| float32 | float64 | float128
| complex64 | complex128 | complex256
Note
When using a 2D array, overloads of the function involved are created to accept both C-style and Fortran-style arrays.
To avoid generating too many functions, one can force the memory layout using order(C) or order(F) after the
array decalaration, as in int[:,:] order(C).
The same syntax can be used to export global variable (in read only mode):
#pythran export var_name
In a similar manner to the Python import statement, it’s possible to chain the export, as in:
#pythran export var_name0, var_name1, function_name(argument_type0)
Multiple overloads can be specified for the same Python function:
#pythran export function_name(argument_type0)
#pythran export function_name(argument_type1)
In the case of function with default parameters, you can either omit the parameter, and in that case it uses the default one, or explicitly state it’s argument type:
#pythran export function_name()
#pythran export function_name(argument_type0)
#pythran export function_name(argument_type0, argument_type1)
def function_name(a0=1, a1=True):
pass
When specifying multiple overloads, instead of listing them, you can use the or operator to list the alternatives, as in:
#pythran export function_name(type0 or type1, type2, type3 or type4)
which is exactly equivalent to:
#pythran export function_name(type0, type2, type3)
#pythran export function_name(type0, type2, type4)
#pythran export function_name(type1, type2, type3)
#pythran export function_name(type1, type2, type4)
Easy enough, isn’t it?
Note
Multiline exports are supported, just use comments to protect them, as in:
#pythran export river_boa(int,
# float,
# bool)
Note
It is in fact possible to analyse a code without specifications, but you
cannot go further that generic (a.k.a. heavily templated) c++ code. Use the
-e switch!
Instead of writing the export lines in the .py file, it is possible to
write them, without the #pythran prefix, inside a file that has the same
path has the .py file, but with the .pythran extension. For instance,
file I_love.py can have its export lines in the I_love.pythran file, using the syntax:
export function_name(argument_type*)
Pythran tries hard to produce code that has the same observable behavior as the original Python code. Unfortunately it’s not always possible:
Pythran does not support heterogeneous containers (except tuples).
There is no BigInt support. All integer operations are performed on
np.int_, which maps to C long type. Beware that as a consequence, the
size of this type is system-dependent.
In most cases (with the notable exception of numpy.ndarray), Pythran is
working on a deep copy of the original Python arguments. This copy shares no memory
relationship with the original object, which means that modifying the
argument content in Pythran won’t modify the original argument content.
Likewise, objects generated by Pythran cannot share reference (in the sense
of is) with one of the input argument.
Of course, this limitation doesn’t apply to non exported functions.
As Pythran translates the Python code in native code that only depends on
libpython for data translation, it can release the GIL during the actual
function run. And that’s what it does :-) Put an another way, you can rip some
speedup at the Python level just by spawning multiple threading.Thread.
The magic function %%pythran is made available to ipython users through an
extension. The extension is located in the extensions/ directory
and can be loaded using IPython’s magic function:
%load_ext pythran.magic
Once done, you can pythranize your code from the IPython shell:
%%pythran
#pythran export foo()
def foo(): print 'hello'
You can pass arguments to this magic, as in:
%%pythran(-O2 -fopenmp)
#pythran export foo()
def foo(): print 'hello'
When distributing a Python application with Pythran modules, you can either:
use the module as a regular Python module (a .py file). After all, they
are 100% Python compatible.
use Pythran as a Python to C++ transpiler and then integrate the generated C++ into your package build the same way as you would with other C++ code (see SciPy for an example of doing this with Meson).
Use Pythran’s PythranExtension in order to directly build extension modules
as part of your build with setuptools.
For the last two options above, you should declare pythran as a build dependency
in the pyproject.toml file for your package:
[build-system]
build-backend = "mesonpy" # or: "setuptools.build_meta"
requires = [
"meson-python", # or: "setuptools"
"pythran",
Pythran comes with a PythranExtension class that extends setuptools and
can be used like this to compile Pythan modules into extension modules:
from distutils.core import setup
# These two lines are required to be able to use pythran in the setup.py
import setuptools
setuptools.dist.Distribution(dict(setup_requires='pythran'))
from pythran.dist import PythranExtension, PythranBuildExt
setup(...,
ext_modules=[PythranExtension("mymodule", ["mymodule.py"])],
cmdclass={"build_ext": PythranBuildExt})
PythranBuildExt is optional, but necessary to build extensions with
different C++ compilers. It derives from distuil’s build_ext by default, but
you can change its base class by using PythranBuildExt[base_cls] instead.
all configuration options supported in .pythranrc can also be passed
through the optional config argument, in the form of a list, e.g.
config=['compiler.blas=openblas']
Note
There’s no strong compatibility guarantee between Pythran version at C++ level. As a consequence, a code distrubuted under pythran version 0.x should depend on that exact version, as version 0.y may introduce some changes.
This behavior is likely to change with revisions >= 1.
Python does not have good support for cross compilation; neither does Pythran’s CLI interface. Using the Python-to-C++ transpilation and using a build system like Meson or CMake with solid support for cross compilation to compile the pythran-generated C++ files into extension modules is your best bet.
Note that Pythran itself is a header-only library and hence it mostly does not
matter whether it is installed in the build or host environment. The exception
is the pythran-config tool - only use that if pythran is installed in
the host architecture (i.e., the architecture on which the produced binaries
need to run).
Instead of creating functions that can be used from Python code, Pythran can
produce native functions to be used by other Python extension, using the
Capsule mechanism. To do so, just add the capsule keyword to the export
line:
#pythran export capsule foo(double*, double)
Note that pointer types are only supported within the context of a capsule, as they don’t match any real Python type. Any Pythran type is valid as capsule parameter, but beware that non scalar or pointer types only make sense withing the Pythran context.
Pythran honors the NDEBUG macro. If set through -DNDEBUG (which should
be the default, check python-config --cflags), it disables all assert
statement and doesn’t perform any runtime check for indexing bounds etc.
However, if unset through -UNDEBUG, all assert are executed and
eventually raise an AssertionError. Additionnaly, many internal checks are
done and may fail with a C-ish assertion.
By default Pythran does not generate thread-safe code for non-OpenMP code: reference
counting for automatic deletion of objects is not done atomically by default. It’s
still possible to force pythran to generate thread-safe reference counting by defining
the flag THREAD_SAFE_REF_COUNT via -DTHREAD_SAFE_REF_COUNT. There is a small
performance penalty associated with this.
One can use -o <filename> or --output=<filename> to control the name of
the generated file. If <filename> contains the %{ext} pattern, it is
replaced by the extension that matches your current platform.
A failing compilation? A lust for c++ tangled code? Give a try to the -E
switch that stops the compilation process right after c++ code generation, so
that you can inspect it.
Want more performance? Big fan of -Ofast -march=native? Pythran
_automagically_ forwards these switches to the underlying compiler!
Tired of typing the same compiler switches again and again? Store them in
$XDG_CONFIG_HOME/.pythranrc!
Wants to try your own compiler? Update the CC and CXX fields from your
pythranrc, or set the same environment variables to the right compilers. Environment
variables have greater precedence than configuration file.
Pythran also honors the CXXFLAGS and LDFLAGS environment variables.
The careful reader might have noticed the -p flag from the command line. It
makes it possible to define your own optimization sequence:
pythran -pConstantFolding -pmy_package.MyOptimization
runs the ConstantFolding optimization from pythran.optimizations
followed by a custom optimization found in the my_package package, loaded
from PYTHONPATH.
When importing a Python module, one can check for the presence of the
__pythran__ variable at the module scope to see if the module has been
pythranized:
import foo
if hasattr(foo, '__pythran__'):
print(r'\_o<')
This variable is a tuple that holds three fields:
pythran’s version
compilation date
sha256 value of the input code
OpenMP is a standard set of directives for C, C++ and FORTRAN that makes it easier to turn a sequential program into a multi-threaded one. Pythran translates OpenMP-like code annotation into OpenMP directives:
r=0
#omp parallel for reduction(+:r)
for x,y in zip(l1,l2):
r+=x*y
OpenMP directive parsing is enabled by -fopenmp when using g++ as the
back-end compiler. Be careful with the indentation. It has to be correct!
Alternatively, one can run the great:
$> pythran -ppythran.analyses.ParallelMaps -e as.py
which runs a code analyzer that displays extra information concerning parallel map found in the code.
Pythran can be used to generate raw templated C++ code, without any Python
glue. To do so use the -e switch. It will turn the Python code into C++
code you can call from a C++ program. In that case there is no need for a
particular Pythran specification.
Curious Python developers might want to study how Pythran transforms their
codes. With the -P switch, Pythran optimizes the Python code, prints the
result and stops there. Pythran does not care about PEP 8, so a Python
formatter is often useful:
$> pythran -P arc_distance.py | yapf
.pythranrc¶Pythran checks for a file named .pythranrc and use it to replace the site
configuration. Here are a few tricks!
You can change the default location of the pythran configuration file using the
environment variable PYTHRANRC:
PYTHRANRC=/opt/company/pythran/config.pythranrc pythran arc_distance.py
When PYTHRANRC is set to the empty string, no user-site configuration is
loaded. This can be helpful for reproducible builds.
All the options in the .pythranrc file can be specified when running pythran by using the command line argument –config= .
For example:
pythran --config compiler.blas=pythran-openblas this_file.py
would specify that pythran-openblas is the blas library to use.
Options specified using command-line arguments override the options found in the .pythranrc file
[compiler]¶This section contains compiler flags configuration. For education purpose, the default linux configuration is
[compiler]
defines=
undefs=
include_dirs=
libs=
library_dirs=
cflags=-std=c++11 -fno-math-errno -fvisibility=hidden -fno-wrapv -Wno-unused-function -Wno-int-in-bool-context -Wno-unknown-warning-option
ldflags=-fvisibility=hidden -Wl,-strip-all
blas=blas
CC=
CXX=
ignoreflags=-Wstrict-prototypes
CC:Path to the C compiler to use
CXX:Path to the C++ compiler to use
defines:Preprocessor definitions. Pythran is sensible to USE_XSIMD. It turns on
xsimd vectorization (instead of the
scalar code in xsimd that is used by default).
undefs:Some preprocessor definitions to remove.
include_dirs:Additional include directories to search for headers.
cflags:Additional random compiler flags (-f, -O). Optimization flags generally
go there. The default is to set -std=c++11 for C++11 support.
libs:Libraries to use during the link process. A typical extension
is to add tcmalloc_minimal to use the allocator from
https://code.google.com/p/gperftools/.
library_dirs:Extra directories to search for required libraries.
ldflags:Additional random linker flags.
blas:BLAS library to use. none, pythran-openblas, blas,
openblas, atlas or mkl are viable choices.
none prevents from linking with blas.
pythran-openblas requires the pythran-openblas package, which provides a
statically linked version of OpenBLAS. Other
options are system dependant. Depending on your setup, you may need to
update include_dirs to point to the location of the BLAS headers, e.g.
/usr/include/openblas.
ignoreflags:Space-separated list of compiler flags that should not be forwarded to the
pythran backend compiler when inherited, for instance, from
python-config. For instance -Wstrict-prototypes is a C-only option
that should be pruned.
[pythran]¶This one contains internal configuration settings. Play with it at your own risk!
optimizations:A list of import paths pointing to transformation classes. This contains the optimization pipeline of Pythran! If you design your own optimizations, register them here!
complex_hook:Set this to True for faster and still Numpy-compliant complex
multiplications. Not very portable, but generally works on Linux.
[typing]¶Another internal setting stuff. This controls the accuracy of the typing phase. An extract from the default setting file should convince you not to touch it:
[typing]
# maximum number of combiner per user function
# increasing this value inreases typing accuracy
# but slows down compilation time, to the point of making g++ crash
max_combiner = 2
# above this number of overloads, pythran specifications are considered invalid
# as it generates ultra-large binaries
max_export_overloads = 128
# to the notable exceptions of tuple, pythran sequences are homogeneous. It is
# however possible to store functions objects in a sequence and a variant
# functor is created. This value bounds the number of different types within
# a sequence
max_heterogeneous_sequence_size = 16
[backend]¶This controls some behavior of the C++ backend, so the default should be safe:
# set to true if you want intermediate C++ code to be annotated with a reference
# to the original python code
annotate = false
Supported compiler versions:
g++ version 4.9 and above
clang++ version 3.5 and above
Plenty of them! Seriously, Pythran is software, so it will crash. You
may make it abort in unusual ways! And more importantly, please provide
feedback to serge_sans_paille using its email serge.guelton@telecom-bretagne.eu,
the IRC channel #pythran on OFTC, or the mailing list pythran@freelists.org
glhf!