Pyccel's Quickstart Guide

Pyccel is a static compiler for Python 3, using Fortran or C as a backend language, with a focus on high-performance computing (HPC) applications.

Pyccel's main goal is to resolve the principal bottleneck in scientific computing: the transition from prototype to production. Programmers usually develop their prototype code in a user-friendly interactive language like Python, but their final application requires an HPC implementation and therefore a new production code. In most cases this is written in a statically compiled language like Fortran/C/C++, and it uses SIMD vectorization, parallel multi-threading, MPI parallelization, GPU offloading, etc.

We believe that this expensive process can be avoided, or at least drastically reduced, by using Pyccel to accelerate the most computationally intensive parts of the Python prototype. Not only is the Pyccel-generated Fortran or C code very fast, but it is human-readable; hence the expert programmer can easily profile the code on the target machine and further optimize it.

Some Useful Background

We recall that Python 3 is an interpreted language, dynamically typed and garbage-collected. In particular, it is worth clarifying the difference between an object and a variable (or name):

Python object

Is created by the Python interpreter when object.__new__() is invoked (e.g. as a result of an expression).
Can be either mutable or immutable, but its type never changes.
Resides in memory and has a reference count.
Is accessed through one or more Python variables.
Is destroyed by the garbage collector when its reference count drops to zero.

Python variable

Is a reference to a Python object in memory (which is the variable's value).
Is created with an assignment operation x = expr, where x is the new variable and expr is any expression:
- x will reference the value of expr, therefore increasing the object's reference count
- If the variable x already existed, the interpreter reduces the reference count of its old value
Can be destroyed with the command del x, which reduces its value's reference count
Since a variable can be reassigned to any object, its type could change at run-time; hence we say that Python is dynamically typed.

Statically Typed Languages

A language is statically typed if the type of a variable is known at compile-time instead of run-time. Common examples of statically typed languages include Java, C, C++, FORTRAN, Pascal and Scala. These languages provide a set of built-in types, and provide the means for creating additional user-defined types. It is the programmer's responsibility to explicitly declare the type of every variable used in the code.

How Pyccel works

In order to translate Python 3 code (dynamically typed) to efficient Fortran/C code (statically typed), Pyccel makes a certain number of assumptions, needs additional information from the user, and imposes a few restrictions on the code. The fundamental rule that guides Pyccel's design is that the Python 3 code and the generated Fortran/C code should behave in the same way. Up to round-off errors, this means that a function must return the same output value if the input arguments are the same. Moreover, Pyccel raises some warnings/errors for the conflicts that can occur in some corner cases.

Type Inference

Pyccel uses type inference to calculate the type of all variables through a static analysis of the Python code. It understands all Python literals, and computes the type of an expression that contains previously processed variables. For example:

x = 3        # int
y = x * 0.5  # float
z = y + 1j   # complex

Assumptions and Restrictions

Because the type of a variable must be unique within a given scope, Pyccel cannot support some of the flexibility that Python provides. The following basic restrictions apply:

The type of a variable cannot be changed
The type of a variable cannot depend on an if condition

For example:

if condition:
    y = 1    # int
    z = 1.0  # float
else:
    y = 2    # int     : OK
    z = 3j   # complex : Pyccel raises an error!

Type Annotations

When parsing a function, Pyccel needs to know the type of the input arguments in order to perform type inference, and ultimately compute the type of the output result. (In the case of recursive functions, the return type should also be declared.) Pyccel can then perform type inference on other Python code that uses that function, because the type of its result will already be known.

The programmer has various ways to provide the argument type information to Pyccel. We recommend using Python-style annotations, which have the syntax:

def fun(arg1: 'type1', arg2: 'type2', ..., argN: 'typeN') -> 'return_type':

or to declare Numpy arrays

def fun(arg1: 'type1[:]', arg2: 'type2[:,:]', ..., argN: 'typeN[dimensions]') -> 'return_type':

The number of dimensions of an array is equal to the number of comma-separated colons in the square brackets. So arr[:] means that the array arr has 1 dimension, arr[:,:] means that it has 2 dimensions and so on. In general string type hints must be used to provide Pyccel with information about the number of dimensions.

For scalar variables and arrays Pyccel supports the following data types:

built-in datatypes: bool, int, float, complex
Numpy integer types: int8, int16, int32, int64
Numpy real types: float32, float64, double
Numpy complex types: complex64, complex128

How to use Pyccel

Installation

Pyccel's official releases can be downloaded from PyPI (the Python Package Index) using pip. To get the latest (trunk) version of Pyccel, one can clone the git repository from GitHub and checkout the master branch. Detailed installation instructions are found in the README file.

Command Line Usage

After installation, the pyccel command will be available on a terminal app (iterm or terminal for MacOs, terminal for Linux). After typing pyccel, the usage should be displayed on the terminal; if this is the case then the installation has succeeded. In essence the pyccel command translates the given Python file to a Fortran or C file, and then compiles the generated code to a Python C extension module or a simple executable.

Example 1: "Hello World" program

In this first example we create the simplest Python script hello.py, which prints Hello, world! followed by an empty line. We use the pyccel command to translate the Python code to a C program in file hello.c, which is placed in the new __pyccel__ directory. By default Pyccel also compiles the C code into an executable named hello, which is placed in the same directory as the original file:

$ printf 'if __name__ == "__main__": \n print("Hello, world!\\n")' > hello.py
$ python3 hello.py
Hello, world!

$ pyccel hello.py --language c
$ ./hello
Hello, world!

Here are the contents of the current directory:

$ tree
.
├── __pyccel__
│   ...
│   ├── prog_hello.c
│   ...
├── hello
└── hello.py

And this is the C file generated by Pyccel:

#include <stdlib.h>
#include <stdio.h>

int main()
{
    printf("%s\n", "Hello, world!\n");
    return 0;
}

Example 2: extension module

If the Python file to be accelerated only contains functions and class definitions, Pyccel will treat it as a Python module instead of just a script. Accordingly, it will not generate a program, but rather a Python C extension module which can be imported from Python.

For example, we now consider the Python module mod.py, which reads

def factorial(n : int) -> int:
    if n <= 0: return 1
    else     : return n * factorial(n - 1)

def binomial_coefficient(n : int, k : int):
    num = factorial(n)
    den = factorial(k) * factorial(n - k)
    return num // den

We use the pyccel command to translate mod.py to the C files mod.c and mod.h, which are placed in the new __pyccel__ directory:

$ pyccel mod.py --language c

By default Pyccel also compiles the C code into a Python C extension module named mod.<TAG>.so, which is placed in the same directory as mod.py. To achieve this Pyccel generates the additional files mod_wrapper.c (which interacts directly with the CPython API) and setup_mod.py (which defines the build procedure for the extension module), as well as a build directory.

If the command import mod is now given to the Python interpreter, this will import the Python C extention module mod.<TAG>.so instead of the pure Python module mod.py.

These are the contents of the current directory:

$ tree
.
├── mod.cpython-36m-x86_64-linux-gnu.so
├── mod.py
└── __pyccel__
    ├── build
    │   ├── lib.linux-x86_64-3.6
    │   └── temp.linux-x86_64-3.6
    │       └── mod_wrapper.o
    ├── mod.c
    ├── mod.h
    ├── mod.o
    ├── mod_wrapper.c
    └── setup_mod.py

4 directories, 10 files

This is the C file mod.c generated by Pyccel:

#include <mod.h>
#include <math.h>
#include <stdlib.h>
#include <stdint.h>

/*........................................*/
int64_t factorial(int64_t n)
{
    int64_t Out_0001;
    if (n <= 0)
    {
        Out_0001 = 1;
        return Out_0001;
    }
    else if (1)
    {
        Out_0001 = n * factorial(n - 1);
        return Out_0001;
    }
}
/*........................................*/

/*........................................*/
int64_t binomial_coefficient(int64_t n, int64_t k)
{
    int64_t num;
    int64_t den;
    int64_t Out_0001;
    num = factorial(n);
    den = factorial(k) * factorial(n - k);
    Out_0001 = floor((double)(num) / (double)(den));
    return Out_0001;
}
/*........................................*/

And this is the header file mod.h:

#ifndef MOD_H
#define MOD_H

#include <stdlib.h>
#include <stdint.h>

int64_t factorial(int64_t n);

int64_t binomial_coefficient(int64_t n, int64_t k);

#endif // MOD_Hy

Example 3: matrix-matrix multiplication with OpenMP

Let's now see a more complicated example, where the Python module mod.py contains a function that performs the matrix-matrix multiplication between two arrays a and b, and writes the result into the array c:

The three function's arguments are 2D Numpy arrays of double-precision floating point numbers
Matrices a and c use C ordering (row-major), matrix b uses Fortran ordering (column-major)
Since matrix c is modified by the function, it has intent(inout) in Fortran
Comments starting with #$ omp are translated to OpenMP pragmas

def matmul(a: 'float[:,:](order=C)',
           b: 'float[:,:](order=F)',
           c: 'float[:,:](order=C)'):

    m, p = a.shape
    q, n = b.shape
    r, s = c.shape

    if p != q or m != r or n != s:
        return -1

#$ omp parallel
#$ omp for schedule(runtime)
    for i in range(m):
        for j in range(n):
            c[i, j] = 0.0
            for k in range(p):
                c[i, j] += a[i, k] * b[k, j]
#$ omp end parallel

    return 0

We now translate this file to Fortran, and compile it to a Python C extension module, using the command

$ pyccel mod.py --language fortran

The flag --language fortran may be omitted, as Pyccel uses Fortran as the default backend language. These are the contents of the current directory:

$ tree .
.
├── mod.cpython-36m-x86_64-linux-gnu.so
├── mod.py
└── __pyccel__
    ├── bind_c_mod.f90
    ├── bind_c_mod.o
    ├── build
    │   ├── lib.linux-x86_64-3.6
    │   └── temp.linux-x86_64-3.6
    │       └── mod_wrapper.o
    ├── mod.f90
    ├── mod.mod
    ├── mod.o
    ├── mod_wrapper.c
    └── setup_mod.py

4 directories, 10 files

And this is the Fortran file mod.f90:

module mod

use, intrinsic :: ISO_C_BINDING

implicit none

contains

!........................................
function matmul(a, b, c) result(Out_0001)

  implicit none

  integer(C_INT64_T) :: Out_0001
  real(C_DOUBLE), intent(in) :: a(0:,0:)
  real(C_DOUBLE), intent(in) :: b(0:,0:)
  real(C_DOUBLE), intent(inout) :: c(0:,0:)
  integer(C_INT64_T) :: m
  integer(C_INT64_T) :: p
  integer(C_INT64_T) :: q
  integer(C_INT64_T) :: n
  integer(C_INT64_T) :: r
  integer(C_INT64_T) :: s
  integer(C_INT64_T) :: i
  integer(C_INT64_T) :: j
  integer(C_INT64_T) :: k

  m = size(a, 2, C_INT64_T)
  p = size(a, 1, C_INT64_T)
  q = size(b, 1, C_INT64_T)
  n = size(b, 2, C_INT64_T)
  r = size(c, 2, C_INT64_T)
  s = size(c, 1, C_INT64_T)
  if (p /= q .or. m /= r .or. n /= s) then
    Out_0001 = -1_C_INT64_T
    return
  end if
  !$omp parallel
  !$omp do schedule(runtime)
  do i = 0_C_INT64_T, m-1_C_INT64_T, 1_C_INT64_T
    do j = 0_C_INT64_T, n-1_C_INT64_T, 1_C_INT64_T
      c(j, i) = 0.0_C_DOUBLE
      do k = 0_C_INT64_T, p-1_C_INT64_T, 1_C_INT64_T
        c(j, i) = c(j, i) + a(k, i) * b(k, j)
      end do
    end do
  end do
  !$omp end parallel
  Out_0001 = 0_C_INT64_T
  return

end function matmul
!........................................

end module mod

Interactive Usage with `epyccel`

In addition to the pyccel command, the Pyccel library provides the epyccel Python function, whose name stands for "embedded Pyccel": given a pure Python function f with type annotations, epyccel returns a "pyccelized" function f_fast that can be used in the same Python session. For example:

from pyccel.epyccel import epyccel
from mod import f

f_fast = epyccel(f)

In practice epyccel copies the contents of f into a temporary python file in the __epyccel__ directory. As a result it is important that all imports are written inside the function when using epyccel. Once the file has been copied, epyccel calls the pyccel command to generate a Python C extension module that contains a single pyccelized function. Then finally, it imports this function and returns it to the caller.

Example 4: quicksort algorithm

Let's assume that we have a quicksort function in a pure Python module mod.py:

def quicksort(a: 'float[:]', lo: int, hi: int):
    i = lo
    j = hi
    while i < hi:
        pivot = a[(lo + hi) // 2]
        while i <= j :
            while a[i] < pivot:
                i += 1
            while a[j] > pivot:
                j -= 1
            if i <= j:
                tmp  = a[i]
                a[i] = a[j]
                a[j] = tmp
                i += 1
                j -= 1
        if lo < j:
            quicksort(a, lo, j)
        lo = i
        j = hi

We now import this function from an interactive IPython terminal and pyccelize it with the epyccel command. We then use the two functions (original and pyccelized) to sort a random array of 100 elements. Finally we compare the timings obtained on an Intel Core 3 architecture.

In [1]: from numpy.random import random
In [2]: from mod import quicksort
In [3]: from pyccel.epyccel import epyccel

In [4]: quicksort_fast = epyccel(quicksort)
In [5]: x = random(100)

In [6]: %timeit y = x.copy()
435 ns ± 4.75 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

In [7]: %timeit y = x.copy(); quicksort(y, 0, 99)
280 µs ± 1.5 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

In [8]: %timeit y = x.copy(); quicksort_fast(y, 0, 99)
1.76 µs ± 10 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

In [9]: (280 - 0.435) / (1.76 - 0.435)
Out[9]: 210.99245283018868

After subtracting the amount of time required to create an array copy from the given times, we can conclude that the pyccelized function is approximately 210 times faster than the original Python function.

Other Features

Pyccel's generated code can use parallel multi-threading through OpenMP; please read our documentation for more details.

We are also working on supporting MPI, LAPACK/BLAS, and OpenACC.

In the future we plan to support GPU programming with CUDA and task-based parallelism.

Getting Help

If you face problems with pyccel, please take the following steps:

Consult our documention in the tutorial directory;
Send an email message to pyccel@googlegroups.com;
Open an issue on GitHub.

Thank you!

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

quickstart.md

quickstart.md

Pyccel's Quickstart Guide

Some Useful Background

Python object

Python variable

Statically Typed Languages

Further Reading

How Pyccel works

Type Inference

Assumptions and Restrictions

Type Annotations

How to use Pyccel

Installation

Command Line Usage

Example 1: "Hello World" program

Example 2: extension module

Example 3: matrix-matrix multiplication with OpenMP

Interactive Usage with `epyccel`

Example 4: quicksort algorithm

Other Features

Getting Help

Files

quickstart.md

Latest commit

History

quickstart.md

File metadata and controls

Pyccel's Quickstart Guide

Some Useful Background

Python object

Python variable

Statically Typed Languages

Further Reading

How Pyccel works

Type Inference

Assumptions and Restrictions

Type Annotations

How to use Pyccel

Installation

Command Line Usage

Example 1: "Hello World" program

Example 2: extension module

Example 3: matrix-matrix multiplication with OpenMP

Interactive Usage with epyccel

Example 4: quicksort algorithm

Other Features

Getting Help

Interactive Usage with `epyccel`