Using Parallelism

Cython supports native parallelism through the cython.parallel module. To use this kind of parallelism, the GIL must be released (see Releasing the GIL). It currently supports OpenMP, but later on more backends might be supported.

cython.parallel.prange([start,] stop[, step][, nogil=False][, schedule=None[, chunksize=None]][, num_threads=None])

This function can be used for parallel loops. OpenMP automatically starts a thread pool and distributes the work according to the schedule used. step must not be 0. This function can only be used with the GIL released. If nogil is true, the loop will be wrapped in a nogil section.

Thread-locality and reductions are automatically inferred for variables.

If you assign to a variable in a prange block, it becomes lastprivate, meaning that the variable will contain the value from the last iteration. If you use an inplace operator on a variable, it becomes a reduction, meaning that the values from the thread-local copies of the variable will be reduced with the operator and assigned to the original variable after the loop. The index variable is always lastprivate. Variables assigned to in a parallel with block will be private and unusable after the block, as there is no concept of a sequentially last value.

The schedule is passed to OpenMP and can be one of the following:

static:

If a chunksize is provided, iterations are distributed to all threads ahead of time in blocks of the given chunksize. If no chunksize is given, the iteration space is divided into chunks that are approximately equal in size, and at most one chunk is assigned to each thread in advance.

This is most appropriate when the scheduling overhead matters and the problem can be cut down into equally sized chunks that are known to have approximately the same runtime.

dynamic:

The iterations are distributed to threads as they request them, with a default chunk size of 1.

This is suitable when the runtime of each chunk differs and is not known in advance and therefore a larger number of smaller chunks is used in order to keep all threads busy.

guided:

As with dynamic scheduling, the iterations are distributed to threads as they request them, but with decreasing chunk size. The size of each chunk is proportional to the number of unassigned iterations divided by the number of participating threads, decreasing to 1 (or the chunksize if provided).

This has an advantage over pure dynamic scheduling when it turns out that the last chunks take more time than expected or are otherwise being badly scheduled, so that most threads start running idle while the last chunks are being worked on by only a smaller number of threads.

runtime:

The schedule and chunk size are taken from the runtime scheduling variable, which can be set through the openmp.omp_set_schedule() function call, or the OMP_SCHEDULE environment variable. Note that this essentially disables any static compile time optimisations of the scheduling code itself and may therefore show a slightly worse performance than when the same scheduling policy is statically configured at compile time.

cython.parallel.parallel(num_threads=None)

This directive can be used as part of a with statement to execute code sequences in parallel. This is currently useful to setup thread-local buffers used by a prange. A contained prange will be a worksharing loop that is not parallel, so any variable assigned to in the parallel section is also private to the prange. Variables that are private in the parallel block are unavailable after the parallel block.

Example with thread-local buffers:

from cython.parallel import parallel, prange
from libc.stdlib cimport abort, malloc, free

cdef Py_ssize_t idx, i, n = 100
cdef int * local_buf
cdef size_t size = 10

with nogil, parallel():
    local_buf = <int *> malloc(sizeof(int) * size)
    if local_buf == NULL:
        abort()

    # populate our local buffer in a sequential loop
    for i in xrange(size):
        local_buf[i] = i * 2

    # share the work using the thread-local buffer(s)
    for i in prange(n, schedule='guided'):
        func(local_buf)

    free(local_buf)

Later on sections might be supported in parallel blocks, to distribute code sections of work among threads.

cython.parallel.threadid()

Returns the id of the thread. For n threads, the ids will range from 0 to n-1.

from distutils.core import setup
from distutils.extension import Extension
from Cython.Distutils import build_ext

ext_module = Extension(
    "hello",
    ["hello.pyx"],
    extra_compile_args=['-fopenmp'],
    extra_link_args=['-fopenmp'],
)

setup(
    name = 'Hello world app',
    cmdclass = {'build_ext': build_ext},
    ext_modules = [ext_module],
)

from cython.parallel import prange

cdef int func(Py_ssize_t n):
    cdef Py_ssize_t i

    for i in prange(n, nogil=True):
        if i == 8:
            with gil:
                raise Exception()
        elif i == 4:
            break
        elif i == 2:
            return i

from cython.parallel cimport parallel
cimport openmp

cdef int num_threads

openmp.omp_set_dynamic(1)
with nogil, parallel():
    num_threads = openmp.omp_get_num_threads()
    ...