Cython backend

Cython code is much more complicated than Pythran code… We won’t be able to support all Cython features!

However, a descent set of Cython can be supported. We need to find Python syntaxes for the most useful Cython special syntaxes.

Note that some Cython features are useless in Pythran (for example cdef of local variables, or nogil).

Note that ideally, we want to write Cython code that can be executed without the Cython package and without Cython compilation. It is possible with the “pure Python mode” of Cython. Therefore, we first need examples of Cython code written in this mode.

Note however, that this mode is currently still experimental and that we hit simple Cython bugs which limit a lot what can be done in practice with the Cython backend. For example:

More generally, there are many known bugs in Cython which do not help! For example:

I think at least some of these bugs have to be solved upstream…

Cython syntaxes already supported

cpdef signature with simple (basic and array) types for arguments

cdef for type declaration of local variables: cython.locals

In “pure Python mode”, one can write

@cython.locals(result=np.float64_t, i=cython.int, n=cython.int)
cpdef mysum(np.float64_t[:] arr_input)

With variables annotations (which are removed for Pythran / Numba):

from transonic import boost

@boost
def mysum(arr_input: "float[]"):
    i: int
    n: int = arr_input.shape[0]
    result: float = 0.
    for i in range(n):
        result += arr_input[i]
    return result

Cython decorators

@cython.boundscheck(False)
@cython.wraparound(False)

Currently only for simple functions (no methods).

Cython syntaxes partly supported

Function definition (cdef, cpdef, inline, nogil, return type)

from transonic import boost

@boost(inline=True, nogil=True)
def func(a: "float[]", n: int) -> "void":
    ...

which would translate in Cython as something like:

cpdef inline void func(np.ndarray[np.float_t, ndim=1] a, cython.int n) nogil

  • all function signatures use cpdef (?)

  • boost(inline=True) is supported for functions, see this example.

  • Return type is supported and there is a void type ("void" or np.void).

Fused types

We already have fused types in Transonic. With Transonic, we can already do:

import numpy as np
from transonic import Array, Type, NDim

np_floats = Type(np.float32, np.float64)
N = NDim(2, 3, 4)
A = Array[np_floats, N]
A1 = Array[np_floats, N + 1]
# or simply
A3d = Array[np_floats, "3d"]

However, Cython Fused types are currently very limited.

Even with something as simple as that

from transonic import Array, Type

A = Array[Type(np.float64, np.complex128), "1d"]

def mysum(arr: A):
    result: A.dtype = arr.dtype.type(0.)
    i: int
    for i in range(arr.shape[0]):
        result += arr[i]
    return result

should be translated to this (not supported, see https://github.com/cython/cython/issues/754) Cython code:

import cython

import numpy as np
cimport numpy as np

ctypedef fused T0:
   np.complex128_t
   np.float64_t

ctypedef np.ndarray[T0, ndim=1] A

def mysum(A arr):
    cdef T0 ret = arr.dtype.type(0.)
    cdef cython.int i
    for i in range(arr.shape[0]):
        ret += arr[i]
    return ret

Note that it works with a memoryview… (but not in pure-Python mode!)

Note that another working alternative is:

import cython

import numpy as np
cimport numpy as np

ctypedef fused T0:
   np.complex128_t
   np.float64_t

def mysum(np.ndarray[T0, ndim=1] arr):
    cdef T0 ret = arr.dtype.type(0.)
    cdef cython.int i
    for i in range(arr.shape[0]):
        ret += arr[i]
    return ret

But the corresponding pure-Python version does not work!

More array types (contiguous arrays, C or F order, memoryviews)

I think we should support:

Array[int, NDim(3), "C"]
Array[int, "3d", "C"]
transonic.typeof(np.empty((2, 2, 2)))

Array["int[:, :, ::1]"]
Array[int, "[:, :, ::1]"]
transonic.str2type("int[:, :, ::1]")

and maybe also:

transonic.int64[:, :, ::1]

I tend to think that the default ("int[:,:]") should correspond to "order=C". “Fortran” order and “any” order (contiguous C or F) could be obtained with "order=F" and "order=any".

Strided arrays could be obtained with Array[int, NDim(3), "strided"] or str2type("int[::, ::, ::]").

Note that for Pythran, we could also support:

A_fixed_dim = Array[Type(np.float32, float), "[:, :, 3]"]

For Cython, we need to be able to specify if an array is a np.ndarray or a memoryview. By default, we will use np.ndarray and memoryview could be obtained with:

Array[int, "[:, :, ::1]", "memview"]

Special C types

For example Py_ssize_t (nearly np.intp, which is supported) and void (supported)

from ctypes import c_ssize_t as Py_ssize_t
from transonic import boost

@boost
def func(arr: "float[]", index: Py_ssize_t):
    if n > 1:
        a[n-1] = 0

or just:

from transonic import boost

@boost
def func(arr: "float[]", index: "Py_ssize_t"):
    ...

Cython syntaxes that can be supported quite easily

with nogil:

We could support something like

from transonic import boost, nogil

@boost
def func(n: int):

    with nogil:
        result = n**2

    return result

Of course there is no equivalent in Pythran, so the Pythran backend would have to suppress the with nogil().

Cast

return <DTYPE_t*> myvar

I guess we should follow Cython and its pure Python mode function cython.cast(type, myvar).

Cython syntaxes that will be difficult to support

Pointers and addresses

cdef Py_ssize_t *p_indexer

Definition struct, enum, class

cdef struct Heap:
    Py_ssize_t items
    Py_ssize_t space
    Heapitem *data
    Heapitem **ptrs

or

cdef class Foo:

C allocation

features_carr = <MBLBP*>malloc(features_number * sizeof(MBLBP))