Fused Types (Templates)

Note

This page uses two different syntax variants:

  • Cython specific cdef syntax, which was designed to make type declarations concise and easily readable from a C/C++ perspective.

  • Pure Python syntax which allows static Cython type declarations in pure Python code, following PEP-484 type hints and PEP 526 variable annotations.

    To make use of C data types in Python syntax, you need to import the special cython module in the Python module that you want to compile, e.g.

    import cython
    

    If you use the pure Python syntax we strongly recommend you use a recent Cython 3 release, since significant improvements have been made here compared to the 0.29.x releases.

Fused types allow you to have one type definition that can refer to multiple types. This allows you to write a single static-typed cython algorithm that can operate on values of multiple types. Thus fused types allow generic programming and are akin to templates in C++ or generics in languages like Java / C#.

Note

Fused types are not currently supported as attributes of extension types. Only variables and function/method arguments can be declared with fused types.

Quickstart

char_or_float = cython.fused_type(cython.char, cython.double)



@cython.ccall
def plus_one(var: char_or_float) -> char_or_float:
    return var + 1


def show_me():

    a: cython.char = 127
    b: cython.double = 127
    print('char', plus_one(a))
    print('float', plus_one(b))

This gives:

>>> show_me()
char -128
float 128.0

plus_one(a) “specializes” the fused type char_or_float as a char, whereas plus_one(b) specializes char_or_float as a float.

Declaring Fused Types

Fused types may be declared as follows:

my_fused_type = cython.fused_type(cython.int, cython.float)

This declares a new type called my_fused_type which can be either an int or a double.

Only names may be used for the constituent types, but they may be any (non-fused) type, including a typedef. I.e. one may write:

my_double = cython.typedef(cython.double)
my_fused_type = cython.fused_type(cython.int, my_double)

Using Fused Types

Fused types can be used to declare parameters of functions or methods:

@cython.cfunc
def cfunc(arg: my_fused_type):
    return arg + 1

If the same fused type appears more than once in the function arguments, then they will all have the same specialised type:

@cython.cfunc
def cfunc(arg1: my_fused_type, arg2: my_fused_type):
    # arg1 and arg2 always have the same type here
    return arg1 + arg2

Here, the type of both parameters is either an int, or a double (according to the previous examples), because they use the same fused type name my_fused_type. Mixing different fused types (or differently named fused types) in the arguments will specialise them independently:

def func(x: A, y: B):
    ...

This will result in specialized code paths for all combinations of types contained in A and B, e.g.:

my_fused_type = cython.fused_type(cython.int, cython.double)



my_fused_type2 = cython.fused_type(cython.int, cython.double)


@cython.cfunc
def func(a: my_fused_type, b: my_fused_type2):
    # a and b may have the same or different types here
    print("SAME!" if my_fused_type is my_fused_type2 else "NOT SAME!")
    return a + b

Note

A simple typedef to rename the fused type does not currently work here. See Github issue #4302.

Fused types and arrays

Note that specializations of only numeric types may not be very useful, as one can usually rely on promotion of types. This is not true for arrays, pointers and typed views of memory however. Indeed, one may write:

@cython.cfunc
def myfunc(x: A[:, :]):
    ...

# and

@cython.cfunc
cdef otherfunc(x: cython.pointer(A)):
    ...

Following code snippet shows an example with pointer to the fused type:

my_fused_type = cython.fused_type(cython.int, cython.float)


@cython.cfunc
def func(a: cython.pointer(my_fused_type)):
    print(a[0])

def main():
    a: cython.int = 3
    b: cython.float = 5.0

    func(cython.address(a))
    func(cython.address(b))

Note

In Cython 0.20.x and earlier, the compiler generated the full cross product of all type combinations when a fused type was used by more than one memory view in a type signature, e.g.

def myfunc(A[:] a, A[:] b):
    # a and b had independent item types in Cython 0.20.x and earlier.
    ...

This was unexpected for most users, unlikely to be desired, and also inconsistent with other structured type declarations like C arrays of fused types, which were considered the same type. It was thus changed in Cython 0.21 to use the same type for all memory views of a fused type. In order to get the original behaviour, it suffices to declare the same fused type under different names, and then use these in the declarations:

ctypedef fused A:
    int
    long

ctypedef fused B:
    int
    long

def myfunc(A[:] a, B[:] b):
    # a and b are independent types here and may have different item types
    ...

To get only identical types also in older Cython versions (pre-0.21), a ctypedef can be used:

ctypedef A[:] A_1d

def myfunc(A_1d a, A_1d b):
    # a and b have identical item types here, also in older Cython versions
    ...

Selecting Specializations

You can select a specialization (an instance of the function with specific or specialized (i.e., non-fused) argument types) in two ways: either by indexing or by calling.

Indexing

You can index functions with types to get certain specializations, i.e.:

indexing.py
import cython

fused_type1 = cython.fused_type(cython.double, cython.float)



fused_type2 = cython.fused_type(cython.double, cython.float)


@cython.cfunc
def cfunc(arg1: fused_type1, arg2: fused_type1):
    print("cfunc called:", cython.typeof(arg1), arg1, cython.typeof(arg2), arg2)

@cython.ccall
def cpfunc(a: fused_type1, b: fused_type2):
    print("cpfunc called:", cython.typeof(a), a, cython.typeof(b), b)

def func(a: fused_type1, b: fused_type2):
    print("func called:", cython.typeof(a), a, cython.typeof(b), b)

# called from Cython space
cfunc[cython.double](5.0, 1.0)
cpfunc[cython.float, cython.double](1.0, 2.0)
# Indexing def functions in Cython code requires string names
func["float", "double"](1.0, 2.0)

Indexed functions can be called directly from Python:

>>> import cython
>>> import indexing
cfunc called: double 5.0 double 1.0
cpfunc called: float 1.0 double 2.0
func called: float 1.0 double 2.0
>>> indexing.cpfunc[cython.float, cython.float](1, 2)
cpfunc called: float 1.0 float 2.0
>>> indexing.func[cython.float, cython.float](1, 2)
func called: float 1.0 float 2.0

If a fused type is used as a component of a more complex type (for example a pointer to a fused type, or a memoryview of a fused type), then you should index the function with the individual component and not the full argument type:

@cython.cfunc
def myfunc(x: cython.pointer(A)):
    ...

# Specialize using int, not int *
myfunc[cython.int](myint)

For memoryview indexing from python space we can do the following:

my_fused_type = cython.fused_type(cython.int[:, ::1], cython.float[:, ::1])

def func(array: my_fused_type):
    print("func called:", cython.typeof(array))

my_fused_type[cython.int[:, ::1]](myarray)

The same goes for when using e.g. cython.numeric[:, :].

Calling

A fused function can also be called with arguments, where the dispatch is figured out automatically:

def main():
    p1: cython.double = 1.0
    p2: cython.float = 2.0
    cfunc(p1, p1)          # prints "cfunc called: double 1.0 double 1.0"
    cpfunc(p1, p2)         # prints "cpfunc called: double 1.0 float 2.0"

For a cdef or cpdef function called from Cython this means that the specialization is figured out at compile time. For def functions the arguments are typechecked at runtime, and a best-effort approach is performed to figure out which specialization is needed. This means that this may result in a runtime TypeError if no specialization was found. A cpdef function is treated the same way as a def function if the type of the function is unknown (e.g. if it is external and there is no cimport for it).

The automatic dispatching rules are typically as follows, in order of preference:

  • try to find an exact match

  • choose the biggest corresponding numerical type (biggest float, biggest complex, biggest int)

Built-in Fused Types

There are some built-in fused types available for convenience, these are:

cython.integral # short, int, long
cython.floating # float, double
cython.numeric  # short, int, long, float, double, float complex, double complex

Casting Fused Functions

Note

Pointers to functions are currently not supported by pure Python mode. (GitHub issue #4279)

Fused cdef and cpdef functions may be cast or assigned to C function pointers as follows:

cdef myfunc(cython.floating, cython.integral):
    ...

# assign directly
cdef object (*funcp)(float, int)
funcp = myfunc
funcp(f, i)

# alternatively, cast it
(<object (*)(float, int)> myfunc)(f, i)

# This is also valid
funcp = myfunc[float, int]
funcp(f, i)

Type Checking Specializations

Decisions can be made based on the specializations of the fused parameters. False conditions are pruned to avoid invalid code. One may check with is, is not and == and != to see if a fused type is equal to a certain other non-fused type (to check the specialization), or use in and not in to figure out whether a specialization is part of another set of types (specified as a fused type). In example:

bunch_of_types = cython.fused_type(bytes, cython.int, cython.float)






string_t = cython.fused_type(cython.p_char, bytes, unicode)



@cython.cfunc
def myfunc(i: cython.integral, s: bunch_of_types) -> cython.integral:
    # Only one of these branches will be compiled for each specialization!
    if cython.integral is int:
        print('i is an int')
    elif cython.integral is long:
        print('i is a long')
    else:
        print('i is a short')

    if bunch_of_types in string_t:
        print("s is a string!")
    return i * 2

myfunc(cython.cast(cython.int, 5), b'm')  # will print "i is an int" and "s is a string"
myfunc(cython.cast(cython.long, 5), 3)    # will print "i is a long"
myfunc(cython.cast(cython.short, 5), 3)   # will print "i is a short"

Conditional GIL Acquiring / Releasing

Acquiring and releasing the GIL can be controlled by a condition which is known at compile time (see Conditional Acquiring / Releasing the GIL).

This is most useful when combined with fused types. A fused type function may have to handle both cython native types (e.g. cython.int or cython.double) and python types (e.g. object or bytes). Conditional Acquiring / Releasing the GIL provides a method for running the same piece of code either with the GIL released (for cython native types) and with the GIL held (for python types):

import cython

double_or_object = cython.fused_type(cython.double, object)

def increment(x: double_or_object):
    with cython.nogil(double_or_object is not object):
        # Same code handles both cython.double (GIL is released)
        # and python object (GIL is not released).
        x = x + 1
    return x

increment(5.0)  # GIL is released during increment
increment(5)    # GIL is acquired during increment

__signatures__

Finally, function objects from def or cpdef functions have an attribute __signatures__, which maps the signature strings to the actual specialized functions. This may be useful for inspection:

indexing.py
import cython

fused_type1 = cython.fused_type(cython.double, cython.float)



fused_type2 = cython.fused_type(cython.double, cython.float)


@cython.ccall
def cpfunc(a: fused_type1, b: fused_type2):
    print("cpfunc called:", cython.typeof(a), a, cython.typeof(b), b)
>>> from indexing import cpfunc
>>> cpfunc.__signatures__,
({'double|double': <cyfunction __pyx_fuse_0_0cpfunc at 0x107292f20>, 'double|float': <cyfunction __pyx_fuse_0_1cpfunc at 0x1072a6040>, 'float|double': <cyfunction __pyx_fuse_1_0cpfunc at 0x1072a6120>, 'float|float': <cyfunction __pyx_fuse_1_1cpfunc at 0x1072a6200>},)

Listed signature strings may also be used as indices to the fused function, but the index format may change between Cython versions

>>> specialized_function = cpfunc["double|float"]
>>> specialized_function(5.0, 1.0)
cpfunc called: double 5.0 float 1.0

However, the better way how to index is by providing list of types as mentioned in Indexing section.