A "well-in-time" compiler, using cython
to compile pyx
modules at runtime.
from witty import compile_module
fancy_module_pyx = """
def add(int a, int b):
return a + b
"""
# equivalent to "import fancy_module"
fancy_module = compile_module(fancy_module_pyx)
result = fancy_module.add(3, 2)
print("fancy_module.add(3, 2) =", result)
This module will no longer be needed if/when cython/cython#555 gets merged into Cython.
Compilation at runtime is very handy to modify C/C++/PYX sources based on type information or configurations that are not known during build time. This allows combining the optimizations of C++ template libraries with Python's runtime flexibility, like so:
import witty
source_pxy_template = """
cdef extern from '<vector>' namespace 'std':
cdef cppclass vector[T]:
vector()
void push_back(T& item)
size_t size()
def to_vector(values):
vec = vector[{type}]()
for x in values:
vec.push_back(x)
return vec
"""
fancy_module = witty.compile_module(
source_pxy_template.format(type="float"), language="c++"
)
# create a C++ vector of floats from a list
vec_float = fancy_module.to_vector([0.1, 0.2, 0.3, 1e10])
print(vec_float)
witty
invokes cython
to compile the module given as a PYX string (just like it would compile it during build time). The compiled module ends up in the Cython cache directory, with a hash build from the content of the PYX string. Repeated calls to compile_module
will only invoke the compiler if the exact PYX string has not been compiled before (or if force_rebuild==True
). Compilation is protected by a file lock, i.e., concurrent calls to compile_module
are safe.