pythonpython-3.xcythoncythonize

Cython compilation error "Not allowed in a constant expression"


The following

cimport cython

@cython.boundscheck(False)
def boundtest():
     cdef int r=4
     cdef double l[3]

works fine. But when I try this:

cimport cython

@cython.boundscheck(False)
def boundtest():
     cdef int r=4
     cdef double l[r]

I receive the following error:

[1/1] Cythonizing test.pyx

Error compiling Cython file:
------------------------------------------------------------
...
cimport cython

@cython.boundscheck(False)
def boundtest():
     cdef int r=4
     cdef double l[r]
                   ^
------------------------------------------------------------

test.pyx:13:20: Not allowed in a constant expression
     

The decorator was added due to finding this related stackexchange post and reading the Cython book by Kurt W. Smith. As far as I can tell this should work to tell Cython not to worry about out of bounds errors that may result from having a dynamic indexing variable but for some reason it does not. I have also tried changing boundscheck in the compiler options and globally to no avail.

If it weren't for the Cython documentation claiming to be up to date I would think boundscheck has been depreciated.


Solution

  • The failure has nothing to do with cython.boundscheck.

    Boundchecking is just checking if you try to access an element of an array that isn't there. For example if you have a size 4 array and try to access element 5 - with boundscheck(True) it will give you an Exception, with boundscheck(False) it will result in undefined behavior (probably causing a segmentation fault).

    The reason for the compilation failure is another: You can't create a static array with a dynamic length! The number of elements need to be known at compile time, that's just something enforces (I guess).

    However you could define r to be known at compile time:

    DEF r=4
    
    cimport cython
    
    @cython.boundscheck(False)
    def boundtest():
        cdef double l[r]
    

    You could however simply create a NumPy array and store it in a memoryview variable:

    cimport cython
    import numpy as np
    
    @cython.boundscheck(False)
    def boundtest():
        cdef int r=4
        cdef double[:] l = np.empty(r, dtype=np.double)