Cython calling lapack, error: `Cannot take address of Python variable'

Question

My problem is very similar to that posed here, and so I am trying to implement what was proposed in the answer there, based on this example of calling blas bundled with python. My code is only a slightly modified version of that example.pyx, here's my relevant code (with file f2pyptr.h unchanged) :

#example.pyx (single precision linear system solver via lapack)
import numpy as np
import scipy.linalg.lapack
import cython
cimport numpy as np

cdef extern from "/(pwd)/f2pyptr.h":
void *f2py_pointer(object) except NULL

ctypedef int sgesv_t(
int *N, int *NRHS,
float *A, int* LDA,
int* IPIV,
float *B,
int *LDB,
int *INFO)

cdef sgesv_t *sgesv = <sgesv_t*>f2py_pointer(scipy.linalg.lapack.sgesv._cpointer)

def myfunc_sgesv():
  cdef int N, NRHS, LDA, LDB, INFO
  N = 10
  NRHS = 1
  LDA = 10
  LDB = 10
  cdef np.ndarray[np.float32_t, ndim=2] A = np.ones((N,N), np.float32, order = "F")
  cdef np.ndarray[np.float32_t, ndim=2] B = np.ones((N, NRHS),np.float32, order = "F")
  cdef np.ndarray[np.int_t, ndim=1] IPIV = np.empty((N,), dtype = np.int)
  sgesv(&N, &NRHS, &A[0,0], &LDA, &IPIV, &B[0,0], &LDB, &INFO)*

def myfunc_sgesv():
  cdef int N, NRHS, LDA, LDB, INFO
  N = 10
  NRHS = 1
  LDA = 10
  LDB = 10

  cdef np.ndarray[np.float32_t, ndim=2] A = np.ones((N,N), np.float32, order = "F")
  cdef np.ndarray[np.float32_t, ndim=2] B = np.ones((N, NRHS),np.float32, order = "F")
  cdef np.ndarray[np.int_t, ndim=1] IPIV = np.empty((N,), dtype = np.int)

  sgesv(&N, &NRHS, &A[0,0], &LDA, &IPIV, &B[0,0], &LDB, &INFO)

With the standard setup.py file, I call python setup.py build_ext --inplace

which results in

Compiling example.pyx because it changed.
Cythonizing example.pyx

Error compiling Cython file:
------------------------------------------------------------
...

    cdef np.ndarray[np.float32_t, ndim=2] A = np.ones((N,N), np.float32, order = "F")
    cdef np.ndarray[np.float32_t, ndim=2] B = np.ones((N, NRHS),np.float32, order = "F")
    cdef np.ndarray[np.int_t, ndim=1] IPIV = np.empty((N,), dtype = np.int)

    sgesv(&N, &NRHS, &A[0,0], &LDA, &IPIV, &B[0,0], &LDB, &INFO)
                            ^
------------------------------------------------------------

example.pyx:103:33: Cannot take address of Python variable
Traceback (most recent call last):
File "setup.py", line 5, in <module>
ext_modules = cythonize("example.pyx")
File "/idiap/home/jnewling/.local/lib/python2.7/site-packages/Cython-0.20.1/Cython/Buil/Dependencies.py", line 785, in cythonize
cythonize_one(*args[1:]) File "/idiap/home/jnewling/.local/lib/python2.7/site-packages/Cython-0.20.1/Cython/Buil/Dependencies.py", line 902, in cythonize_one raise CompileError(None, pyx_file) Cython.Compiler.Errors.CompileError: example.pyx

I have the original example.pyx running fine, as well as a modified single precision version (i.e. sgemm), and so cannot see why my sgesv implementation fails.

Any specific diagnosis or suggestions of alternative approaches would be most welcome.

Answer 1

Not sure if this answer is correct, but I will tell you what I tried as it may help you:

I got the error Cannot take address of Python variable at a different place when I compiled your code. I got the error at & IPIV. I then noticed it is an array so I changed it to:

sgesv(&N, &NRHS, &A[0,0], &LDA, &IPIV[0], &B[0,0], &LDB, &INFO). I then got the error Cannot assign type 'int_t *' to 'int *' . Then found the following answer: Passing numpy integer array to c code and changed the code to:

cdef np.ndarray[int, ndim=1, mode='c'] IPIV = np.empty((N,), dtype = ctypes.c_int) sgesv(&N, &NRHS, &A[0,0], &LDA, &IPIV[0], &B[0,0], &LDB, &INFO)

and now it compiles without any errors.