How does one to return a malloc array pointer (or numpy array pointer) in cython back to python3, efficiently.
The cython code works perfectly as long as I don't return the array pointer
I would like:
def double complex* randn_zig(int n):
...
r = malloc(n*n*sizeof(double complex))
...
return r
The c11 (gcc 11) equivalent is:
double complex* randn_zig(int n){
r = malloc(n*n*sizeof(double complex))
return r
}
I have tried <double complex*> randn_zig(int n):
and randn_zig(<double complex*> r, int n):
and other permutations without success so far. The c and cython code version is 5 times as fast as Numby/ pylab randn version if I can find a way to return a pointer to the large 10^6 to 10^10 double complex array.
Your question is similar to this post .
You can use the function below to pass a C pointer to Numpy array. The memory will be freed automatically when the Numpy array is recycled. If you want free the pointer mamully, you should not set NPY_OWNDATA flag.
import numpy as np
cimport numpy as np
cdef pointer_to_numpy_array_complex128(void * ptr, np.npy_intp size):
'''Convert c pointer to numpy array.
The memory will be freed as soon as the ndarray is deallocated.
'''
cdef extern from "numpy/arrayobject.h":
void PyArray_ENABLEFLAGS(np.ndarray arr, int flags)
cdef np.ndarray[np.complex128, ndim=1] arr = \
np.PyArray_SimpleNewFromData(1, &size, np.NPY_COMPLEX128, ptr)
PyArray_ENABLEFLAGS(arr, np.NPY_OWNDATA)
return arr
For reference:
Of couse, you can also use cython memoryview .
import numpy as np
cimport numpy as np
cdef np.complex128_t[:,:] view = <np.complex128_t[:n,:n]> c_pointer
numpy_arr = np.asarray(view)
The code above will transfer C pointer to a numpy array. However this would not free memory automaticlly, you have to free the memory by yourself or it would lead to memory leak!
A further option (in addition to the two options from the top answer: PyArray_SimpleNewFromData
and just returning the typed memoryview without handling the memory) is to use the cython.view.array
class .
This is a fairly low-level class that can be used to wrap existing memory. It has an attribute callback_free_data
where you can set a function to be called on destruction so that it does free the memory (example code here is copied from the documentation):
cdef view.array my_array = view.array(..., mode="fortran", allocate_buffer=False)
my_array.data = <char *> my_data_pointer
# define a function that can deallocate the data (if needed)
my_array.callback_free_data = free
It exposes the buffer protocol so that you can index it, use it with typed memoryviews, or wrap it with a Numpy array (without copying) with np.asarray
. The latter feature may be easier to use than PyArray_SimpleNewFromData
.
I think the best approach is to pass the pointer of an existing array created in Python via NumPy to Cython, otherwise it seems you have to copy the content of the array created by malloc
to another array, like demonstrated in this toy example:
import numpy as np
cimport numpy as np
from libc.stdlib cimport malloc, free
def main():
cdef int i, n=40
cdef double complex *r
cdef np.ndarray[np.complex128_t, ndim=1] a
a = np.zeros(n*n, dtype=np.complex128)
r = <double complex *>malloc(n*n*sizeof(double complex))
for i in range(n*n):
r[i] = 1.
for i in range(n*n):
a[i] = r[i]
free(r)
return a
For gcc 5+ using the C-11 standard ( gcc -std=gnu11 . . . ), the syntax for multi-dimensional malloc and calloc arrays has changed significantly.
A main() procedure, for creating a 2-D, double, complex calloc array r[n][n] for n = 1024 is now:
long n = 1024;
complex double (*r)[n] = calloc(n, sizeof *r);
An example of a Gaussian random number generator randn_box_muller() using a pointer to this calloc array r[n][n] is:
inline static void randn_box_muller(long n, complex double r[][n])
{
long i, j;
register double x, y;
for(i = 0; i < n; i++){
for(j = 0; j < n; j++){
x = 2.*M_PI*dsfmt_genrand_close_open(&dsfmt);
y = sqrt(-2.*log(dsfmt_genrand_close_open(&dsfmt)));
r[i][j] = (cos(x) + I*sin(x))*y;
}
}
return;
}
This relatively new calloc allocation syntax is a bit strange. It works well for 1, 2 and even n dimensional calloc and malloc arrays. Hopefully this will also work in conjunction with Python3. I hope to be testing this shortly.
The technical post webpages of this site follow the CC BY-SA 4.0 protocol. If you need to reprint, please indicate the site URL or the original address.Any question please contact:yoyou2525@163.com.