Shared address space in Cython - openmp

Question

I have the following Cython code:
from cython import parallel
from libc.stdio cimport printf
cdef extern from "unistd.h" nogil:
int usleep(int);
def test_func():
cdef int var = -1
with nogil, parallel.parallel(num_threads=4):
var = parallel.threadid()
usleep(100000)
printf("Var: %d\n", var)
After compiling and launching in python3 console I get:
>>> import test
>>> test.test_func()
Var: 3
Var: 0
Var: 2
Var: 1
That means, each thread has its own address space.
However, I would like to have C++ OpenMP behavior, duplicating this:
#include "omp.h"
#include <stdio.h>
#include <unistd.h>
int main() {
int v=-1;
omp_set_num_threads(4);
#pragma omp parallel
{
v=omp_get_thread_num();
usleep(100000);
printf("Var: %d\n", v);
}
return(0);
}
which outputs something like this:
Var: 3
Var: 3
Var: 3
Var: 3
Thus, question: is it possible to get shared address space in Cython parallelized blocks?

Answer 1

It's a bit of a hacky way of doing it, but using a pointer to var you look to be able to trick it
def test_func2():
cdef int var = -1
cdef int* var_ptr = &var
with nogil, parallel.parallel(num_threads=4):
var_ptr[0] = parallel.threadid()
usleep(100000)
printf("Var: %d\n", var)
If you look at the generated C code then test_func (your version) gives the line
#pragma omp parallel private(__pyx_v_var) num_threads(4)
while test_func2 gives
#pragma omp parallel num_threads(4)
I think this works because you don't directly assign to it, so it isn't picked up by Cython's normal rules for what to make private. There is a risk here - if a future version of Cython got cleverer and make var_ptr private then it wouldn't be initialised at the start of the parallel section (so be careful and check what it's doing!).