prange generates unnecessary python interactions when indexing a 2d view

[NicolasHug]

Indexing a 2d view in a prange will generate some unnecessary python interactions related to the GIL.

Here is a reproducing example. LMK if I can help with anything.

#cython: boundscheck=False
#cython: language_level=3
from cython.parallel import prange

# Indexing a 2d view in a prange loop: generates lots of python interactions.
def sum_2d_a(float[:, :] view_2d):

    cdef:
        float out = 0
        int row_idx = 0

    for row_idx in prange(view_2d.shape[0], nogil=True):
        out += sum_1d_a(view_2d[row_idx, :])
        # lots of python interactions at the end of the loop

    return out

cdef float sum_1d_a(float[:] view_1d) nogil:
    return 3.4  # irrelevant code


# workaround: pass the whole 2d view and the row index. No interactions in this case.
def sum_2d_b(float[:, :] view_2d):

    cdef:
        float out = 0
        int row_idx = 0

    for row_idx in prange(view_2d.shape[0], nogil=True):
        out += sum_1d_b(view_2d, row_idx)
        # no python interaction

    return out

cdef float sum_1d_b(float[:, :] view_2d, int row_idx) nogil:
    return 3.4  # irrelevant code

version : 0.29.6

[robertwb]

Creation and (last) deletion of a slice require the GIL (to prevent the underlying memoryview from being deleted before the slice). If not done internally, this happens in the __PYX_{INC,DEC}_MEMVIEW utilities. Even if we had stronger analysis, we'd have to know that sum_1d_b doesn't cache it away somewhere, so it's hard to see how we'd do better here.

[NicolasHug]

Thanks for your feedback.

I'm a bit confused about memory views needing the GIL, because changing prange to range will remove the python interactions. Even in a with nogil: context manager.