Parallel prange fails on higher number of iterations and is slower

[step135]

from numba import jit, prange
@jit(nopython=True, parallel=True)
def parorig(p):
    o=[]
    for item in prange(p):
        if item%2==0:o.append(2)
        else: o.append(1)
    return o

It is 10 times slower and unstable than if we remove parallel=True.
It works for small numbers, but calling parorig(10000) fails with "The kernel appears to have died. It will restart automatically.".

[farahhariri]

Thanks for reporting the issue @step135 ! I have tried it on my linux machine and I am able to trigger an error for a large input p. I am getting:

malloc(): unsorted double linked list corrupted
Aborted (core dumped)

However, when I remove the parallel=True everything goes smoothly even for larger inputs. What machine are you running on? and what versions of python and numba?

[step135]

It was tested in Jupyter notebook on Lenovo tablet with graphics card from Intel. Anyway, I suppose that GPU acceleration is not supported in numba for Intel? Alternatively, it could cause some problems.

[farahhariri]

So after some debugging, it seems like this is an issue due to Numba not having support for parallelism with lists. Basically as of now you can't use a list with prange.

Suggestion to numba devs: to mark this issue as a bug/feature request as you see fit and open an issue to update the docs indicating that prange is not supported with lists

[sklam]

marking this as a segfault and a duplicate of #2408

[stuartarchibald]

Docs update in #6918

[stuartarchibald]

Closing as duplicate of #2408.