tuple of tuple arguments not allowed in parfor loop

[albop]

The following code

import numpy as np
import numba as nb
from numba import prange

@nb.jit
def my_fun(vals, grid, coef, x):
    return vals[0] + coef[0] + x[0] # + grid[0][0]

@nb.jit(nopython=True, parallel=True)
def run_in_pfor(vals, grid, coef, x):
    t = 0.0
    for h in prange(6):
        res = my_fun(vals, grid, coef, x)
        t += res
    return t

vals = np.random.rand(5)
grid = ((0., 1., 5),)
coef = vals.copy()
x = np.array([0.22])
run_in_pfor(vals, grid, coef, x)

raises an exception ending in

numba.core.errors.LoweringError: Failed in nopython mode pipeline (step: native parfor lowering)
scalar type Tuple(float64, float64, int64) given for non scalar argument #3

File "notuple.py", line 12:
def run_in_pfor(vals, grid, coef, x):
    <source elided>
    t = 0.0
    for h in prange(6):
    ^

During: lowering "id=0[LoopNest(index_variable = parfor_index.6, range = (0, $const20.3, 1))]{40: <ir.Block at /home/pablo/Econforge/interpolation.py/notuple.py (12)>}Var(parfor_index.6, notuple.py:12)" at /home/pablo/Econforge/interpolation.py/notuple.py (12)

full stacktrace attached : stacktrace.txt .

If I replace grid = ((0., 1., 5),) by grid = (0., 1., 5) it works (it is not used by the function anyway).

[albop]

Original discussion here: EconForge/interpolation.py#107 (comment)

[guilhermeleobas]

Hi @albop, thanks for the report. I can confirm the bug in Numba and only works if parallel=True.

I'll put this topic for discussion at the next triage meeting. I'll check if this is a limitation by design, or if it is actually a bug.

[DrTodd13]

Hi @albop, thanks for the report. I can confirm the bug in Numba and only works if parallel=True.

I'll put this topic for discussion at the next triage meeting. I'll check if this is a limitation by design, or if it is actually a bug.

Limitation by design. For a long time, parallel region didn't support tuples at all but eventually we added the ability to support tuples but the implementation is necessarily very ugly. Supporting nested tuples would be ugly squared. There is some talk of replacing the internal gufunc mechanism to not have this tuple restriction at which point this problem will go away.

[albop]

Fair enough. Thanks for the clarification @DrTodd13 . I can certainly find an ugly workaround in the meantime ;-)