Thinking about OpenMP.

[rainwoodman]

The old openmp-like support was dropped in the single-file rewrite. Time to think about this more carefully.

• Is this feasible? Can this be more 'omp' like if we play with bytecodes? Anything useful from https://github.com/snoack/python-goto ?

@omp.parallel(private=['a', 'b'], shared=['b'], reduce=[operator.add, 'c'])
def parallel_section(a, b, c):
    if omp.master:
        b[...] = 0
        c[...] = 0
    omp.barrier()
    for i in omp.range(0, 1000, schedule='static'):
        c[...] += i
        a[...] += i
        with omp.critical:
             b[...] += i

a = numpy.empty((,), dtype='i8')
b = numpy.empty((,), dtype='i8')
c = numpy.empty((,), dtype='i8')

parallel_section(1, b, c)

assert c == (0 + 999)*1000 / 2
assert b == (0 + 999)*1000 / 2

• How to handle exceptions? ProcessGroup can probably do this well enough.

[rainwoodman]

We need a way to ensure MapReduce dispatches exactly one payload per process. An ordered at the begining of the work function should be able to achieve this.