Shows how CUDA memory addressing could look for a block of (dimx, dimy) threads.

cuda.py

from __future__ import print_function
import sys


if len(sys.argv) == 1:
    dimx, dimy = 4, 3
elif len(sys.argv) == 3:
    dimx, dimy = [int(var) for var in sys.argv[1:3]]
else:
    sys.exit('Wrong amount of parameters')


def fill():
    def row(nr):
        return [(nr, i) for i in range(dimy)]
    return [row(i) for i in range(dimx)]


def pprint(array):
    from math import log10, ceil
    length = int(ceil(log10(dimx * dimy)))
    frmt_str = "{}d".format(length)

    def format_fcn(elem):
        return str(elem) if type(elem) is tuple else format(elem, frmt_str)

    for row in array:
        print(" ".join(map(format_fcn, row)))


def convert(array):
    def new_row(row):
        return [x * dimy + y for x, y in row]
    return [new_row(row) for row in array]


def main():
    before = fill()
    pprint(before)
    after = convert(before)
    pprint(after)


if __name__ == '__main__':
    main()

Run with python cuda.py $dimx $dimy or without parameters for the default (4, 3).