<[EMAIL PROTECTED]> wrote: ... > array elements. Powers and subtractions in array operations seem to be > free in Fortran but very expensive in Python with Numeric.
Right, particularly raising to power: a good part of your observations (not all) is accounted for by the fact that Python doesn't perform strength reduction. Taking a trivial example (everything is slow since I'm using a laptop in lowpower mode, but the ratio is meaningful): kallisti:/tmp alex$ /usr/bin/python timeit.py -s'import Numeric; xs=Numeric.ones(999); sum=Numeric.sum' 'sum(xs*xs)' 10000 loops, best of 3: 67.2 usec per loop kallisti:/tmp alex$ /usr/bin/python timeit.py -s'import Numeric; xs=Numeric.ones(999); sum=Numeric.sum' 'sum(xs**2)' 1000 loops, best of 3: 835 usec per loop I guess this plays to the strengths of old fogeys like me -- having grown up in a world where even good (-for-the-time) Fortran compilers (!) often didn't do strength reduction, it just wouldn't _occur_ to me to write x**2 where I mean x*x. ((Memo to self: point this out in the Optimization section of the Nutshell's 2nd ed, since it IS a reasonably frequent performance trap)). I remember once working with a good scientist, back in the early 80's, on getting good running times for his Fortran program coming from some (I believe) CDC machine to an IBM mainframe, and achieving almost an order of magnitude worth of gain in some important function by doing manual strength reduction in just a few statements; apparently the CDC Fortran compiler _did_ do strength reduction (I don't recall whether it was specifically x**2 -> x*x). I believe, but I don't have one at hand to check, that scipy.weave may be better in this regard. Alex -- http://mail.python.org/mailman/listinfo/python-list
