Following on David's note that IDL is faster when accessing columns
rather than rows, I decided to test exactly how much of a difference
it makes. The result is quite interesting... for a 2D array:
IDL> q = fltarr(10000,10000)
IDL> a = systime(/sec) & p=total(q,1) & print, systime(/sec)-a
0.38179994
IDL> a = systime(/sec) & p=total(q,2) & print, systime(/sec)-a
0.80439496
So basically a factor of 2. But for a 3D array:
IDL> q = fltarr(300,300,300)
IDL> a = systime(/sec) & p=total(q,1) & print, systime(/sec)-a
0.10734391
IDL> a = systime(/sec) & p=total(q,2) & print, systime(/sec)-a
0.12382507
IDL> a = systime(/sec) & p=total(q,3) & print, systime(/sec)-a
0.15879583
Early dimensions are still better, but there's not nearly the same
difference. I'm guessing that this is because the memory gap between
contiguous junks is bigger even when running over the first index?
Going to 4D:
IDL> q = fltarr(80,80,80,80)
IDL> a = systime(/sec) & p=total(q,1) & print, systime(/sec)-a
0.18169713
IDL> a = systime(/sec) & p=total(q,2) & print, systime(/sec)-a
0.20045400
IDL> a = systime(/sec) & p=total(q,3) & print, systime(/sec)-a
0.22229695
IDL> a = systime(/sec) & p=total(q,4) & print, systime(/sec)-a
0.30576181
Hmmm... slight improvement for each of the first 3 dimensions, but a
big penalty on the 4th! Not sure why, but it's very repeatable. In 5D
we reach saturation:
IDL> q = fltarr(35,35,35,35,35)
IDL> a = systime(/sec) & p=total(q,1) & print, systime(/sec)-a
0.26922297
IDL> a = systime(/sec) & p=total(q,2) & print, systime(/sec)-a
0.32297802
IDL> a = systime(/sec) & p=total(q,3) & print, systime(/sec)-a
0.35127592
IDL> a = systime(/sec) & p=total(q,4) & print, systime(/sec)-a
0.35729289
IDL> a = systime(/sec) & p=total(q,5) & print, systime(/sec)-a
0.35672593
I guess that means that this is the point where each successive chunk
is always too far to be in cache and so there's a constant penalty to
go get it from the bus each time?
Anyway, thought you guys might be interested in that. I guess the
moral of the story is that in a 2D array you should always operate on
the first index if possible, but in higher dimensions it doesn't
matter so much.
-Jeremy.
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
