Re: Copying rows in a two dimensional array.

["Hibou57 (Yannick Duchêne)" <yannick_duchene@yahoo.fr>]

Le Sun, 14 Feb 2010 01:42:13 +0100, jonathan <johnscpg@googlemail.com> a  
écrit:
> First example: we eigen-decompose an N x N = 2048 x 2048 matrix.
> The data storage matrix is M x M = (1024+Padding) x (1024+Padding)
> Here is the running time in seconds of an iterative jacobi
> eigen-decomposition:
>
>    2048x2048:  322 sec, gnat      (Padding=24)
>    2048x2048: 1646 sec, gnat      (Padding=0)
>    2048x2048: 1632 sec, gfortran  (Padding=0)
>    2048x2048: 1492 sec, ifort     (Padding=0)
>
> The observed 500% slowdown in the absence of padded arrays is
> unacceptable, even if it is a rare event (occurring only on 2**p sized
> data arrays). In fact it's not all that rare ... more comments on
> that below.  (BTW, ifort is the INTEL fortran compiler, all
> optimizations at Max. gfortran is the gcc variant.)
So this is mostly about representation clauses finally. Is that it ?

Do not know if you already know this document (as I remember I picked it  
up from some one thread at comp.lang.ada), I've talked about on the other  
fr.c.l.a :
http://research.scee.net/files/presentations/gcapaustralia09/Pitfalls_of_Object_Oriented_Programming_GCAP_09.pdf
I had pointed about frames  #17, #18, #19 et #20, which contains good  
source of inspiration. Hope this could help you to figure a path.

You've posted a long list of tests-bench and observations. I did not  
looked at every thing, but hope I will have a more closer look at it later.

-- 
No-no, this isn't an oops ...or I hope (TM) - Don't blame me... I'm just  
not lucky