Re: Copying rows in a two dimensional array.

[jonathan <johnscpg@googlemail.com>]

On Feb 14, 1:54 am, Hibou57 (Yannick Duchêne)
<yannick_duch...@yahoo.fr> wrote:
> Le Sun, 14 Feb 2010 01:42:13 +0100, jonathan <johns...@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)

should be of course:

  First example: we eigen-decompose an N x N = 2048 x 2048 matrix.
  The data storage matrix is M x M = (2048+Padding) x (2048+Padding)

> 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...
> 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.

Yes, I remembered this, probably from an old post of yours.  I wanted
to cite it when
when I posted earlier, but I could not find the site.  This is not
something you
forget quickly (frames 17 and 18):

   1980: RAM latency ~ 1 cycle
   2009: RAM latency ~ 400+ cycles

It's the heart of the matter, and it is just getting worse.  Helps
convince me
anyway that I did not waste time on an unimportant matter! In
numerical linear
algebra the usual solution is to restructure matrices as a collection
of
blocks.  That has a few of its own problems though. Minor footnote: I
did
some tests on Intel's new nehalem CPU's.  Vastly improved performance
on these
multi-megabyte arrays.  Problem not cured though. Don't know enough to
say more about it.
Thanks for the reminder.

Jonathan