gm2
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Gm2] Numerical Library : GM2 issue with multidimensional array para


From: Michael Riedl
Subject: Re: [Gm2] Numerical Library : GM2 issue with multidimensional array parameter
Date: Sun, 4 Mar 2018 21:56:51 +0100
User-agent: Mozilla/5.0 (X11; Linux i686; rv:52.0) Gecko/20100101 Thunderbird/52.6.0

Hallo Gaius,

thanks for taking care about this. I will make some extensive tests as soon as all modules run well. And your remark out unoptimized code is right without doubt - I only added the number for reference. Tried to benchmark with a purely 2-dimensional version of the Jacobi routine, unfortunately the compiler crashed. Will send out an example file soon.

Michael

PS: Currently I am working on some test routines to have a more systematic testing of some routines - will present the results soon.


Am 02.03.2018 um 12:33 schrieb Gaius Mulley:
Michael Riedl <address@hidden> writes:

Hi Michael,

RTExceptions.mod:639:9:invalid address referenced in invalidloc

Perhaps you find some time to have a look on that.
sure - I can reproduce this on my machine.  Will fix it.

Just as an interesting point I also compared the timings of both the
REAL64 and REAL96 variant (see third attachment). The tasks is to
compute the eigensystem of a 256x256 matrix 100 times by a classical
jacobi routine.

As expected in one of my former emails I see a huge time penalty for
the 12 BYTE reals (approx 1.8 times longer runtime if not optimized,
2.7 times longer runtime if optimized with -O2). Keeping in mind that
using a linarised array for the input matrix for the Jacobi
diagonalisation routine is memory but not cache optimal I expect an
even more dramatic difference for the other routines (e.g. for the SVD
routines not yet part of the routines I did sent you).
interesting timings indeed.  I'd ignore timing unoptimised code as it is
not designed to run fast but to compile fast and provide very accurate
debugging information.

Encouraging that the gm2 REAL/LONGREAL are both faster and more accurate
(wider) than xds.  It might be interesting to see if your test can be
compiled with

    -O2 -fm2-whole-program

or

    -O3 -fm2-whole-program

I see good results when compiling a game engine

regards,
Gaius




reply via email to

[Prev in Thread] Current Thread [Next in Thread]