More examples

[Oleg Batrashev]

Hi,

I've commited more examples, the largest is 64k unknowns (256x256 grid size)
I actually moved all examples to the "examples" directory,

I have found C++ procedural code and CG with that (I think Vladimir's work?).
I can see some C++ classes (Ains?) but it wraps (Reins?) C code and CG
uses C++ procedural code (Vladimirs?). as well as come classes.
Yes, and Reins CG code (currently in SVN) seems to have memory leak
and this can also be in other codes as well.

Seems like 30%+30%+30% but complete mess :)

Feel free to correct me via email or MSN.

Oleg

[Oleg Batrashev]

>
> I've commited more examples, the largest is 64k unknowns (256x256 grid size)
> I actually moved all examples to the "examples" directory,

I tried the largest example with DOUG and with tolerance 1E-18 (square
root if r*r) it takes

~25 seconds, ~1100 iterations with just CG
~4 seconds, 1 iteration with PCG and 1-level preconditioner

But the second test is very unfair and does not show anything, because
it is too small example and it actually uses direct solver - UMFPACK
(I do not want to get into details about preconditioners).

Oleg

PS. I think that doing that in C and C++ in parallel may be good
thing, because it will show where in C one can hack (optimize) that is
difficult to include into C++ code.
PPS. I can write C++ interfaces for Vector and Matrix which I would
like to see but it is even better to discuss it in MSN tomorrow with
the people concerned about C++ implementation.

[Lauri Tulmin]

current code from svn took

~13 seconds & 482 iterations with same error. 1000 iterations took ~27 seconds.

[Oleg Batrashev]

On Tue, Apr 29, 2008 at 1:08 AM, Lauri Tulmin <tul...@gmail.com> wrote:
> current code from svn took
> ~13 seconds & 482 iterations with same error. 1000 iterations took ~27
> seconds.

Are you sure, number of iterations must be the same. 1E-18 was the
root square of r*r, so r*r=1E-36

Otherwise seems ok, lets see what is possible with 6 SPUs.

Oleg

[Oleg Batrashev]

> I have found C++ procedural code and CG with that (I think Vladimir's work?).
> I can see some C++ classes (Ains?) but it wraps (Reins?) C code and CG
> uses C++ procedural code (Vladimirs?). as well as come classes.
> Yes, and Reins CG code (currently in SVN) seems to have memory leak
> and this can also be in other codes as well.

Seems like I completely messed the names, sorry :)

Oleg