Re: OSG? was Re: [Flightgear-devel] Re: Graphics Engine

[Christian M. <mail@ChristianMayer.de>]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA256

Mathias Fr=F6hlich schrieb:
>> The sg* vectors and matrices are created in such a way that they'll
>> offer the higest possible performance and compatability for using Open=
GL.
> column major like fortran :)
> Well, I worked, together with some collegues, on getting a linux cluste=
r into=20
> the top500 some time ago :)

Good to hear.

(BTW: my diploma thesis [=3D roughly a masters thesis] was the creation o=
f
cache oblivious matrix operations in C++ using the space filling Peano
curve; I could beat the Intel Math Kernel Library in optimal cache use -
und even performance wise when I used only x87 instructions :) I put the
code and the thesis at http://tifammy.sf.net/
Space filling curves have very interesting properties for high
performance computing - not only for cache efficiency but also for
partitioning of problems for parallel computers / clusters)


> Yep, I believe that this small vector set will even provide more perfor=
mance=20
> since it will just work with all ss?g* functions natively. Instead of t=
hat=20
> horrible mix of sg* and simgear point3d datatypes we have at the moment=
,=20
> which do not interface well and thus needs masses of hand coded copies =
from=20
> one datatype to the other. You can just use such a thing as a drop in=20
> replacement for any sg type.

At the times I did active FGFS development I also didn't like the many
different vector classes.

> It takes me regularily needless time to clearify what I have in my hand=
s if I=20
> get a Point3D datatype.

Different classes (or just class names) sound like an good idea.


It might be interesting to offer the use of SSE values internally. This
could give a little performance boost for modern IA-32 processors.

CU,
Christian