Re: CUDA in Octave 322

[Thomas Ilnseher]

Am Freitag, den 11.09.2009, 12:17 +0200 schrieb Jaroslav Hajek:
> You can, of course, use CUDA in your
> own C++ extensions, but you can't distribute those, because CUDA is
> not free software.
Ok, I might be wrong here, but AFAIK Octave uses LAPACK.

As GPL still allows linking against CS "platform libraries", you can
still link Octave against a CS CUDA enabled LAPACK Lib (granted that
such a thing exists) when you define LAPACK as a platform library. 

Same thing as running *ANY* GPL program on Windowze... (it will link
against some CS win32 libs dynamically).

Same thing as running a GPL prog on Linux w/ NV or AMD binary driver
installed.
... And So On (tm).

It's the question whether LAPACK is a platform library. IMHO it is (you
can substitute ATLAS with IMKL AMD something which are CSd).


Now you only need a LAPACK lib that exploits CUDA.

Problem here is that you have to  copy data to the GPU's memory nad
back, call the GPU and so on. This probably gives a reasonable overhead,
so you'd only accelerate things this way when operating with *LARGE*
vectors / matrices.
-- 
Thomas Ilnseher <address@hidden>