Re: nVidia FX Support?

On Thu, 2004-05-06 at 20:09, Julian Elischer wrote:
> On Thu, 6 May 2004, Jason Andresen wrote:
> 
> > Kenneth Culver wrote:
> > 
> > > Quoting Jason Andresen <jandrese_at_mitre.org>:
> > > 
> > >> Will Andrews wrote:
> > >>
> > >>> On Wed, May 05, 2004 at 11:28:57AM +0200, Dimitry Andric wrote:
> > >>>
> > >>>> The problem is that NVidia hasn't updated its binary-only FreeBSD
> > >>>> driver since July 1st, 2003.  Any cards released after this date will
> > >>>> not work properly, or at all, with this driver.  So please go bug
> > >>>> NVidia. :)
> > >>>>
> > >>>> But, see also this thread:
> > >>>> http://lists.freebsd.org/pipermail/freebsd-threads/2003-June/000530.html
> > >>>>
> > >>>>
> > >>>> which might explain why it takes forever for new drivers to arrive. :(
> > >>
> > >>
> > >> Ugh, reading through that thread I got a bad bad feeling that it will be
> > >> ages before we see another NVidia driver.  The NVidia engineer is
> > >> unwilling to proceed without his %gs register, and the FreeBSD engineers
> > >> are unwilling to give it to him because the fast thread switching needs
> > >> it for good performance.  The NVidia engineers needs the %gs register
> > >> for good performance, and will not release drivers that are 10-50%
> > >> slower than the Windows ones.
> > >>
> > >> It looks like the whole situation is at a standstill at the moment,
> > >> which means we get no new drivers until the whole thing is sorted out.
> > >> It's really distressing to see how frustrated the NVidia engineer was by
> > >> the end of the thread.
> > >>
> > > I'm pretty sure that the FreeBSD people are working on TLS, which will 
> > > allow the
> > > nvidia people to be happy.
> > 
> > Maybe I'm reading it wrong, but it seems like TLS is what got them in 
> > all of this trouble in the first place.  It seems that FreeBSD's 
> > implementation uses the %gs register, which NVidia also wants to use.
> 
> %gs is not theirs to use.. linux also uses the %gs register
> for thread identification
> 
> however when TLS is implemented then they will 'piggyback' on th thread
> sytem, (so to speak).

For what its worth, TLS works nicely in the P4 kse branch for i386.