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

Re: ATI linux drivers

Randall Hopper wrote:

I've seen a number of Usenet, mailing list, and cgshaders posts indicating
it works fine for some folks.  A few others mentioned problems, but to what
extent all of these developers rigorously exercised the extensions, knew
what they were doing, and might have been dealing with bugs in past
versions of the Cg compilers is an unknown.  It at least sounds like a
promising approach.
There have certainly been a bunch of errors - and design changes - in
various revisions of the Cg compiler.  It's not a rock solid design yet.

However, that hinges on an assumption that all the latest ATI PS
stuff is exposed to OpenGL through an ATI-specific fragment shader
Yep - if it's not, you're *doomed*.

I made that assumption at first, but after further net research
last night I don't think this is correct.  In the ATI world, it appears
ATI_fragment_shader is roughly the PS 1.4-level OpenGL hook (pre-R300), and
ARB_fragment_program is the PS 2.0-level hook (R300).
Hmmm - OK.

Now whether the ARB extension was flowed to Linux as well as Windows is
another open question...thus my original post.
It seems suprising that ATI would have changed that aspect of the drivers
in porting them to Linux.   Surely the situation must be the same as it is
with nVidia that the Linux and Windoze drivers are pretty much identical.
What's most likely (if it's not supported under Linux) is that the Linux
driver is simply lagging one or two rev levels behind the Windoze driver
and we could reasonably expect it to catch up in a month or two.

 |One problem may be that the length of a fragment program on the Radion
 |is quite small (compared to the NV30 at least) - and it's possible that
 |the overhead of using a high level language would swamp it's abilities.

Possibly.  Depends on the application of course and how similar/different
it is from Doom ;-) It will be very interesting to see how longer programs
with specific sets of instructions affect FPS performance at various
resolutions on the HW-FP cards.  Longer FPs may be possible, but are they
really practical yet.
Yep.  The Radion executes instructions at a slower rate than the NV30 - but
it's doing things like full floating point where the NV30 only does 'half-float'
at full speed.  That bodes ill for a high level language.  The Radion only has
full float - so you have a choice -

  Highest speed: NV30 half-float,
  Slower: Radion full-float,
  Slowest of all: NV30 full-float.

Radeon R300 PS  : 64 ALU instructions, 32 texture instructions, and 4 levels
                  of dependent texture read
NV30 PS         : 1024 instructions
Yep.  The NV30 positively screams "The Right Thing" - but it's essentially
unobtainable to us mere mortals.  There are a couple of them for sale on eBay
at $620 !

(We have an NV30 QuadroFX at work now BTW!  It arrived late on Monday and
I've been snowed in ever since so I havn't been able to play with it.)
---------------------------- Steve Baker -------------------------
HomeEmail: <sjbaker1@airmail.net>    WorkEmail: <sjbaker@link.com>
HomePage : http://www.sjbaker.org
Projects : http://plib.sf.net    http://tuxaqfh.sf.net
           http://tuxkart.sf.net http://prettypoly.sf.net
GCS d-- s:+ a+ C++++$ UL+++$ P--- L++++$ E--- W+++ N o+ K? w--- !O M- V-- PS++ PE- Y-- PGP-- t+ 5 X R+++ tv b++ DI++ D G+ e++ h--(-) r+++ y++++