On 10/10/2012 09:08 AM, Brian Paul wrote:
On 10/10/2012 08:59 AM, Kenneth Graunke wrote:
On 10/10/2012 07:42 AM, Jose Fonseca wrote:
I think certain versions of SPEC viewperf rely on NV_vertex_program.
See http://www.mesa3d.org/viewperf.html
We had some internal hacks to support just the bare minimum of to
run some of these tests, but they were not accepted on mesa proper.
(There is some bug report on fdo about it).
Jose
Ugh. I'd forgotten about SPECviewperf.
I guess this begs the question: do we care?
According to that page, viewperf11 is a buggy application (using
extensions without checking for them), and to get it working properly,
we'd need to implement two more legacy extensions that aren't
necessary for anything else. Or add the minimum required and driconf
workarounds to falsely advertise them.
In my experience, viewperf is extremely frustrating to work with and
isn't useful for testing either correctness nor performance. The only
reason anyone appears to care is that it's some kind of "industry
standard" benchmark. These days, however, it seems more people care
about glbenchmark.com's benchmarks, 3DMarkMobileES 2.0, and various
games. At least on my team, no one is measuring us against
SPECviewperf.
Do people still care about viewperf on your side?
Unfortunately, the people that review VMware's products (Workstation,
Fusion, etc) often run Viewperf and we've been dinged by reviewers
when they find issues with it. Part of the motivation for creating
http://www.mesa3d.org/viewperf.html was to educate reviewers about the
issues with Viewperf 11.
I've reported the VP11 issues to SPEC and was told that they'd be
addressed for Viewperf 12. Unfortunately, there's no way for the
public to review/test VP12 before it's released (at least not without
paying a very hefty membership fee) so we have to just cross our
fingers that VP12 will be better implemented than VP11.
Can we please hold off on this change for just a bit while we review
the situation?
OK, I think it's good news.
I hadn't looked at VP in a while but it looks like all the
vertex/fragment programs are actually of the ARB variety, not the NV
variety.
The catia-03 test uses GL_NV_fragment_program2 and
GL_NV_vertex_program3 (without checking if they're actually
supported!) but those extensions are layered on
GL_ARB_vertex/fragment_program, not GL_NV_vertex/fragment_program.
The VP source code definitely has calls to glProgramStringARB() but I
don't see any calls to glLoadProgramNV() which is what NV programs use.
So, I think we're OK with viewperf.
I'll do a review of Eric's patches too...
-Brian
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
