Announcement

Collapse
No announcement yet.

"Ask ATI" dev thread

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • Originally posted by cjr2k3 View Post
    Bridgman already say that, in Linux, CrossFire is App specific, no scissor, no anything on other apps.
    It doesn't matter what the default mode is. What matters is that it's possible to set overrides for profiles in *DirectX* programs in CrossFire so surely it must be to make a similar system for OpenGL. Then you could manually set just the kind of a CrossFire profile for the program (read: every program that happens to use OpenGL at the time) that your heart desires, it wouldn't need to have its specific own profile. I was not asking how it's currently being done, I was asking if it's too much work to code in the profile overrides.

    Comment


    • you could edit atiogl.xml and tell us what is happening - adding
      <CrossFireCaps>0x00000001</CrossFireCaps>
      to your favorite app - or adding your favorite app.

      Comment


      • First off - I know glxgears isn't a benchmark. That said, my HD3200 is only getting ~540 fps running the 8.561 driver under linux 2.6.28, 64bit.
        My ancient radeon 9600 gets 8500+ FPS under the opensource 'radeon' driver. With a difference that huge, something must be screwed up.

        Lightsmark 2008 64bit in linux gets an average of 14 (fourteen!) FPS, with lows of 4-7 FPS when doing radiosity calculations.

        Lightsmark 2008 under windows XP (32bit, same hardware) gets more like 30-40fps average. I can't give you the exact number since I'm not at that box right now. Suffice to say it's an extremely noticeable difference.

        The only oddity I noticed is that the fglrx kernel driver is preemption-unsafe in a few functions - kas_spin_lock, kas(Get/Set)ExecutionLevel and the whole usage of kasExecutionLevel in general is meaningless since you could be preempted after setting it on one core and end up on another before even returning from the function. I don't think that has anything to do with the performance issues, since I was using a non-preemptable kernel before and had the same low framerates.

        Comment


        • Originally posted by harik View Post
          First off - I know glxgears isn't a benchmark. That said, my HD3200 is only getting ~540 fps running the 8.561 driver
          I've noticed some pretty odd results on my RadeonHD 3870 with the 8.561 too. Just after booting I results roughly 7600, then wait a bit and rerun, results drop to 6500, then still wait a bit and rerun, results are about 3600. Wonder if this is a power management feature or a rapidly degrading performance bug. If it's a power management feature, my question is, how do I turn this off?
          I've a 2.6.28 as 64bit too.

          Comment


          • I'm actually not sure any frame rate results higher than 120 mean anything, since there aren't any monitors that refresh that fast. Do you really want drivers optimized for a case that isn't useful? Fast clears/flips at the expense of speed in more complex rendering would not be the best tradeoff.

            Comment


            • We *really* need a better benchmark for Linux, something that makes good use of textures and shaders, probably coded for somewhere around GL 1.3 so that all of the open and proprietary drivers will run it without fallbacks.

              Harik; I'll try to check tomorrow re: what kind of frame rates I get with my RV620, which is similar to the 3D engine in a 780 / HD3200. The 8500fps number for the 9600 seems really high, are you sure you had the same window size on both tests ?
              Test signature

              Comment


              • Originally posted by bridgman View Post
                We *really* need a better benchmark for Linux, something that makes good use of textures and shaders, probably coded for somewhere around GL 1.3 so that all of the open and proprietary drivers will run it without fallbacks.
                I agree, we do need a better benchmark. Lightsmark unfortunately requires OpenGL 2.0. The ones that kill the framerate are realtime radiocity, color bleeding, area lights, penumbra shadows, and the flyby of the cage. That set of tests gets 10 FPS at 1280x1024 _AND_ 320x200! Something is very wrong there - that should have at least some difference as the size of the scene increases. The other scenes vary as you'd expect - 320x200 had a peak FPS of over 250 for some of them, compared to mid 20s in 1280. Sadly lightsmark does not log individual framerates so I have to take notes of everything but the overall average.

                Originally posted by bridgman View Post
                Harik; I'll try to check tomorrow re: what kind of frame rates I get with my RV620, which is similar to the 3D engine in a 780 / HD3200. The 8500fps number for the 9600 seems really high, are you sure you had the same window size on both tests ?
                I'll doublecheck when I get home tonight. It's really a terrible benchmark and the more I think about it the less it seems to matter. I'd be much more interested in your lightsmark 2008 score, since that actually tests more of the hardware.

                I'm going to recompile my kernel without preempt to see if that helps - since it looks like important bits of the driver are operating under the assumption they won't be interrupted.

                Comment


                • I lost a lot of time today with car problems; probably won't get a chance to run on 620 until Monday. Sorry...
                  Test signature

                  Comment


                  • I haven't followed ATi's linux drivers for long, what's the problem with requiring OGL 2.0? Does ATi not even support 2.0?

                    Comment


                    • Originally posted by bridgman View Post
                      I lost a lot of time today with car problems; probably won't get a chance to run on 620 until Monday. Sorry...
                      Don't you wish you could just plug a laptop into the ECU and see what's going on yourself?

                      Comment

                      Working...
                      X