Announcement

Collapse
No announcement yet.

Early testers with r300-r500 GPUs wanted

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

  • Early testers with r300-r500 GPUs wanted

    Hi,

    today I've pushed preliminary support for ARB_occlusion_query extension.
    It's based on mesa radeon-rewrite branch. With this extension enabled r300 driver advertises GL 1.5. Yupi!

    Currently I've tested it only on my RV535 (X1650) without KMS. Everything seems to be working. Unfortunately ARB_oq for SW TCL is broken now, and proper implementation will take some time. It also probably won't work with KMS enabled.

    Besides some conformance testing apps I've tested it under sauerbraten game - it gives nice performance boost (especially on bigger maps), so there is an extra motivation for you to test - it can make your favorite game playable

    Please test, and post your results (tested apps, performance boosts/regressions).

    You can find the code here: http://cgit.freedesktop.org/~osiris/mesa/ (occlusion-query branch)

  • #2
    Here is a quick guide:
    • you need the 2.6.30 or later kernel (or else Occlusion Query will be disabled).
      Ubuntu 9.04 users: download (just the linux-image-* for your architecture), install (with sudo dpkg -i) and boot the latest 2.6.30.x kernel from: http://kernel.ubuntu.com/~kernel-ppa/mainline/?C=M;O=D
    • make sure you have the development environment to build mesa. Ubuntu 9.04 users:
      Code:
      sudo apt-get build-dep mesa; sudo apt-get install git-core
    • then:
      Code:
      cd ~
      git clone git://anongit.freedesktop.org/mesa/mesa
      cd mesa
      ./autogen.sh --disable-gallium --with-dri-drivers="swrast,r300"
      make
      export LIBGL_DRIVERS_DIR=~/mesa/lib

    Test on a RV530 (313MHz GPU - 300 MHz memory) with an Intel Core Duo T2600 @ 2.16Ghz:
    • arbocclude demo from mesa-osiris/progs/demos works fine;
    • sauerbraten also works fine but it's slower (with OQ disabled ~85fps, with OQ enabled ~79fps).

    EDIT:
    As discussed on IRC performance improvement is related to CPU and GPU clock ratio. Downclocking my CPU to 1GHz with sauerbraten gives the OQ version better performance (~65 fps) than without OQ (~47fps), at least with sauerbraten and when not using KMS. OQ improves performance only with a "fast" GPU and a "slow" CPU.

    EDIT 2:
    Occlusion query support was merged in mesa master in 2009-08-15 and will be included in mesa 7.6.
    Last edited by oibaf; 08-18-2009, 10:52 AM.

    Comment


    • #3
      oibaf, might want to prepend: sudo apt-get build-dep mesa; sudo apt-get install git-core

      Comment


      • #4
        Originally posted by tormod View Post
        oibaf, might want to prepend: sudo apt-get build-dep mesa; sudo apt-get install git-core
        Done, thanks.

        Comment


        • #5
          Hi,

          I've experienced a performance drop using OQ, I tested it with an XPress 200M.
          I tried open arena and tremulous. With tremulous, the drop was about 50%.

          I also would like to thank you for your amazing work on the r300 serie !

          Comment


          • #6
            Osiris was working with another tester a day or two ago on IRC... it seems that OQ gives a performance boost if you have a slow CPU and a fast GPU, but actually slows things down if you have a fast CPU and slow GPU (because the CPU can do the calculations in less time than the GPU).

            I think the consensus was to add an option to enable/disable OQ. Not sure if it is possible to speed up OQ on the GPU.

            Comment


            • #7
              Originally posted by bridgman View Post
              it seems that OQ gives a performance boost if you have a slow CPU and a fast GPU, but actually slows things down if you have a fast CPU and slow GPU (because the CPU can do the calculations in less time than the GPU).
              Sounds like this might be interesting for me -- Radeon 9800 Pro with an 800MHz DEC Alpha. Or is a 9800 considered slow now?

              Comment


              • #8
                I'm pretty sure a 9800 would still count as "fast" in this case; it should be ~6x the performance of an Xpress200 anyways.

                Comment

                Working...
                X