Announcement

Collapse
No announcement yet.

Why The Radeon Gallium3D Performance Is Down

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

  • Why The Radeon Gallium3D Performance Is Down

    Phoronix: Why The Radeon Gallium3D Performance Is Down

    After yesterday's article about the Grinch that stole the Radeon Gallium3D performance, here's three offending commits since Mesa 7.10 that are causing the open-source Radeon Gallium3D driver to run slower than it should.

    http://www.phoronix.com/vr.php?view=16875

  • #2
    Very nice, Michael! Enjoy your beer

    Comment


    • #3
      Excellent

      Good job, Michael! I hope your detective work helps to rectify the problem.

      During the holidays I plan to use the PTS for the first time to see how it works for me. But it's obvious it is of great use to the Linux community.

      Happy holidays, everyone!

      Comment


      • #4
        Who would have guessed?

        It's not like any of those commits have already been found to be broken, is it.

        Oh, wait...

        [r300g] Flickering user interface in WoW

        Raised back in October and bisected to 363ff844753c46ac9c13866627e096b091ea81f8.

        Can we please now revert this? How much more damage does it need to cause!?

        Comment


        • #5
          Such articles reminds me to disable ABP on phoronix.com!!

          Well deserved beer!

          (only thing that is missing, is info about notifing proper mesa developers about your findings, but I guess they read Phoronix regularly anyway)
          Last edited by przemoli; 12-23-2011, 11:20 AM.

          Comment


          • #6
            Thaks Michael!!! Good job.
            I hope to Marek...

            Comment


            • #7
              Originally posted by przemoli View Post
              (only thing that is missing, is info about notifing proper mesa developers about your findings, but I guess they read Phoronix regularly anyway)
              If there's no real bugreport and just this article, I can only guess the level of annoyance this causes them, expecially when random "helpful" people also point them at it.

              Comment


              • #8
                Reasons for moving to winsys:
                http://lists.freedesktop.org/archive...st/010263.html

                It looks like it was expected to see a 0-15% performance boost, but because GEM_WAIT was disabled, it ended up becoming a performance regression..
                Disabing GEM_WAIT is probably temporary, so it's probably nothing to really worry about.

                On a side note, the Gallium drivers have in the past needed more CPU usage to hit the higher framerates than the Catalyst drivers did.. So fixes like this that reduce overhead gets me all excited.
                Last edited by Sidicas; 12-23-2011, 11:41 PM.

                Comment


                • #9
                  chrisr, it is reverted in this branch, which I will push soon:
                  git://people.freedesktop.org/~mareko/mesa radeon-perf-fix
                  It fixes the performance regression in Nexuiz and hopefully even the bug you reported.

                  Sidicas, the performance improvement is unrelated to the regression in r300g. The regression is caused by a commit that's merely a cleanup.

                  Michael, thanks for doing this. The commit "winsys/radeon: move GEM domains out of the drivers into winsys" will be reverted, fixing Nexuiz.
                  The commit "r300g: implement fences using dummy relocations" won't be reverted. It properly implements glFinish(), which is used by openarena to eliminate the input lag. It's important for playability of the game.

                  Comment


                  • #10
                    Thank you Michael for a well written, very informative article!!!

                    Comment


                    • #11
                      Kudos to both Michael and Marek.

                      Comment


                      • #12
                        Compiling it now...

                        Originally posted by marek View Post
                        chrisr, it is reverted in this branch, which I will push soon:
                        git://people.freedesktop.org/~mareko/mesa radeon-perf-fix
                        It fixes the performance regression in Nexuiz and hopefully even the bug you reported.
                        Thanks, Marek. This bug affects both R600 and R300 and so I'm really hoping that it will be fixed before Mesa 7.12 is released.

                        Comment


                        • #13
                          Originally posted by marek View Post
                          chrisr, it is reverted in this branch, which I will push soon:
                          git://people.freedesktop.org/~mareko/mesa radeon-perf-fix
                          It fixes the performance regression in Nexuiz and hopefully even the bug you reported.

                          Sidicas, the performance improvement is unrelated to the regression in r300g. The regression is caused by a commit that's merely a cleanup.

                          Michael, thanks for doing this. The commit "winsys/radeon: move GEM domains out of the drivers into winsys" will be reverted, fixing Nexuiz.
                          The commit "r300g: implement fences using dummy relocations" won't be reverted. It properly implements glFinish(), which is used by openarena to eliminate the input lag. It's important for playability of the game.
                          Unfortunately 93f4e3cb6c1ca303ee1f5c2a2491a8eff33f2633 regress 0 A.D.: fps go from ~30 to ~20.

                          Comment


                          • #14
                            Originally posted by oibaf View Post
                            Unfortunately 93f4e3cb6c1ca303ee1f5c2a2491a8eff33f2633 regress 0 A.D.: fps go from ~30 to ~20.
                            Forgeot to add that I am using kernel 2.6.39.3 and my card info:
                            Code:
                            r300: DRM version: 2.9.0, Name: ATI RV530, ID: 0x71c5, GB: 1, Z: 2
                            r300: GART size: 509 MB, VRAM size: 256 MB
                            r300: AA compression RAM: YES, Z compression RAM: YES, HiZ RAM: YES

                            Comment


                            • #15
                              Originally posted by oibaf View Post
                              Unfortunately 93f4e3cb6c1ca303ee1f5c2a2491a8eff33f2633 regress 0 A.D.: fps go from ~30 to ~20.
                              In that case, why don't you test whether the first commit, which it partially reverted, improved 0 A.D. performance?

                              Offtopic: If only my HD3200 worked with mesa trunk... bug 43405: Evergreen support broke it
                              --there's a fix, but noone's tested on Evergreen. (guess I'll test if git head works on HD3200 later today, though there's very little chance of it).

                              Comment

                              Working...
                              X