Announcement

Collapse
No announcement yet.

AMD RadeonSI Gallium3D Performance Has A Long Way To Go

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

  • AMD RadeonSI Gallium3D Performance Has A Long Way To Go

    Phoronix: AMD RadeonSI Gallium3D Performance Has A Long Way To Go

    With Fedora 19 presenting a nice "out of the box" experience for AMD Radeon HD 7000 series graphics using the open-source RadeonSI Gallium3D driver, benchmarks of the open-source driver were done and compared to previous generations of AMD hardware. Sadly, there's still much work ahead for the Radeon HD 7000 series driver in being able to catch up with the hardware supported under the mature R600 Gallium3D driver.

    http://www.phoronix.com/vr.php?view=18828

  • #2
    During the RadeonSI Gallium3D testing, for a few of the tests there was also incorrect rendering. Xonotic appeared to be most problematic for RadeonSI
    Why no screenshots?

    Here are some screenshots of how problematic it looks for me (for some reason "ultra" was buggy, but in the past it definitely worked, so this is high):
    http://imgur.com/a/Ixwbx

    In your system description: http://openbenchmarking.org/embed.ph...ha=44bebd4&p=2
    Schrdinger's cat. Come on, you have unicode support, don't you?

    Which llvm version? I'm assuming fedora has 3.3?

    I would suggest testing with 3.10 and with drm-next-3.11-wip-5 and with llvm 3.4 svn/git right ahead.
    Last edited by ChrisXY; 06-28-2013, 02:16 PM.

    Comment


    • #3
      Also a possibility is the fact Fedora doesnt ship libtxc for S3TC support. I dont know if any of these games ship with S3TC textures, but Nexuiz might-- which would explain that at least partially.

      Comment


      • #4
        Very disappointing indeed. The 5830 runs much faster than the 7950.

        Comment


        • #5
          Things/questions that we should probably keep in mind:

          1) Llano and other AMD APUs default to a very low clock speed in the VBIOS... Could SI be doing the same?
          2) All of the SI cards were using GLAMOR, while the other cards were using EXA
          3) Do we know if LLVM was compiled as a debug build in this Fedora build? Debug builds of LLVM are much slower at compiling shaders
          4) S3TC probably isn't enabled in this Fedora build, which probably hurts all architectures, but maybe not all are affected equally.
          5) There's still lots of room for improvement in the LLVM back-end for all radeon generations
          6) Was the r600g driver using llvm for shader compilation or the TGSI back-end?

          The results are disappointing, yes... but how much of this is due to deficiencies in the code, and how much is due to differences in the test setup between card generations?

          Edit:
          Regarding #1, I found this:
          http://www.techpowerup.com/vgabios/i...E&memSize=1024

          Most 7850 1GB cards there default to high clocks. But a few default to clocks that are lower than their optimum-performance levels.
          Last edited by Veerappan; 06-28-2013, 03:52 PM.

          Comment


          • #6
            No 2D benchmarks :/

            Sad that 2D performance was completely ignored in this benchmark round.
            It would have been really interesting to see how well the completely re-done 2D driver architecture RadeonSI performs ... although I have to admit I do not have high hopes

            Comment


            • #7
              Originally posted by Veerappan View Post
              Things/questions that we should probably keep in mind:

              1) Llano and other AMD APUs default to a very low clock speed in the VBIOS... Could SI be doing the same?
              2) All of the SI cards were using GLAMOR, while the other cards were using EXA
              3) Do we know if LLVM was compiled as a debug build in this Fedora build? Debug builds of LLVM are much slower at compiling shaders
              4) S3TC probably isn't enabled in this Fedora build, which probably hurts all architectures, but maybe not all are affected equally.
              5) There's still lots of room for improvement in the LLVM back-end for all radeon generations
              6) Was the r600g driver using llvm for shader compilation or the TGSI back-end?

              The results are disappointing, yes... but how much of this is due to deficiencies in the code, and how much is due to differences in the test setup between card generations?
              About Fedora... by now, with release so close, Debug SHOULD be turned off by now unless they are waiting until the day before to recompile all the images with debugging disabled.

              Comment


              • #8
                dont forget that 7XXX cards bios tend to start at lowest power profile, you need to change it to high, for example as root before benchmarking

                Code:
                echo "high" > /sys/class/drm/card0/device/power_profile
                I am sure that speed difference ~8x because of that. My 7850 work almost perfect here, after that trick. 2D scrolling, gnome-shell transitions, all become very smooth.

                Comment


                • #9
                  Originally posted by Linuxhippy View Post
                  Sad that 2D performance was completely ignored in this benchmark round.
                  It would have been really interesting to see how well the completely re-done 2D driver architecture RadeonSI performs ... although I have to admit I do not have high hopes
                  2D will be covered separately in its own article.
                  Michael Larabel
                  http://www.michaellarabel.com/

                  Comment


                  • #10
                    Originally posted by Michael View Post
                    2D will be covered separately in its own article.
                    When you do the 2D article, could you make sure to include some of the r600g cards with both EXA and GLAMOR modes? It'd help us to see how much of the performance difference is due to GLAMOR, and how much is due to the drivers.

                    Comment


                    • #11
                      Do one covering radeon.dpm=1

                      Comment


                      • #12
                        Originally posted by storm_st View Post
                        dont forget that 7XXX cards bios tend to start at lowest power profile, you need to change it to high, for example as root before benchmarking

                        Code:
                        echo "high" > /sys/class/drm/card0/device/power_profile
                        I am sure that speed difference ~8x because of that. My 7850 work almost perfect here, after that trick. 2D scrolling, gnome-shell transitions, all become very smooth.
                        You're right about the power profiles... I guess that VBIOS info was misleading.

                        From /sys/kernel/debug/dri/0/radeon_pm_info (root-only) for my 7850:

                        default engine clock: 860000 kHz
                        current engine clock: 149990 kHz
                        default memory clock: 1200000 kHz
                        current memory clock: 149990 kHz
                        voltage: 1075 mV
                        PCIE lanes: 16

                        Comment


                        • #13
                          Hm..... http://openbenchmarking.org/result/1...UT-1306287SO77
                          drm-next-3.11-wip-5, llvm svn 185061, latest libdrm git, mesa git, xf86-video-ati-git, glamor git

                          Maybe it was because of kwin's xrender compositing or maybe it was because of prime.
                          (kwin opengl compositing has the problem that often it only shows a black screen for 3d using prime until you disable and reenable compositing with alt+shift+f12)

                          xonotic on ultra was bugged, most of the level did not render. But every other test was rendering fine.
                          Code:
                          warning: failed to translate tgsi opcode DDX to LLVM
                          Failed to translate shader from TGSI to LLVM
                          EE si_state.c:1951 si_shader_select - Failed to build shader variant (type=1) -22
                          warning: failed to translate tgsi opcode DDX to LLVM
                          Failed to translate shader from TGSI to LLVM
                          EE si_state.c:1951 si_shader_select - Failed to build shader variant (type=1) -22
                          warning: failed to translate tgsi opcode DDX to LLVM
                          Failed to translate shader from TGSI to LLVM
                          EE si_state.c:1951 si_shader_select - Failed to build shader variant (type=1) -22
                          warning: failed to translate tgsi opcode DDX to LLVM
                          Failed to translate shader from TGSI to LLVM
                          EE si_state.c:1951 si_shader_select - Failed to build shader variant (type=1) -22
                          warning: failed to translate tgsi opcode DDX to LLVM
                          Failed to translate shader from TGSI to LLVM
                          EE si_state.c:1951 si_shader_select - Failed to build shader variant (type=1) -22
                          warning: failed to translate tgsi opcode DDX to LLVM
                          Failed to translate shader from TGSI to LLVM
                          EE si_state.c:1951 si_shader_select - Failed to build shader variant (type=1) -22
                          warning: failed to translate tgsi opcode DDX to LLVM
                          Failed to translate shader from TGSI to LLVM
                          EE si_state.c:1951 si_shader_select - Failed to build shader variant (type=1) -22
                          warning: failed to translate tgsi opcode DDX to LLVM
                          Failed to translate shader from TGSI to LLVM
                          EE si_state.c:1951 si_shader_select - Failed to build shader variant (type=1) -22
                          warning: failed to translate tgsi opcode DDX to LLVM
                          Failed to translate shader from TGSI to LLVM
                          EE si_state.c:1951 si_shader_select - Failed to build shader variant (type=1) -22
                          warning: failed to translate tgsi opcode DDX to LLVM
                          Failed to translate shader from TGSI to LLVM
                          EE si_state.c:1951 si_shader_select - Failed to build shader variant (type=1) -22
                          warning: failed to translate tgsi opcode DDX to LLVM
                          Failed to translate shader from TGSI to LLVM
                          EE si_state.c:1951 si_shader_select - Failed to build shader variant (type=1) -22
                          warning: failed to translate tgsi opcode DDX to LLVM
                          Failed to translate shader from TGSI to LLVM
                          EE si_state.c:1951 si_shader_select - Failed to build shader variant (type=1) -22



                          And this dpm really needs the auto poweroff of hybrid gpus that are not used. dpm definitely seems to work as the fan goes off now and then, but then after a little while it goes on again. And while the test was running it did not go off.
                          Last edited by ChrisXY; 06-28-2013, 04:28 PM.

                          Comment


                          • #14
                            "6450 beats the 7850"

                            Go 6450!

                            Comment


                            • #15
                              Originally posted by ChrisXY View Post
                              xonotic on ultra was bugged, most of the level did not render. But every other test was rendering fine.
                              Code:
                              warning: failed to translate tgsi opcode DDX to LLVM
                              If those DDX opcode warnings are the problem, then it might be solved pretty soon. I think i've seen patches on the mailing list implementing DDX and DDY.

                              I also think that's one of the last things needed before GL3 support is done, although there might be a few others.

                              Comment

                              Working...
                              X