Announcement

Collapse
No announcement yet.

mesa compiled with ekopath

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

  • mesa compiled with ekopath

    I have just compiled mesa with EKOpath, and did a quick ut2004 bench.
    h/w
    • athlon x2 3800
    • and rv670 (hd3850) AGP


    s/w
    • gcc 4.5.2 with "-O3 -march=native -pipe -fgraphite-identity -floop-interchange -floop-strip-mine -floop-block -ffast-math" with --enable-asm
    • ekopath with "-O3 -ffast-math -pipe" (-Ofast use -ipa and that seg faults here) with --disable-asm
    • mesa git master r600g


    results
    br-bridgeoffate?spectatoronly=1?numbots=8?quickstart= 1?attractcam=1 -benchmark -seconds=60 -nosound
    • gcc
      45.673248 / 114.558311 / 195.075378 fps -- Score = 79.787827 rand[1450720973]
      75.656464 / 114.558418 / 317.631805 fps -- Score = 80.039886 rand[1450720973]
      73.394325 / 114.273186 / 317.922424 fps -- Score = 80.037209 rand[1450720973]
    • ekopath
      44.639256 / 116.853668 / 194.466736 fps -- Score = 79.757851 rand[413195207]
      76.800453 / 114.661110 / 324.195709 fps -- Score = 80.041740 rand[1450720973]
      72.066681 / 114.808113 / 329.388702 fps -- Score = 80.036087 rand[1450720973]

    DM-Deck17?spectatoronly=1?numbots=4?quickstart=1?attr actcam=1 -benchmark -seconds=60 -nosound
    • gcc
      40.071953 / 143.693680 / 395.564453 fps -- Score = 80.000160 rand[1717337507]
      40.100864 / 143.590195 / 401.297119 fps -- Score = 80.000221 rand[1717337507]
      24.770853 / 143.774368 / 246.122559 fps -- Score = 79.863129 rand[1717337507]
    • ekopath
      40.836777 / 144.144852 / 399.184967 fps -- Score = 79.997742 rand[1717337507]
      40.543324 / 141.011581 / 406.091919 fps -- Score = 79.966469 rand[1717337507]
      40.658440 / 142.379562 / 401.178375 fps -- Score = 80.000214 rand[1717337507]

  • #2
    Originally posted by Qaridarium
    ?? this numbers confuse me??? what goes wrong in that round??
    I don't know, lag spikes (ut2004 tends to do that in-game here).

    I have bender bench using this script
    • gcc
      Benchmark Results, Screen Size 1920 x 1061

      Overall Score (FPS)
      gl : 23.0949 fps
      render : 10.54 sec

      Spin wireframe view, subsurf monkey, 4 subsurf levels : 52.9539 fps
      Spin solid view, subsurf monkey, 4 subsurf levels : 29.1804 fps
      Spin solid view, 1000 monkes : 7.5862 fps
      Spin wire view, 1000 monkes : 16.7646 fps
      OpenGL image load & free, 256x256 px : 500.1257 fps
      OpenGL image load & free, 512x512 px : 181.8605 fps
      OpenGL image load & free, 1024x1024 px : 54.9584 fps
      OpenGL image load & free, 2048x2048 px : 13.1438 fps
      Raytracing with AO and area light, 2 threads : 6.87 sec
      Shadowbuffer light, 1 threads : 14.20 sec

      Benchmark Results, Screen Size 1920 x 1061

      Overall Score (FPS)
      gl : 23.6326 fps
      render : 9.86 sec

      Spin wireframe view, subsurf monkey, 4 subsurf levels : 53.0788 fps
      Spin solid view, subsurf monkey, 4 subsurf levels : 29.2730 fps
      Spin solid view, 1000 monkes : 7.6336 fps
      Spin wire view, 1000 monkes : 16.9094 fps
      OpenGL image load & free, 256x256 px : 526.4266 fps
      OpenGL image load & free, 512x512 px : 188.7218 fps
      OpenGL image load & free, 1024x1024 px : 54.9580 fps
      OpenGL image load & free, 2048x2048 px : 14.2892 fps
      Raytracing with AO and area light, 1 threads : 6.87 sec
      Shadowbuffer light, 8 threads : 12.85 sec

      Benchmark Results, Screen Size 1920 x 1061

      Overall Score (FPS)
      gl : 23.3041 fps
      render : 10.96 sec

      Spin wireframe view, subsurf monkey, 4 subsurf levels : 52.9383 fps
      Spin solid view, subsurf monkey, 4 subsurf levels : 29.2563 fps
      Spin solid view, 1000 monkes : 7.6252 fps
      Spin wire view, 1000 monkes : 16.8076 fps
      OpenGL image load & free, 256x256 px : 500.1257 fps
      OpenGL image load & free, 512x512 px : 181.8605 fps
      OpenGL image load & free, 1024x1024 px : 54.0669 fps
      OpenGL image load & free, 2048x2048 px : 13.5902 fps
      Raytracing with AO and area light, 1 threads : 6.87 sec
      Shadowbuffer light, 4 threads : 15.05 sec
    • ekopath
      Benchmark Results, Screen Size 1920 x 1061

      Overall Score (FPS)
      gl : 23.7918 fps
      render : 10.73 sec

      Spin wireframe view, subsurf monkey, 4 subsurf levels : 55.1772 fps
      Spin solid view, subsurf monkey, 4 subsurf levels : 30.4204 fps
      Spin solid view, 1000 monkes : 7.9725 fps
      Spin wire view, 1000 monkes : 17.8420 fps
      OpenGL image load & free, 256x256 px : 476.3061 fps
      OpenGL image load & free, 512x512 px : 181.8605 fps
      OpenGL image load & free, 1024x1024 px : 52.9229 fps
      OpenGL image load & free, 2048x2048 px : 12.9397 fps
      Raytracing with AO and area light, 2 threads : 6.82 sec
      Shadowbuffer light, 1 threads : 14.65 sec

      Benchmark Results, Screen Size 1920 x 1061

      Overall Score (FPS)
      gl : 23.9342 fps
      render : 9.81 sec
      Spin wireframe view, subsurf monkey, 4 subsurf levels : 55.4491 fps
      Spin solid view, subsurf monkey, 4 subsurf levels : 30.4848 fps
      Spin solid view, 1000 monkes : 7.9918 fps
      Spin wire view, 1000 monkes : 17.8225 fps
      OpenGL image load & free, 256x256 px : 500.1257 fps
      OpenGL image load & free, 512x512 px : 181.8613 fps
      OpenGL image load & free, 1024x1024 px : 52.9227 fps
      OpenGL image load & free, 2048x2048 px : 13.1958 fps
      Raytracing with AO and area light, 2 threads : 6.87 sec
      Shadowbuffer light, 8 threads : 12.75 sec

      Benchmark Results, Screen Size 1920 x 1061

      Overall Score (FPS)
      gl : 23.7750 fps
      render : 10.32 sec
      Spin wireframe view, subsurf monkey, 4 subsurf levels : 55.3894 fps
      Spin solid view, subsurf monkey, 4 subsurf levels : 30.4539 fps
      Spin solid view, 1000 monkes : 7.9685 fps
      Spin wire view, 1000 monkes : 17.7847 fps
      OpenGL image load & free, 256x256 px : 500.1257 fps
      OpenGL image load & free, 512x512 px : 178.6163 fps
      OpenGL image load & free, 1024x1024 px : 52.9230 fps
      OpenGL image load & free, 2048x2048 px : 12.9230 fps
      Raytracing with AO and area light, 2 threads : 6.87 sec
      Shadowbuffer light, 2 threads : 13.76 sec

    Comment


    • #3
      It may be interesting to test it with llvmpipe, which should be more CPU bound.

      Comment


      • #4
        I could get llvmpipe to compile with EKO.

        Comment


        • #5
          All results are invalid, both mesa were compiled by gcc (but with different flags), I will redo everything and post again.

          Comment


          • #6
            This time I make sure it's using the right binary. A side note today mesa's git seems slow on max fps.


            gcc

            49.344784 / 114.962814 / 214.564774 fps -- Score = 79.898849 rand[1450720973]
            45.837265 / 115.310844 / 195.620300 fps -- Score = 79.830368 rand[1450720973]
            45.391266 / 114.646248 / 198.058014 fps -- Score = 79.811752 rand[1450720973]

            ekopath

            47.145302 / 108.609818 / 196.764694 fps -- Score = 79.751328 rand[1450720973]
            47.171112 / 108.511078 / 198.167053 fps -- Score = 79.770752 rand[1450720973]
            47.352142 / 108.524216 / 199.882401 fps -- Score = 79.770531 rand[1450720973]

            Comment


            • #7
              Maybe you should add --disable-asm to the gcc build also, to not give it an advantage. Eventually the asm could be fixed to support also eko (or the other way).

              Comment


              • #8
                Originally posted by oibaf View Post
                Maybe you should add --disable-asm to the gcc build also, to not give it an advantage. Eventually the asm could be fixed to support also eko (or the other way).
                Yes, I thought it was a typo. If gcc is building the asm bits while eko is not, isn't this comparing apples to oranges? It's useful to know what, as of now, provides the biggest absolute performance, but it doesn't really compare the code produced by both compilers, does it? Then, if the numbers are right, would this mean ekopath makes the asm parts somewhat redundant?

                Comment


                • #9
                  Updated info about the builds. The following setting were used for those last numbers.

                  gcc 4.5.2
                  CFLAGS="-O2 -march=native -pipe -fgraphite-identity -floop-interchange -floop-strip-mine -floop-block -ffast-math"
                  --disable-asm

                  ekopath 4.0.10 (binary)
                  CFLAGS="-O3 -ffast-math -pipe"
                  --disable-asm

                  Comment

                  Working...
                  X