Announcement

Collapse
No announcement yet.

AMD's A-Sync DMA Code Makes For Fast Performance

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

  • AMD's A-Sync DMA Code Makes For Fast Performance

    Phoronix: AMD's A-Sync DMA Code Makes For Fast Performance

    After the benchmarks of the Radeon Gallium3D sub-allocator that in some tests yields more than a 25% performance boost, initial testing was done of the new AMD a-sync DMA engine support for the open-source Radeon driver...

    http://www.phoronix.com/vr.php?view=MTI1MDE

  • #2
    Damn! I can't wait to see all of the performance improvement for this one....

    But I wonder how bad the regressions, if any, are....

    Great work guys!

    Comment


    • #3
      That's great. If driver will be tweaked next 5 years it will maybe become usable.

      Comment


      • #4
        Additional to that it is quite a bit more efficient than the shader engine when you just want to copy some data from A to B, or just clear a specific region of memory (memcpy/memset).
        Yay! The blitter is back!

        Comment


        • #5
          Originally posted by JS987 View Post
          That's great. If driver will be tweaked next 5 years it will maybe become usable.
          It's quite usable already. Even better is, you can run the radeon driver when you're on bleeding edge.

          Comment


          • #6
            So yeah, Unigine went from 2.7 to 23fps, which is a huge improvement. I started the gaming-free test suite earlier on a 3.7 kernel, and I should have the drm-next results after work. (I believe that includes nexuiz, urban terror, smoking guns, warsow, padman, and maybe a few others).

            I'm not sure how much of this improvement is due solely to the Async DMA engine, and how much is other improvements between 3.7 and drm-next (I noticed a few other things in the changelog that might help). The kernel is the ONLY thing that has changed between runs, and honestly I don't care so much about the source of the improvement as that it's there at all. It would be possible to build the kernel from git using only this patch series as the difference, but for now I'm using the Ubuntu daily drm-next build from 12/11/2012

            Comment


            • #7
              Well, I read in this article: http://www.phoronix.com/scan.php?pag...reaction&num=1
              a report by a user saying that his Unigine Heaven FPS frame-rate on R600g dropped from about 25 FPS to just 3 FPS following the "fix abysmal performance" patch.
              A now we return to 23 fps...

              It's been years we wait the open source driver to compete with the proprietary one.
              Maybe it is an unending wait.

              I hope they will at least code all the missing Opengl functions.

              Comment


              • #8
                Catching up

                Maybe the open source driver is catching up with the proprietary driver.

                I would like to see a open source vs proprietary driver benchmark.

                Comment


                • #9
                  Originally posted by mannerov View Post
                  Well, I read in this article: http://www.phoronix.com/scan.php?pag...reaction&num=1


                  A now we return to 23 fps...
                  This patch:
                  http://cgit.freedesktop.org/~agd5f/l...f65bd4be512627
                  should also improve things above and beyond async DMA.

                  Comment


                  • #10
                    WOW! I am testing it!
                    Ungine - 27

                    And OilRush I can play on ultra quality with hight texture 32-55 FPS!

                    Michael Larabel you need new complex test!

                    Comment


                    • #11
                      Fun times ahead it seems. Will be interesting to see where things are at when the dust settles on all of these patches.

                      Comment


                      • #12
                        I've finished the tests that I was planning on running. Heaven improved the most, but when I ran Xonotic at 1920x1200 at the highest settings, the framerate went from 9fps to 29... Not bad for just a kernel upgrade with a few hundred lines of changed code. I'm assuming that running Xonotic at max settings caused my GPU to run out of available VRAM and cause lots of swapping between system/vram, which is something that this new feature should help out with.

                        The tests include:
                        Heaven
                        Nexuiz
                        OpenArena
                        World of Padman
                        Smoking Guns
                        Tremulous
                        Urban Terror
                        Warsow
                        Reaction Quake 3
                        Xonotic

                        http://openbenchmarking.org/result/1...SU-ASYNCDMAT84

                        Comment


                        • #13
                          Originally posted by Veerappan View Post
                          **tests**
                          Correct URL: http://openbenchmarking.org/result/1...SU-ASYNCDMAT84

                          Thanks for testing!!

                          It seems opensource has ~60% performance already.
                          Does your driver already have Marek HyperZ patches?

                          I assume Xonotic uses more modern OpenGL features, that are done in software or ignored now - hence low fps.
                          You can prove me right or wrong, if you test OpenArena 0.8.5 (old GL) and compare how it scales to 0.8.8 (more modern GL).

                          PS
                          You can safely drop Urban Terror, its bottlenecking piece of dust.

                          PPS
                          I wonder how it scales on 5850/70 // 6950/70 hardware...
                          Thanks again!

                          Comment


                          • #14
                            Originally posted by crazycheese View Post
                            It seems opensource has ~60% performance already.
                            60% of what? What baseline are you using?

                            Comment


                            • #15
                              Originally posted by crazycheese View Post
                              Correct URL: http://openbenchmarking.org/result/1...SU-ASYNCDMAT84
                              It seems opensource has ~60% performance already.
                              Does your driver already have Marek HyperZ patches?
                              No, this doesn't have the HyperZ patches applied. This was as of the following commit (http://cgit.freedesktop.org/mesa/mes...it/?id=3392f2f)

                              I honestly don't remember if this had the suballocator patches hand-applied before I started, but I think it did. I can check again by re-running the reaction quake tests on the old kernel and checking the performance against my original reaction quake 3 tests.

                              Originally posted by crazycheese View Post
                              I assume Xonotic uses more modern OpenGL features, that are done in software or ignored now - hence low fps.
                              You can prove me right or wrong, if you test OpenArena 0.8.5 (old GL) and compare how it scales to 0.8.8 (more modern GL).
                              Running that currently. Probably won't have the results posted until morning.

                              Originally posted by crazycheese View Post
                              PPS
                              I wonder how it scales on 5850/70 // 6950/70 hardware...
                              Thanks again!
                              Couldn't tell you. The only other radeon hardware I've got is the HD4200 IGP built into my motherboard, and the Llano chip in my HTPC (which I'm not especially thrilled about messing with the software installation on).

                              Comment

                              Working...
                              X