Announcement

Collapse
No announcement yet.

Gallium3D LLVMpipe Starts To Smoke

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

  • Gallium3D LLVMpipe Starts To Smoke

    Phoronix: Gallium3D LLVMpipe Starts To Smoke

    While it's rare for a few days to pass at Phoronix without pulling the latest development code for Mesa / Gallium3D and the Linux kernel DRM in order to run updated Radeon, Intel, and Nouveau Linux graphics benchmarks, LLVMpipe isn't benchmarked as commonly. LLVMpipe is the new CPU-based software acceleration method for the Gallium3D that leverages the Low-Level Virtual Machine to provide better performance than the classic Mesa software rasterizer or Softpipe. Fortunately, upon running a brand new set of tests, the results show a bit more promise but there is still much work ahead.

    http://www.phoronix.com/vr.php?view=16453

  • #2
    It's nice to see the state of llvmpipe, and it's very impressive for a purely software renderer. But how on Earth did you manage to get 20 fps with HD 5450 in OpenArena?

    I get close to 60fps in 1080p with an HD 4550 (which is about equal in terms of performance), and this has been the case for about a year. 25fps at 1024x768 is something I can't achieve even with low power profile and anti-performance tweaks.

    Since all tests but one are capped at exactly 60 fps, I'm guessing it's the vsync issue again.

    Comment


    • #3
      It would be interesting to test this on the 6core i7 again and compare against the last test to see the improvements

      Comment


      • #4
        Why comparing LLVMpipe (which is a software pipe) to RadeonHD and Nouveau (which are hardware accelerated pipes) ?
        Why not comparing LLVMpipe to Mesa, to see the real improvements between the two solutions?

        These are real questions, I might have misunderstood some subtlety about the rendering pipe...

        Comment


        • #5
          Originally posted by FireBurn View Post
          It would be interesting to test this on the 6core i7 again and compare against the last test to see the improvements
          +1
          Compare Mesa and LLVMpipe with CPU with more and more cores, to see if LLVM handles it better than GCC. (but I make a lot of assumptions here, correct me if I'm wrong)

          Comment


          • #6
            Originally posted by Creak View Post
            Why comparing LLVMpipe (which is a software pipe) to RadeonHD and Nouveau (which are hardware accelerated pipes) ?
            Why not comparing LLVMpipe to Mesa, to see the real improvements between the two solutions?

            These are real questions, I might have misunderstood some subtlety about the rendering pipe...
            I agree, dedicated hardware should always win

            Comment


            • #7
              Originally posted by FireBurn View Post
              I agree, dedicated hardware should always win
              That first OpenArena test looks a bit suspicious to me. Not in the testing methodology, but that we may have an opportunity to identify a CPU bottleneck in the Gallium/LLVM/Radeon architecture.

              One of these days I'm hoping to starting profiling the r600 driver to see if there's any way I can help. Otherwise, I'll probably start working on fleshing out the GSoC clover implementation a bit (implementing built-in math functions and stuff, conformance testing, etc).

              Comment


              • #8
                BTW... I'd like to buy a new graphic card, is it good to take an ATI now? I prefer their "open" policy, but I don't know if the open source driver is good enough.
                For instance, I have an nVidia card at the moment and if I use Nouveau, the temperature gets higher and higher, way above 100C.
                It took me some time to understand why my PC was freezing without any reason...

                Since RadeonHD drivers are based on open specs and aren't retro-engineered, are they more stable?

                I don't play huge games on my Linux, but I do 3D programming.

                Comment


                • #9
                  Originally posted by Veerappan View Post
                  One of these days I'm hoping to starting profiling the r600 driver to see if there's any way I can help
                  I'd be glad to help you in this as soon as I've got an ATI card (see previous message).

                  Comment


                  • #10
                    I continue to wonder if this will ever breathe any life into the non-T&L Radeon Xpress 1100/200M

                    I suppose the desired behavior would be to fall back to LLVMPipe instead of the current soft pipe for operations not supported by the GPU.

                    Comment


                    • #11
                      Doesn't the r300g driver already do that ?

                      Comment


                      • #12
                        Originally posted by Veerappan View Post
                        One of these days I'm hoping to starting profiling the r600 driver to see if there's any way I can help. Otherwise, I'll probably start working on fleshing out the GSoC clover implementation a bit (implementing built-in math functions and stuff, conformance testing, etc).
                        Nothing jumped at me when profiling the PP shaders, but then again the Brazos platform is very balanced.

                        Comment


                        • #13
                          Originally posted by Creak View Post
                          BTW... I'd like to buy a new graphic card, is it good to take an ATI now? I prefer their "open" policy, but I don't know if the open source driver is good enough.
                          For instance, I have an nVidia card at the moment and if I use Nouveau, the temperature gets higher and higher, way above 100C.
                          It took me some time to understand why my PC was freezing without any reason...

                          Since RadeonHD drivers are based on open specs and aren't retro-engineered, are they more stable?

                          I don't play huge games on my Linux, but I do 3D programming.
                          I wouldn't hesitate to go with an AMD card at the moment. Up until 2 months ago, I was running a Radeon 4770 on the open source drivers without any issues. I upgraded to a Radeon 6850 last month, and for work reasons I need to use the Catalyst drivers (paid OpenCL development). As soon as Clover is capable of doing what I need it to do, I'll switch back to r600g+clover.

                          For desktop usage and lightweight 3D, there's nothing wrong with the OSS radeon drivers. If you need GL 3/4 features, you might need to use the Catalyst drivers. Both sets of drivers have been quite stable for me in Ubuntu 11.04 and Windows 7 x64.

                          Comment


                          • #14
                            Originally posted by Creak View Post
                            Why comparing LLVMpipe (which is a software pipe) to RadeonHD and Nouveau (which are hardware accelerated pipes) ?
                            Why not comparing LLVMpipe to Mesa, to see the real improvements between the two solutions?
                            Because that would be a worthless comparison. We already know that Mesa3D is slow and pretty much absolutely useless for anything other than being a reference-test renderer. Comparing the speed of anything to that would be like comparing the land speed of a vehicle to that of an office building.

                            Showing the speed of LLVMPipe on its own is useful. We can see that it hits playable framerates in some cases. Putting the hardware results in there helps give you an idea of just how fast LLVMPipe is, e.g. that it's comparable in speed to a low-end hardware renderer (well, with the FOSS drivers, at least). That's actually meaningful. That tells you that LLVMPipe is serious stuff. You now both have solid numbers telling you if LLVMPipe is fast enough for your needs and you have a meaningful reference point for just how fast it is compared to buying an extra piece of GPU hardware.

                            Comment


                            • #15
                              Originally posted by elanthis View Post
                              Because that would be a worthless comparison. We already know that Mesa3D is slow and pretty much absolutely useless for anything other than being a reference-test renderer. Comparing the speed of anything to that would be like comparing the land speed of a vehicle to that of an office building.

                              Showing the speed of LLVMPipe on its own is useful. We can see that it hits playable framerates in some cases. Putting the hardware results in there helps give you an idea of just how fast LLVMPipe is, e.g. that it's comparable in speed to a low-end hardware renderer (well, with the FOSS drivers, at least). That's actually meaningful. That tells you that LLVMPipe is serious stuff. You now both have solid numbers telling you if LLVMPipe is fast enough for your needs and you have a meaningful reference point for just how fast it is compared to buying an extra piece of GPU hardware.
                              IMHO, if you have an i5 with 4 cores and more than 3GHz per core, you are not likely to have a slower GPU than what you can achieve using LLVMpipe. Neither are you using (almost sure) a bad supported GPU, like an Unichrome. I'm not even sure if a VIA chipset from that ages can support an i5 processor.
                              So, the comparison made in the benchmark is still useless.

                              Comment

                              Working...
                              X