Announcement

Collapse
No announcement yet.

Valve Lands A Last Minute AMD ACO Improvement For Mesa 24.0

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

  • Valve Lands A Last Minute AMD ACO Improvement For Mesa 24.0

    Phoronix: Valve Lands A Last Minute AMD ACO Improvement For Mesa 24.0

    This week prior to the Mesa 24.0 feature freeze / code branching, a notable merge request landed that had been worked on the past few months by one of Valve's open-source Linux graphics driver developers...

    Phoronix, Linux Hardware Reviews, Linux hardware benchmarks, Linux server benchmarks, Linux benchmarking, Desktop Linux, Linux performance, Open Source graphics, Linux How To, Ubuntu benchmarks, Ubuntu hardware, Phoronix Test Suite

  • #2
    It's just hilarious when game shaders compile faster with ACO on your mobile APU than on your 13700K RTX 4090 Windows system.
    Not to mention AMD D3D12/amdvlk LLPC. And I guess ACO RustiCL support is also just a matter of time, which probably will make ROCm (which btw. just hangs my 7800 XT) look bad too...

    Comment


    • #3
      Originally posted by aufkrawall View Post
      It's just hilarious when game shaders compile faster with ACO on your mobile APU than on your 13700K RTX 4090 Windows system.
      Not to mention AMD D3D12/amdvlk LLPC. And I guess ACO RustiCL support is also just a matter of time, which probably will make ROCm (which btw. just hangs my 7800 XT) look bad too...
      What's also hilarious is that shader compilation for AMD Linux users is only getting faster (due to continuous ACO improvements and RadeonSi also getting ACO soon as well), while for AMD Windows users it's only getting slower probably because LLPC has started replacing AMD's proprietary compiler in the Windows drivers. I've seen so many complains about game stutters with the new DXNAVI DX 9-11 drivers. I've also personally tested the new OpenGL Windows driver and I felt that shader compilation speed is slower than with the old GL driver.

      Comment


      • #4
        AMD should just open source their Windows driver too at this point...

        Comment


        • #5
          Originally posted by OneTimeShot View Post
          AMD should just open source their Windows driver too at this point...
          And no one will support it...

          Comment


          • #6
            I'm confused.. Isn't GPL the new default and a performance optimized replacement for ACO in ​RADV_PERFTEST​ with RADV?

            Comment


            • #7
              Kjell No. RADV_PERFTEST can toggle different features and not just the shader compiler. GPL ("graphics pipeline libraries" if you want to read up on it) is something else than ACO entirely and its main use is for reducing hitching in DXVK/VKD3D and maybe game console emulators.
              Last edited by kiffmet; 13 January 2024, 09:50 PM.

              Comment


              • #8
                As I understand it, this is about increasing resource utilisation by fetching data from memory earlier so you don't sit around twiddling as much.
                GPL is about using new parts of the geometry engine to increase resource throughput when you have the data available.

                As you make something faster, the bottleneck tends to move to somewhere else. Optimization is an eternal game of whack-a-mole

                Comment


                • #9
                  Originally posted by OneTimeShot View Post
                  AMD should just open source their Windows driver too at this point...
                  amdvlk is the same vulkan driver that is used in windows.

                  Comment


                  • #10
                    How hard is it for AMD to join efforts with radv and replace amdvlk/llpc with it? Can they make radv work on Windows or it's too Linux specific?

                    And by the way, Intel handles anv on Linux separately from their Windows driver somehow and doesn't sweat about it. So at least in theory AMD could do the same thing at least anywhere where radv/aco could work.

                    Comment

                    Working...
                    X