Announcement

Collapse
No announcement yet.

AMDGPU Has Another Round Of Updates Ahead Of Linux 5.2

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

  • #21

    Yesterday I got my hands on a (single) 144hz freesync monitor.
    ​​​​​​​Pluged it in and got an flickering nightmare. Tried with DVI, HDMI and DP. No luck.

    Deleted amd* boot options and all work! Got 144hz w freesync now
    ​​​​

    Comment


    • #22
      Originally posted by Etherman View Post
      Deleted amd* boot options and all work! Got 144hz w freesync now
      ​​​​
      Because your VRAM likely runs constantly with full clocks (which is the "viable" workaround for this issue).

      It sucks that this issue exists, but it's probably hardware related and their R&D fund is quite limited, compared to other firms. I'll join the complaints however, if Navi would be a step backwards compared to Vega in this matter.

      Comment


      • #23
        Originally posted by Space Heater View Post
        Yeah it's pretty ridiculous at this point, but I don't think the kernel developers are ignoring these issues because they just don't care. Rather I strongly believe that management behind the scenes is (as usual) driving priorities and is focusing on new hardware enablement. It's very possible that the developers are not given any time to work on (critical) longstanding issues like this, so they can only do give attention to them in their spare time.
        Didn't they just post openings for like a dozen developers, a couple months ago? It will take time to hire and train those folks, but their Linux GPU team will hopefully have noticeably more bandwidth.

        Comment


        • #24
          Originally posted by clapbr View Post

          afaik on Polaris dpm=1 is default. I'd love to run it at full memory clock reliably but even for that I need workarounds for anything that trigger modesets since any clock changes reset then (resolution changes, display on/off, .sleep/suspend, X session restart etc). So I have 3 stupid options: hook all those edge cases, constantly flood sysfs with mclock changes (which triggers random hangs) or constantly poll the mclock and fix it when it resets.
          Have you tried overclocking the lowest memory state a little bit? It could be as simple as bumping it from 300mhz to like 500mhz...and I doubt you'd even need to up the voltage any...maybe 5-10mv at most.

          My 580 has the opposite problem -- it has high voltages, gets hot fast, and throttles down which causes random issues, especially when gaming, but ProjectM and other gpu intensive tasks can trigger it too. I've undervolted my highest five states to 1075mv, "overclocked" the memory to the 2000mhz that most other 580s use, and undervolted it's highest state a hair down to 875mv. Haven't needed to bother with anything else though I know I could get it a tiny bit more power efficient.

          After all of that it went from hitting 80C-85C and throttling to maxing at 70C without throttling and fixed all my clocking issues that caused other stuff to act screwy. I also enable dpm=1 just in case...it's weird, but sometimes just having the supposed default value for something explicitly set can help.

          Here's my current kernel command line if it'll help: "zswap.enabled=1 zswap.compressor=lz4 zswap.max_pool_percent=25 zswap.zpool=z3fold amdgpu.dpm=1 amdgpu.dc=1 amdgpu.ppfeaturemask=0xfffd7fff zfs.zfs_arc_max=8589934592"

          Comment


          • #25
            phoronix

            What did I type or do to trigger that unapproved post?

            Comment


            • #26
              Originally posted by coder View Post
              Didn't they just post openings for like a dozen developers, a couple months ago? It will take time to hire and train those folks, but their Linux GPU team will hopefully have noticeably more bandwidth.
              While I do hope these new hires will be helping with general amdgpu work and bugfixes, the job openings were specifically for machine learning and gpu compute if I remember correctly. So it's possible that they will be purely focused on enterprise workloads, and will not be assigned to improve desktop/laptop use-cases for amdgpu.

              Comment


              • #27
                Originally posted by skeevy420 View Post
                phoronixWhat did I type or do to trigger that unapproved post?
                It's probably a black box to Michael.

                The spam filter is garbage. I hope it at least catches a fair amount real spam. Its sensitivity doesn't seem at all dependent on the user, for one thing. You'd think it would be more permissive for users with high post counts and zero true positives.

                Comment


                • #28
                  Originally posted by Space Heater View Post
                  While I do hope these new hires will be helping with general amdgpu work and bugfixes, the job openings were specifically for machine learning and gpu compute if I remember correctly. So it's possible that they will be purely focused on enterprise workloads, and will not be assigned to improve desktop/laptop use-cases for amdgpu.
                  IIRC, bridgman said they were actually broader than that description. Also, hiring folks to work in those areas could free up bandwidth in the rest of the team to focus on more basic features and functionality.

                  Comment


                  • #29
                    Originally posted by Space Heater View Post
                    While I do hope these new hires will be helping with general amdgpu work and bugfixes, the job openings were specifically for machine learning and gpu compute if I remember correctly. So it's possible that they will be purely focused on enterprise workloads, and will not be assigned to improve desktop/laptop use-cases for amdgpu.
                    It is a long learning curve to get into the amdgpu driver code. You need to study hardware details of every SOC and there are thousands of files, often implementing almost the same feature. It is easy to copy an old driver code to a new but that increases the code complexity and side effects when you forget to update all affected files. We see that constantly in bug reports. Object oriented software design would simplify a lot, but they are C programmers.

                    Comment


                    • #30
                      Originally posted by skeevy420 View Post
                      amdgpu.dpm=1 amdgpu.dc=1
                      Those are enabled by default for Polaris. amdgpu.ppfeaturemask=0xfffd7fff is needed to make powerplay work at 4K60Hz. My Asus Expedition OC RX 570 4GB works fine with the AMD wip kernel.

                      Comment

                      Working...
                      X