AMD Sends In Their Initial AMDGPU Driver Updates For Linux 5.2

[Kraut]

R9 290 has 320GB/s of memory bandwidth, with lower mclk it is 32GB/s if I am correct. Assuming a 4K monitor, 3840*2160*4*120 = 4GB/s. In theory, the lower mclk has enough bandwidth to feed multiple 4K 120Hz displays. In practice, it may not be enough due to memory contention and overheads, so testing is needed to tell whether it works.

This is not about bandwidth. This is about the constant time it takes for the power mode (at last for vram) to switch. If your monitors vblank period is shorter then that you can get flickering each time the power mode changes.

[pixo]

As for the modesetting DDX - what GPU are you using? If you use modesetting for AMD, I believe you lose features, and I'm not sure what the 3D acceleration situation would be. I used to use the modesetting driver when I had an Intel GPU - and it ( and therefore glamor ) certainly worked well in that setup. This is definitely an AMD driver issue ... but as I've just discovered, only under X.

Adaptive-sync is not supported by modesetting DDX but that is all i am aware of.
Patched xserver using modesetting DDX can use DRM leases for multiseat with hw acceleration for every seat via 1 GPU.
That does not work for amdgpu DDX and I could not make it work. Will give it another shot this weekend.

[MrCooper]

I've seen the same on my new Ryzen / Vega 10 laptop. New gtk3 windows briefly render a window full of garbage before behaving properly. I'm also seeing rendering glitches in gtktextview and gtksourceview widgets, which is fixed when I drag-select the content.

Like https://bugs.freedesktop.org/110214 ?

Not sure if it's been reported, but it's hard to see how they wouldn't be aware of it ...

If something isn't reported, the default assumption has to be nobody else is aware of it.

Are you also using Arch Linux? Is anyone using a different distro seeing these issues? Just to be clear, I'm not trying to blame it on Arch, just narrowing down the circumstances, as it's not happening for everybody. I wonder if it might be related to building xserver with meson, not sure any distro other than Arch is doing that yet.

[MrCooper]

Adaptive-sync is not supported by modesetting DDX but that is all i am aware of.

There's also TearFree, just for one more example. The amdgpu driver is generally more polished and robust.

Patched xserver using modesetting DDX can use DRM leases for multiseat with hw acceleration for every seat via 1 GPU.
That does not work for amdgpu DDX and I could not make it work.

That requires porting at least the DRM_IOCTL_CRTC_QUEUE_SEQUENCE related changes from modesetting as well, maybe more.

[atomsymbol]

This is not about bandwidth. This is about the constant time it takes for the power mode (at last for vram) to switch. If your monitors vblank period is shorter then that you can get flickering each time the power mode changes.

I don't understand how [the constant time it takes for the power mode (at last for vram) to switch] is preventing you to manually select the lowest mclk+sclk for non-demanding desktop work with a mid-range or high-end AMD GPU.

[Kraut]

I don't understand how [the constant time it takes for the power mode (at last for vram) to switch] is preventing you to manually select the lowest mclk+sclk for non-demanding desktop work with a mid-range or high-end AMD GPU.

Sure, if you manually overwrite the auto power management and stick to lower power mode it should work fine. (If not that is a driver bug or developer oversight)
It's just the default auto power management setting that force high power mode when switching could cause flickering on your connected screens.

[Cuthred O'Reilly]

Sure, if you manually overwrite the auto power management and stick to lower power mode it should work fine. (If not that is a driver bug or developer oversight)
It's just the default auto power management setting that force high power mode when switching could cause flickering on your connected screens.

My apologies for the long post; I'm not yet smart enough to handle attachments or links.

I have an RX 580 with 2 monitors, primary @ 75Hz & secondary @ 60Hz. I have ppfeaturemask set to allow undervolting which is very beneficial for the RX 580. For me, the flickering from automatic memory p-state switching is unbearable; the manual workaround is essential.

Leaning heavily on the thought process behind grmat's amdgpu-fancontrol I've written a similar script to manage the memory p-states. This switches between memory p-states based upon gpu_busy_percent, but only does so very reluctantly. It works well for my situation.

By the way, I'm a long-time lurker; Phoronix is a great resource.

#!/bin/bash

# Copyright George Scorer, March 2019.
# You may copy and/or modify and/or use this script as you see fit. There is no warranty of any kind.

# Each memory p-state switch causes a screen flicker. Tweak these variables to match
# your personal 'flicker aversion' vs 'efficiency and silence desire' trade-off.
# These variables are integers.
GPU_BUSY_UP=80                 # The gpu_busy_percent at and above which we consider the gpu is busy
GPU_BUSY_DOWN=10               # The gpu_busy_percent at and below which we consider the gpu is idle
LOW_MEM_STATE=0                # Choose between 0 & 1
STANDARD_SLEEP_INTERVAL=1      # in seconds. How frequently we should poll the core p-state.
DOWN_DELAY=10                  # in seconds. How long the gpu should appear to be idle before we drop down.
# These variables are floating point.
CONCENTRATE_SLEEP_INTERVAL=0.1 # in seconds. How frequently we should poll when we're proposing to jump up.
UP_DELAY=0.5                   # in seconds. How long the gpu should appear to be busy before we jump up.

# Sysfs paths here are hardcoded for one amdgpu card at card0; adjust as needed.
FILE_PERF_LEVEL=/sys/class/drm/card0/device/power_dpm_force_performance_level
FILE_MEM_P_STATE=/sys/class/drm/card0/device/pp_dpm_mclk
FILE_GPU_BUSY_PERCENT=/sys/class/drm/card0/device/gpu_busy_percent


# check for root privileges
if [ $UID -ne 0 ]
then
  echo "Writing to sysfs requires root privileges; relaunch as root"
  exit 1
fi

function get_started {
# Set gpu performance level control to manual
# echo "Setting performance level control to manual"
  echo "manual" > "$FILE_PERF_LEVEL"

# Read the current gpu busy percentage and set a corresponding initial memory p-state

  GPU_BUSY_PERCENT="$(<$FILE_GPU_BUSY_PERCENT)"

  if [ "$GPU_BUSY_PERCENT" -ge "$GPU_BUSY_UP" ]; then
    MEM_P_STATE=2
  else
    MEM_P_STATE=$LOW_MEM_STATE
  fi

  echo "$MEM_P_STATE" > "$FILE_MEM_P_STATE"
  echo "New gpu memory p-state: $MEM_P_STATE"
  PROPOSED_MEM_P_STATE=$MEM_P_STATE
  SLEEP_INTERVAL=$STANDARD_SLEEP_INTERVAL
}


function check_gpu_busy_percent {

  GPU_BUSY_PERCENT="$(<$FILE_GPU_BUSY_PERCENT)"

# Propose what the corresponding memory p-state should be
  OLD_PROPOSED_MEM_P_STATE=$PROPOSED_MEM_P_STATE
  PROPOSED_MEM_P_STATE=$MEM_P_STATE
  if [ "$GPU_BUSY_PERCENT" -ge "$GPU_BUSY_UP" ]; then
    PROPOSED_MEM_P_STATE=2
  elif [ "$GPU_BUSY_PERCENT" -le "$GPU_BUSY_DOWN" ]; then
    PROPOSED_MEM_P_STATE=$LOW_MEM_STATE
  fi

  if [ "$PROPOSED_MEM_P_STATE" -eq "$MEM_P_STATE" ]; then
#   We're happy in this memory state - go back to sleep
    SLEEP_INTERVAL=$STANDARD_SLEEP_INTERVAL
  else
#   We want to change so determine whether we're already counting down.    
    if [ "$PROPOSED_MEM_P_STATE" -ne "$OLD_PROPOSED_MEM_P_STATE" ]; then
#     Start a new countdown
      if [ "$PROPOSED_MEM_P_STATE" -eq 2 ]; then
#       We might be going up; check more frequently to avoid jumping on a spike. (bad)
        CHANGE_COUNTDOWN=$UP_DELAY
        SLEEP_INTERVAL=$CONCENTRATE_SLEEP_INTERVAL
      else
        CHANGE_COUNTDOWN=$DOWN_DELAY
        SLEEP_INTERVAL=$STANDARD_SLEEP_INTERVAL
      fi
    else
      CHANGE_COUNTDOWN=$( echo "$CHANGE_COUNTDOWN - $SLEEP_INTERVAL" | bc )
    fi
      
    if ((  $( echo "$CHANGE_COUNTDOWN <= 0.0" | bc ) )); then
#   The countdown has reached 0 so change the memory p-state.
      MEM_P_STATE=$PROPOSED_MEM_P_STATE
#      echo "manual" > "$FILE_PERF_LEVEL"
      echo "$MEM_P_STATE" > "$FILE_MEM_P_STATE"
      echo "New gpu memory p-state: $MEM_P_STATE"
      
    fi
  fi

#   echo "Old  Prop  Mem  Busy %  Countdown"
#   echo "  $OLD_PROPOSED_MEM_P_STATE     $PROPOSED_MEM_P_STATE    $MEM_P_STATE       $GPU_BUSY_PERCENT         $CHANGE_COUNTDOWN"
  }

function reset_on_fail {
  echo "Exiting, setting memory p-state to 2"
  echo "manual" > "$FILE_PERF_LEVEL"
  echo "2" > "$FILE_MEM_P_STATE"
  exit 1
}

# always try to fix memory p-state 2 on failure
trap "reset_on_fail" SIGINT SIGTERM

function run_daemon {
  while :; do
#   If mem_p_state in sysfs differs from our calculated MEM_P_STATE then restart.
#     Causes: external change to power_dpm_force_performance_level, system sleep, a change to display configuration or a commit of a gpu voltage or clock frequency change.
#     Note: we're only testing once every $SLEEP_INTERVAL seconds so it might take us a long time to spot a momentary change in mem_p_state.
    SYS_MEM_P_STATE="$(grep -F '*' $FILE_MEM_P_STATE)"
    SYS_MEM_P_STATE=${SYS_MEM_P_STATE:0:1}

    if [[ $(< $FILE_PERF_LEVEL) != "manual" ]]; then
      echo "Something's changed memory performance level; taking back control."
      get_started
    fi
    if [[ $MEM_P_STATE != $SYS_MEM_P_STATE ]]; then
      echo "Something's meddling with gpu memory p-states; taking back control."
      get_started
    fi
    sleep $SLEEP_INTERVAL
    check_gpu_busy_percent
  done
}

get_started

# start the loop
run_daemon

[dwagner]

Sure, if you manually overwrite the auto power management and stick to lower power mode it should work fine. (If not that is a driver bug or developer oversight)

The manual sclk/mclk overwrite used to work in some kernel versions, then broke, worked again, broke again, currently it is broken for about the last 4 months or so of amd-staging-drm-next history.
While it did work, the manually overwritten values were used only until the next mode change, DPMS-off or suspend/resume.
It's a mess. And there is absolutely no plausible reason to see why these features work-then-break-again every few months. I have really lost all hope that this development process will converge into something stable.

[atomsymbol]

While it did work, the manually overwritten values were used only until the next mode change, DPMS-off or suspend/resume.

A script for suspend/resume:

$ cat $HOME/bin/shared/suspend-computer 
#!/bin/bash

# suspend 
sync
sudo sh -c "echo mem > /sys/power/state"

# resume
if [ $(hostname) == "desktop" ]; then
    sleep 1
    gpu-powersave  # Script which selects lowest mclk+sclk
fi

Or gpu clocks can be recovered in a script which is run (e.g: by systemd) when the machine resumes from suspend.

[debianxfce]

The manual sclk/mclk overwrite used to work in some kernel versions, then broke, worked again, broke again, currently it is broken for about the last 4 months or so of amd-staging-drm-next history.

When you compare logs of the AMD wip kernel and amd-staging-drm-next, you will see that the amd-staging-drm-next kernel:
- uses older mainline kernel as base
- does have much less new patches

Conclusion: use the AMD wip kernel.