jlay/tuned-amdgpu
1
0
Fork 0
Code Issues Pull requests Projects Activity

fix clock profile choice consistency when moving profiles

Browse source
This commit is contained in:
Josh Lay 2022-08-03 00:47:43 -05:00
parent 6214034e59
commit ca723c732b
Signed by: jlay
GPG key ID: B265E45CACAD108A
3 changed files with 57 additions and 37 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

53
playbook.yml
View file

@ -6,35 +6,48 @@
# 0.5 = 50% # 0.5 = 50%
# 1.0 = 100% (of card power capability, not stock limits) # 1.0 = 100% (of card power capability, not stock limits)
# should not exceed 1.0, must be a float. driver will do some rounding/stepping # should not exceed 1.0, must be a float. driver will do some rounding/stepping
power_max_custom_multi: 1.0 # used to control the effective power limit in the 'custom' AMDGPU tuned profile # default is ~87% (281/323) on my 6900XT. check '/sys/class/drm/card*/device/hwmon/hwmon*/power1_cap_default'
power_max_default_multi: 0.75 # used to limit GPU power to some lower percentage on default perf mode power_max_custom_multi: 0.928793 # calculated to ~300W. used to control the effective power limit in the non-default AMDGPU tuned profiles
# minimum/maximum GPU clocks using 'powerplay'. power_max_default_multi: 0.82 # (typically) used to limit GPU power to some lower percentage on default perf mode/profile
#
# minimum/maximum GPU clocks using 'powerplay'
# these do not apply to the resulting 'amdgpu-default' tuned profile
# only the others (eg: 'VR'/'custom') under amdgpu_profiles (below)
# effective clocks are dynamically adjusted by the driver in this range - based on utilization # effective clocks are dynamically adjusted by the driver in this range - based on utilization
# can find baseline values in the '/sys/class/drm/{{ card }}/device/pp_od_clk_voltage' file # can find baseline values in the '/sys/class/drm/{{ card }}/device/pp_od_clk_voltage' file
# OD_RANGE indicates acceptable value ranges for SCLK (core) and MCLK (memory) # OD_RANGE indicates acceptable value ranges for SCLK (core) and MCLK (memory)
# Under 'OD_SCLK' and 'OD_MCLK' you will see 0/1. These are the minimum/maximum values for the respective component. # Under 'OD_SCLK' and 'OD_MCLK' you will see 0/1. These are the minimum/maximum values for the respective component.
gpu_clock_min: "2300" gpu_clock_min: "2000"
gpu_clock_max: "2600" gpu_clock_max: "2615"
#
# memory clocks are static, we only set a 'max' - high-refresh rate display quirk # memory clocks are static, we only set a 'max' - high-refresh rate display quirk
# this allows the memory clock to be adjusted # this allows the memory clock to be adjusted
gpumem_clock_max: "1075" gpumem_clock_max: "1075"
#
# optional, applies offset to GPU voltage, eg: '+100' = to boost GPU core voltage 100mV or 0.1V. for the 'custom' GPU profile. # optional, applies offset to GPU voltage, eg: '+100' = to boost GPU core voltage 100mV or 0.1V. for the 'custom' GPU profile.
gpu_mv_offset: "+50" gpu_mv_offset: "-25"
#
# the card for 'tuned' to control # the card for 'tuned' to control
# this is *usually* 'card0', but may differ. # this is *usually* 'card0', but may differ.
# check '/dev/dri/card*' # check '/dev/dri/card*'
card: 'card0' # default to card0 card: 'card0' # default to card0
base_profiles: # list of source tuned profiles available on Fedora (TODO: should dynamically discover?) # list of source tuned profiles available on Fedora (TODO: should dynamically discover)
- balanced # these are further modified with secondary (templated) profiles, see 'templates/tuned.conf.j2' # further modified with AMD GPU power/clock parameters, creating new profiles. eg: 'balanced-amdgpu-VR'
base_profiles:
- balanced
- desktop - desktop
- latency-performance - latency-performance
- network-latency - network-latency
- network-throughput - network-throughput
- powersave - powersave
- virtual-host - virtual-host
amdgpu_profiles: # statically defined mapping of the contents in /sys/class/drm/{{ card }}/device/pp_power_profile_mode amdgpu_profiles: # statically defined mapping of the contents in /sys/class/drm/card*/device/pp_power_profile_mode
- { name: 'default', value: 0 } default: # more may be added, but do not remove default/custom. new profiles require a script template, see 'templates'
- { name: 'custom', value: 6 } pwrmode: 0
VR:
pwrmode: 4
custom:
pwrmode: 6
handlers: handlers:
- name: restart tuned - name: restart tuned
ansible.builtin.service: ansible.builtin.service:
@ -85,30 +98,34 @@
- name: create custom profile directories - name: create custom profile directories
file: file:
state: directory state: directory
path: /etc/tuned/{{ item.1 }}-amdgpu-{{ item.0.name }} path: /etc/tuned/{{ item.1 }}-amdgpu-{{ item.0.key }}
mode: "0755" mode: "0755"
with_nested: with_nested:
- "{{ amdgpu_profiles }}" - "{{ lookup('dict', amdgpu_profiles) }}"
- "{{ base_profiles }}" - "{{ base_profiles }}"
- name: template AMDGPU clock control scripts (default/GPU profile dependency) - name: template AMDGPU control/reset scripts
template: template:
src: templates/amdgpu-clock.sh.j2 src: templates/amdgpu-clock.sh.j2
dest: /etc/tuned/{{ item.1 }}-amdgpu-{{ item.0.name }}/amdgpu-clock.sh dest: /etc/tuned/{{ item.1 }}-amdgpu-{{ item.0.key }}/amdgpu-clock.sh
owner: root owner: root
group: root group: root
mode: "0755" mode: "0755"
with_nested: with_nested:
- "{{ amdgpu_profiles }}" - "{{ lookup('dict', amdgpu_profiles) }}"
- "{{ base_profiles }}" - "{{ base_profiles }}"
notify: restart tuned notify: restart tuned
- name: template custom tuned profiles - name: template custom tuned profiles
template: template:
src: templates/tuned.conf.j2 src: templates/tuned.conf.j2
dest: /etc/tuned/{{ item.1 }}-amdgpu-{{ item.0.name }}/tuned.conf dest: /etc/tuned/{{ item.1 }}-amdgpu-{{ item.0.key }}/tuned.conf
owner: root owner: root
group: root group: root
mode: "0644" mode: "0644"
with_nested: with_nested:
- "{{ amdgpu_profiles }}" - "{{ lookup('dict', amdgpu_profiles) }}"
- "{{ base_profiles }}" - "{{ base_profiles }}"
notify: restart tuned notify: restart tuned
- name: ensure tuned is enabled
service:
name: tuned
enabled: yes

33
templates/amdgpu-clock.sh.j2
View file

@ -3,23 +3,24 @@
# configures GPU power/clock characteristics # configures GPU power/clock characteristics
# clocks/power in 3D are dynamic based on need/usage # clocks/power in 3D are dynamic based on need/usage
# #
# for 'amdgpu-default' tuned profiles, this will reset the characteristics to default
# for others this will apply overclocking settings -- leaving clock choices to the associated power profile (eg: VR)
#
# rendered by Ansible with environment-appropriate values: # rendered by Ansible with environment-appropriate values:
# card #, eg: card0 # card #, eg: card0
# path to discovered sysfs device files (power/clock/voltage control) # path to discovered sysfs device files (power/clock/voltage control)
# #
# this sets the minimums / maximums for a specific generation of GPU
# settings may need adjusted
#
# AMDGPU driver/sysfs references: # AMDGPU driver/sysfs references:
# https://01.org/linuxgraphics/gfx-docs/drm/gpu/amdgpu.html # https://01.org/linuxgraphics/gfx-docs/drm/gpu/amdgpu.html
# https://docs.kernel.org/gpu/amdgpu/thermal.html # https://docs.kernel.org/gpu/amdgpu/thermal.html
{# begin the templated script for 'default' profiles to reset state #}
{% if 'default' in item.0.key %}
{% if 'default' in item.0.name %}
# set power state transition heuristics to default # set power state transition heuristics to default
echo '{{ item.0.value }}' | tee /sys/class/drm/{{ card }}/device/pp_power_profile_mode echo '{{ item.0.value.pwrmode }}' | tee /sys/class/drm/{{ card }}/device/pp_power_profile_mode
# set control mode back to auto # set control mode back to auto
# attempts to dynamically set optimal power profile for conditions # attempts to dynamically set optimal power profile for (load) conditions
echo 'auto' | tee /sys/class/drm/{{ card }}/device/power_dpm_force_performance_level echo 'auto' | tee /sys/class/drm/{{ card }}/device/power_dpm_force_performance_level
# reset any existing profile clock changes # reset any existing profile clock changes
@ -28,14 +29,17 @@ echo 'r' | tee /sys/class/drm/{{ card }}/device/pp_od_clk_voltage
# give default profile {{ power_max_default_multi * 100.0 |int }}% of the max power capability # give default profile {{ power_max_default_multi * 100.0 |int }}% of the max power capability
# {{ power_cap_default|int/1000000 }} Watts of {{ power_max|int/1000000 }} total # {{ power_cap_default|int/1000000 }} Watts of {{ power_max|int/1000000 }} total
echo '{{ power_cap_default }}' | tee {{ powercap_set.files.0.path }} echo '{{ power_cap_default }}' | tee {{ powercap_set.files.0.path }}
{% elif 'custom' in item.0.name %} {% else %}
{# begin the templated script for non-default AMD GPU profiles, eg: 'VR' or '3D_FULL_SCREEN' #}
# set manual control mode # set manual control mode
# allow control via 'pp_dpm_mclk', 'pp_dpm_sclk', 'pp_dpm_pcie', and 'pp_power_profile_mode' files # allows control via 'pp_dpm_mclk', 'pp_dpm_sclk', 'pp_dpm_pcie', 'pp_dpm_fclk', and 'pp_power_profile_mode' files
# only interested in 'pp_power_profile_mode' for power and 'pp_dpm_mclk' for memory clock (flickering).
# GPU clocks are dynamic based on (load) condition
echo 'manual' | tee /sys/class/drm/{{ card }}/device/power_dpm_force_performance_level echo 'manual' | tee /sys/class/drm/{{ card }}/device/power_dpm_force_performance_level
# set power state transition heuristics to custom/manual # set power state transition heuristics to '{{ item.0.key }}' profile
# looked up from amdgpu_profiles variable using 'with_nested' loop in task echo '{{ item.0.value.pwrmode }}' | tee /sys/class/drm/{{ card }}/device/pp_power_profile_mode
echo '{{ item.0.value }}' | tee /sys/class/drm/{{ card }}/device/pp_power_profile_mode
# give this profile {{ power_max_custom_multi * 100.0 |int }}% of the max power capability # give this profile {{ power_max_custom_multi * 100.0 |int }}% of the max power capability
# {{ power_cap_custom|int/1000000 }} Watts of {{ power_max|int/1000000 }} total # {{ power_cap_custom|int/1000000 }} Watts of {{ power_max|int/1000000 }} total
@ -58,9 +62,8 @@ echo 'vo {{ gpu_mv_offset }}' | tee /sys/class/drm/{{ card }}/device/pp_od_clk_v
# commit the changes # commit the changes
echo 'c' | tee /sys/class/drm/{{ card }}/device/pp_od_clk_voltage echo 'c' | tee /sys/class/drm/{{ card }}/device/pp_od_clk_voltage
# force GPU core/shader into the highest profile: {{ gpu_clock_max }}Mhz # force GPU memory into highest clock (fix flickering)
echo '1' | tee /sys/class/drm/{{ card }}/device/pp_dpm_sclk # pp_dpm_*clk settings are unintuitive, giving profiles that may be used
# opt not to set the others (eg: sclk/fclk) - those should remain for benefits from the curve
# force GPU memory into highest profile
echo '3' | tee /sys/class/drm/{{ card }}/device/pp_dpm_mclk echo '3' | tee /sys/class/drm/{{ card }}/device/pp_dpm_mclk
{% endif %} {% endif %}

2
templates/tuned.conf.j2
View file

@ -1,6 +1,6 @@
[main] [main]
include={{ item.1 }} include={{ item.1 }}
summary={{ item.1 }} + TCP/RAID tweaks + AMDGPU pp_power_profile_mode = {{ item.0.value }} ({{ item.0.name }}) summary={{ item.1 }} + TCP/RAID tweaks + AMDGPU pp_power_profile_mode = {{ item.0.value.pwrmode }} ({{ item.0.key }})
[sysctl] [sysctl]
net.core.default_qdisc=fq net.core.default_qdisc=fq