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

modernize/simplify: OC assumes "high" perf profile

Browse source
This commit is contained in:
Josh Lay 2023-04-08 12:28:51 -05:00
parent f2e4923658
commit bb03fc2cc2
Signed by: jlay
GPG key ID: B265E45CACAD108A
6 changed files with 71 additions and 57 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

12
README.md
View file

@ -1,6 +1,6 @@
# tuned-amdgpu
Hacky solution to integrate AMDGPU power profile control in `tuned` with Ansible
Hacky solution to integrate AMDGPU power control and overclocking in `tuned` with Ansible
Takes a list of existing `tuned` profiles and creates new ones based on them. These new profiles include AMDGPU power/clock parameters
@ -10,7 +10,7 @@ $ grep -ls ^connected /sys/class/drm/*/status | grep -o card[0-9] | sort | uniq
card1
```
_Warning_: This is only tested with `RX6000` series GPUs, it is probable that older AMD GPUs will not work properly. Use at your own risk!
_Warning_: This is only tested with `RX6000` series GPUs, it is probable that other generations will *not* work properly. Use at your own risk!
## Profiles
@ -19,10 +19,7 @@ An example of the output/provided profiles follow
| Output profile | Description |
|:---|---|
| `balanced-amdgpu-default` | Includes the (assumed) existing `balanced` tuned profile.<br/><br/>Only adjusts the GPU power limit (typically lower). Clocks/voltage curve remain the default. |
| `desktop-amdgpu-VR` | Includes the (assumed) existing `desktop` tuned profile.<br/><br/>Adjusts the GPU power limit, clocks, _and_ the voltage curve.<br/><br/>Uses the predefined `VR` profile in the driver. See `/sys/class/drm/card*/device/pp_power_profile_mode` |
| `latency-performance-amdgpu-custom` | Includes the existing `latency-performance` tuned profile.<br/><br/>Like the existing GPU profiles (eg: _VR)), this also adjusts the GPU power limit, clocks, _and_ the voltage curve.<br/><br/>This differs by using the `custom` profile in the driver. This opens up further tweaking of the power/clock heuristics through the driver (currently manual). see: [pp-dpm](https://docs.kernel.org/gpu/amdgpu/thermal.html#pp-dpm) |
**Note**: This is non-exhaustive, see the variables `base_profiles` and `amdgpu_profiles` below for the (default) sources of the merged profile mapping
| `desktop-amdgpu-overclock` | Includes the (assumed) existing `desktop` tuned profile.<br/><br/>Adjusts the GPU power limit, clocks, _and_ the voltage curve. |
## Notable variables
@ -35,5 +32,4 @@ These are the variables you're likely to want to change. They are defined in [p
| gpumem_clock_static | Sets the _static_ memory clock for the GPU (in `MHz`). This is *not* the _effective_ data rate. That is a multiple of this depending on the type of VRAM.<br/><br/>To avoid flickering this does *not* change dynamically with load. | `1050`, results in just over `1GHz`; mild overclock<br/><br/>Actual effective clock depends on this being multiplied against the data/pump rate of the `GDDR?` GPU memory |
| gpu_mv_offset | GPU core voltage offset. Takes +/- some integer in millivolts. Can be used to both over _and_ under volt. | `-50` (undervolt `50mV` or `0.05V`) |
| base_profiles | List of base tuned profiles to clone in the new AMDGPU profiles. Defaults based on `Fedora` | <ul><li>`balanced`</li><li>`desktop`</li><li>`latency-performance`</li><li>`network-latency`</li><li>`network-throughput`</li><li>`powersave`</li><li>`virtual-host`</li>|
| amdgpu_profiles | Dictionary mapping the AMDGPU power profiles found in `/sys/class/drm/card*/device/pp_power_profile_mode` and custom power limits.<br/><br>For each item, two keys: `pwrmode` and `pwr_cap_multi`.<br/><br/>`pwrmode` maps to the number assigned in `/sys` above.<br/>`pwr_cap_multi` is a multiplier against board power capability. Must be a float, eg: `0.5` for *50%* | <pre>default:<br/> pwrmode: 0<br/> pwr_cap_multi: 0.75<br/> # 75% relatively safe default<br/>VR:<br/> pwrmode: 4<br/> pwr_cap_multi: 0.8<br/> # 80%, likely slight boost<br/>custom:<br/> pwrmode: 6<br/> pwr_cap_multi: 1.0<br/> # 100%, full GPU board capability<br/> # warning: significantly increased heat</pre>|
| gpu_power_multi | Dictionary with two keys, `default` and `overclock`. Expects two floats to set a power limit relative to the board _capability_. Example: `1.0` is full board capability, `0.5` is 50%.|

45
playbook.yml
View file

@ -8,10 +8,13 @@
# note: 'gpu_*' vars only apply with the 'custom' suffixed profiles created by this tooling
# profiles based on the 'default' amdgpu power profile mode use default clocks
gpu_clock_min: "500" # default 500
gpu_clock_max: "2600" # default 2529
gpu_clock_max: "2615" # default 2529
gpumem_clock_static: "1075"
gpu_power_multi:
default: 0.789473684210526 # 255W - slightly reduced, maintains clocks well
overclock: 0.869969040247678 # 281W - real default, board supports up to 323W (1.0)
# optional, applies offset (+/-) to GPU voltage by provided mV
gpu_mv_offset: "-50"
# gpu_mv_offset: "-25"
# '-50' undervolts GPU core voltage 50mV or 0.05V
#
# list of source tuned profiles available on Fedora (TODO: should dynamically discover)
@ -29,24 +32,24 @@
# ref: https://www.kernel.org/doc/html/v4.20/gpu/amdgpu.html#pp-power-profile-mode
# 'pwr_cap_multi' is multiplied against board *limit* to determine profile wattage; 0.5 = 50%
# values below reflect my 6900XT
amdgpu_profiles:
default:
pwrmode: 0
pwr_cap_multi: 0.789473684210526 # 255W - slightly reduced, maintains clocks well
3D:
pwrmode: 1
pwr_cap_multi: 0.869969040247678 # 281W - default
powersave:
pwrmode: 2
pwr_cap_multi: 0.869969040247678
VR:
pwrmode: 4
pwr_cap_multi: 0.869969040247678
compute:
pwrmode: 5
pwr_cap_multi: 0.869969040247678
custom:
pwrmode: 6
pwr_cap_multi: 1.0 # 323W - full capability
# amdgpu_profiles:
# default:
# pwrmode: 0
# pwr_cap_multi: 0.789473684210526 # 255W - slightly reduced, maintains clocks well
# 3D:
# pwrmode: 1
# pwr_cap_multi: 0.869969040247678 # 281W - default
# powersave:
# pwrmode: 2
# pwr_cap_multi: 0.869969040247678
# VR:
# pwrmode: 4
# pwr_cap_multi: 0.869969040247678
# compute:
# pwrmode: 5
# pwr_cap_multi: 0.869969040247678
# custom:
# pwrmode: 6
# pwr_cap_multi: 1.0 # 323W - full capability
# both dictionaries are merged to create new 'tuned' profiles. eg:
# 'balanced-amdgpu-default', 'balanced-amdgpu-3D', 'balanced-amdgpu-video'

22
roles/tuned_amdgpu/defaults/main.yml
View file

@ -8,8 +8,20 @@ board_watts: "{{ power_max | int / 1000000 }}"
# internals for profile power calculations
# item in the context of the with_nested loops in the play
profile_name: "{{ item.0.key }}"
profile_percentage: "{{ (item.0.value.pwr_cap_multi * 100.0) | round(2) }}"
profile_multi: "{{ item.0.value.pwr_cap_multi }}"
profile_microwatts: "{{ power_max | float * profile_multi | float }}"
profile_watts: "{{ profile_microwatts | int / 1000000 }}"
profile_name: "{{ item.0 }}"
# determine percentage for human-friendly comments
power_default_pct: "{{ (gpu_power_multi.default * 100.0) | round(2) }}"
power_oc_pct: "{{ (gpu_power_multi.overclock * 100.0) | round(2) }}"
# in microWatts, actually written to sysfs
power_default_mw: "{{ (power_max | float) * (gpu_power_multi.default | float) }}"
power_oc_mw: "{{ (power_max | float) * (gpu_power_multi.overclock | float) }}"
# wattages - more human-friendly comments
power_default_watts: "{{ (power_default_mw | int) / 1000000 }}"
power_oc_watts: "{{ (power_oc_mw | int) / 1000000 }}"
amdgpu_profiles:
- default
- overclock

12
roles/tuned_amdgpu/tasks/main.yml
View file

@ -63,22 +63,22 @@
- name: Create custom profile directories
ansible.builtin.file:
state: directory
path: /etc/tuned/{{ item.1 }}-amdgpu-{{ item.0.key }}
path: /etc/tuned/{{ item.1 }}-amdgpu-{{ item.0 }}
mode: "0755"
with_nested:
- "{{ lookup('dict', amdgpu_profiles) }}"
- "{{ amdgpu_profiles }}"
- "{{ base_profiles }}"
become: true
- name: Template AMDGPU control/reset scripts
ansible.builtin.template:
src: templates/amdgpu-clock.sh.j2
dest: /etc/tuned/{{ item.1 }}-amdgpu-{{ item.0.key }}/amdgpu-clock.sh
dest: /etc/tuned/{{ item.1 }}-amdgpu-{{ item.0 }}/amdgpu-clock.sh
owner: root
group: root
mode: "0755"
with_nested:
- "{{ lookup('dict', amdgpu_profiles) }}"
- "{{ amdgpu_profiles }}"
- "{{ base_profiles }}"
notify: Restart tuned
become: true
@ -86,12 +86,12 @@
- name: Template custom tuned profiles
ansible.builtin.template:
src: templates/tuned.conf.j2
dest: /etc/tuned/{{ item.1 }}-amdgpu-{{ item.0.key }}/tuned.conf
dest: /etc/tuned/{{ item.1 }}-amdgpu-{{ item.0 }}/tuned.conf
owner: root
group: root
mode: "0644"
with_nested:
- "{{ lookup('dict', amdgpu_profiles) }}"
- "{{ amdgpu_profiles }}"
- "{{ base_profiles }}"
notify: Restart tuned
become: true

29
roles/tuned_amdgpu/templates/amdgpu-clock.sh.j2
View file

@ -20,9 +20,6 @@ CARD=$(/usr/bin/grep -ls ^connected /sys/class/drm/*/status | /usr/bin/grep -o '
{# begin the templated script for 'default' profiles to reset state #}
{% if 'default' in profile_name %}
# set power state transition heuristics to default
echo '{{ item.0.value.pwrmode }}' | tee /sys/class/drm/"${CARD}"/device/pp_power_profile_mode
# set control mode back to auto
# attempts to dynamically set optimal power profile for (load) conditions
echo 'auto' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_level
@ -30,23 +27,20 @@ echo 'auto' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_le
# reset any existing profile clock changes
echo 'r' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
# give '{{ profile_name }}' profile ~{{ profile_percentage }}% (rounded) of the max power capability
# {{ profile_watts }} Watts of {{ board_watts }} total
echo '{{ profile_microwatts | int }}' | tee '{{ powercap_set.files.0.path }}'
# give '{{ profile_name }}' profile ~{{ power_default_pct }}% (rounded) of the max power capability
# {{ power_default_watts }} Watts of {{ board_watts }} total
echo '{{ power_default_mw | int }}' | tee '{{ powercap_set.files.0.path }}'
{% else %}
{# begin the templated script for non-default AMD GPU profiles, eg: 'VR' or '3D_FULL_SCREEN' #}
# set manual control mode
# allows control via 'pp_dpm_mclk', 'pp_dpm_sclk', 'pp_dpm_pcie', 'pp_dpm_fclk', and 'pp_power_profile_mode' files
# only interested in 'pp_power_profile_mode' for power and 'pp_dpm_mclk' for memory clock (flickering).
# GPU clocks are dynamic based on (load) condition
echo 'manual' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_level
#echo 'manual' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_level
# set power state transition heuristics to '{{ profile_name }}' profile
echo '{{ item.0.value.pwrmode }}' | tee /sys/class/drm/"${CARD}"/device/pp_power_profile_mode
# give '{{ profile_name }}' profile ~{{ profile_percentage }}% (rounded) of the max power capability
# {{ profile_watts }} Watts of {{ board_watts }} total
echo '{{ profile_microwatts | int }}' | tee '{{ powercap_set.files.0.path }}'
# give '{{ profile_name }}' profile ~{{ power_oc_pct }}% (rounded) of the max power capability
# {{ power_oc_watts }} Watts of {{ board_watts }} total
echo '{{ power_oc_mw | int }}' | tee '{{ powercap_set.files.0.path }}'
# set the minimum GPU clock
echo 's 0 {{ gpu_clock_min }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
@ -54,7 +48,8 @@ echo 's 0 {{ gpu_clock_min }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_v
# set the maximum GPU clock
echo 's 1 {{ gpu_clock_max }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
# set the maximum GPU *memory* clock
# set the minimum / maximum GPU *memory* clock - force it high
echo 'm 0 {{ gpumem_clock_static }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
echo 'm 1 {{ gpumem_clock_static }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
{% if gpu_mv_offset is defined %}
@ -68,5 +63,9 @@ echo 'c' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
# force GPU memory into highest clock (fix flickering)
# 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
echo '3' | tee /sys/class/drm/"${CARD}"/device/pp_dpm_mclk
# echo '3' | tee /sys/class/drm/"${CARD}"/device/pp_dpm_mclk
# note 4/8/2023: instead of 'manual'... deal with broken power management, force clocks to high
# ref: https://gitlab.freedesktop.org/drm/amd/-/issues/1500
echo 'high' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_level
{% endif %}

8
roles/tuned_amdgpu/templates/tuned.conf.j2
View file

@ -1,8 +1,10 @@
[main]
include={{ item.1 }}
summary={{ item.1 }} + TCP/RAID tweaks + AMDGPU pp_power_profile_mode = {{ item.0.value.pwrmode }} ({{ item.0.key }})
summary={{ item.1 }} + TCP/RAID tweaks + AMDGPU {{ item.0 }}
[sysctl]
# allow regular users to see the kernel ring buffer
kernel.dmesg_restrict=0
net.core.default_qdisc=fq
# 'bbr2' requires a [modified] supporting kernel - stock Fedora kernels do *not* support it (currently)
# eg: 'kernel-xanmode-edge' from COPR 'rmnscnce/kernel-xanmod'
@ -25,4 +27,6 @@ script=${i:PROFILE_DIR}/amdgpu-clock.sh
[ssdnosched]
type=disk
devices_udev_regex=(ID_ATA_ROTATION_RATE_RPM=0)
elevator=none
# elevator=none
elevator=kyber
# elevator=mq-deadline