#!/bin/bash
# script for tuned AMDGPU clock control
# configures GPU power/clock characteristics
# 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:
#   card #, eg: card0
#   path to discovered sysfs device files (power/clock/voltage control)
#
# AMDGPU driver/sysfs references:
#   https://01.org/linuxgraphics/gfx-docs/drm/gpu/amdgpu.html
#   https://docs.kernel.org/gpu/amdgpu/thermal.html

{# done this way to avoid issues with the card number possibly shifting after playbook run #}
# dynamically determine the connected GPU using the DRM subsystem
CARD=$(/usr/bin/grep -ls ^connected /sys/class/drm/*/status | /usr/bin/grep -o 'card[0-9]' | /usr/bin/sort | /usr/bin/uniq | /usr/bin/sort -h | /usr/bin/tail -1)

function get_hwmon_dir() {
    CARD_DIR="/sys/class/drm/${1}/device/"
    for CANDIDATE in "${CARD_DIR}"/hwmon/hwmon*; do
            if [[ -f "${CANDIDATE}"/power1_cap ]]; then
                    # found a valid hwmon dir
                    echo "${CANDIDATE}"
            fi
    done
}

# determine the hwmon directory
HWMON_DIR=$(get_hwmon_dir "${CARD}")

# read all of the power profiles, used to get the IDs for assignment later
PROFILE_MODES=$(< /sys/class/drm/"${CARD}"/device/pp_power_profile_mode)

# get power capability; later used determine limits
read -r -d '' POWER_CAP < "$HWMON_DIR"/power1_cap_max

# enable THP; profile enables the 'vm.compaction_proactiveness' sysctl
# improves allocation latency
echo 'always' | tee /sys/kernel/mm/transparent_hugepage/enabled

{# begin the templated script for 'default' profiles to reset state #}
{% if 'default' in profile_name %}
# 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

# reset any existing profile clock changes
echo 'r' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage

# adjust power limit using multiplier against board capability
POWER_LIM_DEFAULT=$(/usr/bin/awk -v m="$POWER_CAP" -v n={{ gpu_power_multi.default }} 'BEGIN {printf "%.0f", (m*n)}')
echo "$POWER_LIM_DEFAULT" | tee "${HWMON_DIR}/power1_cap"

# extract the power-saving profile ID number
PROF_POWER_SAVING_NUM=$(/usr/bin/awk '$0 ~ /POWER_SAVING.*:/ {print $1}' <<< "$PROFILE_MODES")

# reset power/clock heuristics to power-saving
echo "${PROF_POWER_SAVING_NUM}" | tee /sys/class/drm/"${CARD}"/device/pp_power_profile_mode
{% else %}
{# begin the templated script for 'overclocked' AMD GPU profiles based on the existing tuned profiles #}
# set the minimum GPU clock - for best performance, this should be near the maximum
# RX6000 series power management *sucks*
echo 's 0 {{ gpu_clock_min }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage

# set the maximum GPU clock
echo 's 1 {{ gpu_clock_max }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage

# set the GPU *memory* clock
# normally this would appear disregarded, memory clocked at the minimum allowed by the overdrive (OD) range
# it follows the core clock; if both 0/1 profiles for _it_ are high enough, the memory will follow
echo 'm 1 {{ gpumem_clock_static }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
{% if gpu_mv_offset is defined %}

# offset GPU voltage {{ gpu_mv_offset }}mV
echo 'vo {{ gpu_mv_offset }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
{% endif %}

# commit the changes
echo 'c' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage

# force GPU core and memory into highest clocks (fix flickering and poor power management)
# 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
echo 'manual' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_level

# adjust power limit using multiplier against board capability
POWER_LIM_OC=$(/usr/bin/awk -v m="$POWER_CAP" -v n={{ gpu_power_multi.overclock }} 'BEGIN {printf "%.0f", (m*n)}')
echo "$POWER_LIM_OC" | tee "${HWMON_DIR}/power1_cap"

#   pp_dpm_*clk settings are unintuitive, giving profiles that may be used
echo '1' | tee /sys/class/drm/"${CARD}"/device/pp_dpm_sclk
echo '3' | tee /sys/class/drm/"${CARD}"/device/pp_dpm_mclk
echo '2' | tee /sys/class/drm/"${CARD}"/device/pp_dpm_fclk
echo '2' | tee /sys/class/drm/"${CARD}"/device/pp_dpm_socclk

# extract the VR power profile ID number
PROF_VR_NUM=$(/usr/bin/awk '$0 ~ /VR.*:/ {print $1}' <<< "$PROFILE_MODES")

# force 'overclocked' profile to 'VR' power/clock heuristics
# latency/frame timing seemed favorable with relatively-close minimum clocks
echo "${PROF_VR_NUM}" | tee /sys/class/drm/"${CARD}"/device/pp_power_profile_mode

# note 4/8/2023: instead of 'manual'... try dealing with broken power management, force clocks to high
# ref: https://gitlab.freedesktop.org/drm/amd/-/issues/1500
# followup: doesn't work that well in practice, still flaky on clocks/frame times
#echo 'high' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_level
{% endif %}