WattMan in 4.17, does it work for vega10?

[cde1]

From preliminary testing, by setting power_dpm_force_performance_level to manual first, and then setting pp_power_profile_mode to the various profiles (0 to 4), nothing is changed in the power consumption and the sclk still hovers between 1536 and 1630.

I'm not sure what I might be missing. Ideally I'd like a way to simply set the TDP for a given card and have amdgpu automatically choose appropriate clocks. For instance if I set sclk manually to 1401 I almost halve the consumed power with a very small decrease in performance, so this shows a lot of potential for optimisation.

(and to anyone who's wondering: no, I'm not mining cryptocurrencies with this card. I run other compute workloads).

[cde1]

As it turns out the issue with manually forcing sclk to a given frequency is that when the GPU is not used, power consumption is quite higher (+30W at the wall). So instead I've written a patch to vega10_find_highest_dpm_level() that prevents the two highers power levels from being used and still allows the lowest power level to be used when idle.

--- ./drivers/gpu/drm/amd/powerplay/hwmgr/vega10_hwmgr.c~    2018-05-19 23:48:31.447618885 +0200
+++ ./drivers/gpu/drm/amd/powerplay/hwmgr/vega10_hwmgr.c    2018-05-19 23:57:36.837027441 +0200
@@ -3535,7 +3535,7 @@
     uint32_t i = 0;
 
     if (table->count <= MAX_REGULAR_DPM_NUMBER) {
-        for (i = table->count; i > 0; i--) {
+        for (i = table->count - 2; i > 0; i--) {
             if (table->dpm_levels[i - 1].enabled)
                 return i - 1;
         }

With this patch, I have a decrease of about 100W at the wall when running LuxMark, with a score of about 25000 instead of ~26400 (so a decrease of 5% in performance).

[cde1]

As it turns out the issue with manually forcing sclk to a given frequency is that when the GPU is not used, power consumption is quite higher (+30W at the wall). So instead I've written a patch to vega10_find_highest_dpm_level() that prevents the two highers power levels from being used and still allows the lowest power level to be used when idle.

--- ./drivers/gpu/drm/amd/powerplay/hwmgr/vega10_hwmgr.c~ 2018-05-19 23:48:31.447618885 +0200
+++ ./drivers/gpu/drm/amd/powerplay/hwmgr/vega10_hwmgr.c 2018-05-19 23:57:36.837027441 +0200
@@ -3535,7 +3535,7 @@
uint32_t i = 0;

if (table->count <= MAX_REGULAR_DPM_NUMBER) {
- for (i = table->count; i > 0; i--) {
+ for (i = table->count - 2; i > 0; i--) {
if (table->dpm_levels[i - 1].enabled)
return i - 1;
}

With this patch, I have a decrease of about 100W at the wall when running LuxMark, with a score of about 25000 instead of ~26400 (so a decrease of 5% in performance).

For the record, here's a script to recompile amdgpu.ko on Debian without having to go through the (very slow) dpkg-buildpackage:

#!/bin/sh
KVER=`uname -r`
sudo apt-get install module-assistant bison flex libelf-dev libssl-dev bc
sudo m-a prepare
apt-get source linux-image-$KVER
cd linux-*
perl -i -pe 's/i = table->count/i = table->count - 2/g' drivers/gpu/drm/amd/powerplay/hwmgr/vega10_hwmgr.c
cp /usr/src/linux-headers-$KVER/Module.symvers .
make oldconfig
make prepare
make modules_prepare
make SUBDIRS=scripts/mod
make SUBDIRS=drivers/gpu/drm/amd/amdgpu modules
sudo cp drivers/gpu/drm/amd/amdgpu/amdgpu.ko /lib/modules/$KVER/kernel/drivers/gpu/drm/amd/amdgpu/amdgpu.ko
sudo depmod -a

Incidentally, the latest 18.30 amdgpu package (amdgpu-pro-18.30-633530.tar.xz) has the annoying habit of checking wether it's running on Ubuntu 18.04 or not, this check is done by amdgpu-core_18.30-633530_all.deb:

if [ "$VERSION_ID" != "18.04" ] ; then
    >&2 echo "ERROR: This package can only be installed on Ubuntu 18.04."
    exit 1
fi

This can be fixed by temporarily adding VERSION_ID=18.04 to /etc/os-release when installing this package (for Vega, only opencl-amdgpu-pro-icd_18.30-633530_amd64.deb amdgpu-pro-core_18.30-633530_all.deb and amdgpu-core_18.30-633530_all.deb are required for OpenCL support, in fact it might even be easier to extract libamdocl64.so and amdocl64.icd rather than installing those .deb).

[cde1]

Thanks debianxfce!

[cde1]

Additionally, I found out that the intel_pstates driver (enabled by default in the "active" state) does not allow the userspace governor. It has to be either disabled (but then turbo states are unavailable as acpi-cpufreq doesn't support them), or set in passive mode (intel_pstate=passive) which enables the use of the userspace governor and forcing a given frequency with cpupower frequency-set.