Create a gist now

Instantly share code, notes, and snippets.

What would you like to do?
This gist will show you how to tune your Intel-based Skylake, Kabylake and beyond Integrated Graphics Core for performance and reliability through GuC and HuC firmware usage on Linux.

Tuning Intel Skylake and beyond for optimal performance and feature level support on Linux:

Note that on Skylake, Kabylake (and the now cancelled "Broxton") SKUs, functionality such as power saving, GPU scheduling and HDMI audio have been moved onto binary-only firmware, and as such, the GuC and the HuC blobs must be loaded at run-time to access this functionality.

Enabling GuC and HuC on Skylake and above requires a few extra parameters be passed to the kernel before boot.

Instructions provided for both Fedora and Ubuntu (including Debian):

Note that the firmware for these GPUs is often packaged by your distributor, and as such, you can confirm the firmware blob's availability by running:

1. On Fedora:

rpm -ql linux-firmware | fgrep i915

Sample output:

/usr/lib/firmware/i915 
/usr/lib/firmware/i915/bxt_dmc_ver1_07.bin 
/usr/lib/firmware/i915/kbl_dmc_ver1.bin
/usr/lib/firmware/i915/kbl_dmc_ver1_01.bin 
/usr/lib/firmware/i915/skl_dmc_ver1_23.bin 
/usr/lib/firmware/i915/skl_guc_ver1.bin 
/usr/lib/firmware/i915/skl_guc_ver4.bin 
/usr/lib/firmware/i915/skl_guc_ver6.bin 
/usr/lib/firmware/i915/skl_guc_ver6_1.bin 
/usr/share/doc/linux-firmware/LICENSE.i915

2. On Ubuntu:

dpkg -L linux-firmware | fgrep i915

It may be wise to install the linux-firmware-nonfree package as it may contain extra firmware blobs on Ubuntu releases prior to 16.04LTS. Don't install this on 16.04LTS and above.

Now, to the real meat:

GUC loading on Skylake and above:

Add this to the kernel boot parameters on /etc/default/grub:

intel_pstate=skylake_hwp 

The intel_pstate=skylake_hwp option enables Intel's Hardware P-States, known commonly by the marketing moniker "SpeedShift". Note that Intel's IOMMU for VT-d (enabled in the kernel by passing intel_iommu=on to the kernel boot options) cannot be used together with frame buffer compression (i915.enable_fbc=1 argument passed to the i915). Frame Buffer Compression will be disabled automatically to prevent screen tearing in this case.

Then, update grub on Fedora:

(Run commands as root):

For EFI boot (More common):

grub2-mkconfig -o /boot/efi/EFI/fedora/grub.cfg

And if you're still booting up in legacy BIOS mode on an MBR-style partitioning scheme or with CSM enabled on Fedora for whatever reason:

grub2-mkconfig -o /boot/grub2/grub.cfg

On Ubuntu:

Simply run:

update-grub

Then reboot.

Your Skylake processor graphics will work just fine.

You can also add this to: /etc/modprobe.d/i915.conf

In the following syntax:

options i915.enable_guc_loading=1 i915.enable_guc_submission=1

Other safe options to pass are enable_rc6=1, enable_fbc=1, enable_psr=1, disable_power_well=0 and semaphores=1

Reference:

See modinfo output for i915 for available GuC options:

modinfo i915 | grep guc

Further notes:

A list of all options along with short descriptions and default values can be generated with the following command:

$ modinfo -p i915

To check which options are currently enabled, run:

# systool -m i915 -av

You will note that many options default to -1, resulting in per-chip power-saving defaults. It is however possible to configure more aggressive powersaving by using module options. Warning: Diverting from the defaults will mark the kernel as tainted from Linux 3.18 onwards. This basically implies using other options than the per-chip defaults is considered experimental and not supported by the developers.

The following set of options should be generally safe to enable:

/etc/modprobe.d/i915.conf

options i915.enable_rc6=1 i915.enable_fbc=1 i915.semaphores=1

RC6 sleep modes (enable_rc6):

You can experiment with higher values for enable_rc6, but your GPU may not support them or the activation of the other options:

The available enable_rc6 values are a bitmask with bit values RC6=1, RC6p=2, RC6pp=4[4] - where "RC6p" and "RC6pp" are lower power states.

To confirm the current running RC6 level, you can look in sysfs:

# cat /sys/class/drm/card0/power/rc6_enable

If the value read is a lower number than expected, the other RC6 level are probably not supported. Passing drm.debug=0xe to the kernel boot options will add DRM debugging information to the kernel log - possibly including a line like this:

[drm:sanitize_rc6_option] Adjusting RC6 mask to 1 (requested 7, valid 1)

Framebuffer compression (enable_fbc):

Framebuffer compression may be unreliable or unavailable on Intel GPU generations before Sandy Bridge (generation 6). This results in messages logged to the system journal similar to this one:

kernel: drm: not enough stolen space for compressed buffer, disabling.

Tear-free video:

With the SNA acceleration method enabled, tearing may be observed. To fix this, enable the "TearFree" option in the driver by adding the following line to your xorg.conf (or ideally, a sub-configuration file under xorg.conf.d) configuration file:

Option "TearFree" "true"

Debugging:

Simply inspect dmesg:

dmesg | grep drm

And also look at the output of:

journalctl -b -o short-monotonic -k

Screen corruption observed when waking up from suspend

This is often observed as font and screen corruption in GTK+ applications (missing glyphs after suspend/resume).

Should you experience missing font glyphs in GTK+ applications, the following workaround might help. Edit /etc/environment to add the following line:

/etc/environment
COGL_ATLAS_DEFAULT_BLIT_MODE=framebuffer

See this bug here for more details.

Thanks and regards,

Brainiarc7

Which configuration file needs the tear free option?

Owner

Brainiarc7 commented May 19, 2017

The xorg.conf file. Updated to reflect that.

Thanks for the info, now i have enabled guc and huc firmware in gentoo.
Do you happen to now why are disk unrecognized (kernel panic) if i set the gpu as modules in the kernel?

thanks

Owner

Brainiarc7 commented Aug 3, 2017

Hello @josedellabosca,

No.

It might be a bug in the kernel package you're running.

pcordes commented Aug 27, 2017

In recent kernels (for example 4.12), intel_pstate=skylake_hwp isn't listed as an option (https://github.com/torvalds/linux/blob/v4.12/Documentation/admin-guide/kernel-parameters.txt#L1604). Maybe it was removed, but that string doesn't appear in the source anywhere.

The pstate driver Just Works on my desktop (using hwp), and has since at least 4.10 (when I got this hardware) without using any options.

pcordes commented Aug 27, 2017

On my Skylake i7-6700k, I'm getting

[drm:i915_gem_init_stolen [i915]] *ERROR* conflict detected with stolen region: [0x00000000c0000000 - 0x00000000c8000000]

in my kernel log, with i915 options all at their default settings. (Linux 4.12.8-2-ARCH, but I've seen this on 4.10 and 11). My firmware blobs are the latest versions.

Am I missing out on any performance here? Linux drm loads the DMC 1.26 firmware, but I guess GuC may not even be enabled by default?

Owner

Brainiarc7 commented Aug 29, 2017

So, according to commit 023f807989f4d6a076258f4efe6d1da0d7aaeb29
Author: Daniel Vetter daniel.vetter@ffwll.ch
Date: Wed Jul 19 12:00:43 2017 +0200

drm/i915: More stolen quirking 

The regression you're seeing should be fixed.

You may want to run the latest Xorg and Mesa stack on your host. If you're on Ubuntu, run the xorg-edgers PPA and the Oibaf PPA.

Since you seem to be on Arch, keep your host machine up to date. By default, Arch runs the latest and greatest mesa stack available.

Do i need to something else to these parameter to take place? I rebooted and they didn't take place.
I'm in an Manjaro install

Owner

Brainiarc7 commented Nov 27, 2017

Hello @GrayJack,

What kernel version are you running?
And what's the output of lspci ?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment