Skip to content

Install Intel VA-API drivers by device, not by marketing name - #9824

Open
ifigueroap wants to merge 1 commit into
omacom:quattrofrom
ifigueroap:fix/intel-vaapi-detection
Open

ifigueroap wants to merge 1 commit into
omacom:quattrofrom
ifigueroap:fix/intel-vaapi-detection

Conversation

@ifigueroap

@ifigueroap ifigueroap commented Sep 2, 2026 •

Copy link
Copy Markdown

Problem

install/hardware/intel/video-acceleration.sh decides which VA-API driver to install by matching the lspci line against an allowlist of marketing names:

if [[ ${INTEL_GPU,,} =~ (hd\ graphics|uhd\ graphics|xe|iris|arc|panther\ lake) ]]; then

Intel ships many recent parts as a bare [Intel Graphics] with no Iris/UHD/Arc branding, so the allowlist matches nothing and the leaf silently installs no VA-API driver at all.

25 device IDs across 13 families miss, per pci.ids:

Family Examples
Raptor Lake-P/U 8086:a7aa a7ab a7ac a7ad
Meteor Lake-M/P 7d40 7d45 7d60 7dd5
Arrow Lake-U/S/P/H 7d41 7d67 7dd1 b640
Lunar Lake 6420 64b0
Alder Lake-N 46d3 46d4
Wildcat Lake fd80 fd81
Battlemage G21/G31 e202 e20d e210 e215 e216 e220 e221

Battlemage is discrete Arc and misses despite the arc token, because pci.ids does not spell out the brand for those boards.

panther lake is in that allowlist because this same failure was already hit on one family and patched there rather than at the root — so it regressed again for every other family.

Evidence from an affected machine

Intel Raptor Lake-P [8086:a7ab], Omarchy 4.0.2. After a normal install:

  • vpl-gpu-rt — not installed (3rd package on that omarchy-pkg-add line, so the line never ran)
  • libvpl — installed, but as an ffmpeg dependency
  • intel-media-driver — not installed
  • vulkan-intel — installed

That last one is the contrast. install/hardware/vulkan.sh runs in the same install and matches on vendor ((VGA|Display).*Intel); it worked. This leaf matches on model names; it did not.

The practical result was no iHD_drv_video.so on disk at all, so the only VA-API driver present was NVIDIA's — which on a hybrid laptop means video decodes on the dGPU and every frame crosses the PRIME boundary. That surfaced as flicker while scrolling video (see #9812, which is the other half of this and independent of it).

Fix

Invert the test. Any Intel display device gets intel-media-driver; only the pre-Broadwell parts still need a name, and that set stopped growing in 2014. New hardware now works by default instead of waiting for another token to be added.

Also select the device with lspci -d 8086::0300. The old grep -iE 'vga|3d|display' matched unrelated lines whose hex device ID contains 3d — on the reporting machine it pulled in a PCI bridge, Device a73d, alongside the GPU.

Tradeoff

Pre-Broadwell parts that are not named GMA — Ivy Bridge, Haswell — now get intel-media-driver, which cannot drive them, leaving decode in software. That is exactly what they already got from matching no token at all, so this is not a regression, but it is not a fix for them either. Handling them properly needs generation detection rather than name matching, which felt like a separate change.

Tests

test/shell.d/intel-video-acceleration-test.sh stubs lspci and omarchy-pkg-add and covers each missed family, discrete Battlemage, the GMA fallback, no-Intel-GPU, and the previously-matching names (Iris Xe, UHD, Panther Lake) as regression cases.

Against the current script it fails exactly as reported:

expected: pkg-add intel-media-driver libvpl vpl-gpu-rt
actual:
not ok - Raptor Lake reports no marketing name and still gets a driver

test/cli passes. test/shell has 3 failures (config-test, snapper-test, unowned-system-paths-test) that reproduce identically on unmodified quattro.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BT3H7mj63CxzHWKLd4apWk

Fixes #13987

video-acceleration.sh matched the lspci line against an allowlist of
marketing names: hd graphics, uhd graphics, xe, iris, arc, panther lake.
Intel ships many recent parts as a bare "[Intel Graphics]" with no such
branding, so the allowlist matched nothing and the leaf silently installed
no VA-API driver at all.

25 device IDs across 13 families miss, including Raptor Lake-P/U, all of
Meteor, Arrow and Lunar Lake, Alder Lake-N, Wildcat Lake, and discrete
Battlemage -- which misses despite the "arc" token, because pci.ids does not
spell out the brand for those boards. "panther lake" is in that list because
the same failure was already hit on one family and patched there rather than
at the root, so it regressed again for everyone else.

Invert the test: any Intel display device gets intel-media-driver, and only
the pre-Broadwell parts still need a name. That set stopped growing in 2014,
so new hardware now works by default instead of waiting for another token.

Select the device with lspci -d 8086::0300 as well. The old grep -iE
'vga|3d|display' also matched unrelated lines whose hex device ID contains
"3d" -- on the reporting machine it pulled in a PCI bridge, "Device a73d".

Pre-Broadwell parts that are not named GMA, such as Ivy Bridge and Haswell,
now get intel-media-driver, which cannot drive them and leaves decoding in
software. That is what they already got from matching no token at all, so
this is not a regression, but it is not a fix for them either.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BT3H7mj63CxzHWKLd4apWk
@omarchybot

Copy link
Copy Markdown
Collaborator

Reviewed at 8d4a8294 by Claude Opus 5.5 in Claude Code, with a second opinion from Codex Medium. Nothing was run on a worker, for the reason at the end.

The bug is real. In the current pci.ids (hwdata 0.411), every one of the 25 device IDs in the description is named <family> [Intel Graphics], and none of those names matches the allowlist on quattro, so these machines get no VA-API driver at all. #13987 (Meteor Lake) and #11958 (Wildcat Lake) report the same failure from real machines.

#8514 fixes the same bug a different way: it picks the driver from the PCI device ID in sysfs. Both models compared the two side by side, and both conclude that #8514 is the one to take forward:

  • install/hardware/intel/video-acceleration.sh:8 uses lspci -d 8086::0300, which matches only class 0300 (VGA compatible controller). An Intel GPU that is not the boot display reports class 0380 (Display controller) instead. That covers an iGPU on a desktop where a discrete card is primary, and a secondary Arc card. The display pattern on quattro catches these today; with this change they would get no driver. Install Intel VAAPI by PCI generation on hybrid GPUs #8514 accepts any 0x03* class. The test's lspci stub ignores its arguments, so the suite cannot see this.
  • Ivy Bridge and Haswell parts that are not named GMA would get intel-media-driver, which cannot drive them. Two examples are 0166 ("3rd Gen Core processor Graphics Controller") and 0416 ("4th Gen Core Processor Integrated Graphics Controller"). The description already names this as the tradeoff. Install Intel VAAPI by PCI generation on hybrid GPUs #8514 gives these parts libva-intel-driver.
  • This change reaches only new installs. Install Intel VAAPI by PCI generation on hybrid GPUs #8514 adds a migration, so machines already installed without the driver get it on their next update.

#8514 is much larger: a 120-line table of ID ranges, against 12 lines here. The second opinion also found one ancient part it would misroute to iHD (8086:7121, i810). It still covers the hardware this change misses. The second opinion was given the class 0380 question to check, so its agreement on that point is not independent. On Haswell and Ivy Bridge and on the migration, it reached the same reading as mine.

Only one of the two should land, so this pull request was not run on a worker. Choosing between competing fixes is the maintainer's call, and this is waiting on them.

@omarchybot

Copy link
Copy Markdown
Collaborator

Reviewed again at 8d4a8294 by Claude Opus 5.5 in Claude Code, with a second opinion from Codex Medium. This branch has not changed since the last review, but a third fix for the same bug has opened since then: #14252, which adds intel graphics to the name allowlist and a migration.

Both models compared all three. The conclusion is the same as before: #8514 is the one to take forward. It is the oldest of the three, and it is the only one that handles Intel GPUs reporting class 0380, gives Ivy Bridge and Haswell libva-intel-driver, and repairs existing installs. The previous comment is public on this pull request, so the second opinion's agreement is not independent.

This pull request does have one thing #8514 lacks: test cases for the families that pci.ids names only [Intel Graphics]. #8514's detector test covers one Meteor Lake ID (7d55), but none of the Raptor Lake-P/U, Arrow Lake, Lunar Lake, Alder Lake-N, Wildcat Lake or Battlemage IDs listed in this description, and no class 0380 device. Those cases are being carried into #8514's test, credited to @ifigueroap with a Co-Authored-By trailer, as part of #8514's own review.

Nothing was run on a worker. The choice between these fixes is the maintainer's, and this is waiting on them.

omarchybot added a commit to atavacron/omarchy that referenced this pull request Oct 13, 2026
The name regex this replaces missed every part whose pci.ids entry is a bare "[Intel Graphics]": Raptor Lake-P/U, Meteor Lake, Arrow Lake, Lunar Lake, Alder Lake-N, Wildcat Lake and discrete Battlemage. Pin all 25 of those ids to intel-media-driver, and pin that an Intel GPU reporting display class 0380, as a non-boot GPU does, still gets its package. The ids and the class 0380 case come from omacom#9824.

Co-Authored-By: Ismael Figueroa <ifigueroap@gmail.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

bug Something isn't working

Projects

None yet

2 participants