Table of Contents

fourier — patch series in detail

Snapshot of the four sibling patch series as of 2026-04-28 evening. Index: fourier.

chromium-fourier (4 patches, validated end-to-end on ohm)

Reference target: ohm — PineTab2 / RK3566 / Mali-G52 / panfrost / mainline kernel 6.19.10 / KDE Plasma 6.6.4 Wayland.

  1. enable-v4l2-decoder-default.patch — flip kAcceleratedVideoDecodeLinux default
  2. wayland-allow-direct-egl-gles2.patch — re-allow direct EGL/GLES2 in WaylandSurfaceFactory
  3. nv12-external-oes-on-modifier-external-only.patch — pick GL_TEXTURE_EXTERNAL_OES for NV12 dmabufs whose modifier is external_only
  4. v4l2-capture-pool-floor-at-16.patch — floor V4L2 capture pool at 16 buffers

Result on ohm: bbb_1080p30_h264.mp4 plays end-to-end on KDE Plasma 6.6.4 Wayland, ~81 % combined chrome CPU vs ~131 % pre-patch baseline (~3.8× reduction).

Side-effect: Brave's YouTube playback also feels markedly snappier on ohm — patches 2 + 3 are not chromium-fourier-binary-specific, they help any Wayland Chromium-derived browser through the panfrost external_only modifier path, and patch from the kwin layer (see below) helps every wp_linux_dmabuf-v1 client.

Build: cross-compiled on CT 220 chromium-builder-x86 (Proxmox container on data) — see fourier:hosts.

qt6-base-fourier (3 patches, validated on ohm + fresnel)

Pure spec-correctness fix. OpenGL ES 3.0 deprecated GL_ALPHA as a glTexImage2D internalFormat; Qt 6 was hard-coding it via the QT_CONFIG(opengles2) build branch with no runtime ES-version check. Affects every Qt 6 GLES3 client on Mali-class hardware.

  1. 0001-qopengltextureglyphcache-pick-GL_R8-on-ES3.patch
  2. 0002-qrhigles2-RED_OR_ALPHA8-pick-GL_R8-on-ES3.patch
  3. 0003-qopengltextureuploader-pick-GL_R8-on-ES3.patch

Zero GL_INVALID_VALUE in journal post-relogin. Cleanly upstream-targetable to bugreports.qt.io + Gerrit; submission body is written and parked in upstream-submissions/qt6-base-fourier/qt-bug-report.md.

kwin-fourier (2 patches — one shipped, one staged)

  1. 0001-transaction-bypass-watchDmaBuf-fence-wait.patch — diagnostic bypass; this is what identified the bug
  2. 0002-transaction-poll-dmabuf-fd-directly-upstream-shape.patch — the actual upstream-bound shape: poll the dmabuf fd directly instead of going through EXPORT_SYNC_FILE + QSocketNotifier round-trip; validated on ohm

Both are shaped to address the per-frame DMA_BUF_IOCTL_EXPORT_SYNC_FILE overhead in KWin's compositor scheduling path. Side-effect benefit: every wp_linux_dmabuf-v1 client (chrome, brave, mpv, VLC) feels markedly snappier on Mali-class hardware.

The discovery story is in arch/chromium-fourier/KWIN_PIVOT.md in gitea/marfrit-packages — see also journey.

The upstream-shape patch's submission body is at upstream-submissions/kwin-fourier/kde-mr-body.md.

firefox-fourier 150.0.1 (4 patches, partial validation on fresnel)

  1. patch 1 — gfxinfo fourcc advertise (verified working)
  2. patch 2 — libwrapper hwdevice route (compiled in but not engaging at runtime)
  3. patch 3 — ffmpegvideo route (compiled in but not engaging at runtime — root cause identified today)
  4. patch 4 — pref default (verified working)

Today (2026-04-28) the runtime defect on patch 3 was identified via MOZ_LOG: patch 3 only walked the generic codec's hw_configs. libavcodec on ALARM registers v4l2_request as separate AVCodec entries (h264_v4l2request etc.). Patch 3 has been corrected to probe by name first, falling back to hw_configs. Rebuild pending on boltzmann.

Where to find sources