fourier

A campaign of patches that makes mainline-Linux Wayland video playback on ARM Rockchip SoCs work end-to-end on a "boring upstream" stack — no vendor MPP, no Mali blob, no panfork, no 5.10 BSP kernel, no ANGLE-on-Vulkan-only Mali fallback, no NV12-to-AR24 software conversion in the renderer.

Started as one patch series for chromium V4L2 hardware decode. Became four, because each layer of the stack had its own contributing bug. The patches each live in their own directory of this repo — see "Repository plan" below).

What we found, layer by layer

Layer	Bug	Patch
chromium runtime gate	media::kAcceleratedVideoDecodeLinux defaults to disabled when USE_VAAPI isn't set, even on a use_v4l2_codec=true build. Silent fallback to ffmpeg software decode.	enable-v4l2-decoder-default.patch
chromium ozone GL surface advert	WaylandSurfaceFactory::GetAllowedGLImplementations() strips kGLImplementationEGLGLES2, blocking direct-EGL / panfrost native dmabuf import surfacing.	wayland-allow-direct-egl-gles2.patch
chromium NV12 dmabuf import	OzoneImageGLTexturesHolder::GetBinding picks GL_TEXTURE_2D for V4L2-produced NV12 dmabufs. ANGLE rejects YUV EGLImage on non-EXTERNAL_OES target, force-falls to NV12→AR24 software conversion. Skia Ganesh handles GL_TEXTURE_EXTERNAL_OES natively; this branch was just missed.	nv12-external-oes-on-modifier-external-only.patch
chromium V4L2 capture pool depth	V4L2VideoDecoder::ContinueChangeResolution requests num_codec_reference_frames + 2 = 6 buffers for H.264 main — exactly equal to chrome's wayland pipeline depth. First scheduling jitter exhausts the pool and hard-stalls the decoder.	v4l2-capture-pool-floor-at-16.patch
Qt 6 (QOpenGLTextureGlyphCache, QRhiGles2, QOpenGLTextureUploader)	QT_CONFIG(opengles2) build branch unconditionally picks GL_ALPHA as glTexImage2D internalFormat. GL_ALPHA was deprecated in OpenGL ES 3.0 (only sized formats valid). On Mali ES 3.2 contexts every text-glyph cache upload errors GL_INVALID_VALUE; KWin's debug callback fills the journal and adds enough load to mask other bugs.	qt6-base-fourier
KWin 6.6.4 transaction scheduling	Transaction::watchDmaBuf calls DMA_BUF_IOCTL_EXPORT_SYNC_FILE on every imported dmabuf and parks the transaction on a QSocketNotifier(POLLIN). For V4L2-produced dmabufs the kernel returns a stub fence (because vb2 doesn't populate dma_resv — see kernel layer below). Even when the fence is real, the export-then-poll roundtrip adds milliseconds per frame.	kwin-fourier
Linux kernel (videobuf2-core, hantro, rockchip-rga)	vb2 doesn't propagate V4L2 buffer-state-done into the dmabuf's dma_resv exclusive fence. Wayland compositors (and any other userspace that thinks it's doing implicit-sync the modern way) get a stub fence representing nothing real. The architectural hole that the KWin layer above is papering over.	open

Validation (2026-04-28)

End-to-end smooth 1080p30 H.264 playback under KDE Plasma 6.6.4 Wayland on ohm (PineTab2 / RK3566 / Mali-G52 / panfrost mesa 26.0.5 / mainline kernel 6.19.10) with the full patch chain installed:

• bbb_1080p30_h264.mp4 plays through to EOF, no stall, no slideshow.
• Combined Chromium CPU: ~81 % across all chrome procs vs ~131 % baseline pre-patch (NV12→AR24 software conversion). Playback at ~34 % combined was briefly observed pre-stall and is reachable again on a less jitter-prone compositor session.
• KWin Wayland session also feels markedly snappier for unrelated video clients (Brave on YouTube, mpv, VLC) — the kwin-fourier watchDmaBuf bypass is a general-purpose latency reduction for every wp_linux_dmabuf client on this hardware.

Repository layout

fourier/
├── README.md                   (this file)
├── LICENSE
├── docs/
│   ├── dmabuf-zero-copy.md     # NV12 external_only investigation;
│   │                           # chronicles the full discovery thread
│   └── playback-howto.md       # mpv / gstreamer / ffmpeg recipes
├── tools/
│   └── dmabuf-modifiers.c      # tiny EGL modifier-table dumper
├── chromium-fourier/           # 4 patches; per-board PKGBUILD
│   ├── pinetab2/               #   RK3566 / Mali-G52 / panfrost — validated
│   ├── rk3399/                 #   RK3399 / Mali-T860 / panfrost — validated
│   └── rk3588/                 #   RK3588 / Mali-G610 / panthor  — validated
├── qt6-base-fourier/           # 3 patches against qtbase 6.11.0
│   └── PKGBUILD                # pacman epoch=1 to dominate Arch's pkgrel
└── kwin-fourier/               # 1 patch against kwin 6.6.4 (diagnostic)
    └── PKGBUILD                # pacman epoch=1

Each *-fourier/ subdirectory carries a PKGBUILD plus the patch files. The PKGBUILDs target pacman / Arch Linux (ARM and x86_64); the patches themselves are distribution-agnostic and apply unchanged to vanilla upstream sources.

Building

cd chromium-fourier/pinetab2     # or rk3399, rk3588
makepkg -si

Downloads the upstream chromium release tarball, applies the four fourier patches, runs gn gen + ninja, installs into /usr/lib/chromium/ plus a launcher shim at /usr/bin/chromium. Plan for 10+ hours of build time on an RK3566 — and roughly twice the RAM-to-disk swap headroom that pacman's normal heuristics suggest. Cross-compiling from x86_64 (the target_cpu="x64" branch in the same PKGBUILD) is significantly faster.

For the Qt 6 and KWin components:

cd qt6-base-fourier && makepkg -si
cd kwin-fourier && makepkg -si

Both inherit from their upstream Arch packaging and apply just the fourier patches on top, so they're proportionally cheap rebuilds (qt6-base ~30 min on aarch64, kwin ~15 min). Both bump epoch=1 so they dominate normal Arch upstream pkgrel updates until the patches land upstream.

Patches (chromium part)

enable-v4l2-decoder-default.patch

Flips media::kAcceleratedVideoDecodeLinux (user-visible feature name AcceleratedVideoDecoder) to FEATURE_ENABLED_BY_DEFAULT when the build sets BUILDFLAG(USE_V4L2_CODEC), symmetric with the existing USE_VAAPI arm.

wayland-allow-direct-egl-gles2.patch

Re-allows kGLImplementationEGLGLES2 (--use-gl=egl, no ANGLE shim) in WaylandSurfaceFactory::GetAllowedGLImplementations(). The downstream dispatcher in CreateViewGLSurface already handles that implementation; only the advertised list was tightened upstream.

Caveat: with is_official_build=false, the gn default for dcheck_always_on is true, and the direct EGL path FATALs at gl_context_egl.cc:241 (DCHECK(!global_texture_share_group_)). The PKGBUILD therefore explicitly sets dcheck_always_on=false.

nv12-external-oes-on-modifier-external-only.patch

Extends OzoneImageGLTexturesHolder::GetBinding to pick GL_TEXTURE_EXTERNAL_OES whenever the EGL driver advertises the pixmap's DRM modifier as external_only for the given fourcc, in addition to the existing format.PrefersExternalSampler() branch. Adds a static helper NativePixmapEGLBinding::ModifierRequiresExternalOES that queries eglQueryDmaBufModifiersEXT and caches the answer per (fourcc, modifier) tuple.

v4l2-capture-pool-floor-at-16.patch

Floors v4l2_num_buffers at 16 in V4L2VideoDecoder::ContinueChangeResolution when V4L2 is the buffer allocator. Default (num_codec_reference_frames + 2) is 6 for H.264 main, equal to chrome's wayland compositor pipeline depth — first scheduling jitter event exhausts the pool. The std::max preserves correctness for high-DPB codecs (HEVC, VP9 with deeper reference counts).

Runtime

The launcher shim (/usr/bin/chromium) defaults to:

--ozone-platform=wayland
--use-gl=angle --use-angle=gles
--enable-features=AcceleratedVideoDecoder
--disable-features=Vulkan

Vulkan is disabled by default because panvk on RK3566 (Mali-G52 Bifrost) returns VK_ERROR_INCOMPATIBLE_DRIVER on chromium's probe and breaks V4L2 dispatch downstream. Override paths if you're working on a board where Vulkan actually works (rk3588/ defaults Vulkan-on):

User flag	Effect
--enable-features=Vulkan	Enable Vulkan (panthor / others)
--use-vulkan=native	Pick the native Vulkan backend
--use-vulkan=swiftshader	Force the SwANGLE Vulkan backend
--disable-features=Vulkan	Explicit re-disable (rarely needed)
--use-angle=vulkan	Run ANGLE on Vulkan

The launcher detects any of those on the command line and skips its own --disable-features=Vulkan, so the user's intent always wins.

Adjacent

• docs/dmabuf-zero-copy.md — the panfrost / panthor external_only investigation; the original deep-dive that produced patch 3 and pointed at the rest of the chain.
• docs/playback-howto.md — mpv / gstreamer / ffmpeg invocations that work on the same V4L2 stack outside the browser.
• tools/dmabuf-modifiers.c — a 100-line EGL probe that dumps the available NV12 modifiers and their external_only flags. Useful for sanity-checking what your panfrost / panthor / mesa version advertises before you start patching.

A separate Firefox-side effort (firefox-fourier, in marfrit-packages) carries the same V4L2 stateless unlock for Firefox 150 — different patch shape because Firefox's media-rdd / RemoteVideoDecoder plumbing is independent.

License

The patches are released under the same BSD-3-Clause as Chromium itself. See LICENSE. The qt6-base-fourier and kwin-fourier patches are released under the LGPL-2.1+ matching their upstreams.