enricobuehler
a6e51215fd
The native VAAPI decoder now runs end to end: pf-vaadec's plans go into libva's buffers, the surface comes back as DRM-PRIME dmabufs, and the presenter imports them exactly as it does the FFmpeg rung's. Pin-only — `PUNKTFUNK_DECODER=native-vaapi` — for the reason M5's D3D11VA rung was: `auto` admission is earned with hardware parity and a soak, and this rung has decoded nothing yet. libva is dlopen'd rather than linked, so the pf-lxcheck2 container compiles and clippies the whole thing without libva-dev, and a machine without a VAAPI runtime gets a clean refusal instead of a packaging dependency. The surface pool is not the slot map. `SlotMap::assign` hands out the lowest free slot, and a slot freed by an access unit's own removals is free by the time that unit's picture takes it — measured at 225 of the vendored vector's 250 access units. A surface bound by slot index would therefore decode, on nine frames in ten, into the surface still holding the picture on screen. So `plan_to_va` now takes the decode target as a parameter, bound by the caller at activation time the way pf-vkdecode binds a pool image, and a surface is free only when no live picture is bound to it, no output is owed for it, and no consumer holds it. Measured rather than transcribed, as everywhere else here: layout-probe.c grew the export descriptor (312 bytes, objects[4]/layers[4]), the buffer-type enumerators — VASliceParameterBufferType is 4 and VASliceDataBufferType is 5, not the 3 and 4 that counting off the header suggests — and the config, attribute and generic-value layouts. All pinned as compile-time assertions, which is how the 12-byte VAGenericValue in the first draft was caught: the C union holds a pointer, so it is 8-aligned and 16 bytes. The plane walk lives in pf-vaadec, pure and unit-tested on macOS, because it is the one structure the DRIVER writes and we read: SEPARATE_LAYERS returns NV12 as two layers, and taking layers[0] is the green screen this project has already paid for. It also refuses what it cannot express rather than guessing — a bogus object count, a plane naming an object that is not there, objects disagreeing on tiling. Own DecodedImage variant, same payload type. The physical hand-off is identical to the FFmpeg rung's, so the presenter keeps ONE arm and one demotion streak; the variant exists so the compiler asks which rung decoded wherever that matters. Both D3D11VA rungs share a variant and `1573a987` had to fix the consequence afterwards — a "native" soak that could silently have been an FFmpeg soak. Here the four uncovered matches were compile errors. Buffers are destroyed by us, not by vaEndPicture: va.h is explicit that the user must call vaDestroyBuffer, and the libva 0.x behaviour is long gone. Leaking two per picture at 60 fps exhausts the driver's store in minutes. pf-vaadec's presenter headroom was 4, written against no consumer. The Vulkan rung had already measured the client pipeline at four to seven held frames; it is 8 now, pinned to that crate's constant so a re-measurement moves both. Gates: macOS fmt/clippy/341 tests/cargo doc, and in the container clippy -D warnings over six crates, 795 tests, workspace check. Hardware legs are still owed — no AMD/Mesa or Intel box was reachable.
punktfunk-session
The Vulkan session binary: one stream per invocation in an SDL3 window — no UI toolkit,
no widgets, terminal stats. The power-user / gamescope stream client, and the stage-2
presenter of the Linux client re-architecture (punktfunk-planning:
linux-client-rearchitecture.md).
This binary is deliberately dumb: a renderer the front-ends call INTO — the GTK shell
(punktfunk-client), the WinUI shell, and the punktfunk CLI all spawn it through the
same brain (pf_client_core::orchestrate), which resolves policy (profiles, settings,
wake) and hands the result down, normally as a --resolved-spec file. It reads the
shared stores only as the compat fallback for a bare hand-launched invocation.
punktfunk-session --connect host[:port] [--fp HEX] [--launch id] [--fullscreen] [--stats]
punktfunk-session --browse host[:port] [--mgmt PORT] [--fullscreen]
--browse opens the console game library (the Skia coverflow over the animated aurora)
instead of connecting: A launches the focused title as a stream in the same window,
session end returns to the library, B quits (Gaming Mode returns). Paired hosts only —
pairing is the desktop client / Decky plugin's job. PUNKTFUNK_FAKE_LIBRARY=<file.json>
feeds canned entries with no host (portrait paths starting with / load from disk).
Reads the same identity / known-hosts / settings stores as the desktop client
(punktfunk-client), so enrolling on either side makes the other work; this binary never
connects to a host it has no pinned fingerprint for (--fp HEX overrides the store).
Pairing is punktfunk pair <host> — the CLI, which ships alongside this binary in every
package and needs no window and no toolkit either. punktfunk-session --pair still works
for one release (someone's provisioning script calls it today) but prints a deprecation
notice: pairing is a trust ceremony and belongs to the brain, not a renderer.
Stdout is the machine interface: {"ready":true} after the first presented frame,
stats: … once per second while the overlay tier isn't Off (always the full detailed
text, whatever the OSD shows; --stats forces the overlay on), one
{"error"|"ended": …} JSON line on the way out. Logs go to stderr. Exit codes: 0
clean end, 2 connect failed, 3 trust rejected / pairing required, 4 presenter
init failed.
In-stream keys match the desktop client: click captures input (Ctrl+Alt+Shift+Q releases), Ctrl+Alt+Shift+D disconnects, F11 toggles fullscreen; the controller escape chord (L1+R1+Start+Select, hold to disconnect) works the same.
The default build carries the Skia console UI (ui feature): the stats OSD and capture
hint render in-window. Ctrl+Alt+Shift+S cycles the OSD tier live — Off → Compact (one
line: fps · latency · Mb/s) → Normal (mode + end-to-end percentiles) → Detailed (decoder
path + per-stage latency equation); any tier but Off also emits the stdout mirror.
--no-default-features is the ~5 MB power-user build — same streaming, stats on stdout
only, no Skia anywhere in the dependency tree.
Decode follows the Settings preference (auto is vendor-ordered: hardware Vulkan Video →
VAAPI → software on Linux, hardware Vulkan Video → D3D11VA → software on Windows, with
VAAPI/D3D11VA first on Intel; on H.264 and HEVC the native pf-vkdecode Vulkan decoder
is tried immediately before FFmpeg-Vulkan): the Vulkan decoders run on the presenter's own
device where the stack supports it (every vendor, zero copy); VAAPI dmabufs import
per-plane elsewhere (D3D11VA textures on Windows); software is the universal fallback.
10-bit Main10 and HDR10 are advertised (VIDEO_CAP_10BIT|HDR): P010 decodes through the
native, FFmpeg-Vulkan, VAAPI/D3D11VA and software paths alike, and PQ streams present
on an HDR10/ST.2084 swapchain when the desktop offers one (KDE HDR, gamescope) or
tone-map in-shader to SDR when it doesn't (PUNKTFUNK_TONEMAP_PEAK tunes the rolloff,
default ≈1000 nits). The host still gates the upgrade behind its PUNKTFUNK_10BIT
policy.
Debug/bisect knobs: PUNKTFUNK_DECODER=native-vulkan|native-vaapi|native-d3d11va|vulkan|vaapi|d3d11va|software
(the three native-* values pin this program's own decoders; native-vaapi also takes
PUNKTFUNK_VAAPI_DEVICE=/dev/dri/renderDNNN to choose the GPU), PUNKTFUNK_PRESENT_MODE= mailbox|fifo|immediate|fifo_relaxed (default MAILBOX, FIFO where the surface offers no
MAILBOX — AMD on Windows), PUNKTFUNK_VK_DEVICE=<index> (multi-GPU), and
PUNKTFUNK_HW_FAULT=import (fault every VAAPI dmabuf import — proves the three-strike
demotion to software on healthy hardware).
PUNKTFUNK_AU_FAULT=drop|truncate|flip[:period] deliberately corrupts decoder input on the
native Vulkan lane (default period 60 — one AU a second at 60 fps; inert everywhere else, and
inert entirely if the value doesn't parse). drop swallows the AU, so the next one references a
picture that was never decoded — the bitstream planner catches it immediately. truncate delivers
a picture whose slice data stops mid-frame and flip alters one byte deep in the payload: both
parse perfectly, so only the driver's per-frame decode-status query can see them, and neither is
visible at all on a driver without queryResultStatusSupport or on any FFmpeg lane. Watch the
result on the Detailed stats line's integrity: term (damaged = concealment the planner caught,
refused = AUs the decoder rejected outright, driver-failed = the hardware's own verdict, run
= consecutive frames with no picture, worst run = the longest such stretch of the session — the
once-a-second run sample misses the bad moment almost every time — and no driver status = this
device cannot answer the driver question at all). A session that lands on any other lane says so
in the log rather than faulting silently.
Note that PUNKTFUNK_AU_DUMP records the AU as it arrived from the HOST, while the fault injector
runs later, at the native decoder's own entry. On a faulted run the dump is therefore the clean
bitstream — reconstruct the damaged bytes from the spec if you need them (the injector is pure and
deterministic).