DirectAudio from its own release: driver set per Proton audio interface (CachyOS fix), the Steam client's own DirectAudio sink, binaries out of tree
Games
- Two driver sets ship from the pinned DirectAudio release (tools/directaudio/release.env =
directaudio-linux-v1.1.0, sha256-checked by tools/directaudio/fetch.sh in CI and
build_local.sh): linux-wine11 for Valve Proton 11 / Experimental / GE-Proton, and
linux-wine11-systhread for Proton-CachyOS, whose Wine carries the system-thread mmdevapi
interface. steam-compatibility picks by whether the Proton's own winepulse.so imports
PsCreateSystemThread, never by the version string (both say wine-11.0; the wrong set is
silence). No matching set: DirectAudio off, Proton's audio runs, logged.
- Proton-CachyOS's launcher pre-sets WINE_AUDIO_DRIVER=pulse,alsa and its mmdevapi reads that
before the registry (Valve's and GE's do not read it): the wrapper exports
WINE_AUDIO_DRIVER=directaudio,pulse as well as writing the registry. On CachyOS this also
makes DirectAudio engage on a game's first launch.
- A wineserver from an earlier Steam verb can save its in-memory registry over the appended
Audio value; the wrapper says so when one is live. Writing through it (wine reg add) and
creating a missing prefix first (proton run wineboot) were tried on the device and withdrawn:
a second Proton/Wine run inside Steam's launch sequence got launches aborted.
- Tests: both sets, no set, no prefix (34 pass).
Steam client audio
- DirectAudio = module-directaudio-native-sink, the game driver's AAudio engine inside the
PulseAudio daemon, one step from Android (adaptive buffer, two-burst floor, primed ring with
fades, reopen on route change, watchdog, limiter). Default. Shared output = the previous relay
route, kept. New pref clientAudioRoute; the old boolean is not read. If the native sink cannot
open its stream the daemon keeps running (.nofail) and ensureClientSink loads the plain AAudio
sink instead, logged to audio.log. Audio bundle stamp bumped.
Out of tree
- app/src/main/assets/directaudio/ is generated and ignored; tools/directaudio-relay/ and the
vendored module-directaudio-sink.c + da_relay_proto.h are gone; tools/aaudio-sink keeps only
module-aaudio-sink (the fallback). The release carries the relay and sink field fixes by
MaxsTechReview (#338), credited there and in its AUTHORS.
Device-proven on an AYANEO Pocket FIT: client sink 12 ms / 4 ms bursts, 0 underruns; FEZ on
cachyos-11.0-20261005-slr on DirectAudio (took pulse before); FEZ on Proton Experimental
unchanged; Curse of the Dead Gods on CachyOS from a cold prefix, DirectAudio on first launch.
Co-authored-by: The412Banner <the412banner@users.noreply.github.com>
* Build the DirectAudio relay from source
The relay helper shipped as a prebuilt libdirectaudiorelay.so with no source in the
repo. Its source (The412Banner/directaudio @ 3548a4a, the commit the bundled driver
names in version.txt) now lives in tools/directaudio-relay, unmodified, and CI and
build_local.sh compile it into jniLibs the way proot is built. It shares
da_relay_proto.h with the PulseAudio sink, which is byte-identical to its own copy.
* DirectAudio relay: prime before playing, one consumer, safe buffer floors
- The output stream started before the client had its ring, so every connection
opened with a run of short callbacks: six clicks per stream on a RedMagic, even idle.
The callback now plays silence until the ring holds its target, and only a real
underrun afterwards counts.
- An underrun ends the block with a short fade and waits for the ring to refill to
target before resuming with a fade in, instead of zero-padding every short callback.
Each callback spent starved grows the target, so one stall sizes it for the next.
With a producer freezing 60 ms every 2 s: 14 underruns before, 1 after.
- During a reopen the outgoing and incoming streams could both drain the ring. Only
the promoted stream reads it now; any other plays silence.
- The ring target starts at 40 ms instead of two bursts (8 ms) and keeps what it
learned across reopens.
- The device buffer is at least two bursts; 12 ms was a single burst on devices that
grant a 20 ms legacy stream.
* PulseAudio: the classic path plays through the native-rate AAudio sink
The classic path loaded the prebuilt 0.1.5 module, which opened AAudio at the daemon's
default 44.1 kHz: Android gave it a legacy stream (20 ms bursts) and resampled, after
PulseAudio had already resampled the 48 kHz clients down. Its data callback also waited
on the daemon's IO thread, and it started at two bursts and grew only after an audible
underrun. The app's own module-aaudio-sink, built by CI into the same bundle, now
carries that path:
- The stream opens at the device's rate (48 kHz) with the buffer at two bursts or
30 ms, whichever is more, still growing a burst per underrun.
- A write that times out part way keeps the rest for the next write instead of
dropping it.
- A stream reopened after a route change asks for the rate the first one got, so the
sink keeps its format instead of unloading.
* Audio: DirectAudio sink at audio priority, deeper client buffers, no decay probing
- module-directaudio-sink's IO thread asks for nice -16 like module-aaudio-sink's, so
the client's producer keeps up under game load.
- PULSE_LATENCY_MSEC=60 for the session: libpulse clients under FEX (the Steam
client, winepulse when DirectAudio is off) keep enough queued to ride out a late
wakeup.
- BANNER_AUDIO_DIRECT_DECAY=0: the driver's decay shrinks the device buffer every
10 s of quiet until it underruns again, a click by design.
Both sit ahead of Downloads/droiddeck-env and per-game environment, which still win.
* Audio: drop the session-wide PULSE_LATENCY_MSEC
On an AYN Thor it took the Steam client's stream to 110 ms (60 ms tlength on top of the
sink's 40 ms) with no underruns to show for it, and the delay was audible. Clients pick
their own buffers again; the per-game option still sets it.
* PulseAudio sinks: wait in the poll loop, never in a blocking call
Clients' audio reaches a sink as messages its IO thread handles in pa_rtpoll_run. The AAudio
sink spent each iteration inside a blocking AAudioStream_write for up to a burst, and the
DirectAudio sink inside a futex wait for up to a burst, so a client's data sat unread and its
stream ran dry: a Chromium-shaped client (44.1 kHz float, 30 ms tlength, 10 ms minreq) had
304 underflows in 30 s on the AAudio sink, heard as crackling in the Steam client. Bursts of
20 ms (a legacy stream, an AYN Thor) made it worst.
- module-aaudio-sink writes without blocking, renders only what the device has room for, and
comes back after half a burst. 0 underflows in 30 s, with and without 8 busy cores.
- module-directaudio-sink tops the ring up on the same half-burst timer instead of sleeping
on the relay's futex.
* DirectAudio relay: vendor the source the shipped binary was built from
The import took 3548a4a, the commit the bundled driver names in version.txt, but the relay
binary that shipped came from e4727e9 on the same branch (feat/linux-relay-mic): its strings
match a build of that source exactly. 3548a4a has no --mic-fifo, so with the microphone on the
relay printed its usage and exited, and DirectAudio played nothing. This is e4727e9 as
published; the fixes made on top of 3548a4a follow in the next commit.
* DirectAudio relay: prime before playing, one consumer, safe buffer floors (on e4727e9)
The same changes as ca5d39b, carried onto the e4727e9 source: playback waits until the ring
holds its target, an underrun fades out and re-primes, starved callbacks grow the target, only
the promoted stream drains the ring, the ring target starts at 40 ms and survives reopens, and
the device buffer is at least two bursts.
* Steam client audio: DirectAudio by default
A fresh install plays the Steam client through the DirectAudio relay. Choosing Classic in the
Steam cog is stored and kept; only an install that never chose falls back to DirectAudio.
A second sink module: connects to the relay's socket (lazily, the relay comes up after the
daemon; a clocked null sink until then), sends the hello, maps the render ring it gets back and
keeps it topped up to the relay's target with a futex wait per burst. Every AAudio call, the
adaptive buffer and the route handling stay in the relay, outside proot. With DirectAudio on the
daemon loads this sink instead of module-aaudio-sink; the bundle carries both.
Blocking-write design: the IO thread pulls one burst at a time from the sink and hands it to a
blocking AAudio output stream, whose buffer paces the thread; latency comes from the stream's
playback timestamp; with adaptive=1 the buffer grows a burst per underrun up to a cap. Same
module arguments as before, so the daemon's config line is unchanged. build.sh compiles it with
the NDK against the app's own libpulse* and a small config.h; a workflow checks it on every push.