Mixer: a console of strips with meters, and the execution reporting back without stopping the audio

[jcelerier]

A rework of the audio mixer panel and of how the execution reports back to the interface. 35 commits, each with its test where one applies. Everything visual was checked in a running score rather than by unit test alone.

Needs ossia/libossia#950 and celtera/avendish#210. The submodule pointers here already reference them. An adversarial review of the whole branch was run before opening this. Its findings are fixed in the last commits, except one that is kept by design (see below).

Mixer panel

• One console. The sections are Buses, Inputs, Outputs, Mapped and Virtual. Each folds away behind a vertical tab on its side. The master /out/main is pinned on the right. Every row lines up across strips (title, channel count, meter and fader, readout, controls).
• Meters. score::LevelMeter is a multichannel meter with peak hold, an RMS line and clip marks. It keeps the widest channel count seen, so a width change does not make it jump.
• Bus strips:
  • mute, solo and propagate;
  • balance and output routing;
  • an upmix menu (repeat the channels, or add silent ones, up to N);
  • per-channel gains for a wide bus;
  • a channel count and CPU readout;
  • selection linked with the score.
• Port strips: hardware inputs and outputs, mapped ports (metered on their hardware channels) and virtual ports. The "+" in the Mapped and Virtual tabs adds a port, and a strip's menu edits or removes it. Input meters show the level after the gain it is read with.
• Widths. A strip's width is set from its menu, and every strip's from a tab's menu. The toolbar volume follows the master.

Model and commands

• Mute and solo on intervals are undoable and saved. Solo keeps the soloed bus's ancestors and descendants heard. Both act through the existing interval mute, which disables everything in the interval.
• Upmix mode and channel count on audio outlets, with commands and serialization.
• Port gains. The gain of a port of the audio device (Audio::SetPortGain; score-plugin-audio now has commands) is one undo step when a fader or the toolbar volume is released. It is saved as the node's audio-gain attribute and restored on load and on reconnection.
• Gain, pan and upmix reach the running outlet of every process, not only of intervals (SetupContext::register_mixing_impl). The model offered them on every audio outlet.
• Older binary data. Binary data written before mute, solo and upmix existed still reads: DataStreamWriter::atDelimiterOrEnd makes added fields optional.

Execution → interface

• Execution::Telemetry (score-plugin-engine), with Process::TelemetryInterface in score-lib-process, reads libossia's telemetry arena at the document's coarse rate, under the execution-update setting. Meters can be asked for on outlets, inlets, hardware inputs and outputs, and virtual ports. The benchmark goes through bench taps, which lifts the old per-50-ticks hash-map copy.
• Moved onto it or onto triple buffers:
  • interval playheads (only start and stop remain commands, since executing() has model effects);
  • avendish control values;
  • Faust control and display values;
  • YSFX sliders;
  • the pattern sequencer's step;
  • LV2 control outputs.
• LV2. The LV2 interface read voice 0's buffers on the UI thread, every tick, while the audio thread wrote them. Its MIDI for the plug-in window is now copied at the end of the tick, into a bounded queue, only while the window is open.
• Gating. avendish and JS messages to the interface follow the execution-update setting.

Real-time safety and lifetimes

• No stopping the audio. Nothing stops the audio to make a change safe. Changes to what the callback reads run between two ticks (audio_engine::run_between_ticks), or are prepared and swapped in. The audio device edits, adds and removes ports this way while playing, in place, because running processes hold pointers to them.
• Removing a node while playing used to crash as soon as a process had written to it.
  • Devices announce a subtree before removing it (DeviceInterface::nodeAboutToBeRemoved).
  • The meters under it let go first.
  • When a process port may point there, the ports drop those addresses between two ticks, once for the whole removal.
  • The audio device is followed across its reconnections.
• Showing the audio device in the device explorer reconnected it, which stopped the audio during playback. That happened when adding a port or recording a gain from the mixer. A reconnection with the same engine and channels now leaves its tree alone.

Also

• Master gain. It is applied once the tick is summed, and a run starts at it.
• AudioDevice::portsChanged is emitted when a port is added, edited or removed.
• Dummy backend. It has the channels the settings ask for.
• Sliders. A value typed in a slider's popup ends the gesture.
• VU Meter. It reads its inlet through the telemetry, and its Levels output stops allocating once warm.

Worth a second opinion

• Solo as a mute. Solo silences the other buses through the interval mute, as decided, so their MIDI, OSC, video and automation stop too.
• AudioDevice::reconnect now returns early when the tree was built for the same engine (held as a weak_ptr) and channel counts. updateSettings disconnects first, so a settings change still rebuilds, and an engine restart is another engine.
• Removal filter. Deciding whether a removal needs the between-ticks pass uses the model ports' addresses: same device, and either a path prefix or a pattern. This is conservative on patterns.
• Plug-in feedback gating (avendish, JS) is decided when the execution is set up, so toggling the setting mid-run applies from the next play.
• Pattern destinations. Addresses written as patterns are still resolved against a tree the interface may be deleting. This predates the branch.

Tests

New or extended:

• test_unit_telemetry_exec: meters, a virtual port removed while playing, 200 ports removed while playing (once, under a second, with and without a device rebuild), and gain, pan and upmix propagation;
• test_unit_audio_port_edit: live edits, and a gain that is undone, redone and survives a JSON reload;
• test_integration_interval_mute_solo;
• test_unit_avnd_control_outputs_ui;
• test_unit_faust_dsp: generators with and without a cable;
• test_unit_lv2_loading: control outputs through the triple buffer;
• test_unit_datastream_compat.

Known red before any of this, unchanged: js_remove_cable (the test profile has no default package).

🤖 Generated with Claude Code

https://claude.ai/code/session_01HKbSZft9tvQ2sm9KeSchMX

[jcelerier]

The AppImage aarch64 failure on 63e2c52 comes from score-addon-librediffusion: it vendored ossia::triple_buffer in namespace ossia, and the avendish binding now includes libossia's own header. Fixed there in ossia/score-addon-librediffusion#24 (backward compatible, can be merged first); CI clones that add-on's default branch, so this job needs #24 merged.

[coveralls]

Coverage is 48.723% — feature/mixer into master. No base build found for master.

[edumeneses]

This is Edu's Claude assistant, posting on Edu's behalf.

Rebased onto master (bf4c7f5ee0), together with its submodule PRs, so that it merges again. Same 38 commits, same messages and authorship. Every submodule bump points at the rebased counterpart of the libossia commit it pointed at before, so each step still builds against its own libossia.

Submodules

• Mixer: gain stages, telemetry arena, port changes between ticks libossia#950 rebased onto 15a6c68 (the libossia that score master points to). One conflict, in dummy_protocol.hpp: the PR's channel buffers are kept, and the run thread is master's ossia::thread.
• ossia: hand the control values to the host through triple buffers celtera/avendish#210 cherry-picks cleanly onto avendish main (ec179e3, score master's pointer), but nobody here can push to celtera/avendish, so the rebased commit is in the draft ossia: hand the control values to the host through triple buffers (#210 on main) celtera/avendish#218. Score points at it (e248fec), which refs/pull/218/head makes fetchable. Merging either use plain keys (no modifiers) for interpolation/snapshots/etc... #210 (once rebased) or Allow to open .scorejson file with i-score directly #218 works.

Where master and the PR overlapped

• Toolbar volume (audio: the toolbar volume shows the master gain): master's showVolume() (6bcc0ff) did what syncVolume() does. It is gone, and its two call sites (after a rebind, on engine start) call syncVolume(). That keeps m_shownVolume, the "before" of the undoable gain, in step.
• Saved master volume (AudioDevice): master restores the saved value of a node, and carries the live values over a reconnect. The PR restores its audio-gain attribute and swaps the ports between ticks. Both are kept: setupNode(), then the saved value, then applyGain(); on reconnect the carried-over live values come last.
• Device lifetimes (execution: let go of removed nodes…, against master's 69469be and 6890fcc): master's onDeviceChanged() is the one handler. That covers bool register/unregister, triggers and conditions rebuilt, and the wait on the execution queue. It now also does the PR's m_telemetry->forgetUnder(old), and the PR's on_deviceChanged() is gone. The connections live in followDevice(), which also connects nodeAboutToBeRemoved (direct). The audio device is followed from the constructor and skipped in on_deviceAdded(), as in the PR. DeviceInterface has both deviceClearing and nodeAboutToBeRemoved, emitted in that order by disconnect().
• RegisteredPorts: master's several connections per port plus the PR's mixing map.
• DocumentPlugin.cpp: bench_state.hpp replaces bench_map.hpp; master's algorithms.hpp stays.
• LV2 (lv2: stop reading the running plug-in's buffers…): master's close-window drain goes through plugin_events->queue, and master's new LV2 test hooks get the on_finished(node) overload.
• ExecutionEditCommandsTest (master, 20474a0) stopped a graph interval that was never running. Since commands now go out only when the running state changes, that queued nothing; the test ticks the interval once before stopping it. Folded into execution: the playheads of the intervals go through the telemetry.
• Tests CMakeLists.txt and FaustDspTest.cpp: both sides' additions kept.

Checked

• Full Debug build with tests (clang 19, Linux).
• Tests: the whole suite, then every failure rerun at -j2 on this branch and on a master build of the same configuration. No test fails here that passes on master. What fails on both is GPU, media or PipeWire tests on a headless box, plus Avnd_regex (3/11) and one scriptable_dynamic_ports case, already failing on master.
• One thing for the PR itself, not the rebase: TelemetryExecTest "updates stop with the execution-update setting" turns ExecutionUpdate off and only turns it back on after its checks pass. ctest keeps settings in build/test-home/config across tests and runs, so if a check fails under load the setting stays off. Every later test that needs execution updates then fails: playheads, telemetry, avnd control outputs, record_while_playing, device_enumeration_lifecycle. A guard that restores it on scope exit would avoid that.

The pre-rebase head was e0dffba0ea, and libossia#950's was f9e87ad5e0.

Since this rebase, master moved avendish to main (a1f6471, 4603c1c), so GitHub shows the avendish pointer as conflicting again. Nothing else conflicts, and e248fec merges into a1f6471 cleanly; rebasing #218 onto it and bumping the pointer here clears it.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Lc7tYGHRn6TmaxMEfBTswW

[edumeneses]

This is Edu's Claude assistant, posting on Edu's behalf.

Rebased once more, onto master 4b2f7f45ef. Master had moved avendish to main (a1f6471, 4603c1c), which was the only conflict. The 38 commits are unchanged except for the avendish bump in avnd: read the control values from the node's triple buffers. It now points at 4c2b41cd, which is avendish#210's commit on top of a1f6471 (cherry-picked cleanly, in celtera/avendish#218 again). libossia is unchanged (7e4f52c, ossia/libossia#950).

Full Debug build passes. I reran the telemetry, mixer, avnd, playhead, LV2 and audio tests. Only Avnd_regex, Avnd_point_tracker and Avnd_utilities fail, and they fail on master too.

[edumeneses]

Hi JM, I pushed 08e951e to this branch. It goes with ossia/libossia#950 51236ae45: Edu wanted the /out/N and /in/N faders to act as channel faders for everything their channel gets, rather than only for what is sent straight to that address. The details are in this comment on #950.

Here, the commit:

• scales the input meters per channel with read_gain;
• updates the help text of the Inputs/Outputs tabs;
• bumps the submodule.

The audio, port edit, telemetry, routing and outlet tests pass. If you want it split from #2302, tell us.

— Edu's Claude assistant