Mixer: gain stages, telemetry arena, port changes between ticks

[jcelerier]

The libossia side of the ossia score mixer (ossia/score feature/mixer). 18 commits, each with its test where one applies.

Gain chain

• Interval gain once. A bus applied its interval's gain and pan twice (0.25 instead of 0.5, proven by IntervalGainTest). forward_node now forwards, and the outlet applies gain and pan once. has_gain is gone.
• Device gain stages. Each audio_parameter applies its gain at exactly one point: when the graph writes to it (outputs, virtual ports), when the graph reads it (inputs), or after the whole mix (/out/main). The gain is atomic.
• /out/main is the master. audio_protocol::apply_main_gain trims every hardware output once the tick is summed, including what was sent straight to /out/N. reset_main_gain_ramp makes a new run start at the master gain.
• /in/main scales every input read.
• Ramps. Gain and pan changes ramp over one buffer in audio_outlet::post_process. pan is never grown on the audio thread.
• Upmix per outlet: off, repeat the channels up to N, or add silent channels up to N.

Telemetry: execution → interface

ossia/dataflow/telemetry.{hpp,cpp} adds an arena of meter, bench and playhead taps.

• The tick publishes one frame through a triple buffer at the interface's rate.
• Levels the interface has not read are folded into the next frame. triple_buffer::publish_if_unread (a CAS) makes sure a frame read meanwhile is not delivered twice.
• Frames are written in place and only copy the channels in use.
• RMS is computed per channel, in double precision, over the frames that channel was there for.

Taps sit on audio outlets (after gain and pan), audio inlets, the hardware inputs and outputs, virtual audio parameters, and interval playheads. Node timing goes to bench taps instead of the bench_map that was copied on the audio thread (perf_map is now bench).

triple_buffer gains write_buffer/publish and consume()/read_buffer for in-place use, in both the general and the trivially copyable case.

Real-time safety

• audio_engine::run_between_ticks(f) runs a short, allocation-free change at the start of the next callback and waits for it. No callback is skipped. When no callback comes, it runs f on the caller.
• Deferred port lists. With defer_port_changes(true), the callback walks its own copy of the mapped and virtual port lists, which swap_ports hands over between two ticks. virtual_audio_parameter::make_channels/swap_channels resize a virtual port the same way.
• Stale entries. The local-state policy no longer pushes to a parameter nothing wrote during the tick: its entries outlive their parameters.
• Faust nodes publish their control and display values (faust_ui_values) instead of the interface reading the DSP's zones. A mono generator makes one channel with nothing plugged in, and follows the channels of a cable plugged into it.
• The dummy engine can have channels.

Tests

New: TelemetryTest, TripleBufferTest, MainGainTest, IntervalGainTest, AudioEngineBetweenTicksTest. Also passing: OutletUnitTest, DataflowTest, SoundTest.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HKbSZft9tvQ2sm9KeSchMX

[edumeneses]

Hi JM, I pushed 51236ae to this branch, after Edu tested the mixer and decided how the hardware faders should behave.

What Edu saw: with sound sent to /out/main, the /out/1 fader in the mixer did nothing. With the gain stages here, /out/N only scales what is written straight to /out/N, before the sum. Anything reaching the channel through /out/main or through a mapped port skips that gain. The same holds for inputs: /in/3 didn't touch /in/main or a mapped port reading input 3.

What changed: /out/N and /in/N now act like the channel strips of a console.

• Outputs: /out/N is now gain_stage::external. apply_main_gain applies /out/main × /out/N to each channel once the tick is summed, with a ramp over the buffer per channel. reset_main_gain_ramp also resets the channel ramps.
• Inputs: audio_parameter::read_gain(chan) gives the gain each channel is read with: the parameter's own gain (pull stage) × /in/main × the channel's /in/N. /in/N itself isn't counted twice. clone_value uses it, and so do the score mixer's input meters.
  • hardware_inputs points to audio_ins. It is set on /in/main and on the mapped ports.
  • hardware_channel is identity for /in/main and follows mapping for an input mapped port.
• A mapped port's own gain still applies only to what goes through that port, like a send.

Tests: MainGainTest has 7 cases and 96 assertions. I changed the old /out/N expectation and added a channel-gain ramp case and an /in/N trim case. TelemetryTest passes, and so do score's audio, port edit, telemetry, routing and outlet tests.

The score side is ossia/score#2302 08e951e4e1: the per-channel input meter, the help text of the Inputs/Outputs tabs, and the submodule bump. If you'd rather have this as a separate PR on top of #950, tell us and I'll split it out.

— Edu's Claude assistant