Skip to content

TESTING.md

Latest commit

 

History

History
47 lines (39 loc) · 2.48 KB

TESTING.md

File metadata and controls

47 lines (39 loc) · 2.48 KB
Raw

What is tested

The rules are in REALTIME_SAFETY.md. This is where things stand against them.

Check State
No allocation in the callback Tested. A trap on new / delete around a graph with every module, guarded feedback, a live neural model and incoming MIDI notes.
30 minutes of audio through that graph Passes offline (SIGNALPATCH_SOAK_BLOCKS=1687500). A soak on a live device with xrun counting has not been automated.
ASan + UBSan Clean over the whole suite (build-asan/).
TSan Clean over the suite (build-tsan/). The app itself has not been run under TSan with a real device.
Every module renders finite output Tested with changing block sizes. NaN injection is only tested for the delay and the Feedback Guard.
Rapid graph edits while rendering Tested (add, connect, compile and remove in a loop with parameter bursts).
Callback timing Measured in the app: average and worst block in the header.
Linux and Windows CI builds the app and runs the suite on both, from a bare clone.

The suite has 53 cases. Every example patch is loaded, compiled and rendered. Rendering on several cores is checked to be bit-identical to one core, allocation-free on the helpers, and clean under TSan. Beyond the above they cover save and load, device changes with missing channels, undo, the looper and 4-track at odd speeds, recordings across sample rates, the clock, MIDI, tuning of the pluck and synth, the TONE3000 client, and the inspector's ranges and curves.

Not verified

  • Windows with real audio hardware. It has run on one ROG Ally; nothing was measured there.
  • The gamepad code on a real pad.
  • Touch mode on a real touch screen.
  • The Flatpak manifest since the app moved to GLFW.
  • Heavy WaveNet captures at 64 samples on anything slower than a desktop Zen 3 core. SIGNALPATCH_BENCH_NAM_DIR measures a folder of models.

Audits

Twice so far the code has been read end to end for real-time safety, DSP errors, thread lifetime, persistence, device handling and undo. The second pass (September 2026, over the code added since v0.2.0) turned up data loss in recordings, a crash on quit, several MIDI bugs and a looper undo that could stall the callback. The fixes are listed under v0.3.1 in CHANGELOG.md.

One lesson worth keeping: the Windows CI found a use-after-realloc that Linux and ASan never showed, because MSVC grows vectors by 1.5x and libstdc++ doubles them. Do not hold pointers into PatchDocument::getNodes() across a mutation.