Epok Pulse — SPU music driver implementation and verification

Authoring

SoundFont-derived MusicSequences default to Music driver → Epok Pulse in PSX music conversion. Existing recipes without a driver field select this path when cooked again. Keep the existing SoundBank reference and AudioComponent or Blueprint play call. Source snapshots and UUIDs are preserved. Choosing Software reference restores the previous runtime synthesis implementation.

Execution model

The editor resolves MIDI ordering, pedals, note lifetimes, instrument regions, layer admission, pitch/gain modulation and PSX ADSR rates. EPSQ v3 contains absolute microsecond timestamps, atomic note groups, starts, release/cut, pitch/volume/send changes and loop boundaries. Identical tone templates are deduplicated. EPSB v2 remains the sample-bank format.

The IRQ dispatches commands to the SPU and shares the 24 hardware voices with SFX. It performs neither SoundFont envelope/LFO evaluation nor MIDI note-off matching. Hardware ADSR and ADPCM playback run in the SPU. Native-only games omit the reference synthesizer's working states and MIDI FIFO. Legacy/mixed games retain the compatibility service. All paths retain bounded priorities and whole-layer stealing, deferred key-on, source activation, bank pins and reverb ownership.

Epok Pulse is an original driver built around editor-compiled commands and hardware SPU envelopes. Register/timer behavior was checked against PSX SPU documentation and timer documentation.

Timer 0 now uses system clock with target 64000 (about 529 Hz). Its old DIV8 flag did not divide Timer 0, so the old configuration was approximately 2000 Hz, not 250 Hz. Target 65535 was rejected by emulator testing: target/overflow coincidence did not deliver audio callbacks in the installed Redux core. Timer 2 still measures elapsed CPU/8 time and Timer 1 supplies the wrap guard.

Measured result, 2026-09-17

Installed PCSX-Redux, interpreter, 2 MiB RAM; an isolated copy of Denise's arena and hm-bonneforces-exe. Seven emulated seconds measured after warm-up. Same scene/camera/song/48,000-byte sample budget in all cases; profiling HUD and motion interpolation disabled in all three copies to let the reference driver pass the minimum heap/stack guard. The benchmark did not modify the source demo.

This is about 77.5% less measured service CPU, not a zero-cost audio claim. Playing the native song adds about 2.06 percentage points over its idle service. Counters exclude BIOS interrupt-dispatch overhead; all three cases use the same corrected timer. No clock fault or runtime diagnostic occurred. These numbers are emulator evidence for this workload, not a console-wide or hardware guarantee.

With the original HUD/interpolation settings, a separate native/silent run gave 33.50/34.36 FPS. On MIPS, native instance/physical working arrays occupy 2,320 + 960 bytes; this excludes bank/sequence data, audio infrastructure and stack. The tested native sequence is 120,824 bytes plus a 55,552-byte bank (47,936 sample bytes uploaded to SPU RAM), without prepared software note-state tables.

Reproduce with python tests/integration/verify_native_music.py PATH_TO_DENISE_PROJECT after building the editor. It copies inputs to a new owned .epok/ directory and starts only its own headless emulator on an unused port; it does not change the input project or terminate an existing emulator.

Fidelity and bounds

• Hardware ADSR is quantized. Source delay is omitted and hold is folded into decay; conversion and Target Preview report affected layers. Source Preview remains the reference for the authored SoundFont behavior.
• Controllers that dynamically retime an active volume envelope fail native conversion with a Software reference diagnostic. Pan, gain, pitch bend, tuning, tempo and sustain/sostenuto semantics are resolved offline.
• Pitch/gain automation is sampled at 250 Hz plus MIDI event boundaries and only changed register values are emitted. Excessive data fails the 256 KiB / 32,768 command budget; an authored pass is limited to ten minutes. Loops can repeat.
• The actual song reaches the 16-physical-voice ceiling including release tails; 2,613 offline voice steals are reported for the whole song. Its analyzed 11-note MIDI peak is not a measurement of physical voices including tails. The new driver does not silently shorten the source's release setting.
• Target Preview decodes cooked ADPCM and the compiled commands with a hardware ADSR model. It is not a bit-exact SPU emulator: linear interpolation replaces Gaussian interpolation; wet reverb, hardware key-on latency and SFX contention are not simulated.
• Real console validation and a subjective full-song listening review remain.

Host regression coverage includes conversion/audition, pedal release timing across tempo changes, bend/pan automation, repeated loops, malformed/truncated input, cancellation, ADSR fitting/chunk invariance, native/mixed playback, atomic layers, SFX stealing, source disable, gain, bank retirement, legacy sequence playback and the existing instrument allocator/bank tests.