Files

6.9 KiB
Raw Permalink Blame History

Analog sound routing and Quadrapassel regression

Technical evidence extract for the VCC Plus white paper. Only reviewed contracts, mechanisms, methods, results and limitations are retained. Original validation dates and configuration limits remain historical; this export is not a new test run.

Cause and correction

The native sampler mixed cartridge output into the speaker regardless of the PIA sound-enable control and selected analog input. An idle Speech/Sound output has unsigned midpoint 128. The old path emitted that as a large DC signal; joystick polling switched the mux to the cassette branch and back, chopping it into an audible waveform even though the game had disabled analog sound.

The sampler now follows the hardware routing:

Analog selector Source while sound is enabled
0 Six-bit CoCo DAC
1 Cassette input
2 Cartridge SND
3 Ground / silence

The sound-enable control gates that analog path. Single-bit sound and the modeled Orchestra-90 external RCA outputs remain independent. Cartridge PCM is centered on 128; a zero sample remains a valid negative peak, distinct from an absent/inactive sound device. The current DAC register is read even if the guest last wrote it during a muted joystick measurement. Final float mixing saturates rather than wrapping a signed packed word. Cartridge clocks and cassette advancement continue while muted or unselected.

The native sampler uses the new read-only PIA sound snapshot and separate cartridge speaker/external outputs. The original Windows packed mixer remains unchanged. Existing DAC, single-bit and Orchestra-90 gain conventions are retained.

Hardware cross-checks: MAME's CoCo implementation describes the selector map with its Tandy service-manual citation; its CoCo machine wiring connects PIA selection, sound-enable inhibit and the separate single-bit path. No MAME implementation was imported.

Unchanged real game reproduction

The optional tests/macos/audio-game-validation.cpp boots a supplied DOS disk through ordinary keyboard input, with local CoCo 3/Disk BASIC, Speech/Sound and Orchestra-90 ROMs. The source disk is read-only host input; guest writes affect only an owned memory copy. No text, registers or guest RAM are injected to create the observed results. The matrix uses 512 KiB, 1x overclock and both emulated CPUs, with a mounted stopped cassette and FD-502 in startup slot 4.

The real-game matrix ran on arm64. The test used Quadrapassel QUADPSL01.DSK, SHA-256 4ac171e87c51dfd99b302b9e035a91a59c030144e5ae0661a6d2dfc84ba3a63f. Its repository remained clean and its source disk hash stayed unchanged. After 1,800 settling fields, each idle capture spans 120 fields:

Configuration Before Corrected
FD-502 only, either CPU Exact digital silence Exact digital silence
FD-502 + Orchestra-90, either CPU Exact digital silence Exact digital silence
FD-502 + Speech/Sound, 6809 84,827 / 88,651 nonzero frames; 1,409 transitions; AC RMS 0.101581 Every sample zero
FD-502 + both cartridges, 6809 Same repeating erroneous signal Every sample zero
FD-502 + Speech/Sound, 6309 84,818 / 88,651 nonzero frames; 1,421 transitions; AC RMS 0.101695 Every sample zero
FD-502 + both cartridges, 6309 Same repeating erroneous signal Every sample zero

With both cartridges present, normal Enter and D/right keyboard input started the game and produced movement effects on both CPUs. Peak was 0.242188 and the final half-second was exactly silent. Thus the correction does not globally mute the game. Raw PCM, framebuffer captures and logs remain private and outside Git.

Example reproduction after building the current core through Xcode:

DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer \
  xcrun clang++ -std=c++17 -O2 -mmacosx-version-min=14.0 \
  tests/macos/audio-game-validation.cpp /path/to/libVCCCore.a \
  -o /tmp/audio-game-validation
/tmp/audio-game-validation /path/to/ROMs /path/to/game.DSK \
  /tmp/audio-game-captures --expect-quiet

The quiet-title and movement assertions apply to a DOS-boot game with the same Enter-to-start and D-to-move controls. It is an explicit optional compatibility check, not a general assertion that every game's title screen should be silent.

Regression qualification

  • All 67 portable core checks passed, including the new authored routing regression on both guest CPUs. It covers all mux states, the sound gate, joystick-style selector changes, current DAC latch, unsigned midpoint and zero cartridge samples, independent one-bit/Orchestra stereo, and saturation.

  • Seventeen focused XCTest checks passed on arm64 and again on x86_64 under Rosetta: audio buffer/scheduling/routing, cassette images and monitor, cartridge behavior, real output-device pause/restart, and volume changes retaining clocks.

  • Real-ROM speech and AY tone checks passed on both CPUs: 22 phonemes, 443,259 PCM frames, speech range 0.504 and tone range 0.332. Authored speech fixtures now explicitly select and enable cartridge audio. Test WAV writers retain actual signed PCM rather than subtracting another midpoint. Both speech WAV hashes remain identical to the previously qualified greeting capture recorded in the Speech/Sound guide.

  • The guest Speech/Sound SDK checks passed on both CPUs with missing/absent and real firmware, phrases, joystick polling, tones and noise. Orchestra-90 SDK checks retained exact channel plateaus and approximately 3,996 frames/s at normal speed and 3,986–3,988 at fast speed on both CPUs.

  • The isolated native playback test passed with both real cartridges, the unchanged game, and 25% volume. Across three seconds, emulated fields advanced from 2,336 to 2,515 and the real output device consumed another 132,182 sample frames. Nonzero-signal count remained zero and audio-engine start count stayed one. No mute or engine restart manufactured silence.

The native test is opt-in through VCC_AUDIO_GAME_DISK; ordinary UI runs skip it without that supplied disk or private ROMs. --audio-game-listening ROM_DIRECTORY DOS_BOOT_DISK also selects the same temporary fixture and automatically pauses. This is native output-consumption and digital-signal evidence, not a new human listening confirmation, physical Intel test or macOS 14 runtime qualification. The user's physical-machine observation remains the independent comparison.

The native case also passed independently after reopening its single main window through Window → VCC UI Tests (45.061 seconds). The test now uses that public reopening path and raises its window before capturing, avoiding saved closed-window state and unrelated foreground-window captures. The final window capture was visually checked against the game's splash screen and both sound cartridges. The universal normal Debug app built successfully; no DMG or installation was performed.