pymusicassembler

A standalone, pure-Python reader, writer, and player for Music Assembler C64 SID songs (the Richard Bayliss / Eric Campbell era player). It parses a Music Assembler tune into a typed song model, runs the playroutine to produce byte-exact per-frame SID register output, renders WAV through an emulated SID, and writes songs back to a byte-identical tune image.

Read/write/play/register-log is pure stdlib; only WAV rendering needs the optional audio extra (pyresidfp).

pip install pymusicassembler          # reader/writer/player/reglog
pip install pymusicassembler[audio]   # + WAV rendering via pyresidfp

Quick start

import pymusicassembler as ma

song = ma.read("tune.sid")            # PSID/.sid or bare .prg

# Per-frame SID register writes (changed registers only, after frame 0).
for writes in ma.iter_frames(song, max_frames=50 * 60):
    ...                               # writes: list[(register, value)]

# Forward-filled 25-register-per-frame snapshot grid (the oracle form).
grid = ma.render_grid(song, nframes=400)

# Register log (clock reg val triples), and WAV via an emulated SID.
ma.write_reglog(ma.iter_register_writes(song, max_frames=2500), "tune.reglog")
ma.render_wav(song, "tune.wav", seconds=30)   # needs the audio extra

# Edit and re-emit, byte-stable.
song.patterns[0].data[0] = ma.constants.DUR_MIN  # tweak a pattern byte
ma.write(song, "tune-edited.sid")

CLI:

pymusicassembler info   tune.sid
pymusicassembler reglog tune.sid tune.reglog --seconds 30
pymusicassembler wav    tune.sid tune.wav    --seconds 30 --model 8580
pymusicassembler resave tune.sid out.sid     # byte-identical round-trip

Public API

• read(src) -> Song / parse(bytes) -> Song — read a .sid/.prg.
• write(song, dst, container="auto"|"psid"|"prg") — write a tune image; auto re-emits the original container byte-identically.
• build(song) -> bytes — the raw C64 image; validate(song).
• Player(song) — .play_frame() -> list[(reg, val)], .regs.
• iter_frames(song, max_frames) — per-frame writes.
• render_grid(song, nframes) -> list[list[int]] — forward-filled grid.
• render_samples / render_wav / write_wav — audio (pyresidfp extra).
• iter_register_writes / read_reglog / write_reglog / RegWrite.
• Model: Song, Orderlist, OrderEntry, Pattern, Instrument, and the decoded pattern events NoteEvent / InstrumentEvent / DurationEvent (Pattern.events()).
• Errors: MusicAssemblerError, SidParseError, SongValidationError.

The Music Assembler format

Three independent per-voice tracks. The player binary is identical across tunes; only its DATA and the addresses that data lives at differ, so the reader discovers each table base from the player code's 16-bit operands (relocation-safe — no fixed-address assumption):

• Orderlists — one per voice, 2-byte entries (pattern_id, transpose << 4 | repeat), terminated by $FF (loop to start); pattern id $FE is a silence marker.
• Patterns — player bytecode: < $60 note (+ a control byte, with two inline argument bytes for a filter ($80 bit) or vibrato/arp ($20 bit) command); $60–$7F bare duration; $80–$9F instrument select (id = value & $1F); $A0–$FE long duration; $FF terminator.
• Instruments — 8-byte records: attack/decay, sustain/release, waveform, pulse-width init + sweep step, arp speed + step, porta/effect flags.
• Per-frame DSP — tempo divider, 16-bit pulse-width sweep, vibrato/arp frequency accumulators, the gate/retrigger idiom, and a self-modified filter-cutoff sweep.

Notes vs the decompiled reference (decompile.c)

The player was recovered from its disassembly and validated byte-exact against the preframr-sidtrace oracle. Two facts a naive read of the Ghidra decompile (decompile.c) gets wrong, recovered from the raw disassembly:

• The bare-duration token ($60–$7F) DOES advance the pattern cursor (the $10c3 JMP $1187); the decompile's inlining made it look like an early return.
• Init does not clear most work RAM. The init routine zeroes only $1081–$1090 (per-voice cursors / pattern id) and reloads pattern id / transpose / repeat; every other per-voice work byte AND the self-modified filter envelope ($129e accumulator, $1296 gate, $12a0 step, the $1265/$126a voice-0 trigger-reset immediates) keep their loaded-image values. A tune saved mid-playback bakes live state in, so the player seeds from those image bytes — otherwise its first frames diverge.

Byte-exact verdict

Validated against the preframr-sidtrace register oracle (PAL, 19656 cycles/frame, forward-filled, one leading silent play-call aligned over):

Round-trip: build(read(image)) == image and the full .sid container re-emits byte-identically for every tune; read(write(song)) == song.

Tests

Test tunes are HVSC copyright works and are never committed. They are fetched on demand (python scripts/fetch_tunes.py) into a gitignored cache; the byte-exact tests validate against the live preframr-sidtrace binary when $SIDTRACE_BIN is set, else against the committed frozen oracle grids in tests/fixtures/, so CI passes with neither the binary nor the tunes.

pip install -e ".[dev]"
./run_tests.sh
SIDTRACE_BIN=/path/to/sidtrace ./run_tests.sh   # validate live, too

License

Apache 2.0 — see LICENSE.