Game Boy sound hardware (APU)

The game-boy's sound unit, the APU, has four channels that are each built differently — "unlike some other sound chips, such as the C64's SID or the Atari 5200's POKEY, each sound channel is specialized" — plus a VIN input for cartridge audio that no licensed game used. Every channel produces a 4-bit value. The register-level facts below come from pan-docs, the musician-facing names from lsdj (sources: s-pandocs-audio-overview, s-pandocs-audio-registers, s-pandocs-audio-details, s-lsdj-manual).

The four channels#

In LSDj terms: PU1 and PU2 carry leads and chords, and the sweep makes PU1 the pulse bass-drum channel; WAV runs the soft synth, the sample kits and speech (volume 0/25/50/100 %); NOI plays 15-bit noise or 7-bit "pitched" noise (source: s-lsdj-manual).

Register map#

NRxy: x = channel (5 = global), y = register. Rule of thumb: NRx1 length timer and duty, NRx2 volume and envelope, NRx3 period low byte, NRx4 trigger, length enable and period high bits.

(source: s-pandocs-audio-registers; PCM registers: s-pandocs-audio-details)

Timing and formulas#

• One DIV-APU tick is 1/512 s: the envelopes run at 64 Hz, the length timers at 256 Hz, the CH1 sweep at 128 Hz. APU timing does not change in CGB double-speed mode; on the Super Game Boy 1 the whole APU runs about 2.4 % fast (sharp), corrected on the SGB2 (sources: s-pandocs-audio-details, s-pandocs-audio-overview).
• Pitch is a period value v (0–2047) that works as a negative period: pulse frequency = 131072 ÷ (2048 − v) Hz, wave = 65536 ÷ (2048 − v) Hz — at the same value the wave channel is one octave lower. Pan Docs' examples: $500 → 170.67 Hz pulse / 85.33 Hz wave, $740 → 682.67 / 341.33 Hz. Hence the lowest pulse pitch is 64 Hz and the lowest wave pitch 32 Hz (computed from the formulas), and pitch resolution is fine at low values and coarse near the top (pitch-slide-and-portamento). Both ends are visible in lsdj's own pitch table: its last entries repeat — … F9 07 F9 07 F9 07 FA 07 FA 07 FA 07 FB 07 …, several semitones sharing one period value — and a spreadsheet asking for 55 Hz gets the register value −335 (sources: s-little-scale-lsdj-tuning-tables, s-little-scale-koto-tuning; see rom-retuning). Period writes take effect only after the current sample (source: s-pandocs-audio-registers).
• The note grid is software. The chip has no notion of a semitone: lsdj and nanoloop each hold a table of precomputed period values in ROM and index it from the note column — 108 entries (216 bytes) in LSDj 3.8.9 at $6D3A, plus a second 216-byte table at $6E12 recomputed for the SGB1's fast clock, and 144 bytes at $730A in Nanoloop 1.3d. Overwriting it retunes the whole instrument (alternative-tunings, rom-retuning; sources: s-little-scale-lsdj-tuning-tables, s-little-scale-nanoloop-18-tet). The SGB table's first entry is 00 00: at a clock 2.4 % fast, the lowest note of the standard table is below what the register can express at all.
• The crystal is the tuning reference, and it is modifiable. Every formula above is written against the DMG's 4.194304 MHz clock, and that clock sets pitch as directly as it sets everything else: "the crystal oscillator is responsible for controlling all timing aspects of the Game Boy including the frequency pitches of the pulse oscillators. As a result, the underclocked Game Boy plays back just under an octave lower than its unmodified counterpart." Just under is the problem — the machine is then neither transposed nor in tune with anything else. little-scale's fix moves the software grid to match the hardware: a replacement lsdj pitch table computed for a round 4.000000 MHz (= 2 × 2 MHz), so a Game Boy underclocked to exactly half its nominal rate lands an exact octave down and in tune. Decoded against 4 MHz the table's first octave is 65.41, 69.29, 73.40 … 123.52 Hz (12-TET from C2, A = 110.04 Hz); at the machine's real 2 MHz those same bytes sound one octave lower, C2 = 32.71 Hz. Hardware pitch and ROM pitch table have to be changed together (rom-retuning, alternative-tunings; source: s-little-scale-underclocking-and-retuning).
• Length timer: counts at 256 Hz from the written value and switches the channel off at 64 (CH1, CH2, CH4) or 256 (CH3) — LSDj's LENGTH (source: s-pandocs-audio-overview).
• Noise: a 15-bit LFSR — bit 15 ← XNOR of bits 0 and 1, shift right, the bit shifted out selects 0 or the envelope volume; in 7-bit mode the new bit is also copied into bit 7. The LFSR is zeroed on trigger; if its active part is all ones it locks up silent, which happens when switching from 15- to 7-bit mode shortly after a trigger — cured by retriggering (source: s-pandocs-audio-details).

Signal path and pops#

Generator (digital 0–15) → DAC (linear to −1…+1, inverted: digital 0 = analog +1) → mixer (adds the channels chosen in NR51, ±4 per side) → NR50 scaling → volume knob → a high-pass filter per side. The HPF removes the DC offset of silent channels and off-centre waves; it is mildest on the DMG, stronger on the Color and strongest on the Advance (more waveform distortion). Every DC jump pops: switching a DAC on or off, adding or removing a channel in NR51, changing NR50. The clean way to silence a channel is $08 into NRx2 (volume 0 with the DAC alive) followed by a retrigger. The Game Boy Advance mixes digitally, has no DACs, inverts CH3's DAC (a loud spike when it is disabled), and its extra audio features are unavailable to Game Boy programs (source: s-pandocs-audio-details).

Quirks that reach the tracker#

Timing in LSDj is tick-based — at 125 BPM there are 50 ticks per second (source: s-lsdj-manual); the hardware's own clocks (64/128/256 Hz) run independently of it.

The four output levels cost resolution, not just loudness: the wave channel's volume "lower[s] the volume by shifting each sample and discarding the lower bits. So in effect, wave channel only has 3 and 2 bits of resolution for the lower volume settings" — nitro2k01, matching Pan Docs' description of NR32 as a right shift. The master volume (LSDj's M) does not have that problem but moves everything on that stereo side, which is why the workaround is to pan the wave channel alone to one side and ride M there: 8 steps instead of 4 (source: s-chipmusic-lsdj-wav-channel-release).

What 32 × 4 bits buys you#

Two numbers first. Sixteen amplitude levels against 65,536 for 16-bit audio is "part of what gives the Wave channel its characteristically crunchy sound" — the crunch is quantisation, not a filter (LSDj's default saw reads 8ECDCCBBAAA999888776665554433231; source: s-intense-tech-01-wave-synth-deep-dive-part-1). And the space those 32 nibbles span is 16^32 = 2^128 ≈ 3.4 × 10^38 waveforms: little-scale built a Max patch in 2009 to play all of them at 100 per second and 440 Hz, and worked out that it would finish in "1,079,028,307,100,000,000,000,000,000 centuries". The practical reading is not that the channel is limitless but that searching is not a method — the useful waveforms are a vanishingly thin, musically chosen slice, which is why hand-drawing recipes are worth writing down at all (wavetable-programming; source: s-little-scale-all-4-bit-waveforms).

The wave channel's one cycle of 32 samples is also a budget, and two techniques spend it in opposite directions. Drawing a second waveform at a whole-number multiple of the fundamental inside the same 32 samples gives a second, independently playable note — the ghost-channel — but only at harmonic intervals, "so as to fit in the same amount of space in the wave sample", and at the price of amplitude: two waves summed into 16 levels leaves neither loud (source: s-hackaday-gameboy-channelhacking).

The 32-sample cycle is exactly why the intervals are harmonic. The second waveform has to complete a whole number of cycles inside the same 32 samples, so overtone index x means a rider period of 32/x samples and a step every 16/x — for the octave-plus-fifth (x = 3) that is 5⅓ samples, which is not a whole number, so the drawn pattern comes out uneven and the interval aliases. nitro2k01's verdict from the 2019 thread: "In practice, perhaps only index 3 (octave + fifth) is really useful, unless you want to get waves that sound 'bad'" — the higher overtones are also "perfect intervals and slightly detuned from our modern equal temperament tuning" (harmonic-series, alternative-tunings). The amplitude cost is stated as a rule on the same thread: "you must create more headroom by sacrificing some of your sine wave's amplitude", and bryface's way round it is to bring every other instrument down instead (levels-and-headroom). The technique itself is on ghost-channel (source: s-chipmusic-lsdj-ghost-channel).

Escaping that limit by rewriting wave RAM under a playing note does not work at musician level. A 2019 attempt to pre-compute summed wave buffers and swap them at the end of each cycle was abandoned: "the WAV channel is buggy as hell, and produces large clicks when swapping out waves … at the frequencies we'd be writing to the wave sample buffer, it will sound like a constant, out of tune note whose pitch varies based on how hard the CPU is working", the timing "must be incredibly precise, and there's not so much documentation", and it costs a lot of CPU (source: s-hackaday-gameboy-channelhacking) — the same rule the register table states as "wave RAM can only be rewritten while the channel is stopped". The pulse channels turned out to be the softer target: holding them at the lowest frequency and highest duty, then rewriting the amplitude from a summed sine table under interrupt, splits each into two software voices — six voices in total, at 7 amplitude levels instead of 15.

Related#

game-boy · lsdj · nanoloop · pan-docs · alternative-tunings · rom-retuning · pulse-width-modulation · wavetable-programming · ghost-channel · harmonic-series · adsr-envelope · retrigger · stereo-panning · instrument-design · sid · lsdpatch · chip-samples · little-scale · levels-and-headroom

Sources#

s-pandocs-audio-overview · s-pandocs-audio-registers · s-pandocs-audio-details · s-hackaday-gameboy-channelhacking · s-lsdj-manual · s-intense-tech-17-the-joys-of-noise · s-intense-tech-12-scoping-out-new-features · s-intense-tech-19-new-noise-and-910-news · s-intense-tech-20-lsdpatch · s-intense-tech-21-whats-new-in-lsdj-92 · s-little-scale-lsdj-tuning-tables · s-little-scale-koto-tuning · s-little-scale-nanoloop-18-tet · s-little-scale-underclocking-and-retuning · s-intense-tech-01-wave-synth-deep-dive-part-1 · s-little-scale-all-4-bit-waveforms · s-defensemech-every-lsdj-kick · s-little-scale-lsdj-kits · s-chipmusic-lsdj-ghost-channel · s-chipmusic-lsdj-wav-channel-release · s-defensemech-noise-channel-92l