diff --git a/doc/spec_SGU.md b/doc/spec_SGU.md index 10a09fb..16faa06 100644 --- a/doc/spec_SGU.md +++ b/doc/spec_SGU.md @@ -35,7 +35,8 @@ WavetableとPCMは、位相の整数部が指す標本 `a` と次の標本 `b` - Wavetable: 次の標本はインデックス`+1`（`0xFF`の次は`0x00`）。 - PCM: 次の標本はアドレス`+1`。ただし`+1`が PCM_END 以上になる場合は、ループ有効なら PCM_LOOP の標本、無効なら現在の標本を使う（一回再生の末尾は保持）。 - Phase変調は位相の整数部へ加算する。Ring変調は補間後の値に適用する。Slew Rate／Step は補間後のQ8.8値を入力とし、v2.1では状態の小数部を出力に残す。Hard Sync／Windowの標本置換は補間しない。 -- Noise、SQV4 は補間しない。 +- SQV4: デコード後の標本（フラグメント音量適用後）を同じ式で補間する。次の標本は次のフレームの先頭でもよい。最後の標本の次は、ループ有無にかかわらず最後の標本を保持する（ループ先頭へは補間しない）。 +- Noise は補間しない。 ## 3. インターフェース diff --git a/memory/2026-10-24-issue-158-sqv4-interpolation.md b/memory/2026-10-24-issue-158-sqv4-interpolation.md new file mode 100644 index 0000000..bd9498e --- /dev/null +++ b/memory/2026-10-24-issue-158-sqv4-interpolation.md @@ -0,0 +1,7 @@ +# SGU v2.1: SQV4 の線形補間 (#158) + +- SQV4 の補間は新しい版ではなく、v2.1（3WS8PNA）の補間機能の不足分の追加。`SguVersion` やチップ名は上げない。v1/v2 は従来と bit 一致（`frac` を無視）。 +- 次の標本がフレーム境界（63→64）を越えるため、デコード済みフレームを 2 枚（`frame & 1`）保持する。デコーダは順方向にしか進まないので、`ensure_sqv4_frame` は目的フレームより後ろへ戻るときだけ先頭から再デコードする。 +- 最後の標本の次の tap は保持（ループ先頭へは補間しない）。ループ先頭の標本を読むとデコーダ再開が毎 tick 起きるため。 +- SQV4 の音量は Q8.8 のまま `q8 * volume / (255*256)` で適用する（整数 q8 なら従来の `sample*volume/255` と一致）。 +- 検証: `cargo test --lib devices::sgu`（`sqv4_two_tap_interpolation_crosses_frames_and_holds_at_end`）。外部 golden に依存する SQV4 テストは無効のまま。 diff --git a/src/devices/sgu/s3w2.rs b/src/devices/sgu/s3w2.rs index f6b6459..b29565c 100644 --- a/src/devices/sgu/s3w2.rs +++ b/src/devices/sgu/s3w2.rs @@ -130,10 +130,12 @@ pub struct Channel { sqv4_sample_count: usize, sqv4_frame_size: usize, sqv4_loop_sample: usize, - sqv4_cached_frame: Option, + // Two 64-sample frames are kept (indexed by frame & 1) so a sample can be + // interpolated with the first sample of the following frame. + sqv4_cached_frame: [Option; 2], sqv4_next_frame: usize, sqv4_decoder: Sqv4State, - sqv4_frame: [i16; 64], + sqv4_frame: [[i16; 64]; 2], // Internal state pub old_phase: u64, @@ -177,10 +179,10 @@ impl Default for Channel { sqv4_sample_count: 0, sqv4_frame_size: 0, sqv4_loop_sample: 0, - sqv4_cached_frame: None, + sqv4_cached_frame: [None; 2], sqv4_next_frame: 0, sqv4_decoder: Sqv4State::default(), - sqv4_frame: [0; 64], + sqv4_frame: [[0; 64]; 2], old_phase: 0, lfsr_state: 0x12D4_803C, phase: 0.0, @@ -260,10 +262,10 @@ impl Channel { self.sqv4_sample_count = 0; self.sqv4_frame_size = 0; self.sqv4_loop_sample = 0; - self.sqv4_cached_frame = None; + self.sqv4_cached_frame = [None; 2]; self.sqv4_next_frame = 0; self.sqv4_decoder = Sqv4State::default(); - self.sqv4_frame.fill(0); + self.sqv4_frame = [[0; 64]; 2]; } } @@ -644,7 +646,7 @@ impl S3w2Sound { WaveformType::Pcm => self.generate_pcm_sample(ch), WaveformType::Noise => i32::from(self.generate_noise_sample(ch)) << 8, WaveformType::DmaPcm => i32::from(self.generate_dma_pcm_sample(ch)) << 8, - WaveformType::Sqv4 => i32::from(self.generate_sqv4_sample(ch)) << 8, + WaveformType::Sqv4 => self.generate_sqv4_sample(ch), WaveformType::FifoDac => { // 16-bit sample, fixed gain: VOLUME does not apply. let sample = self.generate_fifo_sample(ch); @@ -656,7 +658,7 @@ impl S3w2Sound { self.channels[ch].last_sample = (q8 >> 8) as i16; let volume = self.channels[ch].volume as i32; if waveform == WaveformType::Sqv4 { - ((q8 >> 8) * volume / 255) as i16 + (q8 * volume / (255 * 256)) as i16 } else { (q8 * volume / (4 * 256)) as i16 } @@ -866,23 +868,59 @@ impl S3w2Sound { channel.sqv4_sample_count = header.samples_per_channel; channel.sqv4_frame_size = header.frame_size; channel.sqv4_loop_sample = loop_sample; - channel.sqv4_cached_frame = None; + channel.sqv4_cached_frame = [None; 2]; channel.sqv4_next_frame = 0; channel.sqv4_decoder = Sqv4State::default(); true } - fn generate_sqv4_sample(&mut self, ch: usize) -> i16 { + /// Decode frames in order until `target` is cached (restarting from frame 0 + /// when it lies before the decoder position). + fn ensure_sqv4_frame(channel: &mut Channel, pcm_ram: &[u8], start: usize, target: usize) { + if channel.sqv4_cached_frame[target & 1] == Some(target) { + return; + } + if target < channel.sqv4_next_frame { + channel.sqv4_decoder = Sqv4State::default(); + channel.sqv4_next_frame = 0; + channel.sqv4_cached_frame = [None; 2]; + } + let frame_size = channel.sqv4_frame_size; + let variant = channel.sqv4_variant.unwrap(); + let version = channel.sqv4_version; + while channel.sqv4_next_frame <= target { + let decode_frame = channel.sqv4_next_frame; + let fragment_offset = start + 16 + decode_frame * frame_size; + let volume = pcm_ram[fragment_offset] as i32; + let payload = &pcm_ram[fragment_offset + 1..fragment_offset + frame_size]; + let count = (channel.sqv4_sample_count - decode_frame * 64).min(64); + for within in 0..count { + let code = sqv4::unpack_code(payload, within, variant.bits()); + let reconstructed = + sqv4::decode_code(version, variant, &mut channel.sqv4_decoder, code); + channel.sqv4_frame[decode_frame & 1][within] = (reconstructed * volume / 255) as i16; + } + channel.sqv4_cached_frame[decode_frame & 1] = Some(decode_frame); + channel.sqv4_next_frame += 1; + } + } + + /// Returns the decoded sample in Q8.8 (sample units << 8). v2.1 blends with + /// the next decoded sample; the tap after the last sample holds the last + /// one (no loop wrap, which would force a decoder restart on every tick). + fn generate_sqv4_sample(&mut self, ch: usize) -> i32 { if !self.channels[ch].sqv4_validated && !self.initialize_sqv4(ch) { return 0; } + let interpolate = self.version.has_3ws8pna(); let start = self.channels[ch].pcm_start_addr as usize; - let mut sample_index = (self.channels[ch].phase * 256.0) as usize; let sample_count = self.channels[ch].sqv4_sample_count; if sample_count == 0 { self.channels[ch].active = false; return 0; } + let (sample_index, frac) = Self::split_phase(self.channels[ch].phase); + let mut sample_index = sample_index as usize; if sample_index >= sample_count { if self.channels[ch].pcm_control & 0x02 == 0 { self.channels[ch].active = false; @@ -891,38 +929,20 @@ impl S3w2Sound { sample_index = self.channels[ch].sqv4_loop_sample; self.channels[ch].phase = sample_index as f64 / 256.0; } - let frame_index = sample_index / 64; - let within_frame = sample_index % 64; - let frame_size = self.channels[ch].sqv4_frame_size; - let variant = self.channels[ch].sqv4_variant.unwrap(); - let version = self.channels[ch].sqv4_version; let pcm_ram = self.pcm_ram.borrow(); - let channels = &mut self.channels; - let channel = &mut channels[ch]; - if channel.sqv4_cached_frame != Some(frame_index) { - if frame_index < channel.sqv4_next_frame { - channel.sqv4_decoder = Sqv4State::default(); - channel.sqv4_next_frame = 0; - } - while channel.sqv4_next_frame <= frame_index { - let decode_frame = channel.sqv4_next_frame; - let fragment_offset = start + 16 + decode_frame * frame_size; - let volume = pcm_ram[fragment_offset] as i32; - let payload = &pcm_ram[fragment_offset + 1..fragment_offset + frame_size]; - let count = (sample_count - decode_frame * 64).min(64); - for within in 0..count { - let code = sqv4::unpack_code(payload, within, variant.bits()); - let reconstructed = - sqv4::decode_code(version, variant, &mut channel.sqv4_decoder, code); - channel.sqv4_frame[within] = (reconstructed * volume / 255) as i16; - } - channel.sqv4_cached_frame = Some(decode_frame); - channel.sqv4_next_frame += 1; - } - } - let sample = channel.sqv4_frame[within_frame]; + let channel = &mut self.channels[ch]; + Self::ensure_sqv4_frame(channel, &pcm_ram, start, sample_index / 64); + let current = i32::from(channel.sqv4_frame[(sample_index / 64) & 1][sample_index % 64]); + let q8 = if interpolate && frac != 0 && sample_index + 1 < sample_count { + let next_index = sample_index + 1; + Self::ensure_sqv4_frame(channel, &pcm_ram, start, next_index / 64); + let next = i32::from(channel.sqv4_frame[(next_index / 64) & 1][next_index % 64]); + current * (256 - frac) + next * frac + } else { + current << 8 + }; channel.phase += (channel.frequency as f64) / (SOUND_CLOCK as f64) / 8.0; - sample + q8 } fn generate_noise_sample(&mut self, ch: usize) -> i16 { @@ -1130,6 +1150,48 @@ mod tests { assert_eq!(sound.generate_sample(0), 50 - 128); } + fn sqv4_sound(version: SguVersion, bytes: &[u8]) -> S3w2Sound { + let mut sound = + S3w2Sound::with_version(Rc::new(RefCell::new(vec![0; PCM_RAM_SIZE])), version); + for (offset, byte) in bytes.iter().copied().enumerate() { + sound.write_pcm_ram(256 + offset as u32, byte); + } + let end = 256 + bytes.len() as u32; + for (base, address) in [(0x810, 256u32), (0x813, end), (0x816, 256)] { + sound.write_register(base, (address >> 16) as u8); + sound.write_register(base + 1, (address >> 8) as u8); + sound.write_register(base + 2, address as u8); + } + sound.write_register(0x800, 0x0B); // 3000 Hz => 0.5 sample per call + sound.write_register(0x801, 0xB8); + sound.write_register(0x802, 4); + sound.write_register(0x803, 255); + sound.write_register(0x819, 0x01); + sound + } + + #[test] + fn sqv4_two_tap_interpolation_crosses_frames_and_holds_at_end() { + let pcm: Vec = (0..130).map(|i| i * 40).collect(); + let bytes = super::sqv4::encode_interleaved(&pcm, 1, 8_000, Sqv4Variant::H, 255).unwrap(); + let decoded = super::sqv4::decode(&bytes).unwrap().samples; + assert_eq!(decoded.len(), 130); + + let mut v2 = sqv4_sound(SguVersion::V2, &bytes); + let mut v21 = sqv4_sound(SguVersion::V2_1, &bytes); + for k in 0..129usize { + // Even calls land on a sample, odd calls halfway to the next one. + assert_eq!(v2.generate_sample(0), decoded[k], "v2 sample {k}"); + assert_eq!(v2.generate_sample(0), decoded[k], "v2 half {k}"); + assert_eq!(v21.generate_sample(0), decoded[k], "v2.1 sample {k}"); + let half = ((i32::from(decoded[k]) * 128 + i32::from(decoded[k + 1]) * 128) / 256) as i16; + assert_eq!(v21.generate_sample(0), half, "v2.1 half {k}"); + } + // The last sample has no following tap: it is held, not blended. + assert_eq!(v21.generate_sample(0), decoded[129]); + assert_eq!(v21.generate_sample(0), decoded[129]); + } + #[test] fn fifo_dac_is_16_bit_fixed_gain_with_decimation_and_pan() { let mut sound = v21_sound();