diff --git a/crates/audio/Cargo.toml b/crates/audio/Cargo.toml index ca4f5e509..92e38cfe5 100644 --- a/crates/audio/Cargo.toml +++ b/crates/audio/Cargo.toml @@ -41,6 +41,8 @@ aac-decode = ["craft-media/bitstream"] # The AAC encoder (`aac::Encoder`), and AAC in M4A through `encode` and `FileEncoder`. aac-encode = ["craft-media/bitstream", "craft-media/isobmff"] opus-decode = [] +# ADPCM (`adpcm`): IMA WAV, Microsoft, Yamaha and Flash (SWF) decoders and encoders. +adpcm = ["craft-media/bitstream"] # QOA (Quite OK Audio) files: decoder and encoder in the `qoa` module, each on its own. qoa = ["qoa-decode", "qoa-encode"] qoa-decode = [] @@ -111,6 +113,10 @@ required-features = ["decode", "encode"] name = "mp2_encoder_oracle" required-features = ["mpa-encode"] +[[test]] +name = "adpcm_oracle" +required-features = ["adpcm"] + [[test]] name = "flac_decode" required-features = ["flac-decode", "flac-encode"] diff --git a/crates/audio/README.md b/crates/audio/README.md index 166f3f817..cfe10d9d3 100644 --- a/crates/audio/README.md +++ b/crates/audio/README.md @@ -45,6 +45,7 @@ No default features. The shared types (`AudioBuffer`, `AudioInfo`, `ChannelLayou | `mp3-decode` | `mp3::Decoder`, `mp3::InfoTag` | `mpa-decode` | | `mp3-encode` | `mp3::Encoder`; `encode`, `FileEncoder` for MP3 | craft-media `bitstream` | | `mpa-encode` | `mpa::encoder::Encoder` (Layer II / MP2 frames) | `mpa-decode`, `mp3-encode` | +| `adpcm` | `adpcm::{ima, ms, yamaha, swf}`: decoders and encoders for IMA WAV, Microsoft, Yamaha and Flash ADPCM | craft-media `bitstream` | | `gsm-decode`, `gsm-encode` (`gsm`) | `gsm::Decoder`, `gsm::Encoder`, `gsm::decode`, `gsm::encode`: GSM 06.10 full-rate speech, bit-exact, raw `.gsm` and WAV (GSM 6.10) packings | none | | `vmd-decode` | `vmd::Decoder`: Sierra VMD sound (8-bit PCM and 16-bit DPCM), frame by frame from `craft_media::vmd` | none | | `peaks` | `peaks::Peaks` | none | diff --git a/crates/audio/src/adpcm/ima.rs b/crates/audio/src/adpcm/ima.rs new file mode 100644 index 000000000..dcf6958f7 --- /dev/null +++ b/crates/audio/src/adpcm/ima.rs @@ -0,0 +1,247 @@ +//! IMA ADPCM in WAV blocks (format tag 0x0011). +//! +//! A block is, per channel, a 4-byte header (the first sample as `i16` and the step index as `u8`, +//! then a zero byte), then groups of 4 bytes per channel in channel order, each holding eight +//! 4-bit codes, low nibble first. A block of `block_align` bytes holds +//! `(block_align - 4 * channels) * 2 / channels + 1` samples per channel. A code is a sign bit and +//! three magnitude bits: the decoder adds `step/8` plus `step`, `step/2` and `step/4` for each set +//! magnitude bit from the top, then moves the step index by `[-1, -1, -1, -1, 2, 4, 6, 8]`. + +use super::{Error, Result, clip16, step_at}; + +const INDEX_ADJUST: [i32; 8] = [-1, -1, -1, -1, 2, 4, 6, 8]; + +/// The state of one channel's decoder. +#[derive(Clone, Copy)] +struct State { + predictor: i32, + index: i32, +} + +impl State { + /// The sample after code `nibble`. + fn step(&mut self, nibble: u8) -> i16 { + let step = step_at(self.index); + let mut diff = step >> 3; + if nibble & 1 != 0 { + diff += step >> 2; + } + if nibble & 2 != 0 { + diff += step >> 1; + } + if nibble & 4 != 0 { + diff += step; + } + if nibble & 8 != 0 { + diff = -diff; + } + self.predictor = clip16(self.predictor + diff); + self.index = (self.index + INDEX_ADJUST.get(usize::from(nibble & 7)).copied().unwrap_or(0)).clamp(0, 88); + self.predictor as i16 + } + + /// The code that brings the predictor closest to `sample`, applied to the state. + fn encode(&mut self, sample: i16) -> u8 { + let step = step_at(self.index); + let mut diff = i32::from(sample) - self.predictor; + let mut code = 0u8; + if diff < 0 { + code = 8; + diff = -diff; + } + let mut s = step; + for bit in [4u8, 2, 1] { + if diff >= s { + code |= bit; + diff -= s; + } + s >>= 1; + } + self.step(code); + code + } +} + +/// Samples per channel in a block of `block_align` bytes; `None` when the layout is impossible +/// (a block must hold the headers and whole groups of eight samples, and there are 1 to 8 +/// channels). +pub fn samples_per_block(block_align: usize, channels: usize) -> Option { + if !(1..=8).contains(&channels) { + return None; + } + let header = 4 * channels; + let body = block_align.checked_sub(header)?; + if body == 0 || !body.is_multiple_of(4 * channels) { + return None; + } + let per = (body / channels).checked_mul(2)?.checked_add(1)?; + // Callers size interleaved buffers from this value; the full count must fit too. + per.checked_mul(channels)?; + Some(per) +} + +/// Decode one block, appending its samples interleaved by channel to `out`. +pub fn decode_block(block: &[u8], channels: usize, out: &mut Vec) -> Result<()> { + let per = samples_per_block(block.len(), channels).ok_or(Error::Malformed("block size does not fit the channel count"))?; + let mut states = Vec::with_capacity(channels); + for ch in 0..channels { + let h = block.get(4 * ch..4 * ch + 4).ok_or(Error::Malformed("truncated block header"))?; + let index = i32::from(h[2]); + if index > 88 { + return Err(Error::Malformed("step index above 88")); + } + states.push(State { predictor: i32::from(i16::from_le_bytes([h[0], h[1]])), index }); + } + let start = out.len(); + out.resize(start + per * channels, 0); + let samples = out.get_mut(start..).ok_or(Error::Malformed("output"))?; + for (ch, s) in states.iter().enumerate() { + if let Some(o) = samples.get_mut(ch) { + *o = s.predictor as i16; + } + } + let body = block.get(4 * channels..).ok_or(Error::Malformed("truncated block"))?; + for (g, group) in body.chunks_exact(4 * channels).enumerate() { + for (ch, bytes) in group.as_chunks::<4>().0.iter().enumerate() { + let Some(state) = states.get_mut(ch) else { continue }; + for (k, b) in bytes.iter().enumerate() { + for (half, nibble) in [b & 0xF, b >> 4].into_iter().enumerate() { + let t = 1 + g * 8 + k * 2 + half; + if let Some(o) = samples.get_mut(t * channels + ch) { + *o = state.step(nibble); + } + } + } + } + } + Ok(()) +} + +/// Decode a stream of blocks of `block_align` bytes. A final partial block is decoded as far as it +/// holds whole groups. +pub fn decode(data: &[u8], channels: usize, block_align: usize) -> Result> { + samples_per_block(block_align, channels).ok_or(Error::Malformed("block size does not fit the channel count"))?; + let mut out = Vec::new(); + for block in data.chunks(block_align) { + if block.len() == block_align { + decode_block(block, channels, &mut out)?; + } else { + let (header, group) = (4 * channels, 4 * channels); + let whole = block.len().saturating_sub(header) / group * group; + if whole > 0 { + decode_block(block.get(..header + whole).unwrap_or(block), channels, &mut out)?; + } + } + } + Ok(out) +} + +/// The starting step index for a channel whose first samples are `x`: the index whose first +/// codes follow `x` most closely. +fn start_index(first: i16, x: &[i16]) -> i32 { + let mut best = (i64::MAX, 0); + for index in 0..=88 { + let mut st = State { predictor: i32::from(first), index }; + let mut err = 0i64; + for &s in x.iter().take(16) { + let code = st.encode(s); + let _ = code; + err += (i64::from(s) - i64::from(st.predictor)).pow(2); + } + if err < best.0 { + best = (err, index); + } + } + best.1 +} + +/// Encode one block from exactly `samples_per_block * channels` interleaved samples. +pub fn encode_block(samples: &[i16], channels: usize, block_align: usize) -> Result> { + let per = samples_per_block(block_align, channels).ok_or(Error::InvalidInput("block size does not fit the channel count"))?; + if samples.len() != per * channels { + return Err(Error::InvalidInput("a block needs exactly samples_per_block samples per channel")); + } + let mut block = vec![0u8; block_align]; + let mut states = Vec::with_capacity(channels); + for ch in 0..channels { + let col: Vec = samples.iter().skip(ch).step_by(channels).copied().collect(); + let first = col.first().copied().unwrap_or(0); + let st = State { predictor: i32::from(first), index: start_index(first, col.get(1..).unwrap_or(&[])) }; + if let Some(h) = block.get_mut(4 * ch..4 * ch + 4) { + h[..2].copy_from_slice(&first.to_le_bytes()); + h[2] = st.index as u8; + } + states.push(st); + } + let groups = (block_align - 4 * channels) / (4 * channels); + for g in 0..groups { + for ch in 0..channels { + let Some(state) = states.get_mut(ch) else { continue }; + for k in 0..4 { + let mut byte = 0u8; + for half in 0..2 { + let t = 1 + g * 8 + k * 2 + half; + let nibble = state.encode(samples.get(t * channels + ch).copied().unwrap_or(0)); + byte |= nibble << (4 * half); + } + if let Some(b) = block.get_mut(4 * channels + g * 4 * channels + ch * 4 + k) { + *b = byte; + } + } + } + } + Ok(block) +} + +/// Encode interleaved samples as blocks of `block_align` bytes. The last block is padded with +/// copies of the last sample, so a decoder returns up to one block more than was given. +pub fn encode(samples: &[i16], channels: usize, block_align: usize) -> Result> { + let per = samples_per_block(block_align, channels).ok_or(Error::InvalidInput("block size does not fit the channel count"))?; + if !samples.len().is_multiple_of(channels) { + return Err(Error::InvalidInput("the sample count is not a multiple of the channel count")); + } + let mut out = Vec::new(); + for chunk in samples.chunks(per * channels) { + if chunk.len() == per * channels { + out.extend(encode_block(chunk, channels, block_align)?); + } else { + let mut padded = chunk.to_vec(); + let last: Vec = chunk.get(chunk.len().saturating_sub(channels)..).unwrap_or(&[]).to_vec(); + while padded.len() < per * channels { + padded.extend(last.iter().copied().chain(std::iter::repeat(0)).take(channels)); + } + out.extend(encode_block(&padded, channels, block_align)?); + } + } + Ok(out) +} + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn layout() { + assert_eq!(samples_per_block(256, 1), Some(505)); + assert_eq!(samples_per_block(512, 2), Some(505)); + assert_eq!(samples_per_block(4, 1), None); + assert_eq!(samples_per_block(100, 0), None); + assert_eq!(samples_per_block(258, 2), None); + } + + #[test] + fn unrepresentable_interleaved_layout_is_rejected() { + let align = usize::MAX - 7; + assert_eq!(samples_per_block(align, 2), None); + assert!(encode(&[], 2, align).is_err()); + assert!(encode_block(&[], 2, align).is_err()); + } + + #[test] + fn silence_stays_silent() { + let block = encode_block(&vec![0; 505], 1, 256).unwrap(); + let mut out = Vec::new(); + decode_block(&block, 1, &mut out).unwrap(); + assert!(out.iter().all(|&s| s == 0)); + } +} diff --git a/crates/audio/src/adpcm/mod.rs b/crates/audio/src/adpcm/mod.rs new file mode 100644 index 000000000..da552d3cd --- /dev/null +++ b/crates/audio/src/adpcm/mod.rs @@ -0,0 +1,56 @@ +//! ADPCM: a decoder and an encoder for four families, each on whole blocks or streams of `i16` +//! samples interleaved by channel (feature `adpcm`). +//! +//! - [`ima`]: IMA / DVI ADPCM in WAV blocks (format tag 0x0011): 4 bits per sample, a 4-byte header +//! per channel and groups of eight samples per channel. +//! - [`ms`]: Microsoft ADPCM in WAV blocks (format tag 0x0002): 4 bits per sample, adaptive +//! two-tap prediction with the seven standard coefficient pairs. +//! - [`yamaha`]: Yamaha ADPCM (the YM2608 / AICA family): a raw stream of 4-bit codes. +//! - [`swf`]: the ADPCM of Flash (SWF) audio: 2 to 5 bits per sample, in blocks of 4096 samples. +//! +//! Written from the public descriptions of the algorithms: the IMA Digital Audio Focus and +//! Technical Working Groups' *Recommended Practices for Enhancing Digital Audio Compatibility* +//! (IMA ADPCM), Microsoft's *Multimedia Standards Update* (the WAVE format tags and the block +//! layout of Microsoft ADPCM), the SWF file format specification (the ADPCM section) and the +//! Yamaha ADPCM-B description. ffmpeg is only run as a black-box oracle in the tests. +//! +//! Decoders return errors, never panic, on damaged blocks. Encoders are greedy: they choose each +//! code to bring the decoder's output closest to the sample. + +pub mod ima; +pub mod ms; +pub mod swf; +pub mod yamaha; + +/// What went wrong decoding or encoding ADPCM. +#[non_exhaustive] +#[derive(Debug, thiserror::Error, PartialEq, Eq)] +pub enum Error { + /// The data is damaged, truncated or does not fit the stated block layout. + #[error("malformed ADPCM: {0}")] + Malformed(&'static str), + /// The settings or samples given to an encoder cannot be written. + #[error("ADPCM cannot encode: {0}")] + InvalidInput(&'static str), +} + +/// Result with this module's [`Error`]. +pub type Result = std::result::Result; + +/// IMA ADPCM step sizes (the 89-entry table of the IMA recommended practices). +pub(crate) const STEP_TABLE: [i32; 89] = [ + 7, 8, 9, 10, 11, 12, 13, 14, 16, 17, 19, 21, 23, 25, 28, 31, 34, 37, 41, 45, 50, 55, 60, 66, 73, 80, 88, 97, 107, 118, 130, 143, 157, 173, 190, + 209, 230, 253, 279, 307, 337, 371, 408, 449, 494, 544, 598, 658, 724, 796, 876, 963, 1060, 1166, 1282, 1411, 1552, 1707, 1878, 2066, 2272, 2499, + 2749, 3024, 3327, 3660, 4026, 4428, 4871, 5358, 5894, 6484, 7132, 7845, 8630, 9493, 10442, 11487, 12635, 13899, 15289, 16818, 18500, 20350, + 22385, 24623, 27086, 29794, 32767, +]; + +/// Clamp to the 16-bit sample range. +pub(crate) fn clip16(v: i32) -> i32 { + v.clamp(i32::from(i16::MIN), i32::from(i16::MAX)) +} + +/// The step size at `index` (clamped to the table). +pub(crate) fn step_at(index: i32) -> i32 { + STEP_TABLE.get(index.clamp(0, 88) as usize).copied().unwrap_or(7) +} diff --git a/crates/audio/src/adpcm/ms.rs b/crates/audio/src/adpcm/ms.rs new file mode 100644 index 000000000..057f93549 --- /dev/null +++ b/crates/audio/src/adpcm/ms.rs @@ -0,0 +1,244 @@ +//! Microsoft ADPCM in WAV blocks (format tag 0x0002). +//! +//! A block starts, for all channels in turn, with one predictor index (a byte, 0 to 6), then one +//! initial step (`delta`, `i16`), then each channel's second-last sample (`sample1`, `i16`) and +//! last sample (`sample2`, `i16`) before the block; the block's first two samples per channel are +//! `sample2` then `sample1`. The rest is 4-bit two's-complement codes, the first channel in the +//! high nibble. A block of `block_align` bytes holds `(block_align - 7 * channels) * 2 / channels +//! + 2` samples per channel. Each sample is predicted as `(sample1 * c1 + sample2 * c2) / 256` (rounding toward zero) +//! from the block's coefficient pair, plus `code * delta`; then `delta` becomes +//! `max(16, delta * a / 256)`, where `a` is the table entry for the code (`ADAPT`). + +use super::{Error, Result, clip16}; + +/// The seven coefficient pairs `(c1, c2)` every Microsoft ADPCM stream uses. +pub const COEFFICIENTS: [(i32, i32); 7] = [(256, 0), (512, -256), (0, 0), (192, 64), (240, 0), (460, -208), (392, -232)]; +/// How the step moves with the code's unsigned value (the index is the 4-bit pattern). +/// The largest step. The reference keeps it in 16 bits and no real stream gets near this bound; it +/// keeps every product in range however long a damaged block runs. +const MAX_DELTA: i32 = 1 << 24; +const ADAPT: [i32; 16] = [230, 230, 230, 230, 307, 409, 512, 614, 768, 614, 512, 409, 307, 230, 230, 230]; + +#[derive(Clone, Copy)] +struct State { + c1: i32, + c2: i32, + delta: i32, + s1: i32, + s2: i32, +} + +impl State { + fn step(&mut self, nibble: u8) -> i16 { + let signed = i32::from(nibble) - if nibble & 8 != 0 { 16 } else { 0 }; + let predicted = (self.s1 * self.c1 + self.s2 * self.c2) / 256; + let sample = clip16(predicted + signed * self.delta); + self.s2 = self.s1; + self.s1 = sample; + let adapt = i64::from(ADAPT.get(usize::from(nibble)).copied().unwrap_or(256)); + let new_delta = (adapt * i64::from(self.delta)) >> 8; + self.delta = (new_delta.min(i64::from(MAX_DELTA)) as i32).max(16); + sample as i16 + } + + /// The code that brings the output closest to `sample`, applied to the state. + fn encode(&mut self, sample: i16) -> u8 { + let predicted = (self.s1 * self.c1 + self.s2 * self.c2) / 256; + let err = i32::from(sample) - predicted; + // round to the nearest multiple of delta + let q = if err >= 0 { (err + self.delta / 2) / self.delta } else { -((-err + self.delta / 2) / self.delta) }; + let code = (q.clamp(-8, 7) & 0xF) as u8; + self.step(code); + code + } +} + +/// Samples per channel in a block of `block_align` bytes; `None` for an impossible layout (1 or 2 +/// channels, and room for the header and at least one code byte). +pub fn samples_per_block(block_align: usize, channels: usize) -> Option { + if !(1..=2).contains(&channels) { + return None; + } + let body = block_align.checked_sub(7 * channels).filter(|&b| b > 0)?; + // codes pair up across channels + let per = (body.checked_mul(2)? / channels).checked_add(2)?; + // Include the header samples and all channels before callers size their buffers. + per.checked_mul(channels)?; + Some(per) +} + +/// Decode one block, appending its samples interleaved by channel to `out`. +pub fn decode_block(block: &[u8], channels: usize, out: &mut Vec) -> Result<()> { + let per = samples_per_block(block.len(), channels).ok_or(Error::Malformed("block size does not fit the channel count"))?; + let header = block.get(..7 * channels).ok_or(Error::Malformed("truncated block header"))?; + let mut states = Vec::with_capacity(channels); + let word = |i: usize| i32::from(i16::from_le_bytes([header.get(i).copied().unwrap_or(0), header.get(i + 1).copied().unwrap_or(0)])); + for ch in 0..channels { + let pred = usize::from(header.get(ch).copied().unwrap_or(0)); + let &(c1, c2) = COEFFICIENTS.get(pred).ok_or(Error::Malformed("predictor index above 6"))?; + states.push(State { c1, c2, delta: word(channels + 2 * ch), s1: word(3 * channels + 2 * ch), s2: word(5 * channels + 2 * ch) }); + } + let start = out.len(); + out.resize(start + per * channels, 0); + let samples = out.get_mut(start..).ok_or(Error::Malformed("output"))?; + for (ch, s) in states.iter().enumerate() { + if let Some(a) = samples.get_mut(ch) { + *a = s.s2 as i16; + } + if let Some(b) = samples.get_mut(channels + ch) { + *b = s.s1 as i16; + } + } + let codes = block.get(7 * channels..).ok_or(Error::Malformed("truncated block"))?; + let mut t = 2 * channels; + for byte in codes { + for nibble in [byte >> 4, byte & 0xF] { + let ch = t % channels; + if let (Some(o), Some(state)) = (samples.get_mut(t), states.get_mut(ch)) { + *o = state.step(nibble); + } + t += 1; + } + } + Ok(()) +} + +/// Decode a stream of blocks of `block_align` bytes. A final partial block is decoded in full if +/// it holds a header and a code byte. +pub fn decode(data: &[u8], channels: usize, block_align: usize) -> Result> { + samples_per_block(block_align, channels).ok_or(Error::Malformed("block size does not fit the channel count"))?; + let mut out = Vec::new(); + for block in data.chunks(block_align) { + if block.len() == block_align || samples_per_block(block.len(), channels).is_some() { + decode_block(block, channels, &mut out)?; + } + } + Ok(out) +} + +/// The squared error of encoding `col` (one channel) from state `st`. +fn trial(mut st: State, col: &[i16]) -> i64 { + let mut err = 0i64; + for &s in col { + st.encode(s); + err += (i64::from(s) - i64::from(st.s1)).pow(2); + } + err +} + +/// Encode one block from exactly `samples_per_block * channels` interleaved samples. Each channel +/// gets the coefficient pair and initial step that code it with the least error. +pub fn encode_block(samples: &[i16], channels: usize, block_align: usize) -> Result> { + let per = samples_per_block(block_align, channels).ok_or(Error::InvalidInput("block size does not fit the channel count"))?; + if samples.len() != per * channels { + return Err(Error::InvalidInput("a block needs exactly samples_per_block samples per channel")); + } + let mut block = vec![0u8; block_align]; + let mut states = Vec::with_capacity(channels); + for ch in 0..channels { + let col: Vec = samples.iter().skip(ch).step_by(channels).copied().collect(); + let (x0, x1) = (col.first().copied().unwrap_or(0), col.get(1).copied().unwrap_or(0)); + let rest = col.get(2..).unwrap_or(&[]); + let mut best: Option<(i64, usize, i32)> = None; + for (p, &(c1, c2)) in COEFFICIENTS.iter().enumerate() { + for delta in [16, 64, 256, 1024] { + let st = State { c1, c2, delta, s1: i32::from(x1), s2: i32::from(x0) }; + let e = trial(st, rest); + if best.is_none_or(|(b, _, _)| e < b) { + best = Some((e, p, delta)); + } + } + } + let (_, p, delta) = best.unwrap_or((0, 0, 16)); + let &(c1, c2) = COEFFICIENTS.get(p).unwrap_or(&(256, 0)); + states.push(State { c1, c2, delta, s1: i32::from(x1), s2: i32::from(x0) }); + let (d, a, b) = (delta as i16, x1, x0); + let put = |block: &mut Vec, at: usize, v: i16| { + if let Some(w) = block.get_mut(at..at + 2) { + w.copy_from_slice(&v.to_le_bytes()); + } + }; + if let Some(h) = block.get_mut(ch) { + *h = p as u8; + } + put(&mut block, channels + 2 * ch, d); + put(&mut block, 3 * channels + 2 * ch, a); + put(&mut block, 5 * channels + 2 * ch, b); + } + let mut t = 2 * channels; + for byte in block.iter_mut().skip(7 * channels) { + let mut b = 0u8; + for half in 0..2 { + let ch = t % channels; + let nibble = match states.get_mut(ch) { + Some(st) => st.encode(samples.get(t).copied().unwrap_or(0)), + None => 0, + }; + b |= nibble << (4 * (1 - half)); + t += 1; + } + *byte = b; + } + Ok(block) +} + +/// Encode interleaved samples as blocks of `block_align` bytes. The last block is padded with +/// copies of the last frame, so a decoder returns up to one block more than was given. +pub fn encode(samples: &[i16], channels: usize, block_align: usize) -> Result> { + let per = samples_per_block(block_align, channels).ok_or(Error::InvalidInput("block size does not fit the channel count"))?; + if !samples.len().is_multiple_of(channels) { + return Err(Error::InvalidInput("the sample count is not a multiple of the channel count")); + } + let mut out = Vec::new(); + for chunk in samples.chunks(per * channels) { + let mut padded = chunk.to_vec(); + let last: Vec = chunk.get(chunk.len().saturating_sub(channels)..).unwrap_or(&[]).to_vec(); + while padded.len() < per * channels { + padded.extend(last.iter().copied().chain(std::iter::repeat(0)).take(channels)); + } + out.extend(encode_block(&padded, channels, block_align)?); + } + Ok(out) +} + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn layout() { + assert_eq!(samples_per_block(256, 1), Some(500)); + assert_eq!(samples_per_block(512, 2), Some(500)); + assert_eq!(samples_per_block(7, 1), None); + assert_eq!(samples_per_block(100, 3), None); + } + + #[test] + fn unrepresentable_layouts_are_rejected() { + for (align, channels) in [(usize::MAX / 2 + 7, 1), (usize::MAX / 2 + 14, 2)] { + assert_eq!(samples_per_block(align, channels), None); + assert!(encode(&[], channels, align).is_err()); + assert!(encode_block(&[], channels, align).is_err()); + } + } + + #[test] + fn a_long_run_of_large_codes_cannot_overflow() { + // a step that starts at its largest 16-bit value and grows with every large code + let mut block = vec![0x77u8; 7 + 2000]; + block[..7].copy_from_slice(&[0, 255, 127, 0, 0, 0, 0]); + for b in block.iter_mut().skip(7) { + *b = 0x88; + } + let mut out = Vec::new(); + decode_block(&block, 1, &mut out).unwrap(); + assert_eq!(out.len(), 2000 * 2 + 2); + } + + #[test] + fn bad_predictor_index_is_an_error() { + let mut block = vec![0u8; 64]; + block[0] = 7; + assert!(decode_block(&block, 1, &mut Vec::new()).is_err()); + } +} diff --git a/crates/audio/src/adpcm/swf.rs b/crates/audio/src/adpcm/swf.rs new file mode 100644 index 000000000..2fc780b1d --- /dev/null +++ b/crates/audio/src/adpcm/swf.rs @@ -0,0 +1,168 @@ +//! The ADPCM of Flash (SWF) audio, 2 to 5 bits per sample. +//! +//! The stream starts with two bits, the code size minus 2. It continues in blocks of 4096 samples +//! per channel: each block starts, for every channel, with the first sample (16 bits, signed) and +//! the step index (6 bits, 0 to 88), then 4095 codes per channel, interleaved by sample. All fields +//! are packed most significant bit first. A code is a sign bit and `bits - 1` magnitude bits; the +//! decoder adds `step >> (bits - 1)` plus `step >> k` for each magnitude bit set `k` places below +//! the top one, negated when the sign is set, and moves the step index by the adjustment table of +//! the code size (negative for small magnitudes, positive for large ones). The last block may +//! be shorter. + +use super::{Error, Result, clip16, step_at}; +use craft_media::bitstream::{BitReader, BitWriter}; + +/// Samples per channel in a full block. +pub const BLOCK_SAMPLES: usize = 4096; + +/// The index adjustment for each magnitude, per code size. +fn adjust(bits: u32) -> &'static [i32] { + match bits { + 2 => &[-1, 2], + 3 => &[-1, -1, 2, 4], + 4 => &[-1, -1, -1, -1, 2, 4, 6, 8], + _ => &[-1, -1, -1, -1, -1, -1, -1, -1, 1, 2, 4, 6, 8, 10, 13, 16], + } +} + +#[derive(Clone, Copy)] +struct State { + predictor: i32, + index: i32, +} + +impl State { + fn apply(&mut self, code: u32, bits: u32) -> i16 { + let step = step_at(self.index); + let mag = code & ((1 << (bits - 1)) - 1); + let mut diff = step >> (bits - 1); + for k in 0..bits - 1 { + if mag & (1 << (bits - 2 - k)) != 0 { + diff += step >> k; + } + } + if code & (1 << (bits - 1)) != 0 { + diff = -diff; + } + self.predictor = clip16(self.predictor + diff); + self.index = (self.index + adjust(bits).get(mag as usize).copied().unwrap_or(0)).clamp(0, 88); + self.predictor as i16 + } + + fn encode(&mut self, sample: i16, bits: u32) -> u32 { + let step = step_at(self.index); + let mut diff = i32::from(sample) - self.predictor; + let mut code = 0u32; + if diff < 0 { + code = 1 << (bits - 1); + diff = -diff; + } + let mut s = step; + for k in 0..bits - 1 { + if diff >= s { + code |= 1 << (bits - 2 - k); + diff -= s; + } + s >>= 1; + } + self.apply(code, bits); + code + } +} + +/// Decode a stream. `channels` is 1 or 2; the code size comes from the stream's first two bits. +/// The result holds every sample up to the end of the data (a last block cut short decodes as +/// far as it goes). +pub fn decode(data: &[u8], channels: usize) -> Result> { + if !(1..=2).contains(&channels) { + return Err(Error::Malformed("Flash ADPCM has one or two channels")); + } + let mut r = BitReader::new(data); + let bits = r.read_bits(2).map_err(|_| Error::Malformed("empty stream"))? + 2; + let mut out = Vec::new(); + let header_bits = channels * 22; + while r.bits_left() >= header_bits { + let mut states = Vec::with_capacity(channels); + for _ in 0..channels { + let sample = r.read_bits(16).map_err(|_| Error::Malformed("truncated block"))? as u16 as i16; + let index = r.read_bits(6).map_err(|_| Error::Malformed("truncated block"))? as i32; + if index > 88 { + return Err(Error::Malformed("step index above 88")); + } + out.push(sample); + states.push(State { predictor: i32::from(sample), index }); + } + 'block: for _ in 1..BLOCK_SAMPLES { + if r.bits_left() < channels * (bits as usize) { + break 'block; + } + for st in states.iter_mut() { + match r.read_bits(bits) { + Ok(code) => out.push(st.apply(code, bits)), + Err(_) => break 'block, + } + } + } + } + Ok(out) +} + +/// Encode interleaved samples with `bits` (2 to 5) bits per sample. Blocks hold 4096 samples per +/// channel; the stream ends at a byte boundary. +pub fn encode(samples: &[i16], channels: usize, bits: u32) -> Result> { + if !(1..=2).contains(&channels) { + return Err(Error::InvalidInput("Flash ADPCM has one or two channels")); + } + if !(2..=5).contains(&bits) { + return Err(Error::InvalidInput("Flash ADPCM codes have 2 to 5 bits")); + } + if !samples.len().is_multiple_of(channels) { + return Err(Error::InvalidInput("the sample count is not a multiple of the channel count")); + } + let mut w = BitWriter::new(); + w.write_bits(bits - 2, 2); + for block in samples.chunks(BLOCK_SAMPLES * channels) { + let mut states = Vec::with_capacity(channels); + for ch in 0..channels { + let col: Vec = block.iter().skip(ch).step_by(channels).copied().collect(); + let first = col.first().copied().unwrap_or(0); + // the index whose first codes follow the signal most closely + let mut best = (i64::MAX, 0); + for index in 0..=63 { + let mut st = State { predictor: i32::from(first), index }; + let mut err = 0i64; + for &s in col.iter().skip(1).take(16) { + st.encode(s, bits); + err += (i64::from(s) - i64::from(st.predictor)).pow(2); + } + if err < best.0 { + best = (err, index); + } + } + w.write_bits(u32::from(first as u16), 16); + w.write_bits(best.1 as u32, 6); + states.push(State { predictor: i32::from(first), index: best.1 }); + } + for frame in block.chunks(channels).skip(1) { + for (st, &s) in states.iter_mut().zip(frame) { + w.write_bits(st.encode(s, bits), bits); + } + } + } + w.align_zero(); + Ok(w.finish()) +} + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn settings_are_checked() { + assert!(encode(&[0], 3, 4).is_err()); + assert!(encode(&[0], 1, 1).is_err()); + assert!(encode(&[0], 1, 6).is_err()); + assert!(decode(&[], 1).is_err()); + assert!(decode(&[0], 0).is_err()); + } +} diff --git a/crates/audio/src/adpcm/yamaha.rs b/crates/audio/src/adpcm/yamaha.rs new file mode 100644 index 000000000..22ea0e77f --- /dev/null +++ b/crates/audio/src/adpcm/yamaha.rs @@ -0,0 +1,113 @@ +//! Yamaha ADPCM (the ADPCM-B of the YM2608 and the AICA): a raw stream of 4-bit codes, no header. +//! +//! Each channel starts at predictor 0 and step 127. A code is a sign bit and three magnitude bits; +//! the sample moves by `step * [1, 3, 5, 7, 9, 11, 13, 15][magnitude] / 8` (negative when the sign +//! is set) and the step is multiplied by `[230, 230, 230, 230, 307, 409, 512, 614][magnitude] / 256`, +//! kept within 127 and 24576. One channel: the low nibble of a byte is the first sample. Two +//! channels: the low nibble is the left channel and the high nibble the right one of the same +//! frame. + +use super::{Error, Result, clip16}; + +const SCALE: [i32; 8] = [230, 230, 230, 230, 307, 409, 512, 614]; + +#[derive(Clone, Copy)] +struct State { + predictor: i32, + step: i32, +} + +impl State { + fn new() -> State { + State { predictor: 0, step: 127 } + } + + fn apply(&mut self, nibble: u8) -> i16 { + let m = usize::from(nibble & 7); + let diff = self.step * (2 * m as i32 + 1) / 8; + self.predictor = clip16(self.predictor + if nibble & 8 != 0 { -diff } else { diff }); + self.step = ((self.step * SCALE.get(m).copied().unwrap_or(256)) >> 8).clamp(127, 24_576); + self.predictor as i16 + } + + fn encode(&mut self, sample: i16) -> u8 { + let err = i32::from(sample) - self.predictor; + let m = (err.abs() * 8 / self.step / 2).clamp(0, 7); + // the magnitude whose reconstruction (2m + 1) * step / 8 is nearest + let mut best = (i32::MAX, 0u8); + for cand in [m - 1, m, m + 1] { + let c = cand.clamp(0, 7); + let recon = self.step * (2 * c + 1) / 8; + let e = (err.abs() - recon).abs(); + if e < best.0 { + best = (e, c as u8); + } + } + let nibble = best.1 | if err < 0 { 8 } else { 0 }; + self.apply(nibble); + nibble + } +} + +/// Decode a stream of `channels` (1 or 2) interleaved channels. +pub fn decode(data: &[u8], channels: usize) -> Result> { + if !(1..=2).contains(&channels) { + return Err(Error::Malformed("Yamaha ADPCM has one or two channels")); + } + let mut states = vec![State::new(); channels]; + let mut out = Vec::with_capacity(data.len() * 2); + let mut t = 0usize; + for byte in data { + for nibble in [byte & 0xF, byte >> 4] { + if let Some(st) = states.get_mut(t % channels) { + out.push(st.apply(nibble)); + } + t += 1; + } + } + Ok(out) +} + +/// Encode interleaved samples. With two channels the frame count must be whole bytes, which it +/// always is; with one channel an odd count is padded with a copy of the last sample. +pub fn encode(samples: &[i16], channels: usize) -> Result> { + if !(1..=2).contains(&channels) { + return Err(Error::InvalidInput("Yamaha ADPCM has one or two channels")); + } + if !samples.len().is_multiple_of(channels) { + return Err(Error::InvalidInput("the sample count is not a multiple of the channel count")); + } + let mut states = vec![State::new(); channels]; + let mut out = Vec::with_capacity(samples.len().div_ceil(2)); + let mut t = 0usize; + let mut padded = samples.to_vec(); + if padded.len() % 2 == 1 { + padded.push(padded.last().copied().unwrap_or(0)); + } + for pair in padded.chunks(2) { + let mut byte = 0u8; + for (half, &s) in pair.iter().enumerate() { + let code = match states.get_mut(t % channels) { + Some(st) => st.encode(s), + None => 0, + }; + byte |= code << (4 * half); + t += 1; + } + out.push(byte); + } + Ok(out) +} + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn channel_count_is_checked() { + assert!(decode(&[0], 0).is_err()); + assert!(decode(&[0], 3).is_err()); + assert!(encode(&[0], 3).is_err()); + assert!(encode(&[0, 0, 0], 2).is_err()); + } +} diff --git a/crates/audio/src/lib.rs b/crates/audio/src/lib.rs index f7ed4b948..97b08f794 100644 --- a/crates/audio/src/lib.rs +++ b/crates/audio/src/lib.rs @@ -30,7 +30,8 @@ //! `aac-encode`), [`alac`] (`alac-decode`, `alac-encode`), [`opus`] (`opus-decode`, and //! [`opus::encoder`] with `opus-encode`), [`ac3`] (`ac3-decode`), [`mpa`] (`mpa-decode`; `mpa-encode`: the Layer II encoder), [`mp3`] //! (`mp3-decode`, `mp3-encode`), [`flac`] (`flac-decode`: the frame and file decoder; -//! `flac-encode`: the packet encoder), [`qoa`] (`qoa-decode`, `qoa-encode`: whole QOA files), +//! `flac-encode`: the packet encoder), [`qoa`] (`qoa-decode`, `qoa-encode`: whole QOA files), [`adpcm`] +//! (`adpcm`: IMA WAV, Microsoft, Yamaha and Flash ADPCM), //! [`gsm`] (`gsm-decode`, `gsm-encode`: GSM 06.10 full-rate speech), [`vmd`] (`vmd-decode`: //! Sierra VMD sound). //! @@ -40,6 +41,8 @@ pub mod aac; #[cfg(any(feature = "ac3-decode", feature = "ac3-encode", feature = "eac3-encode"))] pub mod ac3; +#[cfg(feature = "adpcm")] +pub mod adpcm; #[cfg(any(feature = "alac-decode", feature = "alac-encode"))] pub mod alac; mod buffer; diff --git a/crates/audio/tests/adpcm_oracle.rs b/crates/audio/tests/adpcm_oracle.rs new file mode 100644 index 000000000..c89fe6444 --- /dev/null +++ b/crates/audio/tests/adpcm_oracle.rs @@ -0,0 +1,516 @@ +//! ADPCM (feature `adpcm`): every family decodes what ffmpeg's encoder writes to ffmpeg's own +//! samples, and ffmpeg decodes what our encoder writes to our own decoder's samples (oracle; +//! skipped when ffmpeg is absent, a failure with `CRAFT_MEDIA_REQUIRE_ORACLES=1`); plus the +//! round trip through our own encoder and decoder, and damaged input. + +use craft_audio::adpcm::{ima, ms, swf, yamaha}; +use std::path::Path; +use std::process::Command; + +const RATE: u32 = 22_050; + +/// Interleaved test audio: tones, a sweep, noise bursts, silence and a full-scale square. +fn signal(channels: usize, frames: usize) -> Vec { + let mut seed = 0x2545_F491u32; + let mut out = Vec::with_capacity(frames * channels); + for i in 0..frames { + for c in 0..channels { + seed ^= seed << 13; + seed ^= seed >> 17; + seed ^= seed << 5; + let t = i as f64; + let v = match (i / 1500) % 5 { + 0 => (t * (0.03 + 0.011 * c as f64)).sin() * 9000.0, + 1 => (t * 0.002 * (1.0 + t / 3000.0)).sin() * 14000.0, + 2 => f64::from(seed % 12000) - 6000.0, + 3 => 0.0, + _ => { + if (i / 40) % 2 == 0 { + 20000.0 + } else { + -20000.0 + } + } + }; + out.push(v as i16); + } + } + out +} + +fn snr(a: &[i16], b: &[i16]) -> f64 { + let (mut s, mut n) = (0f64, 0f64); + for (x, y) in a.iter().zip(b) { + s += f64::from(*x).powi(2); + n += (f64::from(*x) - f64::from(*y)).powi(2); + } + 10.0 * (s / n.max(1.0)).log10() +} + +fn max_diff(a: &[i16], b: &[i16]) -> i32 { + a.iter().zip(b).map(|(x, y)| (i32::from(*x) - i32::from(*y)).abs()).max().unwrap_or(0) +} + +#[test] +fn encode_then_decode_is_close_to_the_input() { + for channels in [1usize, 2] { + let x = signal(channels, 20_000); + // (name, decoded, minimum SNR dB, a few dB under what each reaches): the signal has noise bursts + // and full-scale square waves, which codes this small follow loosely + let cases: Vec<(&str, Vec, f64)> = vec![ + ("ima", ima::decode(&ima::encode(&x, channels, 512).unwrap(), channels, 512).unwrap(), 5.0), + ("ms", ms::decode(&ms::encode(&x, channels, 512).unwrap(), channels, 512).unwrap(), 18.0), + ("yamaha", yamaha::decode(&yamaha::encode(&x, channels).unwrap(), channels).unwrap(), 5.0), + ("swf2", swf::decode(&swf::encode(&x, channels, 2).unwrap(), channels).unwrap(), 1.0), + ("swf3", swf::decode(&swf::encode(&x, channels, 3).unwrap(), channels).unwrap(), 3.0), + ("swf4", swf::decode(&swf::encode(&x, channels, 4).unwrap(), channels).unwrap(), 5.0), + ("swf5", swf::decode(&swf::encode(&x, channels, 5).unwrap(), channels).unwrap(), 7.0), + ]; + for (name, y, min) in cases { + assert!(y.len() >= x.len(), "{name}: {} samples for {}", y.len(), x.len()); + let s = snr(&x, &y[..x.len()]); + println!("{name} {channels}ch: SNR {s:.1} dB"); + assert!(s >= min, "{name} {channels}ch: SNR {s:.1} dB under {min}"); + } + } +} + +#[test] +fn damaged_input_is_an_error_or_decodes_never_panics() { + let x = signal(2, 4000); + let streams: Vec<(&str, Vec)> = vec![ + ("ima", ima::encode(&x, 2, 256).unwrap()), + ("ms", ms::encode(&x, 2, 256).unwrap()), + ("yamaha", yamaha::encode(&x, 2).unwrap()), + ("swf", swf::encode(&x, 2, 4).unwrap()), + ]; + let mut seed = 99u32; + for (name, s) in &streams { + for cut in (0..s.len()).step_by(7) { + let part = &s[..cut]; + let _ = match *name { + "ima" => ima::decode(part, 2, 256).map(drop), + "ms" => ms::decode(part, 2, 256).map(drop), + "yamaha" => yamaha::decode(part, 2).map(drop), + _ => swf::decode(part, 2).map(drop), + }; + } + for _ in 0..300 { + let mut d = s.clone(); + seed ^= seed << 13; + seed ^= seed >> 17; + seed ^= seed << 5; + let at = seed as usize % d.len(); + d[at] ^= (seed >> 8) as u8 | 1; + let _ = match *name { + "ima" => ima::decode(&d, 2, 256).map(drop), + "ms" => ms::decode(&d, 2, 256).map(drop), + "yamaha" => yamaha::decode(&d, 2).map(drop), + _ => swf::decode(&d, 2).map(drop), + }; + } + } + // the impossible layouts + assert!(ima::decode(&[0; 10], 2, 10).is_err()); + assert!(ima::decode(&[0; 8], 9, 8).is_err()); + assert!(ms::decode(&[0; 8], 3, 8).is_err()); + assert!(ima::encode(&[0; 3], 2, 256).is_err()); + assert!(ima::encode_block(&[0; 3], 1, 256).is_err()); + assert!(ms::encode_block(&[0; 3], 1, 256).is_err()); + // a block whose step index or predictor is out of range + let mut bad = ima::encode_block(&vec![0; 505], 1, 256).unwrap(); + bad[2] = 200; + assert!(ima::decode_block(&bad, 1, &mut Vec::new()).is_err()); + let mut bad = swf::encode(&[0; 10], 1, 4).unwrap(); + bad[2] = 0xFF; + let _ = swf::decode(&bad, 1); +} + +#[test] +fn huge_layouts_and_big_jumps_are_handled() { + // block sizes whose sample counts do not fit in a usize + assert_eq!(ima::samples_per_block(usize::MAX - 3, 1), None); + assert_eq!(ms::samples_per_block(usize::MAX, 1), None); + assert!(ima::encode(&[], 1, usize::MAX - 3).is_err()); + assert!(ms::encode(&[], 1, usize::MAX).is_err()); + // a Flash stream's header holds the step index in six bits: a jump from silence to full scale + // still decodes to what the encoder's own state says + for bits in [2u32, 3, 4, 5] { + let x = [0i16, 32767, -32768, 0, 0]; + let y = swf::decode(&swf::encode(&x, 1, bits).unwrap(), 1).unwrap(); + // the stream has no length: up to seven padding bits can read as one more code + assert!((x.len()..=x.len() + 3).contains(&y.len()), "{bits} bits: {} samples", y.len()); + assert!(y[1] > 0 && y[2] < y[1], "{bits} bits: {y:?}"); + } + // padding bits at the end of a stream are not a code: no incomplete stereo frame + for bits in [2u32, 3, 4, 5] { + let y = swf::decode(&swf::encode(&[0, 0, 5, -5, 9, 9], 2, bits).unwrap(), 2).unwrap(); + assert_eq!(y.len() % 2, 0, "{bits} bits: {y:?}"); + assert!((6..=8).contains(&y.len()), "{bits} bits: {} samples", y.len()); + } +} + +#[test] +fn partial_final_blocks_are_handled() { + let x = signal(1, 1000); + // IMA: 256-byte blocks hold 505 samples; 1000 samples is two blocks, the second padded + let enc = ima::encode(&x, 1, 256).unwrap(); + assert_eq!(enc.len(), 2 * 256); + assert_eq!(ima::decode(&enc, 1, 256).unwrap().len(), 1010); + // a stream cut inside the last block decodes as far as it goes + let cut = ima::decode(&enc[..256 + 4 + 40], 1, 256).unwrap(); + assert_eq!(cut.len(), 505 + 1 + 80); + let enc = ms::encode(&x, 1, 256).unwrap(); + assert_eq!(ms::decode(&enc, 1, 256).unwrap().len(), 1000); + assert!(ima::encode(&[], 1, 256).unwrap().is_empty()); + assert!(ms::encode(&[], 1, 256).unwrap().is_empty()); + assert!(yamaha::encode(&[], 1).unwrap().is_empty()); + assert_eq!(yamaha::decode(&yamaha::encode(&[5, 6, 7], 1).unwrap(), 1).unwrap().len(), 4); +} + +// ---- the other rounding some ffmpeg versions use for IMA and Flash ADPCM +// +// The IMA recommended practices build the step difference from `step >> 3` and the shifted steps +// of the magnitude bits that are set; some decoders (including older ffmpeg releases) compute +// `((2 * magnitude + 1) * step) >> (bits - 1)` instead, which rounds differently by a few units +// and so drifts apart through the predictor. A stream decodes to ours or to this one. + +fn alt_diff(mag: i32, step: i32, bits: u32) -> i32 { + ((2 * mag + 1) * step) >> (bits - 1) +} + +const INDEX_4: [i32; 8] = [-1, -1, -1, -1, 2, 4, 6, 8]; +const INDEX_5: [i32; 16] = [-1, -1, -1, -1, -1, -1, -1, -1, 1, 2, 4, 6, 8, 10, 13, 16]; +const STEPS: [i32; 89] = [ + 7, 8, 9, 10, 11, 12, 13, 14, 16, 17, 19, 21, 23, 25, 28, 31, 34, 37, 41, 45, 50, 55, 60, 66, 73, 80, 88, 97, 107, 118, 130, 143, 157, 173, 190, + 209, 230, 253, 279, 307, 337, 371, 408, 449, 494, 544, 598, 658, 724, 796, 876, 963, 1060, 1166, 1282, 1411, 1552, 1707, 1878, 2066, 2272, 2499, + 2749, 3024, 3327, 3660, 4026, 4428, 4871, 5358, 5894, 6484, 7132, 7845, 8630, 9493, 10442, 11487, 12635, 13899, 15289, 16818, 18500, 20350, + 22385, 24623, 27086, 29794, 32767, +]; + +fn index_table(bits: u32) -> &'static [i32] { + match bits { + 2 => &[-1, 2], + 3 => &[-1, -1, 2, 4], + 4 => &INDEX_4, + _ => &INDEX_5, + } +} + +/// `ima::decode` with the other rounding. +fn alt_ima_decode(data: &[u8], channels: usize, block_align: usize) -> Vec { + let per = ima::samples_per_block(block_align, channels).expect("layout"); + let mut out = Vec::new(); + for block in data.chunks_exact(block_align) { + let mut pred = vec![0i32; channels]; + let mut idx = vec![0i32; channels]; + let base = out.len(); + out.resize(base + per * channels, 0); + for ch in 0..channels { + pred[ch] = i32::from(i16::from_le_bytes([block[4 * ch], block[4 * ch + 1]])); + idx[ch] = i32::from(block[4 * ch + 2]); + out[base + ch] = pred[ch] as i16; + } + for (g, group) in block[4 * channels..].chunks_exact(4 * channels).enumerate() { + for (ch, bytes) in group.as_chunks::<4>().0.iter().enumerate() { + for (k, b) in bytes.iter().enumerate() { + for (half, nib) in [b & 0xF, b >> 4].into_iter().enumerate() { + let step = STEPS[idx[ch] as usize]; + let d = alt_diff(i32::from(nib & 7), step, 4); + pred[ch] = (pred[ch] + if nib & 8 != 0 { -d } else { d }).clamp(-32768, 32767); + idx[ch] = (idx[ch] + INDEX_4[usize::from(nib & 7)]).clamp(0, 88); + out[base + (1 + g * 8 + k * 2 + half) * channels + ch] = pred[ch] as i16; + } + } + } + } + } + out +} + +/// `swf::decode` with the other rounding (one stream). +fn alt_swf_decode(data: &[u8], channels: usize) -> Vec { + let mut bitpos = 0usize; + let mut get = |n: usize| -> Option { + let mut v = 0u32; + for _ in 0..n { + let byte = *data.get(bitpos / 8)?; + v = (v << 1) | u32::from((byte >> (7 - bitpos % 8)) & 1); + bitpos += 1; + } + Some(v) + }; + let bits = get(2).expect("header") + 2; + let mut out = Vec::new(); + loop { + let mut pred = Vec::new(); + let mut idx = Vec::new(); + for _ in 0..channels { + let (Some(p), Some(i)) = (get(16), get(6)) else { return out }; + out.push(p as u16 as i16); + pred.push(i32::from(p as u16 as i16)); + idx.push(i as i32); + } + for _ in 1..4096 { + for ch in 0..channels { + let Some(code) = get(bits as usize) else { return out }; + let mag = (code & ((1 << (bits - 1)) - 1)) as i32; + let d = alt_diff(mag, STEPS[idx[ch] as usize], bits); + pred[ch] = (pred[ch] + if code >> (bits - 1) != 0 { -d } else { d }).clamp(-32768, 32767); + idx[ch] = (idx[ch] + index_table(bits)[mag as usize]).clamp(0, 88); + out.push(pred[ch] as i16); + } + } + } +} + +/// Assert `theirs` is `ours` or `alt` over their common length (and say which). +fn same_as_either(what: &str, ours: &[i16], alt: &[i16], theirs: &[i16]) { + let n = ours.len().min(theirs.len()); + let (a, b) = (max_diff(&ours[..n], &theirs[..n]), max_diff(&alt[..n.min(alt.len())], &theirs[..n.min(alt.len())])); + println!("{what}: largest difference from ours {a}, from the other rounding {b}"); + assert!(a == 0 || b == 0, "{what}: ffmpeg's decode is neither ours ({a}) nor the other rounding ({b})"); +} + +/// A name, the WAV file around our stream, our decode of it and the other rounding's. +type Stream = (&'static str, Vec, Vec, Vec); + +// ---- ffmpeg as the reference + +fn ffmpeg(ff: &Path, args: &[&str]) { + let out = Command::new(ff).args(["-v", "error", "-nostdin", "-y"]).args(args).output().expect("run ffmpeg"); + assert!(out.status.success(), "ffmpeg {args:?}: {}", String::from_utf8_lossy(&out.stderr)); +} + +fn s(p: &Path) -> String { + p.to_str().expect("path").to_string() +} + +fn write_pcm(path: &Path, x: &[i16]) { + std::fs::write(path, x.iter().flat_map(|v| v.to_le_bytes()).collect::>()).expect("write"); +} + +fn read_pcm(path: &Path) -> Vec { + std::fs::read(path).expect("read").as_chunks::<2>().0.iter().map(|b| i16::from_le_bytes(*b)).collect() +} + +/// The `fmt ` chunk's fields and the `data` chunk of a WAV file. +fn wav_parts(bytes: &[u8]) -> (u16, usize, usize, Vec) { + let mut at = 12; + let (mut tag, mut channels, mut align) = (0, 0, 0); + while at + 8 <= bytes.len() { + let id = &bytes[at..at + 4]; + let len = u32::from_le_bytes([bytes[at + 4], bytes[at + 5], bytes[at + 6], bytes[at + 7]]) as usize; + let body = &bytes[at + 8..(at + 8 + len).min(bytes.len())]; + if id == b"fmt " { + tag = u16::from_le_bytes([body[0], body[1]]); + channels = usize::from(u16::from_le_bytes([body[2], body[3]])); + align = usize::from(u16::from_le_bytes([body[12], body[13]])); + } else if id == b"data" { + return (tag, channels, align, body.to_vec()); + } + at += 8 + len + (len & 1); + } + panic!("no data chunk"); +} + +/// A WAV file around `data`. +fn wav(tag: u16, channels: usize, align: usize, bits: u16, extra: &[u8], data: &[u8]) -> Vec { + let mut fmt = Vec::new(); + fmt.extend(tag.to_le_bytes()); + fmt.extend((channels as u16).to_le_bytes()); + fmt.extend(RATE.to_le_bytes()); + fmt.extend((RATE * align as u32 / 500).to_le_bytes()); + fmt.extend((align as u16).to_le_bytes()); + fmt.extend(bits.to_le_bytes()); + if !extra.is_empty() || tag == 0x11 || tag == 2 { + fmt.extend((extra.len() as u16).to_le_bytes()); + fmt.extend(extra); + } + let mut out = b"RIFF".to_vec(); + out.extend((4 + 8 + fmt.len() as u32 + 8 + data.len() as u32).to_le_bytes()); + out.extend(b"WAVEfmt "); + out.extend((fmt.len() as u32).to_le_bytes()); + out.extend(fmt); + out.extend(b"data"); + out.extend((data.len() as u32).to_le_bytes()); + out.extend(data); + out +} + +#[test] +fn we_decode_what_ffmpeg_encodes_to_the_samples_ffmpeg_decodes() { + let Some(ff) = craft_media_testkit::ffmpeg() else { return }; + let dir = craft_media_testkit::fixture_subdir("adpcm"); + for channels in [1usize, 2] { + let x = signal(channels, 30_000); + let pcm = dir.join(format!("in{channels}.pcm")); + write_pcm(&pcm, &x); + let input = ["-f", "s16le", "-ar", "22050", "-ac", if channels == 1 { "1" } else { "2" }, "-i"]; + for (name, codec, align) in [("ima", "adpcm_ima_wav", 512usize), ("ms", "adpcm_ms", 512), ("yamaha", "adpcm_yamaha", 0)] { + let file = dir.join(format!("{name}{channels}.wav")); + let mut args: Vec = input.iter().map(|a| a.to_string()).collect(); + args.extend([s(&pcm), "-c:a".into(), codec.into()]); + if align > 0 { + args.extend(["-block_size".into(), align.to_string()]); + } + args.push(s(&file)); + ffmpeg(&ff, &args.iter().map(String::as_str).collect::>()); + let back = dir.join(format!("{name}{channels}.out.pcm")); + ffmpeg(&ff, &["-i", &s(&file), "-f", "s16le", "-acodec", "pcm_s16le", &s(&back)]); + let theirs = read_pcm(&back); + let (tag, ch, block, data) = wav_parts(&std::fs::read(&file).expect("read")); + assert_eq!(ch, channels); + let ours = match name { + "ima" => { + assert_eq!(tag, 0x11); + ima::decode(&data, ch, block).expect("decode") + } + "ms" => { + assert_eq!(tag, 2); + ms::decode(&data, ch, block).expect("decode") + } + _ => yamaha::decode(&data, ch).expect("decode"), + }; + let n = ours.len().min(theirs.len()); + assert!(n >= 29_000 * channels, "{name} {channels}ch: {} ours, {} theirs", ours.len(), theirs.len()); + if name == "ima" { + same_as_either(&format!("{name} {channels}ch"), &ours, &alt_ima_decode(&data, ch, block), &theirs); + } else { + let d = max_diff(&ours[..n], &theirs[..n]); + println!("{name} {channels}ch: {n} samples, largest difference from ffmpeg's decode {d}"); + assert_eq!(d, 0, "{name} {channels}ch"); + } + // our encoder is as good as ffmpeg's on the same audio, to within 3 dB + let mine = match name { + "ima" => ima::decode(&ima::encode(&x, ch, block).expect("encode"), ch, block).expect("decode"), + "ms" => ms::decode(&ms::encode(&x, ch, block).expect("encode"), ch, block).expect("decode"), + _ => yamaha::decode(&yamaha::encode(&x, ch).expect("encode"), ch).expect("decode"), + }; + let m = x.len().min(mine.len()).min(theirs.len()); + let (a, b) = (snr(&x[..m], &mine[..m]), snr(&x[..m], &theirs[..m])); + println!("{name} {channels}ch: SNR {a:.1} dB ours, {b:.1} dB ffmpeg's"); + assert!(a + 3.0 >= b, "{name} {channels}ch: our encoder {a:.1} dB against ffmpeg's {b:.1} dB"); + } + } +} + +#[test] +fn ffmpeg_decodes_what_we_encode_to_the_samples_we_decode() { + let Some(ff) = craft_media_testkit::ffmpeg() else { return }; + let dir = craft_media_testkit::fixture_subdir("adpcm"); + for channels in [1usize, 2] { + let x = signal(channels, 12_000); + let coefs: Vec = ms::COEFFICIENTS.iter().flat_map(|&(a, b)| [a as i16, b as i16].into_iter().flat_map(i16::to_le_bytes)).collect(); + let streams: Vec = { + let e = ima::encode(&x, channels, 512).unwrap(); + let per = ima::samples_per_block(512, channels).unwrap() as u16; + let d = ima::decode(&e, channels, 512).unwrap(); + let d_alt = alt_ima_decode(&e, channels, 512); + let ima_wav = wav(0x11, channels, 512, 4, &per.to_le_bytes(), &e); + let e = ms::encode(&x, channels, 512).unwrap(); + let per = ms::samples_per_block(512, channels).unwrap() as u16; + let d2 = ms::decode(&e, channels, 512).unwrap(); + let mut extra = per.to_le_bytes().to_vec(); + extra.extend(7u16.to_le_bytes()); + extra.extend(&coefs); + let ms_wav = wav(2, channels, 512, 4, &extra, &e); + let e = yamaha::encode(&x, channels).unwrap(); + let d3 = yamaha::decode(&e, channels).unwrap(); + let y_wav = wav(0x20, channels, 512 * channels, 4, &[], &e); + vec![("ima", ima_wav, d, d_alt), ("ms", ms_wav, d2.clone(), d2), ("yamaha", y_wav, d3.clone(), d3)] + }; + for (name, bytes, ours, alt) in streams { + let file = dir.join(format!("our_{name}{channels}.wav")); + std::fs::write(&file, bytes).expect("write"); + let back = dir.join(format!("our_{name}{channels}.pcm")); + ffmpeg(&ff, &["-i", &s(&file), "-f", "s16le", "-acodec", "pcm_s16le", &s(&back)]); + let theirs = read_pcm(&back); + let n = ours.len().min(theirs.len()); + assert!(n + 2048 >= ours.len(), "{name} {channels}ch: ffmpeg decoded {} of {}", theirs.len(), ours.len()); + same_as_either(&format!("{name} {channels}ch, our stream"), &ours, &alt, &theirs); + } + } +} + +/// The audio tags of an FLV file: each tag's body without its flags byte. +fn flv_audio(bytes: &[u8]) -> Vec> { + let mut at = 13; + let mut out = Vec::new(); + while at + 11 <= bytes.len() { + let kind = bytes[at]; + let len = (usize::from(bytes[at + 1]) << 16) | (usize::from(bytes[at + 2]) << 8) | usize::from(bytes[at + 3]); + let body = &bytes[at + 11..(at + 11 + len).min(bytes.len())]; + if kind == 8 && body.len() > 1 { + assert_eq!(body[0] >> 4, 1, "SWF ADPCM tag"); + out.push(body[1..].to_vec()); + } + at += 11 + len + 4; + } + out +} + +/// An FLV file of 44.1 kHz audio tags with these ADPCM bodies. +fn flv(stereo: bool, tags: &[Vec]) -> Vec { + let mut out = b"FLV\x01\x04\0\0\0\x09\0\0\0\0".to_vec(); + let flags = (1u8 << 4) | (3 << 2) | (1 << 1) | u8::from(stereo); + for (i, t) in tags.iter().enumerate() { + let len = t.len() + 1; + let ts = (i * 4096 * 1000 / 44_100) as u32; + out.extend([ + 8, + (len >> 16) as u8, + (len >> 8) as u8, + len as u8, + (ts >> 16) as u8, + (ts >> 8) as u8, + ts as u8, + (ts >> 24) as u8, + 0, + 0, + 0, + flags, + ]); + out.extend(t); + out.extend((len as u32 + 11).to_be_bytes()); + } + out +} + +#[test] +fn flash_adpcm_matches_ffmpeg_in_both_directions() { + let Some(ff) = craft_media_testkit::ffmpeg() else { return }; + let dir = craft_media_testkit::fixture_subdir("adpcm"); + for channels in [1usize, 2] { + let x = signal(channels, 4096 * 3); + let pcm = dir.join(format!("swf_in{channels}.pcm")); + write_pcm(&pcm, &x); + // ffmpeg encodes (it needs 44.1 kHz for FLV) + let file = dir.join(format!("swf{channels}.flv")); + ffmpeg(&ff, &["-f", "s16le", "-ar", "44100", "-ac", &channels.to_string(), "-i", &s(&pcm), "-c:a", "adpcm_swf", "-f", "flv", &s(&file)]); + let tags = flv_audio(&std::fs::read(&file).expect("read")); + assert!(tags.len() >= 3, "{} tags", tags.len()); + let back = dir.join(format!("swf{channels}.out.pcm")); + ffmpeg(&ff, &["-i", &s(&file), "-f", "s16le", "-acodec", "pcm_s16le", &s(&back)]); + let theirs = read_pcm(&back); + let ours: Vec = tags.iter().flat_map(|t| swf::decode(t, channels).expect("decode")).collect(); + assert_eq!(ours.len(), 4096 * 3 * channels); + assert_eq!(theirs.len(), 4096 * 3 * channels); + let alt: Vec = tags.iter().flat_map(|t| alt_swf_decode(t, channels)).collect(); + same_as_either(&format!("swf {channels}ch"), &ours, &alt, &theirs); + // ffmpeg decodes ours, one 4096-sample stream per tag + for bits in [2u32, 3, 4, 5] { + let tags: Vec> = x.chunks(4096 * channels).map(|c| swf::encode(c, channels, bits).expect("encode")).collect(); + let ours: Vec = tags.iter().flat_map(|t| swf::decode(t, channels).expect("decode")).collect(); + let file = dir.join(format!("our_swf{channels}_{bits}.flv")); + std::fs::write(&file, flv(channels == 2, &tags)).expect("write"); + let back = dir.join(format!("our_swf{channels}_{bits}.pcm")); + ffmpeg(&ff, &["-i", &s(&file), "-f", "s16le", "-acodec", "pcm_s16le", &s(&back)]); + let theirs = read_pcm(&back); + assert_eq!(ours.len(), 4096 * 3 * channels); + assert_eq!(theirs.len(), 4096 * 3 * channels); + let alt: Vec = tags.iter().flat_map(|t| alt_swf_decode(t, channels)).collect(); + same_as_either(&format!("swf {channels}ch {bits} bits, our stream"), &ours, &alt, &theirs); + } + } +}