diff --git a/Cargo.lock b/Cargo.lock index add90f8c2..dc6d125f6 100644 --- a/Cargo.lock +++ b/Cargo.lock @@ -1843,6 +1843,13 @@ dependencies = [ "image", ] +[[package]] +name = "craft-gzip" +version = "0.1.0" +dependencies = [ + "flate2", +] + [[package]] name = "craft-hdr" version = "0.1.0" @@ -1919,6 +1926,7 @@ dependencies = [ "avif-serialize", "craft-heif", "craft-jpeg", + "craft-media-testkit", "craft-metadata", "craft-video", "crc32fast", @@ -2383,13 +2391,6 @@ dependencies = [ "thiserror 2.0.21", ] -[[package]] -name = "craft-plist" -version = "0.1.0" -dependencies = [ - "thiserror 2.0.21", -] - [[package]] name = "craft-platform" version = "0.1.0" @@ -2407,6 +2408,13 @@ dependencies = [ "zbus", ] +[[package]] +name = "craft-plist" +version = "0.1.0" +dependencies = [ + "thiserror 2.0.21", +] + [[package]] name = "craft-plotter" version = "0.1.0" @@ -2704,6 +2712,7 @@ dependencies = [ "data-url", "flate2", "kurbo 0.13.1", + "quick-xml 0.41.0", "resvg 0.48.1", "roxmltree 0.21.2", "thiserror 2.0.21", @@ -2720,6 +2729,16 @@ dependencies = [ "kurbo 0.13.1", ] +[[package]] +name = "craft-tcraft" +version = "0.1.0" +dependencies = [ + "craft-zip", + "serde", + "serde_json", + "thiserror 2.0.21", +] + [[package]] name = "craft-text" version = "0.1.0" @@ -2849,6 +2868,7 @@ name = "craft-vectorcraft-format" version = "0.1.0" dependencies = [ "flate2", + "proptest", "serde", "serde_json", "thiserror 2.0.21", diff --git a/Cargo.toml b/Cargo.toml index eae758e17..0c9729e00 100644 --- a/Cargo.toml +++ b/Cargo.toml @@ -38,6 +38,7 @@ hayro-syntax = { version = "=0.7.2", default-features = false, features = ["std" bytemuck = "1" kurbo = "0.13" thiserror = "2" +craft-gzip = { path = "crates/gzip" } craft-aaf = { path = "crates/aaf" } craft-atomic = { path = "crates/atomic" } craft-captions = { path = "crates/captions" } @@ -146,6 +147,7 @@ craft-stencilcraft-document = { path = "crates/stencilcraft-document" } craft-svg-paths = { path = "crates/svg-paths" } craft-stream-codecs = { path = "crates/stream-codecs" } craft-text = { path = "crates/text" } +craft-tcraft = { path = "crates/tcraft" } craft-texture = { path = "crates/texture" } craft-tga = { path = "crates/tga" } craft-timeline = { path = "crates/timeline" } diff --git a/crates/audio/Cargo.toml b/crates/audio/Cargo.toml index c411a0357..645246130 100644 --- a/crates/audio/Cargo.toml +++ b/crates/audio/Cargo.toml @@ -38,6 +38,10 @@ 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 = [] +# QOA (Quite OK Audio) files: decoder and encoder in the `qoa` module, each on its own. +qoa = ["qoa-decode", "qoa-encode"] +qoa-decode = [] +qoa-encode = [] opus-encode = [] ac3-decode = ["craft-media/bitstream"] mpa-decode = ["craft-media/bitstream"] @@ -84,6 +88,10 @@ required-features = ["decode", "lossless-encode", "peaks"] name = "codecs" required-features = ["decode", "encode"] +[[test]] +name = "qoa_oracle" +required-features = ["qoa"] + [[example]] name = "info" required-features = ["decode"] diff --git a/crates/audio/src/lib.rs b/crates/audio/src/lib.rs index 7bbfbbcf5..6b7668be6 100644 --- a/crates/audio/src/lib.rs +++ b/crates/audio/src/lib.rs @@ -27,7 +27,8 @@ //! - Codec modules for packet-level use, each with its own features: [`aac`] (`aac-decode`, //! `aac-encode`), [`alac`] (`alac-decode`, `alac-encode`), [`opus`] (`opus-decode`, and //! [`opus::encoder`] with `opus-encode`), [`ac3`] (`ac3-decode`), [`mpa`] (`mpa-decode`), [`mp3`] -//! (`mp3-decode`, `mp3-encode`), [`flac`] (`flac-encode`: the packet encoder). +//! (`mp3-decode`, `mp3-encode`), [`flac`] (`flac-encode`: the packet encoder), [`qoa`] +//! (`qoa-decode`, `qoa-encode`: whole QOA files). //! //! The shared types below build with no features. @@ -49,6 +50,8 @@ pub mod opus; mod pcm; #[cfg(feature = "peaks")] pub mod peaks; +#[cfg(any(feature = "qoa-decode", feature = "qoa-encode"))] +pub mod qoa; #[cfg(feature = "decode-native")] mod read; #[cfg(any(feature = "aac-decode", feature = "alac-decode", feature = "opus-decode", feature = "ac3-decode", feature = "mpa-decode"))] @@ -70,6 +73,8 @@ mod write; /// PCM WAVE and AIFF wrappers for interchange essence. #[cfg(feature = "pcm-encode")] pub mod essence_pcm; +/// Range-backed PCM WAVE access and a 16-bit PCM writer. +pub mod wav_range; #[cfg(feature = "vorbis-encode")] pub use write::{OGG_BITRATES, encode_ogg}; diff --git a/crates/audio/src/qoa.rs b/crates/audio/src/qoa.rs new file mode 100644 index 000000000..a04430137 --- /dev/null +++ b/crates/audio/src/qoa.rs @@ -0,0 +1,411 @@ +//! QOA ("Quite OK Audio"): a decoder (feature `qoa-decode`) and an encoder (feature `qoa-encode`) +//! for whole `.qoa` files, 16-bit, 1 to 8 channels. +//! +//! Written from the published QOA specification (qoaformat.org). A file is an 8-byte header (the +//! magic `qoaf` and the number of samples per channel) followed by frames of up to 5120 samples +//! per channel. Each frame starts with its channel count, sample rate, sample count and size, then +//! the state of one 4-tap sign-sign LMS predictor per channel, then 64-bit slices of 20 samples: +//! a 4-bit scale factor and twenty 3-bit quantised residuals, the channels interleaved slice by +//! slice. +//! +//! Streaming files (a sample count of 0 in the header) are not supported. +//! Samples are `i16`, one `Vec` per channel. + +#[cfg(feature = "qoa-decode")] +const MAX_FRAME_SAMPLES: usize = 5120; +const SLICE_SAMPLES: usize = 20; +const LMS_LEN: usize = 4; +const MAGIC: &[u8; 4] = b"qoaf"; +/// The most channels a file may have. +pub const MAX_CHANNELS: usize = 8; + +/// `dequant[scale factor][quantised residual]`: `round(round((sf + 1)^2.75) * [0.75, -0.75, 2.5, +/// -2.5, 4.5, -4.5, 7, -7])`. +const DEQUANT: [[i32; 8]; 16] = [ + [1, -1, 3, -3, 5, -5, 7, -7], + [5, -5, 18, -18, 32, -32, 49, -49], + [16, -16, 53, -53, 95, -95, 147, -147], + [34, -34, 113, -113, 203, -203, 315, -315], + [63, -63, 210, -210, 378, -378, 588, -588], + [104, -104, 345, -345, 621, -621, 966, -966], + [158, -158, 528, -528, 950, -950, 1477, -1477], + [228, -228, 760, -760, 1368, -1368, 2128, -2128], + [316, -316, 1053, -1053, 1895, -1895, 2947, -2947], + [422, -422, 1405, -1405, 2529, -2529, 3934, -3934], + [548, -548, 1828, -1828, 3290, -3290, 5117, -5117], + [696, -696, 2320, -2320, 4176, -4176, 6496, -6496], + [868, -868, 2893, -2893, 5207, -5207, 8099, -8099], + [1064, -1064, 3548, -3548, 6386, -6386, 9933, -9933], + [1286, -1286, 4288, -4288, 7718, -7718, 12005, -12005], + [1536, -1536, 5120, -5120, 9216, -9216, 14336, -14336], +]; + +/// What went wrong reading or writing a QOA file. +#[non_exhaustive] +#[derive(Debug, thiserror::Error, PartialEq, Eq)] +pub enum Error { + /// The input is not a QOA file, or is damaged or truncated. + #[error("malformed QOA: {0}")] + Malformed(&'static str), + /// A valid file this module does not handle (a streaming file). + #[error("unsupported QOA: {0}")] + Unsupported(&'static str), + /// The samples or settings given to the encoder cannot be written as QOA. + #[error("QOA cannot encode: {0}")] + InvalidInput(&'static str), +} + +/// A decoded file. +#[derive(Debug, Clone, PartialEq, Eq)] +pub struct Audio { + /// Sample rate in Hz (1 to 16 777 215). + pub sample_rate: u32, + /// One `Vec` of samples per channel, all the same length. + pub channels: Vec>, +} + +/// The predictor: the last four samples and four weights. +#[derive(Clone, Copy)] +struct Lms { + history: [i32; LMS_LEN], + weights: [i32; LMS_LEN], +} + +impl Lms { + fn new() -> Self { + Lms { history: [0; LMS_LEN], weights: [0, 0, -(1 << 13), 1 << 14] } + } + + fn predict(&self) -> i32 { + let mut p = 0i32; + for i in 0..LMS_LEN { + p = p.wrapping_add(self.weights[i].wrapping_mul(self.history[i])); + } + p >> 13 + } + + fn update(&mut self, sample: i32, residual: i32) { + let delta = residual >> 4; + for i in 0..LMS_LEN { + self.weights[i] += if self.history[i] < 0 { -delta } else { delta }; + } + self.history.copy_within(1.., 0); + self.history[LMS_LEN - 1] = sample; + } +} + +fn clamp16(v: i32) -> i32 { + v.clamp(-32768, 32767) +} + +#[cfg(feature = "qoa-decode")] +mod decode { + use super::{Audio, DEQUANT, Error, LMS_LEN, Lms, MAGIC, MAX_CHANNELS, MAX_FRAME_SAMPLES, SLICE_SAMPLES, clamp16}; + + fn be16(b: &[u8], i: usize) -> u16 { + u16::from_be_bytes([b[i], b[i + 1]]) + } + + /// Decode a whole `.qoa` file. + /// + /// The declared length is checked against the input before anything is allocated: a slice of + /// 8 bytes holds at most 20 samples, so a file cannot declare more samples than that. + pub fn decode(bytes: &[u8]) -> Result { + if bytes.len() < 8 || &bytes[..4] != MAGIC { + return Err(Error::Malformed("missing the qoaf magic")); + } + let total = u32::from_be_bytes([bytes[4], bytes[5], bytes[6], bytes[7]]) as usize; + if total == 0 { + return Err(Error::Unsupported("streaming files (a sample count of 0)")); + } + if bytes.len() < 16 { + return Err(Error::Malformed("truncated frame header")); + } + let channels = usize::from(bytes[8]); + let sample_rate = u32::from_be_bytes([0, bytes[9], bytes[10], bytes[11]]); + if channels == 0 || channels > MAX_CHANNELS || sample_rate == 0 { + return Err(Error::Malformed("bad channel count or sample rate")); + } + // 8 bytes (20 samples) per channel at the very least, plus the frame headers. + let room = (bytes.len() - 8) / (8 * channels) * SLICE_SAMPLES; + if total > room { + return Err(Error::Malformed("the file is too short for its declared length")); + } + let mut out: Vec> = Vec::new(); + for _ in 0..channels { + let mut v = Vec::new(); + v.try_reserve_exact(total).map_err(|_| Error::Malformed("too many samples"))?; + out.push(v); + } + let mut pos = 8; + let mut done = 0usize; + while done < total { + let head = bytes.get(pos..pos + 8).ok_or(Error::Malformed("truncated frame header"))?; + let rate = u32::from_be_bytes([0, head[1], head[2], head[3]]); + if usize::from(head[0]) != channels || rate != sample_rate { + return Err(Error::Malformed("the channel count or sample rate changes between frames")); + } + let samples = usize::from(be16(head, 4)); + let size = usize::from(be16(head, 6)); + let slices = samples.div_ceil(SLICE_SAMPLES); + if samples == 0 || samples > MAX_FRAME_SAMPLES || samples > total - done { + return Err(Error::Malformed("bad frame sample count")); + } + if size != 8 + 16 * channels + 8 * slices * channels { + return Err(Error::Malformed("frame size does not match its sample count")); + } + let frame = bytes.get(pos..pos + size).ok_or(Error::Malformed("truncated frame"))?; + let mut lms = [Lms::new(); MAX_CHANNELS]; + for (c, state) in lms.iter_mut().take(channels).enumerate() { + let at = 8 + c * 16; + for i in 0..LMS_LEN { + state.history[i] = i32::from(be16(frame, at + 2 * i) as i16); + state.weights[i] = i32::from(be16(frame, at + 8 + 2 * i) as i16); + } + } + let mut at = 8 + 16 * channels; + for s in 0..slices { + let n = (samples - s * SLICE_SAMPLES).min(SLICE_SAMPLES); + for (c, state) in lms.iter_mut().take(channels).enumerate() { + let mut slice = u64::from_be_bytes([ + frame[at], + frame[at + 1], + frame[at + 2], + frame[at + 3], + frame[at + 4], + frame[at + 5], + frame[at + 6], + frame[at + 7], + ]); + at += 8; + let table = &DEQUANT[(slice >> 60) as usize & 15]; + for _ in 0..n { + let dequantised = table[(slice >> 57) as usize & 7]; + slice <<= 3; + let sample = clamp16(state.predict() + dequantised); + out[c].push(sample as i16); + state.update(sample, dequantised); + } + } + } + pos += size; + done += samples; + } + Ok(Audio { sample_rate, channels: out }) + } +} +#[cfg(feature = "qoa-decode")] +pub use decode::decode; + +#[cfg(feature = "qoa-encode")] +mod encode { + use super::{DEQUANT, Error, Lms, MAGIC, MAX_CHANNELS, SLICE_SAMPLES, clamp16}; + + /// The quantised residual (0..8) whose dequantised value is nearest `residual`. + fn quantise(table: &[i32; 8], residual: i32) -> usize { + let mut best = 0; + let mut best_err = i64::MAX; + for (q, &d) in table.iter().enumerate() { + let e = i64::from(residual - d).pow(2); + if e < best_err { + best = q; + best_err = e; + } + } + best + } + + /// Encode one slice of one channel with every scale factor and keep the one with the least + /// squared error. Returns the 64-bit slice; `lms` is advanced to the chosen result. + fn encode_slice(lms: &mut Lms, samples: &[i16]) -> u64 { + let mut best: Option<(i64, u64, Lms)> = None; + for (sf, table) in DEQUANT.iter().enumerate() { + let mut trial = *lms; + let mut slice = sf as u64; + let mut err = 0i64; + for k in 0..SLICE_SAMPLES { + let mut q = 0; + if let Some(&s) = samples.get(k) { + let predicted = trial.predict(); + q = quantise(table, i32::from(s) - predicted); + let sample = clamp16(predicted + table[q]); + err += i64::from(i32::from(s) - sample).pow(2); + trial.update(sample, table[q]); + } + slice = slice << 3 | q as u64; + } + // 4 + 60 bits are in `slice`; the scale factor belongs on top. + let slice = slice & ((1 << 60) - 1) | (sf as u64) << 60; + if best.as_ref().is_none_or(|b| err < b.0) { + best = Some((err, slice, trial)); + } + } + match best { + Some((_, slice, state)) => { + *lms = state; + slice + } + None => 0, + } + } + + /// Encode `channels` (one `Vec` of samples per channel, all the same length) as a `.qoa` file. + pub fn encode(sample_rate: u32, channels: &[Vec]) -> Result, Error> { + let n = channels.len(); + if n == 0 || n > MAX_CHANNELS { + return Err(Error::InvalidInput("QOA stores 1 to 8 channels")); + } + if sample_rate == 0 || sample_rate > 0x00ff_ffff { + return Err(Error::InvalidInput("the sample rate must be 1 to 16777215 Hz")); + } + let total = channels[0].len(); + if channels.iter().any(|c| c.len() != total) { + return Err(Error::InvalidInput("the channels differ in length")); + } + let total32 = u32::try_from(total).map_err(|_| Error::InvalidInput("too many samples"))?; + if total == 0 { + return Err(Error::InvalidInput("QOA cannot store an empty file")); + } + let mut out = Vec::with_capacity(8 + total * n / 2); + out.extend_from_slice(MAGIC); + out.extend_from_slice(&total32.to_be_bytes()); + let mut lms = [Lms::new(); MAX_CHANNELS]; + let mut start = 0; + while start < total { + let len = (total - start).min(5120); + let slices = len.div_ceil(SLICE_SAMPLES); + let size = 8 + 16 * n + 8 * slices * n; + out.push(n as u8); + out.extend_from_slice(&sample_rate.to_be_bytes()[1..]); + out.extend_from_slice(&(len as u16).to_be_bytes()); + out.extend_from_slice(&(size as u16).to_be_bytes()); + // The state a frame starts from is written as 16-bit values, so the predictor + // continues from exactly what a decoder reads. + for state in lms.iter_mut().take(n) { + for v in state.history.iter_mut().chain(state.weights.iter_mut()) { + *v = i32::from(*v as i16); + out.extend_from_slice(&(*v as i16).to_be_bytes()); + } + } + for s in 0..slices { + let from = start + s * SLICE_SAMPLES; + let to = (from + SLICE_SAMPLES).min(start + len); + for (c, state) in lms.iter_mut().take(n).enumerate() { + out.extend_from_slice(&encode_slice(state, &channels[c][from..to]).to_be_bytes()); + } + } + start += len; + } + Ok(out) + } +} +#[cfg(feature = "qoa-encode")] +pub use encode::encode; + +#[cfg(all(test, feature = "qoa-decode", feature = "qoa-encode"))] +mod tests { + use super::*; + + fn tone(frames: usize, ch: usize) -> Vec> { + (0..ch) + .map(|c| (0..frames).map(|i| (((i as f64) * (0.03 + 0.01 * c as f64)).sin() * 12000.0 + ((i * 7919) % 200) as f64) as i16).collect()) + .collect() + } + + fn snr(a: &[i16], b: &[i16]) -> f64 { + let (mut s, mut e) = (0f64, 0f64); + for (&x, &y) in a.iter().zip(b) { + s += f64::from(x).powi(2); + e += (f64::from(x) - f64::from(y)).powi(2); + } + 10.0 * (s / e.max(1.0)).log10() + } + + #[test] + fn dequant_table_matches_its_formula() { + for (sf, row) in DEQUANT.iter().enumerate() { + let scale = ((sf as f64 + 1.0).powf(2.75)).round(); + for (q, base) in [0.75, -0.75, 2.5, -2.5, 4.5, -4.5, 7.0, -7.0].into_iter().enumerate() { + assert_eq!(f64::from(row[q]), (scale * base).round(), "sf {sf} q {q}"); + } + } + } + + #[test] + fn round_trips_with_a_good_signal_to_noise_ratio() { + for (frames, ch) in [(1, 1), (19, 2), (20, 1), (21, 2), (5120, 1), (5121, 2), (12_345, 8)] { + let src = tone(frames, ch); + let bytes = encode(44_100, &src).unwrap(); + let back = decode(&bytes).unwrap(); + assert_eq!((back.sample_rate, back.channels.len(), back.channels[0].len()), (44_100, ch, frames)); + if frames > 1000 { + for (c, (a, b)) in src.iter().zip(&back.channels).enumerate() { + assert!(snr(a, b) > 25.0, "{frames}x{ch} channel {c}"); + } + } + } + } + + #[test] + fn silence_and_full_scale_extremes_survive() { + let silent = vec![vec![0i16; 3000]]; + // No residual value is zero, so silence decodes to a tiny residue, never exactly 0. + let quiet = decode(&encode(8000, &silent).unwrap()).unwrap(); + assert!(quiet.channels[0].iter().all(|&s| s.abs() <= 2), "silence stays within 2 LSB"); + let square: Vec = (0..4000).map(|i| if (i / 50) % 2 == 0 { i16::MAX } else { i16::MIN }).collect(); + let back = decode(&encode(48_000, std::slice::from_ref(&square)).unwrap()).unwrap(); + assert_eq!(back.channels[0].len(), square.len()); + } + + #[test] + fn the_file_has_the_documented_layout() { + let bytes = encode(48_000, &tone(100, 2)).unwrap(); + assert_eq!(&bytes[..4], b"qoaf"); + assert_eq!(u32::from_be_bytes(bytes[4..8].try_into().unwrap()), 100); + assert_eq!(bytes[8], 2); + assert_eq!(&bytes[9..12], &48_000u32.to_be_bytes()[1..]); + assert_eq!(u16::from_be_bytes([bytes[12], bytes[13]]), 100); + // 8 header + 2 channels * 16 state + 5 slices * 2 channels * 8 + assert_eq!(usize::from(u16::from_be_bytes([bytes[14], bytes[15]])), 8 + 32 + 80); + assert_eq!(bytes.len(), 8 + 8 + 32 + 80); + // The first frame starts from the specified predictor state. + assert_eq!(&bytes[16 + 8..16 + 16], &[0, 0, 0, 0, 0xe0, 0, 0x40, 0]); + } + + #[test] + fn invalid_input_to_the_encoder_is_refused() { + assert!(encode(44_100, &[]).is_err()); + assert!(encode(0, &[vec![0; 10]]).is_err()); + assert!(encode(1 << 24, &[vec![0; 10]]).is_err()); + assert!(encode(44_100, &[vec![0; 10], vec![0; 9]]).is_err()); + assert!(encode(44_100, &[vec![]]).is_err()); + assert!(encode(44_100, &vec![vec![0; 10]; 9]).is_err()); + } + + #[test] + fn hostile_input_is_an_error_not_a_panic() { + let good = encode(44_100, &tone(6000, 2)).unwrap(); + for len in 0..good.len().min(600) { + let _ = decode(&good[..len]); + } + assert!(decode(&good[..good.len() - 1]).is_err()); + // A header that declares more samples than the bytes can hold. + let mut huge = good.clone(); + huge[4..8].copy_from_slice(&u32::MAX.to_be_bytes()); + assert!(decode(&huge).is_err()); + // A streaming file, and a frame whose size field lies. + let mut streaming = good.clone(); + streaming[4..8].fill(0); + assert_eq!(decode(&streaming), Err(Error::Unsupported("streaming files (a sample count of 0)"))); + let mut lying = good.clone(); + lying[14] ^= 1; + assert!(decode(&lying).is_err()); + let mut zero_rate = good.clone(); + zero_rate[9..12].fill(0); + assert!(decode(&zero_rate).is_err()); + let mut changed = good; + let second = 8 + usize::from(u16::from_be_bytes([changed[14], changed[15]])); + changed[second] = 1; + assert!(decode(&changed).is_err()); + } +} diff --git a/crates/audio/src/wav_range.rs b/crates/audio/src/wav_range.rs new file mode 100644 index 000000000..33762caf2 --- /dev/null +++ b/crates/audio/src/wav_range.rs @@ -0,0 +1,223 @@ +//! Range-backed PCM WAVE reader and canonical 16-bit writer. +//! +//! Reads chunk headers and only the requested sample range; it never loads a whole file. + +use crate::{AudioError, Result}; +use craft_media::ByteSource; + +/// Whether a prefix identifies a RIFF or RF64 WAVE file. +pub fn sniff(b: &[u8]) -> bool { + b.len() >= 12 && (&b[0..4] == b"RIFF" || &b[0..4] == b"RF64") && &b[8..12] == b"WAVE" +} + +const MAX_CHUNKS: usize = 1 << 16; +const IXML_MAX: u64 = 1 << 20; +/// Maximum sample data returned by a single range read. +pub const MAX_RANGE_BYTES: u64 = 4 << 20; + +fn le16(b: &[u8], o: usize) -> u16 { + b.get(o..o.saturating_add(2)).and_then(|s| s.try_into().ok()).map_or(0, u16::from_le_bytes) +} +fn le32(b: &[u8], o: usize) -> u32 { + b.get(o..o.saturating_add(4)).and_then(|s| s.try_into().ok()).map_or(0, u32::from_le_bytes) +} +fn read_range(source: &impl ByteSource, off: u64, len: usize) -> Result> { + let mut bytes = vec![0; len]; + source.read_at(off, &mut bytes)?; + Ok(bytes) +} + +/// A PCM WAVE file held by an arbitrary random-access source. +pub struct WavRange { + source: S, + /// Number of complete sample frames. + pub frames: u64, + /// Channel count. + pub channels: u16, + /// Sample rate in Hz. + pub rate: u32, + /// Container bits per sample. + pub bits: u16, + /// Whether samples are floating point. + pub float: bool, + /// BWF time reference in samples, from bext or iXML. + pub time_reference: Option, + data_off: u64, +} + +impl WavRange { + /// Open with caller-selected bounds on channel count and sample rate. + pub fn open(source: S, max_channels: u32, max_rate: u32) -> Result { + let len = source.len(); + let mut pos: u64 = 12; + let mut fmt = None; + let mut data = None; + let mut bext_tr = None; + let mut ixml_tr = None; + let mut ds64_data = None; + for _ in 0..MAX_CHUNKS { + let Some(body) = pos.checked_add(8).filter(|&b| b <= len) else { break }; + let mut hdr = [0u8; 8]; + source.read_at(pos, &mut hdr)?; + let id = &hdr[0..4]; + let mut clen = u64::from(le32(&hdr, 4)); + let rest = len - body; + if id == b"data" { + if clen == 0xFFFF_FFFF { + clen = ds64_data.unwrap_or(rest); + } + clen = clen.min(rest); + } + if clen > rest { + break; + } + match id { + b"fmt " if clen >= 16 => { + let b = read_range(&source, body, clen.min(40) as usize)?; + let mut tag = le16(&b, 0); + if tag == 0xFFFE && clen >= 40 { + tag = le16(&b, 24); + } + fmt = Some((tag, le16(&b, 2), le32(&b, 4), le16(&b, 14))); + } + b"ds64" if clen >= 16 => { + let b = read_range(&source, body + 8, 8)?; + ds64_data = Some(u64::from(le32(&b, 0)) | u64::from(le32(&b, 4)) << 32); + } + b"data" => data = Some((body, clen)), + b"bext" if clen >= 346 => { + let b = read_range(&source, body + 338, 8)?; + bext_tr = Some(u64::from(le32(&b, 0)) | u64::from(le32(&b, 4)) << 32); + } + b"iXML" => { + let b = read_range(&source, body, clen.min(IXML_MAX) as usize)?; + ixml_tr = ixml_time_reference(&b); + } + _ => {} + } + pos = body.saturating_add(clen).saturating_add(clen & 1); + } + let (tag, channels, rate, bits) = fmt.ok_or_else(|| AudioError::Malformed("missing fmt chunk".into()))?; + let (data_off, dlen) = data.ok_or_else(|| AudioError::Malformed("missing data chunk".into()))?; + let float = tag == 3; + if !(tag == 1 || float) + || channels == 0 + || rate == 0 + || u32::from(channels) > max_channels + || rate > max_rate + || !matches!(bits, 8 | 16 | 24 | 32 | 64) + { + return Err(AudioError::Unsupported(format!("WAV format tag {tag}, {bits} bits"))); + } + let frames = dlen / (u64::from(channels) * u64::from(bits / 8)); + Ok(Self { source, frames, channels, rate, bits, float, time_reference: bext_tr.or(ixml_tr), data_off }) + } + /// Bytes in one whole frame. + pub fn frame_bytes(&self) -> usize { + usize::from(self.channels) * usize::from(self.bits / 8) + } + /// Read the inclusive range of frames, capped to prevent unbounded allocation. + pub fn read_frames(&self, first: u64, last: u64) -> Result> { + if first > last || last >= self.frames { + return Err(AudioError::Malformed("WAV frame range outside data".into())); + } + let n = (last - first + 1).saturating_mul(self.frame_bytes() as u64); + if n > MAX_RANGE_BYTES { + return Err(AudioError::TooLarge("WAV frame range".into())); + } + let off = self.data_off.saturating_add(first.saturating_mul(self.frame_bytes() as u64)); + read_range(&self.source, off, n as usize) + } + /// Decode a channel from a buffer returned by `read_frames`. + pub fn sample(&self, block: &[u8], frame: usize, ch: usize) -> f32 { + if ch >= usize::from(self.channels) { + return 0.0; + } + let bps = usize::from(self.bits / 8); + let Some(o) = frame.checked_mul(usize::from(self.channels)).and_then(|v| v.checked_add(ch)).and_then(|v| v.checked_mul(bps)) else { + return 0.0; + }; + let Some(b) = block.get(o..o.saturating_add(bps)) else { return 0.0 }; + match (self.bits, self.float, b) { + (8, _, &[a]) => (a as f32 - 128.0) / 128.0, + (16, _, &[a, b]) => i16::from_le_bytes([a, b]) as f32 / 32768.0, + (24, _, &[a, b, c]) => (i32::from_le_bytes([0, a, b, c]) >> 8) as f32 / 8_388_608.0, + (32, false, &[a, b, c, d]) => i32::from_le_bytes([a, b, c, d]) as f32 / 2_147_483_648.0, + (32, true, &[a, b, c, d]) => f32::from_le_bytes([a, b, c, d]), + (64, true, b) => f64::from_le_bytes(b.try_into().unwrap_or([0; 8])) as f32, + _ => 0.0, + } + } +} + +fn ixml_time_reference(x: &[u8]) -> Option { + let text = String::from_utf8_lossy(x); + let field = |tag: &str| -> Option { + let open = format!("<{tag}>"); + let a = text.find(&open)? + open.len(); + let b = a + text[a..].find('<')?; + text[a..b].trim().parse().ok() + }; + let low = field("BWF_TIME_REFERENCE_LOW")?; + Some(low | field("BWF_TIME_REFERENCE_HIGH").unwrap_or(0) << 32) +} + +/// Write interleaved samples as 16-bit PCM WAVE, using symmetric 32767 scaling. +#[cfg(feature = "pcm-encode")] +pub fn write_wav16(samples: &[f32], channels: u16, rate: u32) -> Vec { + let data_len = samples.len().saturating_mul(2); + let mut v = crate::pcm_wav_header(u32::try_from(data_len).unwrap_or(u32::MAX), channels, rate, 16); + v.reserve(data_len); + for s in samples { + v.extend_from_slice(&((s.clamp(-1.0, 1.0) * 32767.0).round() as i16).to_le_bytes()); + } + v +} + +#[cfg(all(test, feature = "pcm-encode"))] +mod tests { + use super::*; + use std::sync::{Arc, Mutex}; + + struct Source { + bytes: Vec, + reads: Arc>>, + } + impl ByteSource for Source { + fn len(&self) -> u64 { + self.bytes.len() as u64 + } + fn read_at(&self, offset: u64, buf: &mut [u8]) -> std::io::Result<()> { + self.reads.lock().unwrap_or_else(std::sync::PoisonError::into_inner).push((offset, buf.len())); + self.bytes.read_at(offset, buf) + } + } + #[test] + fn bounded_reads_and_pcm_values() { + let bytes = write_wav16(&[-1.0, -0.5, 0.0, 0.5, 1.0], 1, 48_000); + let reads = Arc::new(Mutex::new(Vec::new())); + let wav = WavRange::open(Source { bytes, reads: reads.clone() }, 64, 384_000).unwrap(); + assert_eq!((wav.frames, wav.channels, wav.rate, wav.bits), (5, 1, 48_000, 16)); + assert!(reads.lock().unwrap().iter().all(|(_, n)| *n <= 40)); + let data = wav.read_frames(1, 3).unwrap(); + assert_eq!(wav.sample(&data, 0, 0), -0.5); + assert_eq!(wav.sample(&data, 2, 0), 0.5); + assert!(wav.read_frames(0, 5).is_err()); + assert!(wav.read_frames(3, 1).is_err()); + assert_eq!(reads.lock().unwrap().last(), Some(&(46, 6))); + } + #[test] + fn ixml_fallback_accepts_flat_fields() { + let base = write_wav16(&[0.0], 1, 48_000); + let mut bytes = base[..36].to_vec(); + let xml = b"51"; + bytes.extend_from_slice(b"iXML"); + bytes.extend_from_slice(&(xml.len() as u32).to_le_bytes()); + bytes.extend_from_slice(xml); + if xml.len() & 1 != 0 { + bytes.push(0); + } + bytes.extend_from_slice(&base[36..]); + assert_eq!(WavRange::open(bytes.as_slice(), 64, 384_000).unwrap().time_reference, Some((1 << 32) + 5)); + } +} diff --git a/crates/audio/tests/qoa_oracle.rs b/crates/audio/tests/qoa_oracle.rs new file mode 100644 index 000000000..7b5ac2674 --- /dev/null +++ b/crates/audio/tests/qoa_oracle.rs @@ -0,0 +1,62 @@ +//! QOA against ffmpeg's decoder as the reference (oracle; skipped when ffmpeg is absent, a failure +//! with `CRAFT_MEDIA_REQUIRE_ORACLES=1`): ffmpeg decodes the files our encoder writes to exactly +//! the samples our decoder returns for them, over several sizes (partial slices, partial frames, +//! several frames), channel counts and rates. + +use craft_audio::qoa::{decode, encode}; +use std::process::Command; + +fn signal(channels: usize, frames: usize) -> Vec> { + let mut seed = 0x2545_F491u32; + (0..channels) + .map(|c| { + (0..frames) + .map(|i| { + seed ^= seed << 13; + seed ^= seed >> 17; + seed ^= seed << 5; + let tone = ((i as f64) * (0.02 + 0.013 * c as f64)).sin() * 9000.0; + // Tones, noise bursts, silence and full-scale peaks. + match (i / 700) % 4 { + 0 => tone as i16, + 1 => (tone + f64::from(seed % 6000) - 3000.0) as i16, + 2 => 0, + _ => { + if i % 2 == 0 { + i16::MAX + } else { + i16::MIN + } + } + } + }) + .collect() + }) + .collect() +} + +#[test] +fn ffmpeg_decodes_our_files_to_the_same_samples() { + let Some(ffmpeg) = craft_media_testkit::ffmpeg() else { return }; + let dir = craft_media_testkit::fixture_subdir("qoa"); + for (rate, channels, frames) in + [(44_100u32, 1usize, 1usize), (8_000, 1, 19), (48_000, 2, 5_121), (22_050, 2, 12_345), (96_000, 1, 20_000), (44_100, 4, 6_000)] + { + let src = signal(channels, frames); + let bytes = encode(rate, &src).expect("encode"); + let ours = decode(&bytes).expect("decode"); + let file = dir.join(format!("ours_{rate}_{channels}_{frames}.qoa")); + std::fs::write(&file, &bytes).expect("write fixture"); + let out = Command::new(&ffmpeg) + .args(["-v", "error", "-y", "-i"]) + .arg(&file) + .args(["-f", "s16le", "-acodec", "pcm_s16le", "-"]) + .output() + .expect("run ffmpeg"); + assert!(out.status.success(), "ffmpeg failed on {}: {}", file.display(), String::from_utf8_lossy(&out.stderr)); + let theirs: Vec = out.stdout.as_chunks::<2>().0.iter().map(|b| i16::from_le_bytes(*b)).collect(); + let interleaved: Vec = (0..frames).flat_map(|i| ours.channels.iter().map(move |ch| ch[i])).collect(); + assert_eq!(theirs.len(), interleaved.len(), "{rate} Hz, {channels} ch, {frames} frames: sample count"); + assert!(theirs == interleaved, "{rate} Hz, {channels} ch, {frames} frames: samples differ"); + } +} diff --git a/crates/cms/src/builtin.rs b/crates/cms/src/builtin.rs index ca7034286..5a8b03228 100644 --- a/crates/cms/src/builtin.rs +++ b/crates/cms/src/builtin.rs @@ -424,6 +424,18 @@ fn shipped_cmyk(bytes: &[u8], b: Builtin) -> Profile { }) } +/// Generates the compact ICC v2.1 sRGB matrix/TRC profile used by PDFCraft. +/// +/// Unlike [`Builtin::Srgb`]'s v4 profile, this uses a 1024-sample tone curve and +/// v2 `desc`/`text` tags for consumers that require a v2 output intent. The legacy +/// description `sRGB (PdfCraft)` is retained to preserve existing document bytes. +/// The generated profile contains no external assets; callers choose where it is +/// appropriate (including their own PDF/A conformance policy). +/// All sizes are fixed; no input-controlled allocation or recursion is involved. +pub fn srgb_v2() -> Vec { + crate::builtin_v2::srgb() +} + #[cfg(test)] mod tests { use super::*; diff --git a/crates/cms/src/builtin_v2.rs b/crates/cms/src/builtin_v2.rs new file mode 100644 index 000000000..4a9e2e25b --- /dev/null +++ b/crates/cms/src/builtin_v2.rs @@ -0,0 +1,143 @@ +//! A small sRGB ICC profile (version 2, display class, matrix/TRC), written from the published +//! sRGB primaries (IEC 61966-2-1) adapted to D50, with the sRGB tone curve sampled at 1024 +//! points. Used as the PDF/A output intent's destination profile. + +fn s15f16(v: f64) -> [u8; 4] { + ((v * 65536.0).round() as i32).to_be_bytes() +} + +fn xyz(x: f64, y: f64, z: f64) -> Vec { + let mut t = b"XYZ \0\0\0\0".to_vec(); + t.extend(s15f16(x)); + t.extend(s15f16(y)); + t.extend(s15f16(z)); + t +} + +fn desc(text: &str) -> Vec { + let mut t = b"desc\0\0\0\0".to_vec(); + let ascii: Vec = text.bytes().chain(std::iter::once(0)).collect(); + t.extend((ascii.len() as u32).to_be_bytes()); + t.extend(&ascii); + // No Unicode or ScriptCode descriptions. + t.extend([0u8; 4 + 4]); + t.extend([0u8; 2 + 1 + 67]); + t +} + +fn text(text: &str) -> Vec { + let mut t = b"text\0\0\0\0".to_vec(); + t.extend(text.bytes()); + t.push(0); + t +} + +fn curve() -> Vec { + const N: usize = 1024; + let mut t = b"curv\0\0\0\0".to_vec(); + t.extend((N as u32).to_be_bytes()); + for i in 0..N { + let v = i as f64 / (N - 1) as f64; + let lin = if v <= 0.04045 { v / 12.92 } else { ((v + 0.055) / 1.055).powf(2.4) }; + t.extend(((lin * 65535.0).round() as u16).to_be_bytes()); + } + t +} + +/// Fixed-size profile bytes; all tag sizes and loop counts are constants. +pub fn srgb() -> Vec { + // sRGB primaries, Bradford-adapted to the D50 profile connection space. + let tags: Vec<(&[u8; 4], Vec)> = vec![ + (b"desc", desc("sRGB (PdfCraft)")), + (b"cprt", text("No copyright, use freely")), + (b"wtpt", xyz(0.9642, 1.0, 0.8249)), + (b"rXYZ", xyz(0.4361, 0.2225, 0.0139)), + (b"gXYZ", xyz(0.3851, 0.7169, 0.0971)), + (b"bXYZ", xyz(0.1431, 0.0606, 0.7141)), + (b"rTRC", curve()), + ]; + // gTRC and bTRC share rTRC's data. + let count = tags.len() + 2; + let table_len = 4 + count * 12; + let mut offset = 128 + table_len; + let mut table = (count as u32).to_be_bytes().to_vec(); + let mut data = Vec::new(); + let mut trc = (0, 0); + for (sig, body) in &tags { + let len = body.len(); + table.extend(*sig); + table.extend((offset as u32).to_be_bytes()); + table.extend((len as u32).to_be_bytes()); + if *sig == b"rTRC" { + trc = (offset, len); + } + data.extend(body); + // Tags start on 4-byte boundaries. + let pad = (4 - len % 4) % 4; + data.extend(std::iter::repeat_n(0u8, pad)); + offset += len + pad; + } + for sig in [b"gTRC", b"bTRC"] { + table.extend(sig); + table.extend((trc.0 as u32).to_be_bytes()); + table.extend((trc.1 as u32).to_be_bytes()); + } + let size = 128 + table.len() + data.len(); + let mut h = Vec::with_capacity(size); + h.extend((size as u32).to_be_bytes()); + h.extend(b"\0\0\0\0"); // preferred CMM + h.extend([2, 0x10, 0, 0]); // version 2.1 + h.extend(b"mntr"); + h.extend(b"RGB "); + h.extend(b"XYZ "); + h.extend([0u8; 12]); // date + h.extend(b"acsp"); + h.extend(b"\0\0\0\0"); // platform + h.extend([0u8; 4]); // flags + h.extend([0u8; 4]); // manufacturer + h.extend([0u8; 4]); // model + h.extend([0u8; 8]); // attributes + h.extend([0, 0, 0, 0]); // perceptual intent + h.extend(s15f16(0.9642)); + h.extend(s15f16(1.0)); + h.extend(s15f16(0.8249)); + h.extend([0u8; 4]); // creator + h.resize(128, 0); + h.extend(table); + h.extend(data); + h +} + +#[cfg(test)] +mod tests { + #[test] + fn parses_as_v2_with_shared_rgb_curves() { + let bytes = super::srgb(); + let p = crate::Profile::parse(&bytes).unwrap(); + assert_eq!(p.version, (2, 0x10)); + assert_eq!(p.color_space, crate::ColorSpace::Rgb); + assert_eq!(p.pcs, crate::Pcs::Xyz); + assert_eq!(p.description, "sRGB (PdfCraft)"); + let curves = p.trc.unwrap(); + assert_eq!(curves[0], curves[1]); + assert_eq!(curves[1], curves[2]); + assert!(matches!(&curves[0], crate::Curve::Table(v) if v.len() == 1024 && v[0] == 0.0 && v[1023] == 1.0)); + assert_eq!(bytes, super::srgb()); + } + + #[test] + fn profile_is_well_formed() { + let p = super::srgb(); + assert_eq!(u32::from_be_bytes(p[0..4].try_into().unwrap()) as usize, p.len()); + assert_eq!(&p[36..40], b"acsp"); + assert_eq!(&p[12..20], b"mntrRGB "); + let n = u32::from_be_bytes(p[128..132].try_into().unwrap()) as usize; + assert_eq!(n, 9); + for i in 0..n { + let e = 132 + i * 12; + let off = u32::from_be_bytes(p[e + 4..e + 8].try_into().unwrap()) as usize; + let len = u32::from_be_bytes(p[e + 8..e + 12].try_into().unwrap()) as usize; + assert!(off.is_multiple_of(4) && off + len <= p.len(), "tag {i}"); + } + } +} diff --git a/crates/cms/src/lib.rs b/crates/cms/src/lib.rs index 8d4f99f49..5e8a1e268 100644 --- a/crates/cms/src/lib.rs +++ b/crates/cms/src/lib.rs @@ -20,6 +20,7 @@ #![forbid(unsafe_code)] pub mod builtin; +mod builtin_v2; pub mod clut; pub mod curve; pub mod gamut; diff --git a/crates/dxf/src/lib.rs b/crates/dxf/src/lib.rs index 4cf3e46e9..9c912613b 100644 --- a/crates/dxf/src/lib.rs +++ b/crates/dxf/src/lib.rs @@ -11,6 +11,7 @@ use std::io::Write as _; pub mod model; +pub mod mtext; pub mod paths; mod version; pub use version::Version; diff --git a/crates/dxf/src/mtext.rs b/crates/dxf/src/mtext.rs new file mode 100644 index 000000000..9b113aa6c --- /dev/null +++ b/crates/dxf/src/mtext.rs @@ -0,0 +1,103 @@ +//! MTEXT inline-escape handling independent of drawing and text layout. +//! +//! Unicode and MIF escapes are decoded to literal characters, with markup-significant +//! characters re-escaped so subsequent MTEXT parsers cannot treat them as commands. + +/// Whether an inline command takes an argument terminated by `;`. +pub fn takes_arg(code: char) -> bool { + matches!(code, 'f' | 'F' | 'H' | 'W' | 'Q' | 'T' | 'A' | 'C' | 'c' | 'p') +} + +/// Four hex digits at `chars[at..at + 4]`. +fn hex4(chars: &[char], at: usize) -> Option { + let digits = chars.get(at..at.checked_add(4)?)?; + digits.iter().try_fold(0u32, |acc, c| Some(acc * 16 + c.to_digit(16)?)) +} + +/// Push a decoded character so that the tokenizers read it as text, never as markup. +fn push_literal(out: &mut String, c: char) { + if matches!(c, '\\' | '{' | '}') { + out.push('\\'); + } + out.push(c); +} + +/// Decode `\U+hhhh` (one UTF-16 unit; a surrogate pair gives one character, a lone surrogate +/// U+FFFD) and `\M+nhhhh` (double-byte East Asian characters) to the characters they stand +/// for. Malformed escapes remain unchanged for the caller to interpret. +/// An escaped backslash (`\\U+0041`) is not an escape. +pub fn unescape_codes(s: &str) -> std::borrow::Cow<'_, str> { + if !s.contains("\\U+") && !s.contains("\\M+") { + return std::borrow::Cow::Borrowed(s); + } + let chars: Vec = s.chars().collect(); + let mut out = String::with_capacity(s.len()); + let mut i = 0; + while let Some(&c) = chars.get(i) { + if c != '\\' { + out.push(c); + i += 1; + continue; + } + let (n, plus) = (chars.get(i + 1).copied(), chars.get(i + 2).copied()); + if n == Some('U') && plus == Some('+') { + if let Some(unit) = hex4(&chars, i + 3) { + i += 7; + let low = |at: usize| { + (chars.get(at) == Some(&'\\') && chars.get(at + 1) == Some(&'U') && chars.get(at + 2) == Some(&'+')) + .then(|| hex4(&chars, at + 3)) + .flatten() + .filter(|u| (0xDC00..0xE000).contains(u)) + }; + let ch = if (0xD800..0xDC00).contains(&unit) { + match low(i) { + Some(lo) => { + i += 7; + char::from_u32(0x10000 + ((unit - 0xD800) << 10) + (lo - 0xDC00)) + } + None => None, + } + } else { + char::from_u32(unit) + }; + push_literal(&mut out, ch.unwrap_or('\u{fffd}')); + continue; + } + } else if n == Some('M') && plus == Some('+') && chars.get(i + 3).is_some_and(char::is_ascii_digit) && hex4(&chars, i + 4).is_some() { + push_literal(&mut out, crate::mif_escape_at(&chars, i).unwrap_or('\u{fffd}')); + i += 8; + continue; + } + // Any other backslash pair stays as it is (and an escaped backslash stays one). + out.push('\\'); + i += 1; + if let Some(&next) = chars.get(i) { + out.push(next); + i += 1; + } + } + std::borrow::Cow::Owned(out) +} + +#[cfg(test)] +mod tests { + use super::*; + #[test] + fn unicode_and_mif_escapes_remain_literal_markup() { + assert_eq!(unescape_codes(r"a\U+0041\U+D83D\U+DE00"), "aA😀"); + assert_eq!(unescape_codes(r"\M+5BCFE"), "ä»¶"); + assert_eq!(unescape_codes(r"\U+005C\U+007B\U+007D"), r"\\\{\}"); + assert_eq!(unescape_codes(r"\\U+0041"), r"\\U+0041"); + assert_eq!(unescape_codes(r"\U+12"), r"\U+12"); + assert_eq!(unescape_codes(r"\U+D800"), "�"); + } + #[test] + fn parameterized_commands_are_classified_without_relaxing_others() { + for c in "fFHWQTACcp".chars() { + assert!(takes_arg(c)); + } + for c in ['S', 'P', 'Z', '\\'] { + assert!(!takes_arg(c)); + } + } +} diff --git a/crates/gzip/Cargo.toml b/crates/gzip/Cargo.toml new file mode 100644 index 000000000..3d4f04603 --- /dev/null +++ b/crates/gzip/Cargo.toml @@ -0,0 +1,14 @@ +[package] +name = "craft-gzip" +description = "Streaming gzip writer with stored and compressed segments" +version.workspace = true +edition.workspace = true +license.workspace = true +rust-version.workspace = true +repository.workspace = true + +[lints] +workspace = true + +[dependencies] +flate2 = { version = "1", default-features = false, features = ["rust_backend"] } diff --git a/crates/gzip/src/lib.rs b/crates/gzip/src/lib.rs new file mode 100644 index 000000000..33f2a5ac1 --- /dev/null +++ b/crates/gzip/src/lib.rs @@ -0,0 +1,119 @@ +//! Streaming gzip output with optional stored segments. +//! +//! Copyright (c) 2026 ArtCraft Team and the VectorCraft contributors. +//! Extracted from VectorCraft, licensed under MIT OR Apache-2.0. +//! +//! `flush` does not emit a deflate block; call [`Gz::finish`] to complete and flush the +//! stream. After an I/O error, discard the writer: the output may be incomplete. +#![deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable)] + +use flate2::{Compress, Compression, Crc, FlushCompress, Status}; +use std::io::{self, Write}; + +/// Bytes of compressed output gathered before they go to the writer. +const OUT_CHUNK: usize = 64 << 10; + +/// A gzip stream, written as [`flate2::write::GzEncoder`] at the default level writes it (the +/// same header and the same data; the deflate blocks are the same too where flate2 deflates with +/// miniz_oxide, not with zlib-rs, which writes them by how the data is fed), except that data which +/// is compressed already ([`Gz::packed`]) is stored as it is: the default level spends a second +/// or more per photo on it for next to nothing, and makes it slower to unpack. Where the level +/// changes the deflate blocks so far end with a sync flush and a fresh compressor writes on, +/// which makes one valid deflate stream that every reader unpacks. +pub struct Gz { + out: W, + deflate: Compress, + /// What `deflate` writes: stored blocks (packed data), else the default level. + stored: bool, + crc: Crc, + buf: Vec, + /// Bytes written (unpacked). + total: u64, +} + +impl Gz { + /// Write a gzip header and construct a writer. + pub fn new(mut out: W) -> io::Result { + // flate2's header for the default level: no name, no time, unknown OS. + out.write_all(&[0x1f, 0x8b, 8, 0, 0, 0, 0, 0, 0, 255])?; + Ok(Self { + out, + deflate: Compress::new(Compression::default(), false), + stored: false, + crc: Crc::new(), + buf: Vec::with_capacity(OUT_CHUNK), + total: 0, + }) + } + + /// Run the compressor over `input` with `flush` until it has taken it all (and, flushing, + /// written everything out). + fn deflate(&mut self, mut input: &[u8], flush: FlushCompress) -> io::Result<()> { + loop { + if self.buf.len() == self.buf.capacity() { + self.out.write_all(&self.buf)?; + self.buf.clear(); + } + let (before_in, before_out) = (self.deflate.total_in(), self.deflate.total_out()); + let status = self.deflate.compress_vec(input, &mut self.buf, flush).map_err(io::Error::other)?; + let taken = usize::try_from(self.deflate.total_in() - before_in).unwrap_or(input.len()).min(input.len()); + input = input.get(taken..).unwrap_or_default(); + let made = self.deflate.total_out() - before_out; + let full = self.buf.len() == self.buf.capacity(); + match flush { + FlushCompress::None if input.is_empty() => return Ok(()), + FlushCompress::Finish if status == Status::StreamEnd => return Ok(()), + // A flush is done once it left room in the output. + FlushCompress::Sync if input.is_empty() && !full => return Ok(()), + _ => {} + } + if taken == 0 && made == 0 && !full { + return Err(io::Error::other("the compressor made no progress")); + } + } + } + + /// Return the number of uncompressed bytes supplied. + pub fn written(&self) -> u64 { + self.total + } + + /// What is written next is compressed already (stored as it is), or not (compressed at the + /// default level). + pub fn packed(&mut self, packed: bool) -> io::Result<()> { + if packed == self.stored { + return Ok(()); + } + self.deflate(&[], FlushCompress::Sync)?; + self.deflate = Compress::new(if packed { Compression::none() } else { Compression::default() }, false); + self.stored = packed; + Ok(()) + } + + /// Complete the gzip trailer and flush the underlying writer. + pub fn finish(mut self) -> io::Result<()> { + self.deflate(&[], FlushCompress::Finish)?; + self.out.write_all(&self.buf)?; + self.out.write_all(&self.crc.sum().to_le_bytes())?; + self.out.write_all(&self.crc.amount().to_le_bytes())?; + self.out.flush() + } +} + +impl Write for Gz { + fn write(&mut self, buf: &[u8]) -> io::Result { + self.write_all(buf)?; + Ok(buf.len()) + } + + fn write_all(&mut self, buf: &[u8]) -> io::Result<()> { + self.total = self.total.checked_add(buf.len() as u64).ok_or_else(|| io::Error::other("uncompressed byte count overflow"))?; + self.crc.update(buf); + self.deflate(buf, FlushCompress::None) + } + + fn flush(&mut self) -> io::Result<()> { + // Flushing would end a deflate block: the stream is flushed when it is finished. + Ok(()) + } +} diff --git a/crates/gzip/tests/writer.rs b/crates/gzip/tests/writer.rs new file mode 100644 index 000000000..38981ad0a --- /dev/null +++ b/crates/gzip/tests/writer.rs @@ -0,0 +1,77 @@ +use std::io::{self, Read, Write}; + +use craft_gzip::Gz; +use flate2::{Compression, read::GzDecoder, write::GzEncoder}; + +fn decode(bytes: &[u8]) -> Vec { + let mut result = Vec::new(); + GzDecoder::new(bytes).read_to_end(&mut result).unwrap(); + result +} + +#[test] +fn matches_flate2_without_mode_changes() { + let data = b"The quick brown fox jumps over the lazy dog. ".repeat(10000); + let mut expected = GzEncoder::new(Vec::new(), Compression::default()); + let mut actual = Vec::new(); + { + let mut gz = Gz::new(&mut actual).unwrap(); + for chunk in data.chunks(131) { + expected.write_all(chunk).unwrap(); + gz.write_all(chunk).unwrap(); + } + assert_eq!(gz.written(), data.len() as u64); + gz.finish().unwrap(); + } + assert_eq!(actual, expected.finish().unwrap()); + assert_eq!(decode(&actual), data); +} + +#[test] +fn mixed_segments_round_trip_and_stored_bytes_remain_uncompressed() { + let plain = b"repeat me ".repeat(20000); + let packed: Vec = (0..200_000).map(|i| (i * 31) as u8).collect(); + let mut bytes = Vec::new(); + { + let mut gz = Gz::new(&mut bytes).unwrap(); + gz.write_all(&plain).unwrap(); + gz.packed(true).unwrap(); + gz.packed(true).unwrap(); + gz.write_all(&packed).unwrap(); + gz.packed(false).unwrap(); + gz.write_all(&plain).unwrap(); + assert_eq!(gz.written(), (plain.len() * 2 + packed.len()) as u64); + gz.finish().unwrap(); + } + let expected = [plain.as_slice(), packed.as_slice(), plain.as_slice()].concat(); + assert_eq!(decode(&bytes), expected); + assert!(bytes.windows(1024).any(|window| window == &packed[..1024])); + assert!(bytes.len() < expected.len()); +} + +struct FailAfter { + remaining: usize, +} +impl Write for FailAfter { + fn write(&mut self, data: &[u8]) -> io::Result { + if self.remaining == 0 { + return Err(io::Error::other("injected failure")); + } + let n = self.remaining.min(data.len()); + self.remaining -= n; + Ok(n) + } + fn flush(&mut self) -> io::Result<()> { + Err(io::Error::other("injected flush failure")) + } +} + +#[test] +fn propagates_header_body_and_flush_failures() { + assert!(Gz::new(FailAfter { remaining: 0 }).is_err()); + let mut body = Gz::new(FailAfter { remaining: 10 }).unwrap(); + body.packed(true).unwrap(); + assert!(body.write_all(&vec![42; 200_000]).is_err()); + let empty = Gz::new(FailAfter { remaining: usize::MAX }).unwrap(); + assert_eq!(empty.finish().unwrap_err().to_string(), "injected flush failure"); +} diff --git a/crates/image-codecs/Cargo.toml b/crates/image-codecs/Cargo.toml index df71d9860..a8872c0d3 100644 --- a/crates/image-codecs/Cargo.toml +++ b/crates/image-codecs/Cargo.toml @@ -47,6 +47,7 @@ exr = { version = "1.73", default-features = false, features = ["rayon"] } proptest = { workspace = true } image = { version = "0.25", default-features = false, features = ["png", "jpeg", "tiff"] } sha2 = { version = "0.10", default-features = false, features = ["std"] } +craft-media-testkit = { workspace = true } [target.'cfg(not(target_arch = "wasm32"))'.dev-dependencies] craft-video = { workspace = true, features = ["av1-decode"] } diff --git a/crates/image-codecs/README.md b/crates/image-codecs/README.md index aaf7f54b4..aa0027666 100644 --- a/crates/image-codecs/README.md +++ b/crates/image-codecs/README.md @@ -56,6 +56,7 @@ craft-image-codecs = { git = "https://github.com/storytold/craft-libs", rev = "< | Cineon | yes | yes | U16 (10-bit code values read and written exactly through U16) | Gray, RGB | no | no | no | no | no | no | no | built in | | DICOM | yes | **no** | U8, U16 | Gray, RGB | no | no | no | no | no | no | n/a | built in | | Amiga IFF (ILBM/PBM) | yes | yes (24-bit ILBM) | U8 | RGB, RGBA | yes (1-bit mask plane on write) | no | no | no | no | no | no | built in | +| JPEG-LS (`.jls`) | yes (any NEAR; one scan per component) | yes (lossless) | U8, U16 | Gray, RGB | no | no | no | no | no | no | no | built in | | AVIF | **no** | only with feature `avif` | U8 | RGB, RGBA | yes | no | no | no | no | no | yes | `image` + `ravif` | | HEIF/HEIC | with feature `heif` | **no** | U8, U16 (10/12-bit decodes to U16) | RGB, RGBA | yes (auxiliary alpha) | yes (`colr` prof) | yes | yes | no | no | n/a | `heic-rs` | diff --git a/crates/image-codecs/src/bmp_probe.rs b/crates/image-codecs/src/bmp_probe.rs new file mode 100644 index 000000000..59c78ca1e --- /dev/null +++ b/crates/image-codecs/src/bmp_probe.rs @@ -0,0 +1,145 @@ +//! Header-only BMP dimensions and pixel-data coverage, without allocating a canvas. + +/// Coverage of the declared pixel array. Row byte counts include row padding; RLE coverage +/// counts the pixels drawn by commands and may be smaller than the canvas. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub struct Coverage { + pub pixels: u64, + pub row_bytes: Option<(u64, u64)>, +} + +/// Inspect pixel data for the given dimensions without decoding a bitmap. The row tuple is +/// `(available, needed)`; callers choose their own sparse-canvas and allocation policy. +pub fn coverage(bytes: &[u8], w: u32, h: u32) -> Coverage { + match bmp_layout(bytes) { + BmpLayout::Rle { offset, rle4 } => Coverage { pixels: bmp_rle_pixels(bytes, offset, rle4, w, h), row_bytes: None }, + BmpLayout::Rows { offset, bits } => { + let needed = u64::from(w).saturating_mul(u64::from(bits)).div_ceil(32).saturating_mul(4).saturating_mul(u64::from(h)); + let held = bytes.len().saturating_sub(offset) as u64; + Coverage { pixels: u64::from(w) * u64::from(h), row_bytes: Some((held, needed)) } + } + } +} + +enum BmpLayout { + /// RLE8 or RLE4 commands. + Rle { offset: usize, rle4: bool }, + /// Rows of `bits` per pixel, each padded to 4 bytes. + Rows { offset: usize, bits: u16 }, +} + +/// The layout the decoder reads, from the file header and the bitmap header: a 12-byte core header +/// (bits per pixel at 24, never compressed) or a 40-byte or larger one (bits at 28, compression at +/// 30). Fields the decoder would have refused are read as 1 bit per pixel at offset 0, the most +/// lenient layout. +fn bmp_layout(bytes: &[u8]) -> BmpLayout { + let u16_at = |i: usize| bytes.get(i..i.checked_add(2)?).and_then(|b| b.try_into().ok()).map(u16::from_le_bytes); + let u32_at = |i: usize| bytes.get(i..i.checked_add(4)?).and_then(|b| b.try_into().ok()).map(u32::from_le_bytes); + let offset = u32_at(10).and_then(|o| usize::try_from(o).ok()).unwrap_or(0); + let (bits, compression) = match u32_at(14) { + Some(12) => (u16_at(24), Some(0)), + _ => (u16_at(28), u32_at(30)), + }; + match compression { + Some(1) => BmpLayout::Rle { offset, rle4: false }, + Some(2) => BmpLayout::Rle { offset, rle4: true }, + _ => BmpLayout::Rows { offset, bits: bits.filter(|&b| b > 0).unwrap_or(1) }, + } +} + +/// The pixels an RLE8 or RLE4 BMP's commands draw on a `w` × `h` canvas, read the way the decoder +/// paints them: a run is clipped at the end of its row, an end-of-row code moves to the next row, a +/// delta skips pixels (left blank), and an end-of-bitmap code, a truncated command or the last row +/// ends the picture. Every command takes at least 2 bytes, so the walk is linear in the file. +fn bmp_rle_pixels(bytes: &[u8], offset: usize, rle4: bool, w: u32, h: u32) -> u64 { + let (w, h) = (u64::from(w), u64::from(h)); + let (mut pos, mut x, mut y, mut covered) = (offset, 0u64, 0u64, 0u64); + while y < h { + let (Some(&count), Some(&value)) = (bytes.get(pos), bytes.get(pos.saturating_add(1))) else { + break; + }; + pos = pos.saturating_add(2); + match (count, value) { + // end of row + (0, 0) => (x, y) = (0, y.saturating_add(1)), + // end of bitmap + (0, 1) => break, + // delta: right and down + (0, 2) => { + let (Some(&dx), Some(&dy)) = (bytes.get(pos), bytes.get(pos.saturating_add(1))) else { + break; + }; + pos = pos.saturating_add(2); + (x, y) = (x.saturating_add(u64::from(dx)), y.saturating_add(u64::from(dy))); + } + // absolute: `value` pixels stored one per byte (RLE8) or nibble (RLE4), padded to 2 bytes + (0, n) => { + let len = if rle4 { usize::from(n).div_ceil(2) } else { usize::from(n) }; + pos = pos.saturating_add(len + (len & 1)); + if pos > bytes.len() { + break; + } + covered = covered.saturating_add(u64::from(n).min(w.saturating_sub(x))); + x = x.saturating_add(u64::from(n)); + } + // a run of `count` pixels + (n, _) => { + covered = covered.saturating_add(u64::from(n).min(w.saturating_sub(x))); + x = x.saturating_add(u64::from(n)); + } + } + } + covered +} + +/// Reads BMP `(width, height)` from the DIB header without allocating. +pub fn dimensions(bytes: &[u8]) -> Option<(u32, u32)> { + if bytes.get(0..2) != Some(b"BM") { + return None; + } + let dib = u32::from_le_bytes(bytes.get(14..18)?.try_into().ok()?); + let (w, h) = if dib == 12 { + (u16::from_le_bytes(bytes.get(18..20)?.try_into().ok()?) as u32, u16::from_le_bytes(bytes.get(20..22)?.try_into().ok()?) as u32) + } else { + let w = i32::from_le_bytes(bytes.get(18..22)?.try_into().ok()?); + let h = i32::from_le_bytes(bytes.get(22..26)?.try_into().ok()?); + (u32::try_from(w).ok()?, h.unsigned_abs()) + }; + Some((w, h)) +} + +#[cfg(test)] +mod tests { + use super::*; + + fn bmp(compression: u32, pixels: &[u8]) -> Vec { + let mut b = vec![0; 54]; + b[..2].copy_from_slice(b"BM"); + b[10..14].copy_from_slice(&54u32.to_le_bytes()); + b[14..18].copy_from_slice(&40u32.to_le_bytes()); + b[18..22].copy_from_slice(&4i32.to_le_bytes()); + b[22..26].copy_from_slice(&2i32.to_le_bytes()); + b[28..30].copy_from_slice(&8u16.to_le_bytes()); + b[30..34].copy_from_slice(&compression.to_le_bytes()); + b.extend_from_slice(pixels); + b + } + + #[test] + fn dimensions_and_padded_rows() { + let b = bmp(0, &[0; 7]); + assert_eq!(dimensions(&b), Some((4, 2))); + assert_eq!(coverage(&b, 4, 2), Coverage { pixels: 8, row_bytes: Some((7, 8)) }); + assert_eq!(dimensions(&b[..20]), None); + } + + #[test] + fn rle_counts_clipped_runs_deltas_and_complete_absolute_blocks() { + let b = bmp(1, &[3, 7, 0, 0, 0, 2, 1, 0, 0, 3, 1, 2, 3, 0, 0, 1]); + assert_eq!(coverage(&b, 4, 2).pixels, 6); + let truncated = bmp(1, &[0, 4, 1, 2]); + assert_eq!(coverage(&truncated, 4, 2).pixels, 0); + let nibble = bmp(2, &[0, 3, 0x12, 0x30, 0, 1]); + assert_eq!(coverage(&nibble, 4, 2).pixels, 3); + } +} diff --git a/crates/image-codecs/src/codecs.rs b/crates/image-codecs/src/codecs.rs index 0288ae22b..af91aa536 100644 --- a/crates/image-codecs/src/codecs.rs +++ b/crates/image-codecs/src/codecs.rs @@ -13,6 +13,7 @@ pub(crate) mod heif; pub(crate) mod iff; pub(crate) mod jp2; pub(crate) mod jpeg; +pub(crate) mod jpegls; pub(crate) mod jxl; pub(crate) mod pcx; pub(crate) mod png; diff --git a/crates/image-codecs/src/codecs/jpegls.rs b/crates/image-codecs/src/codecs/jpegls.rs new file mode 100644 index 000000000..b405e71ed --- /dev/null +++ b/crates/image-codecs/src/codecs/jpegls.rs @@ -0,0 +1,870 @@ +//! JPEG-LS (ITU-T T.87 | ISO/IEC 14495-1, "LOCO-I"): lossless and near-lossless gray and RGB +//! images, 2 to 16 bits per sample. +//! +//! Read: baseline JPEG-LS with the default or `LSE`-preset coding parameters, any `NEAR`, and +//! scans that code one component each (`ILV` 0), which is also how this module writes. Line- and +//! sample-interleaved scans, colour transforms and mapping tables are reported as unsupported. +//! Write: lossless (`NEAR` 0), 8 or 16 bits, gray or RGB, one scan per component. +//! +//! Samples of 9 to 15 bits are scaled to 16 bits on read (the top bits are repeated into the low +//! ones), so a decoded image always spans the full range of its sample type. + +use crate::error::CodecError; +use crate::fidelity::Plan; +use crate::format::Format; +use crate::image::{ChannelLayout, Image, SampleType}; +use crate::options::Limits; + +const F: Format = Format::JpegLs; +const SOF55: u8 = 0xF7; +const LSE: u8 = 0xF8; +const RESET_DEFAULT: i32 = 64; +const MIN_C: i32 = -128; +const MAX_C: i32 = 127; +/// `J[RUNindex]`: the log2 of the run length a run-index step codes. +const J: [u32; 32] = [0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 9, 10, 11, 12, 13, 14, 15]; +/// Context slots: the regular contexts (365 of the 405 slots the index below can reach), then the two +/// run-interruption contexts. +const RI_BASE: usize = 405; +const CONTEXTS: usize = RI_BASE + 2; + +fn bad(reason: impl std::fmt::Display) -> CodecError { + CodecError::malformed(F, reason) +} + +/// Whether `b` is a JPEG-LS file: an SOI followed by marker segments up to an SOF55 frame header. +pub(crate) fn looks_like_jpegls(b: &[u8]) -> bool { + if !b.starts_with(&[0xFF, 0xD8]) { + return false; + } + let mut pos = 2; + for _ in 0..64 { + let (Some(&0xFF), Some(&marker)) = (b.get(pos), b.get(pos + 1)) else { return false }; + match marker { + SOF55 => return true, + 0xDA | 0xD9 => return false, + 0xC0..=0xCF | 0xDB..=0xDF => return false, + 0x01 | 0xD0..=0xD7 | 0xFF => pos += if marker == 0xFF { 1 } else { 2 }, + _ => { + let Some(len) = b.get(pos + 2..pos + 4).map(|l| usize::from(u16::from_be_bytes([l[0], l[1]]))) else { return false }; + pos += 2 + len; + } + } + } + false +} + +/// The coding parameters of one scan, derived as T.87 section C.2 specifies. +#[derive(Clone, Copy)] +struct Params { + maxval: i32, + near: i32, + range: i32, + qbpp: u32, + limit: u32, + t1: i32, + t2: i32, + t3: i32, + reset: i32, +} + +fn ceil_log2(v: i32) -> u32 { + let mut n = 0; + while (1i64 << n) < i64::from(v) { + n += 1; + } + n +} + +impl Params { + /// `preset` holds the `LSE` values (`maxval`, `t1`, `t2`, `t3`, `reset`), 0 where absent. + fn new(precision: u32, near: i32, preset: [i32; 5]) -> Result { + let default_max = (1i32 << precision) - 1; + let maxval = if preset[0] == 0 { default_max } else { preset[0] }; + if maxval < 1 || maxval > default_max || near < 0 || near > (maxval / 2).min(255) { + return Err(bad("invalid MAXVAL or NEAR")); + } + let clamp = |v: i32, lo: i32, hi: i32| if v > hi || v < lo { lo } else { v }; + let (d1, d2, d3) = if maxval >= 128 { + let factor = (maxval.min(4095) + 128) / 256; + let t1 = clamp(factor + 2 + 3 * near, near + 1, maxval); + let t2 = clamp(factor * 4 + 3 + 5 * near, t1, maxval); + let t3 = clamp(factor * 17 + 4 + 7 * near, t2, maxval); + (t1, t2, t3) + } else { + let factor = 256 / (maxval + 1); + let t1 = clamp((3 / factor + 3 * near).max(2), near + 1, maxval); + let t2 = clamp((7 / factor + 5 * near).max(3), t1, maxval); + let t3 = clamp((21 / factor + 7 * near).max(4), t2, maxval); + (t1, t2, t3) + }; + let t1 = if preset[1] == 0 { d1 } else { preset[1] }; + let t2 = if preset[2] == 0 { d2 } else { preset[2] }; + let t3 = if preset[3] == 0 { d3 } else { preset[3] }; + let reset = if preset[4] == 0 { RESET_DEFAULT } else { preset[4] }; + if !(near < t1 && t1 <= t2 && t2 <= t3 && t3 <= maxval) || !(3..=255).contains(&reset) { + return Err(bad("invalid threshold or RESET preset")); + } + let range = (maxval + 2 * near) / (2 * near + 1) + 1; + let bpp = ceil_log2(maxval + 1).max(2); + Ok(Params { maxval, near, range, qbpp: ceil_log2(range), limit: 2 * (bpp + bpp.max(8)), t1, t2, t3, reset }) + } +} + +/// Adaptive state of one scan: contexts, bias and the run index. +struct State { + a: [i32; CONTEXTS], + b: [i32; CONTEXTS], + c: [i32; CONTEXTS], + n: [i32; CONTEXTS], + nn: [i32; 2], + run_index: usize, +} + +impl State { + fn new(p: &Params) -> State { + let a0 = ((p.range + 32) / 64).max(2); + State { a: [a0; CONTEXTS], b: [0; CONTEXTS], c: [0; CONTEXTS], n: [1; CONTEXTS], nn: [0; 2], run_index: 0 } + } +} + +fn quantise_gradient(d: i32, p: &Params) -> i32 { + if d <= -p.t3 { + -4 + } else if d <= -p.t2 { + -3 + } else if d <= -p.t1 { + -2 + } else if d < -p.near { + -1 + } else if d <= p.near { + 0 + } else if d < p.t1 { + 1 + } else if d < p.t2 { + 2 + } else if d < p.t3 { + 3 + } else { + 4 + } +} + +fn predict(ra: i32, rb: i32, rc: i32) -> i32 { + if rc >= ra.max(rb) { + ra.min(rb) + } else if rc <= ra.min(rb) { + ra.max(rb) + } else { + ra + rb - rc + } +} + +/// The error quantised for `NEAR` and reduced into the modulo range, as the encoder codes it. +fn reduce(mut e: i32, p: &Params) -> i32 { + if p.near > 0 { + e = if e > 0 { (e + p.near) / (2 * p.near + 1) } else { -(p.near - e) / (2 * p.near + 1) }; + } + if e < 0 { + e += p.range; + } + if e >= (p.range + 1) / 2 { + e -= p.range; + } + e +} + +/// The reconstructed sample for a decoded (sign-adjusted) error. +fn reconstruct(px: i32, sign: i32, e: i32, p: &Params) -> i32 { + let mut r = px + sign * e * (2 * p.near + 1); + if r < -p.near { + r += p.range * (2 * p.near + 1); + } else if r > p.maxval + p.near { + r -= p.range * (2 * p.near + 1); + } + r.clamp(0, p.maxval) +} + +fn golomb_k(n: i32, a: i32) -> u32 { + let mut k = 0; + while (i64::from(n) << k) < i64::from(a) { + k += 1; + } + k +} + +/// Bits of the entropy-coded segment, with T.87's stuffing: after an `0xFF` byte the next byte +/// carries only 7 bits. +struct BitReader<'a> { + data: &'a [u8], + pos: usize, + cur: u32, + left: u32, + prev_ff: bool, +} + +impl<'a> BitReader<'a> { + fn new(data: &'a [u8]) -> Self { + BitReader { data, pos: 0, cur: 0, left: 0, prev_ff: false } + } + + fn bit(&mut self) -> Result { + if self.left == 0 { + let byte = *self.data.get(self.pos).ok_or_else(|| bad("the entropy-coded data ends early"))?; + self.pos += 1; + if self.prev_ff { + if byte & 0x80 != 0 { + return Err(bad("a marker inside entropy-coded data")); + } + self.cur = u32::from(byte); + self.left = 7; + } else { + self.cur = u32::from(byte); + self.left = 8; + } + self.prev_ff = byte == 0xFF; + } + self.left -= 1; + Ok((self.cur >> self.left) & 1) + } + + fn bits(&mut self, n: u32) -> Result { + let mut v = 0; + for _ in 0..n { + v = v << 1 | self.bit()?; + } + Ok(v) + } + + /// A limited-length Golomb code: the mapped error. + fn golomb(&mut self, k: u32, limit: u32, qbpp: u32) -> Result { + let max_zeros = limit.checked_sub(qbpp + 1).ok_or_else(|| bad("invalid LIMIT"))?; + let mut zeros = 0; + while self.bit()? == 0 { + zeros += 1; + if zeros > max_zeros { + return Err(bad("a Golomb code is longer than LIMIT")); + } + } + let v = if zeros < max_zeros { (zeros << k) + self.bits(k)? } else { self.bits(qbpp)? + 1 }; + i32::try_from(v).map_err(|_| bad("mapped error out of range")) + } +} + +/// Decode one component scan of `w` x `h` samples into `out` (row-major). +fn decode_scan(data: &[u8], w: usize, h: usize, p: &Params, out: &mut [u16]) -> Result<(), CodecError> { + let mut r = BitReader::new(data); + let mut st = State::new(p); + let mut prev = vec![0i32; w + 2]; + let mut cur = vec![0i32; w + 2]; + for y in 0..h { + // Ra at the first sample is Rb; the slot at index 0 is Rc there, left over from the line above. + cur[0] = prev[1]; + if y > 0 { + prev[w + 1] = prev[w]; + } + let mut x = 0usize; + while x < w { + let (ra, rb, rc, rd) = (cur[x], prev[x + 1], prev[x], if y == 0 { 0 } else { prev[x + 2] }); + let (q1, q2, q3) = (quantise_gradient(rd - rb, p), quantise_gradient(rb - rc, p), quantise_gradient(rc - ra, p)); + if q1 == 0 && q2 == 0 && q3 == 0 { + x = decode_run(&mut r, &mut st, p, &prev, &mut cur, ra, x, w)?; + continue; + } + let (sign, q1, q2, q3) = if q1 < 0 || (q1 == 0 && (q2 < 0 || (q2 == 0 && q3 < 0))) { (-1, -q1, -q2, -q3) } else { (1, q1, q2, q3) }; + let q = (q1 * 81 + (q2 + 4) * 9 + (q3 + 4)) as usize; + let k = golomb_k(st.n[q], st.a[q]); + let m = r.golomb(k, p.limit, p.qbpp)?; + let swap = p.near == 0 && k == 0 && 2 * st.b[q] <= -st.n[q]; + let e = if swap { + if m & 1 == 1 { (m - 1) >> 1 } else { -(m >> 1) - 1 } + } else if m & 1 == 0 { + m >> 1 + } else { + -((m + 1) >> 1) + }; + let px = (predict(ra, rb, rc) + sign * st.c[q]).clamp(0, p.maxval); + let rx = reconstruct(px, sign, e, p); + update_regular(&mut st, q, e, p); + cur[x + 1] = rx; + x += 1; + } + let row = &mut out[y * w..(y + 1) * w]; + for (o, &v) in row.iter_mut().zip(&cur[1..=w]) { + *o = v as u16; + } + // The line just decoded becomes the line above; its slot 0 is the Rc of the next line's first sample. + std::mem::swap(&mut prev, &mut cur); + } + Ok(()) +} + +fn update_regular(st: &mut State, q: usize, e: i32, p: &Params) { + st.b[q] += e * (2 * p.near + 1); + st.a[q] += e.abs(); + if st.n[q] == p.reset { + st.a[q] >>= 1; + st.b[q] = if st.b[q] >= 0 { st.b[q] >> 1 } else { -((1 - st.b[q]) >> 1) }; + st.n[q] >>= 1; + } + st.n[q] += 1; + if st.b[q] <= -st.n[q] { + if st.c[q] > MIN_C { + st.c[q] -= 1; + } + st.b[q] += st.n[q]; + if st.b[q] <= -st.n[q] { + st.b[q] = -st.n[q] + 1; + } + } else if st.b[q] > 0 { + if st.c[q] < MAX_C { + st.c[q] += 1; + } + st.b[q] -= st.n[q]; + if st.b[q] > 0 { + st.b[q] = 0; + } + } +} + +/// Run-interruption variables update shared by both directions. +fn update_run(st: &mut State, q: usize, e: i32, em: i32, ri_type: i32, p: &Params) { + let i = q - RI_BASE; + if e < 0 { + st.nn[i] += 1; + } + st.a[q] += (em + 1 - ri_type) >> 1; + if st.n[q] == p.reset { + st.a[q] >>= 1; + st.n[q] >>= 1; + st.nn[i] >>= 1; + } + st.n[q] += 1; +} + +/// The run-interruption Golomb parameter and the mapping flag of an error. +fn run_k(st: &State, q: usize, ri_type: i32) -> u32 { + let temp = if ri_type == 1 { st.a[q] + (st.n[q] >> 1) } else { st.a[q] }; + golomb_k(st.n[q], temp) +} + +#[allow(clippy::too_many_arguments)] +fn decode_run( + r: &mut BitReader, + st: &mut State, + p: &Params, + prev: &[i32], + cur: &mut [i32], + ra: i32, + x: usize, + w: usize, +) -> Result { + let remaining = w - x; + let mut run = 0usize; + while r.bit()? == 1 { + let step = 1usize << J[st.run_index]; + let count = step.min(remaining - run); + run += count; + if count == step && st.run_index < 31 { + st.run_index += 1; + } + if run == remaining { + break; + } + } + if run < remaining { + let rest = r.bits(J[st.run_index])? as usize; + run += rest; + if run > remaining { + return Err(bad("a run passes the end of the line")); + } + } + for v in &mut cur[x + 1..x + 1 + run] { + *v = ra; + } + if run == remaining { + return Ok(w); + } + // The sample that ended the run: its own Rb, the sample above it. + let rb = prev[x + run + 1]; + let ri_type = i32::from((ra - rb).abs() <= p.near); + let q = RI_BASE + ri_type as usize; + let k = run_k(st, q, ri_type); + let limit = p.limit - J[st.run_index] - 1; + let em = r.golomb(k, limit, p.qbpp)?; + let tmp = em + ri_type; + let cond = k == 0 && 2 * st.nn[ri_type as usize] < st.n[q]; + let map = tmp & 1; + let mag = (tmp + map) / 2; + let e = if (cond && map == 1) || (!cond && map == 0) { mag } else { -mag }; + let (px, sign) = if ri_type == 1 { (ra, 1) } else { (rb, if ra > rb { -1 } else { 1 }) }; + // The decoder holds the sign-adjusted error; the sign flips the reconstruction. + let rx = reconstruct(px, sign, e, p); + update_run(st, q, e, em, ri_type, p); + if st.run_index > 0 { + st.run_index -= 1; + } + cur[x + run + 1] = rx; + Ok(x + run + 1) +} + +// ---- encoder ---- + +struct BitWriter { + out: Vec, + cur: u32, + used: u32, + cap: u32, +} + +impl BitWriter { + fn new() -> Self { + BitWriter { out: Vec::new(), cur: 0, used: 0, cap: 8 } + } + + fn put(&mut self, value: u32, n: u32) { + for i in (0..n).rev() { + // A run of more than 32 zero bits (a long escape code) has no bits of `value` to shift out. + let bit = if i < 32 { (value >> i) & 1 } else { 0 }; + self.cur = self.cur << 1 | bit; + self.used += 1; + if self.used == self.cap { + let byte = self.cur as u8; + self.out.push(byte); + self.cap = if byte == 0xFF { 7 } else { 8 }; + self.cur = 0; + self.used = 0; + } + } + } + + /// Pad the last byte with zero bits; a closing 0xFF is followed by a zero byte. + fn finish(mut self) -> Vec { + if self.used > 0 { + let pad = self.cap - self.used; + self.put(0, pad); + } + if self.out.last() == Some(&0xFF) { + self.out.push(0); + } + self.out + } + + fn golomb(&mut self, m: i32, k: u32, limit: u32, qbpp: u32) { + let m = m as u32; + let max_zeros = limit - qbpp - 1; + let q = m >> k; + if q < max_zeros { + self.put(0, q); + self.put(1, 1); + self.put(m & ((1 << k) - 1), k); + } else { + self.put(0, max_zeros); + self.put(1, 1); + self.put(m - 1, qbpp); + } + } +} + +fn encode_scan(plane: &[u16], w: usize, h: usize, p: &Params) -> Vec { + let mut wr = BitWriter::new(); + let mut st = State::new(p); + let mut prev = vec![0i32; w + 2]; + let mut cur = vec![0i32; w + 2]; + for y in 0..h { + cur[0] = prev[1]; + if y > 0 { + prev[w + 1] = prev[w]; + } + let line = &plane[y * w..(y + 1) * w]; + let mut x = 0usize; + while x < w { + let (ra, rb, rc, rd) = (cur[x], prev[x + 1], prev[x], if y == 0 { 0 } else { prev[x + 2] }); + let ix = i32::from(line[x]); + let (q1, q2, q3) = (quantise_gradient(rd - rb, p), quantise_gradient(rb - rc, p), quantise_gradient(rc - ra, p)); + if q1 == 0 && q2 == 0 && q3 == 0 { + x = encode_run(&mut wr, &mut st, p, line, &prev, &mut cur, ra, x); + continue; + } + let (sign, q1, q2, q3) = if q1 < 0 || (q1 == 0 && (q2 < 0 || (q2 == 0 && q3 < 0))) { (-1, -q1, -q2, -q3) } else { (1, q1, q2, q3) }; + let q = (q1 * 81 + (q2 + 4) * 9 + (q3 + 4)) as usize; + let px = (predict(ra, rb, rc) + sign * st.c[q]).clamp(0, p.maxval); + let e = reduce(sign * (ix - px), p); + let k = golomb_k(st.n[q], st.a[q]); + let swap = p.near == 0 && k == 0 && 2 * st.b[q] <= -st.n[q]; + let m = if swap { + if e >= 0 { 2 * e + 1 } else { -2 * (e + 1) } + } else if e >= 0 { + 2 * e + } else { + -2 * e - 1 + }; + wr.golomb(m, k, p.limit, p.qbpp); + cur[x + 1] = reconstruct(px, sign, e, p); + update_regular(&mut st, q, e, p); + x += 1; + } + std::mem::swap(&mut prev, &mut cur); + } + wr.finish() +} + +#[allow(clippy::too_many_arguments)] +fn encode_run(wr: &mut BitWriter, st: &mut State, p: &Params, line: &[u16], prev: &[i32], cur: &mut [i32], ra: i32, x: usize) -> usize { + let w = line.len(); + let mut run = 0usize; + while x + run < w && (i32::from(line[x + run]) - ra).abs() <= p.near { + cur[x + run + 1] = ra; + run += 1; + } + let mut count = run; + while count >= 1usize << J[st.run_index] { + wr.put(1, 1); + count -= 1usize << J[st.run_index]; + if st.run_index < 31 { + st.run_index += 1; + } + } + if x + run == w { + if count > 0 { + wr.put(1, 1); + } + return w; + } + wr.put(0, 1); + wr.put(count as u32, J[st.run_index]); + let ix = i32::from(line[x + run]); + let rb = prev[x + run + 1]; + let ri_type = i32::from((ra - rb).abs() <= p.near); + let q = RI_BASE + ri_type as usize; + let (px, sign) = if ri_type == 1 { (ra, 1) } else { (rb, if ra > rb { -1 } else { 1 }) }; + let e = reduce(sign * (ix - px), p); + let k = run_k(st, q, ri_type); + let n_neg = st.nn[ri_type as usize]; + let map = i32::from((k == 0 && e > 0 && 2 * n_neg < st.n[q]) || (e < 0 && 2 * n_neg >= st.n[q]) || (e < 0 && k != 0)); + let em = 2 * e.abs() - ri_type - map; + wr.golomb(em, k, p.limit - J[st.run_index] - 1, p.qbpp); + cur[x + run + 1] = reconstruct(px, sign, e, p); + update_run(st, q, e, em, ri_type, p); + if st.run_index > 0 { + st.run_index -= 1; + } + x + run + 1 +} + +/// Writes a lossless JPEG-LS file: 8 or 16 bits, gray or RGB, one scan per component. +pub(crate) fn encode(src: &Image, plan: Plan) -> Result, CodecError> { + let enc = |msg: String| CodecError::encode(F, msg); + let img = src.converted(plan.layout, plan.sample); + let nc = match img.layout() { + ChannelLayout::Gray => 1, + ChannelLayout::Rgb => 3, + l => return Err(enc(format!("no JPEG-LS layout for {l:?}"))), + }; + let precision: u32 = match img.sample_type() { + SampleType::U8 => 8, + SampleType::U16 => 16, + s => return Err(enc(format!("JPEG-LS stores 8 or 16 bits per sample, not {s:?}"))), + }; + let (w, h) = img.dimensions(); + let (Ok(w16), Ok(h16)) = (u16::try_from(w), u16::try_from(h)) else { + return Err(enc(format!("dimensions exceed 65535x65535: {w}x{h}"))); + }; + let (wu, hu) = (w as usize, h as usize); + let data = img.data(); + if data.len() != wu * hu * nc * (precision as usize / 8) { + return Err(enc("the image has the wrong number of samples".into())); + } + let params = Params::new(precision, 0, [0; 5]).map_err(|_| enc("invalid coding parameters".into()))?; + let mut out = vec![0xFF, 0xD8, 0xFF, SOF55]; + out.extend_from_slice(&((8 + 3 * nc) as u16).to_be_bytes()); + out.push(precision as u8); + out.extend_from_slice(&h16.to_be_bytes()); + out.extend_from_slice(&w16.to_be_bytes()); + out.push(nc as u8); + for c in 0..nc { + out.extend_from_slice(&[c as u8 + 1, 0x11, 0]); + } + let mut plane = vec![0u16; wu * hu]; + for c in 0..nc { + for (i, v) in plane.iter_mut().enumerate() { + let at = i * nc + c; + *v = if precision == 8 { u16::from(data[at]) } else { u16::from_ne_bytes([data[2 * at], data[2 * at + 1]]) }; + } + out.extend_from_slice(&[0xFF, 0xDA, 0, 8, 1, c as u8 + 1, 0, 0, 0, 0]); + out.extend_from_slice(&encode_scan(&plane, wu, hu, ¶ms)); + } + out.extend_from_slice(&[0xFF, 0xD9]); + Ok(out) +} + +struct Frame { + precision: u32, + width: usize, + height: usize, + components: Vec, +} + +fn be16(b: &[u8], i: usize) -> Result { + b.get(i..i + 2).map(|s| usize::from(u16::from_be_bytes([s[0], s[1]]))).ok_or_else(|| bad("truncated header")) +} + +pub(crate) fn decode(b: &[u8], limits: &Limits) -> Result { + if !looks_like_jpegls(b) { + return Err(bad("not a JPEG-LS file")); + } + let mut pos = 2; + let mut frame: Option = None; + let mut preset = [0i32; 5]; + let mut planes: Vec> = Vec::new(); + let mut scans_done = 0; + loop { + // Find the next marker. + while b.get(pos) != Some(&0xFF) { + if pos >= b.len() { + return finish(frame, planes, scans_done); + } + pos += 1; + } + while b.get(pos + 1) == Some(&0xFF) { + pos += 1; + } + let Some(&marker) = b.get(pos + 1) else { return finish(frame, planes, scans_done) }; + if marker == 0xD9 { + return finish(frame, planes, scans_done); + } + if marker == 0x00 || (0xD0..=0xD7).contains(&marker) || marker == 0x01 { + pos += 2; + continue; + } + let len = be16(b, pos + 2)?; + let seg = b.get(pos + 4..pos + 2 + len).ok_or_else(|| bad("truncated marker segment"))?; + match marker { + SOF55 => { + if frame.is_some() || seg.len() < 6 { + return Err(bad("bad or repeated frame header")); + } + let (precision, height, width, nf) = (u32::from(seg[0]), be16(seg, 1)?, be16(seg, 3)?, usize::from(seg[5])); + if !(2..=16).contains(&precision) || width == 0 || height == 0 || !matches!(nf, 1 | 3) || seg.len() < 6 + 3 * nf { + return Err(bad("unsupported frame header")); + } + if height == 0 { + return Err(CodecError::unsupported(F, "a frame height set later by a DNL marker")); + } + let layout = if nf == 1 { ChannelLayout::Gray } else { ChannelLayout::Rgb }; + let sample = if precision <= 8 { SampleType::U8 } else { SampleType::U16 }; + limits.check(width as u32, height as u32, layout, sample)?; + let components = (0..nf).map(|i| seg[6 + 3 * i]).collect(); + frame = Some(Frame { precision, width, height, components }); + } + LSE => { + if seg.first() == Some(&1) && seg.len() >= 11 { + for (i, p) in preset.iter_mut().enumerate() { + *p = be16(seg, 1 + 2 * i)? as i32; + } + } else if seg.first() != Some(&1) { + return Err(CodecError::unsupported(F, "mapping tables and colour transforms (LSE)")); + } + } + 0xDA => { + let f = frame.as_ref().ok_or_else(|| bad("a scan before the frame header"))?; + let ns = usize::from(*seg.first().ok_or_else(|| bad("empty scan header"))?); + if seg.len() < 1 + 2 * ns + 3 || ns != 1 { + return Err(CodecError::unsupported(F, "interleaved scans (ILV 1 and 2)")); + } + let (cs, tm, near, ilv) = (seg[1], seg[2], i32::from(seg[3]), seg[4]); + if tm != 0 { + return Err(CodecError::unsupported(F, "mapping tables")); + } + if ilv != 0 && f.components.len() > 1 { + return Err(CodecError::unsupported(F, "interleaved scans (ILV 1 and 2)")); + } + if seg[5] & 0x0f != 0 { + return Err(CodecError::unsupported(F, "a point transform")); + } + let idx = f.components.iter().position(|&c| c == cs).ok_or_else(|| bad("a scan names an unknown component"))?; + if planes.len() <= idx { + planes.resize(f.components.len(), Vec::new()); + } + if !planes[idx].is_empty() { + return Err(bad("a component is coded twice")); + } + let params = Params::new(f.precision, near, preset)?; + let start = pos + 2 + len; + // The scan ends at the next real marker: an 0xFF followed by a byte with its top bit set. + let mut end = start; + while end < b.len() { + if b[end] == 0xFF && b.get(end + 1).is_none_or(|&n| n & 0x80 != 0) { + break; + } + end += 1; + } + let mut samples = Vec::new(); + samples.try_reserve_exact(f.width * f.height).map_err(|_| bad("pixel buffer is too large"))?; + samples.resize(f.width * f.height, 0u16); + decode_scan(&b[start..end], f.width, f.height, ¶ms, &mut samples)?; + planes[idx] = samples; + scans_done += 1; + pos = end; + continue; + } + _ => {} + } + pos += 2 + len; + } +} + +fn finish(frame: Option, planes: Vec>, scans: usize) -> Result { + let f = frame.ok_or_else(|| bad("no frame header"))?; + if scans != f.components.len() || planes.iter().any(Vec::is_empty) || planes.len() != f.components.len() { + return Err(bad("the file ends before every component is coded")); + } + let (w, h, nc) = (f.width, f.height, f.components.len()); + let layout = if nc == 1 { ChannelLayout::Gray } else { ChannelLayout::Rgb }; + let p = f.precision; + let scale = |v: u16| -> u16 { + if p == 16 { + v + } else { + let v = u32::from(v); + ((v << (16 - p)) | (v >> (2 * p).saturating_sub(16).min(15))) as u16 + } + }; + if p <= 8 { + let to8 = |v: u16| -> u8 { if p == 8 { v as u8 } else { (scale(v) >> 8) as u8 } }; + let mut data = Vec::with_capacity(w * h * nc); + for i in 0..w * h { + for plane in &planes { + data.push(to8(plane[i])); + } + } + Image::from_u8(w as u32, h as u32, layout, data) + } else { + let mut data = Vec::with_capacity(w * h * nc); + for i in 0..w * h { + for plane in &planes { + data.push(scale(plane[i])); + } + } + Image::from_u16(w as u32, h as u32, layout, &data) + } +} + +#[cfg(test)] +mod tests { + use super::*; + + fn noise_u16(n: usize, seed: &mut u32, mask: u32) -> Vec { + (0..n) + .map(|_| { + *seed ^= *seed << 13; + *seed ^= *seed >> 17; + *seed ^= *seed << 5; + (*seed & mask) as u16 + }) + .collect() + } + + fn image(w: u32, h: u32, layout: ChannelLayout, wide: bool, kind: u8) -> Image { + let n = w as usize * h as usize * layout.channels(); + let mut seed = 0x1234_5679u32; + let samples: Vec = match kind { + 0 => noise_u16(n, &mut seed, if wide { 0xFFFF } else { 0xFF }), + 1 => vec![if wide { 40_000 } else { 100 }; n], + _ => (0..n).map(|i| if wide { ((i / 3) * 37 % 65_536) as u16 } else { ((i / 3) % 251) as u16 }).collect(), + }; + if wide { + Image::from_u16(w, h, layout, &samples).unwrap() + } else { + Image::from_u8(w, h, layout, samples.iter().map(|&v| v as u8).collect()).unwrap() + } + } + + fn plan(img: &Image) -> Plan { + Plan { layout: img.layout(), sample: img.sample_type() } + } + + #[test] + fn round_trips_noise_flat_and_smooth_images() { + for layout in [ChannelLayout::Gray, ChannelLayout::Rgb] { + for wide in [false, true] { + for kind in 0..3 { + // Widths around the run-length steps, and single-pixel dimensions. + for (w, h) in [(1, 1), (1, 9), (2, 3), (7, 5), (33, 4), (300, 3), (1025, 2)] { + let img = image(w, h, layout, wide, kind); + let bytes = encode(&img, plan(&img)).unwrap(); + assert!(looks_like_jpegls(&bytes)); + let back = decode(&bytes, &Limits::default()).unwrap_or_else(|e| panic!("{layout:?} wide={wide} kind={kind} {w}x{h}: {e}")); + assert_eq!(back.data(), img.data(), "{layout:?} wide={wide} kind={kind} {w}x{h}"); + } + } + } + } + } + + #[test] + fn default_parameters_follow_the_standard() { + let p = Params::new(8, 0, [0; 5]).unwrap(); + assert_eq!((p.maxval, p.range, p.qbpp, p.limit, p.t1, p.t2, p.t3, p.reset), (255, 256, 8, 32, 3, 7, 21, 64)); + let p = Params::new(16, 0, [0; 5]).unwrap(); + assert_eq!((p.maxval, p.range, p.qbpp, p.limit, p.t1, p.t2, p.t3), (65535, 65536, 16, 64, 18, 67, 276)); + let p = Params::new(8, 3, [0; 5]).unwrap(); + assert_eq!((p.range, p.t1), (38, 12)); // RANGE = floor((255 + 2 * 3) / 7) + 1 + assert!(Params::new(8, 200, [0; 5]).is_err()); + } + + #[test] + fn detection_tells_jpeg_ls_from_jpeg() { + let img = image(4, 4, ChannelLayout::Gray, false, 0); + let ls = encode(&img, plan(&img)).unwrap(); + assert!(looks_like_jpegls(&ls)); + // A baseline JPEG starts SOI, APP0 / DQT, ... SOF0: not JPEG-LS. + assert!(!looks_like_jpegls(&[0xFF, 0xD8, 0xFF, 0xE0, 0, 4, 0, 0, 0xFF, 0xC0, 0, 11])); + assert!(!looks_like_jpegls(&[0xFF, 0xD8, 0xFF, 0xC0, 0, 11])); + // SOF55 after an application segment still is. + assert!(looks_like_jpegls(&[0xFF, 0xD8, 0xFF, 0xE8, 0, 4, 0, 0, 0xFF, 0xF7, 0, 11])); + assert!(!looks_like_jpegls(&[0xFF, 0xD8])); + } + + #[test] + fn interleaved_scans_and_tables_are_unsupported() { + let img = image(8, 4, ChannelLayout::Rgb, false, 2); + let good = encode(&img, plan(&img)).unwrap(); + // The scan header body is Ns, Cs, Tm, NEAR, ILV, point transform, after the marker and length. + let sos = good.windows(2).position(|w| w == [0xFF, 0xDA]).unwrap() + 4; + let mut line = good.clone(); + line[sos + 4] = 1; + assert!(matches!(decode(&line, &Limits::default()), Err(CodecError::Unsupported { .. }))); + let mut mapped = good.clone(); + mapped[sos + 2] = 1; + assert!(matches!(decode(&mapped, &Limits::default()), Err(CodecError::Unsupported { .. }))); + let mut transform = good; + transform[sos + 5] = 1; + assert!(matches!(decode(&transform, &Limits::default()), Err(CodecError::Unsupported { .. }))); + } + + #[test] + fn hostile_input_is_an_error_not_a_panic() { + for (w, h, layout, wide) in [(23, 9, ChannelLayout::Gray, false), (11, 7, ChannelLayout::Rgb, true)] { + let img = image(w, h, layout, wide, 0); + let good = encode(&img, plan(&img)).unwrap(); + for len in 0..good.len() { + let _ = decode(&good[..len], &Limits::default()); + } + // Flip every bit of the entropy-coded data and of the headers, one at a time. + for i in 0..good.len() { + for bit in 0..8 { + let mut c = good.clone(); + c[i] ^= 1 << bit; + let _ = decode(&c, &Limits::default()); + } + } + } + // A huge frame is stopped by the limits or the length check before any allocation. + let img = image(4, 4, ChannelLayout::Gray, false, 0); + let mut huge = encode(&img, plan(&img)).unwrap(); + let sof = huge.windows(2).position(|w| w == [0xFF, 0xF7]).unwrap(); + huge[sof + 5..sof + 9].copy_from_slice(&[0xFF, 0xFF, 0xFF, 0xFF]); + assert!(decode(&huge, &Limits::default()).is_err()); + } +} diff --git a/crates/image-codecs/src/codecs/png.rs b/crates/image-codecs/src/codecs/png.rs index 4ceb75403..9965dfc30 100644 --- a/crates/image-codecs/src/codecs/png.rs +++ b/crates/image-codecs/src/codecs/png.rs @@ -141,12 +141,12 @@ impl> PngDecodeRows { /// Reads the header and refuses what [`decode`] refuses before any pixel is decoded: sizes /// past `limits`, and a file too small for the image it declares. pub fn new(bytes: B, limits: &Limits) -> Result { - Self::open(bytes, limits, true) + Self::open(bytes, limits, true, false) } /// [`PngDecodeRows::new`], optionally without reserving the whole image: a caller that hands /// the rows on as they come ([`decode_rows`]) never holds them all. - fn open(bytes: B, limits: &Limits, reserve: bool) -> Result { + fn open(bytes: B, limits: &Limits, reserve: bool, bounded_stream: bool) -> Result { let at = std::sync::Arc::new(std::sync::atomic::AtomicU64::new(0)); let bytes = SharedInput(std::sync::Arc::new(bytes)); let input = Tracked { cursor: Cursor::new(bytes.clone()), at: at.clone() }; @@ -187,7 +187,14 @@ impl> PngDecodeRows { let (width, height) = info.frame_control.map_or_else(|| info.size(), |f| (f.width, f.height)); let line = reader.output_line_size(width).ok_or_else(|| err("image too large"))?; // The output layout is known: charge its real bytes per pixel, not RGBA16's worst case. - limits.check(width, height, layout, sample)?; + if bounded_stream { + limits.check_size(width, height)?; + if line > limits.alloc_usize() { + return Err(CodecError::LimitExceeded("one PNG row exceeds max_alloc".into())); + } + } else { + limits.check(width, height, layout, sample)?; + } let height = usize::try_from(height).map_err(|_| err("image too large"))?; if width == 0 || height == 0 { return Err(err("the image is empty")); @@ -234,11 +241,35 @@ impl> PngDecodeRows { /// decode the whole image after all (an interlaced file) calls /// [`PngDecodeRows::reserve_output`] before [`PngDecodeRows::step`]. pub fn new_streaming(bytes: B, limits: &Limits) -> Result { - Self::open(bytes, limits, false) + Self::open(bytes, limits, false, false) + } + + /// Opens a bounded row stream without charging a full image that is never retained. + /// Dimension, pixel-count, plausibility, decoder and metadata limits remain unchanged; + /// `max_alloc` also bounds each row group. It is not a total process-memory limit. + /// Unlike [`Self::new_streaming`], a streamable image may exceed the full-image budget. + /// Non-streamable images still require that budget, as do [`Self::step`] and + /// [`Self::reserve_output`]. The caller must bound any output it retains in its sink. + pub fn new_bounded_streaming(bytes: B, limits: &Limits) -> Result { + let decoder = Self::open(bytes, limits, false, true)?; + if !decoder.streamable() { + decoder.check_full_output()?; + } + Ok(decoder) + } + + fn check_full_output(&self) -> Result<(), CodecError> { + let height = u32::try_from(self.height).map_err(|_| err("image too large"))?; + self.limits.check(self.width, height, self.layout, self.sample)?; + if self.size > self.limits.alloc_usize() { + return Err(CodecError::LimitExceeded("PNG output exceeds max_alloc".into())); + } + Ok(()) } /// Reserves the buffer [`PngDecodeRows::step`] decodes the whole image into. pub fn reserve_output(&mut self) -> Result<(), CodecError> { + self.check_full_output()?; let size = self.size; self.buf.try_reserve_exact(size).map_err(|_| CodecError::LimitExceeded(format!("cannot allocate {size} bytes for the image"))) } @@ -271,6 +302,7 @@ impl> PngDecodeRows { /// Decodes at least one row (the whole image if it is interlaced), then goes on while `more()` /// returns true. `Some` once the image is complete, after which the decoder is spent. pub fn step(&mut self, more: &mut dyn FnMut() -> bool) -> Result, CodecError> { + self.check_full_output()?; if self.done { return Err(err("the image was already decoded")); } @@ -338,7 +370,9 @@ impl> PngDecodeRows { /// `more()` returns true; `Some` with the extras once the file is read to its end, with the /// resolution a format that lacks one takes from its EXIF or XMP. A damaged file fails with /// the error [`decode`] gives, after however many rows were delivered. An orientation (an - /// `eXIf` after the image data may set one) is for the caller to apply. + /// `eXIf` after the image data may set one) is for the caller to apply. A new group size + /// must exceed any partial group retained from an earlier yield; otherwise this returns + /// an error without consuming another row. pub fn step_stream>( &mut self, group: u32, @@ -348,12 +382,29 @@ impl> PngDecodeRows { if self.done { return Err(err("the image was already decoded").into()); } - let group = usize::try_from(group.max(1)).unwrap_or(usize::MAX); + // Tail EXIF may report an orientation after rows have already been delivered; + // that belongs in the returned extras, not a rejection of metadata continuation. + if self.rows == 0 && !self.tail && !self.streamable() { + return Err(err("this PNG requires whole-image decoding").into()); + } + let group = usize::try_from(group.max(1)).unwrap_or(usize::MAX).min(self.height); + if self.rows - self.group_start >= group { + return Err(err("PNG row group is smaller than rows held from the previous step").into()); + } + let group_bytes = group.checked_mul(self.line).ok_or_else(|| err("PNG row group is too large"))?; + if group_bytes > self.limits.alloc_usize() { + return Err(CodecError::LimitExceeded("PNG row group exceeds max_alloc".into()).into()); + } if !self.tail { while self.rows < self.height { let held = self.rows - self.group_start; let at = held.checked_mul(self.line).ok_or_else(|| err("image too large"))?; let end = at.checked_add(self.line).ok_or_else(|| err("image too large"))?; + // A caller may change its group size between yields. Bound the actual held + // bytes as well as the requested group before growing the row buffer. + if end > self.limits.alloc_usize() { + return Err(CodecError::LimitExceeded("PNG row group exceeds max_alloc".into()).into()); + } if self.buf.len() < end { self.buf .try_reserve_exact(end - self.buf.len()) @@ -368,7 +419,7 @@ impl> PngDecodeRows { to_native_endian(row); } self.rows += 1; - if self.rows - self.group_start == group || self.rows == self.height { + if self.rows - self.group_start >= group || self.rows == self.height { let data = self.buf.get(..end).ok_or_else(|| err("invalid PNG row group"))?; sink(u32::try_from(self.group_start).map_err(|_| err("image too large"))?, data)?; self.group_start = self.rows; diff --git a/crates/image-codecs/src/format.rs b/crates/image-codecs/src/format.rs index ecaf945cd..810c1ad48 100644 --- a/crates/image-codecs/src/format.rs +++ b/crates/image-codecs/src/format.rs @@ -44,6 +44,8 @@ pub enum Format { Iff, /// JPEG 2000 (JP2 container or raw J2K codestream): read-only; see [`ASYMMETRIC_EXCEPTIONS`]. Jp2, + /// JPEG-LS (`.jls`): lossless and near-lossless gray and RGB. + JpegLs, } /// What a format can hold **and** what this crate reads/writes for it. @@ -113,7 +115,7 @@ const TIFF_LAYOUTS: &[ChannelLayout] = &[L::Gray, L::GrayA, L::Rgb, L::Rgba, L:: impl Format { /// Every format known to the crate (enabled or not). - pub const ALL: [Format; 24] = [ + pub const ALL: [Format; 25] = [ Format::Png, Format::Jpeg, Format::Jp2, @@ -138,6 +140,7 @@ impl Format { Format::Cineon, Format::Dicom, Format::Iff, + Format::JpegLs, ]; pub fn caps(self) -> FormatCaps { @@ -176,6 +179,7 @@ impl Format { Format::Dicom => "DICOM", Format::Iff => "Amiga IFF", Format::Jp2 => "JPEG 2000", + Format::JpegLs => "JPEG-LS", } } @@ -207,6 +211,7 @@ impl Format { Format::Dicom => &["dcm", "dicom", "dc3"], Format::Iff => &["iff", "lbm", "ilbm"], Format::Jp2 => &["jp2", "j2k", "j2c", "jpf", "jpx"], + Format::JpegLs => &["jls"], } } @@ -236,13 +241,14 @@ impl Format { Format::Dicom => "application/dicom", Format::Iff => "image/x-iff", Format::Jp2 => "image/jp2", + Format::JpegLs => "image/jls", } } /// Maximum encodable dimensions, if the format has a hard limit. pub fn max_dimensions(self) -> Option<(u32, u32)> { match self { - Format::Jpeg | Format::Gif | Format::Iff => Some((65535, 65535)), + Format::Jpeg | Format::Gif | Format::Iff | Format::JpegLs => Some((65535, 65535)), Format::WebP => Some((16384, 16384)), Format::Ico => Some((256, 256)), _ => None, @@ -310,6 +316,7 @@ pub fn caps(format: Format) -> FormatCaps { Format::Cineon => FormatCaps { depths: &[S::U16], layouts: &[L::Gray, L::Rgb], alpha: false, ..base }, Format::Dicom => FormatCaps { read: true, write: false, depths: &[S::U8, S::U16], layouts: &[L::Gray, L::Rgb], alpha: false, ..base }, Format::Iff => FormatCaps { depths: &[S::U8], layouts: &[L::Rgb, L::Rgba], alpha: true, ..base }, + Format::JpegLs => FormatCaps { depths: &[S::U8, S::U16], layouts: &[L::Gray, L::Rgb], alpha: false, ..base }, Format::Jp2 => FormatCaps { read: true, write: false, layouts: &[L::Gray, L::GrayA, L::Rgb, L::Rgba, L::Cmyk, L::CmykA], icc: true, ..base }, } } @@ -321,6 +328,9 @@ pub fn detect(bytes: &[u8]) -> Option { if b.starts_with(b"\x89PNG\r\n\x1a\n") { return Some(Format::Png); } + if crate::codecs::jpegls::looks_like_jpegls(b) { + return Some(Format::JpegLs); + } if b.starts_with(&[0xFF, 0xD8, 0xFF]) { return Some(Format::Jpeg); } diff --git a/crates/image-codecs/src/lib.rs b/crates/image-codecs/src/lib.rs index 6516b9ff2..adeed0825 100644 --- a/crates/image-codecs/src/lib.rs +++ b/crates/image-codecs/src/lib.rs @@ -16,6 +16,7 @@ mod allocation; pub mod apng; +pub mod bmp_probe; pub mod cicp; mod codecs; mod error; @@ -28,6 +29,7 @@ mod inspection; mod lab; mod options; pub mod orientation; +pub mod png_probe; pub mod resolution; pub mod source_settings; pub mod web; @@ -63,7 +65,7 @@ pub use crate::resolution::{exif_resolution, export_exif, export_xmp, photoshop_ pub use crate::source_settings::{WebpEncoding, jpeg_quality, webp_encoding}; pub use half::f16; -use crate::codecs::{dds, dicom, dpx, exr, fits, heif, iff, jp2, jpeg, jxl, pcx, png, pnm, tiff, via_image, webp, xpm}; +use crate::codecs::{dds, dicom, dpx, exr, fits, heif, iff, jp2, jpeg, jpegls, jxl, pcx, png, pnm, tiff, via_image, webp, xpm}; /// Detect the format and decode with default [`Limits`]. pub fn decode(bytes: &[u8]) -> Result { @@ -102,6 +104,7 @@ pub fn decode_as_with(format: Format, bytes: &[u8], opts: &DecodeOptions) -> Res Format::Cineon => dpx::decode_cineon(bytes, l), Format::Dicom => dicom::decode(bytes, l), Format::Iff => iff::decode(bytes, l), + Format::JpegLs => jpegls::decode(bytes, l), Format::OpenExr => exr::decode(bytes, l), Format::Heif => heif::decode(bytes, l, opts.keep_orientation), Format::Jp2 => jp2::decode(bytes, l), @@ -458,6 +461,7 @@ pub fn encode(image: &Image, format: Format, opts: &EncodeOptions) -> Result dpx::encode_cineon(image, plan), Format::Dicom => Err(CodecError::unsupported(format, "DICOM is import-only; save as TIFF or PNG instead")), Format::Iff => iff::encode(image, plan), + Format::JpegLs => jpegls::encode(image, plan), Format::OpenExr => exr::encode(image, plan, opts), Format::Heif => Err(CodecError::unsupported(format, "encoding is not available in this build")), Format::Jp2 => Err(CodecError::unsupported(format, "JPEG 2000 is import-only; save as PNG or TIFF instead")), diff --git a/crates/image-codecs/src/png_probe.rs b/crates/image-codecs/src/png_probe.rs new file mode 100644 index 000000000..7edda454d --- /dev/null +++ b/crates/image-codecs/src/png_probe.rs @@ -0,0 +1,190 @@ +//! Header-only PNG inspection without inflating image data. +use crate::{CodecError, Format}; +fn err(message: &str) -> CodecError { + CodecError::malformed(Format::Png, message) +} + +/// Dimensions declared in a PNG IHDR prefix, without checking the rest of the file. +/// Useful only for an early allocation limit; use [`header`] for structural inspection. +pub fn declared_dimensions(b: &[u8]) -> Option<(u32, u32)> { + if !b.starts_with(b"\x89PNG\r\n\x1a\n") || b.get(12..16) != Some(b"IHDR") { + return None; + } + let w = u32::from_be_bytes(b.get(16..20)?.try_into().ok()?); + let h = u32::from_be_bytes(b.get(20..24)?.try_into().ok()?); + Some((w, h)) +} + +/// Stored dimensions and EXIF orientation. This does not validate CRCs or compressed pixels. +pub fn header(b: &[u8]) -> Result<(u32, u32, u16), CodecError> { + header_with_max_pixels(b, u64::MAX) +} + +/// As [`header`], refusing images over `max_pixels` before scanning their chunks. +pub fn header_with_max_pixels(b: &[u8], max_pixels: u64) -> Result<(u32, u32, u16), CodecError> { + if !b.starts_with(b"\x89PNG\r\n\x1a\n") { + return Err(err("bad signature")); + } + let be32 = |p: usize| b.get(p..p.checked_add(4)?).map(|v| u32::from_be_bytes([v[0], v[1], v[2], v[3]])); + if be32(8) != Some(13) || b.get(12..16) != Some(b"IHDR") { + return Err(err("missing IHDR")); + } + let ihdr = b.get(16..29).ok_or_else(|| err("truncated IHDR"))?; + let (w, h) = (u32::from_be_bytes([ihdr[0], ihdr[1], ihdr[2], ihdr[3]]), u32::from_be_bytes([ihdr[4], ihdr[5], ihdr[6], ihdr[7]])); + let (depth, color, compression, filter, interlace) = (ihdr[8], ihdr[9], ihdr[10], ihdr[11], ihdr[12]); + if w > i32::MAX as u32 || h > i32::MAX as u32 { + return Err(err("dimensions out of range")); + } + if w == 0 || h == 0 { + return Err(err("zero dimension")); + } + if u64::from(w) * u64::from(h) > max_pixels { + return Err(CodecError::LimitExceeded("PNG pixel count".into())); + } + let depth_ok = match color { + 0 => matches!(depth, 1 | 2 | 4 | 8 | 16), + 3 => matches!(depth, 1 | 2 | 4 | 8), + 2 | 4 | 6 => matches!(depth, 8 | 16), + _ => false, + }; + if !depth_ok || compression != 0 || filter != 0 || interlace > 1 { + return Err(err("invalid IHDR")); + } + let (mut p, mut palette, mut idat, mut after_idat, mut exif) = (8usize, false, false, false, None); + while let Some(len) = be32(p) { + let Some(kind) = p.checked_add(4).and_then(|k| b.get(k..k.checked_add(4)?)) else { break }; + if idat && kind != b"IDAT" { + // a chunk follows the image data: the IDAT run is complete + after_idat = true; + } + if kind == b"IEND" { + break; + } + let start = p.saturating_add(8); + let data = (len <= i32::MAX as u32).then(|| start.checked_add(len as usize).and_then(|end| b.get(start..end))).flatten(); + let Some(data) = data else { + if after_idat { + break; // a damaged chunk after the image data: a decode ignores it too + } + return Err(err("truncated")); + }; + match kind { + b"PLTE" => palette = true, + b"IDAT" if after_idat => {} // a stray IDAT after other chunks: the first run is the image + b"IDAT" => { + if color == 3 && !palette { + return Err(err("indexed image without a palette")); + } + idat = true; + } + b"eXIf" if exif.is_none() => exif = Some(data.strip_prefix(b"Exif\0\0").unwrap_or(data)), + _ => {} + } + p = start.saturating_add(data.len()).saturating_add(4); + } + if !idat { + return Err(err("no image data")); + } + if !after_idat { + return Err(err("truncated image data")); + } + Ok((w, h, exif.map_or(1, crate::exif_orientation))) +} + +/// Container color declarations preceding image data. Raw values leave color policy to callers. +#[derive(Clone, Debug, Default, PartialEq)] +pub struct ColorChunks { + pub srgb: bool, + pub gamma: Option, + pub chromaticities: Option<[f64; 8]>, +} + +/// Read bounded color chunks before image data; malformed tails are ignored. +pub fn color_chunks(b: &[u8]) -> Option { + if !b.starts_with(b"\x89PNG\r\n\x1a\n") { + return None; + } + let mut result = ColorChunks::default(); + let mut p = 8usize; + while let Some(length) = b.get(p..p.checked_add(4)?) { + let len = u32::from_be_bytes(length.try_into().ok()?) as usize; + let start = p.checked_add(8)?; + let Some(kind) = b.get(p.checked_add(4)?..start) else { break }; + let Some(payload) = start.checked_add(len).and_then(|end| b.get(start..end)) else { break }; + match kind { + b"sRGB" => result.srgb = true, + b"gAMA" if payload.len() == 4 => { + let v = u32::from_be_bytes(payload.try_into().ok()?) as f32 / 100_000.0; + if v > 0.0 && v.is_finite() { + result.gamma = Some(v); + } + } + b"cHRM" if payload.len() == 32 => { + let mut vals = [0.0; 8]; + for (i, dst) in vals.iter_mut().enumerate() { + let raw = i32::from_be_bytes(payload[i * 4..i * 4 + 4].try_into().ok()?); + *dst = f64::from(raw) / 100_000.0; + } + result.chromaticities = Some(vals); + } + b"IDAT" | b"IEND" => break, + _ => {} + } + let Some(next) = start.checked_add(len).and_then(|end| end.checked_add(4)) else { break }; + p = next; + } + Some(result) +} + +#[cfg(test)] +mod tests { + use super::*; + + fn chunk(out: &mut Vec, kind: &[u8; 4], data: &[u8]) { + out.extend_from_slice(&(data.len() as u32).to_be_bytes()); + out.extend_from_slice(kind); + out.extend_from_slice(data); + out.extend_from_slice(&[0; 4]); // Header inspection deliberately does not validate CRCs. + } + + fn png(color: u8) -> Vec { + let mut bytes = b"\x89PNG\r\n\x1a\n".to_vec(); + let mut ihdr = Vec::new(); + ihdr.extend_from_slice(&2u32.to_be_bytes()); + ihdr.extend_from_slice(&3u32.to_be_bytes()); + ihdr.extend_from_slice(&[8, color, 0, 0, 0]); + chunk(&mut bytes, b"IHDR", &ihdr); + bytes + } + + #[test] + fn header_checks_structure_and_reads_late_orientation_without_crc_validation() { + let mut bytes = png(2); + chunk(&mut bytes, b"IDAT", &[1]); + let exif = b"II\x2a\x00\x08\x00\x00\x00\x01\x00\x12\x01\x03\x00\x01\x00\x00\x00\x06\x00\x00\x00\x00\x00\x00\x00"; + chunk(&mut bytes, b"eXIf", exif); + chunk(&mut bytes, b"IEND", &[]); + assert_eq!(header(&bytes).ok(), Some((2, 3, 6))); + assert!(matches!(header_with_max_pixels(&bytes, 5), Err(CodecError::LimitExceeded(_)))); + assert_eq!(declared_dimensions(&bytes), Some((2, 3))); + assert!(header(&bytes[..34]).is_err()); + let mut indexed = png(3); + chunk(&mut indexed, b"IDAT", &[1]); + chunk(&mut indexed, b"IEND", &[]); + assert!(header(&indexed).is_err()); + indexed[25] = 7; + assert!(header(&indexed).is_err()); + } + + #[test] + fn reads_color_chunks_only_before_image_data() { + let mut bytes = png(2); + chunk(&mut bytes, b"gAMA", &45455u32.to_be_bytes()); + chunk(&mut bytes, b"IDAT", &[1]); + chunk(&mut bytes, b"sRGB", &[0]); + let color = color_chunks(&bytes).unwrap(); + assert_eq!(color.gamma, Some(0.45455)); + assert!(!color.srgb); + assert!(color_chunks(b"not png").is_none()); + } +} diff --git a/crates/image-codecs/tests/it/detect_caps.rs b/crates/image-codecs/tests/it/detect_caps.rs index 640772c83..390eaa41f 100644 --- a/crates/image-codecs/tests/it/detect_caps.rs +++ b/crates/image-codecs/tests/it/detect_caps.rs @@ -62,6 +62,7 @@ fn default_build_formats_all_symmetric() { Format::Dpx, Format::Cineon, Format::Iff, + Format::JpegLs, ]; for f in expected { assert!(caps(f).read && caps(f).write, "{f:?}"); @@ -218,6 +219,21 @@ fn detect_magic_iff() { assert_eq!(detect(b"FORM\0\0\0\x14AIFF"), None); assert!(decode_as(Format::Iff, b"FORM\0\0\0\x14ILBM").is_err()); } +#[test] +fn detect_magic_jpeg_ls() { + use craft_image_codecs::{ChannelLayout, EncodeOptions, Image, encode}; + let img = Image::from_u8(5, 3, ChannelLayout::Gray, (0..15).collect()).unwrap(); + let bytes = encode(&img, Format::JpegLs, &EncodeOptions::default()).unwrap(); + assert_eq!(detect(&bytes), Some(Format::JpegLs)); + assert_eq!(from_extension("scan.JLS"), Some(Format::JpegLs)); + assert_eq!(decode(&bytes).unwrap().data(), img.data()); + assert!(!ASYMMETRIC_EXCEPTIONS.iter().any(|(f, _)| *f == Format::JpegLs)); + assert_eq!(Format::JpegLs.max_dimensions(), Some((65535, 65535))); + // Plain JPEG still is JPEG: SOI, then a baseline frame header. + assert_eq!(detect(&[0xFF, 0xD8, 0xFF, 0xC0, 0, 11]), Some(Format::Jpeg)); + assert!(decode_as(Format::JpegLs, &[0xFF, 0xD8, 0xFF, 0xF7]).is_err()); +} + #[test] fn detect_magic_avif() { assert_eq!(detect(b"\0\0\0\x1cftypavif\0\0\0\0avifmif1miaf"), Some(Format::Avif)); diff --git a/crates/image-codecs/tests/it/jpegls_oracle.rs b/crates/image-codecs/tests/it/jpegls_oracle.rs new file mode 100644 index 000000000..aa98609a1 --- /dev/null +++ b/crates/image-codecs/tests/it/jpegls_oracle.rs @@ -0,0 +1,103 @@ +//! JPEG-LS against ffmpeg as the reference (oracle; skipped when ffmpeg is absent, a failure with +//! `CRAFT_MEDIA_REQUIRE_ORACLES=1`): +//! - files ffmpeg's encoder writes (8- and 16-bit gray: smooth, noisy and flat images, which +//! exercise the regular, run and escape-coded paths) decode to the pixels ffmpeg decodes; +//! - files our encoder writes (gray 8 and 16 bits, RGB 8 bits) decode in ffmpeg to our pixels (ffmpeg +//! cannot decode 16-bit RGB JPEG-LS, so that case is only round-tripped through our own decoder); +//! - ffmpeg writes RGB line-interleaved, which is reported as unsupported. + +use craft_image_codecs::{ChannelLayout, CodecError, EncodeOptions, Format, Image, SampleType, decode, encode}; +use std::path::Path; +use std::process::Command; + +const W: u32 = 66; +const H: u32 = 40; + +fn run(ffmpeg: &Path, args: &[&str]) -> Vec { + let out = Command::new(ffmpeg).args(["-v", "error", "-y"]).args(args).output().expect("run ffmpeg"); + assert!(out.status.success(), "ffmpeg {args:?}: {}", String::from_utf8_lossy(&out.stderr)); + out.stdout +} + +/// `(name, lavfi source)`: smooth, noisy and flat pictures of W x H. +fn sources() -> Vec<(&'static str, String)> { + let s = format!("{W}x{H}"); + vec![ + ("testsrc", format!("testsrc2=size={s}:rate=1")), + ("noise", format!("nullsrc=size={s}:rate=1,geq=lum='random(1)*255':cb=128:cr=128")), + ("flat", format!("color=c=0x336699:size={s}:rate=1")), + ("gradient", format!("gradients=size={s}:rate=1:seed=7")), + ] +} + +#[test] +fn ffmpeg_files_decode_to_ffmpegs_pixels() { + let Some(ffmpeg) = craft_media_testkit::ffmpeg() else { return }; + let dir = craft_media_testkit::fixture_subdir("jpegls"); + for (name, src) in sources() { + for (fmt, ext_bits) in [("gray", 8), ("gray16le", 16)] { + let file = dir.join(format!("ffmpeg_{name}_{ext_bits}.jls")); + let f = file.to_str().expect("path"); + run(&ffmpeg, &["-f", "lavfi", "-i", &src, "-frames:v", "1", "-pix_fmt", fmt, f]); + let theirs = run(&ffmpeg, &["-i", f, "-f", "rawvideo", "-pix_fmt", fmt, "-"]); + let ours = decode(&std::fs::read(&file).expect("read")).unwrap_or_else(|e| panic!("{name} {fmt}: {e}")); + assert_eq!((ours.width(), ours.height(), ours.layout()), (W, H, ChannelLayout::Gray), "{name} {fmt}"); + let want_sample = if ext_bits == 8 { SampleType::U8 } else { SampleType::U16 }; + assert_eq!(ours.sample_type(), want_sample); + let ours_bytes: Vec = if ext_bits == 8 { + ours.data().to_vec() + } else { + ours.data().as_chunks::<2>().0.iter().flat_map(|b| u16::from_ne_bytes(*b).to_le_bytes()).collect() + }; + assert!(ours_bytes == theirs, "{name} {fmt}: pixels differ from ffmpeg's decode"); + } + } +} + +#[test] +fn ffmpegs_rgb_is_line_interleaved_and_unsupported() { + let Some(ffmpeg) = craft_media_testkit::ffmpeg() else { return }; + let dir = craft_media_testkit::fixture_subdir("jpegls"); + let file = dir.join("ffmpeg_rgb.jls"); + run( + &ffmpeg, + &["-f", "lavfi", "-i", &format!("testsrc2=size={W}x{H}:rate=1"), "-frames:v", "1", "-pix_fmt", "rgb24", file.to_str().expect("path")], + ); + assert!(matches!(decode(&std::fs::read(&file).expect("read")), Err(CodecError::Unsupported { .. }))); +} + +#[test] +fn ffmpeg_decodes_our_files_to_our_pixels() { + let Some(ffmpeg) = craft_media_testkit::ffmpeg() else { return }; + let dir = craft_media_testkit::fixture_subdir("jpegls"); + for (name, src) in sources() { + for (layout, fmt8, fmt16) in [(ChannelLayout::Gray, "gray", "gray16le"), (ChannelLayout::Rgb, "rgb24", "rgb48le")] { + for bits in [8, 16] { + let fmt = if bits == 8 { fmt8 } else { fmt16 }; + // Source pixels from ffmpeg, so the test image is realistic. + let raw = run(&ffmpeg, &["-f", "lavfi", "-i", &src, "-frames:v", "1", "-pix_fmt", fmt, "-f", "rawvideo", "-"]); + let img = if bits == 8 { + Image::from_u8(W, H, layout, raw).expect("image") + } else { + let samples: Vec = raw.as_chunks::<2>().0.iter().map(|b| u16::from_le_bytes(*b)).collect(); + Image::from_u16(W, H, layout, &samples).expect("image") + }; + let bytes = encode(&img, Format::JpegLs, &EncodeOptions::default()).expect("encode"); + assert_eq!(decode(&bytes).expect("our own file decodes").data(), img.data(), "{name} {fmt}: round trip"); + let file = dir.join(format!("ours_{name}_{fmt}.jls")); + std::fs::write(&file, &bytes).expect("write"); + // ffmpeg's decoder does not implement 16-bit colour ("Not yet implemented"). + if layout == ChannelLayout::Rgb && bits == 16 { + continue; + } + let theirs = run(&ffmpeg, &["-i", file.to_str().expect("path"), "-f", "rawvideo", "-pix_fmt", fmt, "-"]); + let ours_bytes: Vec = if bits == 8 { + img.data().to_vec() + } else { + img.data().as_chunks::<2>().0.iter().flat_map(|b| u16::from_ne_bytes(*b).to_le_bytes()).collect() + }; + assert!(theirs == ours_bytes, "{name} {fmt}: ffmpeg decodes our file to other pixels"); + } + } + } +} diff --git a/crates/image-codecs/tests/it/main.rs b/crates/image-codecs/tests/it/main.rs index f465a0cbb..4eda74a26 100644 --- a/crates/image-codecs/tests/it/main.rs +++ b/crates/image-codecs/tests/it/main.rs @@ -14,6 +14,7 @@ mod exr_stream; mod fidelity; mod gif_frames; mod heif; +mod jpegls_oracle; mod limits; mod malformed; mod metadata; diff --git a/crates/image-codecs/tests/it/png_jpeg.rs b/crates/image-codecs/tests/it/png_jpeg.rs index 6b213ea83..07c57bf24 100644 --- a/crates/image-codecs/tests/it/png_jpeg.rs +++ b/crates/image-codecs/tests/it/png_jpeg.rs @@ -665,3 +665,97 @@ fn step_stream_gives_the_pixels_and_extras_of_decode() { fn dec_line(w: u32, layout: ChannelLayout, sample: SampleType) -> usize { w as usize * layout.channels() * sample.bytes() } + +#[test] +fn bounded_png_stream_does_not_charge_unretained_full_frame() { + let image = synth(256, 1024, ChannelLayout::Rgba, SampleType::U16, 77, 0.5); + let bytes = encode(&image, Format::Png, &EncodeOptions::default()).unwrap(); + let limits = Limits { max_alloc: 512 * 1024, ..Limits::default() }; + assert!(image.data().len() > limits.max_alloc as usize); + assert!(PngDecodeRows::new(bytes.as_slice(), &limits).is_err()); + assert!(PngDecodeRows::new_streaming(bytes.as_slice(), &limits).is_err()); + let mut decoder = PngDecodeRows::new_bounded_streaming(bytes.as_slice(), &limits).unwrap(); + assert!(decoder.streamable()); + // Opening a stream must not become a way to bypass whole-image allocation guards. + assert!(decoder.reserve_output().is_err()); + assert!(decoder.step(&mut || true).is_err()); + let mut next = 0; + let mut sink = |first: u32, rows: &[u8]| -> Result<(), CodecError> { + assert_eq!(first as usize, next); + assert!(rows.len() <= 3 * 256 * 8); + let start = next * 256 * 8; + assert_eq!(rows, &image.data()[start..start + rows.len()]); + next += rows.len() / (256 * 8); + Ok(()) + }; + while decoder.step_stream(3, &mut || false, &mut sink).unwrap().is_none() {} + assert_eq!(next, 1024); +} + +#[test] +fn bounded_png_stream_keeps_group_dimension_and_fallback_limits() { + let image = synth(256, 1024, ChannelLayout::Rgba, SampleType::U16, 81, 0.5); + let bytes = encode(&image, Format::Png, &EncodeOptions::default()).unwrap(); + let limits = Limits { max_alloc: 512 * 1024, ..Limits::default() }; + let mut decoder = PngDecodeRows::new_bounded_streaming(bytes.as_slice(), &limits).unwrap(); + assert!(decoder.step_stream::(1024, &mut || true, &mut |_, _| panic!("oversized group reached sink")).is_err()); + for strict in [ + Limits { max_width: 255, ..limits }, + Limits { max_height: 1023, ..limits }, + Limits { max_pixels: 256 * 1024 - 1, ..limits }, + Limits { max_alloc: 1024, ..limits }, + ] { + assert!(PngDecodeRows::new_bounded_streaming(bytes.as_slice(), &strict).is_err()); + } + let adam7 = interlaced(&image); + assert!(PngDecodeRows::new_bounded_streaming(adam7.as_slice(), &limits).is_err()); +} + +#[test] +fn bounded_png_stream_rejects_shrinking_below_buffered_rows() { + let image = synth(16, 8, ChannelLayout::Rgba, SampleType::U16, 83, 0.5); + let bytes = encode(&image, Format::Png, &EncodeOptions::default()).unwrap(); + let mut decoder = PngDecodeRows::new_bounded_streaming(bytes.as_slice(), &Limits::default()).unwrap(); + for _ in 0..3 { + assert!(decoder.step_stream::(4, &mut || false, &mut |_, _| panic!("group is incomplete")).unwrap().is_none()); + } + assert!(decoder.step_stream::(1, &mut || true, &mut |_, _| panic!("changed group reached sink")).is_err()); + let mut output = Vec::new(); + decoder + .step_stream::(4, &mut || true, &mut |_, rows| { + output.extend_from_slice(rows); + Ok(()) + }) + .unwrap(); + assert_eq!(output, image.data()); +} + +#[test] +fn bounded_png_stream_preserves_late_exif_across_metadata_yields() { + let image = synth(16, 8, ChannelLayout::Rgba, SampleType::U16, 89, 0.5); + let bytes = encode(&image, Format::Png, &EncodeOptions::default()).unwrap(); + let mut exif = sample_exif(); + exif[18..22].copy_from_slice(&6u32.to_le_bytes()); + let mut chunks = vec![png_chunk(b"eXIf", &exif)]; + for i in 0..150 { + chunks.push(png_chunk(b"tEXt", format!("key{i}\0value").as_bytes())); + } + let bytes = with_chunks_before_iend(&bytes, &chunks); + let mut decoder = PngDecodeRows::new_bounded_streaming(bytes.as_slice(), &Limits::default()).unwrap(); + assert!(decoder.streamable()); + let mut output = Vec::new(); + let extras = loop { + if let Some(extras) = decoder + .step_stream::(2, &mut || false, &mut |_, rows| { + output.extend_from_slice(rows); + Ok(()) + }) + .unwrap() + { + break extras; + } + }; + assert_eq!(output, image.data()); + assert_eq!(extras.orientation, 6); + assert_eq!(extras.meta.text.len(), 150); +} diff --git a/crates/jsonrpc/src/lib.rs b/crates/jsonrpc/src/lib.rs index 9f8128d1d..cfb46ef79 100644 --- a/crates/jsonrpc/src/lib.rs +++ b/crates/jsonrpc/src/lib.rs @@ -195,3 +195,61 @@ where let responses: Vec<_> = batch.iter().filter_map(|v| one(v, limits, &mut on_request, &mut on_notification)).collect(); Ok((!responses.is_empty()).then_some(Value::Array(responses))) } +/// Process a permissive legacy JSON-RPC envelope. This mode accepts arbitrary IDs and +/// parameter shapes, ignores response envelopes, and retains legacy diagnostic messages. +/// The caller supplies dispatch and notification callbacks, and should bound transport input. +pub fn process_legacy(value: &Value, mut on_request: F, mut on_notification: N) -> Option +where + F: FnMut(Request<'_>) -> Result, + N: FnMut(Request<'_>), +{ + fn one(value: &Value, on_request: &mut F, on_notification: &mut N) -> Option + where + F: FnMut(Request<'_>) -> Result, + N: FnMut(Request<'_>), + { + let Some(object) = value.as_object() else { return Some(error_response(Value::Null, RpcError::new(-32600, "message must be an object"))) }; + let id = object.get("id"); + let Some(method) = object.get("method").and_then(Value::as_str) else { + if object.contains_key("result") || object.contains_key("error") { + return None; + } + return Some(error_response(id.cloned().unwrap_or(Value::Null), RpcError::new(-32600, "missing `method`"))); + }; + let request = Request { id, method, params: object.get("params") }; + let Some(id) = id else { + on_notification(request); + return None; + }; + Some(match on_request(request) { + Ok(result) => json!({"jsonrpc": "2.0", "id": id, "result": result}), + Err(error) => error_response(id.clone(), error), + }) + } + if let Some(batch) = value.as_array() { + if batch.is_empty() { + return Some(error_response(Value::Null, RpcError::new(-32600, "empty batch"))); + } + let replies: Vec<_> = batch.iter().filter_map(|v| one(v, &mut on_request, &mut on_notification)).collect(); + (!replies.is_empty()).then_some(Value::Array(replies)) + } else { + one(value, &mut on_request, &mut on_notification) + } +} + +#[cfg(test)] +mod legacy_tests { + use super::*; + #[test] + fn preserves_legacy_batch_routing_and_errors() { + let input = json!([{"id": [1], "method":"echo", "params":5}, {"method":"notice"}, {"result":1}, {}, 7]); + let seen = std::cell::Cell::new(false); + let result = process_legacy(&input, |r| Ok(r.params.cloned().unwrap_or(Value::Null)), |_| seen.set(true)).unwrap(); + assert!(seen.get()); + assert_eq!( + result, + json!([{"jsonrpc":"2.0","id":[1],"result":5}, {"jsonrpc":"2.0","id":null,"error":{"code":-32600,"message":"missing `method`"}}, {"jsonrpc":"2.0","id":null,"error":{"code":-32600,"message":"message must be an object"}}]) + ); + assert_eq!(process_legacy(&json!([]), |_| Ok(Value::Null), |_| {}).unwrap()["error"]["message"], "empty batch"); + } +} diff --git a/crates/mcp/src/lib.rs b/crates/mcp/src/lib.rs index 96cfdc7ae..6ca888fc0 100644 --- a/crates/mcp/src/lib.rs +++ b/crates/mcp/src/lib.rs @@ -34,11 +34,27 @@ pub struct Session { policy: Policy, state: State, capabilities: Value, + versions: Vec, + permissive_initialize: bool, + server_info: Option, + instructions: Option, + initialized: bool, } impl Session { pub fn new(name: impl Into, version: impl Into, policy: Policy) -> Self { - Self { name: name.into(), version: version.into(), policy, state: State::New, capabilities: json!({"tools": {}}) } + Self { + name: name.into(), + version: version.into(), + policy, + state: State::New, + capabilities: json!({"tools": {}}), + versions: vec![PROTOCOL_VERSION.into()], + permissive_initialize: false, + server_info: None, + instructions: None, + initialized: false, + } } /// Set the server capabilities returned at initialization. The value must be an object. @@ -50,6 +66,27 @@ impl Session { Ok(self) } + /// Configure a permissive legacy server: absent or malformed initialize parameters are + /// accepted, supported client versions are echoed, and unknown versions use the first entry. + pub fn with_legacy_initialize(mut self, versions: &[&str], server_info: Value) -> Self { + if !versions.is_empty() { + self.versions = versions.iter().map(|v| (*v).to_owned()).collect(); + } + self.permissive_initialize = true; + self.server_info = Some(server_info); + self + } + + /// Update connection-specific instructions returned by subsequent initialize requests. + pub fn set_instructions(&mut self, instructions: String) { + self.instructions = Some(instructions); + } + + /// Whether an initialized notification has been observed, including before initialize. + pub fn is_initialized(&self) -> bool { + self.initialized + } + pub fn state(&self) -> State { self.state } @@ -64,6 +101,18 @@ impl Session { if self.policy == Policy::Strict && (in_batch || self.state != State::New) { return Some(Err(RpcError::new(-32600, "Invalid Request"))); } + if self.permissive_initialize { + let asked = request.params.and_then(|p| p.get("protocolVersion")).and_then(Value::as_str); + let version = asked + .filter(|v| self.versions.iter().any(|supported| supported == v)) + .unwrap_or_else(|| self.versions.first().map(String::as_str).unwrap_or(PROTOCOL_VERSION)); + let mut result = json!({"protocolVersion": version, "capabilities": self.capabilities, "serverInfo": self.server_info}); + if let Some(instructions) = &self.instructions { + result["instructions"] = json!(instructions); + } + self.state = State::AwaitingInitialized; + return Some(Ok(result)); + } let Some(params) = request.params.and_then(Value::as_object) else { return Some(Err(invalid_initialize())); }; @@ -89,8 +138,11 @@ impl Session { /// Observe a JSON-RPC notification; notifications never receive replies. pub fn notification(&mut self, request: Request<'_>) { - if request.method == "notifications/initialized" && self.state == State::AwaitingInitialized { - self.state = State::Ready; + if request.method == "notifications/initialized" { + self.initialized = true; + if self.state == State::AwaitingInitialized { + self.state = State::Ready; + } } } } @@ -127,6 +179,22 @@ mod tests { Request { id: None, method, params } } + #[test] + fn legacy_initialization_negotiates_and_accepts_missing_parameters() { + let mut s = Session::new("test", "1", Policy::Compatible).with_legacy_initialize(&["new", "old"], json!({"name":"custom", "title":"Title"})); + s.notification(call("notifications/initialized", None)); + assert!(s.is_initialized()); + s.set_instructions("help".into()); + for (params, expected) in + [(None, "new"), (Some(json!({"protocolVersion":"old"})), "old"), (Some(json!({"protocolVersion":"unknown"})), "new")] + { + let result = s.request(call("initialize", params.as_ref()), false).unwrap().unwrap(); + assert_eq!(result["protocolVersion"], expected); + assert_eq!(result["serverInfo"]["title"], "Title"); + assert_eq!(result["instructions"], "help"); + } + } + #[test] fn compatible_negotiation_preserves_existing_clients() { let mut s = Session::new("test", "1", Policy::Compatible); diff --git a/crates/mime/src/base64_compat.rs b/crates/mime/src/base64_compat.rs new file mode 100644 index 000000000..452b276f9 --- /dev/null +++ b/crates/mime/src/base64_compat.rs @@ -0,0 +1,93 @@ +//! Compatibility decoding for legacy transports that accepted mixed Base64 alphabets and padding. +//! +//! This is deliberately permissive and must not be used for signatures, credentials or other +//! contexts requiring canonical Base64. Use [`crate::base64`] for strict RFC 4648 decoding. +use crate::base64::DecodeError; + +/// Decode standard or URL-safe Base64, accepting either alphabet even in the same input. +/// Spaces, CR, LF and `=` are ignored wherever they occur; an incomplete final sextet is +/// discarded. Other bytes are errors. Output allocation is bounded by the input length. +pub fn decode(input: impl AsRef<[u8]>) -> Result, DecodeError> { + let input = input.as_ref(); + let mut out = Vec::new(); + let capacity = (input.len() / 4) * 3 + (input.len() % 4) * 3 / 4; + out.try_reserve(capacity).map_err(|_| DecodeError)?; + let (mut acc, mut bits) = (0u32, 0u32); + for c in input { + let v = match c { + b'A'..=b'Z' => c - b'A', + b'a'..=b'z' => c - b'a' + 26, + b'0'..=b'9' => c - b'0' + 52, + b'+' | b'-' => 62, + b'/' | b'_' => 63, + b'=' | b'\n' | b'\r' | b' ' => continue, + _ => return Err(DecodeError), + }; + acc = (acc << 6) | u32::from(v); + bits += 6; + if bits >= 8 { + bits -= 8; + out.push((acc >> bits) as u8); + acc &= (1 << bits) - 1; + } + } + Ok(out) +} +#[cfg(test)] +mod tests { + #![allow(clippy::unwrap_used)] + use super::*; + #[test] + fn accepts_mixed_alphabets_and_ignores_padding_anywhere() { + assert_eq!(decode("+_8=").unwrap(), [251, 255]); + assert_eq!(decode("Z=m 9\nv\r").unwrap(), b"foo"); + assert_eq!(decode("A").unwrap(), b""); + assert_eq!(decode("AB").unwrap(), [0]); + assert!(decode("Zm9v\t").is_err()); + assert!(decode("Zm9v!").is_err()); + assert!(crate::base64::STANDARD.decode("Z=m9v").is_err()); + } + // Differential against CADCraft's pre-cutover decoder, including malformed legacy inputs. + fn legacy(s: &str) -> Option> { + let mut out = Vec::new(); + let (mut acc, mut bits) = (0u32, 0u32); + for c in s.bytes() { + let value = match c { + b'A'..=b'Z' => c - b'A', + b'a'..=b'z' => c - b'a' + 26, + b'0'..=b'9' => c - b'0' + 52, + b'+' | b'-' => 62, + b'/' | b'_' => 63, + b'=' | b'\n' | b'\r' | b' ' => continue, + _ => return None, + }; + acc = (acc << 6) | u32::from(value); + bits += 6; + if bits >= 8 { + bits -= 8; + out.push((acc >> bits) as u8); + } + } + Some(out) + } + #[test] + fn matches_the_legacy_decoder_on_seeded_mixed_inputs() { + let alphabet = b"ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/-_= \r\n!\t"; + let mut seed = 0x1234_5678u32; + for n in 0..50_000 { + let mut text = String::new(); + for _ in 0..n % 45 { + seed = seed.wrapping_mul(1_664_525).wrapping_add(1_013_904_223); + text.push(char::from(alphabet[(seed as usize) % alphabet.len()])); + } + assert_eq!(decode(&text).ok(), legacy(&text), "{text:?}"); + } + } + + #[test] + fn round_trip_rfc_vectors_with_the_shared_encoder() { + for data in [&b""[..], b"f", b"fo", b"foo", b"foobar", &[0, 255, 127]] { + assert_eq!(decode(crate::base64::STANDARD.encode(data)).unwrap(), data); + } + } +} diff --git a/crates/mime/src/lib.rs b/crates/mime/src/lib.rs index 72a930b8b..d3ff9f257 100644 --- a/crates/mime/src/lib.rs +++ b/crates/mime/src/lib.rs @@ -28,6 +28,7 @@ //! says) is the caller's: this crate has no notion of a mailbox or a document. pub mod base64; +pub mod base64_compat; pub mod charset; #[cfg(feature = "detect")] pub mod detect; diff --git a/crates/psd/src/text.rs b/crates/psd/src/text.rs index 6ba299bca..3a9e14952 100644 --- a/crates/psd/src/text.rs +++ b/crates/psd/src/text.rs @@ -150,3 +150,6 @@ mod tests { assert_eq!(t.style.size, Some(12.0)); } } + +/// Editable type-tool envelope with PhotoCraft-compatible recovery and geometry limits. +pub mod envelope; diff --git a/crates/psd/src/text/envelope.rs b/crates/psd/src/text/envelope.rs new file mode 100644 index 000000000..73345fa83 --- /dev/null +++ b/crates/psd/src/text/envelope.rs @@ -0,0 +1,309 @@ +//! Editable PSD Type Tool Object Setting (`TySh`) envelope, extracted from PhotoCraft. +//! App-specific text styles, layout, units and editing policy stay with the consumer. +//! Parsing intentionally retains PhotoCraft's optional-warp and partial-bounds recovery. +//! Existing read-only [`super::TypeLayer`] behavior is unchanged. + +use crate::descriptor::{Class, Descriptor, Id, ReferenceItem, UnicodeString, Value as D, VersionedDescriptor}; +use crate::error::{PsdError, Result}; + +/// Maximum encoded envelope size, bounding hostile input and serialized output. +pub const MAX_BYTES: usize = 8 * 1024 * 1024; + +/// Parsed `TySh` contents. +#[derive(Clone, Debug, PartialEq)] +pub struct TySh { + /// Affine transform in PSD order: xx, xy, yx, yy, tx, ty. + pub transform: [f64; 6], + /// Text descriptor, including text content and styling data. + pub text: Descriptor, + /// Optional warp descriptor retained from the Type Tool block. + pub warp: Option, + /// The four trailing big-endian f32 values (left, top, right, bottom) after the warp + /// descriptor. Photoshop writes zeros for unwarped type and the text-space bounds (the + /// `bounds` descriptor's rectangle) for warped type. + pub bounds: [f32; 4], +} + +/// Whether a type transform can reach rendering. Pixel coordinates use i32 in the originating PhotoCraft consumer, so non-finite and unrepresentable affine terms are rejected (#4758). +pub fn transform_is_safe(t: &[f64; 6]) -> bool { + t.iter().all(|v| v.is_finite() && v.abs() <= f64::from(i32::MAX)) +} + +/// Parses a `TySh` block. +pub fn parse_tysh(data: &[u8]) -> Option { + if data.len() > MAX_BYTES { + return None; + } + let f = |at: usize| data.get(at..at + 8).map(|b| f64::from_be_bytes(b.try_into().unwrap_or([0; 8]))); + let mut m = [0.0; 6]; + for (i, v) in m.iter_mut().enumerate() { + *v = f(2 + i * 8)?; + } + if !transform_is_safe(&m) { + return None; + } + let (text, used) = VersionedDescriptor::parse_prefix(data.get(52..)?).ok()?; + let rest = data.get(52 + used..).unwrap_or(&[]); + let (warp, wused) = match rest.get(2..).and_then(|r| VersionedDescriptor::parse_prefix(r).ok()) { + Some((w, n)) => (Some(w.descriptor), n + 2), + None => (None, 0), + }; + let tail = rest.get(wused..).unwrap_or(&[]); + let mut bounds = [0f32; 4]; + for (i, b) in bounds.iter_mut().enumerate() { + if let Some(x) = tail.get(i * 4..i * 4 + 4) { + *b = f32::from_be_bytes(x.try_into().unwrap_or([0; 4])); + } + } + if bounds.iter().any(|x| !x.is_finite()) { + return None; + } + Some(TySh { transform: m, text: text.descriptor, warp, bounds }) +} + +/// `tysh` moved by (`dx`, `dy`): only the transform's translation changes. The text, its styles, +/// the warp and the trailing bounds (which are in text space) stay byte for byte, so a moved +/// Photoshop type layer keeps everything Photoshop wrote. `None` when the block is too short. +pub fn translate_tysh(tysh: &[u8], dx: f64, dy: f64) -> Option> { + if tysh.len() > MAX_BYTES || tysh.len() < 50 { + return None; + } + let mut transform = [0.0; 6]; + for (i, x) in transform.iter_mut().enumerate() { + *x = f64::from_be_bytes(tysh.get(2 + i * 8..10 + i * 8)?.try_into().ok()?); + } + transform[4] += dx; + transform[5] += dy; + if !transform_is_safe(&transform) { + return None; + } + let mut out = Vec::new(); + out.try_reserve_exact(tysh.len()).ok()?; + out.extend_from_slice(tysh); + // Version (2 bytes), then xx, xy, yx, yy, tx, ty as big-endian f64: tx at 34, ty at 42. + for (at, d) in [(34, dx), (42, dy)] { + let b = out.get_mut(at..at + 8)?; + let v = f64::from_be_bytes(b.try_into().ok()?) + d; + b.copy_from_slice(&v.to_be_bytes()); + } + Some(out) +} + +/// Serializes a `TySh` block. +pub fn write_tysh(t: &TySh) -> Result> { + if !transform_is_safe(&t.transform) || t.bounds.iter().any(|x| !x.is_finite()) { + return Err(PsdError::LimitExceeded("TySh geometry")); + } + let default; + let warp = match &t.warp { + Some(w) => w, + None => { + default = default_warp(); + &default + } + }; + let mut budget = Budget(78); // versions, transform and trailing bounds + budget.descriptor(&t.text, 0)?; + budget.descriptor(warp, 0)?; + let mut v = Vec::new(); + v.try_reserve_exact(budget.0).map_err(|_| PsdError::OutOfMemory("TySh output"))?; + v.extend_from_slice(&1u16.to_be_bytes()); + for x in t.transform { + v.extend_from_slice(&x.to_be_bytes()); + } + v.extend_from_slice(&50u16.to_be_bytes()); + v.extend_from_slice(&16u32.to_be_bytes()); + t.text.write(&mut v); + v.extend_from_slice(&1u16.to_be_bytes()); + v.extend_from_slice(&16u32.to_be_bytes()); + warp.write(&mut v); + for b in t.bounds { + v.extend_from_slice(&b.to_be_bytes()); + } + Ok(v) +} + +// Count before cloning or serializing: bound both recursive trees and total output. +struct Budget(usize); +impl Budget { + fn add(&mut self, n: usize) -> Result<()> { + self.0 = self.0.checked_add(n).filter(|n| *n <= MAX_BYTES).ok_or(PsdError::LimitExceeded("TySh bytes"))?; + Ok(()) + } + fn unicode(&mut self, s: &UnicodeString) -> Result<()> { + self.add(4)?; + self.add(s.0.len().checked_mul(2).ok_or(PsdError::LimitExceeded("TySh string"))?) + } + fn id(&mut self, id: &Id) -> Result<()> { + self.add(4)?; + self.add(id.as_bytes().len()) + } + fn class(&mut self, c: &Class) -> Result<()> { + self.unicode(&c.name)?; + self.id(&c.class_id) + } + fn descriptor(&mut self, d: &Descriptor, depth: usize) -> Result<()> { + if depth > 32 { + return Err(PsdError::LimitExceeded("TySh depth")); + } + self.unicode(&d.name)?; + self.id(&d.class_id)?; + self.add(4)?; + for (k, v) in &d.items { + self.id(k)?; + self.value(v, depth + 1)?; + } + Ok(()) + } + fn value(&mut self, v: &D, depth: usize) -> Result<()> { + if depth > 32 { + return Err(PsdError::LimitExceeded("TySh depth")); + } + self.add(4)?; + match v { + D::Descriptor(d) | D::GlobalObject(d) => self.descriptor(d, depth), + D::ObjectArray(o) => { + self.add(4)?; + self.descriptor(&o.body, depth) + } + D::List(xs) => { + self.add(4)?; + for x in xs { + self.value(x, depth + 1)?; + } + Ok(()) + } + D::Reference(xs) => { + self.add(4)?; + for x in xs { + self.add(4)?; + match x { + ReferenceItem::Property { class, key } => { + self.class(class)?; + self.id(key)?; + } + ReferenceItem::Class(c) => self.class(c)?, + ReferenceItem::Enumerated { class, type_id, value } => { + self.class(class)?; + self.id(type_id)?; + self.id(value)?; + } + ReferenceItem::Offset { class, .. } | ReferenceItem::Identifier { class, .. } | ReferenceItem::Index { class, .. } => { + self.class(class)?; + self.add(4)?; + } + ReferenceItem::Name { class, name } => { + self.class(class)?; + self.unicode(name)?; + } + } + } + Ok(()) + } + D::Text(x) => self.unicode(x), + D::Class(c) | D::GlobalClass(c) => self.class(c), + D::Enumerated { type_id, value } => { + self.id(type_id)?; + self.id(value) + } + D::UnitFloats { values, .. } => { + self.add(8)?; + self.add(values.len().checked_mul(8).ok_or(PsdError::LimitExceeded("TySh values"))?) + } + D::RawData(x) | D::Alias(x) | D::Path(x) => { + self.add(4)?; + self.add(x.len()) + } + D::Double(_) | D::LargeInteger(_) => self.add(8), + D::UnitFloat { .. } => self.add(12), + D::Integer(_) => self.add(4), + D::Boolean(_) => self.add(1), + } + } +} + +fn default_warp() -> Descriptor { + Descriptor::new("warp") + .with("warpStyle", enumv("warpStyle", "warpNone")) + .with("warpValue", D::Double(0.0)) + .with("warpPerspective", D::Double(0.0)) + .with("warpPerspectiveOther", D::Double(0.0)) + .with("warpRotate", enumv("Ornt", "Hrzn")) +} + +fn enumv(t: &str, v: &str) -> D { + D::Enumerated { type_id: Id::new(t), value: Id::new(v) } +} + +#[cfg(test)] +mod tests { + use super::*; + fn sample() -> TySh { + TySh { + transform: [1.0, 0.0, 0.0, 1.0, 25.0, 30.0], + text: Descriptor::new("TxLr").with("EngineData", D::RawData(b"preserved".to_vec())), + warp: Some(default_warp()), + bounds: [-1.0, 2.0, 30.0, 40.0], + } + } + #[test] + fn roundtrip_preserves_descriptors_transform_and_bounds() { + let t = sample(); + let b = write_tysh(&t).unwrap(); + assert_eq!(parse_tysh(&b), Some(t)); + let shifted = translate_tysh(&b, 7.0, -3.0).unwrap(); + assert_eq!(&shifted[..34], &b[..34]); + assert_eq!(&shifted[50..], &b[50..]); + assert_eq!(parse_tysh(&shifted).unwrap().transform, [1.0, 0.0, 0.0, 1.0, 32.0, 27.0]); + } + #[test] + fn truncations_and_unsafe_transforms_preserve_recovery_contract() { + let mut t = sample(); + let bytes = write_tysh(&t).unwrap(); + let (_, n) = VersionedDescriptor::parse_prefix(&bytes[52..]).unwrap(); + for cut in 0..52 + n { + assert!(parse_tysh(&bytes[..cut]).is_none(), "{cut}"); + } + let short = parse_tysh(&bytes[..52 + n]).unwrap(); + assert!(short.warp.is_none()); + assert_eq!(short.bounds, [0.0; 4]); + for cut in 0..50 { + assert!(translate_tysh(&bytes[..cut], 1.0, 2.0).is_none()); + } + for bad in [f64::NAN, f64::INFINITY, f64::NEG_INFINITY, f64::from(i32::MAX) + 1.0] { + for i in 0..6 { + t = sample(); + t.transform[i] = bad; + assert!(write_tysh(&t).is_err()); + } + } + t = sample(); + t.warp = None; + assert!(parse_tysh(&write_tysh(&t).unwrap()).unwrap().warp.is_some()); + let partial = parse_tysh(&bytes[..bytes.len() - 4]).unwrap(); + assert_eq!(partial.bounds, [-1.0, 2.0, 30.0, 0.0]); + } + #[test] + fn hostile_output_and_translations_are_refused_before_copying() { + let bytes = write_tysh(&sample()).unwrap(); + for bad in [f64::NAN, f64::INFINITY, f64::MAX] { + assert!(translate_tysh(&bytes, bad, 0.0).is_none()); + } + let mut t = sample(); + t.bounds[0] = f32::NAN; + assert!(write_tysh(&t).is_err()); + t = sample(); + t.text = Descriptor::new("TxLr").with("data", D::RawData(vec![0; MAX_BYTES])); + assert!(write_tysh(&t).is_err()); + let oversized = vec![0; MAX_BYTES + 1]; + assert!(translate_tysh(&oversized, 0.0, 0.0).is_none()); + assert!(parse_tysh(&oversized).is_none()); + let mut nested = D::Integer(1); + for _ in 0..34 { + nested = D::List(vec![nested]); + } + t = sample(); + t.text = Descriptor::new("TxLr").with("nest", nested); + assert!(write_tysh(&t).is_err()); + } +} diff --git a/crates/svg/Cargo.toml b/crates/svg/Cargo.toml index 3615a719c..3b778da47 100644 --- a/crates/svg/Cargo.toml +++ b/crates/svg/Cargo.toml @@ -12,6 +12,7 @@ workspace = true [dependencies] craft-xml = { workspace = true } +quick-xml = { version = "0.41", default-features = false } craft-image-codecs = { workspace = true } # SVGZ (gzip), pure-Rust backend. flate2 = { version = "1", default-features = false, features = ["rust_backend"] } diff --git a/crates/svg/src/drawing_write.rs b/crates/svg/src/drawing_write.rs new file mode 100644 index 000000000..91b94d790 --- /dev/null +++ b/crates/svg/src/drawing_write.rs @@ -0,0 +1,56 @@ +//! SVG polyline and triangle path attributes for display-list exporters. +use crate::write::num; +use std::fmt::{self, Display}; + +/// A fixed-precision polyline attribute streamed into an SVG writer. +pub struct Points { + iter: I, + decimals: usize, +} +/// Format points without allocating an intermediate attribute string. Precision is capped at 32. +pub fn points + Clone>(iter: I, decimals: usize) -> Points { + Points { iter, decimals: decimals.min(32) } +} +impl + Clone> Display for Points { + fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { + for [x, y] in self.iter.clone() { + write!(f, "{},{} ", num(x, self.decimals), num(y, self.decimals))?; + } + Ok(()) + } +} +/// A fixed-precision sequence of closed triangle subpaths. +pub struct Triangles { + iter: I, + decimals: usize, +} +/// Format independent triangles without allocating an intermediate attribute string. Precision is capped at 32. +pub fn triangles + Clone>(iter: I, decimals: usize) -> Triangles { + Triangles { iter, decimals: decimals.min(32) } +} +impl + Clone> Display for Triangles { + fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { + for [[ax, ay], [bx, by], [cx, cy]] in self.iter.clone() { + write!( + f, + "M{} {}L{} {}L{} {}Z", + num(ax, self.decimals), + num(ay, self.decimals), + num(bx, self.decimals), + num(by, self.decimals), + num(cx, self.decimals), + num(cy, self.decimals) + )?; + } + Ok(()) + } +} +#[cfg(test)] +mod tests { + use super::*; + #[test] + fn fixed_precision_attributes_keep_independent_triangles() { + assert_eq!(points([[1.0, 2.0], [3.0, 4.0]].into_iter(), 2).to_string(), "1.00,2.00 3.00,4.00 "); + assert_eq!(triangles([[[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]]].into_iter(), 1).to_string(), "M1.0 2.0L3.0 4.0L5.0 6.0Z"); + } +} diff --git a/crates/svg/src/guard.rs b/crates/svg/src/guard.rs index 073bfe34c..f3573e443 100644 --- a/crates/svg/src/guard.rs +++ b/crates/svg/src/guard.rs @@ -178,6 +178,10 @@ pub fn too_wide(svg: &str, most: usize, most_work: usize) -> bool { pub struct Prescan<'s> { /// Whether the elements nest deeper than the cap. The pass stops there. pub too_deep: bool, + /// Deepest element nesting observed (markup entities count ten times). + pub depth: usize, + /// Deepest stack of unclosed source elements, excluding self-closing leaves. + pub open_depth: usize, /// The text with each data URL in an element's attribute values left out: a `data:` value's /// quotes kept, a `url(…)`'s parentheses and quotes. Borrowed when there is none, and the /// whole text when `too_deep`. An embedded picture is a data URL and most of the text when @@ -250,7 +254,7 @@ pub fn prescan(svg: &str, most: usize) -> Prescan<'_> { (b.len(), false) }; // (now, deepest) of the elements, and of the markup inside the document type. - let (mut elements, mut entities) = ((0usize, 0usize), (0usize, 0usize)); + let (mut elements, mut entities, mut open_depth) = ((0usize, 0usize), (0usize, 0usize), 0usize); let mut doctype_end = 0; let mut i = 0; while let Some(at) = b.get(i..).and_then(|rest| rest.iter().position(|c| *c == b'<')).map(|at| i + at) { @@ -287,8 +291,14 @@ pub fn prescan(svg: &str, most: usize) -> Prescan<'_> { end }; if elements.1.saturating_add(entities.1.saturating_mul(10)) > most { - return Prescan { too_deep: true, markers: Cow::Borrowed(svg) }; + return Prescan { + too_deep: true, + depth: elements.1.saturating_add(entities.1.saturating_mul(10)), + open_depth, + markers: Cow::Borrowed(svg), + }; } + open_depth = open_depth.max(elements.0); } let markers = if data_values.is_empty() { Cow::Borrowed(svg) @@ -303,7 +313,7 @@ pub fn prescan(svg: &str, most: usize) -> Prescan<'_> { text.push_str(svg.get(from..).unwrap_or_default()); Cow::Owned(text) }; - Prescan { too_deep: false, markers } + Prescan { too_deep: false, depth: elements.1.saturating_add(entities.1.saturating_mul(10)), open_depth, markers } } #[cfg(test)] diff --git a/crates/svg/src/lib.rs b/crates/svg/src/lib.rs index 982e4f82b..25bf880cd 100644 --- a/crates/svg/src/lib.rs +++ b/crates/svg/src/lib.rs @@ -18,7 +18,10 @@ #![forbid(unsafe_code)] pub mod color; +pub mod drawing_write; pub mod guard; +pub mod neutral_document; +pub mod path_data; pub mod paths; pub mod preflight; pub mod sniff; diff --git a/crates/svg/src/neutral_document.rs b/crates/svg/src/neutral_document.rs new file mode 100644 index 000000000..17f352058 --- /dev/null +++ b/crates/svg/src/neutral_document.rs @@ -0,0 +1,407 @@ +//! A bounded SVG document walk that preserves geometric primitive kinds for CAD import. +//! +//! Coordinates remain in SVG user units. Each shape carries the composed transform into +//! millimetres with the page Y axis pointing up; consumers map shapes into their own model. +use crate::path_data::{self, Point, Subpath}; +use quick_xml::events::{BytesStart, Event as XmlEvent}; +use quick_xml::{Reader, XmlVersion}; +use std::collections::BTreeMap; + +const MAX_DEPTH: usize = 64; +const MAX_SHAPES: usize = 500_000; +const MAX_POINTS: usize = 8_000_000; +const MAX_POLY_POINTS: usize = 100_001; +const SKIP: &[&[u8]] = &[ + b"defs", + b"symbol", + b"clipPath", + b"mask", + b"pattern", + b"marker", + b"linearGradient", + b"radialGradient", + b"filter", + b"style", + b"text", + b"image", + b"title", + b"desc", + b"metadata", +]; +const CONTAINERS: &[&[u8]] = &[b"g", b"svg", b"a", b"switch"]; + +/// Input failure or resource bound. +#[derive(Debug, thiserror::Error)] +#[error("{0}")] +pub struct Error(pub String); +/// A shape's original geometry before its affine transform. +#[derive(Debug, Clone, PartialEq)] +pub enum Primitive { + Rect { x: f64, y: f64, width: f64, height: f64 }, + Circle { center: Point, radius: f64 }, + Ellipse { center: Point, rx: f64, ry: f64 }, + Line { a: Point, b: Point }, + Polyline { points: Vec, closed: bool }, + Path(Vec), +} +/// A primitive with its layer and SVG affine matrix `[a,b,c,d,e,f]`. +#[derive(Debug, Clone, PartialEq)] +pub struct Shape { + pub primitive: Primitive, + pub transform: [f64; 6], + pub layer: String, +} +/// The shapes, encountered layer names, and skipped-feature counts. `visit` streams shapes +/// to its sink, so its returned `shapes` is empty; `import` collects them. +#[derive(Debug, Default, PartialEq)] +pub struct Document { + pub shapes: Vec, + pub layers: Vec, + pub skipped: BTreeMap, +} +/// Features left out by this geometry-only import. +#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord)] +pub enum Skipped { + Text, + Image, + Use, + Style, + RoundedCorners, +} + +fn attr(e: &BytesStart<'_>, name: &[u8]) -> Option { + for a in e.attributes() { + let Ok(a) = a else { return None }; + if a.key.into_inner() == name { + return a.normalized_value(XmlVersion::Explicit1_0).ok().map(|v| v.into_owned()); + } + } + None +} +fn hidden(e: &BytesStart<'_>) -> bool { + attr(e, b"display").as_deref() == Some("none") || attr(e, b"style").is_some_and(|s| s.replace(' ', "").contains("display:none")) +} +fn length(s: &str) -> Option { + let s = s.trim(); + let cut = s.find(|c: char| !(c.is_ascii_digit() || matches!(c, '.' | '-' | '+' | 'e' | 'E'))).unwrap_or(s.len()); + let (n, unit) = s.split_at(cut); + let v: f64 = n.trim().parse().ok()?; + if !v.is_finite() { + return None; + } + Some( + v * match unit.trim() { + "" | "px" | "mm" => 1.0, + "cm" => 10.0, + "in" => 25.4, + "pt" => 25.4 / 72.0, + "pc" => 25.4 / 6.0, + _ => 25.4 / 96.0, + }, + ) +} +fn num(e: &BytesStart<'_>, name: &[u8], default: f64) -> f64 { + attr(e, name).as_deref().and_then(length).unwrap_or(default) +} +const ID: [f64; 6] = [1., 0., 0., 1., 0., 0.]; +fn mul(a: [f64; 6], b: [f64; 6]) -> [f64; 6] { + [ + a[0] * b[0] + a[2] * b[1], + a[1] * b[0] + a[3] * b[1], + a[0] * b[2] + a[2] * b[3], + a[1] * b[2] + a[3] * b[3], + a[0] * b[4] + a[2] * b[5] + a[4], + a[1] * b[4] + a[3] * b[5] + a[5], + ] +} +fn translate(x: f64, y: f64) -> [f64; 6] { + [1., 0., 0., 1., x, y] +} +fn scale(x: f64, y: f64) -> [f64; 6] { + [x, 0., 0., y, 0., 0.] +} +fn rotate(a: f64) -> [f64; 6] { + let (s, c) = a.to_radians().sin_cos(); + [c, s, -s, c, 0., 0.] +} +fn numbers(s: &str) -> impl Iterator + '_ { + path_data::Numbers::new(s) +} +fn transform(s: &str) -> Option<[f64; 6]> { + let mut out = None; + let mut rest = s.trim(); + while !rest.is_empty() { + let open = rest.find('(')?; + let name = rest[..open].trim(); + let close = rest[open..].find(')')? + open; + // No transform function accepts more than six arguments. + let args: Vec<_> = numbers(&rest[open + 1..close]).take(7).collect(); + let m = match (name, args.as_slice()) { + ("matrix", [a, b, c, d, e, f]) => [*a, *b, *c, *d, *e, *f], + ("translate", [x]) => translate(*x, 0.), + ("translate", [x, y]) => translate(*x, *y), + ("scale", [x]) => scale(*x, *x), + ("scale", [x, y]) => scale(*x, *y), + ("rotate", [a]) => rotate(*a), + ("rotate", [a, x, y]) => mul(mul(translate(*x, *y), rotate(*a)), translate(-x, -y)), + ("skewX", [a]) => [1., 0., a.to_radians().tan(), 1., 0., 0.], + ("skewY", [a]) => [1., a.to_radians().tan(), 0., 1., 0., 0.], + _ => return None, + }; + out = Some(out.map_or(m, |prev| mul(prev, m))); + rest = rest[close + 1..].trim_start_matches(|c: char| c == ',' || c.is_whitespace()); + } + out +} +fn viewport(e: &BytesStart<'_>) -> [f64; 6] { + let vb = attr(e, b"viewBox").and_then(|raw| { + let n: Vec<_> = numbers(&raw).take(5).collect(); + match n.as_slice() { + [x, y, w, h] if *w > 0. && *h > 0. => Some((*x, *y, *w, *h)), + _ => None, + } + }); + let root = |name: &[u8]| { + let raw = attr(e, name)?; + let mm = length(&raw)?; + let unit = raw.trim().trim_start_matches(|c: char| c.is_ascii_digit() || matches!(c, '.' | '-' | '+' | 'e' | 'E')).trim(); + Some(if matches!(unit, "" | "px") { mm * 25.4 / 96. } else { mm }) + }; + let unit = match (vb, root(b"width"), root(b"height")) { + (Some((_, _, w, _)), Some(x), _) if x > 0. => x / w, + (Some((_, _, _, h)), _, Some(y)) if y > 0. => y / h, + _ => 25.4 / 96., + }; + let (x, y, h) = vb.map_or((0., 0., 0.), |(x, y, _, h)| (x, y, h)); + [unit, 0., 0., -unit, -x * unit, (y + h) * unit] +} +fn layer_name(raw: &str) -> String { + let s: String = raw.trim().chars().map(|c| if c.is_ascii_alphanumeric() || matches!(c, '-' | '_' | '$') { c } else { '_' }).take(128).collect(); + let s = s.trim_matches('_'); + if s.is_empty() { "SVG".into() } else { s.into() } +} +/// A layer declaration or a drawable primitive in document order. +pub enum Event { + Layer(String), + Shape(Shape), +} + +struct Walker<'a> { + doc: Document, + stack: Vec<[f64; 6]>, + layer: String, + skip: usize, + points: usize, + shapes: usize, + sink: &'a mut dyn FnMut(Event) -> Result<(), Error>, +} +impl Walker<'_> { + fn current(&self) -> [f64; 6] { + self.stack.last().copied().unwrap_or(ID) + } + fn note(&mut self, what: Skipped) { + *self.doc.skipped.entry(what).or_default() += 1; + } + fn skipped(name: &[u8]) -> Option { + match name { + b"text" => Some(Skipped::Text), + b"image" => Some(Skipped::Image), + _ => None, + } + } + fn start(&mut self, name: &[u8], e: &BytesStart<'_>) -> Result<(), Error> { + if name == b"style" { + self.note(Skipped::Style); + } + if self.skip > 0 { + self.skip += 1; + return Ok(()); + } + if hidden(e) { + self.skip = 1; + return Ok(()); + } + if SKIP.contains(&name) { + if let Some(s) = Self::skipped(name) { + self.note(s); + } + self.skip = 1; + return Ok(()); + } + if !CONTAINERS.contains(&name) { + return self.element(name, e); + } + if name == b"svg" + && self.stack.len() == 1 + && let Some(root) = self.stack.first_mut() + { + *root = viewport(e); + } + if name == b"g" + && attr(e, b"inkscape:groupmode").as_deref() == Some("layer") + && let Some(raw) = attr(e, b"inkscape:label").or_else(|| attr(e, b"id")) + { + let layer = layer_name(&raw); + if layer == "0" || self.doc.layers.contains(&layer) { + self.layer = layer; + } else if self.doc.layers.len() < 255 { + self.doc.layers.push(layer.clone()); + (self.sink)(Event::Layer(layer.clone()))?; + self.layer = layer; + } + } + if self.stack.len() >= MAX_DEPTH { + return Err(Error(format!("SVG: more than {MAX_DEPTH} nested groups"))); + } + let next = attr(e, b"transform").as_deref().and_then(transform).map_or(self.current(), |t| mul(self.current(), t)); + self.stack.push(next); + Ok(()) + } + fn element(&mut self, name: &[u8], e: &BytesStart<'_>) -> Result<(), Error> { + if self.skip > 0 { + return Ok(()); + } + if SKIP.contains(&name) || hidden(e) { + if !hidden(e) + && let Some(s) = Self::skipped(name) + { + self.note(s); + } + return Ok(()); + } + let p = |x, y| Point::new(num(e, x, 0.), num(e, y, 0.)); + let primitive = match name { + b"use" => { + self.note(Skipped::Use); + return Ok(()); + } + b"rect" => { + let (x, y, w, h) = (num(e, b"x", 0.), num(e, b"y", 0.), num(e, b"width", 0.), num(e, b"height", 0.)); + if w <= 0. || h <= 0. { + return Ok(()); + } + if num(e, b"rx", 0.) > 0. || num(e, b"ry", 0.) > 0. { + self.note(Skipped::RoundedCorners); + } + Primitive::Rect { x, y, width: w, height: h } + } + b"circle" => { + let r = num(e, b"r", 0.); + if r <= 0. { + return Ok(()); + } + Primitive::Circle { center: p(b"cx", b"cy"), radius: r } + } + b"ellipse" => { + let (rx, ry) = (num(e, b"rx", 0.), num(e, b"ry", 0.)); + if rx <= 0. || ry <= 0. { + return Ok(()); + } + Primitive::Ellipse { center: p(b"cx", b"cy"), rx, ry } + } + b"line" => Primitive::Line { a: p(b"x1", b"y1"), b: p(b"x2", b"y2") }, + b"polyline" | b"polygon" => { + let Some(raw) = attr(e, b"points") else { return Ok(()) }; + let n: Vec<_> = numbers(&raw).take(MAX_POLY_POINTS * 2).collect(); + let pts: Vec<_> = n.as_chunks::<2>().0.iter().map(|[x, y]| Point::new(*x, *y)).collect(); + if pts.len() < 2 { + return Ok(()); + } + Primitive::Polyline { points: pts, closed: name == b"polygon" } + } + b"path" => { + let Some(raw) = attr(e, b"d") else { return Ok(()) }; + Primitive::Path(path_data::parse_path(&raw).map_err(|e| Error(e.to_string()))?) + } + _ => return Ok(()), + }; + let cost = match &primitive { + Primitive::Polyline { points, .. } => points.len(), + Primitive::Path(paths) => paths.iter().map(|p| p.segs.len()).sum(), + _ => 1, + }; + self.points = self.points.checked_add(cost).filter(|n| *n <= MAX_POINTS).ok_or_else(|| Error("SVG: too much generated geometry".into()))?; + if self.shapes >= MAX_SHAPES { + return Err(Error(format!("SVG: more than {MAX_SHAPES} elements"))); + } + let matrix = attr(e, b"transform").as_deref().and_then(transform).map_or(self.current(), |t| mul(self.current(), t)); + self.shapes += 1; + (self.sink)(Event::Shape(Shape { primitive, transform: matrix, layer: self.layer.clone() })) + } +} +/// Walk supported SVG geometry without retaining a second copy of the document's shapes. +/// The sink receives each layer before its first shape and each shape in input order. +pub fn visit(text: &str, sink: &mut dyn FnMut(Event) -> Result<(), Error>) -> Result { + let mut walker = Walker { doc: Document::default(), stack: vec![ID], layer: "0".into(), skip: 0, points: 0, shapes: 0, sink }; + let mut reader = Reader::from_str(text); + reader.config_mut().trim_text(true); + loop { + match reader.read_event().map_err(|e| Error(format!("SVG: {e}")))? { + XmlEvent::Start(e) => walker.start(e.name().into_inner(), &e)?, + XmlEvent::Empty(e) => walker.element(e.name().into_inner(), &e)?, + XmlEvent::End(e) => { + if walker.skip > 0 { + walker.skip -= 1; + } else if CONTAINERS.contains(&e.name().into_inner()) && walker.stack.len() > 1 { + walker.stack.pop(); + } + } + XmlEvent::Eof => break, + _ => {} + } + } + Ok(walker.doc) +} +/// Read the supported geometry into an owned document. +pub fn import(text: &str) -> Result { + let mut shapes = Vec::new(); + let mut doc = visit(text, &mut |event| { + if let Event::Shape(shape) = event { + shapes.push(shape); + } + Ok(()) + })?; + doc.shapes = shapes; + Ok(doc) +} +#[cfg(test)] +mod tests { + #![allow(clippy::unwrap_used)] + use super::*; + #[test] + fn keeps_primitives_transform_layers_and_skips() { + let svg = r##"x"##; + let d = import(svg).unwrap(); + assert_eq!(d.layers, ["cut"]); + assert_eq!(d.shapes.len(), 3); + assert_eq!(d.shapes[0].primitive, Primitive::Circle { center: Point::new(3., 4.), radius: 2. }); + assert_eq!(d.shapes[0].transform, [1., 0., 0., -1., 2., 9.]); + assert!(matches!(d.shapes[1].primitive, Primitive::Ellipse { .. })); + assert!(matches!(&d.shapes[2].primitive,Primitive::Path(p) if matches!(p[0].segs[0],path_data::Segment::Arc { .. }))); + assert_eq!(d.skipped.get(&Skipped::Text), Some(&1)); + assert_eq!(d.skipped.get(&Skipped::Use), Some(&1)); + } + #[test] + fn streaming_delivers_layers_before_shapes_and_stops_on_error() { + let s = r#""#; + let mut events = Vec::new(); + let result = visit(s, &mut |event| { + match event { + Event::Layer(name) => events.push(name), + Event::Shape(_) => { + events.push("shape".into()); + return Err(Error("stop".into())); + } + } + Ok(()) + }); + assert!(result.is_err()); + assert_eq!(events, ["cut", "shape"]); + } + + #[test] + fn refuses_excessive_nesting() { + let s = format!("{}{}", "".repeat(100), "".repeat(100)); + assert!(import(&s).is_err()); + } +} diff --git a/crates/svg/src/path_data.rs b/crates/svg/src/path_data.rs new file mode 100644 index 000000000..96d0db771 --- /dev/null +++ b/crates/svg/src/path_data.rs @@ -0,0 +1,456 @@ +//! SVG path data as neutral line, cubic and elliptical arc segments. +//! +//! Relative commands and smooth controls are resolved, while arcs remain arcs. +//! The caller chooses how to map the result to its own geometry model. +use std::ops::{Add, Mul, Sub}; + +/// A point in the path's original user coordinate system. +#[derive(Debug, Clone, Copy, PartialEq)] +pub struct Point { + /// Horizontal coordinate. + pub x: f64, + /// Vertical coordinate. + pub y: f64, +} +impl Point { + /// The origin. + pub const ZERO: Self = Self { x: 0.0, y: 0.0 }; + /// Construct a point. + pub const fn new(x: f64, y: f64) -> Self { + Self { x, y } + } +} +impl Add for Point { + type Output = Self; + fn add(self, rhs: Self) -> Self { + Self::new(self.x + rhs.x, self.y + rhs.y) + } +} +impl Sub for Point { + type Output = Self; + fn sub(self, rhs: Self) -> Self { + Self::new(self.x - rhs.x, self.y - rhs.y) + } +} +impl Mul for Point { + type Output = Self; + fn mul(self, rhs: f64) -> Self { + Self::new(self.x * rhs, self.y * rhs) + } +} + +/// Invalid or oversized path data. +#[derive(Debug, thiserror::Error)] +#[error("{0}")] +pub struct Error(pub String); +type Result = std::result::Result; +const MAX_PATH_SEGMENTS: usize = 100_000; +const MAX_CURVE_POINTS: usize = 100_001; + +pub(crate) struct Numbers<'a> { + s: &'a [u8], + i: usize, +} + +impl<'a> Numbers<'a> { + pub(crate) fn new(s: &'a str) -> Self { + Self { s: s.as_bytes(), i: 0 } + } + fn skip_seps(&mut self) { + while matches!(self.s.get(self.i), Some(b' ' | b'\t' | b'\r' | b'\n' | b',')) { + self.i += 1; + } + } + + fn next_num(&mut self) -> Option { + self.skip_seps(); + let start = self.i; + if matches!(self.s.get(self.i), Some(b'+' | b'-')) { + self.i += 1; + } + let mut digits = 0; + let mut dot = false; + let mut exp = false; + while let Some(c) = self.s.get(self.i) { + match c { + b'0'..=b'9' => { + digits += 1; + self.i += 1; + } + b'.' if !dot && !exp => { + dot = true; + self.i += 1; + } + b'e' | b'E' if !exp && digits > 0 => { + exp = true; + self.i += 1; + if matches!(self.s.get(self.i), Some(b'+' | b'-')) { + self.i += 1; + } + } + _ => break, + } + } + if digits == 0 { + self.i = start; + return None; + } + std::str::from_utf8(&self.s[start..self.i]).ok()?.parse::().ok().filter(|v| v.is_finite()) + } +} + +impl Iterator for Numbers<'_> { + type Item = f64; + fn next(&mut self) -> Option { + self.next_num() + } +} + +/// One `M …` run of a `d` attribute. +#[derive(Debug, Clone, PartialEq)] +pub struct Subpath { + /// Start point. + pub start: Point, + /// Segments in drawing order. + pub segs: Vec, + /// Whether a close-path command closed this subpath. + pub closed: bool, +} + +/// A path segment; arc values retain the SVG endpoint parameterization. +#[derive(Debug, Clone, PartialEq)] +pub enum Segment { + Line(Point), + Cubic(Point, Point, Point), + Arc { rx: f64, ry: f64, rot: f64, large: bool, sweep: bool, end: Point }, +} + +/// Tokens of a `d` attribute: command letters and numbers, in order. +enum Tok { + Cmd(u8), + Num(f64), +} + +struct Tokenizer<'a> { + s: &'a [u8], + i: usize, + /// Start of the most recently returned token. + tok_start: usize, +} + +impl<'a> Tokenizer<'a> { + fn new(s: &'a str) -> Self { + Tokenizer { s: s.as_bytes(), i: 0, tok_start: 0 } + } + + /// Read an arc flag: a single `0` or `1` character, which needs no separator after it. + fn next_flag(&mut self) -> Option { + let mut nums = Numbers { s: self.s, i: self.i }; + nums.skip_seps(); + let v = match self.s.get(nums.i)? { + b'0' => 0.0, + b'1' => 1.0, + _ => return None, + }; + self.i = nums.i + 1; + Some(v) + } + + fn next_tok(&mut self) -> Option { + let mut nums = Numbers { s: self.s, i: self.i }; + nums.skip_seps(); + self.i = nums.i; + self.tok_start = self.i; + let c = self.s.get(self.i).copied()?; + if c.is_ascii_alphabetic() { + self.i += 1; + return Some(Tok::Cmd(c)); + } + let tok = nums.next_num().map(Tok::Num); + self.i = nums.i; + tok + } +} + +/// Split `d` into subpaths, resolving relative commands and the S/T control reflections. +pub fn parse_path(d: &str) -> Result> { + let mut tk = Tokenizer::new(d); + let mut lookahead: Option = None; + let mut cmd = 0u8; + let mut cur = Point::ZERO; + let mut start = Point::ZERO; + let mut segs: Vec = Vec::new(); + let mut out: Vec = Vec::new(); + let closed = false; + let mut prev_cubic: Option = None; + let mut prev_quad: Option = None; + let mut segments = 0usize; + + let next_num = |tk: &mut Tokenizer, look: &mut Option| -> Result { + match look.take() { + Some(Tok::Num(v)) => Ok(v), + Some(other @ Tok::Cmd(_)) => { + *look = Some(other); + Err(Error("SVG: path command is missing coordinates".into())) + } + None => match tk.next_tok() { + Some(Tok::Num(v)) => Ok(v), + Some(other @ Tok::Cmd(_)) => { + *look = Some(other); + Err(Error("SVG: path command is missing coordinates".into())) + } + None => Err(Error("SVG: path command is missing coordinates".into())), + }, + } + }; + + while let Some(tok) = lookahead.take().or_else(|| tk.next_tok()) { + let rel = match tok { + Tok::Cmd(c) => { + if c.eq_ignore_ascii_case(&b'z') { + if !segs.is_empty() { + out.push(Subpath { start, segs, closed: true }); + segs = Vec::new(); + } + cur = start; + cmd = 0; + continue; + } + cmd = c; + c.is_ascii_lowercase() + } + tok @ Tok::Num(_) => { + if cmd == 0 { + return Err(Error("SVG: path data does not start with a command".into())); + } + lookahead = Some(tok); + cmd.is_ascii_lowercase() + } + }; + let abs = cmd.to_ascii_uppercase(); + let count = match abs { + b'M' | b'L' | b'T' => 2, + b'H' | b'V' => 1, + b'C' => 6, + b'S' | b'Q' => 4, + b'A' => 7, + _ => return Err(Error(format!("SVG: unknown path command `{}`", cmd as char))), + }; + let mut args = [0.0f64; 7]; + for (idx, slot) in args.iter_mut().enumerate().take(count) { + if abs == b'A' && (idx == 3 || idx == 4) { + // Flags are single characters; re-scan a pending number token from its start. + if matches!(lookahead, Some(Tok::Num(_))) { + lookahead = None; + tk.i = tk.tok_start; + } + if lookahead.is_none() { + *slot = tk.next_flag().ok_or_else(|| Error("SVG: path command is missing coordinates".into()))?; + continue; + } + } + *slot = next_num(&mut tk, &mut lookahead)?; + } + let point = |x: f64, y: f64| if rel { cur + Point::new(x, y) } else { Point::new(x, y) }; + match abs { + b'M' => { + if !segs.is_empty() { + out.push(Subpath { start, segs, closed: false }); + segs = Vec::new(); + } + cur = point(args[0], args[1]); + start = cur; + cmd = if rel { b'l' } else { b'L' }; + prev_cubic = None; + prev_quad = None; + } + b'L' => { + cur = point(args[0], args[1]); + segs.push(Segment::Line(cur)); + prev_cubic = None; + prev_quad = None; + } + b'H' => { + cur = Point::new(if rel { cur.x + args[0] } else { args[0] }, cur.y); + segs.push(Segment::Line(cur)); + prev_cubic = None; + prev_quad = None; + } + b'V' => { + cur = Point::new(cur.x, if rel { cur.y + args[0] } else { args[0] }); + segs.push(Segment::Line(cur)); + prev_cubic = None; + prev_quad = None; + } + b'C' => { + let (c1, c2, p) = (point(args[0], args[1]), point(args[2], args[3]), point(args[4], args[5])); + segs.push(Segment::Cubic(c1, c2, p)); + prev_cubic = Some(c2); + prev_quad = None; + cur = p; + } + b'S' => { + let c1 = prev_cubic.map_or(cur, |c2| cur * 2.0 - c2); + let (c2, p) = (point(args[0], args[1]), point(args[2], args[3])); + segs.push(Segment::Cubic(c1, c2, p)); + prev_cubic = Some(c2); + prev_quad = None; + cur = p; + } + b'Q' => { + let (q, p) = (point(args[0], args[1]), point(args[2], args[3])); + segs.push(Segment::Cubic(cur + (q - cur) * (2.0 / 3.0), p + (q - p) * (2.0 / 3.0), p)); + prev_quad = Some(q); + prev_cubic = None; + cur = p; + } + b'T' => { + let reflect = prev_quad.map_or(cur, |q| cur * 2.0 - q); + let p = point(args[0], args[1]); + segs.push(Segment::Cubic(cur + (reflect - cur) * (2.0 / 3.0), p + (reflect - p) * (2.0 / 3.0), p)); + prev_quad = Some(reflect); + prev_cubic = None; + cur = p; + } + b'A' => { + let end = point(args[5], args[6]); + segs.push(Segment::Arc { rx: args[0].abs(), ry: args[1].abs(), rot: args[2], large: args[3] != 0.0, sweep: args[4] != 0.0, end }); + cur = end; + prev_cubic = None; + prev_quad = None; + } + _ => return Err(Error("SVG: unsupported path command".into())), + } + segments += 1; + if segments > MAX_PATH_SEGMENTS { + return Err(Error("SVG: path has too many segments".into())); + } + if segs.len() > MAX_CURVE_POINTS { + return Err(Error("SVG: path has too many segments".into())); + } + } + if !segs.is_empty() { + out.push(Subpath { start, segs, closed }); + } + Ok(out) +} +/// The centre and signed sweep of a circular arc from `p0` to `p1` in a Y-up frame, per the +/// SVG endpoint parameterisation (F.6.5). Radii that cannot span the chord are scaled up, +/// as the specification requires. +pub fn arc_center_circular(p0: Point, p1: Point, r: f64, large: bool, sweep: bool) -> Option<(Point, f64)> { + if r <= 0.0 || !r.is_finite() { + return None; + } + let d = p1 - p0; + let dist = d.x.hypot(d.y); + let r = if dist > 2.0 * r { dist / 2.0 } else { r }; + if dist <= f64::EPSILON { + return None; + } + let mid = (p0 + p1) * 0.5; + let h = (r * r - (dist * dist) / 4.0).max(0.0).sqrt(); + let n = Point::new(-d.y / dist, d.x / dist); + let sign = if large != sweep { 1.0 } else { -1.0 }; + let center = mid + n * (h * sign); + let a0 = (p0.y - center.y).atan2(p0.x - center.x); + let a1 = (p1.y - center.y).atan2(p1.x - center.x); + let ccw = (a1 - a0).rem_euclid(std::f64::consts::TAU); + Some((center, pick_sweep(ccw, large, sweep))) +} + +/// The signed sweep (CCW positive) for a chord sweep `ccw` in `[0, 2Ï€)`: the centre is +/// already chosen so that the magnitude matches `large`, so only the direction is left. +fn pick_sweep(ccw: f64, large: bool, sweep: bool) -> f64 { + let tau = std::f64::consts::TAU; + let eps = f64::EPSILON; + if sweep { + if ccw <= eps { if large { tau } else { eps } } else { ccw } + } else if ccw >= tau - eps { + if large { -tau } else { -eps } + } else { + ccw - tau + } +} + +/// Endpoint → centre parameterisation for a rotated elliptical arc (SVG F.6.5), in the +/// frame the points are given in. Returns the centre, the start parameter and the signed +/// sweep. +pub fn elliptical_params(p0: Point, p1: Point, rx: f64, ry: f64, phi: f64, large: bool, sweep: bool) -> Option<(Point, f64, f64)> { + if rx <= 0.0 || ry <= 0.0 { + return None; + } + let (sphi, cphi) = phi.sin_cos(); + let dx = (p0.x - p1.x) / 2.0; + let dy = (p0.y - p1.y) / 2.0; + let x1 = cphi * dx + sphi * dy; + let y1 = -sphi * dx + cphi * dy; + let (rx, ry) = { + let lambda = x1 * x1 / (rx * rx) + y1 * y1 / (ry * ry); + if lambda > 1.0 { + let s = lambda.sqrt(); + (rx * s, ry * s) + } else { + (rx, ry) + } + }; + let num = rx * rx * ry * ry - rx * rx * y1 * y1 - ry * ry * x1 * x1; + let den = rx * rx * y1 * y1 + ry * ry * x1 * x1; + let coef = if sweep != large { 1.0 } else { -1.0 } * (num.max(0.0) / den.max(f64::MIN_POSITIVE)).sqrt(); + let cxp = coef * rx * y1 / ry; + let cyp = -coef * ry * x1 / rx; + let center = Point::new(cphi * cxp - sphi * cyp + (p0.x + p1.x) / 2.0, sphi * cxp + cphi * cyp + (p0.y + p1.y) / 2.0); + let angle = |ux: f64, uy: f64, vx: f64, vy: f64| { + let dot = ux * vx + uy * vy; + let det = ux * vy - uy * vx; + let denom = (ux.hypot(uy) * vx.hypot(vy)).max(f64::MIN_POSITIVE); + (dot / denom).clamp(-1.0, 1.0).acos().copysign(det) + }; + let (ux, uy) = ((x1 - cxp) / rx, (y1 - cyp) / ry); + let (vx, vy) = ((-x1 - cxp) / rx, (-y1 - cyp) / ry); + let theta = angle(1.0, 0.0, ux, uy); + let ccw = angle(ux, uy, vx, vy).rem_euclid(std::f64::consts::TAU); + Some((center, theta, pick_sweep(ccw, large, sweep))) +} + +#[cfg(test)] +mod tests { + #![allow(clippy::unwrap_used)] + use super::*; + + #[test] + fn resolves_relative_and_smooth_commands_without_flattening_arcs() { + let paths = parse_path("M1 2 l3 4 C5 6 7 8 9 10 s2 3 4 5 A6 7 30 0110 12 z").unwrap(); + assert_eq!(paths.len(), 1); + assert!(paths[0].closed); + assert_eq!(paths[0].start, Point::new(1.0, 2.0)); + assert_eq!(paths[0].segs[0], Segment::Line(Point::new(4.0, 6.0))); + assert_eq!(paths[0].segs[2], Segment::Cubic(Point::new(11.0, 12.0), Point::new(11.0, 13.0), Point::new(13.0, 15.0))); + assert_eq!(paths[0].segs[3], Segment::Arc { rx: 6.0, ry: 7.0, rot: 30.0, large: false, sweep: true, end: Point::new(10.0, 12.0) }); + } + + #[test] + fn smooth_quadratic_consumes_both_coordinates() { + let paths = parse_path("M0 0 Q3 3 6 0 T12 0").unwrap(); + assert_eq!(paths[0].segs[1], Segment::Cubic(Point::new(8.0, -2.0), Point::new(10.0, -2.0), Point::new(12.0, 0.0))); + } + + #[test] + fn arc_centers_keep_radius_correction_and_sweep() { + let (center, delta) = arc_center_circular(Point::new(0.0, 0.0), Point::new(10.0, 0.0), 2.0, false, true).unwrap(); + assert!((center.x - 5.0).abs() < 1e-10); + assert!(center.y.abs() < 1e-10); + assert!((delta - std::f64::consts::PI).abs() < 1e-10); + let (center, start, sweep) = elliptical_params(Point::new(0.0, 0.0), Point::new(10.0, 0.0), 2.0, 1.0, 0.0, false, true).unwrap(); + assert!((center.x - 5.0).abs() < 1e-10); + assert!(center.y.abs() < 1e-10 && start.is_finite()); + assert!((sweep - std::f64::consts::PI).abs() < 1e-10); + } + + #[test] + fn rejects_invalid_flags_and_limits_segments() { + assert!(parse_path("M0 0 A1 1 0 5 1 2 2").is_err()); + assert!(parse_path("M0 0 L").is_err()); + let excessive = format!("M0 0{}", " L1 1".repeat(MAX_PATH_SEGMENTS)); + assert!(parse_path(&excessive).is_err()); + } +} diff --git a/crates/svg/src/preflight.rs b/crates/svg/src/preflight.rs index 3ab381ff3..0c2d3a7b1 100644 --- a/crates/svg/src/preflight.rs +++ b/crates/svg/src/preflight.rs @@ -71,6 +71,12 @@ pub fn text<'a>(bytes: &'a [u8], limits: &Limits) -> Result, Error> /// Parses already-decoded SVG text and checks it against `limits`. pub fn parse_text<'a>(svg: &'a str, limits: &Limits) -> Result, Error> { + parse_text_with_depth(svg, limits).map(|(xml, _, _)| xml) +} + +/// [`parse_text`], with the source nesting: the deepest element nesting (markup entities count +/// ten times) and the deepest stack of open elements (a self-closing leaf opens none). +pub fn parse_text_with_depth<'a>(svg: &'a str, limits: &Limits) -> Result<(roxmltree::Document<'a>, usize, usize), Error> { if svg.len() > limits.bytes { return Err(Error::Text(crate::svgz::Error::TooLarge(limits.bytes))); } @@ -83,9 +89,38 @@ pub fn parse_text<'a>(svg: &'a str, limits: &Limits) -> Result Result<(), Error> { + let limits = Limits { depth: 4, ..Default::default() }; + let (document, depth, open_depth) = parse_text_with_depth("", &limits)?; + assert!(document.root_element().has_tag_name("svg")); + // Every element is a level (svg, g, path); a self-closing leaf opens none (svg, g). + assert_eq!(depth, 3); + assert_eq!(open_depth, 2); + let simple = parse_text("", &limits)?; + assert!(simple.root_element().has_tag_name("svg")); + Ok(()) + } + + #[test] + fn source_depth_and_xml_errors_keep_their_specific_kind() { + let shallow = Limits { depth: 1, ..Default::default() }; + assert!(matches!(parse_text_with_depth("", &shallow), Err(Error::TooDeep(1)))); + + let normal = Limits::default(); + assert!(matches!(parse_text_with_depth("", &normal), Err(Error::Xml(_)))); + } } diff --git a/crates/svg/src/tree.rs b/crates/svg/src/tree.rs index cb33137b8..9cf85c3bd 100644 --- a/crates/svg/src/tree.rs +++ b/crates/svg/src/tree.rs @@ -112,13 +112,19 @@ pub enum Error { /// usvg's own error. #[error("invalid SVG: {0}")] Parse(String), + /// The root viewport is missing or has dimensions this SVG helper cannot preserve. + #[error("invalid SVG root viewport: {0}")] + Viewport(String), + /// The rewritten SVG would exceed the caller's output byte budget. + #[error("rewritten SVG exceeds {0} bytes")] + OutputTooLarge(usize), } /// The SVG text of the bytes of an SVG or SVGZ file, checked against `limits`: what [`parse`] /// hands usvg. For a caller that edits the text before it parses ([`parse_text`]). pub fn text<'a>(bytes: &'a [u8], limits: &Limits) -> Result, Error> { let text = svgz::text(bytes, limits.bytes)?; - checked(&text, limits)?; + preflight(&text, limits)?; Ok(text) } @@ -130,8 +136,7 @@ pub fn parse(bytes: &[u8], limits: &Limits, options: &usvg::Options<'_>) -> Resu /// [`parse`] for SVG text the caller already holds. pub fn parse_text(svg: &str, limits: &Limits, options: &usvg::Options<'_>) -> Result { - let xml = checked(svg, limits)?; - usvg::Tree::from_xmltree(&xml, options).map_err(|e| Error::Parse(e.to_string())) + preflight(svg, limits)?.parse(options) } /// Read the root SVG's unitless width and height without constructing a rendering tree. @@ -139,8 +144,8 @@ pub fn parse_text(svg: &str, limits: &Limits, options: &usvg::Options<'_>) -> Re /// Lengths with units, percentages, missing dimensions, and non-SVG roots return `Ok(None)`; /// malformed or over-limit XML returns the same bounded parsing error as [`parse_text`]. pub fn root_unitless_size(svg: &str, limits: &Limits) -> Result, Error> { - let xml = checked(svg, limits)?; - let root = xml.root_element(); + let checked = preflight(svg, limits)?; + let root = checked.xml.root_element(); if root.tag_name().name() != "svg" { return Ok(None); } @@ -152,8 +157,30 @@ pub fn root_unitless_size(svg: &str, limits: &Limits) -> Result(svg: &'s str, limits: &Limits) -> Result, Error> { +/// SVG text after shared byte, entity, attribute, XML and expansion-depth checks. +/// Keep it opaque so consumers use the same checked tree and don't reparse XML locally. +pub struct Preflight<'s> { + xml: roxmltree::Document<'s>, + depth: usize, + /// [`height`]: the nesting usvg converts, references followed. + height: usize, +} + +impl Preflight<'_> { + /// The deepest stack of open source elements, before references are followed; a + /// self-closing leaf opens no level (`` is 2). + pub fn depth(&self) -> usize { + self.depth + } + + /// Converts this already-checked XML into the configured usvg tree. + pub fn parse(self, options: &usvg::Options<'_>) -> Result { + usvg::Tree::from_xmltree(&self.xml, options).map_err(|e| Error::Parse(e.to_string())) + } +} + +/// Checks already-decoded SVG text once, retaining the parsed XML for both depth policy and tree conversion. +pub fn preflight<'s>(svg: &'s str, limits: &Limits) -> Result, Error> { let preflight_limits = crate::preflight::Limits { bytes: limits.bytes, depth: limits.depth, @@ -163,7 +190,7 @@ fn checked<'s>(svg: &'s str, limits: &Limits) -> Result, attribute_work: limits.attribute_work, xml_nodes: limits.xml_nodes, }; - let xml = crate::preflight::parse_text(svg, &preflight_limits).map_err(|e| match e { + let (xml, _, depth) = crate::preflight::parse_text_with_depth(svg, &preflight_limits).map_err(|e| match e { crate::preflight::Error::Text(e) => Error::Text(e), crate::preflight::Error::Entities => Error::Entities, crate::preflight::Error::EntityMarkup => Error::EntityMarkup, @@ -171,10 +198,69 @@ fn checked<'s>(svg: &'s str, limits: &Limits) -> Result, crate::preflight::Error::TooWide(attributes) => Error::TooWide(attributes), crate::preflight::Error::Xml(msg) => Error::Xml(msg), })?; - if height(&xml, limits.tree_depth)?.is_none() { + let Some(height) = height(&xml, limits.tree_depth)? else { return Err(Error::TooDeep(limits.tree_depth)); + }; + Ok(Preflight { xml, depth, height }) +} + +/// Dimensions of an SVG viewport in SVG user units. +#[derive(Clone, Copy, Debug, PartialEq)] +pub struct Viewport { + pub width: f64, + pub height: f64, +} + +/// Wraps a validated SVG root in a new viewport while preserving the original root as content. +/// The original root must have finite, positive pixel or point dimensions, and the wrapped +/// output must still pass [`preflight`] under `limits`. +pub fn rewrite_root_viewport(svg: &str, target: Viewport, limits: &Limits, output_limit: usize) -> Result { + if !target.width.is_finite() || !target.height.is_finite() || target.width <= 0.0 || target.height <= 0.0 { + return Err(Error::Viewport("target dimensions must be finite and positive".into())); } - Ok(xml) + let checked = preflight(svg, limits)?; + let root = checked.xml.root_element(); + if root.tag_name().name() != "svg" { + return Err(Error::Viewport("root element is not svg".into())); + } + let root_dimension = |key: &str| { + let value = root.attribute(key)?; + let value = value.trim(); + let (number, scale) = match value.strip_suffix("pt") { + Some(points) => (points, 4.0 / 3.0), + None => (value.strip_suffix("px").unwrap_or(value), 1.0), + }; + number.parse::().ok().map(|n| n * scale).filter(|n| n.is_finite() && *n > 0.0) + }; + let (width, height) = root_dimension("width") + .zip(root_dimension("height")) + .ok_or_else(|| Error::Viewport("source has no finite pixel or point dimensions".into()))?; + let range = root.range(); + let preamble = svg.get(..range.start).ok_or_else(|| Error::Viewport("invalid XML preamble range".into()))?; + let inner = svg.get(range).ok_or_else(|| Error::Viewport("invalid root element range".into()))?; + let prefix = format!( + r#""#, + target.width, target.height, width, height + ); + let len = preamble + .len() + .checked_add(prefix.len()) + .and_then(|n| n.checked_add(inner.len())) + .and_then(|n| n.checked_add(6)) + .ok_or(Error::OutputTooLarge(output_limit))?; + if len > output_limit { + return Err(Error::OutputTooLarge(output_limit)); + } + let mut output = String::new(); + output.try_reserve_exact(len).map_err(|_| Error::OutputTooLarge(output_limit))?; + output.push_str(preamble); + output.push_str(&prefix); + output.push_str(inner); + output.push_str(""); + // The wrapper adds a level, so a source at the depth limit can come out over it: check the + // output under the limits `parse_text` applies rather than return text it would refuse. + preflight(&output, limits)?; + Ok(output) } /// Properties whose value an element's descendants use too: the paint server they name is @@ -376,8 +462,8 @@ impl ImageResolver { /// Nested SVGs are checked automatically. The scan shares the normal XML limits and /// streaming-validates URL encoding without allocating another decoded image buffer. pub fn check_document(&self, svg: &str) -> Result<(), Error> { - let xml = checked(svg, &self.limits)?; - account_invalid_data_urls(&xml, &self.used); + let checked = preflight(svg, &self.limits)?; + account_invalid_data_urls(&checked.xml, &self.used); Ok(()) } @@ -520,8 +606,8 @@ fn embedded(used: &Used, limits: &Limits, data: &[u8], options: &usvg::Options<' return None; } used.bytes.fetch_add(text.len(), Ordering::Relaxed); - let xml = match checked(&text, limits) { - Ok(xml) => xml, + let checked = match preflight(&text, limits) { + Ok(checked) => checked, // Decoding UTF-16 or Latin-1 can exceed the text cap after the // raw/inflated bytes fitted. This is a resource refusal, not invalid XML. Err(Error::Text(svgz::Error::TooLarge(_))) => { @@ -538,10 +624,8 @@ fn embedded(used: &Used, limits: &Limits, data: &[u8], options: &usvg::Options<' return None; } }; - let Some(xml_depth) = height(&xml, limits.tree_depth).ok().flatten() else { - used.svg_limit.fetch_add(1, Ordering::Relaxed); - return None; - }; + // The nesting usvg converts, references followed, not the source depth alone. + let xml_depth = checked.height; let total_depth = used.xml_depth.load(Ordering::Relaxed).checked_add(xml_depth); let Some(total_depth) = total_depth.filter(|depth| *depth <= limits.tree_depth) else { used.svg_limit.fetch_add(1, Ordering::Relaxed); @@ -550,8 +634,8 @@ fn embedded(used: &Used, limits: &Limits, data: &[u8], options: &usvg::Options<' let previous_depth = used.xml_depth.swap(total_depth, Ordering::Relaxed); let depth = used.depth.fetch_add(1, Ordering::Relaxed).saturating_add(1); used.svg_depth.fetch_max(depth, Ordering::Relaxed); - account_invalid_data_urls(&xml, used); - let tree = usvg::Tree::from_xmltree(&xml, options); + account_invalid_data_urls(&checked.xml, used); + let tree = checked.parse(options); used.xml_depth.store(previous_depth, Ordering::Relaxed); used.depth.fetch_sub(1, Ordering::Relaxed); match tree { diff --git a/crates/svg/src/tree/tests.rs b/crates/svg/src/tree/tests.rs index 103c0fe7c..35cb13a2e 100644 --- a/crates/svg/src/tree/tests.rs +++ b/crates/svg/src/tree/tests.rs @@ -292,7 +292,7 @@ mod resource_limit_usage { assert_eq!(svgz::inflate(&data, cap).unwrap().len(), cap, "raw bytes fit exactly"); assert_eq!(svgz::text(&data, cap).unwrap().len(), decoded.len(), "decoding itself succeeds"); let limits = Limits { bytes: cap, embedded_bytes: decoded.len(), ..shallow() }; - assert!(matches!(checked(&decoded, &limits), Err(Error::Text(svgz::Error::TooLarge(n))) if n == cap)); + assert!(matches!(preflight(&decoded, &limits), Err(Error::Text(svgz::Error::TooLarge(n))) if n == cap)); let image = data_image(&data); let (found, usage) = resolved(&svg(&image), limits, 1); assert!(found.is_empty(), "{encoding}, gzip {gzip}"); @@ -312,7 +312,7 @@ mod resource_limit_usage { let decoded = svgz::decode(&raw).unwrap(); let limits = Limits { bytes: decoded.len(), embedded_bytes: raw.len(), ..shallow() }; assert!(decoded.len() > limits.embedded_bytes); - checked(&decoded, &limits).unwrap(); // The independent per-image cap allows this complete XML. + preflight(&decoded, &limits).unwrap(); // The independent per-image cap allows this complete XML. for gzip in [false, true] { let data = if gzip { svgz::compress(&raw) } else { raw.clone() }; assert_eq!(svgz::inflate(&data, limits.embedded_bytes).unwrap().len(), raw.len()); @@ -584,3 +584,104 @@ mod invalid_image_accounting { assert_eq!(report.raster_limit_images, 1); } } + +#[test] +fn preflight_keeps_source_depth_and_parses_the_checked_tree() { + let text = svg(""); + let checked = preflight(&text, &Limits::default()).unwrap(); + // svg and g are open; the self-closing rect opens no level. + assert_eq!(checked.depth(), 2); + let tree = checked.parse(&options()).unwrap(); + assert_eq!(tree.size().width(), 10.0); + assert!(matches!(preflight(&svg(&groups(30)), &shallow()), Err(Error::TooDeep(24)))); +} + +#[test] +fn preflight_matches_parse_text_refusals() { + for text in [""] { + assert!(matches!(preflight(text, &Limits::default()), Err(Error::Xml(_)))); + assert!(matches!(parse_text(text, &Limits::default(), &options()), Err(Error::Xml(_)))); + } +} + +#[test] +fn rewrite_root_viewport_wraps_the_root_and_keeps_the_preamble() { + let source = concat!( + "\n]>\n", + "", + "" + ); + let out = rewrite_root_viewport(source, Viewport { width: 40.0, height: 30.0 }, &Limits::default(), 4096).unwrap(); + assert!(out.starts_with("\n]>\n")); + let tree = parse_text(&out, &Limits::default(), &options()).unwrap(); + assert_eq!((tree.size().width(), tree.size().height()), (40.0, 30.0)); +} + +#[test] +fn rewrite_root_viewport_reads_points_and_pixels() { + let points = rewrite_root_viewport( + "", + Viewport { width: 1.0, height: 1.0 }, + &Limits::default(), + 4096, + ) + .unwrap(); + assert!(points.contains("viewBox=\"0 0 96 48\"")); + let pixels = rewrite_root_viewport( + "", + Viewport { width: 1.0, height: 1.0 }, + &Limits::default(), + 4096, + ) + .unwrap(); + assert!(pixels.contains("viewBox=\"0 0 12 8\"")); +} + +#[test] +fn rewrite_root_viewport_refusals() { + let limits = Limits::default(); + let one = Viewport { width: 1.0, height: 1.0 }; + let sized = ""; + for target in [ + Viewport { width: f64::NAN, height: 1.0 }, + Viewport { width: 1.0, height: f64::INFINITY }, + Viewport { width: 0.0, height: 1.0 }, + Viewport { width: 1.0, height: -2.0 }, + ] { + assert!(matches!(rewrite_root_viewport(sized, target, &limits, 4096), Err(Error::Viewport(_)))); + } + assert!(matches!(rewrite_root_viewport("", one, &limits, 4096), Err(Error::Viewport(_)))); + for source in [ + "", + "", + "", + "", + "", + ] { + assert!(matches!(rewrite_root_viewport(source, one, &limits, 4096), Err(Error::Viewport(_))), "{source}"); + } + assert!(matches!(rewrite_root_viewport(" assert!(parse_text(&out, &limits, &options()).is_ok(), "{n} groups"), + Err(Error::TooDeep(_)) => refused_at_the_limit = true, + Err(e) => panic!("{n} groups: {e:?}"), + } + } + assert!(refused_at_the_limit, "a source at the depth limit is refused, not wrapped past it"); +} diff --git a/crates/svg/src/write.rs b/crates/svg/src/write.rs index 469a4f8fa..66cb1d6a0 100644 --- a/crates/svg/src/write.rs +++ b/crates/svg/src/write.rs @@ -86,18 +86,36 @@ impl Attribute { pub fn translate_scale(x: f64, y: f64, scale_x: f64, scale_y: f64) -> Self { Self { name: "transform".into(), value: Value::TranslateScale { x, y, scale_x, scale_y } } } + /// A translate/scale transform with three decimal places for positions and four for scales. + pub fn translate_scale_fixed(x: f64, y: f64, scale_x: f64, scale_y: f64) -> Self { + Self::text("transform", format!("translate({} {}) scale({} {})", num(x, 3), num(y, 3), num(scale_x, 4), num(scale_y, 4))) + } } impl Scene { /// Serialize the scene as an SVG 1.1 document. pub fn write(&self) -> String { + self.write_impl(false) + } + + /// Serialize the root viewBox with three decimal places for stable fixed-width output. + pub fn write_fixed_view_box(&self) -> String { + self.write_impl(true) + } + + fn write_impl(&self, fixed: bool) -> String { let mut writer = SvgWriter::new(); + let view_box = if fixed { + format!("{} {} {} {}", num(self.view_box[0], 3), num(self.view_box[1], 3), num(self.view_box[2], 3), num(self.view_box[3], 3)) + } else { + view_box_data(self.view_box) + }; writer .root() .attr("version", "1.1") .attr("width", format!("{}mm", num(self.width_mm, 3))) .attr("height", format!("{}mm", num(self.height_mm, 3))) - .attr("viewBox", view_box_data(self.view_box)) + .attr("viewBox", view_box) .open(); write_elements(&mut writer, &self.elements, 1); writer.line(0); @@ -307,6 +325,14 @@ mod tests { p } + #[test] + fn scene_can_keep_fixed_view_box_without_changing_default_output() { + let scene = Scene { width_mm: 100.0, height_mm: 150.0, view_box: [0.0, 0.0, 100.0, 150.0], elements: vec![] }; + assert!(scene.write().contains("viewBox=\"0 0 100 150\"")); + assert!(scene.write_fixed_view_box().contains("viewBox=\"0.000 0.000 100.000 150.000\"")); + assert_eq!(Attribute::translate_scale_fixed(6.0, 46.0, 0.5, -0.5).value, Value::Text("translate(6.000 46.000) scale(0.5000 -0.5000)".into())); + } + #[test] fn numbers_have_fixed_decimals_and_are_never_non_finite() { assert_eq!(num(1.5, 3).to_string(), "1.500"); diff --git a/crates/tcraft/Cargo.toml b/crates/tcraft/Cargo.toml new file mode 100644 index 000000000..3ce35abfb --- /dev/null +++ b/crates/tcraft/Cargo.toml @@ -0,0 +1,17 @@ +[package] +name = "craft-tcraft" +description = "Streaming binary payloads of TextureCraft project archives" +version.workspace = true +edition.workspace = true +license.workspace = true +rust-version.workspace = true +repository.workspace = true + +[lints] +workspace = true + +[dependencies] +craft-zip = { workspace = true } +serde = { workspace = true } +serde_json = { workspace = true } +thiserror = { workspace = true } diff --git a/crates/tcraft/README.md b/crates/tcraft/README.md new file mode 100644 index 000000000..e5831ef98 --- /dev/null +++ b/crates/tcraft/README.md @@ -0,0 +1,14 @@ +# craft-tcraft + +Streaming neutral `TCTX` binary payloads used inside TextureCraft `.tcraft` projects. Sparse +paint data stores 64×64 tiles of little-endian `u16` components; baked maps store dense +little-endian `f32` values. This crate does not yet define the project JSON schema or ZIP +envelope and does not depend on TextureCraft's app types. + +`read_tiles` validates the expected dimensions and components, bounds tile count, rejects +out-of-range and repeated tile positions, and delivers one owned tile at a time. `read_baked` +validates dimensions and declared length and decodes in bounded chunks. Both consume the same +caller-owned `craft_zip::Budget` used for the surrounding project and reject trailing data. +`write_tiles` and `write_baked` stream their bodies rather than copying an entire image; a caller +must discard an incomplete output if a write fails. Limits and ownership rules beyond this +binary format remain with the project reader. diff --git a/crates/tcraft/src/lib.rs b/crates/tcraft/src/lib.rs new file mode 100644 index 000000000..d7af3e8a6 --- /dev/null +++ b/crates/tcraft/src/lib.rs @@ -0,0 +1,428 @@ +#![deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable)] +#![forbid(unsafe_code)] +//! Streaming `TCTX` tile and baked-map payloads in TextureCraft `.tcraft` archives. +//! +//! This codec handles the neutral binary payload, not the application's project model or ZIP +//! envelope. Reads use a caller-owned budget shared with other archive entries. No reader opens +//! a file or reserves memory based on a claimed payload before checking it. + +use std::collections::{BTreeMap, HashSet}; +use std::io::{self, Read, Write}; + +use craft_zip::Budget; +use serde::{Deserialize, Serialize}; +use serde_json::Value; + +pub const TILE_SIDE: u32 = 64; +pub const TILE_AREA: usize = 64 * 64; +pub const HEADER_BYTES: u64 = 24; +pub const VERSION: u32 = 1; +const MAGIC: &[u8; 4] = b"TCTX"; +const CHUNK: usize = 1 << 16; + +#[derive(Debug, thiserror::Error)] +pub enum Error { + #[error("invalid TCTX payload: {0}")] + Invalid(String), + #[error("TCTX version {0} is newer than supported")] + NewerVersion(u32), + #[error("TCTX budget: {0}")] + Budget(#[from] craft_zip::Exceeded), + #[error("TCTX limit: {0}")] + Limit(String), + #[error("TCTX I/O: {0}")] + Io(#[from] io::Error), +} + +/// Neutral top-level project schema. Nested app-domain values and unknown fields remain JSON +/// so they round-trip without requiring TextureCraft's in-memory model. +#[derive(Clone, Debug, PartialEq, Deserialize)] +pub struct ProjectDocument { + pub format_version: u32, + pub name: String, + pub mesh: String, + pub normal_format: Value, + pub texture_sets: Vec, + pub layers: BTreeMap, + pub next_id: u64, + pub display: Value, + #[serde(default)] + pub cameras: Vec, + #[serde(flatten)] + pub extra: BTreeMap, +} + +impl Serialize for ProjectDocument { + fn serialize(&self, serializer: S) -> Result { + use serde::ser::{Error as _, SerializeMap as _}; + const RESERVED: &[&str] = &["format_version", "name", "mesh", "normal_format", "texture_sets", "layers", "next_id", "display", "cameras"]; + if self.extra.keys().any(|key| RESERVED.contains(&key.as_str())) { + return Err(S::Error::custom("an extra project field duplicates a known field")); + } + let mut map = serializer.serialize_map(Some(9 + self.extra.len()))?; + map.serialize_entry("format_version", &self.format_version)?; + map.serialize_entry("name", &self.name)?; + map.serialize_entry("mesh", &self.mesh)?; + map.serialize_entry("normal_format", &self.normal_format)?; + map.serialize_entry("texture_sets", &self.texture_sets)?; + map.serialize_entry("layers", &self.layers)?; + map.serialize_entry("next_id", &self.next_id)?; + map.serialize_entry("display", &self.display)?; + map.serialize_entry("cameras", &self.cameras)?; + for (key, value) in &self.extra { + map.serialize_entry(key, value)?; + } + map.end() + } +} + +/// Read only the version before an application applies its own recovery policy. Serde's +/// structured parse rejects a duplicate or invalid version field, like the project reader. +pub fn project_version(bytes: &[u8]) -> Result { + #[derive(Deserialize)] + struct Version { + format_version: u32, + } + Ok(serde_json::from_slice::(bytes)?.format_version) +} + +/// Width, height and number of components expected by the enclosing project schema. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub struct Layout { + pub width: u32, + pub height: u32, + pub components: usize, +} + +#[derive(Clone, Copy, Debug)] +struct Header { + layout: Layout, + count: u32, +} + +fn charge(budget: &mut Budget, n: u64) -> Result<(), Error> { + budget.charge(n).map_err(Error::from) +} + +fn read_header(reader: &mut (impl Read + ?Sized), budget: &mut Budget) -> Result { + charge(budget, HEADER_BYTES)?; + let mut bytes = [0u8; HEADER_BYTES as usize]; + reader.read_exact(&mut bytes)?; + if bytes.get(..4) != Some(MAGIC.as_slice()) { + return Err(Error::Invalid("bad magic".into())); + } + let word = |i: usize| bytes.get(i * 4..i * 4 + 4).map(|b| u32::from_le_bytes([b[0], b[1], b[2], b[3]])).unwrap_or(0); + let version = word(1); + if version > VERSION { + return Err(Error::NewerVersion(version)); + } + Ok(Header { layout: Layout { width: word(2), height: word(3), components: word(4) as usize }, count: word(5) }) +} + +fn write_header(writer: &mut (impl Write + ?Sized), layout: Layout, count: u32) -> Result<(), Error> { + let components = u32::try_from(layout.components).map_err(|_| Error::Limit("too many components".into()))?; + writer.write_all(MAGIC)?; + for value in [VERSION, layout.width, layout.height, components, count] { + writer.write_all(&value.to_le_bytes())?; + } + Ok(()) +} + +fn expect_end(reader: &mut (impl Read + ?Sized)) -> Result<(), Error> { + let mut byte = [0u8; 1]; + if reader.read(&mut byte)? != 0 { + return Err(Error::Invalid("trailing data".into())); + } + Ok(()) +} + +fn tile_dimensions(layout: Layout) -> Result<(u32, u32, usize), Error> { + if layout.width == 0 || layout.height == 0 || layout.components == 0 { + return Err(Error::Invalid("empty tile layout".into())); + } + let x = layout.width.div_ceil(TILE_SIDE); + let y = layout.height.div_ceil(TILE_SIDE); + let bytes = TILE_AREA.checked_mul(layout.components).and_then(|n| n.checked_mul(2)).ok_or_else(|| Error::Limit("tile is too large".into()))?; + Ok((x, y, bytes)) +} + +/// Decode tiles one at a time, transferring each tile to the callback. +/// Positions and duplicate tiles are checked before delivery; the entry must end exactly. +pub fn read_tiles( + reader: &mut (impl Read + ?Sized), + expected: Layout, + max_texels: u64, + budget: &mut Budget, + mut tile: impl FnMut(u32, u32, Vec) -> Result<(), Error>, +) -> Result<(), Error> { + let header = read_header(reader, budget)?; + let texels = u64::from(expected.width) * u64::from(expected.height); + if texels == 0 || texels > max_texels { + return Err(Error::Limit("tile image exceeds its texel limit".into())); + } + if header.layout != expected { + return Err(Error::Invalid("tile layout differs from the project".into())); + } + let (tiles_x, tiles_y, bytes_per_tile) = tile_dimensions(expected)?; + if u64::from(header.count) > u64::from(tiles_x) * u64::from(tiles_y) { + return Err(Error::Invalid("more tiles than the image holds".into())); + } + let mut bytes = Vec::new(); + let mut seen = HashSet::new(); + for _ in 0..header.count { + let payload = u64::try_from(bytes_per_tile).map_err(|_| Error::Limit("tile is too large".into()))?.saturating_add(8); + charge(budget, payload)?; + if bytes.is_empty() { + bytes.try_reserve_exact(bytes_per_tile).map_err(|_| Error::Limit("tile allocation failed".into()))?; + bytes.resize(bytes_per_tile, 0); + } + let mut pos = [0u8; 8]; + reader.read_exact(&mut pos)?; + let x = u32::from_le_bytes([pos[0], pos[1], pos[2], pos[3]]); + let y = u32::from_le_bytes([pos[4], pos[5], pos[6], pos[7]]); + if x >= tiles_x || y >= tiles_y || seen.contains(&(x, y)) { + return Err(Error::Invalid(format!("bad or repeated tile ({x}, {y})"))); + } + seen.try_reserve(1).map_err(|_| Error::Limit("tile index allocation failed".into()))?; + seen.insert((x, y)); + reader.read_exact(&mut bytes)?; + let mut values = Vec::new(); + values.try_reserve_exact(bytes_per_tile / 2).map_err(|_| Error::Limit("tile allocation failed".into()))?; + values.extend(bytes.as_chunks::<2>().0.iter().map(|b| u16::from_le_bytes(*b))); + tile(x, y, values)?; + } + expect_end(reader) +} + +/// Write tiles in caller order, without buffering the image. A failure may leave a partial +/// entry; callers writing a project should discard the incomplete output. +pub fn write_tiles<'a>( + writer: &mut (impl Write + ?Sized), + layout: Layout, + count: u32, + tiles: impl IntoIterator, +) -> Result<(), Error> { + let (tiles_x, tiles_y, bytes_per_tile) = tile_dimensions(layout)?; + if u64::from(count) > u64::from(tiles_x) * u64::from(tiles_y) { + return Err(Error::Invalid("more tiles than the image holds".into())); + } + write_header(writer, layout, count)?; + let mut seen = HashSet::new(); + let mut bytes = Vec::new(); + let mut written = 0u32; + for (x, y, values) in tiles { + if written >= count || x >= tiles_x || y >= tiles_y || seen.contains(&(x, y)) || values.len() != bytes_per_tile / 2 { + return Err(Error::Invalid("bad, repeated or incorrectly sized tile".into())); + } + seen.try_reserve(1).map_err(|_| Error::Limit("tile index allocation failed".into()))?; + seen.insert((x, y)); + if bytes.is_empty() { + bytes.try_reserve_exact(bytes_per_tile).map_err(|_| Error::Limit("tile allocation failed".into()))?; + } + bytes.clear(); + bytes.extend(values.iter().flat_map(|v| v.to_le_bytes())); + writer.write_all(&x.to_le_bytes())?; + writer.write_all(&y.to_le_bytes())?; + writer.write_all(&bytes)?; + written += 1; + } + if written != count { + return Err(Error::Invalid("tile count differs from its data".into())); + } + Ok(()) +} + +/// Read a dense f32 map, validating declared size and charging its shared budget first. +pub fn read_baked( + reader: &mut (impl Read + ?Sized), + declared_bytes: u64, + components: usize, + max_texels: u64, + budget: &mut Budget, +) -> Result<(Layout, Vec), Error> { + let header = read_header(reader, budget)?; + let layout = header.layout; + let texels = u64::from(layout.width) * u64::from(layout.height); + if texels == 0 || texels > max_texels || layout.components != components || components == 0 { + return Err(Error::Invalid("bad baked-map dimensions or components".into())); + } + let count = texels.checked_mul(components as u64).ok_or_else(|| Error::Limit("too many map values".into()))?; + let payload = count.checked_mul(4).ok_or_else(|| Error::Limit("map is too large".into()))?; + if declared_bytes != HEADER_BYTES.checked_add(payload).ok_or_else(|| Error::Limit("map is too large".into()))? { + return Err(Error::Invalid(format!("{declared_bytes} bytes for a {}×{}×{} map", layout.width, layout.height, layout.components))); + } + charge(budget, payload)?; + let count = usize::try_from(count).map_err(|_| Error::Limit("too many map values".into()))?; + let mut values = Vec::new(); + let chunk_bytes = CHUNK.min(count) * 4; + let mut bytes = Vec::new(); + bytes.try_reserve_exact(chunk_bytes).map_err(|_| Error::Limit("map chunk allocation failed".into()))?; + bytes.resize(chunk_bytes, 0); + while values.len() < count { + let n = CHUNK.min(count - values.len()); + let chunk = bytes.get_mut(..n * 4).ok_or_else(|| Error::Invalid("bad map chunk".into()))?; + reader.read_exact(chunk)?; + values.try_reserve(n).map_err(|_| Error::Limit("map allocation failed".into()))?; + values.extend(chunk.as_chunks::<4>().0.iter().map(|b| f32::from_le_bytes(*b))); + } + expect_end(reader)?; + Ok((layout, values)) +} + +/// Write a dense f32 map in bounded chunks without constructing a second copy of its payload. +pub fn write_baked(writer: &mut (impl Write + ?Sized), layout: Layout, values: &[f32]) -> Result<(), Error> { + let count = u64::from(layout.width).checked_mul(u64::from(layout.height)).and_then(|n| n.checked_mul(layout.components as u64)); + if layout.width == 0 || layout.height == 0 || layout.components == 0 || count != Some(values.len() as u64) { + return Err(Error::Invalid("map dimensions differ from its data".into())); + } + write_header(writer, layout, 0)?; + let mut bytes = Vec::new(); + bytes.try_reserve_exact(CHUNK * 4).map_err(|_| Error::Limit("map chunk allocation failed".into()))?; + for chunk in values.chunks(CHUNK) { + bytes.clear(); + bytes.extend(chunk.iter().flat_map(|v| v.to_le_bytes())); + writer.write_all(&bytes)?; + } + Ok(()) +} + +#[cfg(test)] +mod tests { + use super::*; + use std::io::Cursor; + + fn header(layout: Layout, count: u32) -> Vec { + let mut bytes = Vec::new(); + write_header(&mut bytes, layout, count).unwrap(); + bytes + } + + #[test] + fn project_schema_keeps_nested_and_future_fields() { + let json = serde_json::json!({ + "format_version": 2, "name": "Stone", "mesh": "mesh.obj", "normal_format": "openGl", + "texture_sets": [{"name": "main", "futureSet": {"keep": true}}], + "layers": {"1": {"kind": "paint", "futureLayer": [1, 2]}}, "next_id": 2, + "display": {"futureDisplay": 4}, "mesh_file": "source.obj", "futureProject": [3] + }); + let bytes = serde_json::to_vec(&json).unwrap(); + assert_eq!(project_version(&bytes).unwrap(), 2); + let document: ProjectDocument = serde_json::from_slice(&bytes).unwrap(); + let mut expected = json; + expected["cameras"] = serde_json::json!([]); + assert_eq!(serde_json::to_value(&document).unwrap(), expected); + let mut duplicate = document; + duplicate.extra.insert("mesh".into(), serde_json::json!("other.obj")); + assert!(serde_json::to_vec(&duplicate).is_err()); + assert!(project_version(br#"{"format_version":1,"format_version":2}"#).is_err()); + assert!(project_version(br#"{"format_version":-1}"#).is_err()); + } + + #[test] + fn baked_map_matches_the_native_little_endian_wire_layout() { + let layout = Layout { width: 2, height: 1, components: 1 }; + let mut expected = b"TCTX".to_vec(); + for value in [1u32, 2, 1, 1, 0] { + expected.extend_from_slice(&value.to_le_bytes()); + } + expected.extend_from_slice(&1.0f32.to_le_bytes()); + expected.extend_from_slice(&(-0.5f32).to_le_bytes()); + let mut actual = Vec::new(); + write_baked(&mut actual, layout, &[1.0, -0.5]).unwrap(); + assert_eq!(actual, expected); + let (decoded, values) = read_baked(&mut Cursor::new(&expected), expected.len() as u64, 1, 2, &mut Budget::new("project", 32)).unwrap(); + assert_eq!(decoded, layout); + assert_eq!(values, [1.0, -0.5]); + expected[4..8].copy_from_slice(&2u32.to_le_bytes()); + assert!(matches!(read_baked(&mut Cursor::new(&expected), 32, 1, 2, &mut Budget::new("project", 32)), Err(Error::NewerVersion(2)))); + } + + #[test] + fn dense_maps_round_trip_bit_exactly_in_bounded_chunks() { + let layout = Layout { width: 300, height: 300, components: 1 }; + let mut values = vec![0.25; 90_000]; + values[123] = f32::from_bits(0x7fc0_1234); + let mut bytes = Vec::new(); + write_baked(&mut bytes, layout, &values).unwrap(); + let mut budget = Budget::new("project", bytes.len() as u64); + let (decoded_layout, decoded) = read_baked(&mut Cursor::new(&bytes), bytes.len() as u64, 1, 100_000, &mut budget).unwrap(); + assert_eq!(decoded_layout, layout); + assert!(decoded.iter().zip(values.iter()).all(|(a, b)| a.to_bits() == b.to_bits())); + assert_eq!(budget.remaining(), 0); + } + + #[test] + fn dense_maps_refuse_false_lengths_limits_and_trailing_data() { + let layout = Layout { width: 2, height: 2, components: 1 }; + let mut bytes = Vec::new(); + write_baked(&mut bytes, layout, &[1.0; 4]).unwrap(); + let mut budget = Budget::new("project", 100); + assert!(read_baked(&mut Cursor::new(&bytes), 999, 1, 4, &mut budget).is_err()); + let mut budget = Budget::new("project", 39); + assert!(matches!(read_baked(&mut Cursor::new(&bytes), 40, 1, 4, &mut budget), Err(Error::Budget(_)))); + let mut extra = bytes.clone(); + extra.push(0); + let mut budget = Budget::new("project", 100); + assert!(read_baked(&mut Cursor::new(extra), 40, 1, 4, &mut budget).is_err()); + } + + #[test] + fn hostile_headers_and_tiles_are_rejected_without_trusting_counts() { + let layout = Layout { width: 64, height: 64, components: 1 }; + assert!(matches!( + read_tiles(&mut Cursor::new(header(layout, 0)), layout, 4095, &mut Budget::new("project", 100), |_, _, _| Ok(())), + Err(Error::Limit(_)) + )); + let mut bad_magic = header(layout, 0); + bad_magic[0] = b'X'; + assert!(read_tiles(&mut Cursor::new(bad_magic), layout, u64::MAX, &mut Budget::new("project", 100), |_, _, _| Ok(())).is_err()); + assert!(read_tiles(&mut Cursor::new(header(layout, 2)), layout, u64::MAX, &mut Budget::new("project", 100), |_, _, _| Ok(())).is_err()); + let other = Layout { width: 65, ..layout }; + assert!(read_tiles(&mut Cursor::new(header(layout, 0)), other, u64::MAX, &mut Budget::new("project", 100), |_, _, _| Ok(())).is_err()); + let mut outside = header(layout, 1); + outside.extend_from_slice(&1u32.to_le_bytes()); + outside.extend_from_slice(&0u32.to_le_bytes()); + assert!(read_tiles(&mut Cursor::new(outside), layout, u64::MAX, &mut Budget::new("project", 10_000), |_, _, _| Ok(())).is_err()); + let mut truncated = header(layout, 1); + truncated.extend_from_slice(&[0; 8]); + assert!(read_tiles(&mut Cursor::new(truncated), layout, u64::MAX, &mut Budget::new("project", 10_000), |_, _, _| Ok(())).is_err()); + let mut extra = header(layout, 0); + extra.push(1); + assert!(read_tiles(&mut Cursor::new(extra), layout, u64::MAX, &mut Budget::new("project", 100), |_, _, _| Ok(())).is_err()); + } + + #[test] + fn sparse_tiles_stream_and_reject_duplicates_and_short_budget() { + let layout = Layout { width: 128, height: 64, components: 1 }; + let mut bytes = header(layout, 2); + for x in [0u32, 1] { + bytes.extend_from_slice(&x.to_le_bytes()); + bytes.extend_from_slice(&0u32.to_le_bytes()); + for _ in 0..TILE_AREA { + bytes.extend_from_slice(&(x as u16 + 1).to_le_bytes()); + } + } + let first = vec![1u16; TILE_AREA]; + let second = vec![2u16; TILE_AREA]; + let mut written = Vec::new(); + write_tiles(&mut written, layout, 2, [(0, 0, first.as_slice()), (1, 0, second.as_slice())]).unwrap(); + assert_eq!(written, bytes, "the streamed writer keeps the native wire layout"); + assert!(write_tiles(&mut Vec::new(), layout, 2, [(0, 0, first.as_slice())]).is_err()); + let mut seen = Vec::new(); + let mut budget = Budget::new("project", bytes.len() as u64); + read_tiles(&mut Cursor::new(&bytes), layout, u64::MAX, &mut budget, |x, y, tile| { + seen.push((x, y, tile.first().copied())); + Ok(()) + }) + .unwrap(); + assert_eq!(seen, [(0, 0, Some(1)), (1, 0, Some(2))]); + let mut duplicate = bytes.clone(); + let second = HEADER_BYTES as usize + 8 + TILE_AREA * 2; + duplicate[second..second + 4].copy_from_slice(&0u32.to_le_bytes()); + assert!(read_tiles(&mut Cursor::new(duplicate), layout, u64::MAX, &mut Budget::new("project", u64::MAX), |_, _, _| Ok(())).is_err()); + assert!(matches!( + read_tiles(&mut Cursor::new(bytes), layout, u64::MAX, &mut Budget::new("project", HEADER_BYTES), |_, _, _| Ok(())), + Err(Error::Budget(_)) + )); + } +} diff --git a/crates/vectorcraft-format/Cargo.toml b/crates/vectorcraft-format/Cargo.toml index e85d90886..c10435289 100644 --- a/crates/vectorcraft-format/Cargo.toml +++ b/crates/vectorcraft-format/Cargo.toml @@ -15,3 +15,6 @@ flate2 = { version = "1", default-features = false, features = ["rust_backend"] serde = { workspace = true } serde_json = { workspace = true } thiserror = { workspace = true } + +[dev-dependencies] +proptest = { workspace = true } diff --git a/crates/vectorcraft-format/src/lib.rs b/crates/vectorcraft-format/src/lib.rs index 051c31a70..b180c046c 100644 --- a/crates/vectorcraft-format/src/lib.rs +++ b/crates/vectorcraft-format/src/lib.rs @@ -34,6 +34,7 @@ //! [`sniff`] reads: a reader of a compressed file decompresses it with its own limits first. pub mod base64; +pub mod stream; use std::collections::{BTreeMap, HashMap}; use std::fmt; @@ -769,10 +770,20 @@ impl<'a> Tail<'a> { /// Write what follows the JSON: [`BLOB_MAGIC`] and the entries, or nothing when there are none. pub fn write_to(&self, mut w: impl Write) -> io::Result<()> { + self.write_to_with(&mut w, |_, _| Ok(())) + } + + /// Write the marker and parts, calling `before_part` before each payload. A compressed + /// writer can use the MIME type to switch between compressed and stored deflate blocks. + pub fn write_to_with(&self, mut w: W, mut before_part: impl FnMut(&mut W, &str) -> io::Result<()>) -> io::Result<()> { if self.parts.is_empty() { return Ok(()); } w.write_all(BLOB_MAGIC)?; - self.parts.iter().try_for_each(|p| w.write_all(p.bytes)) + for part in &self.parts { + before_part(&mut w, &part.mime)?; + w.write_all(part.bytes)?; + } + Ok(()) } } diff --git a/crates/vectorcraft-format/src/stream.rs b/crates/vectorcraft-format/src/stream.rs new file mode 100644 index 000000000..2e8b0695e --- /dev/null +++ b/crates/vectorcraft-format/src/stream.rs @@ -0,0 +1,302 @@ +//! Streaming helpers for the native JSON envelope. +//! +//! Field values are emitted in bounded chunks. The decoded JSON and tail framing stay the same +//! as VectorCraft's previous writer, but compressed bytes can differ when a deflate backend's +//! output depends on input chunking. Gzip readers still validate the CRC and size trailer. +//! +//! Copyright (c) 2026 ArtCraft Team and the VectorCraft contributors. +//! Extracted from VectorCraft, licensed under MIT OR Apache-2.0. + +use serde::Serialize; +use std::collections::{BTreeMap, HashSet}; +use std::io::{self, Write}; + +/// Native file metadata for streaming a JSON envelope. Asset entries should be prepared using +/// the same [`crate::Tail`] whose parts will be written after the JSON. +pub struct FileHeader<'a> { + /// Native format version to write. + pub version: u32, + /// Application generator string. + pub generator: &'a str, + /// Optional saved PDF compatibility preference. + pub pdf_compatible: Option, + /// Optional inline preview. + pub preview: Option<&'a crate::Image>, + /// Embedded profiles keyed by profile name. + pub profiles: &'a BTreeMap<&'a str, crate::Image>, + /// Optional PDF entry. + pub pdf: Option<&'a crate::Image>, +} + +/// Write the native JSON envelope, streaming the document through `document`. The callback +/// returns the image entries after serializing the document, allowing it to collect image keys +/// without keeping a copy of the document JSON. It must write one complete JSON value. +pub fn write_file_json<'a>( + mut out: impl Write, + header: &FileHeader<'a>, + pretty: bool, + document: impl FnOnce(&mut dyn Write) -> io::Result>, +) -> io::Result<()> { + if !(crate::MIN_VERSION..=crate::VERSION).contains(&header.version) { + return Err(io::Error::new(io::ErrorKind::InvalidInput, "unsupported native file version")); + } + let mut env = Envelope::new(pretty); + let format = if header.version < 3 { crate::LEGACY_EXTENSION } else { crate::EXTENSION }; + env.field(&mut out, "format", &format)?; + env.field(&mut out, "version", &header.version)?; + env.field(&mut out, "generator", &header.generator)?; + if let Some(compatible) = header.pdf_compatible { + env.field(&mut out, "pdfCompatible", &compatible)?; + } + if let Some(preview) = header.preview { + env.field(&mut out, "preview", preview)?; + } + if !header.profiles.is_empty() { + env.field(&mut out, "profiles", header.profiles)?; + } + if let Some(pdf) = header.pdf { + env.field(&mut out, "pdf", pdf)?; + } + env.key(&mut out, "document")?; + let images = document(&mut Indent { inner: &mut out, on: pretty })?; + env.field(&mut out, "images", &images)?; + env.finish(&mut out) +} + +/// Indents JSON text written through it one level (two spaces after each newline), as a member +/// of the file's top-level object; newlines only occur between tokens, inside strings they are +/// escaped. +pub struct Indent { + pub inner: W, + pub on: bool, +} + +impl Write for Indent { + fn write(&mut self, buf: &[u8]) -> io::Result { + self.write_all(buf)?; + Ok(buf.len()) + } + + fn write_all(&mut self, buf: &[u8]) -> io::Result<()> { + if !self.on { + return self.inner.write_all(buf); + } + let mut lines = buf.split(|&b| b == b'\n'); + self.inner.write_all(lines.next().unwrap_or_default())?; + for line in lines { + self.inner.write_all(b"\n ")?; + self.inner.write_all(line)?; + } + Ok(()) + } + + fn flush(&mut self) -> io::Result<()> { + self.inner.flush() + } +} + +/// The member that names an image. +const TAG: &[u8] = b"\"key\":"; + +/// Passes JSON text through, collecting the string values of every `"key": "…"` member at any +/// depth (what the images a document uses are named by). Quotes inside JSON strings are escaped, +/// so every match is a real member. +pub struct KeyScan { + pub inner: W, + pub keys: HashSet, + /// Looking for keys (else it only passes the text on). + pub active: bool, + /// How much of [`TAG`] the last bytes matched. + matched: usize, + /// Right after a tag: skipping whitespace to its value. + after_tag: bool, + /// The value string so far, quotes included. + value: Option>, + escaped: bool, +} + +impl KeyScan { + pub fn new(inner: W) -> Self { + Self { inner, keys: HashSet::new(), active: true, matched: 0, after_tag: false, value: None, escaped: false } + } + + fn scan(&mut self, buf: &[u8]) { + if !self.active { + return; + } + let mut i = 0; + while let Some(&b) = buf.get(i) { + // Nothing in progress: skip to the next quote. + if self.matched == 0 && !self.after_tag && self.value.is_none() { + match buf.get(i..).and_then(|r| r.iter().position(|&c| c == b'"')) { + Some(at) => i += at, + None => return, + } + self.matched = 1; + i += 1; + continue; + } + self.step(b); + i += 1; + } + } + + fn step(&mut self, b: u8) { + if let Some(v) = self.value.as_mut() { + v.push(b); + if self.escaped { + self.escaped = false; + } else if b == b'\\' { + self.escaped = true; + } else if b == b'"' + && let Some(v) = self.value.take() + && let Ok(s) = serde_json::from_slice::(&v) + { + self.keys.insert(s); + } + } else if self.after_tag && !b.is_ascii_whitespace() { + self.after_tag = false; + if b == b'"' { + self.value = Some(vec![b]); + } + } + if TAG.get(self.matched) == Some(&b) { + self.matched += 1; + if self.matched == TAG.len() { + self.matched = 0; + self.after_tag = true; + self.value = None; + self.escaped = false; + } + } else { + self.matched = usize::from(b == b'"'); + } + } +} + +impl Write for KeyScan { + fn write(&mut self, buf: &[u8]) -> io::Result { + self.write_all(buf)?; + Ok(buf.len()) + } + + fn write_all(&mut self, buf: &[u8]) -> io::Result<()> { + self.scan(buf); + self.inner.write_all(buf) + } + + fn flush(&mut self) -> io::Result<()> { + self.inner.flush() + } +} + +/// The file's top-level object, written member by member (the document is streamed in). +pub struct Envelope { + pretty: bool, + first: bool, +} + +impl Envelope { + /// Start an object with optional indentation. + pub fn new(pretty: bool) -> Self { + Self { pretty, first: true } + } + + /// Serialize a member directly to the output without buffering the full value. + pub fn field(&mut self, out: &mut impl Write, key: &str, value: &impl Serialize) -> io::Result<()> { + self.key(out, key)?; + // serde_json emits small writes; keep compression and underlying I/O efficient while + // bounding temporary memory independently of the serialized table size. + let mut buffered = io::BufWriter::with_capacity(64 << 10, out); + let indented = Indent { inner: &mut buffered, on: self.pretty }; + if self.pretty { serde_json::to_writer_pretty(indented, value) } else { serde_json::to_writer(indented, value) }.map_err(io::Error::other)?; + buffered.into_inner().map(|_| ()).map_err(|e| e.into_error()) + } + + /// The start of member `key` (a plain name): its value follows, indented one level when + /// pretty (as serde_json would: see [`Indent`]). + pub fn key(&mut self, out: &mut impl Write, key: &str) -> io::Result<()> { + let sep: &[u8] = if std::mem::take(&mut self.first) { b"{" } else { b"," }; + let indent: &[u8] = if self.pretty { b"\n " } else { b"" }; + out.write_all(sep)?; + out.write_all(indent)?; + serde_json::to_writer(&mut *out, key).map_err(io::Error::other)?; + out.write_all(if self.pretty { b": " } else { b":" }) + } + + /// Close the object. + pub fn finish(self, out: &mut impl Write) -> io::Result<()> { + if self.first { + return out.write_all(b"{}"); + } + out.write_all(if self.pretty { b"\n}" } else { b"}" }) + } +} + +#[cfg(test)] +mod tests { + use proptest::prelude::*; + + use super::*; + + /// How the keys were found before saves were streamed: in the whole JSON text at once. + fn string_members(json: &[u8]) -> HashSet { + let mut out = HashSet::new(); + let mut rest = json; + while let Some(at) = rest.windows(TAG.len()).position(|w| w == TAG) { + rest = rest.get(at + TAG.len()..).unwrap_or_default(); + let value = rest.trim_ascii_start(); + if value.first() == Some(&b'"') + && let Some(Ok(s)) = serde_json::Deserializer::from_slice(value).into_iter::().next() + { + out.insert(s); + } + } + out + } + + /// The keys `KeyScan` finds in `json` written in pieces of `chunk` bytes. + fn scanned(json: &[u8], chunk: usize) -> HashSet { + let mut scan = KeyScan::new(Vec::new()); + for piece in json.chunks(chunk.max(1)) { + scan.write_all(piece).unwrap(); + } + assert_eq!(scan.inner, json, "the text passes through"); + scan.keys + } + + #[test] + fn finds_the_keys_the_whole_text_search_found() { + let v = serde_json::json!({"key": "a", "x": [{"key": "b \"q\" \\ é"}, {"monkey": "no"}, {"k": "key"}], "s": "\"key\": \"no\"", "key2": "n", "o": {"key": 3}}); + for json in [serde_json::to_vec(&v).unwrap(), serde_json::to_vec_pretty(&v).unwrap()] { + let want = string_members(&json); + assert_eq!(want, HashSet::from(["a".to_string(), "b \"q\" \\ é".to_string()])); + for chunk in [1, 2, 3, 7, 64, json.len()] { + assert_eq!(scanned(&json, chunk), want, "chunk {chunk}"); + } + } + } + + fn arb_json() -> impl Strategy { + let key = prop::sample::select(vec!["key", "k", "kkey", "key ", "x"]).prop_map(str::to_string); + let leaf = prop_oneof![ + Just(serde_json::Value::Null), + any::().prop_map(serde_json::Value::from), + "[a-z\"\\\\: ké\n]{0,8}".prop_map(serde_json::Value::from), + ]; + leaf.prop_recursive(4, 32, 6, move |inner| { + prop_oneof![ + prop::collection::vec(inner.clone(), 0..5).prop_map(serde_json::Value::from), + prop::collection::vec((key.clone(), inner), 0..5).prop_map(|m| serde_json::Value::Object(m.into_iter().collect())), + ] + }) + } + + proptest! { + #[test] + fn scanning_in_pieces_finds_what_the_whole_text_search_found(v in arb_json(), chunk in 1usize..40, pretty in any::()) { + let json = if pretty { serde_json::to_vec_pretty(&v) } else { serde_json::to_vec(&v) }.unwrap(); + prop_assert_eq!(scanned(&json, chunk), string_members(&json)); + } + } +} diff --git a/crates/vectorcraft-format/tests/stream.rs b/crates/vectorcraft-format/tests/stream.rs new file mode 100644 index 000000000..98f4685e2 --- /dev/null +++ b/crates/vectorcraft-format/tests/stream.rs @@ -0,0 +1,182 @@ +use craft_vectorcraft_format::stream::{Envelope, Indent, KeyScan}; +use std::io::{self, Write}; + +#[test] +fn envelope_streams_nested_json_and_escapes_keys() { + for pretty in [false, true] { + let mut bytes = Vec::new(); + let mut env = Envelope::new(pretty); + env.field(&mut bytes, "quoted\"key", &vec![1, 2]).unwrap(); + env.key(&mut bytes, "document").unwrap(); + let mut output = Indent { inner: &mut bytes, on: pretty }; + if pretty { + serde_json::to_writer_pretty(&mut output, &serde_json::json!({"nested": [3,4]})).unwrap(); + } else { + serde_json::to_writer(&mut output, &serde_json::json!({"nested": [3,4]})).unwrap(); + } + env.finish(&mut bytes).unwrap(); + let value: serde_json::Value = serde_json::from_slice(&bytes).unwrap(); + assert_eq!(value, serde_json::json!({"quoted\"key": [1, 2], "document": {"nested": [3,4]}})); + let mut expected = Vec::new(); + let empty = Envelope::new(pretty); + empty.finish(&mut expected).unwrap(); + assert_eq!(expected, b"{}"); + } +} + +#[test] +fn scan_survives_every_chunk_boundary_and_forwards_data() { + let json = br#"{"key":"one","nested":[{"key":"two \\\"quoted\\\""},{"key":42}],"other":"x"}"#; + for chunk in 1..=json.len() { + let mut scan = KeyScan::new(Vec::new()); + for part in json.chunks(chunk) { + scan.write_all(part).unwrap(); + } + assert_eq!(scan.inner, json); + assert!(scan.keys.contains("one")); + assert_eq!(scan.keys.len(), 2); + } +} + +struct Broken; +impl Write for Broken { + fn write(&mut self, _: &[u8]) -> io::Result { + Err(io::Error::other("failure")) + } + fn flush(&mut self) -> io::Result<()> { + Err(io::Error::other("failure")) + } +} +#[test] +fn forwards_writer_errors() { + assert!(Envelope::new(false).field(&mut Broken, "x", &1).is_err()); + assert!(KeyScan::new(Broken).write_all(b"{}").is_err()); + assert!(Indent { inner: Broken, on: true }.write_all(b"x\ny").is_err()); +} + +#[test] +fn tail_calls_hook_in_order_and_propagates_errors() { + use craft_vectorcraft_format::{BLOB_MAGIC, INLINE_MAX, Tail}; + let first = vec![1; INLINE_MAX]; + let second = vec![2; INLINE_MAX]; + let mut tail = Tail::new(4); + tail.image("image/png", &first, false); + tail.image("application/pdf", &second, false); + let mut out = Vec::new(); + let mut seen = Vec::new(); + tail.write_to_with(&mut out, |_, mime| { + seen.push(mime.to_owned()); + Ok(()) + }) + .unwrap(); + assert_eq!(seen, ["image/png", "application/pdf"]); + assert_eq!(out, [BLOB_MAGIC, &first, &second].concat()); + assert!(tail.write_to_with(Vec::new(), |_, _| Err(io::Error::other("failure"))).is_err()); + let mut called = false; + Tail::new(4) + .write_to_with(Vec::new(), |_, _| { + called = true; + Ok(()) + }) + .unwrap(); + assert!(!called); +} + +#[test] +fn native_schema_order_and_versions() { + use craft_vectorcraft_format::stream::{FileHeader, write_file_json}; + let profiles = Default::default(); + for (version, name) in [(1, "drawcraft"), (4, "vectorcraft")] { + for pretty in [false, true] { + let header = FileHeader { version, generator: "test", pdf_compatible: Some(false), preview: None, profiles: &profiles, pdf: None }; + let mut out = Vec::new(); + write_file_json(&mut out, &header, pretty, |w| { + w.write_all(if pretty { b"{\n \"key\": \"image\"\n}" } else { b"{\"key\":\"image\"}" })?; + Ok(Default::default()) + }) + .unwrap(); + let value: serde_json::Value = serde_json::from_slice(&out).unwrap(); + assert_eq!(value["format"], name); + assert_eq!(value["pdfCompatible"], false); + let text = std::str::from_utf8(&out).unwrap(); + assert!(text.find("\"format\"").unwrap() < text.find("\"version\"").unwrap()); + assert!(text.find("\"document\"").unwrap() < text.find("\"images\"").unwrap()); + } + } + let header = FileHeader { version: 0, generator: "test", pdf_compatible: None, preview: None, profiles: &profiles, pdf: None }; + assert!(write_file_json(Vec::new(), &header, false, |_| Ok(Default::default())).is_err()); +} + +// Oracle for the previous app writer: each member was first serialized to a Vec, then +// indented and appended. Compare decoded file bytes, not backend-dependent gzip bytes. +#[test] +fn streamed_native_file_matches_previous_framing() { + use craft_vectorcraft_format::{ + Tail, + stream::{FileHeader, write_file_json}, + }; + use flate2::{Compression, read::GzDecoder, write::GzEncoder}; + use std::collections::BTreeMap; + use std::io::Read; + + fn old_field(out: &mut Vec, first: &mut bool, pretty: bool, key: &str, value: &impl serde::Serialize) { + out.extend_from_slice(if std::mem::take(first) { b"{" } else { b"," }); + if pretty { + out.extend_from_slice(b"\n "); + } + out.extend_from_slice(format!("\"{key}\"{}", if pretty { ": " } else { ":" }).as_bytes()); + let json = if pretty { serde_json::to_vec_pretty(value) } else { serde_json::to_vec(value) }.unwrap(); + let mut indent = Indent { inner: out, on: pretty }; + indent.write_all(&json).unwrap(); + } + + let payload = vec![42; craft_vectorcraft_format::INLINE_MAX]; + for version in [1, 3, 4] { + for pretty in [false, true] { + let mut tail = Tail::new(version); + let image = tail.image("image/png", &payload, false); + let profile = tail.image("application/vnd.iccprofile", &payload, false); + let pdf = tail.image("application/pdf", &payload, false); + let profiles = BTreeMap::from([("profile", profile)]); + let images = BTreeMap::from([("image", image)]); + let document = serde_json::json!({"layers": [{"key": "image", "name": "A"}]}); + let header = FileHeader { + version, + generator: "VectorCraft test", + pdf_compatible: Some(true), + preview: None, + profiles: &profiles, + pdf: Some(&pdf), + }; + let mut actual = Vec::new(); + write_file_json(&mut actual, &header, pretty, |w| { + if pretty { serde_json::to_writer_pretty(w, &document) } else { serde_json::to_writer(w, &document) }.map_err(io::Error::other)?; + Ok(images.clone()) + }) + .unwrap(); + tail.write_to(&mut actual).unwrap(); + + let mut old = Vec::new(); + let mut first = true; + let format = if version < 3 { "drawcraft" } else { "vectorcraft" }; + old_field(&mut old, &mut first, pretty, "format", &format); + old_field(&mut old, &mut first, pretty, "version", &version); + old_field(&mut old, &mut first, pretty, "generator", &"VectorCraft test"); + old_field(&mut old, &mut first, pretty, "pdfCompatible", &true); + old_field(&mut old, &mut first, pretty, "profiles", &profiles); + old_field(&mut old, &mut first, pretty, "pdf", &pdf); + old_field(&mut old, &mut first, pretty, "document", &document); + old_field(&mut old, &mut first, pretty, "images", &images); + old.extend_from_slice(if pretty { b"\n}" } else { b"}" }); + tail.write_to(&mut old).unwrap(); + assert_eq!(actual, old, "version {version}, pretty {pretty}"); + for bytes in [&old, &actual] { + let mut gzip = GzEncoder::new(Vec::new(), Compression::default()); + gzip.write_all(bytes).unwrap(); + let mut decoded = Vec::new(); + GzDecoder::new(gzip.finish().unwrap().as_slice()).read_to_end(&mut decoded).unwrap(); + assert_eq!(decoded, old); + } + } + } +} diff --git a/crates/video/Cargo.toml b/crates/video/Cargo.toml index c3ebfd722..dae21e5b1 100644 --- a/crates/video/Cargo.toml +++ b/crates/video/Cargo.toml @@ -34,6 +34,8 @@ mpeg2-encode = ["mpeg2-decode"] # MPEG-4 Part 2 video (ISO/IEC 14496-2, the Xvid/DivX format): Simple profile encoder. mpeg4 = ["mpeg4-encode"] mpeg4-encode = ["craft-media/bitstream"] +# FFV1 lossless video (RFC 9043): version 3, range coder, decoder and encoder. +ffv1 = [] # Apple ProRes (SMPTE RDD 36): decoder and encoder. prores = ["craft-media/bitstream"] # H.264 / AVC: decoder and encoder; `h264-decode` or `h264-encode` for one of them. @@ -143,6 +145,10 @@ required-features = ["mpeg4-encode"] name = "mpeg2" required-features = ["mpeg2-decode"] +[[test]] +name = "ffv1" +required-features = ["ffv1"] + [[test]] name = "prores" required-features = ["prores"] diff --git a/crates/video/src/ffv1/README.md b/crates/video/src/ffv1/README.md new file mode 100644 index 000000000..90312dbe5 --- /dev/null +++ b/crates/video/src/ffv1/README.md @@ -0,0 +1,52 @@ +# FFV1 + +Clean-room FFV1 lossless video codec, written from RFC 9043. Feature `ffv1`. + +## What is supported + +| | Decoder | Encoder | +|---|---|---| +| Version | 3 (the one with a configuration record) | 3 | +| Coder | range coder, default or custom state transition table | range coder, default table | +| Colour | YCbCr (any 2^n chroma subsampling, gray, optional transparency), RGB through the JPEG 2000 reversible colour transform (optional transparency) | the same, RGB up to 15 bits | +| Depth | 8 to 16 bits per sample | 8 to 16 bits (RGB 8 to 15) | +| Slices | any raster whose boundaries are multiples of the chroma subsampling | the same | +| Frames | keyframes | keyframes only (`intra` 1) | +| Error detection | slice and configuration record CRCs, checked | written | + +Not supported, and reported as `Error::Unsupported`: the Golomb-Rice coder (`coder_type` 0), +versions 0 and 1 (parameters inside each keyframe, no configuration record), inter frames +(`keyframe` 0), and slice rasters whose boundaries fall inside a chroma sample. + +## Using it + +The configuration record goes in the container (Matroska `CodecPrivate`, the AVI `strf` +extension, an MP4 `glbl` box); each frame is one packet. + +```text +let enc = Encoder::new(&template_frame, EncoderConfig::default())?; +let config = enc.config_record(); +let packet = enc.encode(&frame)?; + +let dec = Decoder::new(&config, width, height)?; +let frame = dec.decode(&packet)?; +``` + +## Notes on the format + +- The Parameters in the configuration record are range coded with the *default* state transition + table. A custom table (`coder_type` 2) applies to the slices only. +- Each Quantization Table starts from fresh states; a Quantization Table Set is five of them. +- A slice is one range-coded stream: for the first slice the `keyframe` bit, then the slice header, + then the samples. Cb and Cr share one set of context states; transparency has its own. +- The context is the sum of five quantised differences. With the default tables this encoder + writes 15125 contexts; FFmpeg refuses tables that multiply to more than 32768, so the encoder + stays below that although the RFC's limit (on half the product) is looser. +- The encoder ends a stream with the top of its final range, so bytes after it (a CRC in the + configuration record) cannot change what decodes. + +## Verification + +`tests/ffv1.rs`: round trips over every shape, hostile input (every truncation and bit flip), +and ffmpeg as the oracle in both directions (skipped without ffmpeg, a failure with +`CRAFT_MEDIA_REQUIRE_ORACLES=1`). diff --git a/crates/video/src/ffv1/mod.rs b/crates/video/src/ffv1/mod.rs new file mode 100644 index 000000000..4df60f9a1 --- /dev/null +++ b/crates/video/src/ffv1/mod.rs @@ -0,0 +1,929 @@ +//! FFV1 lossless video codec (RFC 9043): version 3, range coder, intra-only, YCbCr and RGB (JPEG 2000 +//! RCT), 8 to 16 bits per sample, any number of slices. +//! +//! - [`Decoder`]: built from the *configuration record* a container stores (Matroska `CodecPrivate`, +//! AVI `strf` extension, MP4 `glbl`) and the frame size; [`Decoder::decode`] turns one packet into a +//! [`Frame`]. Slice CRCs are checked. +//! - [`Encoder`]: [`Encoder::config_record`] gives the bytes for the container, [`Encoder::encode`] +//! one packet per frame; every frame is a keyframe. +//! +//! Not supported (reported as [`Error::Unsupported`]): the Golomb-Rice coder (`coder_type` 0), +//! versions 0 and 1 (no configuration record, parameters in every keyframe), inter frames +//! (`intra` 0 with `keyframe` 0), and slice rasters whose boundaries are not multiples of the +//! chroma subsampling. See `src/ffv1/README.md`. +//! +//! Implemented from RFC 9043. Frames are planar `u16`: for YCbCr, `[Y, Cb, Cr]` (plus an optional +//! transparency plane); for RGB, `[R, G, B]` (plus transparency), all at full size. + +mod rc; + +use crate::budget::{DECODE_MEMORY_BUDGET, PictureShape}; +use rc::{Decoder as Rac, Encoder as RacEnc, Transitions}; + +const CONTEXT_SIZE: usize = 32; +const MAX_CONTEXT_INPUTS: usize = 5; +const MAX_CONTEXTS: u64 = 32768; + +/// Codec errors. +#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)] +pub enum Error { + /// The data ends early or its sizes do not add up. + #[error("truncated FFV1 data")] + Truncated, + /// The data is not valid FFV1. + #[error("invalid FFV1 data: {0}")] + Invalid(&'static str), + /// Valid FFV1 that this module does not decode. + #[error("unsupported FFV1 feature: {0}")] + Unsupported(&'static str), + /// A slice or the configuration record failed its CRC. + #[error("FFV1 CRC mismatch in {0}")] + Crc(&'static str), + /// The encoder was given a frame or settings it cannot encode. + #[error("invalid encoder input: {0}")] + Input(String), +} + +/// `Result` of this module. +pub type Result = std::result::Result; + +/// The colour space coded (RFC 9043 table 8). +#[derive(Debug, Clone, Copy, PartialEq, Eq)] +pub enum Colorspace { + /// YCbCr, coded plane by plane without a pixel transform. + YCbCr, + /// RGB through the JPEG 2000 reversible colour transform, coded line by line. + Rgb, +} + +/// One frame of planar samples. +#[derive(Debug, Clone, PartialEq, Eq)] +pub struct Frame { + /// Width in pixels. + pub width: u32, + /// Height in pixels. + pub height: u32, + /// Colour space of the planes. + pub colorspace: Colorspace, + /// Bits per sample, 8 to 16. + pub bit_depth: u8, + /// log2 of the horizontal and vertical chroma subsampling (YCbCr only; 0, 0 is 4:4:4). + pub chroma_shift: (u8, u8), + /// The planes, in the order the module doc gives. + pub planes: Vec>, +} + +impl Frame { + /// Whether the planes are Y, Cb, Cr (or R, G, B) rather than Y (gray) alone, or with transparency. + fn plane_count_ok(&self) -> bool { + matches!(self.planes.len(), 1..=4) + } + + fn dims(&self, plane: usize, chroma_planes: bool) -> (usize, usize) { + plane_dims(self.width as usize, self.height as usize, self.colorspace, self.chroma_shift, plane, chroma_planes) + } +} + +/// `(width, height)` of plane `p` for a full frame (or a slice) of `w` x `h`. +fn plane_dims(w: usize, h: usize, cs: Colorspace, shift: (u8, u8), p: usize, chroma_planes: bool) -> (usize, usize) { + if cs == Colorspace::YCbCr && chroma_planes && (p == 1 || p == 2) { (w.div_ceil(1 << shift.0), h.div_ceil(1 << shift.1)) } else { (w, h) } +} + +/// One Quantization Table Set (section 4.1). +#[derive(Clone)] +struct QuantSet { + tables: Vec<[i32; 256]>, + context_count: usize, +} + +/// The parsed "Parameters" of a stream. +#[derive(Clone)] +struct Params { + coder_type: u32, + transitions: Transitions, + colorspace: Colorspace, + bits: u32, + chroma_planes: bool, + shift: (u8, u8), + extra_plane: bool, + num_h: u32, + num_v: u32, + sets: Vec, + initial_states: Vec>>, + ec: bool, + intra: bool, +} + +impl Params { + fn plane_count(&self) -> usize { + 1 + if self.chroma_planes { 2 } else { 0 } + usize::from(self.extra_plane) + } + + /// Quantization Table Set index slot of plane `p` (section 3.6). + fn qset_slot(&self, p: usize) -> usize { + match p { + 0 => 0, + 1 | 2 if self.chroma_planes => 1, + _ => 2, + } + } +} + +fn crc32(data: &[u8]) -> u32 { + // The IEEE polynomial 0x04C11DB7, initial value 0, no reflection, no final inversion. + let mut crc = 0u32; + for &b in data { + crc ^= u32::from(b) << 24; + for _ in 0..8 { + crc = if crc & 0x8000_0000 != 0 { crc << 1 ^ 0x04C1_1DB7 } else { crc << 1 }; + } + } + crc +} + +// ---- Parameters ---- + +fn read_quant_set(rac: &mut Rac, t: &Transitions) -> Result { + let mut scale: u64 = 1; + let mut tables = Vec::with_capacity(MAX_CONTEXT_INPUTS); + for _ in 0..MAX_CONTEXT_INPUTS { + let mut table = [0i32; 256]; + // Each quantization table starts from fresh states. + let mut state = [128u8; CONTEXT_SIZE]; + let mut v: i64 = 0; + let mut k = 0usize; + while k < 128 { + let len = rac.symbol(&mut state, false, t) as usize + 1; + if k + len > 128 { + return Err(Error::Invalid("quantization table run passes its half")); + } + for _ in 0..len { + table[k] = i32::try_from(scale as i64 * v).map_err(|_| Error::Invalid("quantization table overflows"))?; + k += 1; + } + v += 1; + } + for k in 1..128 { + table[256 - k] = -table[k]; + } + table[128] = -table[127]; + tables.push(table); + scale = scale.checked_mul(2 * v as u64 - 1).filter(|&s| s <= 2 * MAX_CONTEXTS).ok_or(Error::Invalid("too many contexts"))?; + } + let context_count = scale.div_ceil(2); + if context_count > MAX_CONTEXTS { + return Err(Error::Invalid("too many contexts")); + } + Ok(QuantSet { tables, context_count: context_count as usize }) +} + +fn parse_params(rac: &mut Rac) -> Result { + let mut state = [128u8; CONTEXT_SIZE]; + // Until the custom table is known, the default table decodes the fields before it. + let t0 = Transitions::default_table(); + let version = rac.symbol(&mut state, false, &t0); + if version != 3 { + return Err(Error::Unsupported("only FFV1 version 3 has a configuration record; versions 0 and 1 are not supported")); + } + let _micro = rac.symbol(&mut state, false, &t0); + let coder_type = rac.symbol(&mut state, false, &t0) as u32; + let mut delta = [0i32; 256]; + if coder_type > 1 { + for d in delta.iter_mut().skip(1) { + *d = rac.symbol(&mut state, true, &t0); + } + } + let transitions = Transitions::new(&delta); + if coder_type == 0 { + return Err(Error::Unsupported("the Golomb-Rice coder (coder_type 0)")); + } + if coder_type > 2 { + return Err(Error::Invalid("unknown coder_type")); + } + let colorspace = match rac.symbol(&mut state, false, &t0) { + 0 => Colorspace::YCbCr, + 1 => Colorspace::Rgb, + _ => return Err(Error::Unsupported("colorspace_type")), + }; + let bits = rac.symbol(&mut state, false, &t0) as u32; + let bits = if bits == 0 { 8 } else { bits }; + let chroma_planes = rac.bit(&mut state[0], &t0); + let (h, v) = (rac.symbol(&mut state, false, &t0), rac.symbol(&mut state, false, &t0)); + let extra_plane = rac.bit(&mut state[0], &t0); + let num_h = rac.symbol(&mut state, false, &t0) as u32 + 1; + let num_v = rac.symbol(&mut state, false, &t0) as u32 + 1; + let set_count = rac.symbol(&mut state, false, &t0); + if !(8..=16).contains(&bits) || h > 3 || v > 3 || h < 0 || v < 0 || num_h > 1024 || num_v > 1024 || !(1..=8).contains(&set_count) { + return Err(Error::Invalid("parameter out of range")); + } + if colorspace == Colorspace::Rgb && (!chroma_planes || h != 0 || v != 0) { + return Err(Error::Unsupported("RGB with chroma_planes 0 or subsampling")); + } + let mut sets = Vec::new(); + for _ in 0..set_count { + // The Parameters are range coded with the default table; the custom one is for the slices. + sets.push(read_quant_set(rac, &t0)?); + } + let mut initial_states = Vec::new(); + for set in &sets { + if rac.bit(&mut state[0], &t0) { + let mut all = Vec::with_capacity(set.context_count); + let mut prev = [128u8; CONTEXT_SIZE]; + for _ in 0..set.context_count { + let mut cur = [0u8; CONTEXT_SIZE]; + for k in 0..CONTEXT_SIZE { + let d = rac.symbol(&mut state, true, &t0); + cur[k] = (i32::from(prev[k]) + d) as u8; + } + prev = cur; + all.push(cur); + } + initial_states.push(Some(all)); + } else { + initial_states.push(None); + } + } + let ec = rac.symbol(&mut state, false, &t0) != 0; + let intra = rac.symbol(&mut state, false, &t0) != 0; + Ok(Params { + coder_type, + transitions, + colorspace, + bits, + chroma_planes, + shift: (h as u8, v as u8), + extra_plane, + num_h, + num_v, + sets, + initial_states, + ec, + intra, + }) +} + +/// The default quantisation tables of this encoder: 11 levels for the two gradients next to the +/// sample and 5 for the other three, which makes 15125 contexts (FFmpeg refuses more than 32768). +fn default_quant_set() -> QuantSet { + // |difference| thresholds: the first |d| of each level above 0 (levels 0..=5 and 0..=2). + let first_of_level: [&[usize]; MAX_CONTEXT_INPUTS] = [&[1, 2, 4, 8, 16], &[1, 2, 4, 8, 16], &[1, 3], &[1, 3], &[1, 3]]; + let mut scale: i32 = 1; + let mut tables = Vec::new(); + let mut counts = [0usize; MAX_CONTEXT_INPUTS]; + for (j, firsts) in first_of_level.iter().enumerate() { + let mut table = [0i32; 256]; + for (k, slot) in table.iter_mut().take(128).enumerate() { + let level = firsts.iter().filter(|&&f| k >= f).count() as i32; + *slot = scale * level; + } + for k in 1..128 { + table[256 - k] = -table[k]; + } + table[128] = -table[127]; + counts[j] = firsts.len() + 1; + scale *= 2 * counts[j] as i32 - 1; + tables.push(table); + } + QuantSet { tables, context_count: (scale as usize).div_ceil(2) } +} + +/// Run lengths (`len - 1` per step) that describe the first half of a table, as written to the stream. +fn quant_runs(table: &[i32; 256]) -> Vec { + let mut runs = Vec::new(); + let mut k = 0usize; + while k < 128 { + let start = k; + while k < 128 && table[k] == table[start] { + k += 1; + } + runs.push((k - start) as i32 - 1); + } + runs +} + +fn write_params(enc: &mut RacEnc, p: &Params) { + let t = Transitions::default_table(); + let mut state = [128u8; CONTEXT_SIZE]; + enc.symbol(&mut state, 3, false, &t); // version + enc.symbol(&mut state, 4, false, &t); // micro_version: first stable variant + enc.symbol(&mut state, p.coder_type as i32, false, &t); + enc.symbol(&mut state, if p.colorspace == Colorspace::YCbCr { 0 } else { 1 }, false, &t); + enc.symbol(&mut state, p.bits as i32, false, &t); + enc.bit(p.chroma_planes, &mut state[0], &t); + enc.symbol(&mut state, i32::from(p.shift.0), false, &t); + enc.symbol(&mut state, i32::from(p.shift.1), false, &t); + enc.bit(p.extra_plane, &mut state[0], &t); + enc.symbol(&mut state, p.num_h as i32 - 1, false, &t); + enc.symbol(&mut state, p.num_v as i32 - 1, false, &t); + enc.symbol(&mut state, p.sets.len() as i32, false, &t); + for set in &p.sets { + for table in &set.tables { + let mut qstate = [128u8; CONTEXT_SIZE]; + for run in quant_runs(table) { + enc.symbol(&mut qstate, run, false, &t); + } + } + } + for _ in &p.sets { + enc.bit(false, &mut state[0], &t); // states_coded + } + enc.symbol(&mut state, i32::from(p.ec), false, &t); + enc.symbol(&mut state, i32::from(p.intra), false, &t); +} + +// ---- slices ---- + +/// A slice's place in the frame, in pixels. +#[derive(Clone, Copy)] +struct SliceRect { + x: usize, + y: usize, + w: usize, + h: usize, +} + +fn slice_rect(p: &Params, width: usize, height: usize, sx: usize, sy: usize, sw: usize, sh: usize) -> Result { + let (nh, nv) = (p.num_h as usize, p.num_v as usize); + if sx + sw > nh || sy + sh > nv || sw == 0 || sh == 0 { + return Err(Error::Invalid("slice outside the slice raster")); + } + let x0 = sx * width / nh; + let x1 = (sx + sw) * width / nh; + let y0 = sy * height / nv; + let y1 = (sy + sh) * height / nv; + if x1 <= x0 || y1 <= y0 { + return Err(Error::Invalid("empty slice")); + } + Ok(SliceRect { x: x0, y: y0, w: x1 - x0, h: y1 - y0 }) +} + +/// The samples of one plane of a slice with the border of RFC 9043 section 3.1 around them: two +/// columns to the left, one to the right, two rows above. +struct Border { + buf: Vec, + stride: usize, + w: usize, +} + +impl Border { + fn new(w: usize, h: usize) -> Border { + let stride = w + 3; + Border { buf: vec![0; stride * (h + 2)], stride, w } + } + + fn at(&self, x: usize, y: usize) -> usize { + (y + 2) * self.stride + x + 2 + } + + /// Set up the left border of row `y` (the column left of it repeats the row above's first sample). + fn start_row(&mut self, y: usize) { + let row = (y + 2) * self.stride; + self.buf[row] = 0; + self.buf[row + 1] = if y > 0 { self.buf[(y + 1) * self.stride + 2] } else { 0 }; + } + + /// The right border of row `y` repeats its last sample. + fn end_row(&mut self, y: usize) { + let row = (y + 2) * self.stride; + self.buf[row + self.w + 2] = self.buf[row + self.w + 1]; + } + + /// `(context, predictor)` for sample `(x, y)`; `context` is signed (section 3.5). + fn context(&self, x: usize, y: usize, tables: &[[i32; 256]], sixteen_bit_ycbcr: bool) -> (i32, i32) { + let i = self.at(x, y); + let (b, s) = (&self.buf, self.stride); + let (l, ll, t, tl, tr, tt) = (b[i - 1], b[i - 2], b[i - s], b[i - s - 1], b[i - s + 1], b[i - 2 * s]); + let ctx = tables[0][((l - tl) & 255) as usize] + + tables[1][((tl - t) & 255) as usize] + + tables[2][((t - tr) & 255) as usize] + + tables[3][((ll - l) & 255) as usize] + + tables[4][((tt - t) & 255) as usize]; + let (pl, pt, ptl) = if sixteen_bit_ycbcr { + let s16 = |v: i32| if v >= 32768 { v - 65536 } else { v }; + (s16(l), s16(t), s16(tl)) + } else { + (l, t, tl) + }; + (ctx, median(pl, pt, pl + pt - ptl)) + } +} + +fn median(a: i32, b: i32, c: i32) -> i32 { + a.max(b).min(a.min(b).max(c)) +} + +/// The reversible colour transform bits: a plane's samples have `bits + 1` bits. +fn rct_swapped(p: &Params) -> bool { + p.colorspace == Colorspace::Rgb && (9..=15).contains(&p.bits) && !p.extra_plane +} + +fn coded_bits(p: &Params) -> u32 { + if p.colorspace == Colorspace::Rgb { p.bits + 1 } else { p.bits } +} + +/// One plane's samples inside a slice, and which class of states codes them. +struct PlaneCoder { + border: Border, + slot: usize, +} + +/// The adaptive states of one plane class: Y, Cb and Cr together, then transparency (section 3.6). +struct ClassStates { + set: usize, + states: Vec<[u8; CONTEXT_SIZE]>, +} + +fn make_coders(p: &Params, rect: SliceRect, set_idx: &[usize]) -> Result<(Vec, Vec)> { + let mut planes = Vec::new(); + for pl in 0..p.plane_count() { + let (w, h) = plane_dims(rect.w, rect.h, p.colorspace, p.shift, pl, p.chroma_planes); + planes.push(PlaneCoder { border: Border::new(w, h), slot: p.qset_slot(pl) }); + } + let mut classes = Vec::new(); + for slot in 0..=planes.iter().map(|c| c.slot).max().unwrap_or(0) { + let set = *set_idx.get(slot).ok_or(Error::Invalid("missing quant table set index"))?; + let q = p.sets.get(set).ok_or(Error::Invalid("quant table set index out of range"))?; + let states = match p.initial_states.get(set) { + Some(Some(init)) => init.clone(), + _ => vec![[128u8; CONTEXT_SIZE]; q.context_count], + }; + classes.push(ClassStates { set, states }); + } + Ok((planes, classes)) +} + +/// The number of Quantization Table Set indexes in a slice header (section 4.6.5, version 3). +fn set_index_count(p: &Params) -> usize { + 2 + usize::from(p.extra_plane) +} + +fn check_frame_size(p: &Params, w: u32, h: u32) -> Result<()> { + if w == 0 || h == 0 { + return Err(Error::Invalid("empty frame")); + } + let shape = PictureShape { + width: u64::from(w), + height: u64::from(h), + shift_x: u32::from(p.shift.0.min(1)), + shift_y: u32::from(p.shift.1.min(1)), + chroma_planes: if p.chroma_planes { 2 } else { 0 }, + bytes_per_sample: 2, + }; + // The decoded planes and, while a slice decodes, its bordered `i32` copy (2 x the samples at most). + if !shape.fits(5, DECODE_MEMORY_BUDGET) { + return Err(Error::Unsupported("frame size above the memory budget")); + } + if p.shift.0 > 1 || p.shift.1 > 1 { + // Subsampling beyond 2:1 is legal FFV1 (e.g. 4:1:1) but the budget model above is for 2:1. + let extra = (w as u64).checked_mul(h as u64).ok_or(Error::Invalid("frame too large"))?; + if extra > 1 << 28 { + return Err(Error::Unsupported("frame size above the memory budget")); + } + } + Ok(()) +} + +/// A decoder for one stream. +pub struct Decoder { + params: Params, + width: u32, + height: u32, +} + +impl Decoder { + /// Build a decoder from the container's configuration record and the frame size. + pub fn new(config_record: &[u8], width: u32, height: u32) -> Result { + if config_record.len() < 5 { + return Err(Error::Truncated); + } + let (body, _crc) = config_record.split_at(config_record.len() - 4); + if crc32(config_record) != 0 { + return Err(Error::Crc("the configuration record")); + } + let mut rac = Rac::new(body); + let params = parse_params(&mut rac)?; + if rac.consumed() > body.len() + 2 { + return Err(Error::Truncated); + } + check_frame_size(¶ms, width, height)?; + Ok(Decoder { params, width, height }) + } + + /// Decode one frame (a packet from the container). + pub fn decode(&self, packet: &[u8]) -> Result