diff --git a/ATTRIBUTION.md b/ATTRIBUTION.md index ef3b98f31..43d543062 100644 --- a/ATTRIBUTION.md +++ b/ATTRIBUTION.md @@ -100,3 +100,5 @@ and the exact transfer revision are recorded in each vendored crate's SOURCE.jso ## Binary Apple property lists `crates/plist` is extracted from `storytold/craft-anything` (`crates/formats/src/plist.rs` and checked byte reads), at source main `88d1d9cac852aec1c7fd97886db451e005e49d2a`. MIT OR Apache-2.0; source license texts are retained in the crate. Primary import Git authors preserve Brandon Thomas, BFlat, RRFZ and Afloat. Regression imports preserve BFlat. XML, JSON projection, calendar formatting and app policy were not imported into the public API. + +The XML reader in `crates/plist/src/xml.rs` and its base64 compatibility helper are ported from Craft Anything `8f3b775157d331dd31cfa4da911e26d8856b8c77` (same MIT OR Apache-2.0 source licenses). Brandon Thomas is the primary import author; retained hardening credits include Afloat, BFlat, Rahul, RRFZ and NetterSilver. The finite-real fix preserves gmrmad1 as author from Craft Anything #1905 (`d85e104f`); quoted-attribute/line-ending regressions preserve RRFZ. Bounds/adaptation and bounded entity scanning are subsequent Dylan M. Taylor changes. App-specific JSON projections and keyed-archive interpretation remain outside this crate. diff --git a/Cargo.lock b/Cargo.lock index b3325f790..7e188a05f 100644 --- a/Cargo.lock +++ b/Cargo.lock @@ -222,16 +222,6 @@ version = "1.0.104" source = "registry+https://github.com/rust-lang/crates.io-index" checksum = "330a5ed07fa54e4702c9d6c4174f74427fc0ef6e214bbd677ae50a5099946470" -[[package]] -name = "aotuv_lancer_vorbis_sys" -version = "0.1.6" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "5bc4fd1a61860d2f1198b60bedd30910eaffa978f1ee6214dfb24ac70d589225" -dependencies = [ - "cc", - "ogg_next_sys", -] - [[package]] name = "arbitrary" version = "1.5.0" @@ -340,6 +330,12 @@ dependencies = [ "syn 2.0.119", ] +[[package]] +name = "astc-decode" +version = "0.3.1" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "2d6058e6202e20b7fa445692ec4b77ca8eb02508091fcd479ea753d9f60e376b" + [[package]] name = "async-broadcast" version = "0.7.2" @@ -1414,7 +1410,6 @@ dependencies = [ "serde_json", "symphonia", "thiserror 2.0.21", - "vorbis_rs", ] [[package]] @@ -1491,6 +1486,17 @@ dependencies = [ "proptest", ] +[[package]] +name = "craft-deckcraft-format" +version = "0.1.0" +dependencies = [ + "craft-zip", + "serde", + "serde_json", + "thiserror 2.0.21", + "zip", +] + [[package]] name = "craft-designcraft-format" version = "0.1.0" @@ -1929,6 +1935,7 @@ dependencies = [ "craft-jpeg", "craft-media-testkit", "craft-metadata", + "craft-tiff", "craft-video", "crc32fast", "exr", @@ -2758,6 +2765,8 @@ dependencies = [ name = "craft-texture" version = "0.1.0" dependencies = [ + "bytemuck", + "dds", "rayon", "thiserror 2.0.21", ] @@ -2855,6 +2864,14 @@ dependencies = [ "zip", ] +[[package]] +name = "craft-usd" +version = "0.1.0" +dependencies = [ + "lz4_flex", + "thiserror 2.0.21", +] + [[package]] name = "craft-vcard" version = "0.1.0" @@ -3295,6 +3312,19 @@ dependencies = [ "generic-array", ] +[[package]] +name = "dds" +version = "0.2.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "e917a58e17b27d2f15626c490754ccced7f46592dc5b5f39a3013b89cba5f068" +dependencies = [ + "astc-decode", + "bitflags 2.13.2", + "glam", + "resize", + "zerocopy", +] + [[package]] name = "der" version = "0.7.10" @@ -4543,6 +4573,12 @@ dependencies = [ "xml-rs", ] +[[package]] +name = "glam" +version = "0.25.0" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "151665d9be52f9bb40fc7966565d39666f2d1e69233571b71b87791c7e0528b3" + [[package]] name = "glifo" version = "0.1.1" @@ -6054,6 +6090,12 @@ version = "0.1.3" source = "registry+https://github.com/rust-lang/crates.io-index" checksum = "4050469837a6ff301cd14c1f8f24f88549e6d548f24f64e2148eb0f72cebc51f" +[[package]] +name = "lz4_flex" +version = "0.11.6" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "373f5eceeeab7925e0c1098212f2fbc4d416adec9d35051a6ab251e824c1854a" + [[package]] name = "maybe-owned" version = "0.3.4" @@ -6812,15 +6854,6 @@ dependencies = [ "phf 0.11.3", ] -[[package]] -name = "ogg_next_sys" -version = "0.1.5" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "ed2d7a48e247c2bb07e633aefb65a38648ea58c7eedd4e4408a5861721ab049b" -dependencies = [ - "cc", -] - [[package]] name = "oid-registry" version = "0.8.1" @@ -8136,6 +8169,15 @@ dependencies = [ "web-sys", ] +[[package]] +name = "resize" +version = "0.8.9" +source = "registry+https://github.com/rust-lang/crates.io-index" +checksum = "71725ecd5e0197b54fe859055b108688472ab6a358f8fbe5cee4a556b1b5bfea" +dependencies = [ + "rgb", +] + [[package]] name = "resvg" version = "0.45.1" @@ -10639,20 +10681,6 @@ version = "0.9.5" source = "registry+https://github.com/rust-lang/crates.io-index" checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a" -[[package]] -name = "vorbis_rs" -version = "0.5.6" -source = "registry+https://github.com/rust-lang/crates.io-index" -checksum = "49c5da94d280f7a27e8c937e9b73df2da3e23a2583f48471fd8fb4c72f9c1933" -dependencies = [ - "aotuv_lancer_vorbis_sys", - "errno", - "getrandom 0.4.3", - "ogg_next_sys", - "thiserror 2.0.21", - "tinyvec", -] - [[package]] name = "wait-timeout" version = "0.2.1" diff --git a/Cargo.toml b/Cargo.toml index 0c9729e00..48a2edf1d 100644 --- a/Cargo.toml +++ b/Cargo.toml @@ -11,6 +11,7 @@ rust-version = "1.95" repository = "https://github.com/storytold/craft-libs" [workspace.dependencies] +craft-usd = { path = "crates/usd" } craft-smime = { path = "crates/smime" } craft-ber = { path = "crates/ber" } craft-proto = { path = "crates/proto" } @@ -54,6 +55,7 @@ craft-doc-binary = { path = "crates/doc-binary" } craft-docx = { path = "crates/docx" } craft-document-bundle = { path = "crates/document-bundle" } craft-mcmat = { path = "crates/mcmat" } +craft-deckcraft-format = { path = "crates/deckcraft-format" } craft-designcraft-format = { path = "crates/designcraft-format" } craft-dxf = { path = "crates/dxf" } craft-drawingml = { path = "crates/drawingml" } diff --git a/README.md b/README.md index c6e95bd4f..0b4b9a415 100644 --- a/README.md +++ b/README.md @@ -120,6 +120,10 @@ See the [release notes](CHANGELOG.md) for included functionality and known limit pages and version detection, and written back; and `model`, the drawing above the tags (header variables, tables, blocks, layouts, entities) and its writer (R12 to 2018), merged from CADCraft's and VectorCraft's readers. Bounded on hostile input. Mapping the model onto a document stays in each app. +- [`craft-deckcraft-format`](crates/deckcraft-format/README.md): the DeckCraft `.deckcraft` package: + identity and sniffing, the `presentation.json` envelope with the presentation as a type + parameter, NaN-safe JSON and the reader for older files with `null` numbers, member names and + compression, and bounded, budgeted member reads. The presentation model and its mapping stay in DeckCraft. - [`craft-dwg`](crates/dwg/README.md): bounded byte-level DWG↔DXF conversion with diagnostics, capped output and an optional child-process worker; the `acadrust` dependency remains isolated behind bytes and is unavailable on wasm32 and FreeBSD/aarch64. diff --git a/crates/aaf/src/omf.rs b/crates/aaf/src/omf.rs index 57e0f756c..bbd2e8571 100644 --- a/crates/aaf/src/omf.rs +++ b/crates/aaf/src/omf.rs @@ -21,7 +21,7 @@ //! transition's name) are not in the specification; they were first written by FilmCraft and keep //! their names so that existing files read back. -pub(crate) mod bento; +pub mod bento; mod read; mod write; diff --git a/crates/aaf/src/omf/bento.rs b/crates/aaf/src/omf/bento.rs index 00a7e2a26..171196fe1 100644 --- a/crates/aaf/src/omf/bento.rs +++ b/crates/aaf/src/omf/bento.rs @@ -35,8 +35,8 @@ use std::collections::{BTreeMap, HashMap}; -pub(crate) const MAGIC: [u8; 8] = [0xA4, b'C', b'M', 0xA5, b'H', b'd', b'r', 0xD7]; -pub(crate) const LABEL_SIZE: usize = 24; +pub const MAGIC: [u8; 8] = [0xA4, b'C', b'M', 0xA5, b'H', b'd', b'r', 0xD7]; +pub const LABEL_SIZE: usize = 24; // standard object ids const TOC_OBJECT: u32 = 1; @@ -82,7 +82,7 @@ const MAX_VALUE_BYTES: usize = 16 << 20; const MAX_NAME_BYTES: usize = 4 << 20; /// Builds a Bento container. -pub(crate) struct Writer { +pub struct Writer { data: Vec, names: HashMap<(u32, String), u32>, next: u32, @@ -90,6 +90,12 @@ pub(crate) struct Writer { entries: Vec<(u32, u32, u32, Vec)>, } +impl Default for Writer { + fn default() -> Self { + Self::new() + } +} + impl Writer { pub fn new() -> Writer { Writer { data: Vec::new(), names: HashMap::new(), next: MIN_USER_ID, entries: Vec::new() } @@ -217,9 +223,9 @@ pub(crate) struct Prop { } /// A parsed Bento container. -pub(crate) struct Container<'a> { +pub struct Container<'a> { data: &'a [u8], - pub objects: BTreeMap>, + pub(crate) objects: BTreeMap>, names: HashMap, by_name: HashMap, } @@ -229,11 +235,16 @@ fn be32(b: &[u8], at: usize) -> Option { } /// Whether `b` ends with a Bento label. -pub(crate) fn sniff(b: &[u8]) -> bool { +pub fn sniff(b: &[u8]) -> bool { b.len() >= LABEL_SIZE && b[b.len() - LABEL_SIZE..b.len() - LABEL_SIZE + 8] == MAGIC } impl<'a> Container<'a> { + /// Object identifiers in ascending order. + pub fn object_ids(&self) -> impl Iterator + '_ { + self.objects.keys().copied() + } + pub fn open(data: &'a [u8]) -> Result, String> { if !sniff(data) { return Err("no Bento label at the end of the file".into()); @@ -324,7 +335,7 @@ impl<'a> Container<'a> { } /// The bytes of a value. - pub fn value(&self, p: &Prop) -> Result, String> { + pub(crate) fn value(&self, p: &Prop) -> Result, String> { let total = p.segs.iter().try_fold(0usize, |total, s| { let len = match s { Seg::Imm(b) => b.len(), diff --git a/crates/aaf/tests/bento_api.rs b/crates/aaf/tests/bento_api.rs new file mode 100644 index 000000000..93adc33c1 --- /dev/null +++ b/crates/aaf/tests/bento_api.rs @@ -0,0 +1,27 @@ +use craft_aaf::omf::bento::{Container, Writer, sniff}; + +#[test] +fn external_consumer_reads_named_values_and_references() { + let mut writer = Writer::new(); + let first = writer.new_object(); + let second = writer.new_object(); + writer.set(first, "test:value", "test:type", b"long value"); + writer.set_references(first, &[second]); + let bytes = writer.finish(); + assert!(sniff(&bytes)); + let container = Container::open(&bytes).unwrap(); + assert!(container.object_ids().any(|id| id == first)); + assert_eq!(container.get(first, "test:value").as_deref(), Some(b"long value".as_slice())); + assert_eq!(container.resolve(first, second), second); + assert!(Container::open(&bytes[..bytes.len() - 1]).is_err()); +} + +#[test] +fn default_writer_builds_a_readable_empty_container() -> Result<(), String> { + let bytes = Writer::default().finish(); + let container = Container::open(&bytes)?; + let ids: Vec<_> = container.object_ids().collect(); + assert!(!ids.is_empty()); + assert!(ids.windows(2).all(|pair| pair[0] < pair[1])); + Ok(()) +} diff --git a/crates/audio/Cargo.toml b/crates/audio/Cargo.toml index 645246130..859ac20bc 100644 --- a/crates/audio/Cargo.toml +++ b/crates/audio/Cargo.toml @@ -1,6 +1,6 @@ [package] name = "craft-audio" -description = "Audio file I/O and codecs: WAV/AIFF read and write, FLAC encoder, AAC and ALAC decoders and encoders, AC-3/E-AC-3 decoder and encoder, MP3 decoder and encoder, Opus decoder and encoder, MP1/MP2 decoders." +description = "Audio file I/O and codecs: WAV/AIFF read and write, FLAC encoder, AAC and ALAC decoders and encoders, AC-3/E-AC-3 decoder and encoder, MP3 decoder and encoder, Opus decoder and encoder, Vorbis encoder, MP1/MP2 decoders." version.workspace = true edition.workspace = true license.workspace = true @@ -31,8 +31,9 @@ lossless-encode = ["pcm-encode", "flac-encode"] pcm-encode = [] # Encode FLAC files. flac-encode = ["craft-media/bitstream"] -# Encode Ogg Vorbis files (desktop only: builds the reference libvorbis from its C sources). -vorbis-encode = ["dep:vorbis_rs"] +# The pure-Rust Vorbis encoder (`vorbis::encoder`) and Ogg Vorbis files through `encode` and +# `encode_ogg_vorbis`. +vorbis-encode = ["craft-media/ogg"] # Codec modules on their own, for packet-level use. aac-decode = ["craft-media/bitstream"] # The AAC encoder (`aac::Encoder`), and AAC in M4A through `encode` and `FileEncoder`. @@ -72,9 +73,6 @@ rayon = { workspace = true, optional = true } # symphonia keeps its AAC decoder for AAC in CAF, a container craft-media does not read. symphonia = { version = "0.6.1", optional = true, default-features = false, features = ["mp3", "flac", "vorbis", "ogg", "wav", "pcm", "aiff", "isomp4", "aac", "alac", "adpcm", "caf", "mkv"] } -[target.'cfg(not(target_arch = "wasm32"))'.dependencies] -vorbis_rs = { version = "0.5", optional = true } - [dev-dependencies] craft-media-testkit = { workspace = true } proptest = { workspace = true } diff --git a/crates/audio/README.md b/crates/audio/README.md index 63251db8d..a5af4b3f3 100644 --- a/crates/audio/README.md +++ b/crates/audio/README.md @@ -29,7 +29,7 @@ No default features. The shared types (`AudioBuffer`, `AudioInfo`, `ChannelLayou | `decode` | `decode`, `probe`, `Decoder`, `tracks`, `TrackDecoder`; every format in [Format support](#format-support) | `decode-native`, `symphonia` | | `decode-native` | the same, native formats only (others are `Unsupported`); `wav_pcm`, `aiff_pcm` (a file's integer PCM bit for bit, as `RawPcm`) | the six decoders below, craft-media `isobmff`/`matroska`/`ogg`, `log` | | `symphonia` | the fallback for formats with no native decoder; turning it on is the same as `decode` | `decode-native`, symphonia | -| `encode` | every encoder, and every format of `encode` and `FileEncoder` | `lossless-encode`, `aac-encode`, `alac-encode`, `ac3-encode`, `eac3-encode`, `mp3-encode`, `opus-encode` | +| `encode` | every encoder, and every format of `encode` and `FileEncoder` | `lossless-encode`, `aac-encode`, `alac-encode`, `ac3-encode`, `eac3-encode`, `mp3-encode`, `opus-encode`, `vorbis-encode` | | `lossless-encode` | WAV, AIFF and FLAC writing | `pcm-encode`, `flac-encode` | | `pcm-encode` | `encode`, `FileEncoder` for WAV and AIFF; `wav_float_header`; `pcm_wav_file`, `pcm_wav_header`, `pcm_aiff_file`, `pcm_aiff_header` (minimal files around integer PCM bytes, for essence embedded in AAF / OMF) | none | | `flac-encode` | `encode`, `FileEncoder` for FLAC; `flac::Encoder`, the packet encoder for callers that mux FLAC themselves | craft-media `bitstream` | @@ -37,6 +37,7 @@ No default features. The shared types (`AudioBuffer`, `AudioInfo`, `ChannelLayou | `alac-decode` | `alac::Decoder` | none | | `alac-encode` | `alac::Encoder`; `encode`, `FileEncoder` for ALAC in M4A | craft-media `isobmff` | | `opus-decode`, `opus-encode` | `opus::Decoder`/`StreamDecoder`, `opus::Encoder` | none | +| `vorbis-encode` | `vorbis::encoder::Encoder` (pure Rust); `encode_ogg_vorbis`, `encode_ogg` and `encode` for Ogg Vorbis files (whole buffer; not `FileEncoder`) | craft-media `ogg` | | `ac3-decode` | `ac3::Decoder`, `ac3::ProgrammeDecoder` | craft-media `bitstream` | | `ac3-encode`, `eac3-encode` | `ac3::encoder::Encoder`; `encode`, `FileEncoder` for AC-3 and E-AC-3 streams | craft-media `bitstream` | | `mpa-decode` | `mpa::Decoder` | craft-media `bitstream` | @@ -355,8 +356,12 @@ let pre_skip = enc.pre_skip(); // samples the player drops at the start The encoder takes 48 kHz only, and craft-audio has no resampler. -craft-media has no Ogg muxer, so `.opus` files cannot be written yet; Opus goes into -Matroska/WebM or MP4: +Opus goes into Ogg, Matroska/WebM or MP4: + +- **Ogg (`.opus`):** `craft_media::ogg::PageWriter` with `head` and `opus_tags` as the two header + packets, each packet's granule position the samples decoded so far (pre-skip included, as + the input above starts with it), and `finish(Some(pre_skip + length))` to trim the end + (RFC 7845 §4). - **Matroska/WebM:** `head` is the track's `CodecPrivate`, and the pre-skip goes in `TrackSpec::codec_delay_ns` (in nanoseconds at 48 kHz). The Matroska muxer cannot write @@ -368,6 +373,27 @@ Matroska/WebM or MP4: timescale), `segment_duration` the audio's length (movie timescale), `media_rate` `0x0001_0000`. +### Encode Ogg Vorbis + +```rust +use craft_audio::{OggVorbisOptions, encode_ogg_vorbis}; + +// 1–8 channels in Vorbis order (5.1: L C R Ls Rs LFE), any sample rate. +let opts = OggVorbisOptions::new(6.0).with_tag("TITLE", "Take 3").with_tag("ARTIST", "Me"); +let ogg: Vec = encode_ogg_vorbis(&buf, &opts)?; +``` + +The encoder is pure Rust (`vorbis::encoder`, from FilmCraft #4481) and builds for wasm32. Quality +runs from -1 to 10 at a variable bit rate: about 18 kbit/s per channel at -1, 100 at 6 and 140 +at 10 on music. `encode` with `EncodeFormat::Ogg` takes a WAV-order layout and writes it in +Vorbis order; its `RateControl::Bitrate` (64–500 kbit/s) picks the quality whose nominal rate +that is (`quality_for_bitrate`), and `RateControl::Quality` 0.0–1.0 is quality -1 to 10. For +streaming, feed `vorbis::encoder::Encoder` and write its packets with `craft_media::ogg::PageWriter` +(identification, `vorbis_comment`, setup, then the audio with each packet's granule position). + +Not yet: short blocks (transients may pre-echo), channel coupling (stereo costs +about twice mono) and bit-rate management. + ### Waveform overviews ```rust @@ -704,6 +730,7 @@ the full tests with ffmpeg and the vectors required. | MP3 encoder | FilmCraft `filmcraft-mp3enc` [4bec22b](https://github.com/storytold/filmcraft/commit/4bec22bd03bd1d995c9cf72e02a284f96d3d72cd) | | ALAC | Apple's reference implementation [macosforge/alac@c38887c](https://github.com/macosforge/alac/tree/c38887c5c5e64a4b31108733bd79ca9b2496d987/codec), Copyright © 2011 Apple Inc., Apache-2.0; ported to Rust and modified (see the `alac` module docs and [NOTICE](../../NOTICE)) | | Opus encoder | EffectCraft `effectcraft-opusenc` [c618429](https://github.com/storytold/effectcraft/commit/c618429252587525a2a6d4b5bccbe4ec5261e1b9) | +| Vorbis encoder | FilmCraft `filmcraft-vorbis` by @JordyUSA ([storytold/filmcraft#4481](https://github.com/storytold/filmcraft/pull/4481)), with partitions of whole sample frames so libvorbis reads surround | Where apps had overlapping implementations, the one that measured better was kept, with the tests of both: diff --git a/crates/audio/src/aac/config.rs b/crates/audio/src/aac/config.rs index 88be1e214..3d12b0343 100644 --- a/crates/audio/src/aac/config.rs +++ b/crates/audio/src/aac/config.rs @@ -343,6 +343,51 @@ impl AudioSpecificConfig { } } +/// Why an AudioSpecificConfig prefix cannot be written as an ADTS header. +#[derive(Debug, Clone, Copy, PartialEq, Eq)] +pub enum AdtsConfigError { + /// Fewer than two bytes were supplied. + Short, + /// The prefix names a profile, indexed rate or channel configuration ADTS cannot carry. + Unsupported { + /// Audio object type from the prefix. + object_type: u8, + /// Sampling frequency index from the prefix. + sf_index: u8, + /// Channel configuration from the prefix. + channel_config: u8, + }, +} + +/// ADTS fixed fields taken from the prefix of an AudioSpecificConfig. +#[derive(Debug, Clone, Copy, PartialEq, Eq)] +pub struct AdtsConfig { + /// Audio object type, 1 through 4. + pub object_type: u8, + /// Indexed sample rate, 0 through 12. + pub sf_index: u8, + /// Channel configuration, 1 through 7. + pub channel_config: u8, +} +impl AdtsConfig { + /// Parse only the first two ASC bytes. Unlike a decoder's full ASC parser, this accepts all + /// four object types representable by ADTS and rejects configurations needing a PCE. + pub fn from_asc_prefix(asc: &[u8]) -> std::result::Result { + let (Some(&a), Some(&b)) = (asc.first(), asc.get(1)) else { return Err(AdtsConfigError::Short) }; + let object_type = a >> 3; + let sf_index = ((a & 7) << 1) | (b >> 7); + let channel_config = (b >> 3) & 15; + if !(1..=4).contains(&object_type) || sf_index > 12 || !(1..=7).contains(&channel_config) { + return Err(AdtsConfigError::Unsupported { object_type, sf_index, channel_config }); + } + Ok(Self { object_type, sf_index, channel_config }) + } + /// Write a no-CRC ADTS header with variable-rate buffer fullness for one payload. + pub fn header(self, payload_len: usize) -> Result<[u8; 7]> { + AdtsHeader::write(self.object_type, self.sf_index, self.channel_config, payload_len, 0x7ff) + } +} + /// A parsed ADTS header. #[non_exhaustive] #[derive(Debug, Clone, Copy, PartialEq, Eq)] @@ -519,6 +564,15 @@ mod tests { assert_eq!(AudioSpecificConfig::parse(&asc.to_bytes().unwrap()).unwrap(), asc); } + #[test] + fn adts_config_preserves_prefix_and_frame_limits() { + let config = AdtsConfig::from_asc_prefix(&[0x12, 0x10]).unwrap(); + assert_eq!(config.header(100).unwrap(), AdtsHeader::write(2, 4, 2, 100, 0x7ff).unwrap()); + assert!(config.header(8185).is_err()); + assert_eq!(AdtsConfig::from_asc_prefix(&[]), Err(AdtsConfigError::Short)); + assert!(matches!(AdtsConfig::from_asc_prefix(&[0x2a, 0x10]), Err(AdtsConfigError::Unsupported { object_type: 5, .. }))); + } + #[test] fn adts_headers_refuse_fields_they_cannot_hold() { for (object_type, sf_index, channel_config) in [(0, 4, 2), (5, 4, 2), (29, 4, 2), (2, 13, 2), (2, 15, 2), (2, 4, 8), (2, 4, 11), (2, 4, 12)] { diff --git a/crates/audio/src/aac/mod.rs b/crates/audio/src/aac/mod.rs index f423a4e22..fb79ae743 100644 --- a/crates/audio/src/aac/mod.rs +++ b/crates/audio/src/aac/mod.rs @@ -60,7 +60,7 @@ mod tns; #[cfg(feature = "aac-decode")] pub(crate) use config::split_adts_upto; -pub use config::{AdtsHeader, AudioSpecificConfig, ElementType, ProgramConfig, config_layout, split_adts}; +pub use config::{AdtsConfig, AdtsConfigError, AdtsHeader, AudioSpecificConfig, ElementType, ProgramConfig, config_layout, split_adts}; #[cfg(feature = "aac-decode")] pub use decoder::Decoder; #[cfg(feature = "aac-encode")] diff --git a/crates/audio/src/lib.rs b/crates/audio/src/lib.rs index e0660d9c9..e291b1385 100644 --- a/crates/audio/src/lib.rs +++ b/crates/audio/src/lib.rs @@ -13,6 +13,8 @@ //! demuxers and [`TrackDecoder`]; everything else (FLAC, Vorbis, CAF, ADPCM, A-law, µ-law) through //! `symphonia`, which is to be replaced by native decoders. Feature `decode-native` is the //! same without symphonia. +//! - [`encode_ogg_vorbis`] and [`vorbis::encoder`] (feature `vorbis-encode`): Ogg Vorbis files +//! and Vorbis packets from a pure-Rust encoder, 1 to 8 channels, quality -1 to 10, with tags. //! - [`TrackDecoder`] (any decoder feature): one track's packets from any demuxer to planar //! `f32` in WAV channel order. //! - [`encode`], [`FileEncoder`]: WAV/BWF/RF64 and AIFF/AIFF-C (feature `pcm-encode`, which also @@ -56,8 +58,12 @@ pub mod qoa; mod read; #[cfg(any(feature = "aac-decode", feature = "alac-decode", feature = "opus-decode", feature = "ac3-decode", feature = "mpa-decode"))] mod track; +#[cfg(feature = "vorbis-encode")] +pub mod vorbis; #[cfg(any(feature = "decode-native", feature = "pcm-encode"))] mod wav; +/// Bounded WAV/RF64 cue, adtl and XMP marker metadata, without audio decoding. +pub mod wav_markers; #[cfg(any( feature = "pcm-encode", feature = "flac-encode", @@ -70,10 +76,12 @@ mod wav; ))] mod write; +/// Interleaved PCM packet decoding. +pub mod pcm_interleaved; /// 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}; +pub use write::{OGG_BITRATES, OggVorbisOptions, encode_ogg, encode_ogg_vorbis, quality_for_bitrate}; /// The FLAC packet encoder (RFC 9639), for a caller that muxes FLAC itself (into MP4, Matroska or /// its own `.flac` writer): planar integer samples in, frames out, and the STREAMINFO to patch in @@ -276,7 +284,9 @@ pub enum EncodeFormat { Ac3, /// An E-AC-3 stream: up to 7.1 at 32, 44.1 or 48 kHz. Eac3, - /// Ogg Vorbis, whole-buffer encoding only. + /// Ogg Vorbis, 1–8 channels at any sample rate, whole-buffer encoding only. Variable bit + /// rate: a [`RateControl::Bitrate`] picks the quality whose nominal rate it is + /// ([`quality_for_bitrate`]), [`RateControl::Quality`] sets it directly. Ogg, } @@ -377,7 +387,7 @@ impl EncodeFormat { EncodeFormat::Mp3 => cfg!(feature = "mp3-encode"), EncodeFormat::Ac3 => cfg!(feature = "ac3-encode"), EncodeFormat::Eac3 => cfg!(feature = "eac3-encode"), - EncodeFormat::Ogg => cfg!(all(feature = "vorbis-encode", not(target_arch = "wasm32"))), + EncodeFormat::Ogg => cfg!(feature = "vorbis-encode"), } } @@ -398,9 +408,9 @@ impl EncodeFormat { pub fn max_channels(self) -> u16 { match self { EncodeFormat::Wav | EncodeFormat::Aiff => 32_767, - EncodeFormat::Flac | EncodeFormat::Aac | EncodeFormat::Alac | EncodeFormat::Eac3 => 8, + EncodeFormat::Flac | EncodeFormat::Aac | EncodeFormat::Alac | EncodeFormat::Eac3 | EncodeFormat::Ogg => 8, EncodeFormat::Ac3 => 6, - EncodeFormat::Mp3 | EncodeFormat::Ogg => 2, + EncodeFormat::Mp3 => 2, } } @@ -414,7 +424,10 @@ impl EncodeFormat { EncodeFormat::Aac => &[&[0x4], &[0x3], &[0x7], &[0x107], &[0x607], &[0x60F], &[0x70F], &[0xFF]], // Apple's layout tags for 1–8 channels, surrounds as back speakers. EncodeFormat::Alac => &[&[0x4], &[0x3], &[0x7], &[0x107], &[0x37], &[0x3F], &[0x13F], &[0xFF]], - EncodeFormat::Mp3 | EncodeFormat::Ogg => &[&[0x4], &[0x3]], + EncodeFormat::Mp3 => &[&[0x4], &[0x3]], + // The Vorbis I §4.3.9 layouts for 1–8 channels (6.1 with a back centre and sides, 7.1 + // with sides and backs), written in Vorbis order. + EncodeFormat::Ogg => &[&[0x4], &[0x3], &[0x7], &[0x33], &[0x37], &[0x3F], &[0x70F], &[0x63F]], // The coding modes 1/0 to 3/2 with and without an LFE, by channel count. EncodeFormat::Ac3 => &[&[0x4], &[0x3, 0xC], &[0x7, 0x103, 0xB], &[0x107, 0x603, 0xF, 0x10B], &[0x607, 0x10F, 0x60B], &[0x60F]], // The same, and 7.1 with backs and sides. @@ -456,10 +469,10 @@ impl EncodeFormat { } } - /// Whether the format takes [`RateControl::Quality`]: AAC. The other lossy formats are - /// written at a constant bit rate only. + /// Whether the format takes [`RateControl::Quality`]: AAC and Ogg Vorbis. The other lossy + /// formats are written at a constant bit rate only. pub fn supports_quality(self) -> bool { - self == EncodeFormat::Aac + matches!(self, EncodeFormat::Aac | EncodeFormat::Ogg) } /// Whether the format loses audio: AAC, MP3, AC-3 and E-AC-3. These take @@ -478,11 +491,12 @@ pub enum RateControl { /// set of rates (MP3, AC-3) uses the highest one not above it; a rate outside what the /// stream can carry is moved to the nearest one inside. AAC is written at 6 kbit/s or a /// quarter of the sample rate per channel at least: from about half of that down its encoder - /// codes silence. Vorbis treats this as a VBR target (64–500 kbit/s), falling back to - /// a quality preset when that target is unsupported; it does not promise a constant bitrate. + /// codes silence. Vorbis takes 64–500 kbit/s and writes the quality whose nominal rate that is + /// ([`quality_for_bitrate`]); its rate varies with the audio. Bitrate(u32), /// A constant quality from 0.0 (smallest) to 1.0 (best), at whatever bit rate each frame - /// needs. AAC only; values outside the range are clamped, and NaN is an error. + /// needs. AAC and Vorbis (0.0 is Vorbis quality -1, 1.0 quality 10); values outside the range + /// are clamped, and NaN is an error. Quality(f32), } diff --git a/crates/audio/src/opus/decode/celt/mdct.rs b/crates/audio/src/opus/decode/celt/mdct.rs index b3bfcf0f7..7503d9ebe 100644 --- a/crates/audio/src/opus/decode/celt/mdct.rs +++ b/crates/audio/src/opus/decode/celt/mdct.rs @@ -4,141 +4,7 @@ //! producing the N/2 "folded" time samples; the window overlap region is then unfolded against the //! folded tail of the previous block, which the caller keeps in place in the output buffer. -#[derive(Clone, Copy, Default, Debug)] -pub struct Cpx { - pub re: f32, - pub im: f32, -} - -impl Cpx { - #[inline] - fn mul(self, o: Cpx) -> Cpx { - Cpx { re: self.re * o.re - self.im * o.im, im: self.re * o.im + self.im * o.re } - } - #[inline] - fn add(self, o: Cpx) -> Cpx { - Cpx { re: self.re + o.re, im: self.im + o.im } - } - #[inline] - fn sub(self, o: Cpx) -> Cpx { - Cpx { re: self.re - o.re, im: self.im - o.im } - } -} - -/// Mixed-radix (2, 3, 4, 5) forward complex FFT, unscaled. Radix 2, 3 and 4 use dedicated -/// butterflies; radix 5 uses a generic one. -pub struct Fft { - n: usize, - factors: Vec, - twiddle: Vec, -} - -impl Fft { - pub fn new(n: usize) -> Fft { - let mut factors = Vec::new(); - let mut m = n; - for p in [4usize, 2, 3, 5] { - while m.is_multiple_of(p) { - factors.push(p); - m /= p; - } - } - assert_eq!(m, 1, "unsupported FFT size {n}"); - let twiddle = (0..n) - .map(|k| { - let ph = -2.0 * std::f64::consts::PI * k as f64 / n as f64; - Cpx { re: ph.cos() as f32, im: ph.sin() as f32 } - }) - .collect(); - Fft { n, factors, twiddle } - } - - pub fn process(&self, data: &mut [Cpx], scratch: &mut Vec) { - scratch.clear(); - scratch.extend_from_slice(&data[..self.n]); - self.rec(&mut data[..self.n], scratch, 1, self.n, 0, 1); - } - - fn rec(&self, out: &mut [Cpx], inp: &[Cpx], in_stride: usize, n: usize, fi: usize, tw_stride: usize) { - let p = self.factors[fi]; - let m = n / p; - if m == 1 { - for (i, o) in out.iter_mut().enumerate().take(p) { - *o = inp[i * in_stride]; - } - } else { - for q in 0..p { - self.rec(&mut out[q * m..(q + 1) * m], &inp[q * in_stride..], in_stride * p, m, fi + 1, tw_stride * p); - } - } - let tw = &self.twiddle; - match p { - 2 => { - for k in 0..m { - let a0 = out[k]; - let a1 = out[m + k].mul(tw[k * tw_stride]); - out[k] = a0.add(a1); - out[m + k] = a0.sub(a1); - } - } - 4 => { - for k in 0..m { - let t0 = out[k]; - let t1 = out[m + k].mul(tw[k * tw_stride]); - let t2 = out[2 * m + k].mul(tw[2 * k * tw_stride]); - let t3 = out[3 * m + k].mul(tw[3 * k * tw_stride]); - let s0 = t0.add(t2); - let d0 = t0.sub(t2); - let s1 = t1.add(t3); - let d1 = t1.sub(t3); - out[k] = s0.add(s1); - out[2 * m + k] = s0.sub(s1); - // -j * d1 = (d1.im, -d1.re) - out[m + k] = Cpx { re: d0.re + d1.im, im: d0.im - d1.re }; - out[3 * m + k] = Cpx { re: d0.re - d1.im, im: d0.im + d1.re }; - } - } - 3 => { - // sin(-2*pi/3) = -sqrt(3)/2 - let h = tw[self.n / 3].im; - for k in 0..m { - let t0 = out[k]; - let t1 = out[m + k].mul(tw[k * tw_stride]); - let t2 = out[2 * m + k].mul(tw[2 * k * tw_stride]); - let s = t1.add(t2); - let d = t1.sub(t2); - let a = Cpx { re: t0.re - 0.5 * s.re, im: t0.im - 0.5 * s.im }; - out[k] = t0.add(s); - // j * h * d = (-h * d.im, h * d.re) - out[m + k] = Cpx { re: a.re - h * d.im, im: a.im + h * d.re }; - out[2 * m + k] = Cpx { re: a.re + h * d.im, im: a.im - h * d.re }; - } - } - _ => { - // Radix 5 (at most once per transform): generic butterfly with the five roots of - // unity looked up once. - let mut w5 = [Cpx::default(); 5]; - for (q, w) in w5.iter_mut().enumerate() { - *w = tw[q * (self.n / p)]; - } - let mut t = [Cpx::default(); 5]; - for k in 0..m { - t[0] = out[k]; - for q in 1..p { - t[q] = out[q * m + k].mul(tw[q * k * tw_stride]); - } - for q2 in 0..p { - let mut acc = t[0]; - for q in 1..p { - acc = acc.add(t[q].mul(w5[(q * q2) % p])); - } - out[q2 * m + k] = acc; - } - } - } - } - } -} +use crate::opus::fft::{Cpx, Fft}; pub struct Imdct { /// Full MDCT length (2x the number of coefficients). diff --git a/crates/audio/src/opus/decode/celt/rate.rs b/crates/audio/src/opus/decode/celt/rate.rs index 782d68d5c..29e8566c1 100644 --- a/crates/audio/src/opus/decode/celt/rate.rs +++ b/crates/audio/src/opus/decode/celt/rate.rs @@ -1,272 +1,45 @@ //! Pulse cache and bit allocation (RFC 6716 §4.3.3, §4.3.4.1). -use super::tables::{ALLOC, EBANDS, LOG2_FRAC_TABLE, MAX_LM, NB_EBANDS, OVERLAP}; +use super::tables::{EBANDS, LOG2_FRAC_TABLE, MAX_LM, NB_EBANDS, OVERLAP}; use crate::opus::decode::range::{BITRES, RangeDecoder}; -const MAX_PSEUDO: i32 = 40; -const LOG_MAX_PSEUDO: i32 = 6; +pub use crate::opus::rate_common::get_pulses; +#[cfg(test)] +use crate::opus::rate_common::log2_frac; +#[cfg(test)] +pub use crate::opus::rate_common::pvq_v; pub const MAX_FINE_BITS: i32 = 8; const FINE_OFFSET: i32 = 21; pub const QTHETA_OFFSET: i32 = 4; pub const QTHETA_OFFSET_TWOPHASE: i32 = 16; const ALLOC_STEPS: i32 = 6; -/// Number of pulses represented by pseudo-pulse index `i`. -pub fn get_pulses(i: i32) -> i32 { - if i < 8 { i } else { (8 + (i & 7)) << ((i >> 3) - 1) } -} - -/// Conservative (rounded up) log2 of `val` with `frac` fractional bits. -pub fn log2_frac(mut val: u32, mut frac: u32) -> i32 { - let mut l = crate::opus::decode::range::ilog(val); - if val & val.wrapping_sub(1) != 0 { - if l > 16 { - val = ((val - 1) >> (l - 16)) + 1; - } else { - val <<= 16 - l; - } - let mut l = (l - 1) << frac; - loop { - let b = (val >> 16) as i32; - l += b << frac; - val = (val + b as u32) >> b; - val = (val * val + 0x7FFF) >> 15; - if frac == 0 { - break; - } - frac -= 1; - } - l + (val > 0x8000) as i32 - } else { - l -= 1; - l << frac - } -} - -/// PVQ codebook size V(N, K) as u128 (saturating far above 2^32). -pub fn pvq_v(n: usize, k: usize) -> u128 { - // V(N,K) = V(N-1,K) + V(N,K-1) + V(N-1,K-1); computed row by row. - if k == 0 { - return 1; - } - if n == 0 { - return 0; - } - let mut row = vec![0u128; k + 1]; // V(0, k) - row[0] = 1; - for _ in 0..n { - let mut prev_diag = row[0]; // V(n-1, 0) - // V(n, 0) = 1 - for kk in 1..=k { - let up = row[kk]; // V(n-1, kk) - let v = up.saturating_add(row[kk - 1]).saturating_add(prev_diag); - prev_diag = up; - row[kk] = v.min(u128::MAX >> 8); - } - row[0] = 1; - } - row[k] -} - -fn fits_in32(n: usize, k: i32) -> bool { - pvq_v(n, k as usize) <= u32::MAX as u128 -} - -/// Precomputed per-mode data shared by all CELT decoders. +/// Decoder mode: shared allocation tables and decoder-specific inverse transforms. pub struct Mode { - /// log2 of band width in 1/8 bits. - pub log_n: [i32; NB_EBANDS], - /// Index into `cache_bits` for `(LM+1)*NB_EBANDS + band` (-1 when unused). - pub cache_index: Vec, - /// `cache[0]` = max pseudo-pulses; `cache[k]` = bits(k) - 1 in 1/8 bits. - pub cache_bits: Vec, - /// Per-band maximum allocation: `[(2*LM + C-1)*NB_EBANDS + band]`. - pub caps: Vec, + core: crate::opus::rate_common::Mode, pub window: [f32; OVERLAP], pub mdct: [super::mdct::Imdct; MAX_LM + 1], } - impl Mode { - pub fn new() -> Mode { - let mut log_n = [0i32; NB_EBANDS]; - for (i, l) in log_n.iter_mut().enumerate() { - *l = log2_frac((EBANDS[i + 1] - EBANDS[i]) as u32, BITRES); - } - let (cache_index, cache_bits) = compute_pulse_cache(MAX_LM as i32); - let mut m = Mode { - log_n, - cache_index, - cache_bits, - caps: Vec::new(), + pub fn new() -> Self { + let mut mode = Self { + core: crate::opus::rate_common::Mode::new(), window: [0.0; OVERLAP], mdct: [super::mdct::Imdct::new(1920), super::mdct::Imdct::new(960), super::mdct::Imdct::new(480), super::mdct::Imdct::new(240)], }; - m.caps = compute_caps(&m); for i in 0..OVERLAP { let x = std::f64::consts::FRAC_PI_2 * (i as f64 + 0.5) / OVERLAP as f64; let s = x.sin(); - m.window[i] = (std::f64::consts::FRAC_PI_2 * s * s).sin() as f32; + mode.window[i] = (std::f64::consts::FRAC_PI_2 * s * s).sin() as f32; } - m - } - - #[inline] - pub fn cache(&self, lm: i32, band: usize) -> &[u8] { - let idx = self.cache_index[((lm + 1) as usize) * NB_EBANDS + band]; - &self.cache_bits[idx.max(0) as usize..] - } - - pub fn bits2pulses(&self, band: usize, lm: i32, bits: i32) -> i32 { - let cache = self.cache(lm, band); - let mut lo = 0i32; - let mut hi = cache[0] as i32; - let bits = bits - 1; - for _ in 0..LOG_MAX_PSEUDO { - let mid = (lo + hi + 1) >> 1; - if cache[mid as usize] as i32 >= bits { - hi = mid; - } else { - lo = mid; - } - } - let lo_bits = if lo == 0 { -1 } else { cache[lo as usize] as i32 }; - if bits - lo_bits <= cache[hi as usize] as i32 - bits { lo } else { hi } - } - - pub fn pulses2bits(&self, band: usize, lm: i32, pulses: i32) -> i32 { - if pulses == 0 { 0 } else { self.cache(lm, band)[pulses as usize] as i32 + 1 } - } - - /// Per-band allocation caps for this frame (`init_caps`). - pub fn init_caps(&self, lm: usize, c: usize) -> [i32; NB_EBANDS] { - let mut cap = [0i32; NB_EBANDS]; - for (i, cp) in cap.iter_mut().enumerate() { - let n = (EBANDS[i + 1] - EBANDS[i]) << lm; - *cp = ((self.caps[NB_EBANDS * (2 * lm + c - 1) + i] as i32 + 64) * c as i32 * n) >> 2; - } - cap + mode } } - -fn compute_pulse_cache(lm_max: i32) -> (Vec, Vec) { - let nb = NB_EBANDS; - let rows = (lm_max + 2) as usize; - let mut index = vec![-1i16; nb * rows]; - let mut entries: Vec<(usize, i32, usize)> = Vec::new(); // (N, K, offset) - let mut curr = 0usize; - for i in 0..rows { - for j in 0..nb { - let n = (((EBANDS[j + 1] - EBANDS[j]) << i) >> 1) as usize; - index[i * nb + j] = -1; - // Reuse the entry of an earlier band with the same size. - 'search: for k in 0..=i { - for nn in 0..nb { - if k == i && nn >= j { - break; - } - if n == (((EBANDS[nn + 1] - EBANDS[nn]) << k) >> 1) as usize { - index[i * nb + j] = index[k * nb + nn]; - break 'search; - } - } - } - if index[i * nb + j] == -1 && n != 0 { - let mut k = 0; - while k < MAX_PSEUDO && fits_in32(n, get_pulses(k + 1)) { - k += 1; - } - index[i * nb + j] = curr as i16; - entries.push((n, k, curr)); - curr += (k + 1) as usize; - } - } +impl std::ops::Deref for Mode { + type Target = crate::opus::rate_common::Mode; + fn deref(&self) -> &Self::Target { + &self.core } - let mut bits = vec![0u8; curr]; - for &(n, k, off) in &entries { - let maxk = get_pulses(k) as usize; - let tmp = required_bits(n, maxk, BITRES); - for j in 1..=k { - bits[off + j as usize] = (tmp[get_pulses(j) as usize] - 1) as u8; - } - bits[off] = k as u8; - } - (index, bits) -} - -/// Bits (1/2^frac units, rounded up) needed for each K in `0..=maxk` for dimension `n`. -fn required_bits(n: usize, maxk: usize, frac: u32) -> Vec { - let mut out = vec![0i32; maxk + 1]; - if n == 1 { - for o in out.iter_mut().skip(1) { - *o = 1 << frac; - } - } else { - for (k, o) in out.iter_mut().enumerate().skip(1) { - let v = pvq_v(n, k); - *o = log2_frac(v as u32, frac); - } - } - out -} - -fn compute_caps(m: &Mode) -> Vec { - let nb = NB_EBANDS; - let mut caps = vec![0u8; (MAX_LM + 1) * 2 * nb]; - for i in 0..=MAX_LM as i32 { - for c in 1..=2i32 { - for j in 0..nb { - let mut n0 = EBANDS[j + 1] - EBANDS[j]; - let max_bits = if n0 << i == 1 { - (c * (1 + MAX_FINE_BITS)) << BITRES - } else { - let mut lm0 = 0; - if n0 > 2 { - n0 >>= 1; - lm0 -= 1; - } else if n0 <= 1 { - lm0 = i.min(1); - n0 <<= lm0; - } - let pcache = m.cache(lm0, j); - let mut mb = pcache[pcache[0] as usize] as i32 + 1; - let mut n = n0; - for k in 0..(i - lm0) { - mb <<= 1; - let offset = ((m.log_n[j] + ((lm0 + k) << BITRES)) >> 1) - QTHETA_OFFSET; - let num = 459 * ((2 * n - 1) * offset + mb); - let den = ((2 * n - 1) << 9) - 459; - let qb = ((num + (den >> 1)) / den).min(57); - mb += qb; - n <<= 1; - } - if c == 2 { - mb <<= 1; - let offset = ((m.log_n[j] + (i << BITRES)) >> 1) - if n == 2 { QTHETA_OFFSET_TWOPHASE } else { QTHETA_OFFSET }; - let ndof = 2 * n - 1 - (n == 2) as i32; - let p = if n == 2 { 512 } else { 487 }; - let num = p * (mb + ndof * offset); - let den = (ndof << 9) - p; - let qb = ((num + (den >> 1)) / den).min(if n == 2 { 64 } else { 61 }); - mb += qb; - } - let ndof = c * n + if c == 2 && n > 2 { 1 } else { 0 }; - let mut offset = ((m.log_n[j] + (i << BITRES)) >> 1) - FINE_OFFSET; - if n == 2 { - offset += (1 << BITRES) >> 2; - } - let num = mb + ndof * offset; - let den = (ndof - 1) << BITRES; - let qb = ((num + (den >> 1)) / den).min(MAX_FINE_BITS); - mb += (c * qb) << BITRES; - mb - }; - let w = c * ((EBANDS[j + 1] - EBANDS[j]) << i); - let v = (4 * max_bits / w) - 64; - caps[((i * 2 + c - 1) as usize) * nb + j] = v.clamp(0, 255) as u8; - } - } - } - caps } /// Result of the allocation process. @@ -296,89 +69,8 @@ pub struct AllocationInput<'a, 'd> { pub fn compute_allocation(input: AllocationInput<'_, '_>) -> Allocation { let AllocationInput { mode: m, start, end, offsets, cap, trim: alloc_trim, total, channels: c, lm, decoder: dec } = input; - let ci = c as i32; - let lmi = lm as i32; - let mut total = total.max(0); - let mut skip_start = start; - let skip_rsv = if total >= 1 << BITRES { 1 << BITRES } else { 0 }; - total -= skip_rsv; - let mut intensity_rsv = 0; - let mut dual_stereo_rsv = 0; - if c == 2 { - intensity_rsv = LOG2_FRAC_TABLE[end - start]; - if intensity_rsv > total { - intensity_rsv = 0; - } else { - total -= intensity_rsv; - dual_stereo_rsv = if total >= 1 << BITRES { 1 << BITRES } else { 0 }; - total -= dual_stereo_rsv; - } - } - let mut bits1 = [0i32; NB_EBANDS]; - let mut bits2 = [0i32; NB_EBANDS]; - let mut thresh = [0i32; NB_EBANDS]; - let mut trim_offset = [0i32; NB_EBANDS]; - for j in start..end { - let w = EBANDS[j + 1] - EBANDS[j]; - thresh[j] = (ci << BITRES).max(((3 * w) << lmi << BITRES) >> 4); - trim_offset[j] = (ci * w * (alloc_trim - 5 - lmi) * (end as i32 - j as i32 - 1) * (1 << (lmi + BITRES as i32))) >> 6; - if w << lmi == 1 { - trim_offset[j] -= ci << BITRES; - } - } - let nb_alloc = ALLOC.len() as i32; - let mut lo = 1; - let mut hi = nb_alloc - 1; - loop { - let mut done = false; - let mut psum = 0; - let mid = (lo + hi) >> 1; - for j in (start..end).rev() { - let n = EBANDS[j + 1] - EBANDS[j]; - let mut bitsj = (ci * n * (ALLOC[mid as usize][j] as i32)) << lmi >> 2; - if bitsj > 0 { - bitsj = (bitsj + trim_offset[j]).max(0); - } - bitsj += offsets[j]; - if bitsj >= thresh[j] || done { - done = true; - psum += bitsj.min(cap[j]); - } else if bitsj >= ci << BITRES { - psum += ci << BITRES; - } - } - if psum > total { - hi = mid - 1; - } else { - lo = mid + 1; - } - if lo > hi { - break; - } - } - hi = lo; - lo -= 1; - for j in start..end { - let n = EBANDS[j + 1] - EBANDS[j]; - let mut bits1j = (ci * n * (ALLOC[lo as usize][j] as i32)) << lmi >> 2; - let mut bits2j = if hi >= nb_alloc { cap[j] } else { (ci * n * (ALLOC[hi as usize][j] as i32)) << lmi >> 2 }; - if bits1j > 0 { - bits1j = (bits1j + trim_offset[j]).max(0); - } - if bits2j > 0 { - bits2j = (bits2j + trim_offset[j]).max(0); - } - if lo > 0 { - bits1j += offsets[j]; - } - bits2j += offsets[j]; - if offsets[j] > 0 { - skip_start = j; - } - bits2j = (bits2j - bits1j).max(0); - bits1[j] = bits1j; - bits2[j] = bits2j; - } + let crate::opus::rate_common::Bracket { skip_start, bits1, bits2, thresh, total, skip_rsv, intensity_rsv, dual_stereo_rsv } = + crate::opus::rate_common::bracket(start, end, offsets, cap, alloc_trim, total, c, lm); interp_bits2pulses(InterpPulsesInput { mode: m, start, diff --git a/crates/audio/src/opus/decode/celt/tables.rs b/crates/audio/src/opus/decode/celt/tables.rs index f2381f714..b65d05c09 100644 --- a/crates/audio/src/opus/decode/celt/tables.rs +++ b/crates/audio/src/opus/decode/celt/tables.rs @@ -14,6 +14,7 @@ pub const MAX_LM: usize = 3; pub const PREEMPH: f32 = 0.850_006_1; /// Static allocation table (Table 57), indexed `[quality][band]`, in 1/32 bit per MDCT bin. +#[cfg(not(feature = "opus-encode"))] pub const ALLOC: [[u8; NB_EBANDS]; 11] = transpose([ [0, 90, 110, 118, 126, 134, 144, 152, 162, 172, 200], [0, 80, 100, 110, 119, 127, 137, 145, 155, 165, 200], @@ -38,6 +39,7 @@ pub const ALLOC: [[u8; NB_EBANDS]; 11] = transpose([ [0, 0, 0, 0, 0, 1, 1, 1, 1, 20, 104], ]); +#[cfg(not(feature = "opus-encode"))] const fn transpose(t: [[u8; 11]; NB_EBANDS]) -> [[u8; NB_EBANDS]; 11] { let mut out = [[0u8; NB_EBANDS]; 11]; let mut b = 0; diff --git a/crates/audio/src/opus/decode/silk/resampler.rs b/crates/audio/src/opus/decode/silk/resampler.rs index 0671778cc..7d3f803ef 100644 --- a/crates/audio/src/opus/decode/silk/resampler.rs +++ b/crates/audio/src/opus/decode/silk/resampler.rs @@ -4,25 +4,7 @@ //! RFC 6716 Table 54 (and the integer delays used when no rate conversion is needed), so that the //! SILK output lines up with the CELT layer in Hybrid mode. -/// Dot product with eight independent partial sums, so the additions pipeline and the loop -/// vectorises instead of waiting on one 4-cycle add chain. Rounds differently from a -/// sequential sum (the resampler is not normative). -#[inline] -fn dot8(a: &[f32], b: &[f32]) -> f32 { - let mut s = [0f32; 8]; - let (ca, ra) = a.as_chunks::<8>(); - let (cb, rb) = b.as_chunks::<8>(); - for (x, y) in ca.iter().zip(cb) { - for l in 0..8 { - s[l] += x[l] * y[l]; - } - } - let mut acc = ((s[0] + s[4]) + (s[1] + s[5])) + ((s[2] + s[6]) + (s[3] + s[7])); - for (x, y) in ra.iter().zip(rb) { - acc += x * y; - } - acc -} +use crate::opus::dot::dot8; #[derive(Clone)] pub struct Resampler { diff --git a/crates/audio/src/opus/dot.rs b/crates/audio/src/opus/dot.rs new file mode 100644 index 000000000..536209e0a --- /dev/null +++ b/crates/audio/src/opus/dot.rs @@ -0,0 +1,20 @@ +//! Shared fixed-order FIR dot product. + +/// Dot product with eight independent partial sums, which the compiler can vectorise (a single +/// running sum is bound by the latency of the float addition). The sums are combined pairwise. +#[inline] +pub(super) fn dot8(a: &[f32], b: &[f32]) -> f32 { + let mut s = [0f32; 8]; + let (ca, ra) = a.as_chunks::<8>(); + let (cb, rb) = b.as_chunks::<8>(); + for (x, y) in ca.iter().zip(cb) { + for l in 0..8 { + s[l] += x[l] * y[l]; + } + } + let mut acc = ((s[0] + s[4]) + (s[1] + s[5])) + ((s[2] + s[6]) + (s[3] + s[7])); + for (x, y) in ra.iter().zip(rb) { + acc += x * y; + } + acc +} diff --git a/crates/audio/src/opus/encoder/bands.rs b/crates/audio/src/opus/encoder/bands.rs index 604aab595..697332eb2 100644 --- a/crates/audio/src/opus/encoder/bands.rs +++ b/crates/audio/src/opus/encoder/bands.rs @@ -8,7 +8,8 @@ use crate::opus::encoder::range::{BITRES, RangeEncoder, ilog}; use crate::opus::encoder::rate::{Mode, QTHETA_OFFSET, get_pulses}; -use crate::opus::encoder::tables::{EBANDS, NB_EBANDS}; +use crate::opus::encoder::tables::EBANDS; +use crate::opus::encoder::{CeltFrame, rate::Allocation}; pub const SPREAD_NORMAL: usize = 2; @@ -55,7 +56,7 @@ fn exp_rotation1_inv(x: &mut [f32], len: usize, stride: usize, c: f32, s: f32) { if len <= stride { return; } - if len >= 2 * stride + 1 { + if len > 2 * stride { for i in 0..=len - 2 * stride - 1 { let x1 = x[i]; let x2 = x[i + stride]; @@ -96,12 +97,24 @@ pub fn exp_rotation_inv(x: &mut [f32], len: usize, k: usize, spread: usize) { } } +/// Relative slack for the pulse-candidate filter in [`pvq_search`]: far above the f32 rounding of +/// the filter score (a few 1e-7), far below the spread of real scores. +const PVQ_SCORE_SLACK: f32 = 1e-5; + /// Greedy PVQ search: the integer vector `y` with `sum |y| = k` closest in direction to `x`. pub fn pvq_search(x: &[f32], k: usize, y: &mut [i32]) { let n = x.len(); y[..n].fill(0); - let abs: Vec = x.iter().map(|v| v.abs()).collect(); - let sum: f32 = abs.iter().sum(); + // `abs`, `dy` (2 y + 1, infinite in the padding) and the per-pass scores `score`, each padded to + // whole 8-lane chunks. + let np = n.next_multiple_of(8); + let mut buf = vec![0f32; 3 * np]; + let (abs, rest) = buf.split_at_mut(np); + let (dy, score) = rest.split_at_mut(np); + for (a, v) in abs.iter_mut().zip(x) { + *a = v.abs(); + } + let sum: f32 = abs[..n].iter().sum(); if !sum.is_finite() || sum <= 1e-15 { y[0] = k as i32; return; @@ -113,30 +126,69 @@ pub fn pvq_search(x: &[f32], k: usize, y: &mut [i32]) { // Pre-project onto the pyramid (rounding down) and place the rest greedily. let r = (k as f32 - 1.0) / sum; for j in 0..n { - let p = (abs[j] * r).floor() as i32; + // Non-negative and at most k, so the truncating cast is the floor. + let p = (abs[j] * r) as i32; y[j] = p; placed += p as usize; xy += abs[j] * p as f32; yy += (p * p) as f32; } } + for j in 0..np { + dy[j] = if j < n { 2.0 * y[j] as f32 + 1.0 } else { f32::INFINITY }; + } while placed < k { + // Vectorisable scores (xy + |x_j|)^2 / (yy + 2 y_j + 1) and their maximum. + for ((sc, &a), &d) in score.iter_mut().zip(abs.iter()).zip(dy.iter()) { + let t = xy + a; + *sc = t * t / (yy + d); + } + let mut lane_max = [0f32; 8]; + for c in score.as_chunks::<8>().0 { + for i in 0..8 { + lane_max[i] = if c[i] > lane_max[i] { c[i] } else { lane_max[i] }; + } + } + let top = lane_max.iter().fold(0f32, |m, &v| if v > m { v } else { m }); + // The exact cross-multiplied test below, in index order with the first maximum winning, only + // matters for entries within rounding distance of the top score: every other entry is beaten + // by the top one whenever it is reached and never beats it afterwards. Running the test on + // those alone therefore picks the same pulse as running it on all of them. + let filter = top > 1e-30 && top.is_finite(); + let lo = top * (1.0 - PVQ_SCORE_SLACK); + // Usually exactly one entry qualifies; its index falls out of a lane-wise count. + let mut count = [0u32; 8]; + let mut index = [0u32; 8]; + for (ci, c) in score.as_chunks::<8>().0.iter().enumerate() { + for i in 0..8 { + let hit = u32::from(c[i] >= lo); + count[i] += hit; + index[i] += hit * (ci * 8 + i) as u32; + } + } let mut best = 0usize; - let mut best_num = -1f32; - let mut best_den = 1f32; - for j in 0..n { - let num = xy + abs[j]; - let num = num * num; - let den = yy + 2.0 * y[j] as f32 + 1.0; - if num * best_den > best_num * den { - best_num = num; - best_den = den; - best = j; + if filter && count.iter().sum::() == 1 { + best = index.iter().sum::() as usize; + } else { + let mut best_num = -1f32; + let mut best_den = 1f32; + for j in 0..n { + if !filter || score[j] >= lo { + let num = xy + abs[j]; + let num = num * num; + let den = yy + 2.0 * y[j] as f32 + 1.0; + if num * best_den > best_num * den { + best_num = num; + best_den = den; + best = j; + } + } } } xy += abs[best]; yy += 2.0 * y[best] as f32 + 1.0; y[best] += 1; + dy[best] = 2.0 * y[best] as f32 + 1.0; placed += 1; } for j in 0..n { @@ -146,25 +198,45 @@ pub fn pvq_search(x: &[f32], k: usize, y: &mut [i32]) { } } -/// Encodes the PVQ codeword `y` (`sum |y| = k`) as the index the RFC 6716 §4.3.4.2 decoder maps -/// back to `y`. -pub fn encode_pulses(y: &[i32], k: usize, enc: &mut RangeEncoder, v: &mut Vec) { - let n = y.len(); - let w = k + 1; - v.clear(); - v.resize((n + 1) * w, 0); +/// Largest band width and pulse count the encoder codes (176 coefficients, `get_pulses(40)` = 128). +const PVQ_MAX_N: usize = 176; +const PVQ_MAX_K: usize = 128; + +/// Row-major `V(n, k)` for `n <= max_n`, `k <= max_k` with row stride `max_k + 1`, saturated at +/// `u32::MAX` (a codebook that large is never coded, see `rate::fits_in32`). +fn build_pvq_table(max_n: usize, max_k: usize) -> Vec { + let w = max_k + 1; + let mut v = vec![0u32; (max_n + 1) * w]; v[0] = 1; - for nn in 1..=n { + for nn in 1..=max_n { v[nn * w] = 1; - for kk in 1..=k { - let a = v[(nn - 1) * w + kk]; - let b = v[nn * w + kk - 1]; - let c = v[(nn - 1) * w + kk - 1]; - v[nn * w + kk] = a.saturating_add(b).saturating_add(c); + for kk in 1..=max_k { + let sum = v[(nn - 1) * w + kk] as u64 + v[nn * w + kk - 1] as u64 + v[(nn - 1) * w + kk - 1] as u64; + v[nn * w + kk] = sum.min(u32::MAX as u64) as u32; } } - let total = v[n * w + k]; - let vv = |nn: usize, kx: usize| v[nn * w + kx]; + v +} + +fn pvq_table() -> &'static [u32] { + static TABLE: std::sync::OnceLock> = std::sync::OnceLock::new(); + TABLE.get_or_init(|| build_pvq_table(PVQ_MAX_N, PVQ_MAX_K)) +} + +/// Encodes the PVQ codeword `y` (`sum |y| = k`) as the index the RFC 6716 §4.3.4.2 decoder maps +/// back to `y`. +pub fn encode_pulses(y: &[i32], k: usize, enc: &mut RangeEncoder) { + let n = y.len(); + // Shapes outside the coded range (never produced by the band layout) get a local table. + let local; + let (v, w) = if n <= PVQ_MAX_N && k <= PVQ_MAX_K { + (pvq_table(), PVQ_MAX_K + 1) + } else { + local = build_pvq_table(n, k); + (&local[..], k + 1) + }; + let row = |nn: usize| &v[nn * w..(nn + 1) * w]; + let total = row(n)[k] as u64; let mut index = 0u64; let mut kk = k; for j in 0..n { @@ -173,10 +245,11 @@ pub fn encode_pulses(y: &[i32], k: usize, enc: &mut RangeEncoder, v: &mut Vec { pub i: usize, pub spread: usize, pub remaining_bits: i32, - pvq_v: Vec, iy: Vec, tmp: Vec, } @@ -217,7 +289,7 @@ impl BandEnc<'_, '_> { let q = qn as u32; let half = q >> 1; let ft = (half + 1) * (half + 1); - let (fl, fs) = if it <= half { (it * (it + 1) >> 1, it + 1) } else { (ft - ((q + 1 - it) * (q + 2 - it) >> 1), q + 1 - it) }; + let (fl, fs) = if it <= half { ((it * (it + 1)) >> 1, it + 1) } else { (ft - (((q + 1 - it) * (q + 2 - it)) >> 1), q + 1 - it) }; self.enc.encode(fl, fl + fs, ft); itheta = ((it * 16384) / q) as i32; } @@ -287,36 +359,29 @@ impl BandEnc<'_, '_> { iy.clear(); iy.resize(n, 0); pvq_search(&tmp, k, &mut iy); - encode_pulses(&iy, k, self.enc, &mut self.pvq_v); + encode_pulses(&iy, k, self.enc); self.iy = iy; self.tmp = tmp; } } } -/// Codes all band shapes (`quant_all_bands`, encoder side, long blocks). `x` holds `c` channels -/// of `n` normalised MDCT coefficients each. -#[allow(clippy::too_many_arguments)] -pub fn quant_all_bands( - m: &Mode, - start: usize, - end: usize, - x: &[f32], - n_total: usize, - c: usize, - pulses: &[i32; NB_EBANDS], - spread: usize, - dual_stereo: bool, - intensity: usize, - total_bits: i32, - mut balance: i32, - enc: &mut RangeEncoder, - lm: usize, - coded_bands: usize, -) { +/// Shape-coding controls, independent of the selected band allocation. +pub struct BandCoding { + pub spread: usize, + pub total_bits: i32, +} + +/// Codes all band shapes (`quant_all_bands`, encoder side, long blocks). `x` holds +/// `frame.channels` channels of `n_total` normalised MDCT coefficients each. +pub fn quant_all_bands(m: &Mode, frame: CeltFrame, x: &[f32], n_total: usize, allocation: &Allocation, coding: BandCoding, enc: &mut RangeEncoder) { + let CeltFrame { start, end, channels: c, lm } = frame; + let Allocation { pulses, dual_stereo, intensity, coded_bands, balance, .. } = allocation; + let (dual_stereo, intensity, coded_bands, mut balance) = (*dual_stereo, *intensity, *coded_bands, *balance); + let BandCoding { spread, total_bits } = coding; let mm = 1usize << lm; debug_assert!(c == 1 || dual_stereo || intensity <= start, "stereo frames are coded as dual stereo"); - let mut ctx = BandEnc { m, enc, i: 0, spread, remaining_bits: 0, pvq_v: Vec::new(), iy: Vec::new(), tmp: Vec::new() }; + let mut ctx = BandEnc { m, enc, i: 0, spread, remaining_bits: 0, iy: Vec::new(), tmp: Vec::new() }; for i in start..end { ctx.i = i; let xo = mm * EBANDS[i] as usize; @@ -362,7 +427,7 @@ mod tests { x[i + stride] = c * x2 + s * x1; x[i] = c * x1 + ms * x2; } - if len >= 2 * stride + 1 { + if len > 2 * stride { for i in (0..=len - 2 * stride - 1).rev() { let x1 = x[i]; let x2 = x[i + stride]; @@ -466,12 +531,11 @@ mod tests { #[test] fn pvq_index_roundtrip_exhaustive() { - let mut v = Vec::new(); for n in 1..6usize { for k in 1..6usize { for y in all_codewords(n, k) { let mut e = RangeEncoder::new(); - encode_pulses(&y, k, &mut e, &mut v); + encode_pulses(&y, k, &mut e); let buf = e.done(16); let mut d = RangeDecoder::new(&buf); assert_eq!(decode_pulses(n, k, &mut d), y, "n={n} k={k}"); @@ -480,6 +544,165 @@ mod tests { } } + /// The per-call table encoder this module used before the shared table. + fn encode_pulses_reference(y: &[i32], k: usize, enc: &mut RangeEncoder) { + let n = y.len(); + let w = k + 1; + let mut v = vec![0u64; (n + 1) * w]; + v[0] = 1; + for nn in 1..=n { + v[nn * w] = 1; + for kk in 1..=k { + let a = v[(nn - 1) * w + kk]; + let b = v[nn * w + kk - 1]; + let c = v[(nn - 1) * w + kk - 1]; + v[nn * w + kk] = a.saturating_add(b).saturating_add(c); + } + } + let total = v[n * w + k]; + let vv = |nn: usize, kx: usize| v[nn * w + kx]; + let mut index = 0u64; + let mut kk = k; + for j in 0..n { + if kk == 0 { + break; + } + let rem = n - j; + let t = y[j].unsigned_abs() as usize; + let mut base: u64 = (0..kk - t).map(|m| vv(rem - 1, m)).sum(); + if y[j] < 0 { + base += (vv(rem - 1, kk) + vv(rem, kk)) / 2; + } + index += base; + kk -= t; + } + if total >= 2 { + enc.uint(index as u32, total as u32); + } + } + + #[test] + fn pvq_table_encode_matches_per_call_table() { + let mut seed = 0x9E37_79B9_7F4A_7C15u64; + let mut next = move |m: u64| { + seed ^= seed << 13; + seed ^= seed >> 7; + seed ^= seed << 17; + seed % m + }; + let mut checked = 0; + for n in (1..=PVQ_MAX_N).step_by(5).chain([PVQ_MAX_N]) { + for k in (1..=PVQ_MAX_K).step_by(3).chain([PVQ_MAX_K]) { + if crate::opus::encoder::rate::pvq_v(n, k) > u32::MAX as u128 { + continue; + } + for _ in 0..4 { + let mut y = vec![0i32; n]; + for _ in 0..k { + let at = next(n as u64) as usize; + y[at] += if next(2) == 0 { 1 } else { -1 }; + } + // Opposite pulses cancel; top up so that sum |y| = k again. + let mut have: usize = y.iter().map(|v| v.unsigned_abs() as usize).sum(); + while have < k { + let at = next(n as u64) as usize; + y[at] += if y[at] < 0 { -1 } else { 1 }; + have += 1; + } + let (mut a, mut b) = (RangeEncoder::new(), RangeEncoder::new()); + encode_pulses(&y, k, &mut a); + encode_pulses_reference(&y, k, &mut b); + assert_eq!(a.done(16), b.done(16), "n={n} k={k}"); + checked += 1; + } + } + } + assert!(checked > 100, "{checked}"); + } + + /// The scalar greedy search the vectorised one replaced. + fn pvq_search_ref(x: &[f32], k: usize, y: &mut [i32]) { + let n = x.len(); + y[..n].fill(0); + let abs: Vec = x.iter().map(|v| v.abs()).collect(); + let sum: f32 = abs.iter().sum(); + if !sum.is_finite() || sum <= 1e-15 { + y[0] = k as i32; + return; + } + let mut placed = 0usize; + let mut xy = 0f32; + let mut yy = 0f32; + if k > n / 2 { + let r = (k as f32 - 1.0) / sum; + for j in 0..n { + let p = (abs[j] * r).floor() as i32; + y[j] = p; + placed += p as usize; + xy += abs[j] * p as f32; + yy += (p * p) as f32; + } + } + while placed < k { + let mut best = 0usize; + let mut best_num = -1f32; + let mut best_den = 1f32; + for j in 0..n { + let num = xy + abs[j]; + let num = num * num; + let den = yy + 2.0 * y[j] as f32 + 1.0; + if num * best_den > best_num * den { + best_num = num; + best_den = den; + best = j; + } + } + xy += abs[best]; + yy += 2.0 * y[best] as f32 + 1.0; + y[best] += 1; + placed += 1; + } + for j in 0..n { + if x[j] < 0.0 { + y[j] = -y[j]; + } + } + } + + #[test] + fn pvq_search_matches_scalar_reference() { + let mut state = 0x9e3779b97f4a7c15u64; + let mut rnd = move || { + state ^= state << 13; + state ^= state >> 7; + state ^= state << 17; + (state >> 40) as f32 / (1u64 << 24) as f32 * 2.0 - 1.0 + }; + let mut a = vec![0i32; 200]; + let mut b = vec![0i32; 200]; + let mut cases = 0; + for round in 0..6000 { + let n = 1 + (round * 37 + (rnd().abs() * 176.0) as usize) % 176; + let k = 1 + (rnd().abs() * 127.0) as usize; + let scale = [1.0f32, 1e-9, 1e6, 3e-3][round % 4]; + let x: Vec = (0..n) + .map(|_| match round % 6 { + 0 => rnd(), + 1 => if rnd() > 0.8 { rnd() } else { 0.0 }, + 2 => (rnd() * 4.0).round() / 4.0, + 3 => if rnd() > 0.0 { 1.0 } else { -1.0 }, + 4 => rnd().powi(5), + _ => if rnd() > 0.95 { 3.0 * rnd() } else { 1e-4 * rnd() }, + } * scale) + .collect(); + pvq_search(&x, k, &mut a); + pvq_search_ref(&x, k, &mut b); + assert_eq!(a[..n], b[..n], "n={n} k={k} round={round}"); + cases += 1; + } + assert_eq!(cases, 6000); + } + #[test] fn pvq_search_finds_good_direction() { let x: Vec = (0..32).map(|i| ((i as f32) * 0.7).sin()).collect(); diff --git a/crates/audio/src/opus/encoder/energy.rs b/crates/audio/src/opus/encoder/energy.rs index 8d4f357d0..fb5cc5a4b 100644 --- a/crates/audio/src/opus/encoder/energy.rs +++ b/crates/audio/src/opus/encoder/energy.rs @@ -4,8 +4,9 @@ //! The quantised energies are updated with exactly the decoder's arithmetic so the prediction //! state never drifts between encoder and decoder. +use crate::opus::encoder::CeltFrame; use crate::opus::encoder::range::RangeEncoder; -use crate::opus::encoder::rate::MAX_FINE_BITS; +use crate::opus::encoder::rate::{Allocation, MAX_FINE_BITS}; use crate::opus::encoder::tables::{BETA_COEF, BETA_INTRA, E_PROB_MODEL, NB_EBANDS, PRED_COEF, SMALL_ENERGY_ICDF}; const LAPLACE_MINP: u32 = 1; @@ -56,18 +57,15 @@ pub fn laplace_encode(enc: &mut RangeEncoder, value: i32, mut fs: u32, decay: u3 /// `quant_coarse_energy` for one prediction mode. `e` holds the target log2 energies (relative /// to `E_MEANS`), `old_e` the quantised energies, updated in place. `budget` is the frame size in /// bits. -#[allow(clippy::too_many_arguments)] pub fn quant_coarse_energy( - start: usize, - end: usize, + frame: CeltFrame, e: &[f32; 2 * NB_EBANDS], old_e: &mut [f32; 2 * NB_EBANDS], intra: bool, enc: &mut RangeEncoder, - c: usize, - lm: usize, budget: i32, ) { + let CeltFrame { start, end, channels: c, lm } = frame; let prob = &E_PROB_MODEL[lm][intra as usize]; let (coef, beta) = if intra { (0.0, BETA_INTRA) } else { (PRED_COEF[lm], BETA_COEF[lm]) }; let mut prev = [0f32; 2]; @@ -112,14 +110,13 @@ pub fn quant_coarse_energy( /// `quant_fine_energy`: raw bits refining each band's quantised energy. pub fn quant_fine_energy( - start: usize, - end: usize, + frame: CeltFrame, e: &[f32; 2 * NB_EBANDS], old_e: &mut [f32; 2 * NB_EBANDS], fine_quant: &[i32; NB_EBANDS], enc: &mut RangeEncoder, - c: usize, ) { + let CeltFrame { start, end, channels: c, .. } = frame; for i in start..end { let fq = fine_quant[i]; if fq <= 0 { @@ -139,18 +136,16 @@ pub fn quant_fine_energy( /// `quant_energy_finalise`: spends the bits left at the end of the frame on one more bit of /// energy resolution per band, in priority order. -#[allow(clippy::too_many_arguments)] pub fn quant_energy_finalise( - start: usize, - end: usize, + frame: CeltFrame, e: &[f32; 2 * NB_EBANDS], old_e: &mut [f32; 2 * NB_EBANDS], - fine_quant: &[i32; NB_EBANDS], - fine_priority: &[i32; NB_EBANDS], + allocation: &Allocation, mut bits_left: i32, enc: &mut RangeEncoder, - c: usize, ) { + let CeltFrame { start, end, channels: c, .. } = frame; + let Allocation { fine_quant, fine_priority, .. } = allocation; for prio in 0..2 { let mut i = start; while i < end && bits_left >= c as i32 { diff --git a/crates/audio/src/opus/encoder/mdct.rs b/crates/audio/src/opus/encoder/mdct.rs index 6a42cce96..7cac21d93 100644 --- a/crates/audio/src/opus/encoder/mdct.rs +++ b/crates/audio/src/opus/encoder/mdct.rs @@ -9,84 +9,7 @@ use crate::opus::encoder::tables::OVERLAP; -#[derive(Clone, Copy, Default, Debug)] -pub struct Cpx { - pub re: f32, - pub im: f32, -} - -impl Cpx { - #[inline] - fn mul(self, o: Cpx) -> Cpx { - Cpx { re: self.re * o.re - self.im * o.im, im: self.re * o.im + self.im * o.re } - } - #[inline] - fn add(self, o: Cpx) -> Cpx { - Cpx { re: self.re + o.re, im: self.im + o.im } - } -} - -/// Mixed-radix (2, 3, 4, 5) forward complex FFT, unscaled (decimation in time, recursive). -pub struct Fft { - n: usize, - factors: Vec, - twiddle: Vec, -} - -impl Fft { - pub fn new(n: usize) -> Fft { - let mut factors = Vec::new(); - let mut m = n; - for p in [4usize, 2, 3, 5] { - while m.is_multiple_of(p) { - factors.push(p); - m /= p; - } - } - assert_eq!(m, 1, "unsupported FFT size {n}"); - let twiddle = (0..n) - .map(|k| { - let ph = -2.0 * std::f64::consts::PI * k as f64 / n as f64; - Cpx { re: ph.cos() as f32, im: ph.sin() as f32 } - }) - .collect(); - Fft { n, factors, twiddle } - } - - pub fn process(&self, data: &mut [Cpx], scratch: &mut Vec) { - scratch.clear(); - scratch.extend_from_slice(&data[..self.n]); - self.rec(&mut data[..self.n], scratch, 1, self.n, 0, 1); - } - - fn rec(&self, out: &mut [Cpx], inp: &[Cpx], in_stride: usize, n: usize, fi: usize, tw_stride: usize) { - let p = self.factors[fi]; - let m = n / p; - if m == 1 { - for (i, o) in out.iter_mut().enumerate().take(p) { - *o = inp[i * in_stride]; - } - } else { - for q in 0..p { - self.rec(&mut out[q * m..(q + 1) * m], &inp[q * in_stride..], in_stride * p, m, fi + 1, tw_stride * p); - } - } - let big = self.n; - let mut t = [Cpx::default(); 5]; - for k in 0..m { - for (q, tq) in t.iter_mut().enumerate().take(p) { - *tq = out[q * m + k].mul(self.twiddle[(q * k * tw_stride) % big]); - } - for q2 in 0..p { - let mut acc = Cpx::default(); - for (q, tq) in t.iter().enumerate().take(p) { - acc = acc.add(tq.mul(self.twiddle[((q * q2) % p) * (big / p)])); - } - out[q2 * m + k] = acc; - } - } - } -} +use crate::opus::fft::{Cpx, Fft}; /// Forward MDCT producing `n` coefficients from a block of `n + OVERLAP` samples. pub struct Mdct { @@ -199,6 +122,31 @@ mod tests { .collect() } + #[test] + fn fft_matches_dft() { + // Every radix combination, including the 480-point size of the 20 ms frame. + for &n in &[2usize, 3, 4, 5, 6, 8, 12, 15, 16, 20, 24, 30, 48, 60, 120, 240, 480] { + let fft = Fft::new(n); + let x: Vec = (0..n).map(|i| Cpx { re: ((i * 37 % 101) as f32 - 50.0) / 7.0, im: ((i * 53 % 89) as f32 - 44.0) / 9.0 }).collect(); + let mut y = x.clone(); + fft.process(&mut y, &mut Vec::new()); + for k in 0..n { + let (mut re, mut im) = (0f64, 0f64); + for (j, v) in x.iter().enumerate() { + let ph = -2.0 * std::f64::consts::PI * ((j * k) % n) as f64 / n as f64; + re += v.re as f64 * ph.cos() - v.im as f64 * ph.sin(); + im += v.re as f64 * ph.sin() + v.im as f64 * ph.cos(); + } + assert!( + (y[k].re as f64 - re).abs() < 2e-3 && (y[k].im as f64 - im).abs() < 2e-3, + "n={n} k={k}: ({}, {}) vs ({re}, {im})", + y[k].re, + y[k].im + ); + } + } + } + #[test] fn matches_direct_formula() { for &n in &[240usize, 960] { diff --git a/crates/audio/src/opus/encoder/mod.rs b/crates/audio/src/opus/encoder/mod.rs index cbb0d7984..8a3df07f1 100644 --- a/crates/audio/src/opus/encoder/mod.rs +++ b/crates/audio/src/opus/encoder/mod.rs @@ -10,8 +10,9 @@ //! above in the same range-coded frame; constant packet size. //! - **CELT** (fullband): MDCT transform coding; constant packet size. //! -//! From EffectCraft's `effectcraft-opusenc` (storytold/effectcraft@c618429). Public items keep their -//! names, so EffectCraft can re-export this module. +//! From EffectCraft's `effectcraft-opusenc` (storytold/effectcraft@c618429), with its later speed-ups +//! and CELT stage refactor up to storytold/effectcraft@d6c9ad66. Public items keep their names, so +//! EffectCraft can re-export this module. //! //! ``` //! # #[cfg(feature = "opus-encode")] { @@ -26,6 +27,7 @@ //! # } //! ``` +#![deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable)] // The bit-allocation code mirrors the normative integer arithmetic; keep C-like shift expressions. #![allow(clippy::precedence, clippy::int_plus_one)] @@ -36,19 +38,28 @@ mod range; mod rate; mod resample; mod silk; -mod tables; +pub(super) mod tables; use std::sync::OnceLock; -use bands::{SPREAD_NORMAL, quant_all_bands}; +use bands::{BandCoding, SPREAD_NORMAL, quant_all_bands}; use energy::{quant_coarse_energy, quant_energy_finalise, quant_fine_energy}; use mdct::Mdct; use range::{BITRES, RangeEncoder}; -use rate::{Mode as CeltMode, compute_allocation}; +use rate::{AllocationInput, Mode as CeltMode, compute_allocation}; use resample::Downsampler; use silk::SilkEncoder; use tables::{E_MEANS, EBANDS, NB_EBANDS, OVERLAP, PREEMPH, SPREAD_ICDF, TF_SELECT_TABLE, TRIM_ICDF}; +/// Band range, channel count and transform size shared by CELT coding stages. +#[derive(Clone, Copy)] +struct CeltFrame { + start: usize, + end: usize, + channels: usize, + lm: usize, +} + /// Samples per channel in one 20 ms frame at 48 kHz. const N: usize = 960; /// LM of a 20 ms frame (8 short blocks). @@ -338,7 +349,7 @@ impl Encoder { Mode::Celt => { let (log_e, silence) = self.celt_analysis(pcm); let len = self.frame_bytes; - packet.extend_from_slice(&self.encode_celt(RangeEncoder::new(), len, 0, NB_EBANDS, &log_e, silence)); + packet.extend_from_slice(&self.encode_celt(RangeEncoder::new(), len, 0..NB_EBANDS, &log_e, silence)); } Mode::Silk => { let x = self.silk_input(pcm); @@ -368,7 +379,7 @@ impl Encoder { } let (log_e, _) = self.celt_analysis(pcm); let end = if self.bandwidth == Bandwidth::SuperWideband { 19 } else { NB_EBANDS }; - packet.extend_from_slice(&self.encode_celt(enc, len, HYBRID_START_BAND, end, &log_e, false)); + packet.extend_from_slice(&self.encode_celt(enc, len, HYBRID_START_BAND..end, &log_e, false)); } } Ok(packet) @@ -430,10 +441,18 @@ impl Encoder { /// Codes one CELT frame (RFC 6716 §4.3, in bitstream order) of `len` bytes into `enc` /// (which already holds the SILK frame in hybrid mode), coding bands `start..end`, and /// finalises the frame. - #[allow(clippy::too_many_arguments)] - fn encode_celt(&mut self, mut enc: RangeEncoder, len: usize, start: usize, end: usize, log_e: &[f32; 2 * NB_EBANDS], silence: bool) -> Vec { + fn encode_celt( + &mut self, + mut enc: RangeEncoder, + len: usize, + bands: std::ops::Range, + log_e: &[f32; 2 * NB_EBANDS], + silence: bool, + ) -> Vec { let m = celt_mode(); let c = self.channels; + let frame = CeltFrame { start: bands.start, end: bands.end, channels: c, lm: LM }; + let CeltFrame { start, end, .. } = frame; if c == 1 { for i in 0..NB_EBANDS { self.old_band_e[i] = self.old_band_e[i].max(self.old_band_e[NB_EBANDS + i]); @@ -462,10 +481,10 @@ impl Encoder { let mut enc_intra = enc.clone(); let mut old_intra = self.old_band_e; enc_intra.bit_logp(true, 3); - quant_coarse_energy(start, end, log_e, &mut old_intra, true, &mut enc_intra, c, LM, total_bits_raw); + quant_coarse_energy(frame, log_e, &mut old_intra, true, &mut enc_intra, total_bits_raw); enc.bit_logp(false, 3); let mut old_inter = self.old_band_e; - quant_coarse_energy(start, end, log_e, &mut old_inter, false, &mut enc, c, LM, total_bits_raw); + quant_coarse_energy(frame, log_e, &mut old_inter, false, &mut enc, total_bits_raw); if enc_intra.tell_frac() < enc.tell_frac() { enc = enc_intra; self.old_band_e = old_intra; @@ -473,14 +492,14 @@ impl Encoder { self.old_band_e = old_inter; } } else { - quant_coarse_energy(start, end, log_e, &mut self.old_band_e, false, &mut enc, c, LM, total_bits_raw); + quant_coarse_energy(frame, log_e, &mut self.old_band_e, false, &mut enc, total_bits_raw); } // TF resolution: no changes. { let mut budget = total_bits_raw; let mut tell = enc.tell(); let mut logp = 4; - let tf_select_rsv = LM > 0 && tell + logp + 1 <= budget; + let tf_select_rsv = LM > 0 && tell + logp < budget; budget -= tf_select_rsv as i32; for _ in start..end { if tell + logp <= budget { @@ -512,27 +531,11 @@ impl Encoder { enc.icdf(alloc_trim as usize, &TRIM_ICDF, 7); } let bits = (total_bits_raw << BITRES) - enc.tell_frac() - 1; - let alloc = compute_allocation(m, start, end, &offsets, &cap, alloc_trim, bits, c, LM, &mut enc); - quant_fine_energy(start, end, log_e, &mut self.old_band_e, &alloc.fine_quant, &mut enc, c); - quant_all_bands( - m, - start, - end, - &self.freq, - N, - c, - &alloc.pulses, - SPREAD_NORMAL, - alloc.dual_stereo, - alloc.intensity, - total_bits_raw << BITRES, - alloc.balance, - &mut enc, - LM, - alloc.coded_bands, - ); + let alloc = compute_allocation(m, frame, AllocationInput { offsets: &offsets, cap: &cap, trim: alloc_trim, bits }, &mut enc); + quant_fine_energy(frame, log_e, &mut self.old_band_e, &alloc.fine_quant, &mut enc); + quant_all_bands(m, frame, &self.freq, N, &alloc, BandCoding { spread: SPREAD_NORMAL, total_bits: total_bits_raw << BITRES }, &mut enc); let bits_left = total_bits_raw - enc.tell(); - quant_energy_finalise(start, end, log_e, &mut self.old_band_e, &alloc.fine_quant, &alloc.fine_priority, bits_left, &mut enc, c); + quant_energy_finalise(frame, log_e, &mut self.old_band_e, &alloc, bits_left, &mut enc); debug_assert!(enc.tell() <= total_bits_raw, "frame overflow: {} > {}", enc.tell(), total_bits_raw); if c == 1 { let (a, b) = self.old_band_e.split_at_mut(NB_EBANDS); diff --git a/crates/audio/src/opus/encoder/rate.rs b/crates/audio/src/opus/encoder/rate.rs index c360f9da7..3eb9575d4 100644 --- a/crates/audio/src/opus/encoder/rate.rs +++ b/crates/audio/src/opus/encoder/rate.rs @@ -4,254 +4,18 @@ //! spec-derived `filmcraft-opus` decoder (first-party, MIT OR Apache-2.0) step by step; the only //! difference is that the encoder *writes* the band-skip, intensity and dual-stereo symbols. -use crate::opus::encoder::range::{BITRES, RangeEncoder, ilog}; -use crate::opus::encoder::tables::{ALLOC, EBANDS, LOG2_FRAC_TABLE, MAX_LM, NB_EBANDS}; +use crate::opus::encoder::CeltFrame; +use crate::opus::encoder::range::{BITRES, RangeEncoder}; +use crate::opus::encoder::tables::{EBANDS, LOG2_FRAC_TABLE, NB_EBANDS}; -const MAX_PSEUDO: i32 = 40; -const LOG_MAX_PSEUDO: i32 = 6; +#[cfg(test)] +pub use crate::opus::rate_common::pvq_v; +pub use crate::opus::rate_common::{Mode, get_pulses}; pub const MAX_FINE_BITS: i32 = 8; const FINE_OFFSET: i32 = 21; pub const QTHETA_OFFSET: i32 = 4; -pub const QTHETA_OFFSET_TWOPHASE: i32 = 16; const ALLOC_STEPS: i32 = 6; -/// Number of pulses represented by pseudo-pulse index `i`. -pub fn get_pulses(i: i32) -> i32 { - if i < 8 { i } else { (8 + (i & 7)) << ((i >> 3) - 1) } -} - -/// Conservative (rounded up) log2 of `val` with `frac` fractional bits. -pub fn log2_frac(mut val: u32, mut frac: u32) -> i32 { - let mut l = ilog(val); - if val & val.wrapping_sub(1) != 0 { - if l > 16 { - val = ((val - 1) >> (l - 16)) + 1; - } else { - val <<= 16 - l; - } - let mut l = (l - 1) << frac; - loop { - let b = (val >> 16) as i32; - l += b << frac; - val = (val + b as u32) >> b; - val = (val * val + 0x7FFF) >> 15; - if frac == 0 { - break; - } - frac -= 1; - } - l + (val > 0x8000) as i32 - } else { - l -= 1; - l << frac - } -} - -/// PVQ codebook size V(N, K) (saturating far above 2^32). -pub fn pvq_v(n: usize, k: usize) -> u128 { - if k == 0 { - return 1; - } - if n == 0 { - return 0; - } - let mut row = vec![0u128; k + 1]; - row[0] = 1; - for _ in 0..n { - let mut prev_diag = row[0]; - for kk in 1..=k { - let up = row[kk]; - let v = up.saturating_add(row[kk - 1]).saturating_add(prev_diag); - prev_diag = up; - row[kk] = v.min(u128::MAX >> 8); - } - row[0] = 1; - } - row[k] -} - -fn fits_in32(n: usize, k: i32) -> bool { - pvq_v(n, k as usize) <= u32::MAX as u128 -} - -/// Precomputed per-mode data. -pub struct Mode { - /// log2 of band width in 1/8 bits. - pub log_n: [i32; NB_EBANDS], - cache_index: Vec, - cache_bits: Vec, - caps: Vec, -} - -impl Mode { - pub fn new() -> Mode { - let mut log_n = [0i32; NB_EBANDS]; - for (i, l) in log_n.iter_mut().enumerate() { - *l = log2_frac((EBANDS[i + 1] - EBANDS[i]) as u32, BITRES); - } - let (cache_index, cache_bits) = compute_pulse_cache(MAX_LM as i32); - let mut m = Mode { log_n, cache_index, cache_bits, caps: Vec::new() }; - m.caps = compute_caps(&m); - m - } - - #[inline] - pub fn cache(&self, lm: i32, band: usize) -> &[u8] { - let idx = self.cache_index[((lm + 1) as usize) * NB_EBANDS + band]; - &self.cache_bits[idx.max(0) as usize..] - } - - pub fn bits2pulses(&self, band: usize, lm: i32, bits: i32) -> i32 { - let cache = self.cache(lm, band); - let mut lo = 0i32; - let mut hi = cache[0] as i32; - let bits = bits - 1; - for _ in 0..LOG_MAX_PSEUDO { - let mid = (lo + hi + 1) >> 1; - if cache[mid as usize] as i32 >= bits { - hi = mid; - } else { - lo = mid; - } - } - let lo_bits = if lo == 0 { -1 } else { cache[lo as usize] as i32 }; - if bits - lo_bits <= cache[hi as usize] as i32 - bits { lo } else { hi } - } - - pub fn pulses2bits(&self, band: usize, lm: i32, pulses: i32) -> i32 { - if pulses == 0 { 0 } else { self.cache(lm, band)[pulses as usize] as i32 + 1 } - } - - /// Per-band allocation caps for this frame (`init_caps`). - pub fn init_caps(&self, lm: usize, c: usize) -> [i32; NB_EBANDS] { - let mut cap = [0i32; NB_EBANDS]; - for (i, cp) in cap.iter_mut().enumerate() { - let n = (EBANDS[i + 1] - EBANDS[i]) << lm; - *cp = (self.caps[NB_EBANDS * (2 * lm + c - 1) + i] as i32 + 64) * c as i32 * n >> 2; - } - cap - } -} - -fn compute_pulse_cache(lm_max: i32) -> (Vec, Vec) { - let nb = NB_EBANDS; - let rows = (lm_max + 2) as usize; - let mut index = vec![-1i16; nb * rows]; - let mut entries: Vec<(usize, i32, usize)> = Vec::new(); - let mut curr = 0usize; - for i in 0..rows { - for j in 0..nb { - let n = (((EBANDS[j + 1] - EBANDS[j]) << i) >> 1) as usize; - index[i * nb + j] = -1; - 'search: for k in 0..=i { - for nn in 0..nb { - if k == i && nn >= j { - break; - } - if n == (((EBANDS[nn + 1] - EBANDS[nn]) << k) >> 1) as usize { - index[i * nb + j] = index[k * nb + nn]; - break 'search; - } - } - } - if index[i * nb + j] == -1 && n != 0 { - let mut k = 0; - while k < MAX_PSEUDO && fits_in32(n, get_pulses(k + 1)) { - k += 1; - } - index[i * nb + j] = curr as i16; - entries.push((n, k, curr)); - curr += (k + 1) as usize; - } - } - } - let mut bits = vec![0u8; curr]; - for &(n, k, off) in &entries { - let maxk = get_pulses(k) as usize; - let tmp = required_bits(n, maxk, BITRES); - for j in 1..=k { - bits[off + j as usize] = (tmp[get_pulses(j) as usize] - 1) as u8; - } - bits[off] = k as u8; - } - (index, bits) -} - -fn required_bits(n: usize, maxk: usize, frac: u32) -> Vec { - let mut out = vec![0i32; maxk + 1]; - if n == 1 { - for o in out.iter_mut().skip(1) { - *o = 1 << frac; - } - } else { - for (k, o) in out.iter_mut().enumerate().skip(1) { - *o = log2_frac(pvq_v(n, k) as u32, frac); - } - } - out -} - -fn compute_caps(m: &Mode) -> Vec { - let nb = NB_EBANDS; - let mut caps = vec![0u8; (MAX_LM + 1) * 2 * nb]; - for i in 0..=MAX_LM as i32 { - for c in 1..=2i32 { - for j in 0..nb { - let mut n0 = EBANDS[j + 1] - EBANDS[j]; - - let max_bits = if n0 << i == 1 { - c * (1 + MAX_FINE_BITS) << BITRES - } else { - let mut lm0 = 0; - if n0 > 2 { - n0 >>= 1; - lm0 -= 1; - } else if n0 <= 1 { - lm0 = i.min(1); - n0 <<= lm0; - } - let pcache = m.cache(lm0, j); - let mut mb = pcache[pcache[0] as usize] as i32 + 1; - let mut n = n0; - for k in 0..(i - lm0) { - mb <<= 1; - let offset = ((m.log_n[j] + ((lm0 + k) << BITRES)) >> 1) - QTHETA_OFFSET; - let num = 459 * ((2 * n - 1) * offset + mb); - let den = ((2 * n - 1) << 9) - 459; - let qb = ((num + (den >> 1)) / den).min(57); - mb += qb; - n <<= 1; - } - if c == 2 { - mb <<= 1; - let offset = ((m.log_n[j] + (i << BITRES)) >> 1) - if n == 2 { QTHETA_OFFSET_TWOPHASE } else { QTHETA_OFFSET }; - let ndof = 2 * n - 1 - (n == 2) as i32; - let p = if n == 2 { 512 } else { 487 }; - let num = p * (mb + ndof * offset); - let den = (ndof << 9) - p; - let qb = ((num + (den >> 1)) / den).min(if n == 2 { 64 } else { 61 }); - mb += qb; - } - let ndof = c * n + if c == 2 && n > 2 { 1 } else { 0 }; - let mut offset = ((m.log_n[j] + (i << BITRES)) >> 1) - FINE_OFFSET; - if n == 2 { - offset += (1 << BITRES) >> 2; - } - let num = mb + ndof * offset; - let den = (ndof - 1) << BITRES; - let qb = ((num + (den >> 1)) / den).min(MAX_FINE_BITS); - mb += c * qb << BITRES; - mb - }; - let w = c * ((EBANDS[j + 1] - EBANDS[j]) << i); - let v = (4 * max_bits / w) - 64; - caps[((i * 2 + c - 1) as usize) * nb + j] = v.clamp(0, 255) as u8; - } - } - } - caps -} - /// Result of the allocation process. pub struct Allocation { pub coded_bands: usize, @@ -263,125 +27,43 @@ pub struct Allocation { pub fine_priority: [i32; NB_EBANDS], } -/// `clt_compute_allocation`, encoder side. Stereo frames are always coded as dual (L/R) stereo -/// with no intensity bands, and no band is skipped voluntarily. -#[allow(clippy::too_many_arguments)] -pub fn compute_allocation( - m: &Mode, - start: usize, - end: usize, - offsets: &[i32; NB_EBANDS], - cap: &[i32; NB_EBANDS], - alloc_trim: i32, - total: i32, - c: usize, - lm: usize, - enc: &mut RangeEncoder, -) -> Allocation { - let ci = c as i32; - let lmi = lm as i32; - let mut total = total.max(0); - let mut skip_start = start; - let skip_rsv = if total >= 1 << BITRES { 1 << BITRES } else { 0 }; - total -= skip_rsv; - let mut intensity_rsv = 0; - let mut dual_stereo_rsv = 0; - if c == 2 { - intensity_rsv = LOG2_FRAC_TABLE[end - start]; - if intensity_rsv > total { - intensity_rsv = 0; - } else { - total -= intensity_rsv; - dual_stereo_rsv = if total >= 1 << BITRES { 1 << BITRES } else { 0 }; - total -= dual_stereo_rsv; - } - } - let mut bits1 = [0i32; NB_EBANDS]; - let mut bits2 = [0i32; NB_EBANDS]; - let mut thresh = [0i32; NB_EBANDS]; - let mut trim_offset = [0i32; NB_EBANDS]; - for j in start..end { - let w = EBANDS[j + 1] - EBANDS[j]; - thresh[j] = (ci << BITRES).max((3 * w << lmi << BITRES) >> 4); - trim_offset[j] = ci * w * (alloc_trim - 5 - lmi) * (end as i32 - j as i32 - 1) * (1 << (lmi + BITRES as i32)) >> 6; - if w << lmi == 1 { - trim_offset[j] -= ci << BITRES; - } - } - let nb_alloc = ALLOC.len() as i32; - let mut lo = 1; - let mut hi = nb_alloc - 1; - loop { - let mut done = false; - let mut psum = 0; - let mid = (lo + hi) >> 1; - for j in (start..end).rev() { - let n = EBANDS[j + 1] - EBANDS[j]; - let mut bitsj = ci * n * (ALLOC[mid as usize][j] as i32) << lmi >> 2; - if bitsj > 0 { - bitsj = (bitsj + trim_offset[j]).max(0); - } - bitsj += offsets[j]; - if bitsj >= thresh[j] || done { - done = true; - psum += bitsj.min(cap[j]); - } else if bitsj >= ci << BITRES { - psum += ci << BITRES; - } - } - if psum > total { - hi = mid - 1; - } else { - lo = mid + 1; - } - if lo > hi { - break; - } - } - hi = lo; - lo -= 1; - for j in start..end { - let n = EBANDS[j + 1] - EBANDS[j]; - let mut bits1j = ci * n * (ALLOC[lo as usize][j] as i32) << lmi >> 2; - let mut bits2j = if hi >= nb_alloc { cap[j] } else { ci * n * (ALLOC[hi as usize][j] as i32) << lmi >> 2 }; - if bits1j > 0 { - bits1j = (bits1j + trim_offset[j]).max(0); - } - if bits2j > 0 { - bits2j = (bits2j + trim_offset[j]).max(0); - } - if lo > 0 { - bits1j += offsets[j]; - } - bits2j += offsets[j]; - if offsets[j] > 0 { - skip_start = j; - } - bits2j = (bits2j - bits1j).max(0); - bits1[j] = bits1j; - bits2[j] = bits2j; - } - interp_bits2pulses(m, start, end, skip_start, &bits1, &bits2, &thresh, cap, total, skip_rsv, intensity_rsv, dual_stereo_rsv, c, lm, enc) +/// Per-band limits, boosts and available budget for the allocation search. +pub struct AllocationInput<'a> { + pub offsets: &'a [i32; NB_EBANDS], + pub cap: &'a [i32; NB_EBANDS], + pub trim: i32, + pub bits: i32, } -#[allow(clippy::too_many_arguments)] -fn interp_bits2pulses( - m: &Mode, - start: usize, - end: usize, +/// `clt_compute_allocation`, encoder side. Stereo frames are always coded as dual (L/R) stereo +/// with no intensity bands, and no band is skipped voluntarily. +pub fn compute_allocation(m: &Mode, frame: CeltFrame, input: AllocationInput<'_>, enc: &mut RangeEncoder) -> Allocation { + let CeltFrame { start, end, channels: c, lm } = frame; + let AllocationInput { offsets, cap, trim: alloc_trim, bits: total } = input; + let crate::opus::rate_common::Bracket { skip_start, bits1, bits2, thresh, total, skip_rsv, intensity_rsv, dual_stereo_rsv } = + crate::opus::rate_common::bracket(start, end, offsets, cap, alloc_trim, total, c, lm); + let interpolation = + AllocationInterpolation { skip_start, bits1: &bits1, bits2: &bits2, thresh: &thresh, cap, total, skip_rsv, intensity_rsv, dual_stereo_rsv }; + interp_bits2pulses(m, frame, interpolation, enc) +} + +/// Bracketing allocation vectors and signalling reserves for the interpolation stage. +struct AllocationInterpolation<'a> { skip_start: usize, - bits1: &[i32; NB_EBANDS], - bits2: &[i32; NB_EBANDS], - thresh: &[i32; NB_EBANDS], - cap: &[i32; NB_EBANDS], - mut total: i32, + bits1: &'a [i32; NB_EBANDS], + bits2: &'a [i32; NB_EBANDS], + thresh: &'a [i32; NB_EBANDS], + cap: &'a [i32; NB_EBANDS], + total: i32, skip_rsv: i32, - mut intensity_rsv: i32, - mut dual_stereo_rsv: i32, - c: usize, - lm: usize, - enc: &mut RangeEncoder, -) -> Allocation { + intensity_rsv: i32, + dual_stereo_rsv: i32, +} + +fn interp_bits2pulses(m: &Mode, frame: CeltFrame, interpolation: AllocationInterpolation<'_>, enc: &mut RangeEncoder) -> Allocation { + let CeltFrame { start, end, channels: c, lm } = frame; + let AllocationInterpolation { skip_start, bits1, bits2, thresh, cap, mut total, skip_rsv, mut intensity_rsv, mut dual_stereo_rsv } = + interpolation; let ci = c as i32; let alloc_floor = ci << BITRES; let stereo = (c > 1) as i32; @@ -393,7 +75,7 @@ fn interp_bits2pulses( let mut psum = 0; let mut done = false; for j in (start..end).rev() { - let tmp = bits1[j] + (mid * bits2[j] >> ALLOC_STEPS); + let tmp = bits1[j] + ((mid * bits2[j]) >> ALLOC_STEPS); if tmp >= thresh[j] || done { done = true; psum += tmp.min(cap[j]); @@ -411,7 +93,7 @@ fn interp_bits2pulses( let mut psum = 0; let mut done = false; for j in (start..end).rev() { - let mut tmp = bits1[j] + (lo * bits2[j] >> ALLOC_STEPS); + let mut tmp = bits1[j] + ((lo * bits2[j]) >> ALLOC_STEPS); if tmp < thresh[j] && !done { tmp = if tmp >= alloc_floor { alloc_floor } else { 0 }; } else { @@ -500,9 +182,9 @@ fn interp_bits2pulses( if n == 2 { offset += (den << BITRES) >> 2; } - if bits[j] + offset < den * 2 << BITRES { + if bits[j] + offset < (den * 2) << BITRES { offset += nclogn >> 2; - } else if bits[j] + offset < den * 3 << BITRES { + } else if bits[j] + offset < (den * 3) << BITRES { offset += nclogn >> 3; } let mut eb = (bits[j] + offset + (den << (BITRES - 1))).max(0); @@ -513,7 +195,7 @@ fn interp_bits2pulses( eb = eb.min(MAX_FINE_BITS); ebits[j] = eb; fine_priority[j] = (eb * (den << BITRES) >= bits[j] + offset) as i32; - bits[j] -= ci * eb << BITRES; + bits[j] -= (ci * eb) << BITRES; } else { excess = (bit - (ci << BITRES)).max(0); bits[j] = bit - excess; @@ -523,7 +205,7 @@ fn interp_bits2pulses( if excess > 0 { let extra_fine = (excess >> (stereo + BITRES as i32)).min(MAX_FINE_BITS - ebits[j]); ebits[j] += extra_fine; - let extra_bits = extra_fine * ci << BITRES; + let extra_bits = (extra_fine * ci) << BITRES; fine_priority[j] = (extra_bits >= excess - balance) as i32; excess -= extra_bits; } diff --git a/crates/audio/src/opus/encoder/resample.rs b/crates/audio/src/opus/encoder/resample.rs index 9045cb86d..a72589919 100644 --- a/crates/audio/src/opus/encoder/resample.rs +++ b/crates/audio/src/opus/encoder/resample.rs @@ -45,6 +45,8 @@ fn design(fc: f64, beta: f64, delay: usize) -> Vec { h.iter().map(|&v| v as f32).collect() } +use crate::opus::dot::dot8; + /// Decimates one channel of 48 kHz audio by 3, 4 or 6. #[derive(Clone)] pub struct Downsampler { @@ -77,7 +79,7 @@ impl Downsampler { let t = taps - 1 + self.factor * j; let win = &self.buf[t + 1 - taps..=t]; // h is symmetric, so the window can be taken in forward order. - out.push(win.iter().zip(&self.h).map(|(a, b)| a * b).sum()); + out.push(dot8(win, &self.h)); } let n = self.buf.len(); self.buf.copy_within(n - (taps - 1).., 0); @@ -97,6 +99,16 @@ mod tests { (x.iter().map(|&v| (v as f64).powi(2)).sum::() / x.len() as f64).sqrt() } + #[test] + fn dot8_matches_sequential_sum() { + for len in [0usize, 1, 7, 8, 9, 63, 167, 251] { + let a: Vec = (0..len).map(|i| ((i * 37 % 101) as f32 - 50.0) / 50.0).collect(); + let b: Vec = (0..len).map(|i| ((i * 53 % 89) as f32 - 44.0) / 44.0).collect(); + let exact: f64 = a.iter().zip(&b).map(|(&x, &y)| x as f64 * y as f64).sum(); + assert!((dot8(&a, &b) as f64 - exact).abs() < 1e-4, "len {len}"); + } + } + #[test] fn passes_band_and_rejects_aliases() { for &(fs, pass, stop) in &[(16_000u32, 6_000.0, 8_600.0), (12_000, 4_500.0, 6_500.0), (8_000, 3_000.0, 4_500.0)] { diff --git a/crates/audio/src/opus/encoder/silk/analysis.rs b/crates/audio/src/opus/encoder/silk/analysis.rs index 439331076..d80052046 100644 --- a/crates/audio/src/opus/encoder/silk/analysis.rs +++ b/crates/audio/src/opus/encoder/silk/analysis.rs @@ -69,28 +69,55 @@ pub fn levinson(r: &[f64], order: usize) -> Vec { } /// Evaluates the symmetric polynomial `p` (length `d + 1`, `d` even) as the cosine series -/// `e^{jωd/2} P(e^{jω})`. -fn cos_series(p: &[f64], w: f64) -> f64 { +/// `e^{jωd/2} P(e^{jω})`, with `cos_m(m)` giving `cos(m ω)`. +#[inline] +fn cos_series_with(p: &[f64], cos_m: impl Fn(usize) -> f64) -> f64 { let h = (p.len() - 1) / 2; let mut s = p[h]; for k in 0..h { - s += 2.0 * p[k] * ((h - k) as f64 * w).cos(); + s += 2.0 * p[k] * cos_m(h - k); } s } +fn cos_series(p: &[f64], w: f64) -> f64 { + cos_series_with(p, |m| (m as f64 * w).cos()) +} + +/// Number of grid intervals the root search scans over `[0, Ï€]`. +const GRID: usize = 1024; +/// Largest multiple of ω in the series (half the LPC order, at most 16 / 2). +const MAX_H: usize = 8; +/// Bisection steps per root: the bracket shrinks to 3e-9 rad, far below one Q15 NLSF step (9.6e-5 rad). +const BISECT_STEPS: usize = 20; + +/// `cos(m · Ï€ g / GRID)` for every grid point `g` and `m` in `0..=MAX_H`, computed with the same +/// expression as [`cos_series`], so the table values are bit-identical to calling `cos` there. +fn grid_cos() -> &'static [[f64; MAX_H + 1]] { + static TABLE: std::sync::OnceLock> = std::sync::OnceLock::new(); + TABLE.get_or_init(|| { + (0..=GRID) + .map(|g| { + let w = std::f64::consts::PI * g as f64 / GRID as f64; + std::array::from_fn(|m| (m as f64 * w).cos()) + }) + .collect() + }) +} + fn roots(p: &[f64], out: &mut Vec) { - const GRID: usize = 1024; + let grid = grid_cos(); + let at_grid = |g: usize| cos_series_with(p, |m| grid[g][m]); let mut w0 = 0.0; - let mut f0 = cos_series(p, w0); + let mut f0 = at_grid(0); for g in 1..=GRID { let w1 = std::f64::consts::PI * g as f64 / GRID as f64; - let f1 = cos_series(p, w1); + let f1 = at_grid(g); if f0 == 0.0 { out.push(w0); } else if f0 * f1 < 0.0 { let (mut lo, mut hi, mut flo) = (w0, w1, f0); - for _ in 0..40 { + for _ in 0..BISECT_STEPS { let mid = 0.5 * (lo + hi); let fm = cos_series(p, mid); if fm * flo <= 0.0 { @@ -111,6 +138,9 @@ fn roots(p: &[f64], out: &mut Vec) { /// when the root search fails (unstable or degenerate filter). pub fn a2nlsf(a: &[f64]) -> Option> { let d = a.len(); + if d > 2 * MAX_H { + return None; + } let mut c = vec![0f64; d + 2]; c[0] = 1.0; for k in 0..d { @@ -199,9 +229,20 @@ pub fn pitch_search(x: &[f32], start: usize, len: usize, fs_khz: usize) -> Pitch let score = full[(lag - min_lag) as usize]; // Contour: per-subframe lag offsets around the chosen lag. let cb: &[[i32; 4]] = if fs_khz == 8 { &CB_NB_20 } else { &CB_WB_20 }; + // The contour search revisits the same (subframe, lag) pairs for neighbouring base lags and + // codebook entries, so each correlation is computed once. + let span = (18 * fs - min_lag + 1) as usize; + let memo = std::cell::RefCell::new(vec![f64::NAN; 4 * span]); let sub_score = |k: usize, l: i32| -> f64 { let l = l.clamp(min_lag, 18 * fs); - ncorr(x, start + k * sub..start + (k + 1) * sub, l as usize) + let slot = k * span + (l - min_lag) as usize; + let cached = memo.borrow()[slot]; + if !cached.is_nan() { + return cached; + } + let v = ncorr(x, start + k * sub..start + (k + 1) * sub, l as usize); + memo.borrow_mut()[slot] = v; + v }; let mut best = (f64::MIN, lag, 0usize); for l0 in (lag - 2).max(min_lag)..=(lag + 2).min(max_lag) { @@ -319,4 +360,59 @@ mod tests { } } } + + /// The root search as it was before the grid table: `cos` called for every evaluation. + fn roots_reference(p: &[f64], out: &mut Vec) { + let mut w0 = 0.0; + let mut f0 = cos_series(p, w0); + for g in 1..=GRID { + let w1 = std::f64::consts::PI * g as f64 / GRID as f64; + let f1 = cos_series(p, w1); + if f0 == 0.0 { + out.push(w0); + } else if f0 * f1 < 0.0 { + let (mut lo, mut hi, mut flo) = (w0, w1, f0); + for _ in 0..BISECT_STEPS { + let mid = 0.5 * (lo + hi); + let fm = cos_series(p, mid); + if fm * flo <= 0.0 { + hi = mid; + } else { + lo = mid; + flo = fm; + } + } + out.push(0.5 * (lo + hi)); + } + w0 = w1; + f0 = f1; + } + } + + #[test] + fn grid_table_roots_are_bit_identical_to_direct_cos() { + let mut seed = 7u64; + let mut rnd = move || { + seed ^= seed << 13; + seed ^= seed >> 7; + seed ^= seed << 17; + (seed >> 11) as f64 / (1u64 << 53) as f64 * 2.0 - 1.0 + }; + let mut found = 0; + for len in [11usize, 17] { + for _ in 0..300 { + let mut p = vec![0f64; len]; + for k in 0..len / 2 + 1 { + p[k] = rnd(); + p[len - 1 - k] = p[k]; + } + let (mut a, mut b) = (Vec::new(), Vec::new()); + roots(&p, &mut a); + roots_reference(&p, &mut b); + assert_eq!(a.iter().map(|v| v.to_bits()).collect::>(), b.iter().map(|v| v.to_bits()).collect::>()); + found += a.len(); + } + } + assert!(found > 1000, "{found}"); + } } diff --git a/crates/audio/src/opus/encoder/silk/mod.rs b/crates/audio/src/opus/encoder/silk/mod.rs index d82e537e0..0fda4f59a 100644 --- a/crates/audio/src/opus/encoder/silk/mod.rs +++ b/crates/audio/src/opus/encoder/silk/mod.rs @@ -129,6 +129,48 @@ struct Coded { pulses: Vec, } +/// Noise-feedback history (decoded − input), most recent first, as a double-length ring: each +/// sample is stored at `pos` and `pos + MAX_LPC_ORDER`, so `buf[pos..pos + MAX_LPC_ORDER]` is +/// always the contiguous window and a push costs two stores instead of a shift. +struct NfRing { + buf: [f64; 2 * MAX_LPC_ORDER], + /// Index of the most recent sample, always below `MAX_LPC_ORDER`. + pos: usize, +} + +impl NfRing { + fn new(hist: &[f64; MAX_LPC_ORDER]) -> Self { + let mut buf = [0.0; 2 * MAX_LPC_ORDER]; + buf[..MAX_LPC_ORDER].copy_from_slice(hist); + buf[MAX_LPC_ORDER..].copy_from_slice(hist); + NfRing { buf, pos: 0 } + } + + /// The history, most recent first. + fn window(&self) -> &[f64] { + self.buf.get(self.pos..self.pos + MAX_LPC_ORDER).unwrap_or(&[]) + } + + /// Makes `v` the most recent sample and drops the oldest. + fn push(&mut self, v: f64) { + self.pos = self.pos.checked_sub(1).unwrap_or(MAX_LPC_ORDER - 1); + for i in [self.pos, self.pos + MAX_LPC_ORDER] { + if let Some(slot) = self.buf.get_mut(i) { + *slot = v; + } + } + } + + /// The history as a plain array, most recent first. + fn to_array(&self) -> [f64; MAX_LPC_ORDER] { + let mut out = [0.0; MAX_LPC_ORDER]; + for (o, v) in out.iter_mut().zip(self.window()) { + *o = *v; + } + out + } +} + #[derive(Clone)] struct Channel { fs_khz: usize, @@ -237,7 +279,7 @@ impl Channel { let sub = self.sub; let order = self.order; let st = &mut self.dec; - let nf = &mut self.nf; + let mut ring = NfRing::new(&self.nf); let off = if p.signal_type == 0 { offset + INACTIVE_OFFSET } else { offset }; let gamma = nf_gamma(off); let mut nf_c = [0f64; MAX_LPC_ORDER]; @@ -355,7 +397,7 @@ impl Channel { } let lpc_q14 = lpc_pred.saturating_mul(16); // Target excitation (Q14) for this sample. - let fb: f64 = (0..order).map(|j| nf_c[j] * nf[j]).sum(); + let fb: f64 = nf_c.iter().zip(ring.window()).take(order).map(|(c, e)| c * e).sum(); let tgt = (x[n] as f64 + fb).clamp(-32768.0, 32767.0) * inv_g - lpc_q14 as f64 - ltp_q14 as f64; let sgn = if seed_v < 0 { -1.0 } else { 1.0 }; let u = sgn * tgt; @@ -394,11 +436,11 @@ impl Channel { let v = res.saturating_add(lpc_q14); s_lpc[MAX_LPC_ORDER + i] = v; xq[n] = sat16(rshift_round(smulww(v, gain_q10), 8)) as i16; - nf.copy_within(0..MAX_LPC_ORDER - 1, 1); - nf[0] = xq[n] as f64 - x[n] as f64; + ring.push(xq[n] as f64 - x[n] as f64); } s_lpc.copy_within(sub..sub + MAX_LPC_ORDER, 0); } + self.nf = ring.to_array(); st.s_lpc_q14.copy_from_slice(&s_lpc[..MAX_LPC_ORDER]); st.out_buf[..l].copy_from_slice(&xq); (Coded { gain_ix, gains_q16, seed, pulses }, xq) @@ -485,6 +527,14 @@ struct Prep { pred_q13: [i32; 2], } +/// One rate-loop trial: the frame coded at gain offset `offset`. +struct Trial { + offset: f64, + bits: i32, + state: SilkEncoder, + coder: RangeEncoder, +} + /// SILK encoder for one mono or stereo (mid/side) stream. #[derive(Clone)] pub struct SilkEncoder { @@ -635,11 +685,20 @@ impl SilkEncoder { pub fn encode(&mut self, input: &[Vec], target_bits: f64, max_bits: i32, enc: &mut RangeEncoder) { let prep = self.prepare(input); let start_bits = enc.tell(); - let trial = |s: &Self, o: f64| -> i32 { - let mut t = s.clone(); - let mut e = enc.clone(); - t.code(&prep, o, &mut e); - e.tell() - start_bits + // Every trial codes the frame into a clone of the encoder and the range coder. The result + // is kept per offset: the hard-limit check below hits the memo for the chosen offset, and + // the final step adopts that trial instead of coding the frame once more. + let mut trials: Vec = Vec::new(); + let mut trial = |s: &Self, o: f64| -> i32 { + if let Some(t) = trials.iter().find(|t| t.offset == o) { + return t.bits; + } + let mut state = s.clone(); + let mut coder = enc.clone(); + state.code(&prep, o, &mut coder); + let bits = coder.tell() - start_bits; + trials.push(Trial { offset: o, bits, state, coder }); + bits }; // Bisection on the gain offset: the smallest offset (finest quantisation) within budget. let target = target_bits.min(max_bits as f64); @@ -701,8 +760,20 @@ impl SilkEncoder { while trial(self, o) > max_bits && o < 12.0 { o += 1.0; } + match trials.iter().position(|t| t.offset == o) { + Some(i) => { + let t = trials.swap_remove(i); + *self = t.state; + *enc = t.coder; + } + None => self.code(&prep, o, enc), + } + self.end_frame(&prep, o); + } + + /// Advances the input history and the frame counter after frame `prep` was coded at `o`. + fn end_frame(&mut self, prep: &Prep, o: f64) { self.last_offset = o; - self.code(&prep, o, enc); for ch in 0..self.channels { self.ch[ch].push_history(&prep.x[ch]); } @@ -712,3 +783,80 @@ impl SilkEncoder { self.frame += 1; } } + +#[cfg(test)] +mod tests { + use super::*; + + /// Deterministic speech-like test input: two harmonics with a moving pitch, a syllable + /// envelope, noise and short silences, at the internal rate and 16-bit scale. + fn signal(len: usize, fs_khz: usize, seed: u32) -> Vec { + let mut s = seed; + let fs = (fs_khz * 1000) as f64; + (0..len) + .map(|i| { + s = s.wrapping_mul(1_664_525).wrapping_add(1_013_904_223); + let noise = (s >> 8) as f64 / (1u32 << 24) as f64 - 0.5; + let t = i as f64 / fs; + let f0 = 140.0 + 40.0 * (t * 1.3).sin(); + let env = ((t * 4.0).sin().max(0.0)).powf(0.7); + let voiced = (std::f64::consts::TAU * f0 * t).sin() + 0.5 * (std::f64::consts::TAU * 2.0 * f0 * t).sin(); + (8000.0 * env * voiced + 600.0 * noise) as f32 + }) + .collect() + } + + #[test] + fn nf_ring_matches_a_shifted_history() { + let mut hist = [0f64; MAX_LPC_ORDER]; + for (i, h) in hist.iter_mut().enumerate() { + *h = i as f64 * 0.5 - 3.0; + } + let mut ring = NfRing::new(&hist); + assert_eq!(ring.window(), &hist[..]); + for n in 0..100 { + let v = (n as f64 * 0.37).sin() * 100.0; + hist.copy_within(0..MAX_LPC_ORDER - 1, 1); + hist[0] = v; + ring.push(v); + assert_eq!(ring.window(), &hist[..], "after push {n}"); + } + assert_eq!(ring.to_array(), hist); + } + + /// The rate loop adopts the memoised trial for the chosen offset. A second encoder that codes + /// every frame afresh at that offset (what the loop did before the memo) must produce the + /// same bytes, frame after frame, so the adopted encoder and range-coder states are exact. + #[test] + fn adopting_the_memoised_trial_equals_coding_the_frame_again() { + for (fs_khz, channels, target_bits, max_bits) in + [(8, 1, 200.0, 1000), (12, 1, 320.0, 1000), (16, 1, 480.0, 1000), (16, 2, 640.0, 1000), (16, 1, 900.0, 400), (8, 2, 120.0, 160)] + { + let mut memo = SilkEncoder::new(fs_khz, channels); + let mut fresh = memo.clone(); + let l = memo.frame_len(); + let frames = 40; + let left = signal(l * frames, fs_khz, 7); + let right: Vec = signal(l * frames + 37, fs_khz, 11)[37..].iter().map(|v| v * 0.6).collect(); + for f in 0..frames { + let mut input = vec![left[f * l..(f + 1) * l].to_vec()]; + if channels == 2 { + input.push(right[f * l..(f + 1) * l].to_vec()); + } + let mut enc = RangeEncoder::new(); + memo.encode(&input, target_bits, max_bits, &mut enc); + let o = memo.last_offset; + let prep = fresh.prepare(&input); + let mut enc_fresh = RangeEncoder::new(); + fresh.code(&prep, o, &mut enc_fresh); + fresh.end_frame(&prep, o); + assert_eq!(enc.tell(), enc_fresh.tell(), "{fs_khz} kHz {channels} ch frame {f}"); + assert_eq!(enc.done(1275), enc_fresh.done(1275), "{fs_khz} kHz {channels} ch frame {f}"); + for c in 0..channels { + assert_eq!(memo.ch[c].nf, fresh.ch[c].nf, "noise feedback, channel {c}, frame {f}"); + assert_eq!(memo.ch[c].dec.s_lpc_q14, fresh.ch[c].dec.s_lpc_q14, "LPC state, channel {c}, frame {f}"); + } + } + } + } +} diff --git a/crates/audio/src/opus/fft.rs b/crates/audio/src/opus/fft.rs new file mode 100644 index 000000000..300c78d48 --- /dev/null +++ b/crates/audio/src/opus/fft.rs @@ -0,0 +1,130 @@ +//! Shared CELT forward complex FFT. + +#[derive(Clone, Copy, Default, Debug)] +pub struct Cpx { + pub re: f32, + pub im: f32, +} + +impl Cpx { + #[inline] + pub(crate) fn mul(self, o: Cpx) -> Cpx { + Cpx { re: self.re * o.re - self.im * o.im, im: self.re * o.im + self.im * o.re } + } + #[inline] + pub(crate) fn add(self, o: Cpx) -> Cpx { + Cpx { re: self.re + o.re, im: self.im + o.im } + } + #[inline] + pub(crate) fn sub(self, o: Cpx) -> Cpx { + Cpx { re: self.re - o.re, im: self.im - o.im } + } +} + +/// Mixed-radix (2, 3, 4, 5) forward complex FFT, unscaled (decimation in time, recursive). +pub struct Fft { + pub(crate) n: usize, + pub(crate) factors: Vec, + pub(crate) twiddle: Vec, +} + +impl Fft { + pub fn new(n: usize) -> Fft { + let mut factors = Vec::new(); + let mut m = n; + for p in [4usize, 2, 3, 5] { + while m.is_multiple_of(p) { + factors.push(p); + m /= p; + } + } + assert_eq!(m, 1, "unsupported FFT size {n}"); + let twiddle = (0..n) + .map(|k| { + let ph = -2.0 * std::f64::consts::PI * k as f64 / n as f64; + Cpx { re: ph.cos() as f32, im: ph.sin() as f32 } + }) + .collect(); + Fft { n, factors, twiddle } + } + + pub fn process(&self, data: &mut [Cpx], scratch: &mut Vec) { + scratch.clear(); + scratch.extend_from_slice(&data[..self.n]); + self.rec(&mut data[..self.n], scratch, 1, self.n, 0, 1); + } + + fn rec(&self, out: &mut [Cpx], inp: &[Cpx], in_stride: usize, n: usize, fi: usize, tw_stride: usize) { + let p = self.factors[fi]; + let m = n / p; + if m == 1 { + for (i, o) in out.iter_mut().enumerate().take(p) { + *o = inp[i * in_stride]; + } + } else { + for q in 0..p { + self.rec(&mut out[q * m..(q + 1) * m], &inp[q * in_stride..], in_stride * p, m, fi + 1, tw_stride * p); + } + } + // The twiddle index q * k * tw_stride is always below n, so no modulo is needed. + let tw = &self.twiddle; + match p { + 2 => { + for k in 0..m { + let a0 = out[k]; + let a1 = out[m + k].mul(tw[k * tw_stride]); + out[k] = a0.add(a1); + out[m + k] = a0.sub(a1); + } + } + 4 => { + for k in 0..m { + let t0 = out[k]; + let t1 = out[m + k].mul(tw[k * tw_stride]); + let t2 = out[2 * m + k].mul(tw[2 * k * tw_stride]); + let t3 = out[3 * m + k].mul(tw[3 * k * tw_stride]); + let (s0, d0) = (t0.add(t2), t0.sub(t2)); + let (s1, d1) = (t1.add(t3), t1.sub(t3)); + out[k] = s0.add(s1); + out[2 * m + k] = s0.sub(s1); + // The odd outputs combine d0 with -j * d1 = (d1.im, -d1.re) and its negative. + out[m + k] = Cpx { re: d0.re + d1.im, im: d0.im - d1.re }; + out[3 * m + k] = Cpx { re: d0.re - d1.im, im: d0.im + d1.re }; + } + } + 3 => { + // Imaginary part of exp(-2 pi i / 3) = -sqrt(3) / 2. + let h = tw[self.n / 3].im; + for k in 0..m { + let t0 = out[k]; + let t1 = out[m + k].mul(tw[k * tw_stride]); + let t2 = out[2 * m + k].mul(tw[2 * k * tw_stride]); + let (s, d) = (t1.add(t2), t1.sub(t2)); + let a = Cpx { re: t0.re - 0.5 * s.re, im: t0.im - 0.5 * s.im }; + out[k] = t0.add(s); + // j * h * d = (-h * d.im, h * d.re) + out[m + k] = Cpx { re: a.re - h * d.im, im: a.im + h * d.re }; + out[2 * m + k] = Cpx { re: a.re + h * d.im, im: a.im - h * d.re }; + } + } + _ => { + // Generic DFT butterfly (radix 5, once per transform). + let roots: [Cpx; 5] = std::array::from_fn(|q| if q < p { tw[q * (self.n / p)] } else { Cpx::default() }); + let mut t = [Cpx::default(); 5]; + for k in 0..m { + t[0] = out[k]; + for q in 1..p { + t[q] = out[q * m + k].mul(tw[q * k * tw_stride]); + } + for q2 in 0..p { + let mut acc = t[0]; + for q in 1..p { + acc = acc.add(t[q].mul(roots[(q * q2) % p])); + } + out[q2 * m + k] = acc; + } + } + } + } + } +} diff --git a/crates/audio/src/opus/mod.rs b/crates/audio/src/opus/mod.rs index b67e1ef3d..84e81ec8c 100644 --- a/crates/audio/src/opus/mod.rs +++ b/crates/audio/src/opus/mod.rs @@ -7,14 +7,18 @@ //! 20 ms, mono or stereo). From EffectCraft's `effectcraft-opusenc` (storytold/effectcraft@c618429). //! //! Public items keep their names, so FilmCraft and EffectCraft can re-export this module. The two -//! halves come from separate crates and do not share code yet (each has its own range coder, MDCT -//! and band tables). [`Bandwidth`] is one type for both (`encoder::Bandwidth` is the same type); +//! halves share the FFT, resampling dot product and CELT allocation preparation; range coders and +//! mode-specific MDCT processing remain separate. [`Bandwidth`] is one type for both (`encoder::Bandwidth` is the same type); //! the two `Mode` enums keep their apps' variant names and convert into each other. // Index loops and long argument lists follow the specification's pseudo-code (FilmCraft and // EffectCraft allow these lints workspace-wide). #![allow(clippy::needless_range_loop, clippy::too_many_arguments)] +mod dot; +mod fft; +mod rate_common; + #[cfg(feature = "opus-decode")] mod decode; #[cfg(feature = "opus-encode")] diff --git a/crates/audio/src/opus/rate_common.rs b/crates/audio/src/opus/rate_common.rs new file mode 100644 index 000000000..86b27c90a --- /dev/null +++ b/crates/audio/src/opus/rate_common.rs @@ -0,0 +1,362 @@ +//! CELT pulse cache and allocation caps shared by encoder and decoder. + +#[cfg(not(feature = "opus-encode"))] +use super::decode::celt::tables::{ALLOC, EBANDS, LOG2_FRAC_TABLE, MAX_LM, NB_EBANDS}; +#[cfg(feature = "opus-encode")] +use super::encoder::tables::{ALLOC, EBANDS, LOG2_FRAC_TABLE, MAX_LM, NB_EBANDS}; + +const BITRES: u32 = 3; +const MAX_PSEUDO: i32 = 40; +const LOG_MAX_PSEUDO: i32 = 6; +const MAX_FINE_BITS: i32 = 8; +const FINE_OFFSET: i32 = 21; +const QTHETA_OFFSET: i32 = 4; +const QTHETA_OFFSET_TWOPHASE: i32 = 16; +fn ilog(x: u32) -> i32 { + (32 - x.leading_zeros()) as i32 +} + +/// Number of pulses represented by pseudo-pulse index `i`. +pub fn get_pulses(i: i32) -> i32 { + if i < 8 { i } else { (8 + (i & 7)) << ((i >> 3) - 1) } +} + +/// Conservative (rounded up) log2 of `val` with `frac` fractional bits. +pub fn log2_frac(mut val: u32, mut frac: u32) -> i32 { + let mut l = ilog(val); + if val & val.wrapping_sub(1) != 0 { + if l > 16 { + val = ((val - 1) >> (l - 16)) + 1; + } else { + val <<= 16 - l; + } + let mut l = (l - 1) << frac; + loop { + let b = (val >> 16) as i32; + l += b << frac; + val = (val + b as u32) >> b; + val = (val * val + 0x7FFF) >> 15; + if frac == 0 { + break; + } + frac -= 1; + } + l + (val > 0x8000) as i32 + } else { + l -= 1; + l << frac + } +} + +/// PVQ codebook size V(N, K) (saturating far above 2^32). +pub fn pvq_v(n: usize, k: usize) -> u128 { + if k == 0 { + return 1; + } + if n == 0 { + return 0; + } + let mut row = vec![0u128; k + 1]; + row[0] = 1; + for _ in 0..n { + let mut prev_diag = row[0]; + for kk in 1..=k { + let up = row[kk]; + let v = up.saturating_add(row[kk - 1]).saturating_add(prev_diag); + prev_diag = up; + row[kk] = v.min(u128::MAX >> 8); + } + row[0] = 1; + } + row[k] +} + +fn fits_in32(n: usize, k: i32) -> bool { + pvq_v(n, k as usize) <= u32::MAX as u128 +} + +/// Precomputed per-mode data. +pub struct Mode { + /// log2 of band width in 1/8 bits. + pub log_n: [i32; NB_EBANDS], + pub(crate) cache_index: Vec, + pub(crate) cache_bits: Vec, + pub(crate) caps: Vec, +} + +impl Mode { + pub fn new() -> Mode { + let mut log_n = [0i32; NB_EBANDS]; + for (i, l) in log_n.iter_mut().enumerate() { + *l = log2_frac((EBANDS[i + 1] - EBANDS[i]) as u32, BITRES); + } + let (cache_index, cache_bits) = compute_pulse_cache(MAX_LM as i32); + let mut m = Mode { log_n, cache_index, cache_bits, caps: Vec::new() }; + m.caps = compute_caps(&m); + m + } + + #[inline] + pub fn cache(&self, lm: i32, band: usize) -> &[u8] { + let idx = self.cache_index[((lm + 1) as usize) * NB_EBANDS + band]; + &self.cache_bits[idx.max(0) as usize..] + } + + pub fn bits2pulses(&self, band: usize, lm: i32, bits: i32) -> i32 { + let cache = self.cache(lm, band); + let mut lo = 0i32; + let mut hi = cache[0] as i32; + let bits = bits - 1; + for _ in 0..LOG_MAX_PSEUDO { + let mid = (lo + hi + 1) >> 1; + if cache[mid as usize] as i32 >= bits { + hi = mid; + } else { + lo = mid; + } + } + let lo_bits = if lo == 0 { -1 } else { cache[lo as usize] as i32 }; + if bits - lo_bits <= cache[hi as usize] as i32 - bits { lo } else { hi } + } + + pub fn pulses2bits(&self, band: usize, lm: i32, pulses: i32) -> i32 { + if pulses == 0 { 0 } else { self.cache(lm, band)[pulses as usize] as i32 + 1 } + } + + /// Per-band allocation caps for this frame (`init_caps`). + pub fn init_caps(&self, lm: usize, c: usize) -> [i32; NB_EBANDS] { + let mut cap = [0i32; NB_EBANDS]; + for (i, cp) in cap.iter_mut().enumerate() { + let n = (EBANDS[i + 1] - EBANDS[i]) << lm; + *cp = ((self.caps[NB_EBANDS * (2 * lm + c - 1) + i] as i32 + 64) * c as i32 * n) >> 2; + } + cap + } +} + +fn compute_pulse_cache(lm_max: i32) -> (Vec, Vec) { + let nb = NB_EBANDS; + let rows = (lm_max + 2) as usize; + let mut index = vec![-1i16; nb * rows]; + let mut entries: Vec<(usize, i32, usize)> = Vec::new(); + let mut curr = 0usize; + for i in 0..rows { + for j in 0..nb { + let n = (((EBANDS[j + 1] - EBANDS[j]) << i) >> 1) as usize; + index[i * nb + j] = -1; + 'search: for k in 0..=i { + for nn in 0..nb { + if k == i && nn >= j { + break; + } + if n == (((EBANDS[nn + 1] - EBANDS[nn]) << k) >> 1) as usize { + index[i * nb + j] = index[k * nb + nn]; + break 'search; + } + } + } + if index[i * nb + j] == -1 && n != 0 { + let mut k = 0; + while k < MAX_PSEUDO && fits_in32(n, get_pulses(k + 1)) { + k += 1; + } + index[i * nb + j] = curr as i16; + entries.push((n, k, curr)); + curr += (k + 1) as usize; + } + } + } + let mut bits = vec![0u8; curr]; + for &(n, k, off) in &entries { + let maxk = get_pulses(k) as usize; + let tmp = required_bits(n, maxk, BITRES); + for j in 1..=k { + bits[off + j as usize] = (tmp[get_pulses(j) as usize] - 1) as u8; + } + bits[off] = k as u8; + } + (index, bits) +} + +fn required_bits(n: usize, maxk: usize, frac: u32) -> Vec { + let mut out = vec![0i32; maxk + 1]; + if n == 1 { + for o in out.iter_mut().skip(1) { + *o = 1 << frac; + } + } else { + for (k, o) in out.iter_mut().enumerate().skip(1) { + *o = log2_frac(pvq_v(n, k) as u32, frac); + } + } + out +} + +fn compute_caps(m: &Mode) -> Vec { + let nb = NB_EBANDS; + let mut caps = vec![0u8; (MAX_LM + 1) * 2 * nb]; + for i in 0..=MAX_LM as i32 { + for c in 1..=2i32 { + for j in 0..nb { + let mut n0 = EBANDS[j + 1] - EBANDS[j]; + + let max_bits = if n0 << i == 1 { + (c * (1 + MAX_FINE_BITS)) << BITRES + } else { + let mut lm0 = 0; + if n0 > 2 { + n0 >>= 1; + lm0 -= 1; + } else if n0 <= 1 { + lm0 = i.min(1); + n0 <<= lm0; + } + let pcache = m.cache(lm0, j); + let mut mb = pcache[pcache[0] as usize] as i32 + 1; + let mut n = n0; + for k in 0..(i - lm0) { + mb <<= 1; + let offset = ((m.log_n[j] + ((lm0 + k) << BITRES)) >> 1) - QTHETA_OFFSET; + let num = 459 * ((2 * n - 1) * offset + mb); + let den = ((2 * n - 1) << 9) - 459; + let qb = ((num + (den >> 1)) / den).min(57); + mb += qb; + n <<= 1; + } + if c == 2 { + mb <<= 1; + let offset = ((m.log_n[j] + (i << BITRES)) >> 1) - if n == 2 { QTHETA_OFFSET_TWOPHASE } else { QTHETA_OFFSET }; + let ndof = 2 * n - 1 - (n == 2) as i32; + let p = if n == 2 { 512 } else { 487 }; + let num = p * (mb + ndof * offset); + let den = (ndof << 9) - p; + let qb = ((num + (den >> 1)) / den).min(if n == 2 { 64 } else { 61 }); + mb += qb; + } + let ndof = c * n + if c == 2 && n > 2 { 1 } else { 0 }; + let mut offset = ((m.log_n[j] + (i << BITRES)) >> 1) - FINE_OFFSET; + if n == 2 { + offset += (1 << BITRES) >> 2; + } + let num = mb + ndof * offset; + let den = (ndof - 1) << BITRES; + let qb = ((num + (den >> 1)) / den).min(MAX_FINE_BITS); + mb += (c * qb) << BITRES; + mb + }; + let w = c * ((EBANDS[j + 1] - EBANDS[j]) << i); + let v = (4 * max_bits / w) - 64; + caps[((i * 2 + c - 1) as usize) * nb + j] = v.clamp(0, 255) as u8; + } + } + } + caps +} + +/// Shared allocation search before encoder/decoder signalling diverges. +pub struct Bracket { + pub skip_start: usize, + pub bits1: [i32; NB_EBANDS], + pub bits2: [i32; NB_EBANDS], + pub thresh: [i32; NB_EBANDS], + pub total: i32, + pub skip_rsv: i32, + pub intensity_rsv: i32, + pub dual_stereo_rsv: i32, +} + +pub fn bracket( + start: usize, + end: usize, + offsets: &[i32; NB_EBANDS], + cap: &[i32; NB_EBANDS], + alloc_trim: i32, + total: i32, + c: usize, + lm: usize, +) -> Bracket { + let ci = c as i32; + let lmi = lm as i32; + let mut total = total.max(0); + let mut skip_start = start; + let skip_rsv = if total >= 1 << BITRES { 1 << BITRES } else { 0 }; + total -= skip_rsv; + let mut intensity_rsv = 0; + let mut dual_stereo_rsv = 0; + if c == 2 { + intensity_rsv = LOG2_FRAC_TABLE[end - start]; + if intensity_rsv > total { + intensity_rsv = 0; + } else { + total -= intensity_rsv; + dual_stereo_rsv = if total >= 1 << BITRES { 1 << BITRES } else { 0 }; + total -= dual_stereo_rsv; + } + } + let mut bits1 = [0i32; NB_EBANDS]; + let mut bits2 = [0i32; NB_EBANDS]; + let mut thresh = [0i32; NB_EBANDS]; + let mut trim_offset = [0i32; NB_EBANDS]; + for j in start..end { + let w = EBANDS[j + 1] - EBANDS[j]; + thresh[j] = (ci << BITRES).max(((3 * w) << lmi << BITRES) >> 4); + trim_offset[j] = (ci * w * (alloc_trim - 5 - lmi) * (end as i32 - j as i32 - 1) * (1 << (lmi + BITRES as i32))) >> 6; + if w << lmi == 1 { + trim_offset[j] -= ci << BITRES; + } + } + let nb_alloc = ALLOC.len() as i32; + let mut lo = 1; + let mut hi = nb_alloc - 1; + loop { + let mut done = false; + let mut psum = 0; + let mid = (lo + hi) >> 1; + for j in (start..end).rev() { + let n = EBANDS[j + 1] - EBANDS[j]; + let mut bitsj = (ci * n * (ALLOC[mid as usize][j] as i32)) << lmi >> 2; + if bitsj > 0 { + bitsj = (bitsj + trim_offset[j]).max(0); + } + bitsj += offsets[j]; + if bitsj >= thresh[j] || done { + done = true; + psum += bitsj.min(cap[j]); + } else if bitsj >= ci << BITRES { + psum += ci << BITRES; + } + } + if psum > total { + hi = mid - 1; + } else { + lo = mid + 1; + } + if lo > hi { + break; + } + } + hi = lo; + lo -= 1; + for j in start..end { + let n = EBANDS[j + 1] - EBANDS[j]; + let mut bits1j = (ci * n * (ALLOC[lo as usize][j] as i32)) << lmi >> 2; + let mut bits2j = if hi >= nb_alloc { cap[j] } else { (ci * n * (ALLOC[hi as usize][j] as i32)) << lmi >> 2 }; + if bits1j > 0 { + bits1j = (bits1j + trim_offset[j]).max(0); + } + if bits2j > 0 { + bits2j = (bits2j + trim_offset[j]).max(0); + } + if lo > 0 { + bits1j += offsets[j]; + } + bits2j += offsets[j]; + if offsets[j] > 0 { + skip_start = j; + } + bits2j = (bits2j - bits1j).max(0); + bits1[j] = bits1j; + bits2[j] = bits2j; + } + Bracket { skip_start, bits1, bits2, thresh, total, skip_rsv, intensity_rsv, dual_stereo_rsv } +} diff --git a/crates/audio/src/pcm_interleaved.rs b/crates/audio/src/pcm_interleaved.rs new file mode 100644 index 000000000..f08c19202 --- /dev/null +++ b/crates/audio/src/pcm_interleaved.rs @@ -0,0 +1,96 @@ +//! Interleaved PCM packet decoding into caller-owned planar channels. + +/// The sample representation in an interleaved PCM packet. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub struct Format { + /// Bits per sample. + pub bits: u16, + /// IEEE floating-point samples. + pub float: bool, + /// Big-endian sample bytes. + pub big_endian: bool, + /// Signed 8-bit integers (wider integer PCM is signed). + pub signed: bool, +} + +/// Append complete interleaved frames to caller-owned planar channels. Empty channel slices and +/// zero-bit formats append nothing. Incomplete frames are ignored; unsupported widths produce +/// silent samples, matching the container adapters that use this API. +pub fn decode_into(data: &[u8], cfg: Format, out: &mut [Vec]) { + let ch = out.len(); + if cfg.bits == 0 || ch == 0 { + return; + } + let bps = usize::from(cfg.bits).div_ceil(8); + let Some(frame_bytes) = bps.checked_mul(ch).filter(|&n| n > 0) else { return }; + let n = data.len() / frame_bytes; + for channel in out.iter_mut() { + channel.reserve(n); + } + // `n` is whole frames within `data`, so every sample slice below is in bounds. + for i in 0..n { + for (c, o) in out.iter_mut().enumerate() { + let off = (i * ch + c) * bps; + let b = &data[off..off + bps]; + let v = match (cfg.bits, cfg.float, cfg.big_endian) { + (32, true, false) => f32::from_le_bytes([b[0], b[1], b[2], b[3]]), + (32, true, true) => f32::from_be_bytes([b[0], b[1], b[2], b[3]]), + (64, true, false) => f64::from_le_bytes(b.try_into().unwrap_or([0; 8])) as f32, + (64, true, true) => f64::from_be_bytes(b.try_into().unwrap_or([0; 8])) as f32, + (8, _, _) => { + if cfg.signed { + b[0] as i8 as f32 / 128.0 + } else { + (b[0] as f32 - 128.0) / 128.0 + } + } + (16, _, false) => i16::from_le_bytes([b[0], b[1]]) as f32 / 32768.0, + (16, _, true) => i16::from_be_bytes([b[0], b[1]]) as f32 / 32768.0, + (24, _, false) => (i32::from_le_bytes([0, b[0], b[1], b[2]]) >> 8) as f32 / 8_388_608.0, + (24, _, true) => (i32::from_be_bytes([b[0], b[1], b[2], 0]) >> 8) as f32 / 8_388_608.0, + (32, false, false) => i32::from_le_bytes([b[0], b[1], b[2], b[3]]) as f32 / 2_147_483_648.0, + (32, false, true) => i32::from_be_bytes([b[0], b[1], b[2], b[3]]) as f32 / 2_147_483_648.0, + _ => 0.0, + }; + o.push(v); + } + } +} + +#[cfg(test)] +mod tests { + use super::*; + #[test] + fn decodes_endian_and_interleaving_without_touching_partial_frames() { + let mut channels = vec![Vec::new(), Vec::new()]; + let fmt = Format { bits: 16, float: false, big_endian: true, signed: true }; + decode_into(&[0x40, 0, 0xc0, 0, 0xff], fmt, &mut channels); + assert_eq!(channels, [vec![0.5], vec![-0.5]]); + decode_into(&[0; 8], Format { bits: 0, ..fmt }, &mut channels); + assert_eq!(channels, [vec![0.5], vec![-0.5]]); + } + #[test] + fn decodes_sample_representations_and_ignores_invalid_inputs() { + let base = Format { bits: 8, float: false, big_endian: false, signed: false }; + let cases: &[(Format, &[u8], f32)] = &[ + (base, &[192], 0.5), + (Format { signed: true, ..base }, &[192], -0.5), + (Format { bits: 16, ..base }, &[0, 64], 0.5), + (Format { bits: 24, ..base }, &[0, 0, 64], 0.5), + (Format { bits: 24, big_endian: true, ..base }, &[0xc0, 0, 0], -0.5), + (Format { bits: 32, ..base }, &[0, 0, 0, 64], 0.5), + (Format { bits: 32, big_endian: true, ..base }, &[0xc0, 0, 0, 0], -0.5), + (Format { bits: 32, float: true, ..base }, &0.5_f32.to_le_bytes(), 0.5), + (Format { bits: 32, float: true, big_endian: true, ..base }, &(-0.5_f32).to_be_bytes(), -0.5), + (Format { bits: 64, float: true, ..base }, &0.5_f64.to_le_bytes(), 0.5), + (Format { bits: 64, float: true, big_endian: true, ..base }, &(-0.5_f64).to_be_bytes(), -0.5), + (Format { bits: 12, ..base }, &[0xff, 0xff], 0.0), + ]; + for &(format, bytes, expected) in cases { + let mut out = [vec![1.0]]; + decode_into(bytes, format, &mut out); + assert_eq!(out[0], [1.0, expected], "{format:?}"); + } + decode_into(&[1], base, &mut []); + } +} diff --git a/crates/audio/src/vorbis/encoder/book.rs b/crates/audio/src/vorbis/encoder/book.rs new file mode 100644 index 000000000..b392cf27c --- /dev/null +++ b/crates/audio/src/vorbis/encoder/book.rs @@ -0,0 +1,287 @@ +//! Bit packing (Vorbis I §2) and codebooks (§3): Huffman lengths, the codeword assignment the +//! decoder derives from them, VQ lattices (lookup type 1) and the header encoding. + +use super::{Error, Result}; + +/// LSB-first bit packer (Vorbis I §2.1.4): bits fill each byte from its least significant end. +#[derive(Clone, Debug, Default)] +pub struct BitWriter { + buf: Vec, + bits: u64, +} + +impl BitWriter { + pub fn new() -> BitWriter { + BitWriter::default() + } + + /// Writes the low `n` bits of `v` (`n` ≤ 32). + pub fn write(&mut self, v: u32, n: u32) { + let n = n.min(32); + for i in 0..n { + let bit = (v >> i) & 1; + let at = (self.bits / 8) as usize; + if at == self.buf.len() { + self.buf.push(0); + } + if bit != 0 + && let Some(b) = self.buf.get_mut(at) + { + *b |= 1 << (self.bits % 8); + } + self.bits += 1; + } + } + + pub fn flag(&mut self, b: bool) { + self.write(b as u32, 1); + } + + /// Writes a codeword MSB first: the decoder walks the tree from its first bit. + fn codeword(&mut self, word: u32, len: u32) { + for i in (0..len).rev() { + self.write((word >> i) & 1, 1); + } + } + + pub fn bytes(&mut self, data: &[u8]) { + for &b in data { + self.write(u32::from(b), 8); + } + } + + pub fn finish(self) -> Vec { + self.buf + } +} + +/// `ilog` of the specification (§9.2.1): the position of the highest set bit, 0 for 0. +pub fn ilog(v: u32) -> u32 { + 32 - v.leading_zeros() +} + +/// The 32-bit float format of codebook headers (§9.2.2), for values it represents exactly. +pub fn float32_pack(v: f64) -> u32 { + if v == 0.0 || !v.is_finite() { + return 0; + } + let sign = if v < 0.0 { 0x8000_0000u32 } else { 0 }; + let mut m = v.abs(); + let mut exp: i32 = 0; + // a 21-bit mantissa: scale into [2^20, 2^21) + while m >= (1 << 21) as f64 { + m /= 2.0; + exp += 1; + } + while m < (1 << 20) as f64 { + m *= 2.0; + exp -= 1; + } + let mant = (m.round() as u32).min((1 << 21) - 1); + let biased = (exp + 788).clamp(0, 1023) as u32; + sign | (biased << 21) | mant +} + +/// Huffman code lengths (1..=`max_len`) for `weights`, every entry coded. +pub fn huffman_lengths(weights: &[f64], max_len: u32) -> Vec { + let n = weights.len(); + if n <= 1 { + return vec![1; n]; + } + let total: f64 = weights.iter().map(|w| if w.is_finite() { w.max(0.0) } else { 0.0 }).sum::().max(1e-300); + let mut floor = 1e-9; + loop { + let w: Vec = weights.iter().map(|&x| (if x.is_finite() { x.max(0.0) } else { 0.0 }) / total + floor).collect(); + let lens = huffman(&w); + if lens.iter().all(|&l| l <= max_len) || floor > 1.0 { + return lens.into_iter().map(|l| l.clamp(1, max_len)).collect(); + } + floor *= 4.0; + } +} + +fn huffman(w: &[f64]) -> Vec { + // nodes: (weight, children) with leaves first + let n = w.len(); + let mut parent = vec![usize::MAX; 2 * n]; + let mut weight: Vec = w.to_vec(); + let mut alive: Vec = (0..n).collect(); + while alive.len() > 1 { + alive.sort_by(|&a, &b| weight[b].partial_cmp(&weight[a]).unwrap_or(std::cmp::Ordering::Equal).then(b.cmp(&a))); + let (Some(a), Some(b)) = (alive.pop(), alive.pop()) else { break }; + let id = weight.len(); + weight.push(weight[a] + weight[b]); + if parent.len() <= id { + parent.resize(id + 1, usize::MAX); + } + parent[a] = id; + parent[b] = id; + alive.push(id); + } + (0..n) + .map(|i| { + let mut d = 0; + let mut x = i; + while let Some(&p) = parent.get(x) { + if p == usize::MAX { + break; + } + d += 1; + x = p; + } + d + }) + .collect() +} + +/// The codewords the decoder assigns to `lengths` (§3.2.1): in entry order, each entry gets the +/// lowest available codeword of its length (no prefix of an assigned codeword, and no assigned +/// codeword a prefix of it). An error when the lengths do not fit a prefix code. +pub fn assign_codewords(lengths: &[u32]) -> Result> { + // A trie of the assigned codewords: `children[node]` (0 = none), `leaf[node]`. + let mut children: Vec<[usize; 2]> = vec![[0, 0]]; + let mut leaf = vec![false]; + let mut words = Vec::with_capacity(lengths.len()); + for &len in lengths { + if len == 0 || len > 32 { + return Err(Error::Internal("codeword length")); + } + let bits = lowest_free(&children, &leaf, 0, len).ok_or(Error::Internal("over-subscribed codebook"))?; + let mut node = 0usize; + let mut word = 0u32; + for &bit in &bits { + word = (word << 1) | bit as u32; + node = match children[node][bit] { + 0 => { + children.push([0, 0]); + leaf.push(false); + let id = children.len() - 1; + children[node][bit] = id; + id + } + c => c, + }; + } + leaf[node] = true; + words.push(word); + } + Ok(words) +} + +/// The leftmost free path of `depth` bits below `node` (depth-first, 0 before 1). +fn lowest_free(children: &[[usize; 2]], leaf: &[bool], node: usize, depth: u32) -> Option> { + if leaf[node] || depth == 0 { + return None; + } + for bit in 0..2 { + let c = children[node][bit]; + if c == 0 { + let mut p = vec![bit]; + p.resize(depth as usize, 0); + return Some(p); + } + if let Some(mut rest) = lowest_free(children, leaf, c, depth - 1) { + rest.insert(0, bit); + return Some(rest); + } + } + None +} + +/// A codebook: entry lengths and codewords, and for VQ books a lattice (lookup type 1). +#[derive(Clone, Debug)] +pub struct Codebook { + pub dimensions: u32, + pub lengths: Vec, + pub words: Vec, + /// Lattice values per dimension (lookup type 1: `values^dimensions == entries`). + pub lattice: Option, +} + +#[derive(Clone, Copy, Debug)] +pub struct Lattice { + pub minimum: f64, + pub delta: f64, + /// Values per dimension. + pub values: u32, +} + +impl Codebook { + /// A scalar book (no lookup) of `weights.len()` entries. + pub fn scalar(dimensions: u32, weights: &[f64]) -> Result { + let lengths = huffman_lengths(weights, 20); + let words = assign_codewords(&lengths)?; + Ok(Codebook { dimensions, lengths, words, lattice: None }) + } + + /// A lattice VQ book: `values` integers per dimension from `minimum` in steps of `delta`; + /// `weight(v)` is the relative probability of one component `v`. + pub fn lattice(dimensions: u32, values: u32, minimum: i32, delta: i32, weight: impl Fn(i32) -> f64) -> Result { + let entries = values.checked_pow(dimensions).ok_or(Error::Internal("codebook size"))?; + let w: Vec = (0..entries) + .map(|e| { + let mut p = 1.0; + let mut rest = e; + for _ in 0..dimensions { + let t = (rest % values) as i32; + rest /= values; + p *= weight(minimum + t * delta); + } + p + }) + .collect(); + let lengths = huffman_lengths(&w, 20); + let words = assign_codewords(&lengths)?; + Ok(Codebook { dimensions, lengths, words, lattice: Some(Lattice { minimum: f64::from(minimum), delta: f64::from(delta), values }) }) + } + + pub fn entries(&self) -> usize { + self.lengths.len() + } + + /// Writes entry `e`. + pub fn put(&self, w: &mut BitWriter, e: usize) { + if let (Some(&word), Some(&len)) = (self.words.get(e), self.lengths.get(e)) { + w.codeword(word, len); + } + } + + /// The entry of the lattice vector `v` (component 0 is the least significant digit, §3.2.1). + pub fn entry_of(&self, v: &[i32]) -> usize { + let Some(l) = self.lattice else { return 0 }; + let mut e = 0usize; + let mut scale = 1usize; + for &x in v.iter().take(self.dimensions as usize) { + let t = ((f64::from(x) - l.minimum) / l.delta).round().clamp(0.0, f64::from(l.values - 1)) as usize; + e += t * scale; + scale *= l.values as usize; + } + e + } + + /// The header encoding (§3.2.1). + pub fn write_header(&self, w: &mut BitWriter) { + w.write(0x56_4342, 24); + w.write(self.dimensions, 16); + w.write(self.entries() as u32, 24); + w.flag(false); // not ordered + w.flag(false); // not sparse + for &l in &self.lengths { + w.write(l - 1, 5); + } + match self.lattice { + None => w.write(0, 4), + Some(l) => { + w.write(1, 4); + w.write(float32_pack(l.minimum), 32); + w.write(float32_pack(l.delta), 32); + let bits = ilog(l.values - 1).max(1); + w.write(bits - 1, 4); + w.flag(false); // sequence_p + for m in 0..l.values { + w.write(m, bits); + } + } + } + } +} diff --git a/crates/audio/src/vorbis/encoder/floor.rs b/crates/audio/src/vorbis/encoder/floor.rs new file mode 100644 index 000000000..503385bd3 --- /dev/null +++ b/crates/audio/src/vorbis/encoder/floor.rs @@ -0,0 +1,207 @@ +//! Floor type 1 (Vorbis I §7): a piecewise-linear spectral envelope in the decoder's dB scale. +//! +//! The encoder picks a target value at each of the floor's X positions, codes it against the +//! decoder's prediction from its neighbours (step 2 of §7.2.4, inverted), and renders the curve +//! exactly as the decoder will (`render_line`, integer arithmetic), so the residue is computed +//! against the very floor the decoder multiplies it by. + +use super::book::{BitWriter, Codebook, ilog}; + +/// `floor1_multiplier`: 2, so values run 0..128 and step two entries of the dB table. +pub const MULTIPLIER: u32 = 2; +/// The value range for [`MULTIPLIER`] 2. +pub const RANGE: i32 = 128; +const RANGE_BITS: u32 = 10; +/// Points per partition, and partitions. +pub const CLASS_DIM: usize = 4; +pub const PARTITIONS: usize = 8; + +/// Step (in dB) between entries of `floor1_inverse_dB_table`: it runs from 1.0649863e-07 to 1 +/// over 256 entries. +fn db_step() -> f64 { + (1.0f64 / 1.064_986_3e-7).ln() / 255.0 +} + +/// `floor1_inverse_dB_table[i]`. +pub fn inverse_db(i: i32) -> f32 { + (f64::from(i.clamp(0, 255) - 255) * db_step()).exp() as f32 +} + +/// The table index whose amplitude is nearest (in dB) to `a`. +pub fn db_index(a: f32) -> i32 { + if !a.is_finite() || a <= 0.0 { + return 0; + } + ((f64::from(a)).ln() / db_step() + 255.0).round().clamp(0.0, 255.0) as i32 +} + +/// A floor 1 configuration for blocks of `n/2 = 2^RANGE_BITS` bins. +#[derive(Clone, Debug)] +pub struct Floor1 { + /// X positions: 0, 1024, then the partitions' points. + pub xs: Vec, + /// Index of the book coding the point values. + pub book: usize, + /// For each point (in coding order) its low and high neighbours. + neighbours: Vec<(usize, usize)>, + /// Point indices sorted by X. + sorted: Vec, +} + +impl Floor1 { + /// Points spaced about evenly in log frequency over 1..1024. + pub fn new(book: usize) -> Floor1 { + let count = CLASS_DIM * PARTITIONS; + let mut xs = vec![0, 1 << RANGE_BITS]; + let mut last = 0; + for i in 1..=count { + // geometric from 2 to ~1000 bins, at least one bin apart + let x = (2.0f64 * (500.0f64).powf((i - 1) as f64 / (count - 1) as f64)).round() as i32; + let x = x.max(last + 1).min((1 << RANGE_BITS) - 1); + xs.push(x); + last = x; + } + let neighbours = (0..xs.len()) + .map(|i| { + if i < 2 { + return (0, 0); + } + let (mut lo, mut hi) = (0usize, 1usize); + for j in 0..i { + if xs[j] < xs[i] && xs[j] > xs[lo] { + lo = j; + } + if xs[j] > xs[i] && xs[j] < xs[hi] { + hi = j; + } + } + (lo, hi) + }) + .collect(); + let mut sorted: Vec = (0..xs.len()).collect(); + sorted.sort_by_key(|&i| xs[i]); + Floor1 { xs, book, neighbours, sorted } + } + + /// The setup-header encoding (§7.2.2): one class of `CLASS_DIM` points, no subclasses. + pub fn write_header(&self, w: &mut BitWriter) { + w.write(PARTITIONS as u32, 5); + for _ in 0..PARTITIONS { + w.write(0, 4); + } + w.write(CLASS_DIM as u32 - 1, 3); + w.write(0, 2); // no subclasses + w.write(self.book as u32 + 1, 8); + w.write(MULTIPLIER - 1, 2); + w.write(RANGE_BITS, 4); + for &x in &self.xs[2..] { + w.write(x as u32, RANGE_BITS); + } + } + + /// Codes the target values `want` (one per point, 0..RANGE) into `w`; returns the floor curve + /// the decoder renders (`n2` amplitudes). + pub fn encode(&self, w: &mut BitWriter, book: &Codebook, want: &[i32], n2: usize) -> Vec { + let n = self.xs.len(); + let want = |i: usize| want.get(i).copied().unwrap_or(0).clamp(0, RANGE - 1); + let mut fin = vec![0i32; n]; + let mut used = vec![false; n]; + let mut coded = vec![0u32; n]; + fin[0] = want(0); + fin[1] = want(1); + used[0] = true; + used[1] = true; + for i in 2..n { + let (lo, hi) = self.neighbours[i]; + let pred = render_point(self.xs[lo], fin[lo], self.xs[hi], fin[hi], self.xs[i]); + let f = want(i); + let (highroom, lowroom) = (RANGE - pred, pred); + let room = 2 * highroom.min(lowroom); + let val = if f == pred { + 0 + } else { + let d = f - pred; + let inside = if d > 0 { 2 * d } else { -2 * d - 1 }; + if inside < room { + inside + } else if highroom > lowroom { + f - pred + lowroom + } else { + pred - f + highroom - 1 + } + }; + coded[i] = val.clamp(0, RANGE - 1) as u32; + if val != 0 { + used[lo] = true; + used[hi] = true; + used[i] = true; + fin[i] = f; + } else { + fin[i] = pred; + } + } + let bits = ilog(RANGE as u32 - 1); + w.write(fin[0] as u32, bits); + w.write(fin[1] as u32, bits); + for &c in &coded[2..] { + book.put(w, c as usize); + } + self.render(&fin, &used, n2) + } + + /// The decoder's curve (§7.2.4 step 2, `render_line`) as amplitudes. + fn render(&self, fin: &[i32], used: &[bool], n2: usize) -> Vec { + let mut v = vec![0i32; n2]; + let m = MULTIPLIER as i32; + let (mut lx, mut ly) = (0i32, fin[self.sorted[0]] * m); + let (mut hx, mut hy) = (0i32, ly); + for &i in &self.sorted[1..] { + if used[i] { + hy = fin[i] * m; + hx = self.xs[i]; + render_line(lx, ly, hx, hy, &mut v); + lx = hx; + ly = hy; + } + } + if (hx as usize) < n2 { + render_line(hx, hy, n2 as i32, hy, &mut v); + } + v.iter().map(|&i| inverse_db(i)).collect() + } +} + +fn render_point(x0: i32, y0: i32, x1: i32, y1: i32, x: i32) -> i32 { + let dy = y1 - y0; + let adx = (x1 - x0).max(1); + let err = dy.abs() * (x - x0); + let off = err / adx; + if dy < 0 { y0 - off } else { y0 + off } +} + +fn render_line(x0: i32, y0: i32, x1: i32, y1: i32, v: &mut [i32]) { + let dy = y1 - y0; + let adx = x1 - x0; + if adx <= 0 { + return; + } + let base = dy / adx; + let sy = if dy < 0 { base - 1 } else { base + 1 }; + let ady = dy.abs() - base.abs() * adx; + let (mut y, mut err) = (y0, 0); + if let Some(s) = v.get_mut(x0 as usize) { + *s = y; + } + for x in x0 + 1..x1 { + err += ady; + if err >= adx { + err -= adx; + y += sy; + } else { + y += base; + } + if let Some(s) = v.get_mut(x as usize) { + *s = y; + } + } +} diff --git a/crates/audio/src/vorbis/encoder/mdct.rs b/crates/audio/src/vorbis/encoder/mdct.rs new file mode 100644 index 000000000..3802a24d6 --- /dev/null +++ b/crates/audio/src/vorbis/encoder/mdct.rs @@ -0,0 +1,134 @@ +//! Forward MDCT of a windowed block of `n` samples to `n/2` coefficients: +//! `X[k] = Σ x[j] cos(2Ï€/n · (j + 1/2 + n/4)(k + 1/2))`, the transform the Vorbis decoder inverts +//! (Vorbis I §1.3.2, §4.3.9). Computed as a DCT-IV of the folded block through an `n/8`-point +//! complex FFT. + +use std::f64::consts::PI; + +#[derive(Clone, Copy, Default, Debug)] +struct Cpx { + re: f64, + im: f64, +} + +impl Cpx { + fn mul(self, o: Cpx) -> Cpx { + Cpx { re: self.re * o.re - self.im * o.im, im: self.re * o.im + self.im * o.re } + } +} + +/// A forward MDCT for one block size (a power of two, at least 16). +#[derive(Clone, Debug)] +pub struct Mdct { + n: usize, + twiddle: Vec, + post: Vec, + fft_tw: Vec, + rev: Vec, +} + +impl Mdct { + pub fn new(n: usize) -> Mdct { + let m = n / 2; // DCT-IV length + let h = m / 2; // FFT length + let twiddle = (0..h).map(|j| cis(-PI * (4.0 * j as f64 + 1.0) / (4.0 * m as f64))).collect(); + let post = (0..h).map(|k| cis(-PI * k as f64 / m as f64)).collect(); + let fft_tw = (0..h / 2).map(|j| cis(-2.0 * PI * j as f64 / h as f64)).collect(); + let bits = h.trailing_zeros(); + let rev = (0..h).map(|i| if bits == 0 { 0 } else { i.reverse_bits() >> (usize::BITS - bits) }).collect(); + Mdct { n, twiddle, post, fft_tw, rev } + } + + /// `x`: `n` windowed samples; `out`: `n/2` coefficients (scaled by `scale`). + pub fn forward(&self, x: &[f32], out: &mut [f32], scale: f64) { + let n = self.n; + let m = n / 2; + let q = n / 4; + let at = |i: usize| f64::from(x.get(i).copied().unwrap_or(0.0)); + // fold (a, b, c, d) to the DCT-IV input (-c_r - d, a - b_r) + let mut v = vec![0f64; m]; + for (i, s) in v.iter_mut().enumerate() { + *s = if i < q { -at(3 * q - 1 - i) - at(3 * q + i) } else { at(i - q) - at(3 * q - 1 - i) }; + } + // DCT-IV through a half-length complex FFT + let h = m / 2; + let mut z: Vec = (0..h).map(|j| Cpx { re: v[2 * j], im: v[m - 1 - 2 * j] }.mul(self.twiddle[j])).collect(); + self.fft(&mut z); + for (k, c) in z.iter().enumerate() { + let d = c.mul(self.post[k]); + if let Some(o) = out.get_mut(2 * k) { + *o = (d.re * scale) as f32; + } + if let Some(o) = out.get_mut(m - 1 - 2 * k) { + *o = (-d.im * scale) as f32; + } + } + } + + /// In-place radix-2 FFT (`e^{-2Ï€i jk/h}`). + fn fft(&self, a: &mut [Cpx]) { + let h = a.len(); + for i in 0..h { + let j = self.rev[i]; + if i < j { + a.swap(i, j); + } + } + let mut len = 2; + while len <= h { + let step = h / len; + for start in (0..h).step_by(len) { + for k in 0..len / 2 { + let w = self.fft_tw[k * step]; + let u = a[start + k]; + let t = a[start + k + len / 2].mul(w); + a[start + k] = Cpx { re: u.re + t.re, im: u.im + t.im }; + a[start + k + len / 2] = Cpx { re: u.re - t.re, im: u.im - t.im }; + } + } + len *= 2; + } + } +} + +fn cis(t: f64) -> Cpx { + Cpx { re: t.cos(), im: t.sin() } +} + +/// The Vorbis window (§4.3.1) for a block of `n` between blocks of the same size. +pub fn window(n: usize) -> Vec { + (0..n) + .map(|i| { + let s = (PI * (i as f64 + 0.5) / n as f64).sin(); + (PI / 2.0 * s * s).sin() as f32 + }) + .collect() +} + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn matches_the_direct_formula() { + for n in [16usize, 64, 256, 2048] { + let x: Vec = (0..n).map(|i| ((i * 7919 % 211) as f32 - 105.0) / 50.0).collect(); + let mut fast = vec![0f32; n / 2]; + Mdct::new(n).forward(&x, &mut fast, 1.0); + for k in [0usize, 1, 2, n / 4, n / 2 - 1] { + let direct: f64 = + (0..n).map(|j| f64::from(x[j]) * (2.0 * PI / n as f64 * (j as f64 + 0.5 + n as f64 / 4.0) * (k as f64 + 0.5)).cos()).sum(); + assert!((direct - f64::from(fast[k])).abs() < 1e-3 * (1.0 + direct.abs()), "n={n} k={k}: {direct} vs {}", fast[k]); + } + } + } + + #[test] + fn the_window_is_power_complementary() { + let w = window(2048); + for i in 0..1024 { + let s = w[i] * w[i] + w[i + 1024] * w[i + 1024]; + assert!((s - 1.0).abs() < 1e-5, "{i}: {s}"); + } + } +} diff --git a/crates/audio/src/vorbis/encoder/mod.rs b/crates/audio/src/vorbis/encoder/mod.rs new file mode 100644 index 000000000..95be88244 --- /dev/null +++ b/crates/audio/src/vorbis/encoder/mod.rs @@ -0,0 +1,494 @@ +//! Pure-Rust Vorbis I encoder, written from the Vorbis I specification (Xiph.Org, 2020-07-04). +//! +//! No `unsafe` and no C code; builds for `wasm32-unknown-unknown`. ffmpeg is used only as an +//! external test oracle. +//! +//! The stream: one mode of 2048-sample blocks; per channel a floor 1 envelope of 34 points; one +//! residue type 2 over the interleaved channels, partitions of 32 values (a multiple of the +//! channel count for 3, 5, 6 and 7 channels) in seven classes (silent, +//! three single-pass lattice books, and two- and three-pass cascades for large values); no channel +//! coupling. Quality (-1 to 10) sets how far the floor sits from the spectral envelope (below it +//! for tonal regions, above it for noise-like ones, from the spectral flatness), and with it the +//! residue's precision: variable bit rate, like other Vorbis encoders. +//! +//! From FilmCraft's `filmcraft-vorbis` by @JordyUSA (storytold/filmcraft#4481). +//! +//! The packets go into Ogg with [`craft_media::ogg::PageWriter`] and a comment header from +//! [`craft_media::ogg::vorbis_comment`]; [`crate::encode_ogg_vorbis`] does all of it for a +//! whole buffer. +//! +//! ``` +//! use craft_audio::vorbis::encoder::{Encoder, EncoderConfig}; +//! +//! let mut enc = Encoder::new(EncoderConfig::new(48_000, 2, 5.0))?; +//! let tone: Vec = (0..4_800).map(|i| (i as f32 * 0.05).sin() * 0.5).collect(); +//! let mut packets = enc.encode(&[&tone, &tone])?; +//! packets.extend(enc.finish()?); +//! // the last granule position is the input's length, so a decoder trims the padding +//! assert_eq!(packets.last().map(|p| p.1), Some(4_800)); +//! assert_eq!(&enc.identification()[1..7], b"vorbis"); +//! # Ok::<(), craft_audio::vorbis::encoder::Error>(()) +//! ``` + +// Index loops follow the specification, as in FilmCraft (which allows this lint workspace-wide). +#![allow(clippy::needless_range_loop)] + +mod book; +mod floor; +mod mdct; + +use book::{BitWriter, Codebook}; +use floor::Floor1; +use mdct::Mdct; + +/// Why an encoder could not be created or input could not be encoded. +#[non_exhaustive] +#[derive(Clone, Debug, PartialEq, Eq, thiserror::Error)] +pub enum Error { + /// A setting or the input is outside what the encoder takes. + #[error("invalid argument: {0}")] + InvalidArgument(&'static str), + /// An encoder invariant failed (a bug, reported instead of panicking). + #[error("internal error: {0}")] + Internal(&'static str), +} + +/// The encoder's result type. +pub type Result = std::result::Result; + +/// Block size (samples); the short block size of the header is never used. +pub const BLOCK: usize = 2048; +const HOP: usize = BLOCK / 2; +/// Residue values per partition for `channels` channels: 32, or for 3, 5, 6 and 7 channels a +/// multiple of the channel count (48, 40, 48, 56). libvorbis decodes each partition of an +/// interleaved (type 2) residue as if it started on channel 0, so a partition must hold whole +/// sample frames; the size also divides `channels × 1024` and is a multiple of every book's +/// dimension (4). +fn partition_size(channels: usize) -> usize { + match channels { + 3 | 6 => 48, + 5 => 40, + 7 => 56, + _ => 32, + } +} +const CLASSES: usize = 8; +/// Largest residue magnitude the cascade codes: 7·225 + 7·15 + 7. +const MAX_RESIDUE: i32 = 1687; + +/// The quality a NaN or infinite setting reads as. +pub const DEFAULT_QUALITY: f32 = 6.0; + +/// Encoder settings. +#[derive(Clone, Copy, Debug, PartialEq)] +pub struct EncoderConfig { + /// Samples per second, 1 to 768 000. + pub sample_rate: u32, + /// 1 to 8, in Vorbis channel order (§4.3.9). + pub channels: usize, + /// -1 (smallest) to 10 (best); clamped. + pub quality: f32, +} + +impl EncoderConfig { + /// Settings for `channels` channels (in Vorbis order) at `sample_rate` and `quality`. + pub fn new(sample_rate: u32, channels: usize, quality: f32) -> EncoderConfig { + EncoderConfig { sample_rate, channels, quality } + } +} + +/// Book indices in the setup header. +const BOOK_CLASS: usize = 0; +const BOOK_FLOOR: usize = 1; +const BOOK_A1_SPARSE: usize = 2; +const BOOK_A1: usize = 3; +const BOOK_A2: usize = 4; +const BOOK_A4: usize = 5; +const BOOK_F: usize = 6; +const BOOK_M: usize = 7; +const BOOK_C: usize = 8; + +/// Books per class and pass (`None`: nothing coded in that pass). +const CASCADE: [[Option; 3]; CLASSES] = [ + [None, None, None], + [Some(BOOK_A1_SPARSE), None, None], + [Some(BOOK_A1), None, None], + [Some(BOOK_A2), None, None], + [Some(BOOK_A4), None, None], + [Some(BOOK_F), None, None], + [Some(BOOK_M), Some(BOOK_F), None], + [Some(BOOK_C), Some(BOOK_M), Some(BOOK_F)], +]; + +/// A Vorbis encoder for one stream: planar `f32` in, Vorbis packets with their granule +/// positions out. +pub struct Encoder { + cfg: EncoderConfig, + books: Vec, + floor: Floor1, + mdct: Mdct, + window: Vec, + /// Input not yet past a block, per channel; the next block starts at `start`. + pending: Vec>, + /// Read cursor into `pending`: consumed hops are dropped once per call, not per block. + start: usize, + samples_in: u64, + blocks: u64, + finished: bool, + /// Floor offset below the envelope (dB) for tonal and for noise-like regions. + tonal_db: f32, + noise_db: f32, + /// Residue magnitudes below this quantise to 0. + dead_zone: f32, +} + +impl Encoder { + /// An encoder for `cfg`: 1 to 8 channels, a sample rate of 1 to 768 000; the quality is + /// clamped to -1..=10, and NaN or infinite reads as [`DEFAULT_QUALITY`]. + pub fn new(cfg: EncoderConfig) -> Result { + if !(1..=8).contains(&cfg.channels) { + return Err(Error::InvalidArgument("Vorbis encoder: 1 to 8 channels")); + } + if !(1..=768_000).contains(&cfg.sample_rate) { + return Err(Error::InvalidArgument("Vorbis encoder: sample rate")); + } + let quality = if cfg.quality.is_finite() { cfg.quality.clamp(-1.0, 10.0) } else { DEFAULT_QUALITY }; + let cfg = EncoderConfig { quality, ..cfg }; + let books = vec![ + Codebook::scalar(2, &(0..CLASSES * CLASSES).map(|e| class_weight(e / CLASSES) * class_weight(e % CLASSES)).collect::>())?, + Codebook::scalar(1, &(0..floor::RANGE).map(|v| 1.0 / (1.0 + f64::from(v)).powf(1.6)).collect::>())?, + Codebook::lattice(4, 3, -1, 1, |v| if v == 0 { 0.92 } else { 0.04 })?, + Codebook::lattice(4, 3, -1, 1, |v| if v == 0 { 0.5 } else { 0.25 })?, + Codebook::lattice(2, 5, -2, 1, |v| 0.6f64.powi(v.abs()))?, + Codebook::lattice(2, 9, -4, 1, |v| 0.7f64.powi(v.abs()))?, + Codebook::lattice(2, 15, -7, 1, |v| 0.8f64.powi(v.abs()))?, + Codebook::lattice(2, 15, -105, 15, |v| 0.55f64.powi((v / 15).abs()))?, + Codebook::lattice(2, 15, -1575, 225, |v| 0.4f64.powi((v / 225).abs()))?, + ]; + Ok(Encoder { + cfg, + books, + floor: Floor1::new(BOOK_FLOOR), + mdct: Mdct::new(BLOCK), + window: mdct::window(BLOCK), + // the first block is centred on sample 0: the decoder's first output is input sample 0 + pending: vec![vec![0.0; HOP]; cfg.channels], + start: 0, + samples_in: 0, + blocks: 0, + finished: false, + // tones 4 dB (-1) to 30.4 dB (10) under their envelope; noise 14 dB above it (-1) + // to 4.1 dB (10) + tonal_db: 4.0 + 2.4 * (quality + 1.0), + noise_db: -14.0 + 0.9 * (quality + 1.0), + dead_zone: 0.8 - 0.025 * (quality + 1.0), + }) + } + + /// Channels per sample frame. + pub fn channels(&self) -> usize { + self.cfg.channels + } + + /// The quality in use, after clamping. + pub fn quality(&self) -> f32 { + self.cfg.quality + } + + /// Input samples per channel taken so far. + pub fn samples_in(&self) -> u64 { + self.samples_in + } + + /// The identification header packet (§4.2.2): the first packet of the stream. + pub fn identification(&self) -> Vec { + let mut w = BitWriter::new(); + w.write(1, 8); + w.bytes(b"vorbis"); + w.write(0, 32); + w.write(self.cfg.channels as u32, 8); + w.write(self.cfg.sample_rate, 32); + w.write(0, 32); // bit rates: unset (variable) + w.write(0, 32); + w.write(0, 32); + w.write(8, 4); // 256-sample short blocks (unused) + w.write(11, 4); // 2048-sample long blocks + w.flag(true); + w.finish() + } + + /// The setup header packet (§4.2.4): codebooks, floor, residue, mapping and mode. It follows + /// the comment header. + pub fn setup(&self) -> Vec { + let mut w = BitWriter::new(); + w.write(5, 8); + w.bytes(b"vorbis"); + w.write(self.books.len() as u32 - 1, 8); + for b in &self.books { + b.write_header(&mut w); + } + w.write(0, 6); // one time-domain transform placeholder + w.write(0, 16); + w.write(0, 6); // one floor + w.write(1, 16); + self.floor.write_header(&mut w); + w.write(0, 6); // one residue, type 2 + w.write(2, 16); + let end = (self.cfg.channels * HOP) as u32; + w.write(0, 24); + w.write(end, 24); + w.write(partition_size(self.cfg.channels) as u32 - 1, 24); + w.write(CLASSES as u32 - 1, 6); + w.write(BOOK_CLASS as u32, 8); + for c in &CASCADE { + let bits: u32 = c.iter().enumerate().map(|(j, b)| u32::from(b.is_some()) << j).sum(); + w.write(bits & 7, 3); + w.flag(bits > 7); + if bits > 7 { + w.write(bits >> 3, 5); + } + } + for c in &CASCADE { + for b in c.iter().flatten() { + w.write(*b as u32, 8); + } + } + w.write(0, 6); // one mapping, type 0 + w.write(0, 16); + w.flag(false); // one submap + w.flag(false); // no coupling + w.write(0, 2); + w.write(0, 8); // submap 0: time, floor 0, residue 0 + w.write(0, 8); + w.write(0, 8); + w.write(0, 6); // one mode: long blocks, mapping 0 + w.flag(true); + w.write(0, 16); + w.write(0, 16); + w.write(0, 8); + w.flag(true); + w.finish() + } + + /// Encodes planar samples (one slice per channel, all the same length, nominal range ±1); + /// returns the packets completed with their granule positions. NaN or infinite samples, or a + /// channel count or length mismatch, are an error. + pub fn encode(&mut self, planar: &[&[f32]]) -> Result, i64)>> { + if self.finished { + return Err(Error::InvalidArgument("Vorbis encoder: encode after finish")); + } + if planar.len() != self.cfg.channels { + return Err(Error::InvalidArgument("Vorbis encoder: one slice per channel")); + } + let len = planar.first().map_or(0, |p| p.len()); + if planar.iter().any(|p| p.len() != len) { + return Err(Error::InvalidArgument("Vorbis encoder: channels of different lengths")); + } + if planar.iter().any(|p| p.iter().any(|v| !v.is_finite())) { + return Err(Error::InvalidArgument("Vorbis encoder: NaN or infinite sample")); + } + for (b, p) in self.pending.iter_mut().zip(planar) { + b.extend_from_slice(p); + } + self.samples_in += len as u64; + self.drain() + } + + /// Encodes what remains, padded with silence; the last packet's granule position is the + /// input's length, so the decoder trims the padding (§A.2). + pub fn finish(&mut self) -> Result, i64)>> { + if self.finished { + return Ok(Vec::new()); + } + self.finished = true; + // block k ends output at k·HOP: blocks 0..=ceil(len / HOP) + let blocks = self.samples_in.div_ceil(HOP as u64) + 1; + let missing = blocks.saturating_sub(self.blocks) as usize; + self.compact(); + let need = (missing + 1) * HOP; + for b in self.pending.iter_mut() { + if b.len() < need { + b.resize(need, 0.0); + } + } + let mut out = Vec::new(); + for _ in 0..missing { + out.push(self.block()?); + } + Ok(out) + } + + fn drain(&mut self) -> Result, i64)>> { + let mut out = Vec::new(); + while self.pending.first().is_some_and(|b| b.len().saturating_sub(self.start) >= BLOCK) { + out.push(self.block()?); + } + self.compact(); + Ok(out) + } + + /// Drops the hops already encoded from the front of every channel buffer. + fn compact(&mut self) { + for b in self.pending.iter_mut() { + b.drain(..self.start.min(b.len())); + } + self.start = 0; + } + + /// Encodes the block at `start` in `pending` and advances by a hop. + fn block(&mut self) -> Result<(Vec, i64)> { + let ch = self.cfg.channels; + let mut w = BitWriter::new(); + w.write(0, 1); // audio packet; one mode, so no mode number + w.flag(true); // previous and next windows long + w.flag(true); + let mut residue = vec![0i32; ch * HOP]; + let mut any = false; + let mut x = vec![0f32; BLOCK]; + let mut coeffs = vec![0f32; HOP]; + for c in 0..ch { + let src = self.pending.get(c).and_then(|b| b.get(self.start..)).ok_or(Error::Internal("channel buffer"))?; + for (i, v) in x.iter_mut().enumerate() { + *v = src.get(i).copied().unwrap_or(0.0).clamp(-8.0, 8.0) * self.window[i]; + } + self.mdct.forward(&x, &mut coeffs, 2.0 / HOP as f64); + let peak = coeffs.iter().fold(0f32, |m, v| m.max(v.abs())); + // (a NaN peak cannot happen: samples are finite and clamped) + if peak <= 1e-7 { + w.flag(false); // floor unused: the channel is silent + continue; + } + any = true; + w.flag(true); + let want = self.floor_targets(&coeffs); + let curve = self.floor.encode(&mut w, &self.books[BOOK_FLOOR], &want, HOP); + for (k, (&v, &f)) in coeffs.iter().zip(&curve).enumerate() { + let r = if f > 0.0 { v / f } else { 0.0 }; + // a dead zone: values just past half a step are noise peaks, not worth a ±1 + let q = if r.abs() < self.dead_zone { 0.0 } else { r.round() }; + residue[k * ch + c] = if q.is_finite() { (q as i32).clamp(-MAX_RESIDUE, MAX_RESIDUE) } else { 0 }; + } + } + if any { + self.write_residue(&mut w, &residue); + } + let granule = if self.blocks == 0 { + 0 + } else { + let natural = self.blocks.saturating_mul(HOP as u64); + if self.finished { natural.min(self.samples_in) } else { natural } + }; + self.start += HOP; + self.blocks += 1; + Ok((w.finish(), i64::try_from(granule).unwrap_or(i64::MAX))) + } + + /// Floor values at the floor's points: the spectral envelope (power smoothed over a quarter of + /// the frequency, at least 8 bins, on each side) less an offset that grows with the region's tonality, from its + /// spectral flatness: noise needs a few dB of precision, tones much more. + fn floor_targets(&self, coeffs: &[f32]) -> Vec { + let n = coeffs.len(); + let mut cum = vec![0f64; n + 1]; + let mut cum_ln = vec![0f64; n + 1]; + for (i, &v) in coeffs.iter().enumerate() { + let p = f64::from(v) * f64::from(v); + cum[i + 1] = cum[i] + p; + cum_ln[i + 1] = cum_ln[i] + (p + 1e-30).ln(); + } + self.floor + .xs + .iter() + .map(|&x| { + let k = (x as usize).min(n - 1); + let half = (k / 4).max(8); + let (lo, hi) = (k.saturating_sub(half), (k + half + 1).min(n)); + let len = (hi - lo) as f64; + let am = (cum[hi] - cum[lo]) / len; + let gm = ((cum_ln[hi] - cum_ln[lo]) / len).exp(); + // spectral flatness: about -2.5 dB for noise, far below for a tone + let sfm_db = 10.0 * (gm / am.max(1e-30)).max(1e-30).log10(); + let tonality = ((-sfm_db - 2.5) / 20.0).clamp(0.0, 1.0); + let offset = tonality * f64::from(self.tonal_db) + (1.0 - tonality) * f64::from(self.noise_db); + let amp = (am.sqrt() * 10f64.powf(-offset / 20.0)).max(1e-7) as f32; + (floor::db_index(amp) + 1) / floor::MULTIPLIER as i32 + }) + .collect() + } + + /// Residue type 2 (§8.6.2 with the interleaved vector): classes per partition, then the + /// cascade's passes in the decoder's order. + fn write_residue(&self, w: &mut BitWriter, r: &[i32]) { + let partition = partition_size(self.cfg.channels); + let parts = r.len() / partition; + let classes: Vec = r.chunks(partition).map(class_of).collect(); + let per_word = self.books[BOOK_CLASS].dimensions as usize; + for pass in 0..3 { + let mut pc = 0; + while pc < parts { + if pass == 0 { + let mut e = 0usize; + for i in 0..per_word { + e = e * CLASSES + classes.get(pc + i).copied().unwrap_or(0); + } + self.books[BOOK_CLASS].put(w, e); + } + for _ in 0..per_word { + if pc >= parts { + break; + } + let class = classes[pc]; + if let Some(b) = CASCADE[class][pass] { + let book = &self.books[b]; + let dims = book.dimensions as usize; + let part = &r[pc * partition..(pc + 1) * partition]; + for chunk in part.chunks(dims) { + let v: Vec = chunk.iter().map(|&x| pass_value(x, class, pass)).collect(); + book.put(w, book.entry_of(&v)); + } + } + pc += 1; + } + } + } + } +} + +/// Relative frequency of each residue class (for the class book's code lengths). +fn class_weight(c: usize) -> f64 { + [0.3, 0.2, 0.12, 0.12, 0.1, 0.08, 0.06, 0.02].get(c).copied().unwrap_or(0.01) +} + +/// The part of residue value `v` coded in `pass` for `class`: the passes add up to `v`. +fn pass_value(v: i32, class: usize, pass: usize) -> i32 { + let split = |v: i32, step: i32| -> (i32, i32) { + let a = ((f64::from(v) / f64::from(step)).round() as i32).clamp(-7, 7); + (a * step, v - a * step) + }; + match (class, pass) { + (6, 0) => split(v, 15).0, + (6, _) => split(v, 15).1, + (7, 0) => split(v, 225).0, + (7, 1) => split(split(v, 225).1, 15).0, + (7, _) => split(split(v, 225).1, 15).1, + _ => v, + } +} + +/// The residue class of a partition: silent; ±1 with at most four nonzero values, or more; then +/// by the largest magnitude. +fn class_of(p: &[i32]) -> usize { + let m = p.iter().map(|v| v.abs()).max().unwrap_or(0); + match m { + 0 => 0, + 1 if p.iter().filter(|&&v| v != 0).count() <= 4 => 1, + 1 => 2, + 2 => 3, + 3..=4 => 4, + 5..=7 => 5, + 8..=112 => 6, + _ => 7, + } +} + +#[cfg(test)] +mod tests; diff --git a/crates/audio/src/vorbis/encoder/tests.rs b/crates/audio/src/vorbis/encoder/tests.rs new file mode 100644 index 000000000..b1ae87463 --- /dev/null +++ b/crates/audio/src/vorbis/encoder/tests.rs @@ -0,0 +1,285 @@ +//! Unit tests, and ffmpeg (as an external oracle only) decoding the encoder's streams. + +#![allow(clippy::unwrap_used)] + +use std::path::Path; +use std::process::Command; + +use super::book::{BitWriter, assign_codewords, float32_pack}; +use super::{BLOCK, Encoder, EncoderConfig, HOP, partition_size}; + +#[test] +fn codewords_follow_the_specification_example() { + // Vorbis I §3.2.1: lengths 2 4 4 4 4 2 3 3 + let words = assign_codewords(&[2, 4, 4, 4, 4, 2, 3, 3]).unwrap(); + assert_eq!(words, [0b00, 0b0100, 0b0101, 0b0110, 0b0111, 0b10, 0b110, 0b111]); + assert!(assign_codewords(&[1, 1, 1]).is_err(), "over-subscribed"); +} + +#[test] +fn every_book_is_a_complete_prefix_code() { + let e = Encoder::new(EncoderConfig::new(48_000, 2, 5.0)).unwrap(); + for (i, b) in e.books.iter().enumerate() { + let kraft: f64 = b.lengths.iter().map(|&l| 0.5f64.powi(l as i32)).sum(); + assert!((kraft - 1.0).abs() < 1e-9, "book {i}: Kraft sum {kraft}"); + assert!(b.lengths.iter().all(|&l| (1..=20).contains(&l))); + for (a, (&wa, &la)) in b.words.iter().zip(&b.lengths).enumerate() { + for (c, (&wc, &lc)) in b.words.iter().zip(&b.lengths).enumerate() { + if a != c && la <= lc { + assert_ne!(wc >> (lc - la), wa, "book {i}: codeword {a} is a prefix of {c}"); + } + } + } + } +} + +#[test] +fn float32_pack_is_exact_for_lattice_values() { + let unpack = |x: u32| { + let mant = f64::from(x & 0x1F_FFFF); + let exp = ((x >> 21) & 0x3FF) as i32 - 788; + let v = mant * 2f64.powi(exp); + if x & 0x8000_0000 != 0 { -v } else { v } + }; + for v in [1.0, -1.0, -2.0, -4.0, -7.0, -105.0, 15.0, -1575.0, 225.0, 0.5, 3.0] { + assert_eq!(unpack(float32_pack(v)), v); + } + assert_eq!(float32_pack(0.0), 0); +} + +#[test] +fn bits_are_packed_lsb_first() { + let mut w = BitWriter::new(); + w.write(0b1, 1); + w.write(0b10, 2); + w.write(0x1F, 5); + w.write(0xABC, 12); + assert_eq!(w.finish(), [0b1111_1101, 0xBC, 0x0A]); +} + +#[test] +fn hostile_input_is_an_error_not_a_panic() { + assert!(Encoder::new(EncoderConfig::new(48_000, 0, 3.0)).is_err()); + assert!(Encoder::new(EncoderConfig::new(48_000, 9, 3.0)).is_err()); + assert!(Encoder::new(EncoderConfig::new(0, 2, 3.0)).is_err()); + assert_eq!(Encoder::new(EncoderConfig::new(48_000, 1, f32::NAN)).unwrap().cfg.quality, 6.0, "NaN quality uses the default, as the export does"); + let mut e = Encoder::new(EncoderConfig::new(44_100, 2, 3.0)).unwrap(); + assert!(e.encode(&[&[0.0; 4]]).is_err()); + assert!(e.encode(&[&[0.0; 4], &[0.0; 3]]).is_err()); + assert!(e.encode(&[&[f32::NAN; 4], &[0.0; 4]]).is_err()); + let loud = vec![1e30f32; 5000]; + assert!(!e.encode(&[&loud, &loud]).unwrap().is_empty()); + assert!(!e.finish().unwrap().is_empty()); + assert!(e.finish().unwrap().is_empty()); + assert!(e.encode(&[&[0.0; 4], &[0.0; 4]]).is_err()); +} + +#[test] +fn granules_count_the_samples_and_trim_the_end() { + for len in [0usize, 1, 1023, 1024, 1025, 48_000] { + let mut e = Encoder::new(EncoderConfig::new(48_000, 1, 3.0)).unwrap(); + let x = vec![0.1f32; len]; + let mut packets = e.encode(&[&x]).unwrap(); + packets.extend(e.finish().unwrap()); + assert_eq!(packets.len(), len.div_ceil(HOP) + 1, "len {len}"); + // packet k ends output at k·1024 samples; the last one at the input's length + let want: Vec = (0..packets.len()).map(|k| ((k * HOP) as i64).min(len as i64)).collect(); + let got: Vec = packets.iter().map(|p| p.1).collect(); + assert_eq!(got, want, "len {len}"); + } +} + +// --- ffmpeg oracle --- + +/// An Ogg Vorbis file of the three headers and the audio packets, through craft-media's writer. +fn ogg(packets: &[(Vec, i64)]) -> Vec { + let mut w = craft_media::ogg::PageWriter::new(7); + let mut out = Vec::new(); + for (i, (p, g)) in packets.iter().enumerate() { + out.extend(if i < 3 { w.header(p) } else { w.packet(p, *g) }); + } + out.extend(w.finish(None)); + out +} + +fn comment() -> Vec { + craft_media::ogg::vorbis_comment("test", &[]) +} + +/// Encodes `input` (planar) and decodes it with ffmpeg's `decoder`; returns the planar output. +/// ffmpeg's decoders to check: its own `vorbis` always, and `libvorbis` when ffmpeg has it. +fn decoders(ff: &Path) -> Vec<&'static str> { + let list = + Command::new(ff).args(["-hide_banner", "-decoders"]).output().map(|o| String::from_utf8_lossy(&o.stdout).into_owned()).unwrap_or_default(); + if list.lines().any(|l| l.split_whitespace().nth(1) == Some("libvorbis")) { + vec!["vorbis", "libvorbis"] + } else { + eprintln!("SKIPPED (libvorbis decoder): this ffmpeg has no libvorbis; checking its own decoder only"); + vec!["vorbis"] + } +} + +fn round_trip(input: &[Vec], rate: u32, quality: f32, decoder: &str, name: &str) -> Option<(Vec>, usize)> { + let ff = craft_media_testkit::ffmpeg_or_skip("vorbis oracle")?; + let mut e = Encoder::new(EncoderConfig::new(rate, input.len(), quality)).unwrap(); + let mut packets = vec![(e.identification(), 0), (comment(), 0), (e.setup(), 0)]; + let n = input[0].len(); + let mut at = 0; + while at < n { + let end = (at + 4410).min(n); + let s: Vec<&[f32]> = input.iter().map(|c| &c[at..end]).collect(); + packets.extend(e.encode(&s).unwrap()); + at = end; + } + packets.extend(e.finish().unwrap()); + let bytes: usize = packets[3..].iter().map(|p| p.0.len()).sum(); + let dir = craft_media_testkit::fixtures_dir("audio/vorbis-encoder"); + let path = dir.join(format!("{name}.ogg")); + std::fs::write(&path, ogg(&packets)).unwrap(); + let out = Command::new(ff) + .args(["-v", "error", "-xerror", "-c:a", decoder, "-i"]) + .arg(&path) + .args(["-f", "f32le", "-c:a", "pcm_f32le", "-"]) + .output() + .unwrap(); + assert!(out.status.success() && out.stderr.is_empty(), "{name} {decoder}: {}", String::from_utf8_lossy(&out.stderr)); + let pcm: Vec = out.stdout.as_chunks::<4>().0.iter().map(|&b| f32::from_le_bytes(b)).collect(); + let ch = input.len(); + Some(((0..ch).map(|c| pcm.iter().skip(c).step_by(ch).copied().collect()).collect(), bytes)) +} + +fn snr_db(a: &[f32], b: &[f32]) -> f64 { + let (mut s, mut e) = (0f64, 0f64); + for (&x, &y) in a.iter().zip(b) { + s += f64::from(x) * f64::from(x); + e += (f64::from(x) - f64::from(y)).powi(2); + } + 10.0 * (s.max(1e-20) / e.max(1e-20)).log10() +} + +fn music(n: usize, seed: u32) -> Vec { + let mut r = seed; + (0..n) + .map(|i| { + let t = i as f32 / 48_000.0; + r = r.wrapping_mul(1_664_525).wrapping_add(1_013_904_223); + let noise = (r >> 8) as f32 / 8_388_608.0 - 1.0; + let chord: f32 = [220.0f32, 277.18, 329.63, 440.0] + .iter() + .enumerate() + .map(|(k, f)| (0.2 / (k + 1) as f32) * (2.0 * std::f32::consts::PI * f * t).sin()) + .sum(); + chord + 0.01 * noise + }) + .collect() +} + +/// Harmonic chords changing every half second, with decaying noise bursts (music-like). +fn chords(n: usize, seed: u32) -> Vec { + let mut r = seed; + let notes = [220.0f32, 277.2, 329.6, 440.0, 349.2, 523.3, 392.0, 659.3]; + (0..n) + .map(|i| { + let t = i as f32 / 48_000.0; + let chord = (t * 2.0) as usize; + let mut s = 0.0; + for k in 0..3 { + let f = notes[(chord * 3 + k * 2) % notes.len()]; + for h in 1..6 { + s += (2.0 * std::f32::consts::PI * f * h as f32 * t).sin() * 0.08 / h as f32; + } + } + r = r.wrapping_mul(1_664_525).wrapping_add(1_013_904_223); + let env = (-(t * 2.0).fract() * 3.0).exp(); + s * (0.5 + 0.5 * env) + 0.02 * ((r >> 8) as f32 / 8_388_608.0 - 1.0) * env + }) + .collect() +} + +/// ffmpeg's own Vorbis decoder and libvorbis read every stream exactly as long as the input, +/// with SNR floors about 6 dB (chords 4 dB) under what the encoder measured when written, and bit +/// rates near libvorbis's for the same quality. +/// A signal, and per quality the SNR floor and the bit-rate range (kbit/s). +type Case = (&'static str, Vec>, [(f32, f64, std::ops::Range); 5]); + +#[test] +fn ffmpeg_decodes_every_quality() { + let Some(ff) = craft_media_testkit::ffmpeg_or_skip("vorbis oracle") else { return }; + let n = 48_000 * 3; + let cases: [Case; 2] = [ + ( + "music", + vec![music(n, 1), music(n, 2).iter().map(|v| v * 0.7).collect()], + [(-1.0, 3.0, 15..90), (0.0, 6.5, 20..110), (3.0, 12.0, 60..180), (6.0, 18.5, 120..280), (10.0, 25.5, 180..400)], + ), + ( + "chords", + vec![chords(n, 1), chords(n, 2).iter().map(|v| v * 0.8).collect()], + [(-1.0, 4.5, 20..90), (0.0, 6.0, 25..110), (3.0, 10.0, 70..200), (6.0, 14.0, 130..300), (10.0, 19.0, 180..420)], + ), + ]; + for (name, input, levels) in &cases { + for (q, floor, kbps) in levels { + for decoder in decoders(&ff) { + let Some((out, bytes)) = round_trip(input, 48_000, *q, decoder, &format!("{name}{q}")) else { return }; + assert!(out.iter().all(|c| c.len() == n), "{name} q{q} {decoder}: length {}", out[0].len()); + let snr = snr_db(&input[0][4096..], &out[0][4096..]); + assert!(snr >= *floor, "{name} q{q} {decoder}: {snr:.2} dB < {floor}"); + let rate = bytes * 8 / 3 / 1000; + assert!(kbps.contains(&rate), "{name} q{q}: {rate} kbit/s outside {kbps:?}"); + } + } + } +} + +/// Mono, 44.1 kHz, 3.0, 5.1 and 6.1 streams decode too, with libvorbis as well (which reads each +/// residue partition from channel 0); a silent channel costs one bit per block. +#[test] +fn channel_counts_and_rates() { + let Some(ff) = craft_media_testkit::ffmpeg_or_skip("vorbis oracle") else { return }; + let n = 44_100; + for (rate, ch) in [(44_100u32, 1usize), (22_050, 2), (48_000, 3), (48_000, 6), (48_000, 7)] { + let mut input: Vec> = (0..ch).map(|c| chords(n, c as u32 + 1).iter().map(|v| v * (1.0 - 0.1 * c as f32)).collect()).collect(); + if ch == 6 { + input[3] = vec![0.0; n]; + } + for decoder in decoders(&ff) { + let Some((mut out, _)) = round_trip(&input, rate, 5.0, decoder, &format!("ch{ch}_{rate}_{decoder}")) else { return }; + if decoder == "vorbis" { + // ffmpeg's decoder gives WAV order (5.1: FL FR FC LFE BL BR); the stream is in Vorbis + // order (L C R Ls Rs LFE), which libvorbis passes through + let wav_to_vorbis: &[usize] = match ch { + 3 => &[0, 2, 1], + 6 => &[0, 2, 1, 4, 5, 3], + 7 => &[0, 2, 1, 5, 6, 4, 3], + _ => &[0, 1], + }; + if ch > 2 { + out = wav_to_vorbis.iter().map(|&i| out[i].clone()).collect(); + } + } + assert_eq!(out.len(), ch); + for (c, (a, b)) in input.iter().zip(&out).enumerate() { + assert_eq!(b.len(), n, "{ch} ch {rate} Hz {decoder}: channel {c} length"); + if a.iter().all(|&v| v == 0.0) { + assert!(b.iter().all(|v| v.abs() < 1e-6), "silent channel {c}"); + } else { + let snr = snr_db(&a[4096..], &b[4096..]); + assert!(snr > 8.0, "{ch} ch {rate} Hz {decoder} channel {c}: {snr:.1} dB"); + } + } + } + } +} + +/// Every partition holds whole sample frames (libvorbis decodes each one from channel 0), splits +/// the interleaved residue evenly and fits every book's dimension. +#[test] +fn partitions_hold_whole_frames() { + for ch in 1..=8 { + let p = partition_size(ch); + assert_eq!(p % ch, 0, "{ch} channels"); + assert_eq!((ch * BLOCK / 2) % p, 0, "{ch} channels"); + assert_eq!(p % 4, 0, "{ch} channels"); + } +} diff --git a/crates/audio/src/vorbis/mod.rs b/crates/audio/src/vorbis/mod.rs new file mode 100644 index 000000000..29c423f1d --- /dev/null +++ b/crates/audio/src/vorbis/mod.rs @@ -0,0 +1,4 @@ +//! Vorbis I (feature `vorbis-encode`): the pure-Rust [`encoder`]. Decoding Vorbis goes through +//! [`crate::decode`]. + +pub mod encoder; diff --git a/crates/audio/src/wav_markers.rs b/crates/audio/src/wav_markers.rs new file mode 100644 index 000000000..0bf4e7568 --- /dev/null +++ b/crates/audio/src/wav_markers.rs @@ -0,0 +1,721 @@ +//! Markers a WAV file carries, read from its RIFF chunks. +//! +//! Two encodings are read into separate lists for the caller to map and merge: +//! - **Cue points** (the standard): a `cue ` chunk with one point per marker, named by `labl` (or +//! `note`) sub-chunks of a `LIST` `adtl` chunk, and given a length by `ltxt`. +//! - **XMP** (Adobe's tools): an `_PMX` chunk holding an XMP packet whose `xmpDM:markers` list has +//! one entry per marker (`xmpDM:startTime`, `xmpDM:duration`, `xmpDM:name`), counted in frames +//! of `xmpDM:frameRate` (`f44100` for samples at 44.1 kHz), or with an explicit +//! FrameCount rate suffix (`15f24`, `15f30000s1001`). Durations with a distinct rate +//! are converted to the marker start rate without app-specific time quantization. +//! +//! RF64 files (WAVs over 4 GB) carry a `0xFFFFFFFF` size for a chunk too big for 32 bits and give its real size in a +//! `ds64` chunk; those sizes are resolved so the chunks after the audio are still found. +//! +//! The file is untrusted input: every read is bounds-checked, sizes use checked arithmetic, and +//! the number of markers, of `adtl` sub-chunks looked at and the size of the XMP packet are capped. + +use std::collections::{HashMap, HashSet}; + +/// Marker coordinates before app-specific time quantization. Explicit XMP duration +/// rates are normalized to the start coordinate rate; bare counts retain the legacy +/// track/file-rate interpretation. +#[derive(Clone, Debug, PartialEq)] +pub struct Marker { + /// Start in frames at `rate`. + pub start: f64, + /// Duration in frames at `rate`. + pub duration: f64, + /// Frames per second; consumers decide how invalid or absent rates map to their timeline. + pub rate: f64, + /// Label text. + pub name: String, +} +/// Separate source lists: rounding, deduplication, ordering and labels belong to the caller. +#[derive(Clone, Debug, Default, PartialEq)] +pub struct Markers { + /// RIFF cue/adtl markers in file order. + pub cue: Vec, + /// XMP markers in packet order. + pub xmp: Vec, + /// A chunk-inventory or cue-label retention limit omitted metadata. Marker coordinates + /// remain available when only their label was omitted. Existing marker/XMP prefix limits + /// are separate and do not set this flag. + pub limited: bool, +} + +/// Most markers read from one file. +pub const MAX_MARKERS: usize = 10_000; +/// Largest XMP packet read (bytes). +const MAX_XMP: usize = 4 << 20; +/// Most `adtl` sub-chunks (`labl`, `note`, `ltxt`) looked at in one file: a few per cue point is plenty. +const MAX_ADTL_RECORDS: usize = 4 * MAX_MARKERS; +/// Most entries read from an RF64 file's `ds64` size table. +const MAX_DS64_TABLE: usize = 1_024; +/// Maximum retained RIFF chunk descriptors (including empty chunks). +pub const MAX_CHUNKS: usize = 100_000; +/// Maximum UTF-8 bytes in one retained cue label. +pub const MAX_LABEL_BYTES: usize = 64 * 1024; +/// Aggregate cue label bytes, enforced separately for the name map and emitted markers. +pub const MAX_LABEL_TOTAL_BYTES: usize = 4 * 1024 * 1024; + +fn u32_at(b: &[u8], at: usize) -> Option { + let s = b.get(at..at.checked_add(4)?)?; + Some(u32::from_le_bytes(s.try_into().ok()?)) +} + +fn u64_at(b: &[u8], at: usize) -> Option { + let s = b.get(at..at.checked_add(8)?)?; + Some(u64::from_le_bytes(s.try_into().ok()?)) +} + +/// The text of a NUL-terminated label (the rest of the chunk), lossily decoded. +fn label(b: &[u8]) -> Option { + // Look only as far as one label can be retained; never copy an arbitrary chunk body. + let end = b.iter().take(MAX_LABEL_BYTES + 1).position(|&c| c == 0).unwrap_or(b.len()); + if end > MAX_LABEL_BYTES { + return None; + } + let text = String::from_utf8_lossy(b.get(..end)?); + let text = text.trim(); + (text.len() <= MAX_LABEL_BYTES).then(|| text.to_string()) +} + +/// Whether `bytes` start a RIFF/WAVE file. +pub fn is_wav(bytes: &[u8]) -> bool { + matches!(bytes.get(0..4), Some(b"RIFF" | b"RF64")) && bytes.get(8..12) == Some(b"WAVE") +} + +/// The sizes an RF64 file's `ds64` chunk gives for chunks too big for the 32-bit size field. +struct Ds64 { + /// The `data` chunk's size. + data: u64, + /// Other chunks: `(id, size)`. + table: Vec<([u8; 4], u64)>, +} + +impl Ds64 { + /// The real size of chunk `id`, when the file says. + fn size_of(&self, id: &[u8; 4]) -> Option { + if id == b"data" { Some(self.data) } else { self.table.iter().find(|(t, _)| t == id).map(|(_, size)| *size) } + } +} + +/// The `ds64` chunk of an RF64 file (`RF64`, size, `WAVE`, then `ds64`); `None` for any other file or one whose +/// `ds64` doesn't fit. +fn ds64(bytes: &[u8]) -> Option { + if bytes.get(0..4) != Some(b"RF64") || bytes.get(12..16) != Some(b"ds64") { + return None; + } + let size = usize::try_from(u32_at(bytes, 16)?).ok()?; + let body = bytes.get(20..20usize.checked_add(size)?)?; + // riff size (8), data size (8), sample count (8), table length (4), then `(id, size)` entries of 12 bytes + let data = u64_at(body, 8)?; + let entries = usize::try_from(u32_at(body, 24)?).ok()?; + let mut table = Vec::new(); + for i in 0..entries.min(MAX_DS64_TABLE) { + let at = 28usize.checked_add(i.checked_mul(12)?)?; + let id = <[u8; 4]>::try_from(body.get(at..at.checked_add(4)?)?).ok()?; + table.push((id, u64_at(body, at + 4)?)); + } + Some(Ds64 { data, table }) +} + +/// The chunks of a WAVE file as `(id, body)`, stopping at the first one that doesn't fit. +fn chunks<'a>(bytes: &'a [u8], limited: &mut bool) -> Vec<([u8; 4], &'a [u8])> { + let big = ds64(bytes); + let mut out = Vec::new(); + let mut at = 12usize; + while let Some(head) = bytes.get(at..at.saturating_add(8)) { + if out.len() == MAX_CHUNKS { + *limited = true; + break; + } + let (Ok(id), Some(field)) = (<[u8; 4]>::try_from(&head[..4]), u32_at(head, 4)) else { break }; + // In an RF64 file a size field of 0xFFFFFFFF stands for the size `ds64` gives. + let size = match &big { + Some(d) if field == u32::MAX => d.size_of(&id).unwrap_or(u64::from(field)), + _ => u64::from(field), + }; + let start = at.saturating_add(8); + let Some(end) = usize::try_from(size).ok().and_then(|n| start.checked_add(n)) else { break }; + let Some(body) = bytes.get(start..end) else { break }; + out.push((id, body)); + // Chunks are padded to an even size. + at = end.saturating_add(end & 1); + } + out +} + +/// Read bounded cue/adtl and XMP metadata without interpreting application timeline policy. +pub fn read(bytes: &[u8]) -> Markers { + if !is_wav(bytes) { + return Markers::default(); + } + let mut limited = false; + let chunks = chunks(bytes, &mut limited); + let rate = chunks.iter().find(|(id, _)| id == b"fmt ").and_then(|(_, b)| u32_at(b, 4)).map_or(0.0, f64::from); + let cue = cue_markers(&chunks, rate, &mut limited); + let xmp = chunks.iter().find(|(id, _)| id == b"_PMX").map_or_else(Vec::new, |(_, xmp)| xmp_markers(xmp, rate)); + Markers { cue, xmp, limited } +} + +/// Cue points named and sized by the `adtl` list. +fn cue_markers(chunks: &[([u8; 4], &[u8])], rate: f64, limited: &mut bool) -> Vec { + let Some((_, cue)) = chunks.iter().find(|(id, _)| id == b"cue ") else { return Vec::new() }; + let count = (u32_at(cue, 0).unwrap_or(0) as usize).min(MAX_MARKERS); + // The cue points first: `(id, offset in samples)`. + let mut points = Vec::new(); + for i in 0..count { + let Some(point) = cue.get(4usize.saturating_add(i.saturating_mul(24))..) else { break }; + let (Some(id), Some(offset)) = (u32_at(point, 0), u32_at(point, 20)) else { break }; + points.push((id, offset)); + } + // Only the names and lengths of those ids are kept, so a hostile `adtl` list cannot grow them past the cue count, + // and each record is a lookup, not a scan. + let wanted: HashSet = points.iter().map(|(id, _)| *id).collect(); + let mut names: HashMap = HashMap::new(); + let mut lengths: HashMap = HashMap::new(); + let mut name_bytes = 0usize; + let mut records = 0usize; + 'lists: for (_, list) in chunks.iter().filter(|(id, b)| id == b"LIST" && b.get(0..4) == Some(b"adtl")) { + let mut at = 4usize; + while let Some(head) = list.get(at..at.saturating_add(8)) { + let (Ok(id), Some(size)) = (<[u8; 4]>::try_from(&head[..4]), u32_at(head, 4)) else { break }; + let start = at.saturating_add(8); + let Some(end) = start.checked_add(size as usize) else { break }; + let Some(body) = list.get(start..end) else { break }; + records += 1; + if records > MAX_ADTL_RECORDS { + break 'lists; + } + if let Some(cue_id) = u32_at(body, 0).filter(|c| wanted.contains(c)) { + match &id { + // A `labl` is the name; a `note` only when there is none. + b"labl" | b"note" if id == *b"labl" || !names.contains_key(&cue_id) => { + let without_old = name_bytes.saturating_sub(names.get(&cue_id).map_or(0, String::len)); + let text = label(body.get(4..).unwrap_or_default()) + .filter(|text| text.len() <= MAX_LABEL_TOTAL_BYTES.saturating_sub(without_old)) + .unwrap_or_else(|| { + *limited = true; + String::new() + }); + name_bytes = without_old + text.len(); + names.insert(cue_id, text); + } + // The first `ltxt` of a cue point gives its length. + b"ltxt" => { + lengths.entry(cue_id).or_insert_with(|| u32_at(body, 4).unwrap_or(0)); + } + _ => {} + } + } + at = end.saturating_add(end & 1); + } + } + let mut emitted_bytes = 0usize; + points + .into_iter() + .map(|(id, offset)| { + let comment = match names.get(&id) { + Some(text) if text.len() <= MAX_LABEL_TOTAL_BYTES.saturating_sub(emitted_bytes) => { + emitted_bytes += text.len(); + text.clone() + } + Some(_) => { + *limited = true; + String::new() + } + None => String::new(), + }; + let length = lengths.get(&id).copied().unwrap_or(0); + Marker { start: f64::from(offset), duration: f64::from(length), rate, name: comment } + }) + .collect() +} + +/// The text of `name` in an XMP entry, as an attribute (`name="v"`) or an element +/// (`v`). +fn xmp_value(entry: &str, name: &str) -> Option { + // Each entry starts just after ` in_tag = true, + b'>' => in_tag = false, + b'\'' | b'"' if in_tag => { + let quote = bytes[at]; + at += 1; + while at < bytes.len() && bytes[at] != quote { + at += 1; + } + } + _ if in_tag && (at == 0 || xml_space(bytes[at - 1])) && bytes[at..].starts_with(name.as_bytes()) => { + let mut value = at + name.len(); + while value < bytes.len() && xml_space(bytes[value]) { + value += 1; + } + if bytes.get(value) == Some(&b'=') { + value += 1; + while value < bytes.len() && xml_space(bytes[value]) { + value += 1; + } + if let Some("e @ (b'\'' | b'"')) = bytes.get(value) { + let start = value + 1; + let end = bytes.get(start..)?.iter().position(|&b| b == quote)? + start; + return Some(unescape(entry.get(start..end)?)); + } + } + } + _ => {} + } + at += 1; + } + let open = format!("<{name}>"); + let i = entry.find(&open)?; + let rest = entry.get(i + open.len()..)?; + Some(unescape(rest.get(..rest.find('<')?)?)) +} + +fn xml_space(b: u8) -> bool { + matches!(b, b' ' | b'\t' | b'\r' | b'\n') +} + +fn unescape(s: &str) -> String { + let mut out = String::with_capacity(s.len()); + let mut rest = s; + while let Some(i) = rest.find('&') { + out.push_str(rest.get(..i).unwrap_or_default()); + let tail = rest.get(i..).unwrap_or_default(); + let Some(semi) = tail.as_bytes().iter().take(11).position(|b| *b == b';') else { + out.push('&'); + rest = tail.get(1..).unwrap_or_default(); + continue; + }; + let ent = tail.get(1..semi).unwrap_or_default(); + match ent { + "amp" => out.push('&'), + "lt" => out.push('<'), + "gt" => out.push('>'), + "quot" => out.push('"'), + "apos" => out.push('\''), + _ => { + let code = ent + .strip_prefix("#x") + .and_then(|h| u32::from_str_radix(h, 16).ok()) + .or_else(|| ent.strip_prefix('#').and_then(|d| d.parse().ok())); + match code.and_then(char::from_u32) { + Some(c) => out.push(c), + None => out.push_str(tail.get(..=semi).unwrap_or_default()), + } + } + } + rest = tail.get(semi + 1..).unwrap_or_default(); + } + out.push_str(rest); + out.trim().to_string() +} + +/// A frame rate as XMP writes it: `f44100`, or `f30000s1001` (rate over scale). +fn xmp_rate(text: &str) -> Option { + let t = text.trim().strip_prefix('f')?; + let (num, den) = t.split_once('s').unwrap_or((t, "1")); + let (n, d): (f64, f64) = (num.parse().ok()?, den.parse().ok()?); + (n.is_finite() && d.is_finite() && n > 0.0 && d > 0.0 && (n / d).is_finite() && n / d > 0.0).then_some(n / d) +} + +/// Adobe XMP FrameCount: a count optionally followed by `fRATE[sSCALE]`. +/// Bare counts retain the caller's legacy track/file-rate interpretation. +fn xmp_count(text: &str, default_rate: f64) -> Option<(f64, f64)> { + let text = text.trim(); + let (count, rate) = match text.find('f') { + Some(at) => (text.get(..at)?, xmp_rate(text.get(at..)?)?), + None => (text, default_rate), + }; + let count: f64 = count.parse().ok()?; + (count.is_finite() && rate.is_finite() && rate > 0.0).then_some((count, rate)) +} + +/// The markers of an XMP packet's `xmpDM:markers` entries. +fn xmp_markers(packet: &[u8], file_rate: f64) -> Vec { + let text = String::from_utf8_lossy(packet.get(..packet.len().min(MAX_XMP)).unwrap_or_default()); + let mut out = Vec::new(); + // The frame rate belongs to the track an entry follows. + let mut rate = file_rate; + for entry in text.split("= MAX_MARKERS { + break; + } + } + out +} + +#[cfg(test)] +mod tests { + use super::*; + + struct TimedMarker { + time: f64, + duration: f64, + comment: String, + } + fn read(bytes: &[u8]) -> Vec { + let metadata = super::read(bytes); + let time = |frames: f64, rate: f64| if rate > 0.0 && frames.is_finite() { (frames / rate).max(0.0) } else { 0.0 }; + let map = |m: Marker| TimedMarker { time: time(m.start, m.rate), duration: time(m.duration, m.rate), comment: m.name }; + let mut out: Vec<_> = metadata.cue.into_iter().map(map).collect(); + let mut seen: HashSet<_> = out.iter().map(|m| (m.comment.clone(), m.time.to_bits())).collect(); + for m in metadata.xmp.into_iter().map(map) { + if seen.insert((m.comment.clone(), m.time.to_bits())) { + out.push(m); + } + } + out.truncate(MAX_MARKERS); + out.sort_by(|a, b| a.time.total_cmp(&b.time)); + out + } + + fn chunk(id: &[u8; 4], body: &[u8]) -> Vec { + let mut v = id.to_vec(); + v.extend((body.len() as u32).to_le_bytes()); + v.extend(body); + if body.len() % 2 == 1 { + v.push(0); + } + v + } + + fn fmt(rate: u32) -> Vec { + let mut b = Vec::new(); + b.extend(1u16.to_le_bytes()); // PCM + b.extend(1u16.to_le_bytes()); // mono + b.extend(rate.to_le_bytes()); + b.extend((rate * 2).to_le_bytes()); + b.extend(2u16.to_le_bytes()); + b.extend(16u16.to_le_bytes()); + chunk(b"fmt ", &b) + } + + fn wav(chunks: &[Vec]) -> Vec { + let body: Vec = chunks.iter().flatten().copied().collect(); + let mut v = b"RIFF".to_vec(); + v.extend((body.len() as u32 + 4).to_le_bytes()); + v.extend(b"WAVE"); + v.extend(body); + v + } + + fn cue(points: &[(u32, u32)]) -> Vec { + let mut b = (points.len() as u32).to_le_bytes().to_vec(); + for (id, offset) in points { + b.extend(id.to_le_bytes()); + b.extend(0u32.to_le_bytes()); + b.extend(b"data"); + b.extend(0u32.to_le_bytes()); + b.extend(0u32.to_le_bytes()); + b.extend(offset.to_le_bytes()); + } + chunk(b"cue ", &b) + } + + fn labl(id: &[u8; 4], cue_id: u32, text: &str) -> Vec { + let mut b = cue_id.to_le_bytes().to_vec(); + b.extend(text.as_bytes()); + b.push(0); + chunk(id, &b) + } + + fn adtl(subs: &[Vec]) -> Vec { + let mut b = b"adtl".to_vec(); + b.extend(subs.iter().flatten()); + chunk(b"LIST", &b) + } + + #[test] + fn cue_points_are_named_by_the_adtl_list_and_timed_in_samples() { + let file = wav(&[ + fmt(44100), + chunk(b"data", &[0; 8]), + cue(&[(2, 88200), (1, 22050)]), + adtl(&[labl(b"labl", 1, "SEQ01_Prologue"), labl(b"labl", 2, "SEQ02_Vallee_P1_Agriculture")]), + ]); + let m = read(&file); + assert_eq!(m.len(), 2); + assert_eq!((m[0].comment.as_str(), m[0].time), ("SEQ01_Prologue", 0.5)); + assert_eq!((m[1].comment.as_str(), m[1].time), ("SEQ02_Vallee_P1_Agriculture", 2.0)); + } + + #[test] + fn a_note_names_a_cue_only_when_it_has_no_label_and_ltxt_gives_a_length() { + let mut ltxt = 1u32.to_le_bytes().to_vec(); + ltxt.extend(44100u32.to_le_bytes()); + ltxt.extend([0; 16]); + let file = wav(&[ + fmt(44100), + cue(&[(1, 0), (2, 44100)]), + adtl(&[labl(b"note", 1, "from note"), chunk(b"ltxt", <xt), labl(b"labl", 2, "named"), labl(b"note", 2, "ignored")]), + ]); + let m = read(&file); + assert_eq!((m[0].comment.as_str(), m[0].duration), ("from note", 1.0)); + assert_eq!(m[1].comment, "named"); + } + + const XMP_ELEMENTS: &str = r#"MarkersCuef44100441000One & Two13230022050Three"#; + + #[test] + fn xmp_markers_are_read_as_elements_in_frames_of_the_track_rate() { + let file = wav(&[fmt(48000), chunk(b"_PMX", XMP_ELEMENTS.as_bytes())]); + let m = read(&file); + assert_eq!(m.len(), 2); + // 44100 frames at f44100 is one second, whatever the file's own rate. + assert_eq!((m[0].comment.as_str(), m[0].time), ("One & Two", 1.0)); + assert_eq!((m[1].comment.as_str(), m[1].time, m[1].duration), ("Three", 3.0, 0.5)); + } + + #[test] + fn xmp_markers_are_read_as_attributes_and_use_the_file_rate_without_a_track_rate() { + let xmp = r#""#; + let m = read(&wav(&[fmt(48000), chunk(b"_PMX", xmp.as_bytes())])); + assert_eq!((m[0].comment.as_str(), m[0].time), ("Attr é", 0.5)); + } + + #[test] + fn a_marker_in_both_encodings_is_listed_once() { + let xmp = r#""#; + let file = wav(&[fmt(44100), cue(&[(1, 44100)]), adtl(&[labl(b"labl", 1, "Same")]), chunk(b"_PMX", xmp.as_bytes())]); + let names: Vec = read(&file).into_iter().map(|m| m.comment).collect(); + assert_eq!(names, ["Same", "Only XMP"]); + } + + #[test] + fn damaged_or_hostile_files_read_as_far_as_they_go_without_panicking() { + assert!(read(b"").is_empty()); + assert!(read(b"RIFF").is_empty()); + assert!(read(b"not a wav file at all").is_empty()); + let good = wav(&[fmt(44100), cue(&[(1, 44100)]), adtl(&[labl(b"labl", 1, "Kept")])]); + // Every truncation is safe. + for n in 0..good.len() { + let _ = read(&good[..n]); + } + // A cue count far beyond the chunk, and chunk sizes past the end of the file. + let mut huge = 0xFFFF_FFFFu32.to_le_bytes().to_vec(); + huge.extend([0; 24]); + let m = read(&wav(&[fmt(44100), chunk(b"cue ", &huge)])); + assert!(m.len() <= 1); + let mut lying = good.clone(); + lying.extend(b"LIST\xff\xff\xff\xffadtl"); + assert_eq!(read(&lying).len(), 1); + // No sample rate: times are zero, not a division by zero. + let m = read(&wav(&[cue(&[(1, 44100)])])); + assert_eq!(m[0].time, 0.0); + } + + /// A chunk whose size field says `field` (an RF64 file's `0xFFFFFFFF`) while its body is `body`. + fn chunk_with_size_field(id: &[u8; 4], field: u32, body: &[u8]) -> Vec { + let mut v = id.to_vec(); + v.extend(field.to_le_bytes()); + v.extend(body); + if body.len() % 2 == 1 { + v.push(0); + } + v + } + + /// An RF64 `ds64` chunk: the `data` chunk's size and a table of other chunks' sizes. + fn ds64_chunk(data_size: u64, table: &[([u8; 4], u64)]) -> Vec { + let mut b = 0u64.to_le_bytes().to_vec(); // riff size (not used) + b.extend(data_size.to_le_bytes()); + b.extend(0u64.to_le_bytes()); // sample count + b.extend((table.len() as u32).to_le_bytes()); + for (id, size) in table { + b.extend(id); + b.extend(size.to_le_bytes()); + } + chunk(b"ds64", &b) + } + + fn rf64(chunks: &[Vec]) -> Vec { + let body: Vec = chunks.iter().flatten().copied().collect(); + let mut v = b"RF64".to_vec(); + v.extend(u32::MAX.to_le_bytes()); + v.extend(b"WAVE"); + v.extend(body); + v + } + + #[test] + fn rf64_markers_after_a_data_chunk_of_sentinel_size_are_read() { + let audio = [0u8; 8]; + let file = rf64(&[ + ds64_chunk(audio.len() as u64, &[]), + fmt(44100), + chunk_with_size_field(b"data", u32::MAX, &audio), + cue(&[(1, 44100)]), + adtl(&[labl(b"labl", 1, "After the audio")]), + ]); + let m = read(&file); + assert_eq!(m.len(), 1); + assert_eq!((m[0].comment.as_str(), m[0].time), ("After the audio", 1.0)); + } + + #[test] + fn rf64_sizes_of_other_chunks_come_from_the_ds64_table() { + let junk = [7u8; 6]; + let tail = || [cue(&[(1, 8000)]), adtl(&[labl(b"labl", 1, "Found")])].concat(); + let with_entry = rf64(&[ds64_chunk(0, &[(*b"junk", junk.len() as u64)]), fmt(8000), chunk_with_size_field(b"junk", u32::MAX, &junk), tail()]); + assert_eq!(read(&with_entry)[0].comment, "Found"); + // Without the table entry the sentinel is all there is, and what follows is not reached. + let without = rf64(&[ds64_chunk(0, &[]), fmt(8000), chunk_with_size_field(b"junk", u32::MAX, &junk), tail()]); + assert!(read(&without).is_empty()); + } + + #[test] + fn hostile_rf64_size_tables_read_as_far_as_they_go_without_panicking() { + // sizes that overflow when added to an offset + for data in [u64::MAX, u64::MAX - 7, 1 << 63] { + let file = + rf64(&[ds64_chunk(data, &[(*b"junk", u64::MAX)]), fmt(8000), chunk_with_size_field(b"data", u32::MAX, &[0; 4]), cue(&[(1, 0)])]); + assert!(read(&file).is_empty()); + } + // a table that claims far more entries than its chunk holds + let mut liar = ds64_chunk(0, &[]); + liar[8 + 24..8 + 28].copy_from_slice(&u32::MAX.to_le_bytes()); + let _ = read(&rf64(&[liar, fmt(8000), cue(&[(1, 0)])])); + // every truncation of a good RF64 file is safe + let good = rf64(&[ + ds64_chunk(4, &[]), + fmt(44100), + chunk_with_size_field(b"data", u32::MAX, &[0; 4]), + cue(&[(1, 44100)]), + adtl(&[labl(b"labl", 1, "Kept")]), + ]); + assert_eq!(read(&good).len(), 1); + for n in 0..good.len() { + let _ = read(&good[..n]); + } + } + + #[test] + fn a_huge_adtl_list_is_read_in_linear_time_and_only_for_the_cue_points_that_exist() { + // Records for ids that have no cue point are skipped, not stored, and the number looked at is capped. + let junk = |n: u32| (1000..1000 + n).map(|id| labl(b"labl", id, "n")).collect::>(); + let started = std::time::Instant::now(); + // the real label comes first: read + let mut subs = vec![labl(b"labl", 1, "Real")]; + subs.extend(junk(100_000)); + let m = read(&wav(&[fmt(44100), cue(&[(1, 44100)]), adtl(&subs)])); + assert_eq!(m[0].comment, "Real"); + // it comes after more records than are looked at: not read, and the marker is still there + let mut subs = junk(MAX_ADTL_RECORDS as u32 + 10); + subs.push(labl(b"labl", 1, "Too late")); + let m = read(&wav(&[fmt(44100), cue(&[(1, 44100)]), adtl(&subs)])); + assert_eq!((m.len(), m[0].comment.as_str()), (1, "")); + assert!(started.elapsed() < std::time::Duration::from_secs(5), "took {:?}", started.elapsed()); + } + + #[test] + fn metadata_keeps_original_coordinates_and_separate_sources() { + let xmp = br#""#; + let file = wav(&[fmt(44100), cue(&[(1, 1)]), adtl(&[labl(b"labl", 1, "Same")]), chunk(b"_PMX", xmp)]); + let data = super::read(&file); + assert_eq!(data.cue[0].start, 1.0); + assert_eq!(data.cue[0].rate, 44100.0); + assert_eq!(data.xmp[0].start, 1.0); + assert_eq!(data.xmp[0].duration, 2.0); + assert_eq!(data.cue.len(), 1); + assert_eq!(data.xmp.len(), 1); + } + + #[test] + fn repeated_unterminated_entities_preserve_text_with_bounded_lookahead() { + let text = "&".repeat(100_000); + assert_eq!(unescape(&text), text); + assert_eq!(unescape("A A & &unknown;"), "A A & &unknown;"); + } + #[test] + fn repeated_cue_ids_cannot_multiply_label_storage() { + let text = "x".repeat(MAX_LABEL_BYTES); + let points: Vec<_> = (0..MAX_MARKERS).map(|i| (1, i as u32)).collect(); + let file = wav(&[fmt(44100), cue(&points), adtl(&[labl(b"labl", 1, &text)])]); + let metadata = super::read(&file); + assert_eq!(metadata.cue.len(), MAX_MARKERS); + assert_eq!(metadata.cue.iter().map(|m| m.name.len()).sum::(), MAX_LABEL_TOTAL_BYTES); + assert!(metadata.limited); + assert_eq!(metadata.cue[0].name, text); + assert!(metadata.cue[MAX_MARKERS - 1].name.is_empty()); + assert_eq!(metadata.cue[MAX_MARKERS - 1].start, (MAX_MARKERS - 1) as f64); + } + + #[test] + fn cue_name_map_and_single_label_are_bounded() { + let count = MAX_LABEL_TOTAL_BYTES / MAX_LABEL_BYTES + 2; + let text = "x".repeat(MAX_LABEL_BYTES); + let points: Vec<_> = (0..count as u32).map(|i| (i, i)).collect(); + let labels: Vec<_> = (0..count as u32).map(|i| labl(b"labl", i, &text)).collect(); + let metadata = super::read(&wav(&[fmt(44100), cue(&points), adtl(&labels)])); + assert!(metadata.limited); + assert_eq!(metadata.cue.iter().map(|m| m.name.len()).sum::(), MAX_LABEL_TOTAL_BYTES); + let oversized = "x".repeat(MAX_LABEL_BYTES + 1); + let metadata = super::read(&wav(&[fmt(44100), cue(&[(1, 3)]), adtl(&[labl(b"labl", 1, &oversized)])])); + assert!(metadata.limited); + assert_eq!(metadata.cue[0].start, 3.0); + assert!(metadata.cue[0].name.is_empty()); + } + + #[test] + fn empty_chunk_inventory_is_bounded() { + let empty = chunk(b"junk", &[]); + let many = empty.repeat(MAX_CHUNKS + 1); + let metadata = super::read(&wav(&[many, fmt(44100), cue(&[(1, 1)])])); + assert!(metadata.limited); + assert!(metadata.cue.is_empty()); + } + + #[test] + fn entity_lookahead_keeps_previous_semicolon_boundary() { + assert_eq!(unescape("A"), "A"); + assert_eq!(unescape("A"), "A"); + } + #[test] + fn xmp_quotes_spacing_and_attribute_boundaries() { + let packet = br#""#; + let markers = xmp_markers(packet, 48000.0); + assert_eq!(markers.len(), 1); + assert_eq!((markers[0].start, markers[0].duration, markers[0].rate), (24000.0, 12000.0, 48000.0)); + assert_eq!(markers[0].name, "single & quoted"); + let spoof = r#" label=" xmpDM:startTime='7'" otherxmpDM:startTime='8' xmpDM:startTime = '9'/>"#; + assert_eq!(xmp_value(spoof, "xmpDM:startTime").as_deref(), Some("9")); + assert!(xmp_value(" xmpDM:startTime='unterminated", "xmpDM:startTime").is_none()); + } + + #[test] + fn xmp_frame_counts_preserve_explicit_and_mixed_rates() { + let packet = br#""#; + let markers = xmp_markers(packet, 44100.0); + assert_eq!(markers.len(), 3); + assert_eq!((markers[0].start, markers[0].rate), (15.0, 24.0)); + assert!((markers[0].duration / markers[0].rate - 0.5005).abs() < 1e-12); + assert_eq!((markers[1].start, markers[1].duration, markers[1].rate), (48000.0, 24000.0, 48000.0)); + assert_eq!((markers[2].start, markers[2].duration, markers[2].rate), (15.0, 12.0, 24.0)); + for invalid in ["15f0", "15f24s0", "15fNaN", "15finf", "NaN", "1e999", "15f24junk"] { + assert!(xmp_count(invalid, 48000.0).is_none(), "{invalid}"); + } + } +} diff --git a/crates/audio/src/write/mod.rs b/crates/audio/src/write/mod.rs index ce4b90773..a2043d279 100644 --- a/crates/audio/src/write/mod.rs +++ b/crates/audio/src/write/mod.rs @@ -12,7 +12,7 @@ mod pcm; #[cfg(feature = "vorbis-encode")] mod vorbis; #[cfg(feature = "vorbis-encode")] -pub use vorbis::{OGG_BITRATES, encode as encode_ogg}; +pub use vorbis::{OGG_BITRATES, OggVorbisOptions, encode as encode_ogg, encode_ogg_vorbis, quality_for_bitrate}; #[cfg(feature = "pcm-encode")] mod raw; #[cfg(feature = "pcm-encode")] @@ -66,15 +66,22 @@ pub fn encode(buf: &AudioBuffer, opts: &EncodeOptions) -> Result> { EncodeFormat::Aac | EncodeFormat::Alac | EncodeFormat::Mp3 | EncodeFormat::Ac3 | EncodeFormat::Eac3 => coded::encode(buf, opts), #[cfg(feature = "vorbis-encode")] EncodeFormat::Ogg => { - if opts.layout.is_some_and(|layout| !EncodeFormat::Ogg.supports_layout(&layout)) { - return Err(AudioError::Unsupported("Vorbis requires a mono or stereo speaker layout".into())); + let count = u16::try_from(buf.channels.len()).unwrap_or(u16::MAX); + let layout = opts.layout.unwrap_or_else(|| ChannelLayout::default_for(count)); + if layout.channels() != count { + return Err(AudioError::Unsupported(format!("a {}-channel layout for {count} channels", layout.channels()))); } - let kbps = match opts.rate_control { - None => 192, - Some(crate::RateControl::Bitrate(bps)) => bps / 1000, - Some(crate::RateControl::Quality(_)) => return Err(AudioError::Unsupported("Ogg export takes a target bitrate".into())), + let map = vorbis::vorbis_map(&layout)?; + let quality = match opts.rate_control { + None => vorbis::quality_for_bitrate(192, buf.channels.len()), + Some(crate::RateControl::Bitrate(bps)) => { + vorbis::quality_for_bitrate((bps / 1000).clamp(OGG_BITRATES.0, OGG_BITRATES.1), buf.channels.len()) + } + Some(crate::RateControl::Quality(q)) if q.is_nan() => return Err(AudioError::Encode("the Vorbis quality is NaN".into())), + Some(crate::RateControl::Quality(q)) => -1.0 + 11.0 * q.clamp(0.0, 1.0), }; - vorbis::encode(buf, kbps) + let planar: Vec<&[f32]> = map.iter().filter_map(|&i| buf.channels.get(i).map(Vec::as_slice)).collect(); + vorbis::encode_planar(buf.sample_rate, &planar, &OggVorbisOptions::new(quality)) } #[allow(unreachable_patterns)] f => Err(AudioError::Unsupported(format!("writing {f:?}"))), @@ -106,7 +113,7 @@ const fn extensions() -> ([&'static str; 13], usize) { add!(cfg!(feature = "mp3-encode"), "mp3"); add!(cfg!(feature = "ac3-encode"), "ac3"); add!(cfg!(feature = "eac3-encode"), "eac3", "ec3"); - add!(cfg!(all(feature = "vorbis-encode", not(target_arch = "wasm32"))), "ogg"); + add!(cfg!(feature = "vorbis-encode"), "ogg"); (all, n) } @@ -148,7 +155,7 @@ pub fn encode_format_for(name: &str) -> Option { "ac3" => Some(EncodeFormat::Ac3), #[cfg(feature = "eac3-encode")] "eac3" | "ec3" => Some(EncodeFormat::Eac3), - #[cfg(all(feature = "vorbis-encode", not(target_arch = "wasm32")))] + #[cfg(feature = "vorbis-encode")] "ogg" => Some(EncodeFormat::Ogg), _ => None, } diff --git a/crates/audio/src/write/vorbis.rs b/crates/audio/src/write/vorbis.rs index f56d01148..6c19ada77 100644 --- a/crates/audio/src/write/vorbis.rs +++ b/crates/audio/src/write/vorbis.rs @@ -1,83 +1,198 @@ -//! Ogg Vorbis encoding (RFC 5334): the reference libvorbis encoder, built from its C sources by -//! `vorbis_rs`. Stereo or mono, variable bitrate at the requested kbit/s (64–500), with a -//! fallback to a quality preset when the bitrate is not supported for the rate and channel count. +//! Ogg Vorbis files (Vorbis I §A, RFC 5334) from the pure-Rust [`crate::vorbis::encoder`] and +//! craft-media's Ogg page writer: 1 to 8 channels in Vorbis channel order, a quality from -1 to +//! 10 (variable bit rate), and the tags in the comment header. Builds for every target. -use crate::{AudioBuffer, AudioError, Result}; +use craft_media::ogg::{PageWriter, vorbis_comment}; -/// The bitrates an Ogg Vorbis export is asked for, in kbit/s. +use crate::vorbis::encoder::{DEFAULT_QUALITY, Encoder, EncoderConfig, Error as VorbisError}; +use crate::{AudioBuffer, AudioError, ChannelLayout, Result}; + +/// The bit rates [`encode_ogg`](crate::encode_ogg) takes, in kbit/s. Vorbis is variable: the bit rate picks a +/// quality ([`quality_for_bitrate`]), and the file comes out near it for typical music. pub const OGG_BITRATES: (u32, u32) = (64, 500); -/// The block size the encoder processes at a time: small enough that memory stays bounded -/// regardless of file length. -#[cfg(not(target_arch = "wasm32"))] -const BLOCK: usize = 4096; +/// The serial number of the stream when the options do not set one (`LCRF`): one stream per file, +/// so a constant keeps files byte-for-byte reproducible. +const SERIAL: u32 = 0x4C43_5246; -#[cfg(not(target_arch = "wasm32"))] -fn encode_error(e: vorbis_rs::VorbisError) -> AudioError { - AudioError::Encode(e.to_string()) +/// Samples per channel fed to the encoder at a time. +const CHUNK: usize = 4096; + +/// What [`encode_ogg_vorbis`] writes besides the audio. +/// +/// `#[non_exhaustive]`: build it with `OggVorbisOptions::default()` or +/// [`OggVorbisOptions::new`] and the `with_*` methods. +#[non_exhaustive] +#[derive(Debug, Clone, PartialEq)] +pub struct OggVorbisOptions { + /// -1 (smallest) to 10 (best); clamped, and NaN or infinite reads as + /// [`DEFAULT_QUALITY`] (6). + pub quality: f32, + /// The comment header's vendor string. + pub vendor: String, + /// `NAME=value` comments (Vorbis I §5), in order. A name that is empty, or holds `=` or a byte + /// outside 0x20–0x7D, is left out. + pub tags: Vec<(String, String)>, + /// The logical stream's serial number. + pub serial: u32, } -/// Encode `buf` as an Ogg Vorbis stream at about `kbps` (variable, as Vorbis is): stereo, or mono -/// for a mono buffer. A bitrate the encoder has no setup for at that rate and channel count falls -/// back to a quality that suits it. `Err` says why there is nothing. -#[cfg(not(target_arch = "wasm32"))] -pub fn encode(buf: &AudioBuffer, kbps: u32) -> Result> { - use std::num::{NonZeroU8, NonZeroU32}; - - let channels = buf.channels.len(); - if !(1..=2).contains(&channels) { - return Err(AudioError::Unsupported("Ogg Vorbis export requires mono or stereo audio".into())); +impl Default for OggVorbisOptions { + fn default() -> Self { + OggVorbisOptions { quality: DEFAULT_QUALITY, vendor: "craft-audio".into(), tags: Vec::new(), serial: SERIAL } } - if buf.channels.iter().any(|c| c.len() != buf.frames() || c.iter().any(|s| !s.is_finite())) { +} + +impl OggVorbisOptions { + /// The defaults at `quality`. + pub fn new(quality: f32) -> Self { + OggVorbisOptions { quality, ..Self::default() } + } + + /// Sets [`quality`](Self::quality). + pub fn with_quality(mut self, quality: f32) -> Self { + self.quality = quality; + self + } + + /// Sets [`vendor`](Self::vendor). + pub fn with_vendor(mut self, vendor: impl Into) -> Self { + self.vendor = vendor.into(); + self + } + + /// Adds a comment ([`tags`](Self::tags)). + pub fn with_tag(mut self, name: impl Into, value: impl Into) -> Self { + self.tags.push((name.into(), value.into())); + self + } + + /// Sets [`serial`](Self::serial). + pub fn with_serial(mut self, serial: u32) -> Self { + self.serial = serial; + self + } +} + +fn encode_error(e: VorbisError) -> AudioError { + AudioError::Encode(format!("Vorbis: {e}")) +} + +/// Encode `buf` as an Ogg Vorbis file with `opts`. The channels are in Vorbis channel order +/// (Vorbis I §4.3.9: L C R for three, L C R Ls Rs LFE for 5.1, L C R Ls Rs Rl Rr LFE for 7.1). +/// The last granule position is the buffer's length, so a decoder returns exactly its frames. +/// +/// ``` +/// use craft_audio::{AudioBuffer, OggVorbisOptions, encode_ogg_vorbis}; +/// +/// let buf = AudioBuffer::new(48_000, 6, 4_800); // 0.1 s of 5.1 silence +/// let opts = OggVorbisOptions::new(4.0).with_tag("TITLE", "Silence"); +/// let ogg = encode_ogg_vorbis(&buf, &opts).unwrap(); +/// assert!(ogg.starts_with(b"OggS")); +/// ``` +pub fn encode_ogg_vorbis(buf: &AudioBuffer, opts: &OggVorbisOptions) -> Result> { + let planar: Vec<&[f32]> = buf.channels.iter().map(Vec::as_slice).collect(); + encode_planar(buf.sample_rate, &planar, opts) +} + +/// [`encode_ogg_vorbis`] over channels already in Vorbis order. +pub(crate) fn encode_planar(sample_rate: u32, planar: &[&[f32]], opts: &OggVorbisOptions) -> Result> { + if !(1..=8).contains(&planar.len()) { + return Err(AudioError::Unsupported(format!("Ogg Vorbis holds 1 to 8 channels, not {}", planar.len()))); + } + let frames = planar.first().map_or(0, |c| c.len()); + if planar.iter().any(|c| c.len() != frames || c.iter().any(|s| !s.is_finite())) { return Err(AudioError::Encode("Vorbis samples must be finite and channels equally long".into())); } - let (Some(rate), Some(count)) = (NonZeroU32::new(buf.sample_rate), NonZeroU8::new(channels as u8)) else { + if sample_rate == 0 { return Err(AudioError::Encode("there is no audio to encode".into())); - }; - let target = NonZeroU32::new(kbps.clamp(OGG_BITRATES.0, OGG_BITRATES.1) * 1000).unwrap_or(NonZeroU32::MIN); - - let encode = |strategy: vorbis_rs::VorbisBitrateManagementStrategy| -> Result> { - let mut encoder = vorbis_rs::VorbisEncoderBuilder::new_with_serial(rate, count, Vec::new(), 0x4C43_5246) - .bitrate_management_strategy(strategy) - .build() - .map_err(encode_error)?; - let frames = buf.frames(); - let mut at = 0usize; - while at < frames { - let end = (at + BLOCK).min(frames); - let block: Vec<&[f32]> = buf.channels.iter().take(channels).map(|side| side.get(at..end).unwrap_or_default()).collect(); - encoder.encode_audio_block(&block).map_err(encode_error)?; - at = end; + } + let mut enc = Encoder::new(EncoderConfig::new(sample_rate, planar.len(), opts.quality)).map_err(encode_error)?; + let mut pages = PageWriter::new(opts.serial); + let tags: Vec<(&str, &str)> = opts.tags.iter().map(|(k, v)| (k.as_str(), v.as_str())).collect(); + let mut out = pages.header(&enc.identification()); + out.extend(pages.header(&vorbis_comment(&opts.vendor, &tags))); + out.extend(pages.header(&enc.setup())); + // The first audio packet (which outputs nothing) gets a page of its own: a stream whose first + // audio page is also its last would leave decoders to choose between start and end trimming + // (Vorbis I §A.2), and ffmpeg then plays the padding. + let mut first = true; + let mut add = |packets: Vec<(Vec, i64)>, out: &mut Vec| { + for (packet, granule) in packets { + out.extend(pages.packet(&packet, granule)); + if std::mem::take(&mut first) { + out.extend(pages.flush()); + } } - encoder.finish().map_err(encode_error) }; + let mut at = 0; + loop { + let end = (at + CHUNK).min(frames); + let chunk: Vec<&[f32]> = planar.iter().map(|c| c.get(at..end).unwrap_or_default()).collect(); + add(enc.encode(&chunk).map_err(encode_error)?, &mut out); + at = end; + if at >= frames { + break; + } + } + add(enc.finish().map_err(encode_error)?, &mut out); + out.extend(pages.finish(None)); + Ok(out) +} - // Convert Rust unwinds to an error. This does not catch native crashes or aborts. - let attempt = |strategy| { - std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| encode(strategy))) - .unwrap_or_else(|_| Err(AudioError::Encode("the Vorbis encoder failed".into()))) - }; - match attempt(vorbis_rs::VorbisBitrateManagementStrategy::Vbr { target_bitrate: target }) { - Ok(bytes) => Ok(bytes), - Err(_) => attempt(vorbis_rs::VorbisBitrateManagementStrategy::QualityVbr { target_quality: 0.6 }) - .map_err(|e| AudioError::Encode(format!("the Vorbis encoder refused the audio: {e}"))), +/// Nominal kbit/s per channel at each whole quality from -1 to 10: this encoder's rate on +/// music-like test signals at 48 kHz (harmonic chords with noise), averaged. +const KBPS_PER_CHANNEL: [f32; 12] = [18.0, 23.0, 31.0, 43.0, 58.0, 75.0, 89.0, 100.0, 111.0, 121.0, 130.0, 140.0]; + +/// The quality whose nominal bit rate is `kbps` kbit/s for `channels` channels (clamped to -1..=10). +/// Vorbis is variable: the bit rate of a file depends on the audio. +pub fn quality_for_bitrate(kbps: u32, channels: usize) -> f32 { + let per = kbps as f32 / channels.max(1) as f32; + let mut q = -1.0; + for pair in KBPS_PER_CHANNEL.windows(2) { + if let [lo, hi] = *pair + && per <= hi + { + return q + ((per - lo) / (hi - lo)).clamp(0.0, 1.0); + } + q += 1.0; + } + 10.0 +} + +/// Encode `buf` as an Ogg Vorbis stream near `kbps` kbit/s ([`OGG_BITRATES`], clamped): the +/// quality [`quality_for_bitrate`] gives, no tags. 1 to 8 channels in Vorbis channel order. +/// `Err` says why there is nothing. +pub fn encode(buf: &AudioBuffer, kbps: u32) -> Result> { + let kbps = kbps.clamp(OGG_BITRATES.0, OGG_BITRATES.1); + encode_ogg_vorbis(buf, &OggVorbisOptions::new(quality_for_bitrate(kbps, buf.channels.len()))) +} + +/// WAV-order speakers ([`crate::EncodeFormat::layout_masks`]) to Vorbis order, per channel count. +const VORBIS_ORDER: [&[usize]; 8] = + [&[0], &[0, 1], &[0, 2, 1], &[0, 1, 2, 3], &[0, 2, 1, 3, 4], &[0, 2, 1, 4, 5, 3], &[0, 2, 1, 5, 6, 4, 3], &[0, 2, 1, 6, 7, 4, 5, 3]]; + +/// For each Vorbis channel, the channel of audio in `layout` it comes from. +pub(crate) fn vorbis_map(layout: &ChannelLayout) -> Result> { + let n = layout.channels(); + let mask = crate::EncodeFormat::Ogg.layout_masks(n).first().copied(); + let order = usize::from(n).checked_sub(1).and_then(|i| VORBIS_ORDER.get(i)).copied(); + match (mask, order) { + (Some(mask), Some(order)) => crate::layout::channel_map(mask, Some(order), layout) + .ok_or_else(|| AudioError::Unsupported(format!("Vorbis has no layout of {n} channels with these speakers"))), + _ => Err(AudioError::Unsupported(format!("Ogg Vorbis holds 1 to 8 channels, not {n}"))), } } -/// Ogg Vorbis encoding is unavailable on WASM; returns an unsupported-format error. -#[cfg(target_arch = "wasm32")] -pub fn encode(_buf: &AudioBuffer, _kbps: u32) -> Result> { - Err(AudioError::Unsupported("Ogg Vorbis files are exported by the desktop app".into())) -} - -#[cfg(all(test, not(target_arch = "wasm32")))] +#[cfg(test)] mod tests { #![allow(clippy::unwrap_used)] - use super::*; + use super::{OGG_BITRATES, OggVorbisOptions, encode, encode_ogg_vorbis, quality_for_bitrate, vorbis_map}; + use crate::{AudioBuffer, AudioError, ChannelLayout}; #[test] fn rejects_unsupported_or_malformed_channels_without_dropping_audio() { - for channels in [0, 3, 6] { + for channels in [0, 9] { assert!(matches!(encode(&AudioBuffer::new(48_000, channels, 16), 192), Err(AudioError::Unsupported(_)))); } let ragged = AudioBuffer { sample_rate: 48_000, channels: vec![vec![0.0; 16], vec![0.0; 15]] }; @@ -85,46 +200,68 @@ mod tests { assert!(encode(&AudioBuffer::new(0, 1, 16), 192).is_err()); let invalid = AudioBuffer { sample_rate: 48_000, channels: vec![vec![f32::NAN]] }; assert!(encode(&invalid, 192).is_err()); + let infinite = AudioBuffer { sample_rate: 48_000, channels: vec![vec![f32::INFINITY]] }; + assert!(encode(&infinite, 192).is_err()); } #[test] - fn generic_and_direct_encoders_write_mono_and_stereo_ogg() { - for channels in [1, 2] { + fn generic_and_direct_encoders_write_one_to_eight_channels() { + for channels in 1..=8 { let audio = AudioBuffer::new(48_000, channels, 512); let opts = crate::EncodeOptions::new(crate::EncodeFormat::Ogg); assert!(opts.check(48_000, channels).is_ok()); assert_eq!(opts.bitrate(48_000, channels).unwrap(), None); let direct = encode(&audio, 192).unwrap(); - let generic = crate::encode(&audio, &opts).unwrap(); assert!(direct.starts_with(b"OggS")); - assert_eq!(direct, generic); + // silence: the channel order changes nothing, so the generic path writes the same file + let generic = crate::encode(&audio, &opts.with_bitrate(192_000)).unwrap(); + assert_eq!(direct, generic, "{channels} channels"); #[cfg(feature = "decode")] { let (info, decoded) = crate::decode(&generic, Some("ogg")).unwrap(); assert_eq!(info.channels, channels); - assert_eq!(decoded.num_channels(), channels); assert_eq!(decoded.frames(), audio.frames()); } } } - #[cfg(feature = "decode")] + #[test] - fn distinct_signals_survive_full_and_partial_blocks() { - for frames in [4096, 4097, 8193] { - let channels: Vec> = [440.0f32, 880.0] - .into_iter() - .map(|hz| (0..frames).map(|i| (std::f32::consts::TAU * hz * i as f32 / 48_000.0).sin() * 0.4).collect()) - .collect(); - let audio = AudioBuffer { sample_rate: 48_000, channels }; - let bytes = encode(&audio, 192).unwrap(); - let (_, decoded) = crate::decode(&bytes, Some("ogg")).unwrap(); - assert_eq!(decoded.frames(), frames); - assert_eq!(decoded.channels.len(), 2); - for (expected, actual) in audio.channels.iter().zip(&decoded.channels) { - let mse = expected.iter().zip(actual).map(|(a, b)| (a - b).powi(2)).sum::() / frames as f32; - // Vorbis is lossy; this tolerates codec error but rejects silence or swapped channels. - assert!(mse.sqrt() < 0.08, "{frames} frames: RMS error {}", mse.sqrt()); - } + fn the_quality_follows_the_bit_rate() { + assert_eq!(quality_for_bitrate(0, 2), -1.0); + assert_eq!(quality_for_bitrate(10_000, 2), 10.0); + assert!((quality_for_bitrate(200, 2) - 6.0).abs() < 1e-6); + assert!((quality_for_bitrate(600, 6) - 6.0).abs() < 1e-6); + let mut last = -2.0; + for kbps in OGG_BITRATES.0..=OGG_BITRATES.1 { + let q = quality_for_bitrate(kbps, 2); + assert!(q >= last && (-1.0..=10.0).contains(&q)); + last = q; } + assert!(quality_for_bitrate(192, 0).is_finite()); + } + + #[test] + fn layouts_map_to_vorbis_order() { + // 5.1 in WAV order (L R C LFE Ls Rs) is written L C R Ls Rs LFE + assert_eq!(vorbis_map(&ChannelLayout::SURROUND_5_1).unwrap(), [0, 2, 1, 4, 5, 3]); + assert_eq!(vorbis_map(&ChannelLayout::SURROUND_5_1_SIDE).unwrap(), [0, 2, 1, 4, 5, 3]); + assert_eq!(vorbis_map(&ChannelLayout::from_mask(0x63F)).unwrap(), [0, 2, 1, 6, 7, 4, 5, 3]); + assert_eq!(vorbis_map(&ChannelLayout::from_mask(0x7)).unwrap(), [0, 2, 1]); + assert_eq!(vorbis_map(&ChannelLayout::STEREO).unwrap(), [0, 1]); + assert!(vorbis_map(&ChannelLayout::from_mask(0xC)).is_err()); + assert!(vorbis_map(&ChannelLayout::unordered(9)).is_err()); + assert!(vorbis_map(&ChannelLayout::unordered(0)).is_err()); + } + + #[test] + fn options_build_and_tags_reach_the_comment_header() { + let opts = OggVorbisOptions::default().with_quality(2.0).with_vendor("v").with_tag("TITLE", "t").with_tag("bad=name", "x").with_serial(9); + assert_eq!((opts.quality, opts.vendor.as_str(), opts.serial, opts.tags.len()), (2.0, "v", 9, 2)); + let ogg = encode_ogg_vorbis(&AudioBuffer::new(44_100, 1, 100), &opts).unwrap(); + let file = craft_media::ogg::open(&ogg).unwrap(); + let s = &file.streams[0]; + assert_eq!((s.serial, s.saw_eos, s.headers.len()), (9, true, 3)); + assert_eq!(s.headers[1], craft_media::ogg::vorbis_comment("v", &[("TITLE", "t")])); + assert_eq!(s.last_granule, Some(100)); } } diff --git a/crates/audio/tests/audio/encode_format.rs b/crates/audio/tests/audio/encode_format.rs index e91f93623..72749df04 100644 --- a/crates/audio/tests/audio/encode_format.rs +++ b/crates/audio/tests/audio/encode_format.rs @@ -20,8 +20,8 @@ fn only_formats_the_encoder_writes_are_named() { } #[test] -fn ogg_names_follow_the_native_feature() { - let built = cfg!(all(feature = "vorbis-encode", not(target_arch = "wasm32"))); +fn ogg_names_follow_the_feature() { + let built = cfg!(feature = "vorbis-encode"); assert_eq!(EncodeFormat::Ogg.is_built(), built); assert_eq!(encode_format_for("ogg"), built.then_some(EncodeFormat::Ogg)); assert_eq!(craft_audio::ENCODE_EXTENSIONS.contains(&"ogg"), built); diff --git a/crates/audio/tests/codecs/file_formats.rs b/crates/audio/tests/codecs/file_formats.rs index 84169f665..82edf72aa 100644 --- a/crates/audio/tests/codecs/file_formats.rs +++ b/crates/audio/tests/codecs/file_formats.rs @@ -527,7 +527,7 @@ fn the_format_queries_agree_with_the_writer() { // The most channels is the most: one more is refused, by the layouts of a format that has // a fixed set and by the writer of one that has not. let most: Vec = EncodeFormat::ALL.iter().map(|f| f.max_channels()).collect(); - assert_eq!(most, [32_767, 32_767, 8, 8, 8, 2, 6, 8, 2]); + assert_eq!(most, [32_767, 32_767, 8, 8, 8, 2, 6, 8, 8]); for &format in EncodeFormat::ALL { let most = format.max_channels(); match format.fixed_layouts(most) { diff --git a/crates/audio/tests/codecs/main.rs b/crates/audio/tests/codecs/main.rs index c1c9e2131..41829be5c 100644 --- a/crates/audio/tests/codecs/main.rs +++ b/crates/audio/tests/codecs/main.rs @@ -20,3 +20,5 @@ mod opus; mod opusenc; mod routing_oracle; mod track_decoder; +#[cfg(feature = "vorbis-encode")] +mod vorbis_encoder_oracle; diff --git a/crates/audio/tests/codecs/vorbis_encoder_oracle.rs b/crates/audio/tests/codecs/vorbis_encoder_oracle.rs new file mode 100644 index 000000000..342f4fd9a --- /dev/null +++ b/crates/audio/tests/codecs/vorbis_encoder_oracle.rs @@ -0,0 +1,131 @@ +//! ffmpeg as an external oracle for the pure-Rust Vorbis encoder: both of its decoders (its own +//! and libvorbis) read every file exactly as long as the input, surround comes out on the right +//! speakers, and the tags reach ffprobe. Skipped when ffmpeg is not installed. + +use std::path::Path; +use std::process::Command; + +use craft_audio::{AudioBuffer, ChannelLayout, EncodeFormat, EncodeOptions, OggVorbisOptions, encode, encode_ogg_vorbis}; + +/// A different tone on every channel, so each output channel can be traced to its input. +fn tones(rate: u32, channels: usize, frames: usize) -> AudioBuffer { + let channels = (0..channels) + .map(|c| { + let f = 220.0 * (c + 1) as f32; + (0..frames).map(|i| (std::f32::consts::TAU * f * i as f32 / rate as f32).sin() * 0.3).collect() + }) + .collect(); + AudioBuffer { sample_rate: rate, channels } +} + +/// Decodes `file` with ffmpeg's `decoder`; planar output in ffmpeg's (WAV) channel order. +fn ffmpeg_decode(ff: &Path, decoder: &str, file: &[u8], channels: usize, tag: &str) -> Vec> { + let path = crate::testkit::fixtures_dir("vorbis-encoder").join(format!("{tag}.ogg")); + std::fs::write(&path, file).unwrap(); + let out = Command::new(ff) + .args(["-v", "error", "-xerror", "-c:a", decoder, "-i"]) + .arg(&path) + .args(["-f", "f32le", "-c:a", "pcm_f32le", "-"]) + .output() + .unwrap(); + assert!(out.status.success() && out.stderr.is_empty(), "{tag} {decoder}: {}", String::from_utf8_lossy(&out.stderr)); + let pcm: Vec = out.stdout.as_chunks::<4>().0.iter().map(|&b| f32::from_le_bytes(b)).collect(); + (0..channels).map(|c| pcm.iter().skip(c).step_by(channels).copied().collect()).collect() +} + +/// ffmpeg's decoders to check: its own `vorbis` always, and `libvorbis` when ffmpeg has it. +fn decoders(ff: &Path) -> Vec<&'static str> { + let list = + Command::new(ff).args(["-hide_banner", "-decoders"]).output().map(|o| String::from_utf8_lossy(&o.stdout).into_owned()).unwrap_or_default(); + if list.lines().any(|l| l.split_whitespace().nth(1) == Some("libvorbis")) { + vec!["vorbis", "libvorbis"] + } else { + eprintln!("SKIPPED (libvorbis decoder): this ffmpeg has no libvorbis; checking its own decoder only"); + vec!["vorbis"] + } +} + +fn correlation(a: &[f32], b: &[f32]) -> f64 { + let (mut ab, mut aa, mut bb) = (0f64, 0f64, 0f64); + for (&x, &y) in a.iter().zip(b) { + ab += f64::from(x) * f64::from(y); + aa += f64::from(x) * f64::from(x); + bb += f64::from(y) * f64::from(y); + } + ab / (aa * bb).sqrt().max(1e-20) +} + +/// Vorbis order (§4.3.9) of the WAV-order layouts the test writes, per channel count. +const VORBIS_ORDER: [&[usize]; 8] = + [&[0], &[0, 1], &[0, 2, 1], &[0, 1, 2, 3], &[0, 2, 1, 3, 4], &[0, 2, 1, 4, 5, 3], &[0, 2, 1, 5, 6, 4, 3], &[0, 2, 1, 6, 7, 4, 5, 3]]; + +/// 1 to 8 channels through `encode` with a WAV-order layout, at exactly the input's length out of +/// both decoders: ffmpeg's own puts every channel back on its WAV speaker, and libvorbis (through +/// ffmpeg, which passes its order through) gives the file's Vorbis order. +#[test] +fn every_channel_count_decodes_on_the_right_speakers() { + let ff = require_ffmpeg!(); + for channels in 1..=8u16 { + let layout = if channels == 7 { ChannelLayout::from_mask(0x70F) } else { ChannelLayout::default_for(channels) }; + let buf = tones(48_000, channels.into(), 24_000 + 77); + let opts = EncodeOptions::new(EncodeFormat::Ogg).with_layout(layout).with_quality(0.6); + let file = encode(&buf, &opts).unwrap(); + for decoder in decoders(&ff) { + let out = ffmpeg_decode(&ff, decoder, &file, channels.into(), &format!("tones{channels}-{decoder}")); + for (j, got) in out.iter().enumerate() { + let c = if decoder == "libvorbis" { VORBIS_ORDER[usize::from(channels) - 1][j] } else { j }; + assert_eq!(got.len(), buf.frames(), "{channels} ch {decoder}: channel {j} length"); + let best = (0..buf.channels.len()).max_by(|&a, &b| { + correlation(&buf.channels[a][2048..], &got[2048..]).total_cmp(&correlation(&buf.channels[b][2048..], &got[2048..])) + }); + assert_eq!(best, Some(c), "{channels} ch {decoder}: output channel {c} carries input {best:?}"); + let r = correlation(&buf.channels[c][2048..], &got[2048..]); + assert!(r > 0.95, "{channels} ch {decoder}: channel {c} correlation {r:.3}"); + } + } + } +} + +/// The comment header's tags, UTF-8 values included, are what ffprobe reports; a name that is not +/// a valid field name is left out. +#[test] +fn tags_reach_ffprobe() { + let Some(probe) = crate::common::tool("ffprobe").or_else(|| crate::common::skip("ffprobe not found")) else { return }; + let opts = OggVorbisOptions::new(3.0) + .with_tag("TITLE", "SmörgÃ¥sbord ✓") + .with_tag("ARTIST", "Craft") + .with_tag("bad=name", "dropped") + .with_vendor("craft-audio test"); + let file = encode_ogg_vorbis(&tones(44_100, 2, 44_100), &opts).unwrap(); + let path = crate::testkit::fixtures_dir("vorbis-encoder").join("tags.ogg"); + std::fs::write(&path, &file).unwrap(); + let out = Command::new(probe).args(["-v", "error", "-of", "json", "-show_streams", "-show_format"]).arg(&path).output().unwrap(); + assert!(out.status.success(), "{}", String::from_utf8_lossy(&out.stderr)); + let j: serde_json::Value = serde_json::from_slice(&out.stdout).unwrap(); + let s = &j["streams"][0]; + assert_eq!((s["codec_name"].as_str(), s["channels"].as_u64(), s["sample_rate"].as_str()), (Some("vorbis"), Some(2), Some("44100"))); + let tags = &s["tags"]; + assert_eq!((tags["TITLE"].as_str(), tags["ARTIST"].as_str()), (Some("SmörgÃ¥sbord ✓"), Some("Craft"))); + assert!(tags.as_object().unwrap().keys().all(|k| !k.contains("bad")), "{tags}"); + assert_eq!(j["format"]["duration"].as_str().map(|d| d.parse::().unwrap()), Some(1.0)); +} + +/// The lowest and highest qualities and the bit-rate path decode at the exact length, mono at +/// 44.1 kHz and stereo at 22.05 kHz. +#[test] +fn quality_extremes_and_rates_decode_exactly() { + let ff = require_ffmpeg!(); + for (rate, channels, quality) in [(44_100u32, 1usize, -1.0f32), (44_100, 1, 10.0), (22_050, 2, -1.0), (22_050, 2, 10.0), (96_000, 2, 5.0)] { + let buf = tones(rate, channels, rate as usize / 2 + 13); + let file = encode_ogg_vorbis(&buf, &OggVorbisOptions::new(quality)).unwrap(); + for decoder in decoders(&ff) { + let out = ffmpeg_decode(&ff, decoder, &file, channels, &format!("q{quality}-{rate}-{channels}-{decoder}")); + assert!(out.iter().all(|c| c.len() == buf.frames()), "q{quality} {rate} Hz {decoder}: {}", out[0].len()); + let r = correlation(&buf.channels[0][2048..], &out[0][2048..]); + assert!(r > 0.9, "q{quality} {rate} Hz {decoder}: correlation {r:.3}"); + } + } + let low = craft_audio::encode_ogg(&tones(48_000, 2, 48_000), craft_audio::OGG_BITRATES.0).unwrap(); + let high = craft_audio::encode_ogg(&tones(48_000, 2, 48_000), craft_audio::OGG_BITRATES.1).unwrap(); + assert!(low.len() < high.len(), "{} vs {} bytes", low.len(), high.len()); +} diff --git a/crates/cms/src/lib.rs b/crates/cms/src/lib.rs index 5e8a1e268..1ca07eecd 100644 --- a/crates/cms/src/lib.rs +++ b/crates/cms/src/lib.rs @@ -31,9 +31,11 @@ pub mod math; pub mod named; pub mod pipeline; pub mod profile; +pub mod spi1d; pub mod synth; pub mod transform; mod write; +pub mod write_v2; pub use builtin::Builtin; pub use clut::Clut; diff --git a/crates/cms/src/spi1d.rs b/crates/cms/src/spi1d.rs new file mode 100644 index 000000000..ea77bdc88 --- /dev/null +++ b/crates/cms/src/spi1d.rs @@ -0,0 +1,59 @@ +//! Sony Imageworks SPI1D tables, preserving EffectCraft's permissive header syntax. +//! +//! Unknown header lines and trailing text after the closing brace are ignored. Table +//! samples are not clamped or sanitized: callers own their numerical rendering policy. +//! The declared length is validated before storing samples; surplus samples are refused. + +/// Largest accepted table (the existing EffectCraft interpolation limit). +pub const MAX_ENTRIES: usize = 1 << 24; + +/// A one-dimensional RGB lookup table and its common input domain. +#[derive(Clone, Debug, PartialEq)] +pub struct Table { + pub values: Vec<[f32; 3]>, + pub domain: [f32; 2], +} + +/// Parse a SPI1D table with one or three components, expanding monochrome samples to RGB. +/// No allocation is based only on the claimed length. At most `MAX_ENTRIES` RGB entries +/// are retained. Missing final braces remain accepted, as in the source reader. +pub fn parse(text: &str) -> Option { + let (mut from, mut len, mut comps) = ([0.0f32, 1.0], 0usize, 1usize); + let mut values = Vec::new(); + let mut body = false; + let mut sample = [0.0; 3]; + let mut channel = 0; + for line in text.lines().map(str::trim).filter(|l| !l.is_empty()) { + if body { + if line.starts_with('}') { + break; + } + for t in line.split_whitespace() { + if values.len() == len { + return None; + } + sample[channel] = t.parse().ok()?; + channel += 1; + if channel == comps { + values.push(if comps == 1 { [sample[0]; 3] } else { sample }); + channel = 0; + } + } + continue; + } + let mut t = line.split_whitespace(); + match t.next()? { + "{" => { + if !(2..=MAX_ENTRIES).contains(&len) || !matches!(comps, 1 | 3) || from[1] <= from[0] || from.iter().any(|v| !v.is_finite()) { + return None; + } + body = true; + } + "From" => from = [t.next()?.parse().ok()?, t.next()?.parse().ok()?], + "Length" => len = t.next()?.parse().ok()?, + "Components" => comps = t.next()?.parse().ok()?, + _ => {} + } + } + (body && values.len() == len && channel == 0).then_some(Table { values, domain: from }) +} diff --git a/crates/cms/src/write_v2.rs b/crates/cms/src/write_v2.rs new file mode 100644 index 000000000..63933c801 --- /dev/null +++ b/crates/cms/src/write_v2.rs @@ -0,0 +1,195 @@ +//! Bounded ICC v2.1 matrix/TRC display-profile writer, extracted from TextureCraft. +//! Colour-space selection and transfer-curve sampling remain caller policy. + +const D50_XYZ: [f64; 3] = [0.9642, 1.0, 0.8249]; + +/// One common red/green/blue transfer curve. +#[derive(Clone, Copy)] +pub enum ToneCurve<'a> { + /// ICC gamma 1.0 (`curv` with one u8Fixed8 entry). + Linear, + /// Already quantized samples, inclusive endpoints, between 2 and 4096 entries. + Samples(&'a [u16]), +} + +/// Parameters for an ICC v2.1 RGB display profile. +pub struct MatrixProfile<'a> { + /// RGB-to-XYZ matrix already adapted to PCS D50, in row-major order. + pub matrix: [[f64; 3]; 3], + /// Common RGB transfer curve; its storage is shared by all three tags. + pub curve: ToneCurve<'a>, + /// ASCII description, at most 1024 bytes with no embedded NUL. + pub description: &'a str, + /// ASCII copyright text, at most 1024 bytes with no embedded NUL. + pub copyright: &'a str, + /// Raw ICC date fields: year, month, day, hour, minute, second. + pub date: [u16; 6], +} + +fn s15f16(v: f64) -> [u8; 4] { + ((v * 65536.0).round().clamp(f64::from(i32::MIN), f64::from(i32::MAX)) as i32).to_be_bytes() +} + +fn xyz_tag(v: [f64; 3]) -> Vec { + let mut b = b"XYZ \0\0\0\0".to_vec(); + for c in v { + b.extend_from_slice(&s15f16(c)); + } + b +} + +fn text_tag(text: &str) -> Vec { + let mut b = b"text\0\0\0\0".to_vec(); + b.extend_from_slice(text.as_bytes()); + b.push(0); + b +} + +/// `textDescriptionType` with the ASCII description only. +fn desc_tag(text: &str) -> Vec { + let mut b = b"desc\0\0\0\0".to_vec(); + b.extend_from_slice(&(text.len() as u32 + 1).to_be_bytes()); + b.extend_from_slice(text.as_bytes()); + b.push(0); + // No Unicode text and no Macintosh ScriptCode text: empty fields and the 67-byte filler. + b.extend_from_slice(&[0; 4 + 4 + 2 + 1 + 67]); + b +} + +fn curve_tag(curve: ToneCurve<'_>) -> Vec { + let mut b = b"curv\0\0\0\0".to_vec(); + match curve { + ToneCurve::Linear => { + b.extend_from_slice(&1u32.to_be_bytes()); + b.extend_from_slice(&0x0100u16.to_be_bytes()); + } + ToneCurve::Samples(samples) => { + b.extend_from_slice(&(samples.len() as u32).to_be_bytes()); + for sample in samples { + b.extend_from_slice(&sample.to_be_bytes()); + } + } + } + b +} + +/// Write a deterministic ICC v2.1 RGB display profile with one shared RGB TRC. +/// +/// Returns `None` for nonfinite matrix values, descriptions longer than 1024 ASCII +/// bytes or containing NUL, or sampled curves outside 2..=4096 entries. These caps +/// bound every allocation independently of caller input. Matrix values outside the +/// signed 15.16 range saturate, matching the legacy exporter. The caller owns colour +/// adaptation, transfer sampling, names and timestamp policy. +pub fn matrix_profile(params: &MatrixProfile<'_>) -> Option> { + let valid_text = |s: &str| s.len() <= 1024 && s.is_ascii() && !s.contains('\0'); + if !valid_text(params.description) + || !valid_text(params.copyright) + || !params.matrix.iter().flatten().all(|v| v.is_finite()) + || matches!(params.curve, ToneCurve::Samples(s) if !(2..=4096).contains(&s.len())) + { + return None; + } + let m = params.matrix; + let column = |c: usize| [m[0][c], m[1][c], m[2][c]]; + let curve = curve_tag(params.curve); + // The green and blue curves are the red one's entry. + let tags: [(&[u8; 4], Vec); 9] = [ + (b"desc", desc_tag(params.description)), + (b"cprt", text_tag(params.copyright)), + (b"wtpt", xyz_tag(D50_XYZ)), + (b"rXYZ", xyz_tag(column(0))), + (b"gXYZ", xyz_tag(column(1))), + (b"bXYZ", xyz_tag(column(2))), + (b"rTRC", curve.clone()), + (b"gTRC", curve.clone()), + (b"bTRC", curve), + ]; + // Header (128) + tag count (4) + 12 per tag, then the tag data, each tag padded to 4 bytes. + let table_end = 128 + 4 + 12 * tags.len(); + let mut offsets = Vec::new(); + let mut data: Vec = Vec::new(); + for (i, (_, bytes)) in tags.iter().enumerate() { + let shared = i >= 7; + let at = if shared { offsets.get(6).copied().unwrap_or(0) } else { table_end + data.len() }; + offsets.push(at); + if !shared { + data.extend_from_slice(bytes); + while !data.len().is_multiple_of(4) { + data.push(0); + } + } + } + let total = table_end + data.len(); + let mut out = vec![0u8; 128]; + out.get_mut(0..4)?.copy_from_slice(&(total as u32).to_be_bytes()); + out.get_mut(8..12)?.copy_from_slice(&[2, 0x10, 0, 0]); // version 2.1.0 + out.get_mut(12..16)?.copy_from_slice(b"mntr"); + out.get_mut(16..20)?.copy_from_slice(b"RGB "); + out.get_mut(20..24)?.copy_from_slice(b"XYZ "); + // Serialize caller-supplied ICC date fields; no calendar arithmetic. + for (i, value) in params.date.iter().enumerate() { + out.get_mut(24 + i * 2..26 + i * 2)?.copy_from_slice(&value.to_be_bytes()); + } + out.get_mut(36..40)?.copy_from_slice(b"acsp"); + for (k, v) in D50_XYZ.iter().enumerate() { + out.get_mut(68 + 4 * k..72 + 4 * k)?.copy_from_slice(&s15f16(*v)); + } + out.extend_from_slice(&(tags.len() as u32).to_be_bytes()); + for ((sig, bytes), at) in tags.iter().zip(&offsets) { + out.extend_from_slice(*sig); + out.extend_from_slice(&(*at as u32).to_be_bytes()); + out.extend_from_slice(&(bytes.len() as u32).to_be_bytes()); + } + out.extend_from_slice(&data); + Some(out) +} + +#[cfg(test)] +mod tests { + #![allow(clippy::unwrap_used)] + use super::*; + + fn params() -> MatrixProfile<'static> { + MatrixProfile { + matrix: [[0.4361, 0.3851, 0.1431], [0.2225, 0.7169, 0.0606], [0.0139, 0.0971, 0.7141]], + curve: ToneCurve::Linear, + description: "v2", + copyright: "free", + date: [2026, 1, 1, 0, 0, 0], + } + } + + #[test] + fn valid_profile_parses_and_aliases_curves() { + for curve in [ToneCurve::Linear, ToneCurve::Samples(&[0, 32768, 65535])] { + let bytes = matrix_profile(&MatrixProfile { curve, ..params() }).unwrap(); + let p = crate::Profile::parse(&bytes).unwrap(); + assert_eq!(p.version, (2, 0x10)); + assert_eq!(p.description, "v2"); + assert_eq!(&bytes[24..28], &[7, 234, 0, 1]); + assert_eq!(&bytes[208..216], &bytes[220..228]); + assert_eq!(&bytes[208..216], &bytes[232..240]); + assert_eq!(bytes, matrix_profile(&MatrixProfile { curve, ..params() }).unwrap()); + } + } + + #[test] + fn bounded_inputs_reject_before_serializing() { + let long = "x".repeat(1025); + for description in [long.as_str(), "bad\0name", "non-ascii é"] { + assert!(matrix_profile(&MatrixProfile { description, ..params() }).is_none()); + } + for samples in [&[][..], &[0][..], &vec![0; 4097][..]] { + assert!(matrix_profile(&MatrixProfile { curve: ToneCurve::Samples(samples), ..params() }).is_none()); + } + for value in [f64::NAN, f64::INFINITY, f64::NEG_INFINITY] { + let mut p = params(); + p.matrix[0][0] = value; + assert!(matrix_profile(&p).is_none()); + } + let samples = vec![0; 4096]; + let name = "x".repeat(1024); + let bytes = matrix_profile(&MatrixProfile { description: &name, copyright: &name, curve: ToneCurve::Samples(&samples), ..params() }).unwrap(); + assert!(bytes.len() < 12 * 1024); + } +} diff --git a/crates/cms/tests/spi1d.rs b/crates/cms/tests/spi1d.rs new file mode 100644 index 000000000..f4e27d051 --- /dev/null +++ b/crates/cms/tests/spi1d.rs @@ -0,0 +1,32 @@ +use craft_cms::spi1d::{MAX_ENTRIES, Table, parse}; + +#[test] +fn monochrome_and_three_component_samples_preserve_domain() { + assert_eq!( + parse("Version 1\nFrom -1 2\nLength 2\nComponents 1\n{\n-2\n3\n}"), + Some(Table { values: vec![[-2.0; 3], [3.0; 3]], domain: [-1.0, 2.0] }) + ); + assert_eq!( + parse("Length 2\nComponents 3\n{\n0 1\n2 3 4 5\n}"), + Some(Table { values: vec![[0.0, 1.0, 2.0], [3.0, 4.0, 5.0]], domain: [0.0, 1.0] }) + ); +} + +#[test] +fn legacy_header_and_eof_tolerance_is_preserved() { + let table = parse("Ignored arbitrary header\nLength 2 extra\n{\n0 1"); + assert!(table.is_some()); + assert_eq!(table, parse("Length 2\n{\n0 1\n} ignored\ninvalid trailing text")); + assert!(parse("Length 2\n{\nNaN inf\n}").is_some(), "sample sanitization belongs to the caller"); +} + +#[test] +fn invalid_headers_partial_and_surplus_tables_are_refused() { + for header in ["Length 1", "Length 2\nComponents 2", "Length 2\nFrom 2 1", "Length 2\nFrom NaN 1"] { + assert!(parse(&format!("{header}\n{{\n0 1\n}}")).is_none()); + } + assert!(parse(&format!("Length {}\n{{\n0 1", MAX_ENTRIES + 1)).is_none()); + assert!(parse("Length 2\nComponents 3\n{\n0 1 2 3 4").is_none()); + assert!(parse("Length 2\n{\n0 1 2").is_none()); + assert!(parse("Length 2\n{\n").is_none()); +} diff --git a/crates/deckcraft-format/Cargo.toml b/crates/deckcraft-format/Cargo.toml new file mode 100644 index 000000000..d7c6b035f --- /dev/null +++ b/crates/deckcraft-format/Cargo.toml @@ -0,0 +1,18 @@ +[package] +name = "craft-deckcraft-format" +description = "The DeckCraft .deckcraft package: identity, manifest envelope, NaN-safe JSON, member names and bounded reads" +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 } +zip = { version = "9", default-features = false, features = ["deflate-flate2"] } diff --git a/crates/deckcraft-format/README.md b/crates/deckcraft-format/README.md new file mode 100644 index 000000000..94f5d90e8 --- /dev/null +++ b/crates/deckcraft-format/README.md @@ -0,0 +1,25 @@ +# craft-deckcraft-format + +The DeckCraft `.deckcraft` package. A file is a ZIP that holds, in this order: + +- `mimetype`, stored: `application/vnd.storyteller.deckcraft+zip`. Files from before the + SlideCraft rename carry the legacy type, and `sniff` accepts both. +- `presentation.json`, deflated: `{"format":"deckcraft","version":1,"generator":…,"presentation":…}`. + It is compact JSON, and non-finite numbers are written as `0` (`Finite`). +- `media/` and `chart-parts/` binary members. A member whose content type is already + compressed (JPEG, PNG, video, MP3/MP4 audio, anything `zip`) is stored, and any other member is deflated. + +The presentation is a type parameter: DeckCraft owns the model, decides which media and chart +parts become members, and repairs a loaded deck. This crate owns the bytes: + +- `write(presentation, generator, members)` writes the package with `craft-zip`'s deterministic writer. +- `Package::open_manifest(bytes, max)` reads and checks the manifest. A file with no + `presentation.json` or a different `format` is `NotOurs`. `parse_manifest` reads files written + before `Finite`, where serde_json saved NaN as `null`, as 0 within fixed patch and re-parse budgets. +- `Package::read_max` refuses a member that declares more than its limit and stops one that + inflates past it. `read_charged` charges a shared budget for every byte it inflates, including + the bytes of a read that then fails its CRC. +- `Budget` and the `MAX_*` constants are the format's limits: 2 GiB per member, 4 GiB of media and + chart parts, 100,000 items and a 256 MiB manifest. + +There is no `unsafe` and no panics, and it builds for `wasm32-unknown-unknown`. diff --git a/crates/deckcraft-format/src/finite.rs b/crates/deckcraft-format/src/finite.rs new file mode 100644 index 000000000..834108a33 --- /dev/null +++ b/crates/deckcraft-format/src/finite.rs @@ -0,0 +1,185 @@ +//! A serde serializer adapter that writes non-finite floats as 0. +//! +//! serde_json writes NaN and ±infinity as `null`, which doesn't read back as a number: a single +//! NaN anywhere in the model (a rotation, a crop, a stroke width) made a saved file fail to open +//! as "damaged". [`Finite`] wraps any `Serialize` value and replaces those floats on the way out; +//! everything else passes through unchanged. + +use serde::ser::{self, Serialize, Serializer}; + +/// `T`, serialized with non-finite floats written as 0. +pub struct Finite<'a, T: ?Sized>(pub &'a T); + +impl Serialize for Finite<'_, T> { + fn serialize(&self, s: S) -> Result { + self.0.serialize(Wrap(s)) + } +} + +struct Wrap(S); + +macro_rules! pass { + ($($m:ident($t:ty)),* $(,)?) => { + $(fn $m(self, v: $t) -> Result { self.0.$m(v) })* + }; +} + +impl Serializer for Wrap { + type Ok = S::Ok; + type Error = S::Error; + type SerializeSeq = Wrap; + type SerializeTuple = Wrap; + type SerializeTupleStruct = Wrap; + type SerializeTupleVariant = Wrap; + type SerializeMap = Wrap; + type SerializeStruct = Wrap; + type SerializeStructVariant = Wrap; + + pass!( + serialize_bool(bool), + serialize_i8(i8), + serialize_i16(i16), + serialize_i32(i32), + serialize_i64(i64), + serialize_i128(i128), + serialize_u8(u8), + serialize_u16(u16), + serialize_u32(u32), + serialize_u64(u64), + serialize_u128(u128), + serialize_char(char), + serialize_str(&str), + serialize_bytes(&[u8]), + ); + + fn serialize_f32(self, v: f32) -> Result { + self.0.serialize_f32(if v.is_finite() { v } else { 0.0 }) + } + fn serialize_f64(self, v: f64) -> Result { + self.0.serialize_f64(if v.is_finite() { v } else { 0.0 }) + } + fn serialize_none(self) -> Result { + self.0.serialize_none() + } + fn serialize_some(self, v: &T) -> Result { + self.0.serialize_some(&Finite(v)) + } + fn serialize_unit(self) -> Result { + self.0.serialize_unit() + } + fn serialize_unit_struct(self, name: &'static str) -> Result { + self.0.serialize_unit_struct(name) + } + fn serialize_unit_variant(self, name: &'static str, i: u32, variant: &'static str) -> Result { + self.0.serialize_unit_variant(name, i, variant) + } + fn serialize_newtype_struct(self, name: &'static str, v: &T) -> Result { + self.0.serialize_newtype_struct(name, &Finite(v)) + } + fn serialize_newtype_variant(self, name: &'static str, i: u32, variant: &'static str, v: &T) -> Result { + self.0.serialize_newtype_variant(name, i, variant, &Finite(v)) + } + fn serialize_seq(self, len: Option) -> Result { + self.0.serialize_seq(len).map(Wrap) + } + fn serialize_tuple(self, len: usize) -> Result { + self.0.serialize_tuple(len).map(Wrap) + } + fn serialize_tuple_struct(self, name: &'static str, len: usize) -> Result { + self.0.serialize_tuple_struct(name, len).map(Wrap) + } + fn serialize_tuple_variant(self, name: &'static str, i: u32, variant: &'static str, len: usize) -> Result { + self.0.serialize_tuple_variant(name, i, variant, len).map(Wrap) + } + fn serialize_map(self, len: Option) -> Result { + self.0.serialize_map(len).map(Wrap) + } + fn serialize_struct(self, name: &'static str, len: usize) -> Result { + self.0.serialize_struct(name, len).map(Wrap) + } + fn serialize_struct_variant( + self, + name: &'static str, + i: u32, + variant: &'static str, + len: usize, + ) -> Result { + self.0.serialize_struct_variant(name, i, variant, len).map(Wrap) + } + fn is_human_readable(&self) -> bool { + self.0.is_human_readable() + } +} + +macro_rules! compound { + ($tr:ident, $m:ident) => { + impl ser::$tr for Wrap { + type Ok = S::Ok; + type Error = S::Error; + fn $m(&mut self, v: &T) -> Result<(), S::Error> { + self.0.$m(&Finite(v)) + } + fn end(self) -> Result { + self.0.end() + } + } + }; +} +compound!(SerializeSeq, serialize_element); +compound!(SerializeTuple, serialize_element); +compound!(SerializeTupleStruct, serialize_field); +compound!(SerializeTupleVariant, serialize_field); + +impl ser::SerializeMap for Wrap { + type Ok = S::Ok; + type Error = S::Error; + fn serialize_key(&mut self, k: &T) -> Result<(), S::Error> { + self.0.serialize_key(&Finite(k)) + } + fn serialize_value(&mut self, v: &T) -> Result<(), S::Error> { + self.0.serialize_value(&Finite(v)) + } + fn end(self) -> Result { + self.0.end() + } +} + +macro_rules! fields { + ($tr:ident) => { + impl ser::$tr for Wrap { + type Ok = S::Ok; + type Error = S::Error; + fn serialize_field(&mut self, key: &'static str, v: &T) -> Result<(), S::Error> { + self.0.serialize_field(key, &Finite(v)) + } + fn skip_field(&mut self, key: &'static str) -> Result<(), S::Error> { + self.0.skip_field(key) + } + fn end(self) -> Result { + self.0.end() + } + } + }; +} +fields!(SerializeStruct); +fields!(SerializeStructVariant); + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn non_finite_floats_are_written_as_zero_and_nothing_else_changes() { + #[derive(serde::Serialize)] + struct T { + a: f64, + b: Option, + c: Vec<(f32, f64)>, + d: Option, + e: std::collections::BTreeMap, + } + let t = T { a: f64::NAN, b: Some(f64::INFINITY), c: vec![(f32::NEG_INFINITY, 1.5)], d: None, e: [("k".to_string(), f64::NAN)].into() }; + let s = serde_json::to_string(&Finite(&t)).unwrap_or_default(); + assert_eq!(s, r#"{"a":0.0,"b":0.0,"c":[[0.0,1.5]],"d":null,"e":{"k":0.0}}"#); + } +} diff --git a/crates/deckcraft-format/src/lib.rs b/crates/deckcraft-format/src/lib.rs new file mode 100644 index 000000000..b59363c99 --- /dev/null +++ b/crates/deckcraft-format/src/lib.rs @@ -0,0 +1,226 @@ +//! The native `.deckcraft` package: a ZIP holding a stored `mimetype`, `presentation.json` (the +//! manifest: format name, version, generator and the presentation, minus media bytes), and +//! `media/` and `chart-parts/` binary members. Readers ignore unknown fields, so newer +//! files open in older builds with what those builds understand. +//! +//! DeckCraft owns the presentation model and its mapping: which media and chart parts become +//! members, and how a loaded deck is repaired. This crate owns the bytes: the package identity, +//! the manifest envelope, the NaN-safe JSON writer, the reader that opens files with NaN saved as +//! `null`, member names and compression, and the bounded member reads and their budgets. +#![deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable)] +#![forbid(unsafe_code)] + +mod finite; +mod manifest; + +#[cfg(test)] +mod tests; + +use std::io::{Cursor, Read}; + +pub use finite::Finite; +pub use manifest::parse_manifest; + +pub const MIMETYPE: &str = "application/vnd.storyteller.deckcraft+zip"; +/// Files written before the SlideCraft → DeckCraft rename (`.slidecraft`) are still read. +pub const LEGACY_MIMETYPE: &str = concat!("application/vnd.storyteller.", "slide", "craft+zip"); +pub const EXTENSION: &str = "deckcraft"; +/// The manifest's `format`. +pub const FORMAT: &str = "deckcraft"; +/// Format version written into the manifest. +pub const VERSION: u32 = 1; +/// The stored member naming the package type. +pub const MIMETYPE_ENTRY: &str = "mimetype"; +/// The manifest member. +pub const MANIFEST_ENTRY: &str = "presentation.json"; +/// Largest entry we will inflate (guards against zip bombs). +pub const MAX_ENTRY: u64 = 2 << 30; +/// All embedded media and chart sidecars together, inflated (storytold/deckcraft#393). +pub const MAX_TOTAL_MEDIA: u64 = 4 << 30; +/// Media items and chart sidecar parts one file may list. +pub const MAX_MEDIA_ITEMS: usize = 100_000; +/// Largest `presentation.json` we will inflate. The manifest holds no media bytes, so it is far +/// smaller than a media entry may be; inflating and parsing one is what dominates memory. +pub const MAX_MANIFEST: u64 = 256 << 20; + +const MEDIA_PREFIX: &str = "media/"; +const CHART_PARTS_PREFIX: &str = "chart-parts/"; + +/// What one load may inflate. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub struct Budget { + /// Bytes all media and chart members may inflate to together. + pub total: u64, + /// Media items and chart parts the manifest may list. + pub items: usize, + /// Bytes `presentation.json` may inflate to. + pub manifest: u64, +} + +impl Default for Budget { + fn default() -> Self { + Budget { total: MAX_TOTAL_MEDIA, items: MAX_MEDIA_ITEMS, manifest: MAX_MANIFEST } + } +} + +impl Budget { + /// A media budget with the default manifest limit. + pub const fn media(total: u64, items: usize) -> Self { + Budget { total, items, manifest: MAX_MANIFEST } + } +} + +#[derive(Debug, thiserror::Error)] +pub enum FormatError { + #[error("not a DeckCraft file: {0}")] + NotOurs(String), + #[error("damaged file: {0}")] + Damaged(String), + #[error("write failed: {0}")] + Write(String), +} + +pub type Result = std::result::Result; + +/// `presentation.json`: the envelope around the app's presentation. +#[derive(serde::Serialize, serde::Deserialize)] +pub struct Manifest { + pub format: String, + pub version: u32, + pub generator: String, + pub presentation: T, +} + +/// The member holding media item `id`'s bytes. +pub fn media_entry(id: u32) -> String { + format!("{MEDIA_PREFIX}{id}") +} + +/// The member holding the `index`th chart part a save separated out. +pub fn chart_part_entry(index: usize) -> String { + format!("{CHART_PARTS_PREFIX}{index}") +} + +/// Whether `entry` is a name [`chart_part_entry`] gives, exactly: no other member, no path +/// tricks, no leading zeros. +pub fn is_chart_part_entry(entry: &str) -> bool { + entry.strip_prefix(CHART_PARTS_PREFIX).and_then(|s| s.parse::().ok()).is_some_and(|n| entry == chart_part_entry(n)) +} + +/// Whether a member of this content type is stored rather than deflated: it is compressed already. +pub fn is_compressed(content_type: &str) -> bool { + content_type.starts_with("image/jpeg") + || content_type.starts_with("image/png") + || content_type.starts_with("video/") + || content_type.starts_with("audio/mpeg") + || content_type.starts_with("audio/mp4") + || content_type.contains("zip") +} + +/// Does this look like a `.deckcraft` file? +pub fn sniff(bytes: &[u8]) -> bool { + bytes.starts_with(b"PK") && [MIMETYPE, LEGACY_MIMETYPE].iter().any(|m| bytes.windows(m.len()).take(200).any(|w| w == m.as_bytes())) +} + +/// A binary member of a package being written. +#[derive(Clone, Copy, Debug)] +pub struct Member<'a> { + /// Its name, from [`media_entry`] or [`chart_part_entry`]. + pub name: &'a str, + /// Its content type, which decides whether it is stored or deflated. + pub content_type: &'a str, + pub bytes: &'a [u8], +} + +/// The package: `mimetype` (stored), the manifest naming `generator` around `presentation` +/// (deflated, with non-finite numbers written as 0), then `members` in order. +pub fn write(presentation: &T, generator: &str, members: &[Member<'_>]) -> Result> { + use craft_zip::Compression::{Deflated, Stored}; + let w = |e: craft_zip::Error| FormatError::Write(e.to_string()); + let mut z = craft_zip::Writer::new(); + z.add(MIMETYPE_ENTRY, MIMETYPE.as_bytes(), Stored).map_err(w)?; + let m = Manifest { format: FORMAT.to_string(), version: VERSION, generator: generator.to_string(), presentation }; + let json = serde_json::to_vec(&Finite(&m)).map_err(|e| FormatError::Write(e.to_string()))?; + z.add(MANIFEST_ENTRY, &json, Deflated).map_err(w)?; + for member in members { + let compression = if is_compressed(member.content_type) { Stored } else { Deflated }; + z.add(member.name, member.bytes, compression).map_err(w)?; + } + z.finish().map_err(w) +} + +/// A package open for reading. Nothing is inflated until a member is read. +pub struct Package<'a> { + zip: zip::ZipArchive>, +} + +impl<'a> Package<'a> { + /// Index `bytes` as a ZIP. + pub fn open(bytes: &'a [u8]) -> Result { + let zip = zip::ZipArchive::new(Cursor::new(bytes)).map_err(|e| FormatError::NotOurs(e.to_string()))?; + Ok(Package { zip }) + } + + /// Open `bytes` and read its manifest, inflating at most `max_manifest` bytes of it. Missing + /// `presentation.json` or a different `format` is [`FormatError::NotOurs`]. + pub fn open_manifest(bytes: &'a [u8], max_manifest: u64) -> Result<(Manifest, Self)> { + let mut package = Package::open(bytes)?; + if !package.contains(MANIFEST_ENTRY) { + return Err(FormatError::NotOurs(format!("no {MANIFEST_ENTRY}"))); + } + let json = package.read_max(MANIFEST_ENTRY, max_manifest)?; + let m: Manifest = parse_manifest(&json)?; + if m.format != FORMAT { + return Err(FormatError::NotOurs(m.format)); + } + Ok((m, package)) + } + + /// Whether the package has a member named `name`. + pub fn contains(&self, name: &str) -> bool { + self.zip.index_for_name(name).is_some() + } + + /// The size `name` declares, inflated; `None` when there is no such readable member. + pub fn declared_size(&mut self, name: &str) -> Option { + self.zip.by_name(name).ok().map(|f| f.size()) + } + + /// Read one member within [`MAX_ENTRY`]. + pub fn read(&mut self, name: &str) -> Result> { + self.read_max(name, MAX_ENTRY) + } + + /// Read one member, refusing it before inflating if it declares more than `max` bytes, and + /// stopping at `max` if it inflates further than it declared. + pub fn read_max(&mut self, name: &str, max: u64) -> Result> { + self.read_counted(name, max, &mut 0) + } + + /// [`Package::read_max`] within `left` (and [`MAX_ENTRY`]), charging `left` for every byte + /// inflated, including those inflated before a decode or CRC error, so a member that fails + /// late still spends the shared budget it used. + pub fn read_charged(&mut self, name: &str, left: &mut u64) -> Result> { + let mut spent = 0; + let out = self.read_counted(name, (*left).min(MAX_ENTRY), &mut spent); + *left = left.saturating_sub(spent); + out + } + + /// [`Package::read_max`], adding the bytes inflated to `spent` whether or not the read succeeds. + fn read_counted(&mut self, name: &str, max: u64, spent: &mut u64) -> Result> { + let f = self.zip.by_name(name).map_err(|e| FormatError::Damaged(format!("{name}: {e}")))?; + if f.size() > max { + return Err(FormatError::Damaged(format!("{name} is too large"))); + } + let mut out = Vec::with_capacity(usize::try_from(f.size().min(64 << 20)).unwrap_or(0)); + let mut f = f.take(max.saturating_add(1)); + let read = f.read_to_end(&mut out); + *spent = spent.saturating_add(out.len() as u64); + read.map_err(|e| FormatError::Damaged(format!("{name}: {e}")))?; + if out.len() as u64 > max { + return Err(FormatError::Damaged(format!("{name} is too large"))); + } + Ok(out) + } +} diff --git a/crates/deckcraft-format/src/manifest.rs b/crates/deckcraft-format/src/manifest.rs new file mode 100644 index 000000000..03f017ace --- /dev/null +++ b/crates/deckcraft-format/src/manifest.rs @@ -0,0 +1,99 @@ +//! Reading `presentation.json`, including files written with NaN or infinity saved as `null`. + +use serde::de::DeserializeOwned; + +use crate::{FormatError, Manifest, Result}; + +/// Most `null`s standing in for numbers that [`parse_manifest`] patches in one file, and the +/// candidates it tries for each. +const MAX_NULL_FIXES: usize = 256; +const MAX_TRIES_PER_FIX: usize = 32; +/// Bytes [`parse_manifest`] may re-parse in all while patching, so a hostile file full of `null`s +/// can't keep it busy: past this the file is reported damaged. +const MAX_REPARSE_BYTES: usize = 256 * 1024 * 1024; + +/// Where a parse failed, comparable: later is progress. +fn failure(src: &[u8]) -> std::result::Result, (String, (usize, usize))> { + serde_json::from_slice::>(src).map_err(|e| (e.to_string(), (e.line(), e.column()))) +} + +/// Parse `presentation.json`. serde_json writes a NaN or infinite number as `null`, which then +/// fails to read back as a number ("invalid type: null, expected f64"), so DeckCraft could save +/// a file it couldn't open. Each such `null` is read as 0 instead (the app then brings the +/// value into range); anything else that doesn't parse is damage. +/// +/// serde reports the failing `null` exactly, except inside tagged enums (buffered), where it +/// reports the end of the enclosing object: there the `null`s before that point are tried +/// nearest first, and a patch is kept only if the parse then gets further. (Files saved since +/// [`crate::Finite`] never contain such `null`s; this reads older ones.) +pub fn parse_manifest(json: &[u8]) -> Result> { + // Parse the inflated text as it is: only a file that fails on a `null` number is copied for + // patching (the copy is bounded by `MAX_MANIFEST`). + let mut last = match failure(json) { + Ok(m) => return Ok(m), + Err(f) => f, + }; + if !is_null_number(&last.0) { + return Err(FormatError::Damaged(last.0)); + } + let mut src = json.to_vec(); + let mut budget = MAX_REPARSE_BYTES; + 'fixes: for _ in 0..MAX_NULL_FIXES { + let (msg, pos) = &last; + if !is_null_number(msg) { + break; + } + let end = offset(&src, pos.0, pos.1).unwrap_or(src.len()); + let mut progressed = false; + let candidates: Vec = value_nulls_before(&src, end + 4).take(MAX_TRIES_PER_FIX).collect(); + for at in candidates { + let (Some(head), Some(tail)) = (src.get(..at), src.get(at.saturating_add(4)..)) else { continue }; + let Some(left) = budget.checked_sub(src.len()) else { break 'fixes }; + budget = left; + let mut next = Vec::with_capacity(src.len()); + next.extend_from_slice(head); + next.push(b'0'); + next.extend_from_slice(tail); + match failure(&next) { + Ok(m) => return Ok(m), + // The patch shortened the text by 3 bytes before the error: compare positions + // allowing for that. + Err(f) if f.1 > (pos.0, pos.1.saturating_sub(3)) => { + src = next; + last = f; + progressed = true; + break; + } + Err(_) => {} + } + } + if !progressed { + break; + } + } + Err(FormatError::Damaged(last.0)) +} + +/// Is this parse error a `null` where a number belongs (a NaN or infinity serde_json wrote)? +fn is_null_number(msg: &str) -> bool { + msg.starts_with("invalid type: null, expected f") || msg.starts_with("invalid type: null, expected a number") +} + +/// Byte offset of a 1-based `line`:`column`. +fn offset(src: &[u8], line: usize, column: usize) -> Option { + let mut start = 0; + for _ in 1..line { + start += src.get(start..)?.iter().position(|b| *b == b'\n')? + 1; + } + Some((start + column).min(src.len())) +} + +/// Starts of `null` tokens in value position (after `:`, `[` or `,`) before `end`, nearest first. +fn value_nulls_before(src: &[u8], end: usize) -> impl Iterator + '_ { + let end = end.min(src.len()); + (0..end.saturating_sub(3)).rev().filter(move |&i| { + src.get(i..i + 4) == Some(b"null") + && src[..i].iter().rev().find(|b| !b.is_ascii_whitespace()).is_some_and(|b| matches!(b, b':' | b'[' | b',')) + && src.get(i + 4).is_none_or(|b| !b.is_ascii_alphanumeric()) + }) +} diff --git a/crates/deckcraft-format/src/tests.rs b/crates/deckcraft-format/src/tests.rs new file mode 100644 index 000000000..8090d709a --- /dev/null +++ b/crates/deckcraft-format/src/tests.rs @@ -0,0 +1,235 @@ +#![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)] + +use serde::{Deserialize, Serialize}; + +use super::*; + +/// A stand-in for an app's presentation: plain numbers, an optional number and a tagged enum +/// (whose `null`s serde reports only at the end of the enclosing object). +#[derive(Debug, Default, PartialEq, Serialize, Deserialize)] +struct Deck { + #[serde(default)] + width: f64, + #[serde(default)] + zoom: Option, + #[serde(default)] + shapes: Vec, +} + +#[derive(Debug, PartialEq, Serialize, Deserialize)] +#[serde(tag = "kind")] +enum Shape { + Box { x: f64, y: f64 }, + Line { len: f64 }, +} + +fn package(members: &[Member<'_>]) -> Vec { + write(&Deck { width: 960.0, ..Default::default() }, "Test 1", members).unwrap() +} + +#[test] +fn a_package_holds_mimetype_then_manifest_then_members_in_order() { + let png = [0x89, b'P', b'N', b'G']; + let svg = b"".repeat(20); + let bytes = package(&[ + Member { name: &media_entry(7), content_type: "image/png", bytes: &png }, + Member { name: &media_entry(8), content_type: "image/svg+xml", bytes: &svg }, + Member { name: &chart_part_entry(0), content_type: "application/zip", bytes: b"PK" }, + ]); + assert!(sniff(&bytes)); + let mut z = zip::ZipArchive::new(std::io::Cursor::new(bytes.as_slice())).unwrap(); + assert_eq!(z.len(), 5); + let methods: Vec<_> = (0..z.len()) + .map(|i| { + let f = z.by_index(i).unwrap(); + (f.name().unwrap().into_owned(), f.compression()) + }) + .collect(); + use zip::CompressionMethod::{Deflated, Stored}; + assert_eq!( + methods, + [("mimetype", Stored), ("presentation.json", Deflated), ("media/7", Stored), ("media/8", Deflated), ("chart-parts/0", Stored)] + .map(|(n, m)| (n.to_string(), m)) + ); + // The mimetype is the first, stored member, readable in place. + assert_eq!(craft_zip::stored_mimetype(&bytes), Some(MIMETYPE.as_bytes())); + let (m, mut p) = Package::open_manifest::(&bytes, MAX_MANIFEST).unwrap(); + assert_eq!((m.format.as_str(), m.version, m.generator.as_str(), m.presentation.width), (FORMAT, VERSION, "Test 1", 960.0)); + assert_eq!(p.read(&media_entry(8)).unwrap(), svg); + assert_eq!(p.declared_size(&media_entry(7)), Some(4)); + assert_eq!(p.declared_size("media/99"), None); +} + +#[test] +fn the_manifest_is_compact_json_with_its_fields_in_order() { + let bytes = package(&[]); + let mut p = Package::open(&bytes).unwrap(); + let json = String::from_utf8(p.read(MANIFEST_ENTRY).unwrap()).unwrap(); + assert_eq!(json, r#"{"format":"deckcraft","version":1,"generator":"Test 1","presentation":{"width":960.0,"zoom":null,"shapes":[]}}"#); +} + +#[test] +fn non_finite_numbers_are_written_as_zero_and_read_back() { + let deck = Deck { width: f64::NAN, zoom: Some(f64::INFINITY), shapes: vec![Shape::Box { x: f64::NEG_INFINITY, y: 2.0 }] }; + let bytes = write(&deck, "t", &[]).unwrap(); + let (m, _) = Package::open_manifest::(&bytes, MAX_MANIFEST).unwrap(); + assert_eq!(m.presentation, Deck { width: 0.0, zoom: Some(0.0), shapes: vec![Shape::Box { x: 0.0, y: 2.0 }] }); +} + +/// Files written before [`Finite`] (plain serde_json: NaN as `null`) open too, including a +/// `null` inside a tagged enum, where serde reports only the enclosing object. +#[test] +fn older_manifests_with_null_numbers_parse() { + let cases: [(&str, Deck); 3] = [ + (r#"{"format":"deckcraft","version":1,"generator":"old","presentation":{"width":null}}"#, Deck::default()), + ( + r#"{"format":"deckcraft","version":1,"generator":"old","presentation":{"width":3,"shapes":[{"kind":"Box","x":null,"y":4}]}}"#, + Deck { width: 3.0, shapes: vec![Shape::Box { x: 0.0, y: 4.0 }], ..Default::default() }, + ), + ( + "{\n \"format\": \"deckcraft\",\n \"version\": 1,\n \"generator\": \"t\",\n \"presentation\": {\"shapes\": [{\"kind\": \"Line\", \"len\": null}]}\n}", + Deck { shapes: vec![Shape::Line { len: 0.0 }], ..Default::default() }, + ), + ]; + for (json, want) in cases { + let m = parse_manifest::(json.as_bytes()).unwrap_or_else(|e| panic!("{json}: {e}")); + assert_eq!(m.presentation, want, "{json}"); + } +} + +/// Only `null`s where numbers belong are patched: other damage still fails. +#[test] +fn other_manifest_damage_is_still_reported() { + let ok = br#"{"format": "deckcraft", "version": 1, "generator": "t", "presentation": {}}"#; + assert!(parse_manifest::(ok).is_ok()); + let bad = br#"{"format": null, "version": 1, "generator": "t", "presentation": {}}"#; + assert!(matches!(parse_manifest::(bad), Err(FormatError::Damaged(_)))); + assert!(parse_manifest::(&ok[..ok.len() - 3]).is_err()); +} + +#[test] +fn a_manifest_full_of_nulls_is_refused_within_bounds() { + let shapes = (0..2000).map(|_| r#"{"kind":"Box","x":1,"y":"not a number"}"#).collect::>().join(","); + let json = format!(r#"{{"format":"deckcraft","version":1,"generator":"t","presentation":{{"width":null,"shapes":[{shapes}]}}}}"#); + assert!(matches!(parse_manifest::(json.as_bytes()), Err(FormatError::Damaged(_)))); +} + +#[test] +fn what_is_not_a_package_is_not_ours() { + for bytes in [&b""[..], b"PK\x03\x04garbage", &[0u8; 1000]] { + assert!(matches!(Package::open_manifest::(bytes, MAX_MANIFEST), Err(FormatError::NotOurs(_)))); + } + // A ZIP without the manifest, and one from another app. + let mut z = craft_zip::Writer::new(); + z.add("mimetype", b"application/zip", craft_zip::Compression::Stored).unwrap(); + let no_manifest = z.finish().unwrap(); + let err = Package::open_manifest::(&no_manifest, MAX_MANIFEST).err().unwrap(); + assert_eq!(err.to_string(), "not a DeckCraft file: no presentation.json"); + let mut z = craft_zip::Writer::new(); + z.add(MANIFEST_ENTRY, br#"{"format":"other","version":1,"generator":"x","presentation":{}}"#, craft_zip::Compression::Deflated).unwrap(); + let other = z.finish().unwrap(); + let err = Package::open_manifest::(&other, MAX_MANIFEST).err().unwrap(); + assert_eq!(err.to_string(), "not a DeckCraft file: other"); + // A truncated package is an error, not a panic. + let mut b = package(&[]); + b.truncate(b.len() / 2); + assert!(Package::open_manifest::(&b, MAX_MANIFEST).is_err()); +} + +#[test] +fn legacy_slidecraft_packages_are_recognised() { + let mut z = craft_zip::Writer::new(); + z.add(MIMETYPE_ENTRY, LEGACY_MIMETYPE.as_bytes(), craft_zip::Compression::Stored).unwrap(); + assert!(sniff(&z.finish().unwrap())); + assert!(!sniff(b"PK\x03\x04 application/zip")); + assert!(!sniff(MIMETYPE.as_bytes())); +} + +#[test] +fn an_oversized_manifest_is_refused_before_inflating() { + let bytes = package(&[]); + let size = Package::open(&bytes).unwrap().declared_size(MANIFEST_ENTRY).unwrap(); + assert!(Package::open_manifest::(&bytes, size).is_ok()); + let err = Package::open_manifest::(&bytes, size - 1).err().unwrap(); + assert_eq!(err.to_string(), "damaged file: presentation.json is too large"); +} + +#[test] +fn a_member_that_inflates_past_its_declared_size_stops_at_the_limit() { + // Declared sizes are the file's word: a stored member whose header lies is still capped. + let data = vec![7u8; 100]; + let bytes = package(&[Member { name: &media_entry(1), content_type: "image/png", bytes: &data }]); + let mut p = Package::open(&bytes).unwrap(); + assert_eq!(p.read_max(&media_entry(1), 100).unwrap().len(), 100); + assert!(p.read_max(&media_entry(1), 99).is_err()); + let err = p.read("media/2").unwrap_err(); + assert!(matches!(err, FormatError::Damaged(ref m) if m.starts_with("media/2: ")), "{err}"); +} + +#[test] +fn corrupt_members_are_charged_for_what_they_inflated() { + let a = b"PICTURE-AAAAAAAAAAAA"; + let b = b"PICTURE-BBBBBBBBBBBB"; + let mut bytes = package(&[ + Member { name: &media_entry(1), content_type: "image/png", bytes: a }, + Member { name: &media_entry(2), content_type: "image/png", bytes: b }, + ]); + // Corrupt both stored members, so each fails its CRC only after inflating all its bytes. + for tag in [b"PICTURE-AAAA", b"PICTURE-BBBB"] { + let at = bytes.windows(tag.len()).position(|w| w == tag).unwrap(); + bytes[at] ^= 0xff; + } + let mut p = Package::open(&bytes).unwrap(); + let mut left = 30; + assert!(p.read_charged(&media_entry(1), &mut left).is_err()); + assert_eq!(left, 10, "a failed read still spends what it inflated"); + // The second fits a fresh budget on its own, but not what is left: refused before inflating. + let err = p.read_charged(&media_entry(2), &mut left).unwrap_err(); + assert_eq!(err.to_string(), "damaged file: media/2 is too large"); + assert_eq!(left, 10); +} + +#[test] +fn chart_part_names_are_exact() { + assert_eq!(chart_part_entry(3), "chart-parts/3"); + assert!(is_chart_part_entry("chart-parts/0")); + assert!(is_chart_part_entry("chart-parts/12")); + for not in ["media/1", "chart-parts/../0", "chart-parts/00", "chart-parts/-1", "chart-parts/", "chart-parts/+1", "chart-parts/1 "] { + assert!(!is_chart_part_entry(not), "{not}"); + } +} + +#[test] +fn compressed_content_types_are_stored() { + for stored in [ + "image/jpeg", + "image/png", + "video/mp4", + "audio/mpeg", + "audio/mp4", + "application/zip", + "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet+zip", + ] { + assert!(is_compressed(stored), "{stored}"); + } + for deflated in ["image/svg+xml", "image/gif", "audio/wav", "application/octet-stream", "text/plain"] { + assert!(!is_compressed(deflated), "{deflated}"); + } +} + +#[test] +fn duplicate_members_are_a_write_error() { + let err = write( + &Deck::default(), + "t", + &[Member { name: "media/1", content_type: "x", bytes: b"a" }, Member { name: "media/1", content_type: "x", bytes: b"b" }], + ) + .unwrap_err(); + assert!(matches!(err, FormatError::Write(_)), "{err}"); +} + +#[test] +fn the_default_budget_is_the_formats_limits() { + assert_eq!(Budget::default(), Budget { total: 4 << 30, items: 100_000, manifest: 256 << 20 }); + assert_eq!(Budget::media(10, 2), Budget { total: 10, items: 2, manifest: MAX_MANIFEST }); +} diff --git a/crates/dwg/src/lib.rs b/crates/dwg/src/lib.rs index 05c1a673c..c77fdbf4c 100644 --- a/crates/dwg/src/lib.rs +++ b/crates/dwg/src/lib.rs @@ -62,6 +62,10 @@ impl Limits { /// A byte count in megabytes for messages ("9.9 MB", "1024 MB"). #[cfg_attr(no_dwg_bridge, allow(dead_code))] +#[expect( + clippy::cast_precision_loss, + reason = "Megabyte formatting is rounded for display; exact integer precision beyond 2^53 bytes is not meaningful." +)] fn mb(n: u64) -> String { let m = n as f64 / f64::from(1u32 << 20); if m >= 100.0 { format!("{m:.0} MB") } else { format!("{m:.1} MB") } @@ -79,12 +83,17 @@ mod proxy_transport; #[cfg(not(no_dwg_bridge))] mod native { + use std::fmt::Write as _; use std::io::{Cursor, Write}; use std::panic::AssertUnwindSafe; use super::{Limits, mb}; /// Read DWG bytes and return an ASCII DXF rendition, within [`Limits::DEFAULT`]. + /// + /// # Errors + /// + /// Returns an error when the DWG cannot be read or post-processed, or its DXF rendition exceeds the default limits. pub fn dwg_to_dxf(bytes: &[u8]) -> Result, String> { dwg_to_dxf_with(bytes, Limits::DEFAULT) } @@ -92,6 +101,10 @@ mod native { /// [`dwg_to_dxf`] with explicit limits. Both are checked before the expensive step they /// guard: the input size before the DWG reader runs, the DXF size while it is written, so a /// runaway conversion stops at the limit instead of filling memory first. + /// + /// # Errors + /// + /// Returns an error when the input exceeds `limits`, decoding fails, or the DXF output exceeds `limits`. pub fn dwg_to_dxf_with(bytes: &[u8], limits: Limits) -> Result, String> { dwg_to_dxf_noted(bytes, limits).map(|(dxf, _)| dxf) } @@ -100,6 +113,10 @@ mod native { /// failsafe mode: an object record it can't decode is skipped instead of failing the whole /// file, and the notes say how many were skipped and of which types, so a partial read never /// passes for a complete one. + /// + /// # Errors + /// + /// Returns an error when the input exceeds `limits`, decoding fails, or the DXF output exceeds `limits`. pub fn dwg_to_dxf_noted(bytes: &[u8], limits: Limits) -> Result<(Vec, Vec), String> { if bytes.len() > limits.max_dwg_bytes { return Err(format!( @@ -122,13 +139,13 @@ mod native { let mut notes = read_notes(&outcome.stats); let mut doc = outcome.document; notes.extend( - guarded(AssertUnwindSafe(|| application_objects(&mut doc))).map_err(|_| "the DWG application object conversion failed".to_string())?, + guarded(AssertUnwindSafe(|| application_objects(&mut doc))).map_err(|()| "the DWG application object conversion failed".to_string())?, ); // After application_objects, so the proxy objects it makes are numbered as the DXF's. notes.extend( - guarded(AssertUnwindSafe(|| super::proxy_transport::prepare(&mut doc))).map_err(|_| "the DWG proxy conversion failed".to_string())?, + guarded(AssertUnwindSafe(|| super::proxy_transport::prepare(&mut doc))).map_err(|()| "the DWG proxy conversion failed".to_string())?, ); - let orphans = guarded(AssertUnwindSafe(|| post_process(&mut doc))).map_err(|_| "the DWG post-processing failed on this file".to_string())?; + let orphans = guarded(AssertUnwindSafe(|| post_process(&mut doc))).map_err(|()| "the DWG post-processing failed on this file".to_string())?; if orphans > 0 { let (s, were) = if orphans == 1 { ("", "was") } else { ("s", "were") }; notes.push(format!("DWG: {orphans} DGN line style component{s} that no line style definition uses {were} left out")); @@ -632,7 +649,7 @@ mod native { // A real table has at most one merge per cell; more ranges than that come from a // damaged file and would only make this rows × cols × ranges. let max_ranges = rows.saturating_mul(cols).min(1 << 16); - for r in t.merged_ranges.iter().take(max_ranges).cloned().collect::>() { + for r in t.merged_ranges.iter().take(max_ranges).copied().collect::>() { // Clamp to the table: the ranges come from the file. let (bottom, right) = (r.bottom_row.min(rows.saturating_sub(1)), r.right_col.min(cols.saturating_sub(1))); if r.top_row > bottom || r.left_col > right || rows == 0 || cols == 0 || (r.top_row, r.left_col) == (bottom, right) { @@ -715,8 +732,8 @@ mod native { } /// Leave out DGN line style components (`LS*COMPONENT`) that no line style definition uses - /// (#1482). A drawing that once came from MicroStation can carry hundreds of thousands of - /// them, orphaned copies piled up by earlier DGN imports and copy-pastes (AutoCAD's PURGE + /// (#1482). A drawing that once came from `MicroStation` can carry hundreds of thousands of + /// them, orphaned copies piled up by earlier DGN imports and copy-pastes (`AutoCAD`'s PURGE /// removes them as orphaned data): in a real 83 MB site plan they were 888,483 of 889,053 /// components and made its DXF rendition 1.2 GB, over [`Limits::DEFAULT`], so the file /// didn't open at all. No definition that was read refers to them (a definition the reader @@ -846,7 +863,7 @@ mod native { for (name, code, value) in unwritten_header_vars(h) { let written = section.windows(name.len() + 2).any(|w| w.starts_with(name.as_bytes()) && w.ends_with(b"\r\n")); if !written && value.parse::().is_ok_and(f64::is_finite) { - add.push_str(&format!(" 9\r\n{name}\r\n{code:>3}\r\n{value}\r\n")); + let _ = write!(add, " 9\r\n{name}\r\n{code:>3}\r\n{value}\r\n"); } } dxf.splice(end..end, add.into_bytes()); @@ -878,14 +895,22 @@ mod native { /// Oldest DWG version written; older DXF input is written as this version. const MIN_DWG_VERSION: acadrust::DxfVersion = acadrust::DxfVersion::AC1018; - /// Convert DXF bytes into a DWG file of the DXF's version, and at least AutoCAD 2004 + /// Convert DXF bytes into a DWG file of the DXF's version, and at least `AutoCAD` 2004 /// (AC1018). + /// + /// # Errors + /// + /// Returns an error when the DXF cannot be decoded or the DWG cannot be written, including oversized input. pub fn dxf_to_dwg(dxf: &[u8]) -> Result, String> { convert(dxf, None, None).map(|(dwg, _)| dwg) } /// Convert DXF bytes into a DWG file of `version`, a code from [`super::VERSIONS`] ("AC1032"), /// whatever version the DXF declares. + /// + /// # Errors + /// + /// Returns an error for an unsupported version, invalid DXF, oversized input, or a DWG write failure. pub fn dxf_to_dwg_as(dxf: &[u8], version: &str) -> Result, String> { dxf_to_dwg_noted(dxf, version).map(|(dwg, _)| dwg) } @@ -1011,7 +1036,7 @@ mod native { /// The current dimension style (`$DIMSTYLE`) by its name: the DXF gives the name, and the /// handle read with it needn't be the style's in the document, in which case the DWG writer - /// fell back to Standard and AutoCAD opened the drawing with Standard current (#1051). + /// fell back to Standard and `AutoCAD` opened the drawing with Standard current (#1051). fn current_dimstyle(doc: &mut acadrust::CadDocument) { let name = doc.header.current_dimstyle_name.clone(); if let Some(h) = doc.dim_styles.iter().find(|ds| ds.name.eq_ignore_ascii_case(&name)).map(|ds| ds.handle) { @@ -1021,7 +1046,7 @@ mod native { /// The current text style (`$TEXTSTYLE`) by its name, the same way as [`current_dimstyle`]: /// the handle read with the DXF needn't be the style's in the document, in which case the DWG - /// writer fell back to Standard and AutoCAD opened the drawing with Standard current (#1650). + /// writer fell back to Standard and `AutoCAD` opened the drawing with Standard current (#1650). fn current_text_style(doc: &mut acadrust::CadDocument) { let name = doc.header.current_text_style_name.clone(); if let Some(h) = doc.text_styles.iter().find(|ts| ts.name.eq_ignore_ascii_case(&name)).map(|ts| ts.handle) { @@ -1211,7 +1236,7 @@ mod native { }; let (classes, source) = guarded(AssertUnwindSafe(|| (dxf_class_names(&data), super::proxy_transport::source_classes(&data)))) .map_err(|_| "the DXF proxy class conversion failed".to_string())?; - let mut doc = std::panic::catch_unwind(move || acadrust::DxfReader::from_reader(Cursor::new(data)).and_then(|r| r.read())) + let mut doc = std::panic::catch_unwind(move || acadrust::DxfReader::from_reader(Cursor::new(data)).and_then(acadrust::DxfReader::read)) .map_err(|_| "the DXF→DWG conversion failed".to_string())? .map_err(|e| format!("DXF: {e}"))?; match version { @@ -1250,14 +1275,14 @@ mod native { current_layer(&mut doc); unwritten })) - .map_err(|_| "the DXF to DWG conversion failed on this file".to_string())?; + .map_err(|()| "the DXF to DWG conversion failed on this file".to_string())?; let dwg = std::panic::catch_unwind(AssertUnwindSafe(|| acadrust::DwgWriter::write_to_vec(&doc))) .map_err(|_| "the DWG writer failed on this drawing".to_string())? .map_err(|e| format!("DWG write: {e}"))?; Ok((dwg, unwritten)) } - /// AutoCAD's entry name for the drawing variables dictionary (the DXF Reference calls it the + /// `AutoCAD`'s entry name for the drawing variables dictionary (the DXF Reference calls it the /// ACDBVARIABLEDICTIONARY; dictionary keys are case-insensitive). const VARIABLE_DICTIONARY: &str = "AcDbVariableDictionary"; @@ -1273,11 +1298,11 @@ mod native { if valid.contains(&value) { value.to_string() } else { default.to_string() } } - /// Add the drawing variables dictionary AutoCAD always writes into the named object + /// Add the drawing variables dictionary `AutoCAD` always writes into the named object /// dictionary, with the variables it keeps there, when the drawing has none (#1813). /// Vectric Aspire 10.5 reads a 2004 (AC1018) DWG without it as empty ("No data was imported /// from the DXF / DWG file"), though it reads the same drawing as DXF — which has no such - /// dictionary either, as CADCraft's and ezdxf's writers don't make one. AutoCAD's AC1032 DWGs + /// dictionary either, as CADCraft's and ezdxf's writers don't make one. `AutoCAD`'s AC1032 DWGs /// read there without it, so newer output is left as it was, and a dictionary the drawing /// does have keeps its own entries and values. fn add_variable_dictionary(doc: &mut acadrust::CadDocument) { @@ -1326,7 +1351,7 @@ mod native { /// Entities of a layout other than the one in `*Paper_Space` are written owned by their block /// (entity mode 0, with its handle). acadrust writes every paper-space entity with mode 1, - /// "paper space", which AutoCAD reads as the `*Paper_Space` block's: everything drawn on a + /// "paper space", which `AutoCAD` reads as the `*Paper_Space` block's: everything drawn on a /// second layout landed on the first one, and the second stayed empty (#1057). fn other_layout_owners(doc: &mut acadrust::CadDocument) { let others: std::collections::HashSet = doc @@ -1549,11 +1574,11 @@ thread_local! { /// Panics when a test asked for `stage` to fail (see [`PANIC_IN`]); does nothing otherwise. #[cfg(not(no_dwg_bridge))] #[allow(clippy::panic)] -fn fail_in_test(_stage: &str) { +fn fail_in_test(stage: &str) { + #[cfg(not(test))] + let _ = stage; #[cfg(test)] - if PANIC_IN.with(|p| p.get()) == Some(_stage) { - panic!("test-forced panic in {_stage}"); - } + assert!(PANIC_IN.with(std::cell::Cell::get) != Some(stage), "test-forced panic in {stage}"); } /// The error for a DWG version this writer can't create. @@ -1633,6 +1658,8 @@ pub fn dxf_to_dwg_kept_noted(_dxf: &[u8], _version: &str, _kept: &str) -> Result #[cfg(test)] mod tests { + use std::fmt::Write as _; + use super::*; /// The DXF of an empty drawing, enough for the bridge to make a small DWG. @@ -1726,9 +1753,9 @@ mod tests { let mut s = String::from(" 0\nSECTION\n 2\nHEADER\n 9\n$ACADVER\n 1\nAC1032\n 0\nENDSEC\n 0\nSECTION\n 2\nENTITIES\n"); for (text, flags, sub, pts) in dims { s += " 0\nDIMENSION\n 8\n0\n100\nAcDbEntity\n100\nAcDbDimension\n 10\n0\n 20\n0\n 30\n0\n"; - s += &format!(" 11\n50\n 21\n-20\n 31\n0\n 70\n{flags}\n 1\n{text}\n 3\nStandard\n100\n{sub}\n"); - for (code, [x, y]) in pts.iter() { - s += &format!("{code:>3}\n{x}\n{:>3}\n{y}\n{:>3}\n0\n", code + 10, code + 20); + let _ = write!(s, " 11\n50\n 21\n-20\n 31\n0\n 70\n{flags}\n 1\n{text}\n 3\nStandard\n100\n{sub}\n"); + for (code, [x, y]) in *pts { + let _ = write!(s, "{code:>3}\n{x}\n{:>3}\n{y}\n{:>3}\n0\n", code + 10, code + 20); } } s + " 0\nENDSEC\n 0\nEOF\n" @@ -1797,13 +1824,17 @@ mod tests { .collect() } - /// Issue #1482: a DWG that once came from MicroStation can carry hundreds of thousands of DGN + /// Issue #1482: a DWG that once came from `MicroStation` can carry hundreds of thousands of DGN /// line style components no definition uses (a real 83 MB site plan: 888,483 of 889,053, a /// 1.2 GB DXF rendition over the limit, so it didn't open). They're left out with a note, and /// their dictionary entries with them; the definition and every component it reaches — /// through compound and point components, around a cycle — are kept. #[cfg(not(no_dwg_bridge))] #[test] + #[expect( + clippy::too_many_lines, + reason = "This integration test builds and verifies one complete DGN component graph; splitting the fixture would obscure reachability." + )] fn orphaned_dgn_line_style_components_are_left_out() { use acadrust::objects::*; let mut doc = acadrust::CadDocument::new(); diff --git a/crates/dxf/src/lib.rs b/crates/dxf/src/lib.rs index 5f18e3291..c4595bef9 100644 --- a/crates/dxf/src/lib.rs +++ b/crates/dxf/src/lib.rs @@ -199,8 +199,9 @@ pub use codepage::decode_code_page; /// # Errors /// /// Returns an error for empty input, an invalid group code, an incomplete group-code/value pair, -/// truncated binary data, or more than [`MAX_TAGS`] group codes. For ASCII input, invalid numeric -/// text becomes `0.0` for real values or `0` for integer values, and an EOF tag is not required. +/// truncated binary data, or more group codes than the limit returned by [`max_tags()`], which is +/// [`MAX_TAGS`] by default. For ASCII input, invalid numeric text becomes `0.0` for real values or +/// `0` for integer values, and an EOF tag is not required. pub fn parse(bytes: &[u8]) -> Result> { parse_with_limit(bytes, max_tags()) } diff --git a/crates/font-catalog/src/embedding.rs b/crates/font-catalog/src/embedding.rs new file mode 100644 index 000000000..42321824e --- /dev/null +++ b/crates/font-catalog/src/embedding.rs @@ -0,0 +1,85 @@ +//! Permission to embed TrueType outline subsets based on the face's OS/2 `fsType`. + +use skrifa::{FontRef, Tag}; + +/// Whether a face can be embedded as a TrueType subset. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub enum SubsetEmbedding { + /// Installable, editable, or preview-and-print embedding, or no readable `fsType`. + Allowed, + /// Restricted licence embedding. + Restricted, + /// Embedding is allowed only without subsetting. + NoSubsetting, + /// Only bitmap embedding is allowed. + BitmapOnly, + /// The face is invalid or lacks a TrueType outline table. + NotTrueType, +} + +/// Classify a face's TrueType subset embedding permissions. Missing or truncated OS/2 +/// data permits embedding; bitmap-only takes precedence over restricted and no-subsetting. +pub fn subset_embedding(bytes: &[u8], index: u32) -> SubsetEmbedding { + let Ok(face) = FontRef::from_index(bytes, index) else { return SubsetEmbedding::NotTrueType }; + if [b"glyf", b"loca", b"head"].iter().any(|tag| face.table_data(Tag::new(tag)).is_none()) { + return SubsetEmbedding::NotTrueType; + } + let fs_type = face.table_data(Tag::new(b"OS/2")).and_then(|table| { + let data = table.as_bytes(); + Some(u16::from_be_bytes([*data.get(8)?, *data.get(9)?])) + }); + let Some(fs_type) = fs_type else { return SubsetEmbedding::Allowed }; + if fs_type & 0x0200 != 0 { + SubsetEmbedding::BitmapOnly + } else if fs_type & 0x000f == 0x0002 { + SubsetEmbedding::Restricted + } else if fs_type & 0x0100 != 0 { + SubsetEmbedding::NoSubsetting + } else { + SubsetEmbedding::Allowed + } +} + +#[cfg(test)] +mod tests { + use super::*; + use write_fonts::{FontBuilder, types::Tag}; + + fn font(fs_type: Option<&[u8]>, missing: Option<&[u8; 4]>) -> Vec { + let mut builder = FontBuilder::new(); + for tag in [b"glyf", b"loca", b"head"] { + if Some(tag) != missing { + builder.add_raw(Tag::new(tag), vec![0; 4]); + } + } + if let Some(data) = fs_type { + builder.add_raw(Tag::new(b"OS/2"), data.to_vec()); + } + builder.build() + } + + #[test] + fn subset_permissions_and_precedence() { + assert_eq!(subset_embedding(b"junk", 0), SubsetEmbedding::NotTrueType); + for tag in [b"glyf", b"loca", b"head"] { + assert_eq!(subset_embedding(&font(None, Some(tag)), 0), SubsetEmbedding::NotTrueType); + } + assert_eq!(subset_embedding(&font(None, None), 1), SubsetEmbedding::NotTrueType); + assert_eq!(subset_embedding(&font(Some(&[0; 9]), None), 0), SubsetEmbedding::Allowed); + for (bits, expected) in [ + (0x0000u16, SubsetEmbedding::Allowed), + (0x0004, SubsetEmbedding::Allowed), + (0x0008, SubsetEmbedding::Allowed), + (0x0006, SubsetEmbedding::Allowed), + (0x0002, SubsetEmbedding::Restricted), + (0x0100, SubsetEmbedding::NoSubsetting), + (0x0102, SubsetEmbedding::Restricted), + (0x0302, SubsetEmbedding::BitmapOnly), + ] { + let mut os2 = [0; 10]; + os2[8..10].copy_from_slice(&bits.to_be_bytes()); + assert_eq!(subset_embedding(&font(Some(&os2), None), 0), expected, "fsType {bits:#06x}"); + } + assert_eq!(subset_embedding(&font(None, None), 0), SubsetEmbedding::Allowed); + } +} diff --git a/crates/font-catalog/src/lib.rs b/crates/font-catalog/src/lib.rs index c787d3ece..045faebfe 100644 --- a/crates/font-catalog/src/lib.rs +++ b/crates/font-catalog/src/lib.rs @@ -21,14 +21,18 @@ //! ``` mod catalog; +mod embedding; mod env; mod fontconfig; mod registry; mod scan; +/// Bounded SFNT face metadata discovery over caller-provided random-access readers. +pub mod sfnt; mod suitcase; pub mod system; pub use catalog::{Catalog, FontData, FontFileError, Limits, Lister, LoadedFace, Match, Watch}; +pub use embedding::{SubsetEmbedding, subset_embedding}; pub use env::{FontEnv, Os, font_dirs, font_store_dir, office_font_dirs}; pub use fontconfig::fontconfig_dirs; pub use scan::{FaceInfo, LocalizedName, Source, is_font_file, style_italic, style_weight, style_width}; diff --git a/crates/font-catalog/src/sfnt.rs b/crates/font-catalog/src/sfnt.rs new file mode 100644 index 000000000..83efa64dd --- /dev/null +++ b/crates/font-catalog/src/sfnt.rs @@ -0,0 +1,178 @@ +//! Minimal SFNT (TrueType / OpenType / collection) header reader for font discovery: reads only +//! the table directory, `name` and `OS/2` of each face, so scanning a system font folder does not +//! load whole files. + +use std::io::{Read, Seek, SeekFrom}; + +/// The tables skrifa reads glyph outlines from: TrueType, CFF, CFF2 and VARC. The renderer draws +/// no glyph of a face that has none of them, such as a face with bitmaps only or with its outlines +/// in Apple's `hvgl` table. +const OUTLINE_TABLES: [&[u8; 4]; 4] = [b"glyf", b"CFF ", b"CFF2", b"VARC"]; + +/// Names and style attributes of one face in a font file. +#[derive(Clone, Debug, PartialEq)] +pub struct FaceNames { + pub index: u32, + pub family: String, + pub style: String, + pub full_name: String, + pub weight: u16, + pub italic: bool, +} + +fn be16(b: &[u8], o: usize) -> Option { + Some(u16::from_be_bytes([*b.get(o)?, *b.get(o + 1)?])) +} +fn be32(b: &[u8], o: usize) -> Option { + Some(u32::from_be_bytes([*b.get(o)?, *b.get(o + 1)?, *b.get(o + 2)?, *b.get(o + 3)?])) +} + +fn read_at(r: &mut R, off: u64, len: usize) -> Option> { + if len > 4 << 20 { + return None; + } + r.seek(SeekFrom::Start(off)).ok()?; + let mut v = vec![0u8; len]; + r.read_exact(&mut v).ok()?; + Some(v) +} + +/// Read the names of every face in a font file (a single font or a collection). Faces with none +/// of the `OUTLINE_TABLES` are left out. +pub fn read_faces(r: &mut R) -> Vec { + let Some(head) = read_at(r, 0, 12) else { return Vec::new() }; + let offsets: Vec = if &head[0..4] == b"ttcf" { + let n = be32(&head, 8).unwrap_or(0).min(256) as usize; + let Some(t) = read_at(r, 12, n * 4) else { return Vec::new() }; + (0..n).filter_map(|i| be32(&t, i * 4)).collect() + } else { + vec![0] + }; + offsets.iter().enumerate().filter_map(|(i, &o)| read_face(r, o as u64, i as u32)).collect() +} + +/// Names from in-memory font data. +pub fn read_faces_bytes(data: &[u8]) -> Vec { + read_faces(&mut std::io::Cursor::new(data)) +} + +fn read_face(r: &mut R, off: u64, index: u32) -> Option { + let h = read_at(r, off, 12)?; + let tag = &h[0..4]; + if tag != [0, 1, 0, 0] && tag != b"OTTO" && tag != b"true" { + return None; + } + let n = be16(&h, 4)? as usize; + let dir = read_at(r, off + 12, n * 16)?; + if !dir.as_chunks::<16>().0.iter().any(|rec| OUTLINE_TABLES.iter().any(|t| rec.starts_with(*t))) { + return None; + } + let mut name = None; + let mut os2 = None; + for i in 0..n { + let rec = &dir[i * 16..i * 16 + 16]; + let (o, l) = (be32(rec, 8)? as u64, be32(rec, 12)? as usize); + match &rec[0..4] { + b"name" => name = Some((o, l)), + b"OS/2" => os2 = Some((o, l)), + _ => {} + } + } + let (no, nl) = name?; + let nb = read_at(r, no, nl)?; + let get = |id: u16| name_string(&nb, id); + let family = get(16).or_else(|| get(1))?; + let style = get(17).or_else(|| get(2)).unwrap_or_else(|| "Regular".into()); + let full_name = get(4).unwrap_or_else(|| format!("{family} {style}")); + let (mut weight, mut italic) = (400u16, false); + if let Some((oo, ol)) = os2 + && let Some(ob) = read_at(r, oo, ol.min(78)) + { + weight = be16(&ob, 4).unwrap_or(400); + italic = be16(&ob, 62).is_some_and(|fs| fs & 1 != 0); + } + let sl = style.to_ascii_lowercase(); + italic |= sl.contains("italic") || sl.contains("oblique"); + Some(FaceNames { index, family, style, full_name, weight, italic }) +} + +/// A name-table string by id, preferring Windows Unicode English, then any Unicode, then Mac Roman. +fn name_string(b: &[u8], id: u16) -> Option { + let count = be16(b, 2)? as usize; + let storage = be16(b, 4)? as usize; + let mut best: Option<(u8, String)> = None; + for i in 0..count { + let r = 6 + i * 12; + let (plat, enc, lang, nid, len, off) = + (be16(b, r)?, be16(b, r + 2)?, be16(b, r + 4)?, be16(b, r + 6)?, be16(b, r + 8)? as usize, be16(b, r + 10)? as usize); + if nid != id { + continue; + } + let rank = match plat { + 3 if enc == 1 || enc == 10 => { + if best.as_ref().is_some_and(|(current, _)| *current <= if lang == 0x409 { 0 } else { 1 }) { + continue; + } + if lang == 0x409 { 0 } else { 1 } + } + 0 => { + if best.as_ref().is_some_and(|(current, _)| *current <= 2) { + continue; + } + 2 + } + 1 if enc == 0 => { + if best.as_ref().is_some_and(|(current, _)| *current <= 3) { + continue; + } + 3 + } + _ => continue, + }; + // Only decode a record that can improve the selected name. A hostile name table may + // point thousands of records at the same huge string. + let raw = b.get(storage + off..storage + off + len)?; + let s = match plat { + 3 => { + let u: Vec = raw.as_chunks::<2>().0.iter().map(|c| u16::from_be_bytes([c[0], c[1]])).collect(); + String::from_utf16_lossy(&u) + } + 0 => { + let u: Vec = raw.as_chunks::<2>().0.iter().map(|c| u16::from_be_bytes([c[0], c[1]])).collect(); + String::from_utf16_lossy(&u) + } + 1 => raw.iter().map(|&c| if c < 128 { c as char } else { '?' }).collect(), + _ => return None, + }; + best = Some((rank, s)); + } + best.map(|(_, s)| s).filter(|s| !s.trim().is_empty()) +} + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn later_non_improving_name_records_are_not_decoded() { + // The first record is the preferred Windows-English value. The second names an invalid + // range: it must be ignored instead of making the whole name lookup fail. + let mut table = vec![0; 30]; + table[2..4].copy_from_slice(&2u16.to_be_bytes()); + table[4..6].copy_from_slice(&30u16.to_be_bytes()); + let first = 6usize; + table[first..first + 2].copy_from_slice(&3u16.to_be_bytes()); + table[first + 2..first + 4].copy_from_slice(&1u16.to_be_bytes()); + table[first + 4..first + 6].copy_from_slice(&0x409u16.to_be_bytes()); + table[first + 6..first + 8].copy_from_slice(&1u16.to_be_bytes()); + table[first + 8..first + 10].copy_from_slice(&4u16.to_be_bytes()); + let second = first + 12; + table[second..second + 2].copy_from_slice(&3u16.to_be_bytes()); + table[second + 2..second + 4].copy_from_slice(&1u16.to_be_bytes()); + table[second + 4..second + 6].copy_from_slice(&0x409u16.to_be_bytes()); + table[second + 6..second + 8].copy_from_slice(&1u16.to_be_bytes()); + table[second + 8..second + 10].copy_from_slice(&u16::MAX.to_be_bytes()); + table.extend_from_slice(&[0, b'A', 0, b'B']); + assert_eq!(name_string(&table, 1).as_deref(), Some("AB")); + } +} diff --git a/crates/font-catalog/src/tests.rs b/crates/font-catalog/src/tests.rs index 5137a4295..7967cc9c8 100644 --- a/crates/font-catalog/src/tests.rs +++ b/crates/font-catalog/src/tests.rs @@ -568,3 +568,20 @@ fn a_read_is_capped_where_the_file_has_grown() { assert!(matches!(read_capped(&path.with_extension("missing"), 9), Err(FontFileError::Io(_)))); let _ = std::fs::remove_file(&path); } + +#[test] +fn bounded_sfnt_reader_preserves_face_names_and_collection_indices() { + let mut bold = spec("Example", "Bold Italic", 700); + bold.italic = true; + let data = font(&bold); + let faces = crate::sfnt::read_faces_bytes(&data); + assert_eq!(faces.len(), 1); + assert_eq!((&*faces[0].family, &*faces[0].style, faces[0].weight, faces[0].italic), ("Example", "Bold Italic", 700, true)); + let mut hidden = spec("Hidden", "Regular", 400); + hidden.outlines = false; + let combined = collection(&[font(&hidden), data]); + let faces = crate::sfnt::read_faces_bytes(&combined); + assert_eq!(faces.len(), 1); + assert_eq!((faces[0].index, faces[0].family.as_str()), (1, "Example")); + assert!(crate::sfnt::read_faces_bytes(b"invalid").is_empty()); +} diff --git a/crates/i18n/src/fold.rs b/crates/i18n/src/fold.rs index 959c6425a..fb420e1e1 100644 --- a/crates/i18n/src/fold.rs +++ b/crates/i18n/src/fold.rs @@ -15,7 +15,7 @@ //! `IRMAK` and `ırmak`. Plain `str::to_lowercase` follows no locale and turns `İ` into `i` plus a //! combining dot, which nothing typed would match. //! -//! The Turkish i handling and sigma folding come from PdfCraft's `fold_case` (snipeTR, +//! The Turkish i handling and sigma folding come from `PdfCraft`'s `fold_case` (snipeTR, //! nevicia2010-hub). use std::ops::Range; @@ -85,7 +85,7 @@ impl Folder { } /// One cluster, already in NFKD. - fn fold_cluster(&self, cluster: &str, out: &mut String) { + fn fold_cluster(self, cluster: &str, out: &mut String) { let mut chars = cluster.chars().peekable(); while let Some(c) = chars.next() { match c { @@ -363,10 +363,10 @@ mod tests { #[test] fn matches_map_back_to_the_source() { - let f = Folder::new("en"); fn find<'a>(f: Folder, haystack: &'a str, needle: &str) -> Option<&'a str> { f.fold(haystack).find(&f.fold(needle)).map(|r| &haystack[r]) } + let f = Folder::new("en"); let find = |haystack, needle| find(f, haystack, needle); assert_eq!(find("Große Straße", "strasse"), Some("Straße")); assert_eq!(find("Große Straße", "SS"), Some("ß")); diff --git a/crates/i18n/src/format.rs b/crates/i18n/src/format.rs index 14eb8c102..b907b469c 100644 --- a/crates/i18n/src/format.rs +++ b/crates/i18n/src/format.rs @@ -4,7 +4,7 @@ use std::fmt; use fixed_decimal::{Decimal, Sign}; -use icu_decimal::{DecimalFormatter, parts}; +use icu_decimal::{DecimalFormatter, options::DecimalFormatterOptions, parts}; use icu_locale_core::Locale; use icu_locale_core::extensions::unicode::{key, value}; use writeable::{Part, PartsWrite, Writeable}; @@ -46,7 +46,8 @@ impl Format { if digits == Digits::Latin { locale.extensions.unicode.keywords.set(key!("nu"), value!("latn")); } - let decimal = data::provider().and_then(|p| DecimalFormatter::try_new_with_buffer_provider(p, (&locale).into(), Default::default()).ok()); + let decimal = data::provider() + .and_then(|p| DecimalFormatter::try_new_with_buffer_provider(p, (&locale).into(), DecimalFormatterOptions::default()).ok()); let mut format = Format { decimal, group: None, point: None, percent: ("", "%"), marks: Marks::default() }; // The locale's own separators and digits, read from a number that shows them all. let probe = format.parts(&Decimal::try_from_str("-1234567.5").unwrap_or_default()); @@ -212,7 +213,7 @@ impl PartsWrite for Sink<'_> { _ => return result, }; if slot.is_empty() { - *slot = written.to_owned(); + written.clone_into(slot); } } result @@ -370,7 +371,7 @@ mod tests { for (tag, want) in [("en", "50%"), ("de", "50\u{a0}%"), ("fr", "12,5\u{a0}%"), ("tr", "%50"), ("ar-EG", "٥٠٪\u{61c}"), ("fa", "ÛµÛ°Ùª"), ("hi", "50%")] { - let decimals = if tag == "fr" { 1 } else { 0 }; + let decimals = usize::from(tag == "fr"); let ratio = if tag == "fr" { 0.125 } else { 0.5 }; assert_eq!(Format::from_tag(tag).percent(ratio, decimals), want, "{tag}"); } diff --git a/crates/i18n/src/lib.rs b/crates/i18n/src/lib.rs index 53bbffe9e..24ce578a6 100644 --- a/crates/i18n/src/lib.rs +++ b/crates/i18n/src/lib.rs @@ -113,7 +113,9 @@ impl Locale { rows.entry(e.context).or_default().insert(e.source, e.translation); } } - problems.iter().for_each(|p| log::warn!("translation catalog: {p}; row skipped")); + for p in &problems { + log::warn!("translation catalog: {p}; row skipped"); + } Parsed { rows, problems } }) } @@ -483,14 +485,13 @@ fn wrap(initiator: char, value: &str) -> String { PDF if open.last() == Some(&false) => { open.pop(); } - PDF => continue, PDI if isolates > 0 => { while open.pop() == Some(false) { out.push(PDF); } isolates -= 1; } - PDI => continue, + PDF | PDI => continue, _ => {} } out.push(c); @@ -713,11 +714,11 @@ fn plural_rules(code: &str) -> Option<(PluralRules, Vec)> { categories.push(c); } } - categories.sort_by_key(rank); + categories.sort_by_key(|c| rank(*c)); Some((rules, categories)) } -fn rank(c: &PluralCategory) -> u8 { +fn rank(c: PluralCategory) -> u8 { match c { PluralCategory::Zero => 0, PluralCategory::One => 1, diff --git a/crates/i18n/src/number.rs b/crates/i18n/src/number.rs index 1120fbe88..211d9dfe0 100644 --- a/crates/i18n/src/number.rs +++ b/crates/i18n/src/number.rs @@ -80,6 +80,12 @@ impl Format { /// /// Percent signs, units and the space before a unit (`50 %`, `12 px`) are not read; strip them /// first. Anything else is an error rather than a guess. + /// + /// # Errors + /// + /// Returns `NumberError::Empty` for an empty input or one reduced to spaces and a sign, + /// `Unexpected` for an invalid character, `MixedDigits` for mixed digit sets, `Separator` + /// for misplaced marks, and `TooLarge` when the parsed value is not finite. pub fn parse(&self, text: &str) -> Result { let marks = self.marks; let text: String = text.chars().filter(|c| !is_bidi_mark(*c)).collect(); @@ -129,6 +135,11 @@ impl Format { /// The number in `text` as `locale` (a language tag, `de-CH`, `ar_EG.UTF-8`) writes it: /// [`Format::parse`] for [`Format::from_tag`]. Build the [`Format`] once to read many numbers. A /// locale CLDR doesn't know reads as CLDR's root does (`1,234.5`). +/// +/// # Errors +/// +/// Returns the same `NumberError` as [`Format::parse`] for invalid digits, characters, or +/// separators, or when the value is not finite. pub fn parse_number(text: &str, locale: &str) -> Result { Format::from_tag(locale).parse(text) } diff --git a/crates/i18n/src/region.rs b/crates/i18n/src/region.rs index fcdffe8af..15874d67f 100644 --- a/crates/i18n/src/region.rs +++ b/crates/i18n/src/region.rs @@ -298,6 +298,7 @@ fn from_apple(identifier: &str, metric: bool, decimal: &str, grouping: &str, hou /// A grouping separator the system sets: none at all when empty (Windows' `sThousand` the user /// cleared), else one character. #[cfg(any(test, windows, target_vendor = "apple"))] +#[expect(clippy::option_option, reason = "The public setting distinguishes following the locale, no grouping, and a custom separator")] fn grouping(s: &str) -> Option> { if s.is_empty() { Some(None) } else { single_char(s).map(Some) } } diff --git a/crates/i18n/src/scripts.rs b/crates/i18n/src/scripts.rs index 1ae0f29bc..aa39d8917 100644 --- a/crates/i18n/src/scripts.rs +++ b/crates/i18n/src/scripts.rs @@ -113,7 +113,10 @@ pub enum Han { /// Latin. // ponytail: the languages a Craft app ships or could plausibly ship; add a row with the locale. pub fn scripts(tag: &str) -> &'static [Script] { - use Script::*; + use Script::{ + Arab, Armn, Beng, Cyrl, Deva, Ethi, Geor, Grek, Gujr, Guru, Hang, Hani, Hebr, Hira, Kana, Khmr, Knda, Laoo, Latn, Mlym, Mymr, Orya, Sinh, + Syrc, Taml, Telu, Thaa, Thai, Tibt, + }; let tag = Subtags::parse(tag); if let Some(s) = tag.script.as_deref().and_then(subtag_scripts) { return s; @@ -175,7 +178,7 @@ pub fn base_direction(text: &str) -> Option { /// The forms a language tag's Han characters take: `zh-CN`, `zh-SG`, `zh-Hans` and plain `zh` → /// Simplified; `zh-TW`, `zh-Hant` and the other regions CLDR writes in Traditional characters -/// (`zh-US`, `zh-AU`) → Traditional; `zh-HK`, `zh-MO`, `zh-Hant-HK` and `yue` → HongKong; `ja` → +/// (`zh-US`, `zh-AU`) → Traditional; `zh-HK`, `zh-MO`, `zh-Hant-HK` and `yue` → `HongKong`; `ja` → /// Japanese; `ko` → Korean. `None` for a language not written with Han, which leaves the choice /// to the font stack. pub fn han(tag: &str) -> Option { @@ -202,8 +205,14 @@ pub fn han(tag: &str) -> Option { // ponytail: mn-CN → Mong is left out until vertical Mongolian layout is supported. pub(crate) fn region_script(language: &str, region: &str) -> Option<&'static str> { Some(match (language, region) { - ("az", "iq" | "ir") | ("ha", "cm" | "sd") | ("kk", "af" | "cn" | "ir" | "mn") | ("ku", "iq" | "ir" | "lb") => "arab", - ("ky", "cn") | ("ms", "cc") | ("pa", "pk") | ("tg", "pk") | ("uz", "af") => "arab", + ("az", "iq" | "ir") + | ("ha", "cm" | "sd") + | ("kk", "af" | "cn" | "ir" | "mn") + | ("ku", "iq" | "ir" | "lb") + | ("ky", "cn") + | ("ms", "cc") + | ("pa" | "tg", "pk") + | ("uz", "af") => "arab", ("az", "ru") | ("ku", "am" | "az" | "ge" | "tm") | ("ug", "kz" | "mn") | ("uz", "cn") => "cyrl", ("ky", "tr") | ("sr", "me" | "ro" | "tr") => "latn", ("sd", "in") => "deva", @@ -216,7 +225,10 @@ pub(crate) fn region_script(language: &str, region: &str) -> Option<&'static str /// The scripts a script subtag stands for, including the CLDR unions (`jpan`, `kore`). `None` for /// one not in [`Script`], so the language's own scripts apply. fn subtag_scripts(subtag: &str) -> Option<&'static [Script]> { - use Script::*; + use Script::{ + Arab, Armn, Beng, Cyrl, Deva, Ethi, Geor, Grek, Gujr, Guru, Hang, Hani, Hebr, Hira, Kana, Khmr, Knda, Laoo, Latn, Mlym, Mong, Mymr, Orya, + Sinh, Syrc, Taml, Telu, Thaa, Thai, Tibt, + }; Some(match subtag { "hans" | "hant" | "hani" => &[Hani], "jpan" => &[Hani, Hira, Kana], diff --git a/crates/i18n/src/size.rs b/crates/i18n/src/size.rs index 830c09573..e0d45d59d 100644 --- a/crates/i18n/src/size.rs +++ b/crates/i18n/src/size.rs @@ -104,6 +104,10 @@ impl FileSize { } /// The unit `bytes` is written in, the number in it and its digits after the point. + #[expect( + clippy::cast_precision_loss, + reason = "File sizes are displayed as rounded f64 values; exact u64 arithmetic is not required for presentation" + )] fn parts(&self, bytes: u64) -> (SizeUnit, f64, usize) { let (smallest, largest) = (self.smallest.min(self.largest), self.largest); let mut unit = smallest; @@ -137,6 +141,7 @@ impl FileSize { } } + #[expect(clippy::cast_precision_loss, reason = "Unit scales are displayed as f64 values; exact u64 arithmetic is not required for presentation")] fn scale(&self, unit: SizeUnit) -> f64 { (self.step as f64).powi(unit as i32) } diff --git a/crates/image-codecs/Cargo.toml b/crates/image-codecs/Cargo.toml index a8872c0d3..be5cd57e1 100644 --- a/crates/image-codecs/Cargo.toml +++ b/crates/image-codecs/Cargo.toml @@ -15,6 +15,9 @@ corpus = [] default = [] avif = ["image/avif", "dep:ravif", "dep:rav1e", "dep:avif-serialize"] heif = ["dep:craft-heif"] +# AVIF decode: craft-video's clean-room pure-Rust AV1 decoder. Native targets only (on wasm32 a +# decoder panic aborts instead of becoming an error), so wasm32 builds leave AVIF unreadable. +avif-decode = ["dep:craft-video", "craft-video/av1-decode"] [dependencies] crc32fast = "1.5" @@ -32,6 +35,7 @@ zune-core = "0.5" jpeg-encoder = "0.6" craft-jpeg = { workspace = true } craft-metadata = { workspace = true } +craft-tiff = { workspace = true } zune-jpegxl = { version = "=0.5.2", default-features = false, features = ["std"] } craft-heif = { workspace = true, optional = true } ravif = { version = "0.13", default-features = false, optional = true } @@ -41,6 +45,7 @@ hayro-jpeg2000 = { version = "=0.3.5", default-features = false, features = ["st [target.'cfg(not(target_arch = "wasm32"))'.dependencies] rayon = { workspace = true } +craft-video = { workspace = true, optional = true } exr = { version = "1.73", default-features = false, features = ["rayon"] } [dev-dependencies] diff --git a/crates/image-codecs/README.md b/crates/image-codecs/README.md index aa0027666..a0e1afaa4 100644 --- a/crates/image-codecs/README.md +++ b/crates/image-codecs/README.md @@ -1,9 +1,10 @@ # craft-image-codecs Codecs for flat raster image formats, written in pure Rust. This crate defines its own `Image` -type and shares its baseline and extended JPEG decoders through `craft-jpeg`, and its optional -HEIF decoder through `craft-heif`. It builds for `wasm32-unknown-unknown` and does no filesystem -I/O: decoders take `&[u8]` and encoders return `Vec`. +type and shares its baseline and extended JPEG decoders through `craft-jpeg`, its optional HEIF +decoder through `craft-heif`, and its optional AVIF decoder through `craft-video`'s AV1 decoder. +It builds for `wasm32-unknown-unknown` and does no filesystem I/O: decoders take `&[u8]` and +encoders return `Vec`. It follows the "Avoiding GIMP's hole" rules in `docs/architecture.md` §1.1: @@ -57,7 +58,7 @@ craft-image-codecs = { git = "https://github.com/storytold/craft-libs", rev = "< | 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` | +| AVIF | with feature `avif-decode` (native only) | only with feature `avif` | U8, U16 (10/12-bit decodes to U16; PQ/HLG to linear F32) | RGB, RGBA | yes (auxiliary alpha) | read (`colr` prof/rICC) | read | read | no | no | yes | `craft-video` AV1 decoder; `image` + `ravif` to write | | 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` | "Native" means the data is stored and read back without conversion. Anything else is converted by @@ -178,10 +179,20 @@ embedded table (and a PNG's transparent entry) so `photocraft-io` can open them decodes to RGB. Image sequences, overlays and identity derivations, multilayer HEVC and some 4:2:2/4:4:4 streams return `CodecError::Unsupported` or `Malformed`, never wrong pixels. Writing needs an HEVC encoder and every mature one is C, so HEIF is listed in `ASYMMETRIC_EXCEPTIONS`. -* **AVIF.** Encoding uses `ravif`, which is pure Rust. Decoding - needs `dav1d`, which is C. AVIF is therefore read-unsupported, and write support is gated - behind the non-default `avif` feature. In a default build it is neither readable nor writable, - so the symmetric guarantee holds. It is listed in `ASYMMETRIC_EXCEPTIONS`. +* **AVIF.** Encoding uses `ravif`, which is pure Rust, behind the non-default `avif` feature. + Decoding (`codecs::avif_decode`, behind the non-default `avif-decode` feature) reads the HEIF + boxes itself and decodes the AV1 pictures with `craft-video`'s clean-room pure-Rust AV1 decoder: + one `av01` image or a `grid`, 8/10/12-bit, 4:2:0/4:2:2/4:4:4/monochrome, the `nclx` colour (or + the sequence header's) and an ICC profile, an `auxl` alpha plane (`prem` too), the Exif and XMP + items, and `clap`/`irot`/`imir`, which are applied unless `keep_orientation` is set, as for HEIF. + PQ and HLG open as linear light in F32, as HEIF does. Image sequences (animated AVIF), depth and + gain-map auxiliary images and layered progressive images aren't read. It is bounded like the + other decoders: the `Limits` are checked against the declared size before any picture decodes, + against every picture and every output buffer, `max_alloc` caps the AV1 decoder's picture store, + box, item and extent counts are capped, item data can't exceed the file, and a decoder panic is + `CodecError::Malformed`. wasm32 builds leave AVIF unreadable, since a panic there aborts. In a + default build AVIF is neither readable nor writable, so the symmetric guarantee holds; a build + with only one of the two features is listed in `ASYMMETRIC_EXCEPTIONS`. * **Lossy WebP.** Lossless (VP8L, via `image-webp`) is the default. `webp_lossless: false` writes a lossy file with our own clean-room VP8 key-frame encoder (`codecs::vp8`, from RFC 6386; `webp_quality` 0–100 on libwebp's scale sets the quantizer). Alpha travels losslessly diff --git a/crates/image-codecs/fuzz/Cargo.toml b/crates/image-codecs/fuzz/Cargo.toml index 3257a3953..30226c76e 100644 --- a/crates/image-codecs/fuzz/Cargo.toml +++ b/crates/image-codecs/fuzz/Cargo.toml @@ -14,7 +14,7 @@ cargo-fuzz = true [dependencies] libfuzzer-sys = "0.4" -craft-image-codecs = { path = "..", features = ["heif"] } +craft-image-codecs = { path = "..", features = ["heif", "avif-decode"] } # Called directly too: the codec turns its panics into errors (see decode_heif.rs). heic-rs = "=0.1.4" @@ -76,3 +76,10 @@ path = "fuzz_targets/decode_webp.rs" test = false doc = false bench = false + +[[bin]] +name = "decode_avif" +path = "fuzz_targets/decode_avif.rs" +test = false +doc = false +bench = false diff --git a/crates/image-codecs/fuzz/fuzz_targets/decode_avif.rs b/crates/image-codecs/fuzz/fuzz_targets/decode_avif.rs new file mode 100644 index 000000000..9a4b8978d --- /dev/null +++ b/crates/image-codecs/fuzz/fuzz_targets/decode_avif.rs @@ -0,0 +1,14 @@ +#![no_main] + +use libfuzzer_sys::fuzz_target; +use craft_image_codecs::{decode_as_with, probe_size, DecodeOptions, Format, Limits}; + +// The AVIF box reader and craft-video's AV1 decoder parse untrusted files: run with `-timeout=10` +// so a hang is reported, not just a panic. The codec turns an AV1 decoder panic into an error, so +// only the box reader's own panics would reach the fuzzer. Both orientation paths are covered. +fuzz_target!(|data: &[u8]| { + let keep_orientation = data.first().is_some_and(|b| b & 1 == 1); + let limits = Limits { max_width: 4096, max_height: 4096, max_pixels: 1 << 22, max_alloc: 256 << 20 }; + let _ = probe_size(data); + let _ = decode_as_with(Format::Avif, data, &DecodeOptions { limits, keep_orientation }); +}); diff --git a/crates/image-codecs/src/allocation.rs b/crates/image-codecs/src/allocation.rs index fe98122e7..df9afd93f 100644 --- a/crates/image-codecs/src/allocation.rs +++ b/crates/image-codecs/src/allocation.rs @@ -247,3 +247,83 @@ fn read_le24(bytes: &[u8], at: usize) -> Option { let b = bytes.get(at..at.checked_add(3)?)?; Some(u32::from_le_bytes([b[0], b[1], b[2], 0])) } + +/// Probe the stored dimensions of the first PNG, JPEG, BMP or TIFF image, without decoding +/// pixels or applying orientation. TIFF uses the first directory even if it is a thumbnail. +/// The caller's limits bound both the header decoder and the declared output dimensions. +pub fn probe_stored_size(bytes: &[u8], limits: &crate::Limits) -> Result<(u32, u32), CodecError> { + use std::io::Cursor; + let format = crate::detect(bytes).ok_or(CodecError::UnknownFormat)?; + if !matches!(format, Format::Png | Format::Jpeg | Format::Bmp | Format::Tiff) { + return Err(CodecError::unsupported(format, "stored-size probe is not available for this format")); + } + // The JPEG header reader owns a copy of its input before applying image limits. + if format == Format::Jpeg && u64::try_from(bytes.len()).map_or(true, |len| len > limits.max_alloc) { + return Err(CodecError::malformed(format, "JPEG input exceeds the allocation budget")); + } + let image_format = match format { + Format::Png => image::ImageFormat::Png, + Format::Jpeg => image::ImageFormat::Jpeg, + Format::Bmp => image::ImageFormat::Bmp, + Format::Tiff => image::ImageFormat::Tiff, + _ => return Err(CodecError::unsupported(format, "stored-size probe is not available for this format")), + }; + // Preserve detect's BigTIFF selection; image's signature guesser only knows classic TIFF. + let mut reader = image::ImageReader::with_format(Cursor::new(bytes), image_format); + let mut image_limits = image::Limits::default(); + image_limits.max_image_width = Some(limits.max_width); + image_limits.max_image_height = Some(limits.max_height); + image_limits.max_alloc = Some(limits.max_alloc); + reader.limits(image_limits); + let (w, h) = reader.into_dimensions().map_err(|e| CodecError::malformed(format, e))?; + limits.check_size(w, h)?; + Ok((w, h)) +} + +#[cfg(test)] +mod stored_size_tests { + #![allow(clippy::unwrap_used)] + use super::*; + use std::io::Cursor; + + #[test] + fn stored_probe_uses_caller_limits_without_decoding() { + let mut bmp = Cursor::new(Vec::new()); + image::write_buffer_with_format(&mut bmp, &vec![0; 60_000 * 3], 1, 60_000, image::ExtendedColorType::Rgb8, image::ImageFormat::Bmp).unwrap(); + let bytes = bmp.into_inner(); + let limits = crate::Limits::new(u32::MAX, u32::MAX, 128_000_000, 2_048_000_000); + assert_eq!(probe_stored_size(&bytes, &limits).unwrap(), (1, 60_000)); + assert!(probe_stored_size(&bytes, &crate::Limits::new(u32::MAX, 50_000, u64::MAX, 2_048_000_000)).is_err()); + assert!(probe_stored_size(&bytes[..20], &limits).is_err()); + } + + #[test] + fn bigtiff_stored_dimensions_preserve_detected_format() { + let mut out = Cursor::new(Vec::new()); + tiff::encoder::TiffEncoder::new_big(&mut out).unwrap().write_image::(3, 2, &[10; 18]).unwrap(); + let bytes = out.into_inner(); + assert_eq!(crate::detect(&bytes), Some(Format::Tiff)); + assert_eq!(probe_stored_size(&bytes, &crate::Limits::default()).unwrap(), (3, 2)); + } + + #[test] + fn jpeg_input_copy_is_bounded_by_allocation_budget() { + let mut out = Cursor::new(Vec::new()); + image::write_buffer_with_format(&mut out, &[10; 18], 3, 2, image::ExtendedColorType::Rgb8, image::ImageFormat::Jpeg).unwrap(); + let bytes = out.into_inner(); + let mut limits = crate::Limits { max_alloc: bytes.len() as u64, ..Default::default() }; + assert_eq!(probe_stored_size(&bytes, &limits).unwrap(), (3, 2)); + limits.max_alloc -= 1; + assert!(probe_stored_size(&bytes, &limits).is_err()); + } + + #[test] + fn common_stored_headers_are_probed() { + let limits = crate::Limits::new(100, 100, 10_000, 1 << 20); + for format in [image::ImageFormat::Png, image::ImageFormat::Jpeg, image::ImageFormat::Bmp, image::ImageFormat::Tiff] { + let mut out = Cursor::new(Vec::new()); + image::write_buffer_with_format(&mut out, &[10; 18], 3, 2, image::ExtendedColorType::Rgb8, format).unwrap(); + assert_eq!(probe_stored_size(&out.into_inner(), &limits).unwrap(), (3, 2), "{format:?}"); + } + } +} diff --git a/crates/image-codecs/src/cicp.rs b/crates/image-codecs/src/cicp.rs index 290c82bc3..e6911d1b0 100644 --- a/crates/image-codecs/src/cicp.rs +++ b/crates/image-codecs/src/cicp.rs @@ -99,7 +99,7 @@ pub fn linear_to_pq(linear: f32) -> f32 { /// Relative luminance weights of the RGB primaries `code` (BT.2020's for anything else, the /// space BT.2100 defines HLG in). -fn luma_weights(code: u16) -> [f32; 3] { +pub(crate) fn luma_weights(code: u16) -> [f32; 3] { match code { 1 => [0.2126, 0.7152, 0.0722], 12 => [0.2290, 0.6917, 0.0793], diff --git a/crates/image-codecs/src/codecs.rs b/crates/image-codecs/src/codecs.rs index af91aa536..00b94b7ec 100644 --- a/crates/image-codecs/src/codecs.rs +++ b/crates/image-codecs/src/codecs.rs @@ -1,4 +1,6 @@ pub(crate) mod avif; +#[cfg(all(feature = "avif-decode", not(target_arch = "wasm32")))] +pub(crate) mod avif_decode; pub(crate) mod dds; pub(crate) mod dds_bc; pub(crate) mod deep_exr; diff --git a/crates/image-codecs/src/codecs/avif_decode.rs b/crates/image-codecs/src/codecs/avif_decode.rs new file mode 100644 index 000000000..382f6a474 --- /dev/null +++ b/crates/image-codecs/src/codecs/avif_decode.rs @@ -0,0 +1,1350 @@ +//! AVIF (AV1 Image File Format) decoding, behind the non-default `avif-decode` feature, native +//! targets only. +//! +//! The container is HEIF (ISO/IEC 23008-12) over ISO-BMFF. This module reads the boxes itself +//! (`meta`, `pitm`, `iloc`, `iinf`, `iref`, `iprp`/`ipco`/`ipma`, `idat`), and `craft-video`'s +//! clean-room AV1 decoder (`av1-decode`, pure Rust, no `unsafe`) decodes the AV1 still pictures. +//! The container reader, colour conversion and transforms are adapted from @emmertex's LightCraft +//! AVIF importer (storytold/lightcraft#3017). +//! +//! Supported: the primary item as one `av01` image or a `grid` of them; 8-, 10- and 12-bit; +//! 4:2:0, 4:2:2, 4:4:4 and monochrome; full or limited range; the `nclx` colour description (or the +//! AV1 sequence header's) and an ICC profile (`prof` / `rICC`); an alpha auxiliary image (`auxl`, +//! premultiplied with `prem` too); the primary item's Exif and XMP items. `clap`, `irot` and `imir` +//! are applied unless the caller keeps the stored orientation, as for HEIF. PQ and HLG files +//! open as linear light in 32-bit float, as HEIF does. Not read: image sequences (animated AVIF), +//! depth and gain-map auxiliary images, layered (`lsel`/`a1lx`) progressive images. +//! +//! Bounded for hostile input: box nesting, item, property and extent counts are capped; item data +//! can't add up to more than the file; the [`Limits`] are checked against the declared size +//! before any picture is decoded, against every decoded picture and against each output buffer; +//! the AV1 decoder's picture store and the pictures one chunk may show are capped by `max_alloc` +//! too; repeated transforms are refused; and a panic in the decoder is +//! a malformed-file error. wasm32 builds leave AVIF unreadable: a panic there aborts instead of +//! unwinding. +//! +//! Sources: ISO/IEC 23008-12 (HEIF) and the AV1 Image File Format specification v1.1.0 (box +//! layouts), ITU-T H.273 (colour code points). + +use std::panic::{AssertUnwindSafe, catch_unwind}; + +use craft_video::av1::decode as av1; + +use crate::Format; +use crate::cicp::Cicp; +use crate::error::CodecError; +use crate::image::{ChannelLayout, Image, Metadata, SampleType}; +use crate::options::Limits; + +const F: Format = Format::Avif; +/// Deepest box nesting followed (meta → iprp → ipco → property is 4). +const MAX_DEPTH: u32 = 8; +/// Most boxes, items, properties, associations or extents accepted: far above real files. +const MAX_ITEMS: usize = 1 << 16; +/// The AV1 decoder's own picture-store cap (`craft-video`'s default on 64-bit targets). +const AV1_MEMORY_BUDGET: u64 = 4 << 30; +/// The `auxC` types of an alpha plane (AVIF and the older HEVC-in-HEIF spelling). +const ALPHA_URNS: [&[u8]; 2] = [b"urn:mpeg:mpegB:cicp:systems:auxiliary:alpha", b"urn:mpeg:hevc:2015:auxid:1"]; + +type Result = std::result::Result; + +fn bad(m: impl std::fmt::Display) -> CodecError { + CodecError::malformed(F, m) +} + +/// A big-endian reader over a byte slice that never panics. +struct Rd<'a> { + b: &'a [u8], + p: usize, +} + +impl<'a> Rd<'a> { + fn new(b: &'a [u8]) -> Self { + Rd { b, p: 0 } + } + fn take(&mut self, n: usize) -> Result<&'a [u8]> { + let end = self.p.checked_add(n).ok_or_else(|| bad("box overflows"))?; + let s = self.b.get(self.p..end).ok_or_else(|| bad("box truncated"))?; + self.p = end; + Ok(s) + } + fn u8(&mut self) -> Result { + Ok(self.take(1)?.first().copied().unwrap_or(0)) + } + fn u16(&mut self) -> Result { + Ok(self.uint(2)? as u16) + } + fn u32(&mut self) -> Result { + Ok(self.uint(4)? as u32) + } + /// An unsigned big-endian number of `n` (0 to 8) bytes. + fn uint(&mut self, n: usize) -> Result { + if n > 8 { + return Err(bad("field too wide")); + } + Ok(self.take(n)?.iter().fold(0u64, |a, &x| (a << 8) | u64::from(x))) + } + /// A full box's version, its flags skipped. + fn version(&mut self) -> Result { + let v = self.u8()?; + self.take(3)?; + Ok(v) + } + /// An item ID: 16 bits in version 0 boxes, 32 bits otherwise. + fn id(&mut self, wide: bool) -> Result { + if wide { self.u32() } else { self.u16().map(u32::from) } + } + fn cstr(&mut self) -> Result<&'a [u8]> { + let rest = self.rest(); + let n = rest.iter().position(|&c| c == 0).unwrap_or(rest.len()); + let s = self.take(n)?; + if self.p < self.b.len() { + self.p += 1; + } + Ok(s) + } + fn rest(&self) -> &'a [u8] { + self.b.get(self.p..).unwrap_or(&[]) + } +} + +/// The child boxes of `b`: (type, payload). +fn boxes(b: &[u8]) -> Result> { + let mut out = Vec::new(); + let mut p = 0usize; + while p < b.len() { + let mut r = Rd::new(b.get(p..).unwrap_or(&[])); + if r.b.len() < 8 { + break; // trailing padding + } + let size32 = r.u32()?; + let t = r.take(4)?; + let ty = [t[0], t[1], t[2], t[3]]; + let (hdr, size) = match size32 { + 0 => (8usize, r.b.len() as u64), + 1 => (16, r.uint(8)?), + s => (8, u64::from(s)), + }; + let size = usize::try_from(size).map_err(|_| bad("box too large"))?; + if size < hdr || size > r.b.len() { + return Err(bad(format!("box {} has a bad size", String::from_utf8_lossy(&ty)))); + } + out.push((ty, r.b.get(hdr..size).unwrap_or(&[]))); + if out.len() > MAX_ITEMS { + return Err(bad("too many boxes")); + } + p = p.checked_add(size).ok_or_else(|| bad("box overflows"))?; + } + Ok(out) +} + +#[derive(Clone, Debug, Default)] +struct Extent { + offset: u64, + length: u64, +} + +#[derive(Clone, Debug, Default)] +struct Loc { + method: u8, + base: u64, + extents: Vec, +} + +#[derive(Clone, Copy, Debug, PartialEq)] +enum Transform { + /// Clean aperture: width, height, horizontal and vertical offset of the centre. + Clap { w: f64, h: f64, dx: f64, dy: f64 }, + /// Anticlockwise rotation by `n` × 90°. + Rot(u8), + /// Mirror, as HEIF amendment 2, libheif and libavif read it: 0 = top and bottom exchanged, + /// 1 = left and right exchanged. + Mir(u8), +} + +#[derive(Clone, Debug)] +enum Prop { + Av1c(Vec), + Nclx(Cicp), + Icc(Vec), + Ispe(u32, u32), + AuxC(Vec), + T(Transform), + Other, +} + +struct Item { + id: u32, + kind: [u8; 4], + /// `mime` items: the content type. + content_type: Option>, +} + +#[derive(Default)] +struct Container<'a> { + file: &'a [u8], + primary: u32, + items: Vec, + locs: Vec<(u32, Loc)>, + idat: &'a [u8], + props: Vec, + /// item → 1-based property indices, in order. + assoc: Vec<(u32, Vec)>, + /// (type, from, to…) + refs: Vec<([u8; 4], u32, Vec)>, + /// `iloc` extents read so far, across all items (capped at [`MAX_ITEMS`]). + extents: usize, +} + +fn parse(file: &[u8]) -> Result> { + let top = boxes(file)?; + let meta = top.iter().find(|(t, _)| t == b"meta").map(|(_, b)| *b).ok_or_else(|| { + if top.iter().any(|(t, _)| t == b"moov") { + CodecError::unsupported(F, "AVIF image sequences (animated AVIF) aren't read") + } else { + bad("no meta box") + } + })?; + let mut c = Container { file, ..Default::default() }; + let meta = meta.get(4..).ok_or_else(|| bad("meta box truncated"))?; + for (t, b) in boxes(meta)? { + let mut r = Rd::new(b); + match &t { + b"pitm" => { + let v = r.version()?; + c.primary = r.id(v != 0)?; + } + b"iloc" => parse_iloc(&mut r, &mut c)?, + b"iinf" => parse_iinf(&mut r, &mut c)?, + b"iref" => parse_iref(&mut r, &mut c)?, + b"idat" => c.idat = b, + b"iprp" => parse_iprp(b, &mut c, 0)?, + _ => {} + } + } + if c.primary == 0 { + return Err(bad("no primary item")); + } + Ok(c) +} + +fn parse_iinf(r: &mut Rd, c: &mut Container) -> Result<()> { + let v = r.version()?; + let _count = r.id(v != 0)?; + for (t, b) in boxes(r.rest())? { + if &t != b"infe" { + continue; + } + let mut r = Rd::new(b); + let v = r.version()?; + if v < 2 { + continue; + } + let id = r.id(v != 2)?; + r.u16()?; // protection index + let kind = r.take(4)?; + let kind = [kind[0], kind[1], kind[2], kind[3]]; + let _name = r.cstr()?; + let content_type = (&kind == b"mime").then(|| r.cstr().ok().map(<[u8]>::to_vec)).flatten(); + c.items.push(Item { id, kind, content_type }); + if c.items.len() > MAX_ITEMS { + return Err(bad("too many items")); + } + } + Ok(()) +} + +fn parse_iref(r: &mut Rd, c: &mut Container) -> Result<()> { + let v = r.version()?; + for (t, b) in boxes(r.rest())? { + let mut r = Rd::new(b); + let from = r.id(v != 0)?; + let n = r.u16()?; + let mut to = Vec::with_capacity(usize::from(n).min(r.rest().len() / 2 + 1)); + for _ in 0..n { + to.push(r.id(v != 0)?); + } + c.refs.push((t, from, to)); + if c.refs.len() > MAX_ITEMS { + return Err(bad("too many item references")); + } + } + Ok(()) +} + +fn parse_iloc(r: &mut Rd, c: &mut Container) -> Result<()> { + let v = r.version()?; + if v > 2 { + return Err(CodecError::unsupported(F, "iloc version")); + } + let a = r.u8()?; + let b = r.u8()?; + let (off_size, len_size, base_size) = (usize::from(a >> 4), usize::from(a & 15), usize::from(b >> 4)); + let index_size = if v >= 1 { usize::from(b & 15) } else { 0 }; + let count = if v < 2 { u32::from(r.u16()?) } else { r.u32()? }; + for _ in 0..count { + let id = r.id(v >= 2)?; + let method = if v >= 1 { (r.u16()? & 15) as u8 } else { 0 }; + r.u16()?; // data reference index + let base = r.uint(base_size)?; + let n = r.u16()?; + let mut extents = Vec::with_capacity(usize::from(n).min(64)); + for _ in 0..n { + r.uint(index_size)?; + let offset = r.uint(off_size)?; + let length = r.uint(len_size)?; + extents.push(Extent { offset, length }); + // Zero-width offset and length fields read no bytes, so the file size alone doesn't + // bound this. + c.extents += 1; + if c.extents > MAX_ITEMS { + return Err(bad("too many item extents")); + } + } + c.locs.push((id, Loc { method, base, extents })); + if c.locs.len() > MAX_ITEMS { + return Err(bad("too many items")); + } + } + Ok(()) +} + +fn parse_iprp(b: &[u8], c: &mut Container, depth: u32) -> Result<()> { + if depth > MAX_DEPTH { + return Err(bad("boxes nested too deep")); + } + for (t, b) in boxes(b)? { + match &t { + b"ipco" => { + for (t, b) in boxes(b)? { + c.props.push(parse_prop(&t, b)?); + if c.props.len() > MAX_ITEMS { + return Err(bad("too many properties")); + } + } + } + b"ipma" => parse_ipma(b, c)?, + _ => {} + } + } + Ok(()) +} + +fn parse_ipma(b: &[u8], c: &mut Container) -> Result<()> { + let mut r = Rd::new(b); + let v = r.u8()?; + let flags = r.take(3)?; + let wide = flags.get(2).is_some_and(|f| f & 1 != 0); + let n = r.u32()?; + for _ in 0..n { + let id = r.id(v != 0)?; + let k = r.u8()?; + let mut list = Vec::with_capacity(usize::from(k)); + for _ in 0..k { + let idx = if wide { usize::from(r.u16()? & 0x7fff) } else { usize::from(r.u8()? & 0x7f) }; + list.push(idx); + } + c.assoc.push((id, list)); + if c.assoc.len() > MAX_ITEMS { + return Err(bad("too many property associations")); + } + } + Ok(()) +} + +fn parse_prop(t: &[u8; 4], b: &[u8]) -> Result { + let mut r = Rd::new(b); + Ok(match t { + b"av1C" => Prop::Av1c(b.get(4..).unwrap_or(&[]).to_vec()), + b"colr" => match r.take(4)? { + b"nclx" => Prop::Nclx(Cicp { primaries: r.u16()?, transfer: r.u16()?, matrix: r.u16()?, full_range: r.u8()? & 0x80 != 0 }), + b"prof" | b"rICC" => Prop::Icc(r.rest().to_vec()), + _ => Prop::Other, + }, + b"ispe" => { + r.version()?; + Prop::Ispe(r.u32()?, r.u32()?) + } + b"auxC" => { + r.version()?; + Prop::AuxC(r.cstr()?.to_vec()) + } + b"irot" => Prop::T(Transform::Rot(r.u8()? & 3)), + b"imir" => Prop::T(Transform::Mir(r.u8()? & 1)), + b"clap" => { + // Width and height are unsigned fractions; the centre offsets have signed numerators + // (and unsigned denominators), as libavif and libheif read them. + let mut q = |signed: bool| -> Result { + let raw = r.u32()?; + let n = if signed { f64::from(raw as i32) } else { f64::from(raw) }; + let d = r.u32()?; + if d == 0 { + return Err(bad("clap with a zero denominator")); + } + Ok(n / f64::from(d)) + }; + let (w, h, dx, dy) = (q(false)?, q(false)?, q(true)?, q(true)?); + Prop::T(Transform::Clap { w, h, dx, dy }) + } + _ => Prop::Other, + }) +} + +impl Container<'_> { + fn item_type(&self, id: u32) -> Option<[u8; 4]> { + self.items.iter().find(|t| t.id == id).map(|t| t.kind) + } + fn props_of(&self, id: u32) -> impl Iterator { + self.assoc.iter().filter(move |a| a.0 == id).flat_map(|a| a.1.iter()).filter_map(|&i| i.checked_sub(1).and_then(|i| self.props.get(i))) + } + fn data(&self, id: u32) -> Result> { + let loc = self.locs.iter().find(|l| l.0 == id).map(|l| &l.1).ok_or_else(|| bad(format!("item {id} has no location")))?; + let src = match loc.method { + 0 => self.file, + 1 => self.idat, + _ => return Err(CodecError::unsupported(F, "item data by reference to another item")), + }; + let mut out = Vec::new(); + for e in &loc.extents { + let start = loc.base.checked_add(e.offset).ok_or_else(|| bad("item offset overflows"))?; + let start = usize::try_from(start).map_err(|_| bad("item offset out of range"))?; + let end = if e.length == 0 { + src.len() + } else { + usize::try_from(e.length).ok().and_then(|l| start.checked_add(l)).ok_or_else(|| bad("item length overflows"))? + }; + let chunk = src.get(start..end).ok_or_else(|| bad("item data runs past the end of the file"))?; + // Every byte comes from the file, so a real item is never larger than it; repeated or + // overlapping extents that add up to more are refused before they allocate. + if out.len().saturating_add(chunk.len()) > self.file.len() { + return Err(bad("item data is larger than the file")); + } + out.extend_from_slice(chunk); + } + Ok(out) + } + fn refs_from(&self, kind: &[u8; 4], from: u32) -> Vec { + self.refs.iter().filter(|r| &r.0 == kind && r.1 == from).flat_map(|r| r.2.iter().copied()).collect() + } + /// Items with a `kind` reference to `to`. + fn refs_to(&self, kind: &[u8; 4], to: u32) -> impl Iterator { + self.refs.iter().filter(move |r| &r.0 == kind && r.2.contains(&to)).map(|r| r.1) + } + fn ispe(&self, id: u32) -> Option<(u32, u32)> { + self.props_of(id).find_map(|p| if let Prop::Ispe(w, h) = p { Some((*w, *h)) } else { None }) + } + /// The item's transformative properties in association order. HEIF allows each of `clap`, + /// `irot` and `imir` at most once per item; a file that associates one more often (repeated + /// `ipma` entries could otherwise ask for a quarter of a million full-image rotations) is + /// refused before any pixels are touched. + fn transforms(&self, id: u32) -> Result> { + // Three is the most a valid item has, so a fourth is already a repeat. + let ts: Vec = self.props_of(id).filter_map(|p| if let Prop::T(t) = p { Some(*t) } else { None }).take(4).collect(); + for (i, t) in ts.iter().enumerate() { + if ts[..i].iter().any(|u| std::mem::discriminant(u) == std::mem::discriminant(t)) { + return Err(bad("an image item has a transformative property (clap, irot or imir) more than once")); + } + } + Ok(ts) + } + fn nclx(&self, id: u32) -> Option { + self.props_of(id).find_map(|p| if let Prop::Nclx(n) = p { Some(*n) } else { None }) + } + fn icc(&self, id: u32) -> Option> { + self.props_of(id).find_map(|p| if let Prop::Icc(b) = p { Some(b.clone()) } else { None }) + } + /// The primary item's alpha plane: an auxiliary image (`auxl`) whose `auxC` says alpha. + fn alpha_item(&self) -> Option { + self.refs_to(b"auxl", self.primary).find(|&id| { + matches!(self.item_type(id), Some(t) if &t == b"av01" || &t == b"grid") + && self.props_of(id).any(|p| matches!(p, Prop::AuxC(urn) if ALPHA_URNS.contains(&urn.as_slice()))) + }) + } + /// The primary item's colour is premultiplied by its alpha (`prem` from colour to alpha). + fn premultiplied(&self, alpha: u32) -> bool { + self.refs_from(b"prem", self.primary).contains(&alpha) + } + /// The first item of a metadata kind that describes (`cdsc`) the primary item. + fn metadata_item(&self, is: impl Fn(&Item) -> bool) -> Option { + let described: Vec = self.refs_to(b"cdsc", self.primary).collect(); + self.items.iter().filter(|i| is(i)).map(|i| i.id).find(|id| described.contains(id)) + } +} + +/// The size of an image item (`av01` or `grid`) before its transforms. +fn item_size(c: &Container, id: u32) -> Result<(u32, u32)> { + let kind = c.item_type(id).ok_or_else(|| bad("an image item is not described"))?; + let (w, h) = match (c.ispe(id), &kind) { + (_, b"grid") => grid_header(&c.data(id)?)?.2, + (Some(s), b"av01") => s, + (None, b"av01") => return Err(bad("an AV1 image item has no size (ispe)")), + _ => return Err(CodecError::unsupported(F, "the image item is not an AV1 image or a grid")), + }; + if w == 0 || h == 0 { + return Err(bad("zero dimension")); + } + Ok((w, h)) +} + +/// The size after `clap` / `irot`, from a stored size. +fn displayed(mut w: u32, mut h: u32, ts: &[Transform]) -> (u32, u32) { + for t in ts { + match *t { + Transform::Clap { w: cw, h: ch, .. } if cw >= 1.0 && ch >= 1.0 && cw.is_finite() && ch.is_finite() => { + w = (cw.round() as u32).clamp(1, w); + h = (ch.round() as u32).clamp(1, h); + } + Transform::Rot(n) if n % 2 == 1 => std::mem::swap(&mut w, &mut h), + _ => {} + } + } + (w, h) +} + +/// (rows, columns, output size) of a `grid` item. +fn grid_header(d: &[u8]) -> Result<(u32, u32, (u32, u32))> { + let mut r = Rd::new(d); + r.u8()?; // version + let flags = r.u8()?; + let rows = u32::from(r.u8()?) + 1; + let cols = u32::from(r.u8()?) + 1; + let n = if flags & 1 != 0 { 4 } else { 2 }; + let w = r.uint(n)? as u32; + let h = r.uint(n)? as u32; + Ok((rows, cols, (w, h))) +} + +/// The upright size [`decode`] returns (`keep_orientation` off), from the boxes alone. +pub(crate) fn probe(bytes: &[u8], limits: &Limits) -> Result<(u32, u32)> { + let c = parse(bytes)?; + let (w, h) = item_size(&c, c.primary)?; + limits.check_size(w, h)?; + Ok(displayed(w, h, &c.transforms(c.primary)?)) +} + +/// One decoded AV1 picture. +struct Pic { + w: usize, + h: usize, + bits: u8, + ssx: usize, + ssy: usize, + mono: bool, + planes: [Vec; 3], + cicp: Cicp, +} + +fn av1_error(e: av1::Error) -> CodecError { + match e { + av1::Error::Unsupported(w) => CodecError::unsupported(F, format!("AV1: {w}")), + other => bad(format!("AV1: {other}")), + } +} + +/// Decodes the `av01` item `id`, refusing pictures over `limits`. +fn decode_av1(c: &Container, id: u32, limits: &Limits) -> Result { + if c.item_type(id) != Some(*b"av01") { + return Err(CodecError::unsupported(F, "an image tile is not an AV1 image")); + } + // A tile's declared size counts before its picture is decoded too. + if let Some((w, h)) = c.ispe(id) { + limits.check(w, h, ChannelLayout::Rgb, SampleType::U16)?; + } + let data = c.data(id)?; + let config = c.props_of(id).find_map(|p| if let Prop::Av1c(o) = p { Some(o.clone()) } else { None }); + let mut dec = av1::Decoder::new(); + dec.set_memory_budget(limits.max_alloc.min(AV1_MEMORY_BUDGET)); + // Only the first picture is kept, so the pictures a chunk may show are held to the same budget + // before they are built (the decoder allows up to 1 GiB of them otherwise). + dec.set_shown_budget(limits.max_alloc); + let mut pics = catch_unwind(AssertUnwindSafe(|| -> std::result::Result, av1::Error> { + let mut pics = Vec::new(); + // The sequence header normally travels in the item too; the av1C copy covers files + // where it doesn't. + if let Some(cfg) = config.filter(|o| !o.is_empty()) { + pics.extend(dec.decode(&cfg)?); + } + pics.extend(dec.decode(&data)?); + pics.extend(dec.flush_result()?); + Ok(pics) + })) + .map_err(|_| bad("the AV1 decoder failed on this file"))? + .map_err(av1_error)?; + if pics.is_empty() { + return Err(bad("the AV1 data holds no picture")); + } + let p = pics.swap_remove(0); + limits.check(p.width, p.height, ChannelLayout::Rgb, SampleType::U16)?; + let (w, h) = (p.width as usize, p.height as usize); + let (ssx, ssy) = (usize::from(p.subsampling_x.min(1)), usize::from(p.subsampling_y.min(1))); + let need = w.checked_mul(h).ok_or_else(|| bad("picture too large"))?; + let cneed = ((w + ssx) >> ssx) * ((h + ssy) >> ssy); + if need == 0 || p.planes[0].len() < need || (!p.mono_chrome && (p.planes[1].len() < cneed || p.planes[2].len() < cneed)) { + return Err(bad("AV1 picture planes are short")); + } + let cicp = Cicp { + primaries: u16::from(p.color_primaries), + transfer: u16::from(p.transfer_characteristics), + matrix: u16::from(p.matrix_coefficients), + full_range: p.full_range, + }; + Ok(Pic { w, h, bits: p.bit_depth.clamp(8, 16), ssx, ssy, mono: p.mono_chrome, planes: p.planes, cicp }) +} + +/// Kr, Kb of an H.273 matrix code point; `None` = identity (GBR). +fn coefficients(matrix: u16, primaries: u16) -> Result> { + Ok(match matrix { + 0 => None, + 1 => Some((0.2126, 0.0722)), + 4 => Some((0.30, 0.11)), + 7 => Some((0.212, 0.087)), + // 9 is BT.2020 non-constant luminance; 10 (constant luminance) needs other equations and + // is refused below. + 9 => Some((0.2627, 0.0593)), + // Chromaticity-derived non-constant luminance. + 12 => { + let [kr, _, kb] = crate::cicp::luma_weights(primaries); + Some((kr, kb)) + } + // 2 (unspecified: AVIF readers assume BT.601), 5, 6 (BT.601) and reserved values. + 2 | 5 | 6 | 3 | 15.. => Some((0.299, 0.114)), + _ => return Err(CodecError::unsupported(F, format!("matrix coefficients {matrix} (YCgCo, constant luminance or ICtCp)"))), + }) +} + +/// A chroma sample at luma position (x, y), bilinear between chroma sites centred on their +/// luma pairs. +#[inline] +fn chroma_at(plane: &[u16], cw: usize, ch: usize, ssx: usize, ssy: usize, x: usize, y: usize) -> f32 { + let pos = |v: usize, ss: usize, n: usize| -> (usize, usize, f32) { + if ss == 0 { + return (v.min(n - 1), v.min(n - 1), 0.0); + } + let f = ((v as f32 + 0.5) / 2.0 - 0.5).max(0.0); + let i = f as usize; + (i.min(n - 1), (i + 1).min(n - 1), f - i as f32) + }; + let (x0, x1, fx) = pos(x, ssx, cw); + let (y0, y1, fy) = pos(y, ssy, ch); + let s = |xx: usize, yy: usize| f32::from(plane.get(yy * cw + xx).copied().unwrap_or(0)); + let top = s(x0, y0) * (1.0 - fx) + s(x1, y0) * fx; + let bot = s(x0, y1) * (1.0 - fx) + s(x1, y1) * fx; + top * (1.0 - fy) + bot * fy +} + +/// The sample scaling of a picture: luma offset and range, chroma range and midpoint. +struct Levels { + yoff: f32, + yrange: f32, + crange: f32, + mid: f32, +} + +impl Levels { + fn new(bits: u8, full_range: bool) -> Self { + let max = ((1u32 << bits) - 1) as f32; + let scale = f32::from(1u16 << (bits - 8).min(8)); + let (yoff, yrange, crange) = if full_range { (0.0, max, max) } else { (16.0 * scale, 219.0 * scale, 224.0 * scale) }; + Levels { yoff, yrange, crange, mid: f32::from(1u16 << (bits - 1).min(15)) } + } +} + +fn to_u16(v: f32) -> u16 { + (v.clamp(0.0, 1.0) * 65535.0 + 0.5) as u16 +} + +/// Where a decoded picture goes: an interleaved 16-bit canvas of `ch` channels. +struct Canvas { + w: usize, + h: usize, + ch: usize, + data: Vec, +} + +impl Canvas { + fn new(w: u32, h: u32, layout: ChannelLayout, limits: &Limits) -> Result { + limits.check(w, h, layout, SampleType::U16)?; + let (w, h, ch) = (w as usize, h as usize, layout.channels()); + let len = w.checked_mul(h).and_then(|n| n.checked_mul(ch)).ok_or_else(|| bad("image too large"))?; + Ok(Canvas { w, h, ch, data: vec![0; len] }) + } + + /// The picture's YCbCr as non-linear R'G'B' into channels 0..3, its top-left at (x0, y0). + fn put_colour(&mut self, p: &Pic, cicp: Cicp, x0: usize, y0: usize) -> Result<()> { + let lv = Levels::new(p.bits, cicp.full_range); + let k = coefficients(cicp.matrix, cicp.primaries)?; + let (cw, chh) = ((p.w + p.ssx) >> p.ssx, (p.h + p.ssy) >> p.ssy); + for y in 0..p.h.min(self.h.saturating_sub(y0)) { + for x in 0..p.w.min(self.w.saturating_sub(x0)) { + let luma = (f32::from(p.planes[0][y * p.w + x]) - lv.yoff) / lv.yrange; + let (cb, cr) = if p.mono { + (lv.mid, lv.mid) + } else { + (chroma_at(&p.planes[1], cw, chh, p.ssx, p.ssy, x, y), chroma_at(&p.planes[2], cw, chh, p.ssx, p.ssy, x, y)) + }; + let rgb = match k { + // Identity (GBR): Y = G, Cb = B, Cr = R, all three at the luma levels. + None => [(cr - lv.yoff) / lv.yrange, luma, (cb - lv.yoff) / lv.yrange], + Some((kr, kb)) => { + let (cb, cr) = ((cb - lv.mid) / lv.crange, (cr - lv.mid) / lv.crange); + let r = luma + 2.0 * (1.0 - kr) * cr; + let b = luma + 2.0 * (1.0 - kb) * cb; + [r, (luma - kr * r - kb * b) / (1.0 - kr - kb), b] + } + }; + let o = ((y0 + y) * self.w + x0 + x) * self.ch; + for (d, v) in self.data[o..o + 3].iter_mut().zip(rgb) { + *d = to_u16(v); + } + } + } + Ok(()) + } + + /// The picture's luma as alpha into channel 3, its top-left at (x0, y0). + fn put_alpha(&mut self, p: &Pic, x0: usize, y0: usize) { + let lv = Levels::new(p.bits, p.cicp.full_range); + for y in 0..p.h.min(self.h.saturating_sub(y0)) { + for x in 0..p.w.min(self.w.saturating_sub(x0)) { + let a = (f32::from(p.planes[0][y * p.w + x]) - lv.yoff) / lv.yrange; + self.data[((y0 + y) * self.w + x0 + x) * self.ch + 3] = to_u16(a); + } + } + } +} + +/// Decodes the image item `id` (one `av01` or a `grid` of them) into `canvas`, through `put`. +/// Returns the first picture's bit depth and colour description. +fn render( + c: &Container, + id: u32, + limits: &Limits, + canvas: &mut Canvas, + put: &mut dyn FnMut(&mut Canvas, &Pic, usize, usize) -> Result<()>, +) -> Result<(u8, Cicp)> { + if c.item_type(id) != Some(*b"grid") { + let p = decode_av1(c, id, limits)?; + if (p.w, p.h) != (canvas.w, canvas.h) { + return Err(bad("the AV1 picture's size differs from its item's")); + } + put(canvas, &p, 0, 0)?; + return Ok((p.bits, p.cicp)); + } + let (rows, cols, _) = grid_header(&c.data(id)?)?; + let tiles = c.refs_from(b"dimg", id); + if tiles.is_empty() || tiles.len() as u64 != u64::from(rows) * u64::from(cols) { + return Err(bad("grid tile count doesn't match its rows and columns")); + } + let mut first: Option<(usize, usize, u8, Cicp)> = None; + for (i, &t) in tiles.iter().enumerate() { + let p = decode_av1(c, t, limits)?; + let (tw, th, ..) = *first.get_or_insert((p.w, p.h, p.bits, p.cicp)); + if (p.w, p.h) != (tw, th) { + return Err(bad("grid tiles differ in size")); + } + if tw.saturating_mul(cols as usize) < canvas.w || th.saturating_mul(rows as usize) < canvas.h { + return Err(bad("grid tiles don't cover the grid")); + } + put(canvas, &p, (i % cols as usize) * tw, (i / cols as usize) * th)?; + } + first.map(|(_, _, bits, cicp)| (bits, cicp)).ok_or_else(|| bad("grid without tiles")) +} + +/// Applies the container's transforms in order to an interleaved canvas. +fn apply_transforms(mut cv: Canvas, ts: &[Transform]) -> Canvas { + for t in ts { + cv = match *t { + Transform::Clap { w, h, dx, dy } => clap(cv, w, h, dx, dy), + Transform::Rot(n) => (0..n).fold(cv, |cv, _| rotate(cv)), + Transform::Mir(axis) => mirror(cv, axis), + }; + } + cv +} + +fn clap(cv: Canvas, cw: f64, chh: f64, dx: f64, dy: f64) -> Canvas { + if !(cw >= 1.0 && chh >= 1.0 && cw.is_finite() && chh.is_finite() && dx.is_finite() && dy.is_finite()) { + return cv; + } + let (w, h, ch) = (cv.w, cv.h, cv.ch); + let nw = (cw.round() as usize).clamp(1, w); + let nh = (chh.round() as usize).clamp(1, h); + // The centre of the aperture is the image centre plus the offset. + let x0 = ((w as f64 - 1.0) / 2.0 + dx - (nw as f64 - 1.0) / 2.0).round().clamp(0.0, (w - nw) as f64) as usize; + let y0 = ((h as f64 - 1.0) / 2.0 + dy - (nh as f64 - 1.0) / 2.0).round().clamp(0.0, (h - nh) as f64) as usize; + let mut data = Vec::with_capacity(nw * nh * ch); + for y in y0..y0 + nh { + data.extend_from_slice(&cv.data[(y * w + x0) * ch..(y * w + x0 + nw) * ch]); + } + Canvas { w: nw, h: nh, ch, data } +} + +/// 90° anticlockwise: (x, y) → (y, w - 1 - x), the new size h × w. +fn rotate(cv: Canvas) -> Canvas { + let (w, h, ch) = (cv.w, cv.h, cv.ch); + let mut data = vec![0u16; cv.data.len()]; + for y in 0..h { + for x in 0..w { + let d = ((w - 1 - x) * h + y) * ch; + let s = (y * w + x) * ch; + data[d..d + ch].copy_from_slice(&cv.data[s..s + ch]); + } + } + Canvas { w: h, h: w, ch, data } +} + +fn mirror(mut cv: Canvas, axis: u8) -> Canvas { + let (w, h, ch) = (cv.w, cv.h, cv.ch); + if axis == 0 { + for y in 0..h / 2 { + let (top, bottom) = cv.data.split_at_mut((h - 1 - y) * w * ch); + top[y * w * ch..(y + 1) * w * ch].swap_with_slice(&mut bottom[..w * ch]); + } + } else { + for row in cv.data.chunks_exact_mut(w * ch) { + for x in 0..w / 2 { + for c in 0..ch { + row.swap(x * ch + c, (w - 1 - x) * ch + c); + } + } + } + } + cv +} + +/// Straight colour from colour premultiplied by alpha. +fn unpremultiply(cv: &mut Canvas) { + for px in cv.data.as_chunks_mut::<4>().0 { + let a = u32::from(px[3]); + if a == 0 { + continue; + } + for v in &mut px[..3] { + *v = (u32::from(*v) * 65535 / a).min(65535) as u16; + } + } +} + +/// The primary item's Exif (TIFF structure, its offset prefix dropped) and XMP. +fn read_metadata(c: &Container) -> (Option>, Option) { + let exif = c.metadata_item(|i| &i.kind == b"Exif").and_then(|id| c.data(id).ok()).and_then(|d| { + let offset = usize::try_from(u32::from_be_bytes(d.get(..4)?.try_into().ok()?)).ok()?; + let tiff = d.get(4usize.checked_add(offset)?..)?; + (tiff.starts_with(b"II*\0") || tiff.starts_with(b"MM\0*")).then(|| tiff.to_vec()) + }); + let xmp = c + .metadata_item(|i| &i.kind == b"mime" && i.content_type.as_deref() == Some(b"application/rdf+xml")) + .and_then(|id| c.data(id).ok()) + .map(|d| String::from_utf8_lossy(&d).trim_end_matches('\0').to_owned()); + (exif, xmp) +} + +/// Decodes the primary image. `keep_orientation` leaves the container's `clap`, `irot` and +/// `imir` unapplied (the stored pixels), as for HEIF. +pub(crate) fn decode(bytes: &[u8], limits: &Limits, keep_orientation: bool) -> Result { + let c = parse(bytes)?; + let id = c.primary; + let (w, h) = item_size(&c, id)?; + let alpha = c.alpha_item().filter(|&a| item_size(&c, a).is_ok_and(|s| s == (w, h))); + let transforms = c.transforms(id)?; + // The alpha plane is composed in stored orientation and turned with the colour, which is right + // when it has no transforms of its own (what encoders write) or the colour's. Others would put + // transparency in the wrong place, so they are refused rather than ignored. + if let Some(a) = alpha { + let own = c.transforms(a)?; + if !own.is_empty() && own != transforms { + return Err(CodecError::unsupported(F, "an alpha plane with transforms (clap, irot, imir) that differ from the image's")); + } + } + let layout = if alpha.is_some() { ChannelLayout::Rgba } else { ChannelLayout::Rgb }; + // The declared size, before any picture is decoded. + let mut canvas = Canvas::new(w, h, layout, limits)?; + let file_cicp = c.nclx(id); + let mut colour = |cv: &mut Canvas, p: &Pic, x: usize, y: usize| cv.put_colour(p, file_cicp.unwrap_or(p.cicp), x, y); + let (bits, stream_cicp) = render(&c, id, limits, &mut canvas, &mut colour)?; + if let Some(a) = alpha { + render(&c, a, limits, &mut canvas, &mut |cv, p, x, y| { + cv.put_alpha(p, x, y); + Ok(()) + })?; + if c.premultiplied(a) { + unpremultiply(&mut canvas); + } + } + if !keep_orientation { + canvas = apply_transforms(canvas, &transforms); + } + let (w, h) = (canvas.w as u32, canvas.h as u32); + let image = Image::from_u16(w, h, layout, &canvas.data)?; + drop(canvas); + let mut img = if bits > 8 { image } else { image.convert(layout, SampleType::U8) }; + let cicp = file_cicp.unwrap_or(stream_cicp); + img.icc = c.icc(id); + img.cicp = Some(cicp); + let (exif, xmp) = read_metadata(&c); + img.meta = Metadata { + exif: exif.map(|e| if keep_orientation { e } else { crate::orientation::upright_exif(&e).into_owned() }), + xmp: xmp.map(|x| if keep_orientation { x } else { crate::orientation::upright_xmp(&x).into_owned() }), + ..Default::default() + }; + // PQ or HLG: linear light in 32-bit float, as for HEIF. + if cicp.is_hdr() { + limits.check(w, h, layout, SampleType::F32)?; + return crate::cicp::linearize(img, cicp); + } + Ok(img) +} + +#[cfg(test)] +mod tests { + use super::*; + use crate::DecodeOptions; + + const QUAD_420: &[u8] = include_bytes!("../../tests/fixtures/avif/quad-420-bt709.avif"); + const QUAD_420_DAV1D: &[u8] = include_bytes!("../../tests/fixtures/avif/quad-420-bt709.dav1d.yuv"); + const QUAD_GBR_LOSSLESS: &[u8] = include_bytes!("../../tests/fixtures/avif/quad-444-gbr-lossless.avif"); + const QUAD_SOURCE: &[u8] = include_bytes!("../../tests/fixtures/avif/quad-source.rgb"); + + fn rgb8(img: &Image) -> Vec { + img.to_rgba8().as_chunks::<4>().0.iter().flat_map(|p| [p[0], p[1], p[2]]).collect() + } + + fn mean_abs_diff(a: &[u8], b: &[u8]) -> f64 { + assert_eq!(a.len(), b.len()); + a.iter().zip(b).map(|(&x, &y)| (f64::from(x) - f64::from(y)).abs()).sum::() / a.len() as f64 + } + + /// A box of type `t` around `body`. + fn bx(t: &[u8; 4], body: &[u8]) -> Vec { + let mut v = ((body.len() + 8) as u32).to_be_bytes().to_vec(); + v.extend_from_slice(t); + v.extend_from_slice(body); + v + } + + /// A full box: version 0, no flags. + fn fbx(t: &[u8; 4], body: &[u8]) -> Vec { + let mut b = vec![0, 0, 0, 0]; + b.extend_from_slice(body); + bx(t, &b) + } + + /// An `infe` (version 2) for item `id` of `kind`. + fn infe(id: u16, kind: &[u8; 4]) -> Vec { + let mut b = vec![2, 0, 0, 0]; + b.extend_from_slice(&id.to_be_bytes()); + b.extend_from_slice(&[0, 0]); + b.extend_from_slice(kind); + b.push(0); + bx(b"infe", &b) + } + + fn ispe(w: u32, h: u32) -> Vec { + let mut b = w.to_be_bytes().to_vec(); + b.extend_from_slice(&h.to_be_bytes()); + fbx(b"ispe", &b) + } + + /// The AV1 data and `av1C` of an encoded file's primary item. + fn av1_of(file: &[u8]) -> (Vec, Vec) { + let c = parse(file).unwrap(); + let cfg = c.props_of(c.primary).find_map(|p| if let Prop::Av1c(o) = p { Some(o.clone()) } else { None }).unwrap(); + (c.data(c.primary).unwrap(), cfg) + } + + /// An AVIF of `items` (id, type, data, property indices) with `props`, `refs` and primary 1. + /// All item data sits in `idat` (construction method 1). + /// An item: ID, type, data and 1-based property indices. + type TestItem<'a> = (u16, &'a [u8; 4], &'a [u8], &'a [u8]); + /// An item reference: type, from, to. + type TestRef<'a> = (&'a [u8; 4], u16, &'a [u16]); + + fn build(items: &[TestItem], props: &[Vec], refs: &[TestRef]) -> Vec { + let mut iinf = vec![0, 0, 0, 0]; + iinf.extend_from_slice(&(items.len() as u16).to_be_bytes()); + let mut iloc = vec![1, 0, 0, 0, 0x44, 0x00]; + iloc.extend_from_slice(&(items.len() as u16).to_be_bytes()); + let mut ipma = vec![0, 0, 0, 0]; + ipma.extend_from_slice(&(items.len() as u32).to_be_bytes()); + let mut idat = Vec::new(); + for &(id, kind, data, assoc) in items { + iinf.extend(infe(id, kind)); + iloc.extend_from_slice(&id.to_be_bytes()); + iloc.extend_from_slice(&[0, 1, 0, 0, 0, 1]); // method 1, data ref 0, one extent + iloc.extend_from_slice(&(idat.len() as u32).to_be_bytes()); + iloc.extend_from_slice(&(data.len() as u32).to_be_bytes()); + idat.extend_from_slice(data); + ipma.extend_from_slice(&id.to_be_bytes()); + ipma.push(assoc.len() as u8); + ipma.extend(assoc.iter().map(|&p| p | 0x80)); + } + let mut iref = vec![0, 0, 0, 0]; + for &(kind, from, to) in refs { + let mut b = from.to_be_bytes().to_vec(); + b.extend_from_slice(&(to.len() as u16).to_be_bytes()); + to.iter().for_each(|t| b.extend_from_slice(&t.to_be_bytes())); + iref.extend(bx(kind, &b)); + } + let mut meta = vec![0, 0, 0, 0]; + meta.extend(fbx(b"hdlr", b"\0\0\0\0pict\0\0\0\0\0\0\0\0\0\0\0\0\0")); + meta.extend(fbx(b"pitm", &[0, 1])); + meta.extend(bx(b"iinf", &iinf)); + meta.extend(bx(b"iloc", &iloc)); + meta.extend(bx(b"iref", &iref)); + meta.extend(bx(b"iprp", &[bx(b"ipco", &props.concat()), bx(b"ipma", &ipma)].concat())); + meta.extend(bx(b"idat", &idat)); + let mut file = bx(b"ftyp", b"avif\0\0\0\0avifmif1miaf"); + file.extend(bx(b"meta", &meta)); + file + } + + #[test] + fn libaom_420_limited_range_matches_dav1d() { + let img = crate::decode(QUAD_420).unwrap(); + assert_eq!((img.width(), img.height(), img.layout(), img.sample_type()), (33, 17, ChannelLayout::Rgb, SampleType::U8)); + assert_eq!(img.cicp.map(|c| (c.matrix, c.full_range)), Some((1, false)), "BT.709 matrix, limited range"); + assert_eq!(crate::probe_size(QUAD_420).unwrap(), (33, 17)); + // The AV1 picture is libdav1d's, sample for sample. + let c = parse(QUAD_420).unwrap(); + let p = decode_av1(&c, c.primary, &Limits::default()).unwrap(); + assert_eq!((p.w, p.h, p.bits, p.ssx, p.ssy, p.mono), (33, 17, 8, 1, 1, false)); + let planes: Vec = p.planes.iter().flatten().map(|&v| v as u8).collect(); + assert_eq!(planes, QUAD_420_DAV1D); + let ours = rgb8(&img); + // The quadrant centres are the source's colours, give or take the lossy coding. + for (x, y) in [(5usize, 4usize), (27, 4), (5, 13), (27, 13)] { + let i = (y * 33 + x) * 3; + let d = mean_abs_diff(&ours[i..i + 3], &QUAD_SOURCE[i..i + 3]); + assert!(d < 12.0, "({x}, {y}): {:?} vs {:?}", &ours[i..i + 3], &QUAD_SOURCE[i..i + 3]); + } + } + + #[test] + fn libaom_lossless_identity_matrix_is_exact() { + let img = crate::decode(QUAD_GBR_LOSSLESS).unwrap(); + assert_eq!((img.width(), img.height(), img.sample_type()), (33, 17, SampleType::U8)); + assert_eq!(img.cicp.map(|c| (c.matrix, c.full_range)), Some((0, true))); + assert_eq!(rgb8(&img), QUAD_SOURCE); + } + + #[test] + fn limits_are_checked_before_decoding() { + let tight = |max_pixels, max_alloc| DecodeOptions { limits: Limits { max_pixels, max_alloc, ..Limits::default() }, ..Default::default() }; + assert!(matches!(crate::decode_with(QUAD_420, &tight(33 * 17 - 1, u64::MAX)), Err(CodecError::LimitExceeded(_)))); + // The 16-bit intermediate counts: 33 × 17 RGB at 2 bytes. + assert!(matches!(crate::decode_with(QUAD_420, &tight(u64::MAX, 33 * 17 * 6 - 1)), Err(CodecError::LimitExceeded(_)))); + assert!(crate::decode_with(QUAD_420, &tight(33 * 17, Limits::default().max_alloc)).is_ok()); + // `max_alloc` is the AV1 decoder's picture-store budget too. + assert!(crate::decode_with(QUAD_420, &tight(33 * 17, 33 * 17 * 6)).is_err()); + // A declared size over the limits fails from the boxes alone. + let (av1, cfg) = av1_of(QUAD_420); + let huge = build(&[(1, b"av01", &av1, &[1, 2])], &[ispe(1 << 20, 1 << 20), bx(b"av1C", &[&[0x81, 0, 0, 0][..], &cfg].concat())], &[]); + assert!(matches!(crate::decode(&huge), Err(CodecError::LimitExceeded(_)))); + assert!(matches!(crate::probe_size(&huge), Err(CodecError::LimitExceeded(_)))); + } + + #[test] + fn a_picture_that_differs_from_its_ispe_is_refused() { + let (av1, cfg) = av1_of(QUAD_420); + let file = build(&[(1, b"av01", &av1, &[1, 2])], &[ispe(32, 17), bx(b"av1C", &[&[0x81, 0, 0, 0][..], &cfg].concat())], &[]); + assert!(matches!(crate::decode(&file), Err(CodecError::Malformed { .. }))); + } + + #[test] + fn a_grid_is_assembled_from_its_tiles_and_cropped() { + // Two copies of the 33 × 17 picture side by side, cropped to 60 × 15 by the grid. + let (av1, cfg) = av1_of(QUAD_GBR_LOSSLESS); + let av1c = bx(b"av1C", &[&[0x81, 0, 0, 0][..], &cfg].concat()); + let grid = [0u8, 0, 0, 1, 0, 60, 0, 15]; + let file = build( + &[(1, b"grid", &grid, &[1]), (2, b"av01", &av1, &[2, 3]), (3, b"av01", &av1, &[2, 3])], + &[ispe(60, 15), ispe(33, 17), av1c, fbx(b"pixi", &[3, 8, 8, 8])], + &[(b"dimg", 1, &[2, 3])], + ); + assert_eq!(crate::probe_size(&file).unwrap(), (60, 15)); + let img = crate::decode(&file).unwrap(); + assert_eq!((img.width(), img.height()), (60, 15)); + let got = rgb8(&img); + for y in 0..15 { + for x in 0..60 { + let s = (y * 33 + x % 33) * 3; + assert_eq!(&got[(y * 60 + x) * 3..][..3], &QUAD_SOURCE[s..s + 3], "({x}, {y})"); + } + } + // Tiles that can't cover the grid are refused. + let wide = [0u8, 0, 0, 1, 0, 70, 0, 15]; + let file = build( + &[(1, b"grid", &wide, &[1]), (2, b"av01", &av1, &[2, 3]), (3, b"av01", &av1, &[2, 3])], + &[ispe(70, 15), ispe(33, 17), bx(b"av1C", &[&[0x81, 0, 0, 0][..], &cfg].concat())], + &[(b"dimg", 1, &[2, 3])], + ); + assert!(matches!(crate::decode(&file), Err(CodecError::Malformed { .. }))); + } + + #[test] + fn container_rotation_and_mirror_are_applied_unless_kept() { + let (av1, cfg) = av1_of(QUAD_GBR_LOSSLESS); + let av1c = bx(b"av1C", &[&[0x81, 0, 0, 0][..], &cfg].concat()); + let file = build(&[(1, b"av01", &av1, &[1, 2, 3, 4])], &[ispe(33, 17), av1c, bx(b"irot", &[1]), bx(b"imir", &[0])], &[]); + assert_eq!(crate::probe_size(&file).unwrap(), (17, 33)); + let img = crate::decode(&file).unwrap(); + assert_eq!((img.width(), img.height()), (17, 33)); + let got = rgb8(&img); + // Rotated 90° anticlockwise, then top and bottom exchanged: a transpose, (x, y) from (y, x). + for y in 0..33 { + for x in 0..17 { + let s = (x * 33 + y) * 3; + assert_eq!(&got[(y * 17 + x) * 3..][..3], &QUAD_SOURCE[s..s + 3], "({x}, {y})"); + } + } + let kept = crate::decode_with(&file, &DecodeOptions { keep_orientation: true, ..Default::default() }).unwrap(); + assert_eq!((kept.width(), kept.height(), rgb8(&kept)), (33, 17, QUAD_SOURCE.to_vec())); + } + + /// A `clap` box: width, height, horizontal and vertical offset, each numerator then denominator. + fn clap_box(f: [u32; 8]) -> Vec { + bx(b"clap", &f.iter().flat_map(|v| v.to_be_bytes()).collect::>()) + } + + #[test] + fn clean_aperture_sizes_are_unsigned() { + let (av1, cfg) = av1_of(QUAD_GBR_LOSSLESS); + let av1c = bx(b"av1C", &[&[0x81, 0, 0, 0][..], &cfg].concat()); + // A 2 × 2 aperture with its width written as 0x80000000 / 0x40000000, centred one pixel + // left of the image centre (a signed offset numerator: -1 / 1). + let clap = clap_box([0x8000_0000, 0x4000_0000, 2, 1, (-1i32) as u32, 1, 0, 1]); + let file = build(&[(1, b"av01", &av1, &[1, 2, 3])], &[ispe(33, 17), av1c, clap], &[]); + assert_eq!(crate::probe_size(&file).unwrap(), (2, 2)); + let img = crate::decode(&file).unwrap(); + assert_eq!((img.width(), img.height()), (2, 2)); + // Centre (16, 8) moved to (15, 8): the 2 × 2 crop starts at (15, 8) (rounded from 14.5, 7.5). + let got = rgb8(&img); + let (x0, y0) = ((32.0f64 / 2.0 - 1.0 - 0.5).round() as usize, (16.0f64 / 2.0 - 0.5).round() as usize); + for y in 0..2 { + for x in 0..2 { + let s = ((y0 + y) * 33 + x0 + x) * 3; + assert_eq!(&got[(y * 2 + x) * 3..][..3], &QUAD_SOURCE[s..s + 3], "({x}, {y})"); + } + } + } + + #[test] + fn repeated_transforms_are_refused() { + let (av1, cfg) = av1_of(QUAD_GBR_LOSSLESS); + let av1c = bx(b"av1C", &[&[0x81, 0, 0, 0][..], &cfg].concat()); + // the same irot associated twice, and through two ipma entries for the item + let twice = build(&[(1, b"av01", &av1, &[1, 2, 3, 3])], &[ispe(33, 17), av1c.clone(), bx(b"irot", &[1])], &[]); + assert!(matches!(crate::probe_size(&twice), Err(CodecError::Malformed { .. }))); + assert!(matches!(crate::decode(&twice), Err(CodecError::Malformed { .. }))); + let mut many = build(&[(1, b"av01", &av1, &[1, 2, 3])], &[ispe(33, 17), av1c.clone(), bx(b"irot", &[1])], &[]); + let c = parse(&many).unwrap(); + assert_eq!(c.transforms(1).unwrap(), vec![Transform::Rot(1)]); + // 1024 more ipma records for item 1, each repeating the irot 255 times + let mut ipma = vec![0, 0, 0, 0]; + ipma.extend_from_slice(&1024u32.to_be_bytes()); + for _ in 0..1024 { + ipma.extend_from_slice(&[0, 1, 255]); + ipma.extend(std::iter::repeat_n(3u8, 255)); + } + many = build(&[(1, b"av01", &av1, &[1, 2])], &[ispe(33, 17), av1c, bx(b"irot", &[1])], &[]); + let mut c = parse(&many).unwrap(); + parse_ipma(&ipma, &mut c).unwrap(); + assert!(matches!(c.transforms(1), Err(CodecError::Malformed { .. }))); + // each kind once is fine, in any order + let ok = build( + &[(1, b"av01", &av1, &[1, 2, 3, 4, 5])], + &[ + ispe(33, 17), + bx(b"av1C", &[&[0x81, 0, 0, 0][..], &cfg].concat()), + clap_box([33, 1, 17, 1, 0, 1, 0, 1]), + bx(b"irot", &[1]), + bx(b"imir", &[0]), + ], + &[], + ); + assert_eq!(crate::probe_size(&ok).unwrap(), (17, 33)); + } + + #[test] + fn alpha_with_transforms_of_its_own_is_refused() { + let (av1, cfg) = av1_of(QUAD_GBR_LOSSLESS); + let av1c = bx(b"av1C", &[&[0x81, 0, 0, 0][..], &cfg].concat()); + let auxc = fbx(b"auxC", b"urn:mpeg:mpegB:cicp:systems:auxiliary:alpha\0"); + let props = [ispe(33, 17), av1c, auxc, bx(b"imir", &[1])]; + let with = |colour: &[u8], alpha: &[u8]| build(&[(1, b"av01", &av1, colour), (2, b"av01", &av1, alpha)], &props, &[(b"auxl", 2, &[1])]); + // no transforms on the alpha plane, or the colour's: composed, then turned together + let plain = crate::decode(&with(&[1, 2], &[1, 2, 3])).unwrap(); + assert_eq!(plain.layout(), ChannelLayout::Rgba); + let both = crate::decode(&with(&[1, 2, 4], &[1, 2, 3, 4])).unwrap(); + let alpha_free = crate::decode(&with(&[1, 2, 4], &[1, 2, 3])).unwrap(); + assert_eq!(both.to_rgba8(), alpha_free.to_rgba8()); + // an alpha-only mirror would put transparency on the wrong side + assert!(matches!(crate::decode(&with(&[1, 2], &[1, 2, 3, 4])), Err(CodecError::Unsupported { .. }))); + } + + #[test] + fn transforms_rotate_mirror_and_crop() { + // 3 × 2, one channel, value = index + let cv = || Canvas { w: 3, h: 2, ch: 1, data: (0..6).collect() }; + let r = apply_transforms(cv(), &[Transform::Rot(1)]); + // anticlockwise: the right column becomes the top row + assert_eq!((r.w, r.h, r.data), (2, 3, vec![2, 5, 1, 4, 0, 3])); + // axis 0 exchanges top and bottom, axis 1 left and right (libheif's and libavif's reading) + assert_eq!(apply_transforms(cv(), &[Transform::Mir(0)]).data, vec![3, 4, 5, 0, 1, 2]); + assert_eq!(apply_transforms(cv(), &[Transform::Mir(1)]).data, vec![2, 1, 0, 5, 4, 3]); + let c = apply_transforms(cv(), &[Transform::Clap { w: 1.0, h: 1.0, dx: 1.0, dy: 0.0 }]); + assert_eq!((c.w, c.h), (1, 1)); + assert!(c.data == [2] || c.data == [5]); + assert_eq!(displayed(3, 2, &[Transform::Rot(3)]), (2, 3)); + // A non-finite or empty aperture is ignored. + assert_eq!(apply_transforms(cv(), &[Transform::Clap { w: f64::NAN, h: 1.0, dx: 0.0, dy: 0.0 }]).data.len(), 6); + } + + #[test] + fn hostile_files_are_errors_not_panics() { + for file in [QUAD_420, QUAD_GBR_LOSSLESS] { + // every truncation + for n in 0..file.len() { + let _ = crate::probe_size(&file[..n]); + let _ = crate::decode(&file[..n]); + } + // byte flips through the container and the AV1 data + for i in 0..file.len() { + for x in [0x01u8, 0x80, 0xff] { + let mut m = file.to_vec(); + m[i] ^= x; + let _ = crate::probe_size(&m); + let _ = crate::decode(&m); + } + } + } + // box sizes that point far outside the file + let mut b = bx(b"ftyp", b"avif\0\0\0\0avif"); + b.extend_from_slice(b"\xff\xff\xff\xffmeta\0\0\0\0"); + assert!(crate::decode(&b).is_err()); + let mut b = bx(b"ftyp", b"avif\0\0\0\0avif"); + b.extend_from_slice(b"\0\0\0\x01meta\xff\xff\xff\xff\xff\xff\xff\xff"); + assert!(crate::probe_size(&b).is_err()); + } + + #[test] + fn an_animated_avif_is_unsupported() { + let mut b = bx(b"ftyp", b"avis\0\0\0\0avis"); + b.extend(bx(b"moov", &[])); + assert!(matches!(crate::decode(&b), Err(CodecError::Unsupported { format: Format::Avif, .. }))); + } + + #[test] + fn zero_width_iloc_extents_are_capped() { + // iloc v2 with zero-byte offset/length/base/index fields: each extent costs no input bytes + let mut iloc = vec![2, 0, 0, 0, 0, 0]; + iloc.extend_from_slice(&3u32.to_be_bytes()); + for id in 1..=3u32 { + iloc.extend_from_slice(&id.to_be_bytes()); + iloc.extend_from_slice(&[0, 0, 0, 0, 0xff, 0xff]); // method, data ref, 65535 extents + } + let mut meta = vec![0, 0, 0, 0]; + meta.extend(bx(b"iloc", &iloc)); + let mut b = bx(b"ftyp", b"avif\0\0\0\0avif"); + b.extend(bx(b"meta", &meta)); + assert!(matches!(crate::decode(&b), Err(CodecError::Malformed { message, .. }) if message.contains("extents"))); + } + + #[test] + fn item_data_larger_than_the_file_is_refused() { + // one item whose 1000 extents each repeat the whole file (length 0 = to the end) + let mut iloc = vec![1, 0, 0, 0, 0x44, 0]; + iloc.extend_from_slice(&1u16.to_be_bytes()); + iloc.extend_from_slice(&[0, 1, 0, 0, 0, 0]); // id 1, method 0, data ref 0 + iloc.extend_from_slice(&1000u16.to_be_bytes()); + for _ in 0..1000 { + iloc.extend_from_slice(&[0; 8]); + } + let mut meta = vec![0, 0, 0, 0]; + meta.extend(fbx(b"pitm", &[0, 1])); + meta.extend(bx(b"iloc", &iloc)); + let mut b = bx(b"ftyp", b"avif\0\0\0\0avif"); + b.extend(bx(b"meta", &meta)); + let c = parse(&b).unwrap(); + assert!(matches!(c.data(1), Err(CodecError::Malformed { message, .. }) if message.contains("larger than the file"))); + } + + #[test] + fn unsupported_matrices_are_errors() { + for m in [8u16, 10, 11, 13, 14] { + assert!(matches!(coefficients(m, 1), Err(CodecError::Unsupported { .. })), "{m}"); + } + assert_eq!(coefficients(0, 1).unwrap(), None); + assert!(coefficients(12, 9).unwrap().is_some()); + } + + #[cfg(feature = "avif")] + mod with_the_encoder { + use super::*; + use crate::codecs::avif::{HdrTransfer, encode_avif_hdr, encode_avif_with_speed}; + + /// A smooth test card with a saturated band. + fn card(w: usize, h: usize, ch: usize) -> Vec { + (0..w * h) + .flat_map(|i| { + let (x, y) = (i % w, i / w); + let px = + [(x * 255 / (w - 1)) as u8, (y * 255 / (h - 1)) as u8, if x < w / 4 { 230 } else { 40 }, (255 - x * 200 / (w - 1)) as u8]; + px.into_iter().take(ch) + }) + .collect() + } + + #[test] + fn ravif_rgb_round_trip_is_10_bit() { + let (w, h) = (48, 32); + let src = Image::from_u8(w as u32, h as u32, ChannelLayout::Rgb, card(w, h, 3)).unwrap(); + let file = encode_avif_with_speed(&src, 95, 10).unwrap(); + let img = crate::decode(&file).unwrap(); + assert_eq!((img.width(), img.height(), img.layout(), img.sample_type()), (48, 32, ChannelLayout::Rgb, SampleType::U16)); + let diff = mean_abs_diff(&rgb8(&img), &card(w, h, 3)); + assert!(diff < 2.5, "mean difference {diff}"); + } + + #[test] + fn ravif_alpha_round_trips() { + let (w, h) = (40, 24); + let src = Image::from_u8(w as u32, h as u32, ChannelLayout::Rgba, card(w, h, 4)).unwrap(); + let file = encode_avif_with_speed(&src, 95, 10).unwrap(); + let img = crate::decode(&file).unwrap(); + assert_eq!(img.layout(), ChannelLayout::Rgba); + let got = img.to_rgba8(); + let want = card(w, h, 4); + let alpha = |v: &[u8]| v.iter().skip(3).step_by(4).copied().collect::>(); + let diff = mean_abs_diff(&alpha(&got), &alpha(&want)); + assert!(diff < 2.0, "alpha mean difference {diff}"); + } + + #[test] + fn exif_is_read_from_its_item() { + let mut src = Image::from_u8(16, 8, ChannelLayout::Rgb, card(16, 8, 3)).unwrap(); + let exif = b"II*\0\x08\0\0\0\0\0\0\0\0\0".to_vec(); + src.meta.exif = Some(exif.clone()); + let file = encode_avif_with_speed(&src, 80, 10).unwrap(); + let img = crate::decode_with(&file, &DecodeOptions { keep_orientation: true, ..Default::default() }).unwrap(); + assert_eq!(img.meta.exif.as_deref(), Some(exif.as_slice())); + } + + #[test] + fn hdr_round_trips_keep_highlights_above_white() { + let (w, h) = (32usize, 16usize); + // linear BT.2020: left half grey 0.18, right half 4 × SDR white + let rgb: Vec = (0..w * h).flat_map(|i| if i % w < w / 2 { [0.18f32; 3] } else { [4.0; 3] }).collect(); + for t in [HdrTransfer::Pq, HdrTransfer::Hlg] { + let file = encode_avif_hdr(w as u32, h as u32, &rgb, 98, 10, None, t).unwrap(); + let img = crate::decode(&file).unwrap(); + assert_eq!(img.sample_type(), SampleType::F32, "{t:?}"); + assert_eq!(img.cicp.map(|c| c.transfer), Some(Cicp::TRANSFER_LINEAR), "{t:?}"); + let v = img.to_f32_samples().unwrap(); + let at = |x: usize| v[(8 * w + x) * 3 + 1]; + assert!((at(4) / 0.18 - 1.0).abs() < 0.08, "{t:?}: grey {}", at(4)); + assert!((at(w - 4) / 4.0 - 1.0).abs() < 0.08, "{t:?}: highlight {}", at(w - 4)); + } + } + } +} diff --git a/crates/image-codecs/src/codecs/jpeg.rs b/crates/image-codecs/src/codecs/jpeg.rs index bcd86ad5a..15346ea56 100644 --- a/crates/image-codecs/src/codecs/jpeg.rs +++ b/crates/image-codecs/src/codecs/jpeg.rs @@ -195,6 +195,27 @@ enum DataEnd { /// measured on these, not on the file length, which APP padding can inflate without /// adding image data. fn data_end(b: &[u8]) -> (DataEnd, u64) { + data_end_mode(b, false) +} + +/// Whether the marker walk actually reaches the first EOI, with malformed segment lengths refused. +pub(crate) fn reaches_eoi(b: &[u8]) -> bool { + data_end_mode(b, true).0 == DataEnd::Complete +} + +fn find_ff(b: &[u8]) -> Option { + const ONES: u64 = 0x0101_0101_0101_0101; + let (words, tail) = b.as_chunks::<8>(); + for (i, word) in words.iter().enumerate() { + let x = !u64::from_ne_bytes(*word); + if x.wrapping_sub(ONES) & !x & (ONES << 7) != 0 { + return word.iter().position(|&v| v == 0xff).map(|j| i * 8 + j); + } + } + tail.iter().position(|&v| v == 0xff).map(|j| words.len() * 8 + j) +} + +fn data_end_mode(b: &[u8], strict: bool) -> (DataEnd, u64) { // Where the bytes run out: inside the image data once a scan has held some. let cut = |scan_data: bool, scanned: u64| (if scan_data { DataEnd::Truncated } else { DataEnd::Empty }, scanned); let mut i = 2; // past SOI @@ -207,7 +228,7 @@ fn data_end(b: &[u8]) -> (DataEnd, u64) { // segments after the stray bytes (padding among them) are still recognised. Stray // bytes are not scan data, so they never count (junk would buy budget as padding // did); with no marker left, the image data ends here. - let Some(p) = b.get(i..).and_then(|r| r.iter().position(|&v| v == 0xFF)) else { return cut(scan_data, scanned) }; + let Some(p) = b.get(i..).and_then(find_ff) else { return cut(scan_data, scanned) }; i += p; continue; } @@ -216,15 +237,19 @@ fn data_end(b: &[u8]) -> (DataEnd, u64) { 0xD9 => return (DataEnd::Complete, scanned), 0xFF => i += 1, 0x01 | 0xD0..=0xD8 => i += 2, + 0x00 if strict => i += 2, _ => { let Some(len) = b.get(i + 2..i + 4).and_then(|s| <[u8; 2]>::try_from(s).ok()).map(u16::from_be_bytes) else { return cut(scan_data, scanned); }; + if strict && len < 2 { + return cut(scan_data, scanned); + } i = i.saturating_add(2 + usize::from(len)); if code == 0xDA { let start = i; loop { - let Some(p) = b.get(i..).and_then(|r| r.iter().position(|&v| v == 0xFF)) else { + let Some(p) = b.get(i..).and_then(find_ff) else { // EOF inside this scan: everything from the scan start to the end // of the file is scan data (inside a declared segment it may not // be, so only count what remains). @@ -653,6 +678,20 @@ mod tests { } /// `info` reads the frame from the markers, and which decoder the file goes to. + #[test] + fn ff_search_matches_plain_search() { + let mut b: Vec = (0..300u32).map(|i| (i * 7 % 255) as u8).collect(); + for at in [None, Some(0), Some(7), Some(8), Some(15), Some(203), Some(296), Some(299)] { + if let Some(i) = at { + b[i] = 0xff; + } + for start in [0, 1, 5, 9, 290, 299, 300] { + let s = &b[start..]; + assert_eq!(find_ff(s), s.iter().position(|&v| v == 0xff)); + } + } + } + #[test] fn info_reads_the_frame_and_the_decoder() { let img = Image::from_raw(37, 21, ChannelLayout::Rgb, SampleType::U8, vec![128; 37 * 21 * 3]).unwrap(); diff --git a/crates/image-codecs/src/codecs/jpegls.rs b/crates/image-codecs/src/codecs/jpegls.rs index b405e71ed..a5dff1015 100644 --- a/crates/image-codecs/src/codecs/jpegls.rs +++ b/crates/image-codecs/src/codecs/jpegls.rs @@ -654,6 +654,8 @@ pub(crate) fn decode(b: &[u8], limits: &Limits) -> Result { 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)?; + // Each scan retains u16 samples even when the final image is 8-bit. + limits.check_bytes(width as u32, height as u32, 2)?; let components = (0..nf).map(|i| seg[6 + 3 * i]).collect(); frame = Some(Frame { precision, width, height, components }); } @@ -731,7 +733,7 @@ fn finish(frame: Option, planes: Vec>, scans: usize) -> Result u8 { if p == 8 { v as u8 } else { (scale(v) >> 8) as u8 } }; + let to8 = |v: u16| -> u8 { if p == 8 { v as u8 } else { (u32::from(v) * 255 / ((1u32 << p) - 1)) as u8 } }; let mut data = Vec::with_capacity(w * h * nc); for i in 0..w * h { for plane in &planes { @@ -784,6 +786,25 @@ mod tests { Plan { layout: img.layout(), sample: img.sample_type() } } + #[test] + fn low_precision_samples_span_the_full_output_range() { + for precision in 2..=7 { + let max = (1u16 << precision) - 1; + let f = Frame { precision, width: usize::from(max) + 1, height: 1, components: vec![1] }; + let image = finish(Some(f), vec![(0..=max).collect()], 1).unwrap(); + let expected: Vec = (0..=max).map(|v| (u32::from(v) * 255 / u32::from(max)) as u8).collect(); + assert_eq!(image.data(), expected); + } + } + + #[test] + fn eight_bit_scan_intermediate_respects_allocation_limit() { + let img = image(8, 8, ChannelLayout::Gray, false, 0); + let bytes = encode(&img, plan(&img)).unwrap(); + assert!(matches!(decode(&bytes, &Limits::new(8, 8, 64, 127)), Err(CodecError::LimitExceeded(_)))); + assert!(decode(&bytes, &Limits::new(8, 8, 64, 128)).is_ok()); + } + #[test] fn round_trips_noise_flat_and_smooth_images() { for layout in [ChannelLayout::Gray, ChannelLayout::Rgb] { diff --git a/crates/image-codecs/src/format.rs b/crates/image-codecs/src/format.rs index 810c1ad48..26696ae2a 100644 --- a/crates/image-codecs/src/format.rs +++ b/crates/image-codecs/src/format.rs @@ -26,7 +26,8 @@ pub enum Format { OpenExr, /// Radiance RGBE. Hdr, - /// Encode-only behind the `avif` feature; see [`ASYMMETRIC_EXCEPTIONS`]. + /// Encoded behind the `avif` feature and decoded behind `avif-decode` (native targets); see + /// [`ASYMMETRIC_EXCEPTIONS`] for a build with only one of them. Avif, /// HEIF/HEIC (HEVC-coded): read-only; see [`ASYMMETRIC_EXCEPTIONS`]. Heif, @@ -91,8 +92,10 @@ pub const ASYMMETRIC_EXCEPTIONS: &[(Format, &str)] = &[ ), ( Format::Avif, - "AVIF encode uses ravif (pure Rust) but decoding requires dav1d (C); read stays unsupported \ - until a pure-Rust AV1 decoder is viable. Only enabled with the non-default `avif` feature.", + "AVIF encode (ravif and rav1e, pure Rust) is behind the non-default `avif` feature and AVIF \ + decode (craft-video's clean-room pure-Rust AV1 decoder) behind the non-default \ + `avif-decode` feature, native targets only (on wasm32 a decoder panic would abort). A build \ + with only one of them reads or writes AVIF, not both.", ), (Format::Jxl, "Lossless JPEG XL export uses the pure-Rust zune-jpegxl encoder; JPEG XL reading is tracked separately and is not yet available."), ( @@ -304,7 +307,9 @@ pub fn caps(format: Format) -> FormatCaps { Format::Qoi => base, Format::OpenExr => FormatCaps { depths: &[S::F16, S::F32], layouts: RGB_GRAY, ..base }, Format::Hdr => FormatCaps { depths: &[S::F32], layouts: &[L::Rgb], alpha: false, lossy: true, ..base }, - Format::Avif => FormatCaps { read: false, write: cfg!(feature = "avif"), lossy: true, ..base }, + Format::Avif => { + FormatCaps { read: cfg!(all(feature = "avif-decode", not(target_arch = "wasm32"))), write: cfg!(feature = "avif"), lossy: true, ..base } + } Format::Heif => { FormatCaps { read: cfg!(feature = "heif"), write: false, depths: &[S::U8, S::U16], icc: true, exif: true, xmp: true, lossy: true, ..base } } diff --git a/crates/image-codecs/src/jpeg_sef.rs b/crates/image-codecs/src/jpeg_sef.rs new file mode 100644 index 000000000..6ae878013 --- /dev/null +++ b/crates/image-codecs/src/jpeg_sef.rs @@ -0,0 +1,82 @@ +//! Validated Samsung SEF trailer repair for JPEGs missing their final EOI marker. +use std::borrow::Cow; + +/// Whether the first JPEG image reaches an EOI through its marker and scan structure. +pub fn reaches_eoi(bytes: &[u8]) -> bool { + crate::codecs::jpeg::reaches_eoi(bytes) +} + +/// Borrow ordinary input unchanged. For a validated Samsung SEF trailer following an open JPEG +/// scan, remove that trailer and append the missing EOI. The owned copy is made only on repair. +pub fn repair(bytes: &[u8]) -> Cow<'_, [u8]> { + let Some(image) = image_end(bytes).and_then(|end| bytes.get(..end)) else { return Cow::Borrowed(bytes) }; + if reaches_eoi(image) { + return Cow::Borrowed(bytes); + } + let mut fixed = Vec::with_capacity(image.len().saturating_add(2)); + fixed.extend_from_slice(image); + fixed.extend_from_slice(&[0xff, 0xd9]); + let in_scan = craft_metadata::inspect_jpeg(&fixed).is_some_and(|m| m.first_scan_components > 0) && reaches_eoi(&fixed); + if in_scan { Cow::Owned(fixed) } else { Cow::Borrowed(bytes) } +} + +fn image_end(b: &[u8]) -> Option { + let le32 = |at: usize| b.get(at..at.checked_add(4)?)?.try_into().ok().map(|v| u32::from_le_bytes(v) as usize); + let tail = b.len().checked_sub(8)?; + if b.get(tail.checked_add(4)?..)? != b"SEFT" { + return None; + } + let dir = tail.checked_sub(le32(tail)?)?; + if b.get(dir..dir.checked_add(4)?)? != b"SEFH" { + return None; + } + let count = le32(dir.checked_add(8)?)?; + let entries = dir.checked_add(12)?; + if count == 0 || count > tail.checked_sub(entries)? / 12 { + return None; + } + let mut start = dir; + for i in 0..count { + let entry = entries.checked_add(i.checked_mul(12)?)?; + let block = dir.checked_sub(le32(entry.checked_add(4)?)?)?; + let len = le32(entry.checked_add(8)?)?; + let kind = |at: usize| b.get(at.checked_add(2)?..at.checked_add(4)?); + if len < 4 || block.checked_add(len)? > dir || kind(block)? != kind(entry)? { + return None; + } + start = start.min(block); + } + Some(start) +} + +#[cfg(test)] +mod tests { + use super::*; + #[test] + fn repairs_only_valid_sef_after_an_open_scan() { + let image = [0xff, 0xd8, 0xff, 0xc0, 0, 8, 8, 0, 1, 0, 1, 1, 0xff, 0xda, 0, 3, 1, 0x12, 0xff, 0, 0x34]; + let mut b = image.to_vec(); + let block = [0, 0, 0xe0, 0x08, 1, 2, 3, 4]; + b.extend_from_slice(&block); + b.extend_from_slice(b"SEFH"); + b.extend_from_slice(&106u32.to_le_bytes()); + b.extend_from_slice(&1u32.to_le_bytes()); + b.extend_from_slice(&[0, 0, 0xe0, 0x08]); + b.extend_from_slice(&(block.len() as u32).to_le_bytes()); + b.extend_from_slice(&(block.len() as u32).to_le_bytes()); + b.extend_from_slice(&24u32.to_le_bytes()); + b.extend_from_slice(b"SEFT"); + let mut expected = image.to_vec(); + expected.extend_from_slice(&[0xff, 0xd9]); + assert_eq!(repair(&b).as_ref(), expected); + assert!(matches!(repair(&image), Cow::Borrowed(_))); + b[image.len() + 2] = 0; + assert!(matches!(repair(&b), Cow::Borrowed(_))); + } + #[test] + fn eoi_requires_valid_lengths_and_skips_stuffed_bytes() { + assert!(reaches_eoi(&[0xff, 0xd8, 0xff, 0, 0xff, 0xd9])); + assert!(!reaches_eoi(&[0xff, 0xd8, 0xff, 0xe1, 0, 1, 0xff, 0xd9])); + assert!(!reaches_eoi(&[0xff, 0xd8, 0xff, 0xe1, 0, 4, 0xff, 0xd9])); + } +} diff --git a/crates/image-codecs/src/lib.rs b/crates/image-codecs/src/lib.rs index adeed0825..c174c1b18 100644 --- a/crates/image-codecs/src/lib.rs +++ b/crates/image-codecs/src/lib.rs @@ -26,14 +26,16 @@ mod fidelity; mod format; mod image; mod inspection; +pub mod jpeg_sef; mod lab; mod options; pub mod orientation; pub mod png_probe; pub mod resolution; pub mod source_settings; +pub mod tiff_probe; pub mod web; -pub use crate::allocation::probe_allocation_size; +pub use crate::allocation::{probe_allocation_size, probe_stored_size}; pub use crate::cicp::Cicp; pub use crate::cicp::ContentLightLevel; pub use crate::codecs::avif::{HDR_REFERENCE_WHITE_NITS, HdrTransfer, encode_avif_hdr, encode_avif_with_speed}; @@ -109,7 +111,11 @@ pub fn decode_as_with(format: Format, bytes: &[u8], opts: &DecodeOptions) -> Res Format::Heif => heif::decode(bytes, l, opts.keep_orientation), Format::Jp2 => jp2::decode(bytes, l), Format::Jxl => Err(CodecError::unsupported(format, "JPEG XL reading is not yet available; export is supported")), - Format::Gif | Format::Bmp | Format::Tga | Format::Ico | Format::Qoi | Format::Hdr | Format::Avif => via_image::decode(format, bytes, l), + #[cfg(all(feature = "avif-decode", not(target_arch = "wasm32")))] + Format::Avif => codecs::avif_decode::decode(bytes, l, opts.keep_orientation), + #[cfg(not(all(feature = "avif-decode", not(target_arch = "wasm32"))))] + Format::Avif => via_image::decode(format, bytes, l), + Format::Gif | Format::Bmp | Format::Tga | Format::Ico | Format::Qoi | Format::Hdr => via_image::decode(format, bytes, l), }?; finish_format(format, bytes, img, opts) } @@ -166,10 +172,11 @@ fn finish_format(format: Format, bytes: &[u8], mut img: Image, opts: &DecodeOpti } // Turn the pixels upright, like Photoshop: a TIFF records it in the decoded page's own // directory, the others in their EXIF block. The metadata is rewritten to Orientation = 1 - // on the way. HEIF keeps it in its container, and its decoder has already applied it. + // on the way. HEIF and AVIF keep it in their container, and their decoders have already + // applied it. let o = match format { Format::Tiff => tiff::orientation(bytes, None), - Format::Heif => 1, + Format::Heif | Format::Avif => 1, Format::Dpx => dpx::orientation(bytes), Format::Cineon => dpx::cineon_orientation(bytes), _ => img.meta.exif.as_deref().map_or(1, exif_orientation), @@ -198,10 +205,11 @@ pub fn decode_png_rows>( } /// The upright pixel size [`decode`] would return, read from the file's header without decoding -/// pixels, for WebP, OpenEXR (flat) and the GIF/BMP/TGA/ICO/QOI/HDR family. It runs the same +/// pixels, for WebP, OpenEXR (flat), AVIF (with `avif-decode`) and the GIF/BMP/TGA/ICO/QOI/HDR family. It runs the same /// header reads and size checks as the decoder (default [`Limits`]) and swaps the size for an -/// EXIF orientation of 5–8, so wherever [`decode`] succeeds the two agree. Other formats (and -/// deep EXR or AVIF) return [`CodecError::Unsupported`]: decode those to learn their size. +/// EXIF orientation of 5–8 (AVIF: its container's `clap` and `irot`), so wherever [`decode`] +/// succeeds the two agree. Other formats (and deep EXR) return [`CodecError::Unsupported`]: decode +/// those to learn their size. pub fn probe_size(bytes: &[u8]) -> Result<(u32, u32), CodecError> { let format = detect(bytes).ok_or(CodecError::UnknownFormat)?; if !caps(format).read { @@ -211,7 +219,11 @@ pub fn probe_size(bytes: &[u8]) -> Result<(u32, u32), CodecError> { let (w, h, exif) = match format { Format::WebP => webp::probe(bytes, l)?, Format::OpenExr => exr::probe(bytes, l).map(|(w, h)| (w, h, None))?, - Format::Gif | Format::Bmp | Format::Tga | Format::Ico | Format::Qoi | Format::Hdr | Format::Avif => via_image::probe(format, bytes, l)?, + #[cfg(all(feature = "avif-decode", not(target_arch = "wasm32")))] + Format::Avif => codecs::avif_decode::probe(bytes, l).map(|(w, h)| (w, h, None))?, + #[cfg(not(all(feature = "avif-decode", not(target_arch = "wasm32"))))] + Format::Avif => via_image::probe(format, bytes, l)?, + Format::Gif | Format::Bmp | Format::Tga | Format::Ico | Format::Qoi | Format::Hdr => via_image::probe(format, bytes, l)?, _ => return Err(CodecError::unsupported(format, "no header-only size probe")), }; let o = exif.as_deref().map_or(1, exif_orientation); diff --git a/crates/image-codecs/src/tiff_probe.rs b/crates/image-codecs/src/tiff_probe.rs new file mode 100644 index 000000000..1c6889c21 --- /dev/null +++ b/crates/image-codecs/src/tiff_probe.rs @@ -0,0 +1,101 @@ +//! First-directory TIFF inspection without decoding pixels. + +/// A header inspection failure, distinguishing caller limits from unsupported decoder layouts. +#[derive(Debug, thiserror::Error)] +pub enum Error { + /// Dimensions exceed the caller's pixel limit. + #[error("TIFF dimensions {width}x{height} exceed the pixel limit")] + TooLarge { + /// Stored width. + width: u32, + /// Stored height. + height: u32, + }, + /// The shared TIFF decoder does not support this layout. + #[error("unsupported TIFF: {0}")] + Unsupported(&'static str), + /// The file's header, directory or image-data bounds are malformed. + #[error("malformed TIFF: {0}")] + Malformed(String), +} + +/// First-directory values needed to preflight decoding. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub struct DecodeInfo { + /// Stored width. + pub width: u32, + /// Stored height. + pub height: u32, + /// First sample's bit depth, defaulting to eight. + pub bit_depth: u8, + /// Raw orientation tag, if present. + pub orientation: Option, +} + +fn first_directory(bytes: &[u8], max_pixels: u64) -> Result<(craft_tiff::Ifd, craft_tiff::image::ImageInfo), Error> { + let (tiff, complete) = craft_tiff::Tiff::parse_with_status(bytes, &Default::default()).map_err(|e| Error::Malformed(e.to_string()))?; + if !complete { + return Err(Error::Malformed("truncated TIFF".into())); + } + let ifd = tiff.ifds.into_iter().next().ok_or_else(|| Error::Malformed("no IFD".into()))?; + let info = ifd.image().map_err(|e| Error::Malformed(e.to_string()))?; + if info.width == 0 || info.height == 0 { + return Err(Error::Malformed("zero dimension".into())); + } + if u64::from(info.width) * u64::from(info.height) > max_pixels { + return Err(Error::TooLarge { width: info.width, height: info.height }); + } + Ok((ifd, info)) +} + +/// Parse the first directory and enforce the pixel limit before decoding. +/// Unlike [`header`], this does not impose layout or strip checks ahead of the decoder. +pub fn decode_info(bytes: &[u8], max_pixels: u64) -> Result { + let (ifd, info) = first_directory(bytes, max_pixels)?; + Ok(DecodeInfo { + width: info.width, + height: info.height, + bit_depth: info.bits_per_sample.first().copied().unwrap_or(8) as u8, + orientation: ifd.u16(craft_tiff::tags::ORIENTATION), + }) +} + +/// Stored dimensions and orientation of the first directory, after checking the layouts and +/// image-data bounds required by the TIFF decoder. Compressed strips are not decompressed. +pub fn header(bytes: &[u8], max_pixels: u64) -> Result<(u32, u32, u16), Error> { + let (ifd, info) = first_directory(bytes, max_pixels)?; + let bits = info.bits_per_sample.first().copied().unwrap_or(1); + if bits < 8 && info.planar == 2 && info.samples_per_pixel > 1 { + return Err(Error::Unsupported("planar TIFF with samples narrower than a byte")); + } + if !(matches!(info.photometric, 0 | 1 | 2 | 3 | 5) || info.photometric == 6 && info.compression == 7) { + return Err(Error::Unsupported("colour type")); + } + if !matches!(info.compression, 1 | 4 | 5 | 7 | 8 | 32773 | 32946) { + return Err(Error::Unsupported("compression method")); + } + let len = bytes.len() as u64; + if info.offsets.is_empty() + || info.offsets.iter().zip(&info.byte_counts).any(|(&o, &c)| c > 0 && (o >= len || (info.compression == 1 && o.saturating_add(c) > len))) + { + return Err(Error::Malformed("truncated: image data past the end of the file".into())); + } + let orientation = ifd.u16(craft_tiff::tags::ORIENTATION).filter(|o| (1..=8).contains(o)).unwrap_or(1); + Ok((info.width, info.height, orientation)) +} + +#[cfg(test)] +mod tests { + use super::*; + #[test] + fn reads_first_directory_and_enforces_limits_and_bounds() { + let image = crate::Image::from_u8(2, 3, crate::ChannelLayout::Rgb, vec![42; 18]).unwrap(); + let bytes = crate::encode(&image, crate::Format::Tiff, &crate::EncodeOptions::default()).unwrap(); + assert_eq!(header(&bytes, 6).unwrap(), (2, 3, 1)); + let info = decode_info(&bytes, 6).unwrap(); + assert_eq!((info.width, info.height, info.bit_depth), (2, 3, 8)); + assert!(matches!(decode_info(&bytes, 5), Err(Error::TooLarge { width: 2, height: 3 }))); + assert!(matches!(header(&bytes, 5), Err(Error::TooLarge { width: 2, height: 3 }))); + assert!(header(&bytes[..8], 6).is_err()); + } +} diff --git a/crates/image-codecs/tests/fixtures/avif/README.md b/crates/image-codecs/tests/fixtures/avif/README.md new file mode 100644 index 000000000..b8b3a8c2a --- /dev/null +++ b/crates/image-codecs/tests/fixtures/avif/README.md @@ -0,0 +1,12 @@ +# AVIF decode fixtures + +Small AVIF files from an independent encoder (libaom through FFmpeg 7.1), for `src/codecs/avif_decode.rs`. + +- `quad-source.rgb`: the 33×17 source, raw RGB24 four-colour quadrants (red, green, blue, near-white). +- `quad-420-bt709.avif`: 8-bit 4:2:0, limited range, BT.709 matrix, odd size (FFmpeg leaves the primaries and transfer unspecified): + `ffmpeg -f rawvideo -pix_fmt rgb24 -s 33x17 -i quad-source.rgb -c:v libaom-av1 -crf 12 -still-picture 1 -pix_fmt yuv420p -colorspace bt709 -color_primaries bt709 -color_trc iec61966-2-1 -color_range tv quad-420-bt709.avif` +- `quad-420-bt709.dav1d.yuv`: that file's AV1 picture as libdav1d decodes it, planar 8-bit 4:2:0 (Y, then Cb, then Cr): + `ffmpeg -c:v libdav1d -i quad-420-bt709.avif -f rawvideo -pix_fmt yuv420p quad-420-bt709.dav1d.yuv` +- `quad-444-gbr-lossless.avif`: lossless 8-bit 4:4:4 with the identity (GBR) matrix, full range: + `ffmpeg -f rawvideo -pix_fmt rgb24 -s 33x17 -i quad-source.rgb -c:v libaom-av1 -crf 0 -still-picture 1 -pix_fmt gbrp -colorspace rgb -color_range pc quad-444-gbr-lossless.avif`. + It decodes to `quad-source.rgb` exactly. diff --git a/crates/image-codecs/tests/fixtures/avif/quad-420-bt709.avif b/crates/image-codecs/tests/fixtures/avif/quad-420-bt709.avif new file mode 100644 index 000000000..419196cb1 Binary files /dev/null and b/crates/image-codecs/tests/fixtures/avif/quad-420-bt709.avif differ diff --git a/crates/image-codecs/tests/fixtures/avif/quad-420-bt709.dav1d.yuv b/crates/image-codecs/tests/fixtures/avif/quad-420-bt709.dav1d.yuv new file mode 100644 index 000000000..95e4ac464 --- /dev/null +++ b/crates/image-codecs/tests/fixtures/avif/quad-420-bt709.dav1d.yuv @@ -0,0 +1 @@ +OOOOOOOOOOOOOOOOO’’’’’’’’’’’’’’’’OOOOOOOOOOOOOOOOO’’’’’’’’’’’’’’’’OOOOOOOOOOOOOOOOO’’’’’’’’’’’’’’’’OOOOOOOOOOOOOOOOO’’’’’’’’’’’’’’’’OOOOOOOOOOOOOOOOO’’’’’’’’’’’’’’’’OOOOOOOOOOOOOOOOO’’’’’’’’’’’’’’’’OOOOOOOOOOOOOOOOO’’’’’’’’’’’’’’’’OOOOOOOOOOOOOOOOO’’’’’’’’’’’’’’’’OOOOOOOOOOOOOOOOO’’’’’’’’’’’’’’’’HHHHHHHHHHHHHHHHHÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞHHHHHHHHHHHHHHHHHÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞHHHHHHHHHHHHHHHHHÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞHHHHHHHHHHHHHHHHHÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞHHHHHHHHHHHHHHHHHÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞHHHHHHHHHHHHHHHHHÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞHHHHHHHHHHHHHHHHHÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞHHHHHHHHHHHHHHHHHÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞÞppppppppbUUUUUUUbppppppppbUUUUUUUbppppppppbUUUUUUUbmmmmmmmmbUUUUUUUb……………………r_______ŠÍÍÍÍÍÍÍͦ�������¦ÉÉÉÉÉÉÉɦ�������¦ÉÉÉÉÉÉÉɦ�������¦ÉÉÉÉÉÉÉɦ�������‘×××××××ׇ6666666‡×××××××ׇ6666666‡×××××××ׇ6666666‡ÛÛÛÛÛÛÛÛ‡6666666‡¾¾¾¾¾¾¾¾ƒHHHHHHHghhhhhhhhv�������vllllllllv�������vllllllllv�������vllllllllv�������| \ No newline at end of file diff --git a/crates/image-codecs/tests/fixtures/avif/quad-444-gbr-lossless.avif b/crates/image-codecs/tests/fixtures/avif/quad-444-gbr-lossless.avif new file mode 100644 index 000000000..c3b1f1423 Binary files /dev/null and b/crates/image-codecs/tests/fixtures/avif/quad-444-gbr-lossless.avif differ diff --git a/crates/image-codecs/tests/fixtures/avif/quad-source.rgb b/crates/image-codecs/tests/fixtures/avif/quad-source.rgb new file mode 100644 index 000000000..9e77c1dd1 --- /dev/null +++ b/crates/image-codecs/tests/fixtures/avif/quad-source.rgb @@ -0,0 +1 @@ +æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(æ(È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<È<<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Üðððððððððððððððððððððððððððððððððððððððððððððððð<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Üðððððððððððððððððððððððððððððððððððððððððððððððð<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Üðððððððððððððððððððððððððððððððððððððððððððððððð<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Üðððððððððððððððððððððððððððððððððððððððððððððððð<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Üðððððððððððððððððððððððððððððððððððððððððððððððð<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Üðððððððððððððððððððððððððððððððððððððððððððððððð<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Üðððððððððððððððððððððððððððððððððððððððððððððððð<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Ü<Üðððððððððððððððððððððððððððððððððððððððððððððððð \ No newline at end of file diff --git a/crates/image-codecs/tests/it/detect_caps.rs b/crates/image-codecs/tests/it/detect_caps.rs index 390eaa41f..e4868f97c 100644 --- a/crates/image-codecs/tests/it/detect_caps.rs +++ b/crates/image-codecs/tests/it/detect_caps.rs @@ -70,10 +70,15 @@ fn default_build_formats_all_symmetric() { } #[test] -fn avif_never_readable() { - assert!(!caps(Format::Avif).read); +fn avif_is_readable_only_with_the_decode_feature() { + assert_eq!(caps(Format::Avif).read, cfg!(all(feature = "avif-decode", not(target_arch = "wasm32")))); assert_eq!(caps(Format::Avif).write, cfg!(feature = "avif")); - assert!(matches!(decode_as(Format::Avif, b"\0\0\0\x1cftypavif"), Err(CodecError::Unsupported { .. }))); + let header = decode_as(Format::Avif, b"\0\0\0\x1cftypavif"); + if caps(Format::Avif).read { + assert!(matches!(header, Err(CodecError::Malformed { .. })), "{header:?}"); + } else { + assert!(matches!(header, Err(CodecError::Unsupported { .. })), "{header:?}"); + } } #[test] diff --git a/crates/jpeg/src/app14.rs b/crates/jpeg/src/app14.rs new file mode 100644 index 000000000..99ddabac1 --- /dev/null +++ b/crates/jpeg/src/app14.rs @@ -0,0 +1,40 @@ +//! Adobe APP14 transform inspection for JPEG samples. + +/// Read the Adobe APP14 transform byte needed for RGB/CMYK JPEG interpretation. +pub fn adobe_transform(jpeg: &[u8]) -> Option { + let mut at = 2usize; + if jpeg.get(..2) != Some(&[0xFF, 0xD8]) { + return None; + } + while at + 1 < jpeg.len() && jpeg[at] == 0xFF { + let marker = jpeg[at + 1]; + if marker == 0xDA { + return None; + } + let len = usize::from(u16::from_be_bytes([*jpeg.get(at + 2)?, *jpeg.get(at + 3)?])); + if len < 2 { + return None; + } + let end = at.checked_add(2 + len)?; + let segment = jpeg.get(at + 4..end)?; + if marker == 0xEE && segment.get(..5) == Some(b"Adobe") { + return segment.get(11).copied(); + } + at = end; + } + None +} + +#[cfg(test)] +mod tests { + use super::adobe_transform; + #[test] + fn reads_adobe_transform_and_rejects_truncated_segments() { + let mut jpeg = vec![0xff, 0xd8, 0xff, 0xee, 0, 14]; + jpeg.extend_from_slice(b"Adobe\0\x64\0\0\0\0\x02"); + assert_eq!(adobe_transform(&jpeg), Some(2)); + assert_eq!(adobe_transform(&jpeg[..jpeg.len() - 1]), None); + assert_eq!(adobe_transform(&[0xff, 0xd8, 0xff, 0xda]), None); + assert_eq!(adobe_transform(&[0xff, 0xd8, 0xff, 0xee, 0, 1]), None); + } +} diff --git a/crates/jpeg/src/lib.rs b/crates/jpeg/src/lib.rs index 50e1bd625..23434fdd0 100644 --- a/crates/jpeg/src/lib.rs +++ b/crates/jpeg/src/lib.rs @@ -1706,3 +1706,9 @@ mod tests; /// Baseline JPEG encoder (8-bit, Huffman-coded sequential DCT). mod encode; pub use encode::{ChromaSubsampling, EncodeError, encode_baseline}; + +mod app14; +pub use app14::adobe_transform; + +mod rgba; +pub use rgba::{UnsupportedComponents, to_rgba8}; diff --git a/crates/jpeg/src/rgba.rs b/crates/jpeg/src/rgba.rs new file mode 100644 index 000000000..495c5141f --- /dev/null +++ b/crates/jpeg/src/rgba.rs @@ -0,0 +1,44 @@ +//! Conversion of decoded 8-bit JPEG samples to opaque RGBA. + +/// The JPEG component count is not supported by the RGBA conversion. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub struct UnsupportedComponents(pub u8); + +/// Convert grayscale, RGB or CMYK samples to opaque RGBA, ignoring incomplete trailing pixels. +pub fn to_rgba8(samples: &[u8], components: u8) -> Result, UnsupportedComponents> { + let mut rgba = Vec::with_capacity(samples.len().saturating_mul(4)); + match components { + 1 => { + for &gray in samples { + rgba.extend_from_slice(&[gray, gray, gray, 255]); + } + } + 3 => { + for rgb in samples.as_chunks::<3>().0 { + rgba.extend_from_slice(&[rgb[0], rgb[1], rgb[2], 255]); + } + } + 4 => { + for cmyk in samples.as_chunks::<4>().0 { + let (c, m, y, k) = (u16::from(cmyk[0]), u16::from(cmyk[1]), u16::from(cmyk[2]), u16::from(cmyk[3])); + let rgb = [((255 - c) * (255 - k) / 255) as u8, ((255 - m) * (255 - k) / 255) as u8, ((255 - y) * (255 - k) / 255) as u8]; + rgba.extend_from_slice(&[rgb[0], rgb[1], rgb[2], 255]); + } + } + n => return Err(UnsupportedComponents(n)), + } + Ok(rgba) +} + +#[cfg(test)] +mod tests { + use super::*; + #[test] + fn gray_rgb_and_cmyk_conversion_matches_integer_rules() -> Result<(), UnsupportedComponents> { + assert_eq!(to_rgba8(&[12], 1)?, [12, 12, 12, 255]); + assert_eq!(to_rgba8(&[1, 2, 3, 4], 3)?, [1, 2, 3, 255]); + assert_eq!(to_rgba8(&[0, 255, 128, 0], 4)?, [255, 0, 127, 255]); + assert_eq!(to_rgba8(&[], 2), Err(UnsupportedComponents(2))); + Ok(()) + } +} diff --git a/crates/jsonrpc/src/control.rs b/crates/jsonrpc/src/control.rs new file mode 100644 index 000000000..c4c3a0389 --- /dev/null +++ b/crates/jsonrpc/src/control.rs @@ -0,0 +1,57 @@ +//! Bounded compatibility framing for single-line `{id, method, params}` control requests. +//! This is not JSON-RPC: replies use `{id, ok, result}` or `{id, ok, error}` and every +//! request, including one without an id, receives a reply. Method routing stays with the caller. +use serde_json::{Value, json}; +use std::fmt::Display; + +/// Process one bounded control line. Blank lines yield no response. Missing `params` becomes null; +/// arbitrary JSON ids are echoed and malformed JSON has a null id, matching the legacy protocol. +pub fn process_line(line: &str, max_bytes: usize, dispatch: impl FnOnce(&str, &Value) -> Result) -> Option { + if line.len() > max_bytes { + return Some(error(None, "request exceeds byte limit")); + } + let trimmed = line.trim(); + if trimmed.is_empty() { + return None; + } + let value: Value = match serde_json::from_str(trimmed) { + Ok(v) => v, + Err(e) => return Some(error(None, &format!("bad request envelope: {e}"))), + }; + let id = value.get("id").cloned(); + let Some(method) = value.get("method").and_then(Value::as_str) else { + return Some(error(id, "missing `method`")); + }; + let params = value.get("params").cloned().unwrap_or(Value::Null); + Some(match dispatch(method, ¶ms) { + Ok(result) => json!({"id": id, "ok": true, "result": result}).to_string(), + Err(e) => error(id, &e.to_string()), + }) +} + +fn error(id: Option, message: &str) -> String { + json!({"id": id, "ok": false, "error": message}).to_string() +} + +#[cfg(test)] +mod tests { + use super::*; + #[test] + fn preserves_legacy_responses_and_null_defaults() { + let call = + |method: &str, params: &Value| -> Result { if method == "echo" { Ok(params.clone()) } else { Err("unknown") } }; + assert_eq!(process_line(" ", 20, call), None); + assert_eq!(process_line(r#"{"id":[1],"method":"echo"}"#, 100, call).unwrap(), r#"{"id":[1],"ok":true,"result":null}"#); + assert_eq!(process_line(r#"{"method":"bad"}"#, 100, call).unwrap(), r#"{"error":"unknown","id":null,"ok":false}"#); + assert_eq!(process_line(r#"{"method":"echo"}"#, 100, call).unwrap(), r#"{"id":null,"ok":true,"result":null}"#); + assert!(process_line("{", 100, call).unwrap().contains("bad request envelope:")); + assert_eq!(process_line(r#"{"id":3}"#, 100, call).unwrap(), r#"{"error":"missing `method`","id":3,"ok":false}"#); + assert!(process_line("abcdef", 2, call).unwrap().contains("byte limit")); + let invoked = std::cell::Cell::new(false); + let _ = process_line(" ", 2, |_, _| -> Result { + invoked.set(true); + Ok(Value::Null) + }); + assert!(!invoked.get()); + } +} diff --git a/crates/jsonrpc/src/lib.rs b/crates/jsonrpc/src/lib.rs index cfb46ef79..653b536e8 100644 --- a/crates/jsonrpc/src/lib.rs +++ b/crates/jsonrpc/src/lib.rs @@ -8,6 +8,8 @@ #![forbid(unsafe_code)] #![deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable)] +pub mod control; + use serde_json::{Value, json}; use std::io::BufRead; diff --git a/crates/media/README.md b/crates/media/README.md index b7ad5be7a..6663da282 100644 --- a/crates/media/README.md +++ b/crates/media/README.md @@ -447,7 +447,13 @@ page yet. Damage recovery may read bytes again up to twice the file's length plu is read at most about three times over; the packets kept and the header packets of a stream are bounded too, with a warning when a bound is reached. `OpusTiming` gives each Opus packet's start, the pre-skip and the end trimming of the final page. -No muxer. + +The muxer side is `PageWriter`, from FilmCraft's `filmcraft-oggmux` by @JordyUSA +(storytold/filmcraft#4481): it laces the packets of one logical stream into pages (continued +packets, granule positions, BOS/EOS, CRC-32), puts each header packet on pages of its own, closes +audio pages at a byte or packet target, and can replace the final granule position for Ogg Opus +end trimming. `opus_tags` and `vorbis_comment` build the comment headers; tag names that are not +valid field names (Vorbis I §5.2.1) are left out. ## `avi` @@ -655,6 +661,7 @@ Implemented from the SMPTE documents (editions used): | SMPTE ST 386 | 2004 | D-10 (IMX) mapping: container label, sound element type | | SMPTE ST 2019-4 | 2009 | VC-3 (DNxHD / DNxHR) mapping | | SMPTE RDD 44 | 2017 | Apple ProRes mapping | +| Adobe CinemaDNG Image Data Format Specification | 1.0.0.0 (September 2009), §7 | CinemaDNG in MXF: frame-wrapped DNG picture elements (element type 42h), the essence container labels for CFA / LinearRaw content, RGBA descriptor; the labels are the specification's **provisional** values from Adobe's experimental range (byte 9 = 0Fh), not SMPTE-registered | | SMPTE RP 224 / RP 210 | registers | labels: data definitions, picture coding, essence containers, operational patterns; property labels of the primer pack (writer) | | SMPTE ST 330 | 2011 | basic UMIDs of the written packages | @@ -696,7 +703,10 @@ let pcm = mxf.read_pcm(&file, a, 48_000, 1920)?; // planar f32, sample-exact orders by the bitstream (the AVC picture order counts). - **Codecs.** From the picture essence coding label, then the essence container label, then the essence bytes: AVC / AVC-Intra, VC-3, ProRes (profile from the label), MPEG-2, - MPEG-4 visual, JPEG 2000, DV, uncompressed. + MPEG-4 visual, JPEG 2000, DV, uncompressed, CinemaDNG (the CinemaDNG container labels, the + registered MXF-GC TIFF/EP container and TIFF/EP picture coding labels, a 42h picture element + type, or a TIFF header with the DNGVersion tag at the start of the essence; the DNG pictures + themselves are decoded by `craft-raw`). - **Sound.** PCM chunks (frame- or clip-wrapped) with sample counts; `read_pcm` decodes little-endian 8/16/24/32-bit PCM (wave / AES3 descriptors) and ST 331 AES3 elements (D-10). - **Timecode.** The material package timecode component (start frame count, rounded base, @@ -769,7 +779,10 @@ w.finish()?; element types 0x0C / 0x0D), ProRes (RDD 44: coding label profile byte, container `02 1C`), AVC byte stream (ST 381-3: container `02 10 60`, coding label from `profile_idc`; Annex B access units with in-band SPS / PPS — the AVC sub-descriptor is not written), Broadcast Wave PCM - (ST 382: container `02 06 01/02`, element types 0x01 / 0x02). + (ST 382: container `02 06 01/02`, element types 0x01 / 0x02), CinemaDNG (frame-wrapped DNG + datasets, element type 0x42, the provisional container label `0F 0A 0D 0B 02 7F 01/02 00`, an + RGBA descriptor with pixel layout and reference levels of the decoded 16-bit RGB picture and no + coding label; OP-Atom is refused because it is clip-wrapped here and DNG datasets cannot be). - `write_opatom_pcm(&OpAtomPcm, &[i32])` builds a complete in-memory OP-Atom PCM file (AAF media consolidation); `encode_pcm` interleaves planar f32 as little-endian PCM. diff --git a/crates/media/src/mxf/essence.rs b/crates/media/src/mxf/essence.rs index 27f893dc5..74818e0e8 100644 --- a/crates/media/src/mxf/essence.rs +++ b/crates/media/src/mxf/essence.rs @@ -27,6 +27,9 @@ pub enum Codec { Dv, /// Uncompressed pictures (ST 384). Uncompressed, + /// CinemaDNG: one DNG (TIFF) dataset per frame-wrapped picture element (CinemaDNG Image Data + /// Format Specification §7), or TIFF/EP pictures under the registered MXF-GC TIFF/EP labels. + CinemaDng, /// Linear PCM, little-endian (Broadcast Wave and AES3 descriptors, ST 382). Pcm, /// 8-channel AES3 element data (ST 331, D-10 sound). @@ -46,13 +49,14 @@ impl Codec { Codec::Jpeg2000 => "JPEG 2000", Codec::Dv => "DV", Codec::Uncompressed => "Uncompressed", + Codec::CinemaDng => "CinemaDNG", Codec::Pcm => "PCM", Codec::Aes3Element => "AES3 PCM", Codec::Unknown => "unknown", } } /// The crate-wide identifier: VC-3 is [`CodecId::Dnxhd`], PCM and AES3 element data are - /// [`CodecId::Pcm`]; JPEG 2000, DV and uncompressed pictures have no identifier. + /// [`CodecId::Pcm`]; JPEG 2000, DV, uncompressed and CinemaDNG pictures have no identifier. /// /// [`CodecId::Dnxhd`]: crate::CodecId::Dnxhd /// [`CodecId::Pcm`]: crate::CodecId::Pcm @@ -65,15 +69,35 @@ impl Codec { Codec::Mpeg2 => Id::Mpeg2Video, Codec::Mpeg4Visual => Id::Mpeg4Video, Codec::Pcm | Codec::Aes3Element => Id::Pcm, - Codec::Jpeg2000 | Codec::Dv | Codec::Uncompressed | Codec::Unknown => Id::Unknown, + Codec::Jpeg2000 | Codec::Dv | Codec::Uncompressed | Codec::CinemaDng | Codec::Unknown => Id::Unknown, } } /// Every picture is coded independently. pub fn intra_only(&self) -> bool { - matches!(self, Codec::Avc { intra: true } | Codec::Vc3 | Codec::ProRes { .. } | Codec::Jpeg2000 | Codec::Dv | Codec::Uncompressed) + matches!( + self, + Codec::Avc { intra: true } | Codec::Vc3 | Codec::ProRes { .. } | Codec::Jpeg2000 | Codec::Dv | Codec::Uncompressed | Codec::CinemaDng + ) } } +/// The CinemaDNG essence container labels (CinemaDNG Image Data Format Specification 1.0.0.0, +/// §7.2.1): provisional, unregistered labels from Adobe's experimental range (byte 9 = 0F h), +/// `…0F 0A 0D 0B 02 7F cc 00` with content kind `cc` 01 (CFA) or 02 (LinearRaw). +pub const CINEMADNG_CONTAINER_PREFIX: [u8; 6] = [0x0F, 0x0A, 0x0D, 0x0B, 0x02, 0x7F]; + +/// Element type (byte 15 of the picture element key) of a frame-wrapped CinemaDNG picture: the +/// "TIFF" type of ST 331 (CinemaDNG specification §7.1.1.2). +pub const CINEMADNG_ELEMENT_TYPE: u8 = 0x42; + +/// Whether `ul` is a CinemaDNG essence container label (CFA or LinearRaw content), or the +/// registered MXF-GC TIFF/EP container (`…0D 01 03 01 02 14 ww 00`, RP 224). +pub fn is_cinemadng_container(ul: &Ul) -> bool { + ul.is_smpte() + && (ul.item_starts_with(&CINEMADNG_CONTAINER_PREFIX) && matches!(ul.0[14], 0x01 | 0x02) + || ul.item_starts_with(&[0x0D, 0x01, 0x03, 0x01, 0x02, 0x14])) +} + /// Identify a picture codec from the picture essence coding label and the essence container label. pub fn identify_picture(coding: Option
    , container: Option
      , mpeg_descriptor: bool) -> Codec { if let Some(c) = coding.filter(Ul::is_smpte) { @@ -81,6 +105,9 @@ pub fn identify_picture(coding: Option
        , container: Option
          , mpeg_descript // 04 01 02 02 .. : compressed picture coding if b[8..12] == [0x04, 0x01, 0x02, 0x02] { match (b[12], b[13]) { + // TIFF/EP Profile 2 picture coding (RP 224: 03 02 01 01 uncompressed CFA … 04 + // compressed LinearRaw), the coding of DNG pictures + (0x03, 0x02) if b[14] == 0x01 => return Codec::CinemaDng, // 01: MPEG compression (ISO/IEC 13818-2 profiles, 14496-2, 14496-10) (0x01, 0x01..=0x1F) => return Codec::Mpeg2, (0x01, 0x20..=0x2F) => return Codec::Mpeg4Visual, @@ -99,6 +126,9 @@ pub fn identify_picture(coding: Option
            , container: Option
              , mpeg_descript return Codec::Uncompressed; } } + if container.is_some_and(|c| is_cinemadng_container(&c)) { + return Codec::CinemaDng; + } if let Some(c) = container.filter(Ul::is_smpte) && c.item_starts_with(&[0x0D, 0x01, 0x03, 0x01, 0x02]) { @@ -117,8 +147,38 @@ pub fn identify_picture(coding: Option
                , container: Option
                  , mpeg_descript if mpeg_descriptor { Codec::Mpeg2 } else { Codec::Unknown } } +/// Read a TIFF `u16` at offset `o`. The end offset is computed with checked arithmetic, so an +/// untrusted offset near `usize::MAX` (a 32-bit target with IFD offset `0xFFFF_FFFF`) is +/// out of bounds instead of overflowing. +fn tiff_u16(b: &[u8], le: bool, o: usize) -> Option { + let v = b.get(o..o.checked_add(2)?)?; + Some(if le { u16::from_le_bytes([v[0], v[1]]) } else { u16::from_be_bytes([v[0], v[1]]) }) +} + +/// DNG evidence in the leading bytes of a TIFF: a TIFF 6.0 §2 header whose first IFD, when it lies +/// within `b`, holds the DNGVersion tag (DNG 1.x §2, tag 50706). A plain TIFF has no such tag. +fn is_dng(b: &[u8]) -> bool { + let le = match b.get(..4) { + Some(h) if h == b"II*\0" => true, + Some(h) if h == b"MM\0*" => false, + _ => return false, + }; + let u16_at = |o: usize| tiff_u16(b, le, o); + let Some(ifd) = b.get(4..8).map(|v| if le { u32::from_le_bytes([v[0], v[1], v[2], v[3]]) } else { u32::from_be_bytes([v[0], v[1], v[2], v[3]]) }) + else { + return false; + }; + let Some(count) = usize::try_from(ifd).ok().and_then(|o| u16_at(o).map(|n| (o, n))) else { + return false; + }; + (0..usize::from(count.1)).any(|i| u16_at(count.0 + 2 + i * 12) == Some(50706)) +} + /// Identify a picture codec from the first bytes of an essence element. pub fn sniff_picture(b: &[u8]) -> Option { + if is_dng(b) { + return Some(Codec::CinemaDng); + } if b.len() >= 8 && &b[4..8] == b"icpf" { return Some(Codec::ProRes { profile: None }); } @@ -367,6 +427,29 @@ mod tests { assert_eq!(identify_picture(None, None, false), Codec::Unknown); } + #[test] + fn tiff_offset_near_usize_max_is_out_of_bounds() { + let b = [0x49, 0x49, 0x2A, 0x00, 0xFF, 0xFF, 0xFF, 0xFF]; + assert_eq!(tiff_u16(&b, true, usize::MAX), None); + assert_eq!(tiff_u16(&b, true, usize::MAX - 1), None); + assert_eq!(tiff_u16(&b, true, 0), Some(0x4949)); + assert!(!is_dng(&b)); + } + + #[test] + fn identifies_cinemadng() { + // CinemaDNG specification §7.2.1: CFA and LinearRaw containers (any registry version byte) + let adobe = |kind: u8| Ul([0x06, 0x0E, 0x2B, 0x34, 0x04, 0x01, 0x01, 0x01, 0x0F, 0x0A, 0x0D, 0x0B, 0x02, 0x7F, kind, 0x00]); + assert_eq!(identify_picture(None, Some(adobe(1)), false), Codec::CinemaDng); + assert_eq!(identify_picture(None, Some(adobe(2)), false), Codec::CinemaDng); + assert_eq!(identify_picture(None, Some(adobe(3)), false), Codec::Unknown); + // RP 224: MXF-GC frame-wrapped TIFF/EP container, TIFF/EP picture coding + assert_eq!(identify_picture(None, Some(ul(&[0x0D, 1, 3, 1, 2, 0x14, 1])), false), Codec::CinemaDng); + assert_eq!(identify_picture(Some(ul(&[4, 1, 2, 2, 3, 2, 1, 3])), None, false), Codec::CinemaDng); + assert!(Codec::CinemaDng.intra_only()); + assert_eq!(Codec::CinemaDng.name(), "CinemaDNG"); + } + #[test] fn sniffs_essence() { assert_eq!(sniff_picture(&[0, 0, 0, 1, 0x09, 0x10]), Some(Codec::Avc { intra: false })); @@ -374,6 +457,15 @@ mod tests { assert_eq!(sniff_picture(&[0, 0, 2, 0x80, 1]), Some(Codec::Vc3)); assert_eq!(sniff_picture(&[0, 0, 0, 9, b'i', b'c', b'p', b'f']), Some(Codec::ProRes { profile: None })); assert_eq!(sniff_picture(&[1, 2, 3]), None); + // a TIFF header alone is not DNG; an IFD0 with the DNGVersion tag (50706 = 0xC612) is + let tiff_le = |tag: u16| [b"II*\0\x08\0\0\0\x01\0".as_slice(), &tag.to_le_bytes(), &[1, 0, 4, 0, 0, 0, 1, 4, 0, 0], &[0; 8]].concat(); + let tiff_be = |tag: u16| [b"MM\0*\0\0\0\x08\0\x01".as_slice(), &tag.to_be_bytes(), &[0, 1, 0, 0, 0, 4, 1, 4, 0, 0], &[0; 8]].concat(); + assert_eq!(sniff_picture(&tiff_le(0xC612)), Some(Codec::CinemaDng)); + assert_eq!(sniff_picture(&tiff_be(0xC612)), Some(Codec::CinemaDng)); + assert_eq!(sniff_picture(&tiff_le(0x0100)), None); + assert_eq!(sniff_picture(&tiff_be(0x0100)), None); + assert_eq!(sniff_picture(b"II*\0\x08\0\0\0"), None); + assert_eq!(sniff_picture(b"II*\0"), None); } #[test] diff --git a/crates/media/src/mxf/file.rs b/crates/media/src/mxf/file.rs index 15d847c7d..e23471307 100644 --- a/crates/media/src/mxf/file.rs +++ b/crates/media/src/mxf/file.rs @@ -571,8 +571,32 @@ fn descriptor_kind(s: &Set) -> Option { } } +// PictureInfo has one depth for all color components. Do not invent one for a +// malformed layout or unequal RGB precisions. PixelLayout is eight code/size pairs. +fn rgba_component_depth(layout: &[u8]) -> u32 { + if layout.len() != 16 { + return 0; + } + let mut depths = [0u8; 3]; + for pair in layout.as_chunks::<2>().0 { + let index = match pair[0] { + b'R' => 0, + b'G' => 1, + b'B' => 2, + 0 => break, + _ => continue, + }; + if depths[index] != 0 || pair[1] == 0 { + return 0; + } + depths[index] = pair[1]; + } + if depths[0] != 0 && depths.iter().all(|&depth| depth == depths[0]) { u32::from(depths[0]) } else { 0 } +} + fn picture_info(d: &Set) -> PictureInfo { let u = |t| d.u(t).unwrap_or(0) as u32; + let rgba = d.kind() == Some(st::RGBA_DESCRIPTOR); PictureInfo { stored_width: u(tag::STORED_WIDTH), stored_height: u(tag::STORED_HEIGHT), @@ -581,13 +605,13 @@ fn picture_info(d: &Set) -> PictureInfo { frame_layout: u(tag::FRAME_LAYOUT) as u8, field_dominance: u(tag::FIELD_DOMINANCE) as u8, aspect_ratio: d.rational(tag::ASPECT_RATIO).unwrap_or_default(), - component_depth: u(tag::COMPONENT_DEPTH), + component_depth: if rgba { d.get(tag::PIXEL_LAYOUT).map_or(0, rgba_component_depth) } else { u(tag::COMPONENT_DEPTH) }, horizontal_subsampling: u(tag::HORIZONTAL_SUBSAMPLING), vertical_subsampling: u(tag::VERTICAL_SUBSAMPLING), - black_ref_level: d.u(tag::BLACK_REF_LEVEL).map(|v| v as u32), - white_ref_level: d.u(tag::WHITE_REF_LEVEL).map(|v| v as u32), + black_ref_level: d.u(if rgba { tag::COMPONENT_MIN_REF } else { tag::BLACK_REF_LEVEL }).map(|v| v as u32), + white_ref_level: d.u(if rgba { tag::COMPONENT_MAX_REF } else { tag::WHITE_REF_LEVEL }).map(|v| v as u32), alpha_depth: u(tag::ALPHA_SAMPLE_DEPTH), - rgba: d.kind() == Some(st::RGBA_DESCRIPTOR), + rgba, picture_coding: d.ul(tag::PICTURE_ESSENCE_CODING), transfer_characteristic: d.ul(tag::TRANSFER_CHARACTERISTIC), coding_equations: d.ul(tag::CODING_EQUATIONS), @@ -910,7 +934,7 @@ fn build_track( } // A wrong or missing label: trust the essence bytes. if let Some(first) = elements.first() { - let head = read_upto(src, first.value_offset, 64)?; + let head = read_upto(src, first.value_offset, 4096)?; if let Some(sniffed) = sniff_picture(&head) { let agree = std::mem::discriminant(&sniffed) == std::mem::discriminant(&codec); if !agree && !matches!(codec, Codec::Avc { intra: true }) { @@ -918,6 +942,12 @@ fn build_track( } } } + // CinemaDNG §7.1.1.2: a "TIFF" picture element type names DNG pictures + if codec == Codec::Unknown + && elements.first().is_some_and(|e| e.key.0[12] == 0x15 && e.key.0[14] == super::essence::CINEMADNG_ELEMENT_TYPE) + { + codec = Codec::CinemaDng; + } t.codec = codec; t.picture = Some(pi); build_pictures(src, &mut t, elements, index, warnings)?; @@ -1337,3 +1367,24 @@ mod fallback_tracks_tests { assert_eq!(tracks.len(), 1); } } + +#[cfg(test)] +mod rgba_layout_tests { + use super::rgba_component_depth; + + #[test] + fn only_complete_uniform_rgb_layouts_have_one_depth() { + let mut layout = [0; 16]; + layout[..8].copy_from_slice(&[b'B', 16, b'G', 16, b'R', 16, b'A', 8]); + assert_eq!(rgba_component_depth(&layout), 16); + assert_eq!(rgba_component_depth(&layout[..15]), 0); + layout[3] = 8; + assert_eq!(rgba_component_depth(&layout), 0); + layout[3] = 16; + layout[4] = b'G'; + assert_eq!(rgba_component_depth(&layout), 0); + layout[4] = 0; + assert_eq!(rgba_component_depth(&layout), 0); + assert_eq!(rgba_component_depth(&[0; 16]), 0); + } +} diff --git a/crates/media/src/mxf/meta.rs b/crates/media/src/mxf/meta.rs index d02afdd71..2b649d099 100644 --- a/crates/media/src/mxf/meta.rs +++ b/crates/media/src/mxf/meta.rs @@ -204,6 +204,8 @@ pub mod tag { pub const WHITE_REF_LEVEL: u16 = 0x3305; pub const ALPHA_SAMPLE_DEPTH: u16 = 0x3309; pub const PIXEL_LAYOUT: u16 = 0x3401; + pub const COMPONENT_MAX_REF: u16 = 0x3406; + pub const COMPONENT_MIN_REF: u16 = 0x3407; pub const AUDIO_SAMPLING_RATE: u16 = 0x3D03; pub const LOCKED: u16 = 0x3D02; pub const CHANNEL_COUNT: u16 = 0x3D07; diff --git a/crates/media/src/mxf/mod.rs b/crates/media/src/mxf/mod.rs index ca0e8f918..56de637c7 100644 --- a/crates/media/src/mxf/mod.rs +++ b/crates/media/src/mxf/mod.rs @@ -2,7 +2,8 @@ //! ST 377-1 (file format: KLV, partitions, primer pack, header metadata, index tables, random //! index pack), ST 378 (OP1a), ST 390 (OP-Atom), ST 379-1/2 (generic container), ST 381-1 (MPEG //! video mapping), ST 381-3 (AVC mapping), ST 382 (AES3 / Broadcast Wave audio mapping), ST 331 -//! (AES3 element data, D-10 sound), ST 2019-4 (VC-3 mapping) and RDD 44 (ProRes mapping). +//! (AES3 element data, D-10 sound), ST 2019-4 (VC-3 mapping), RDD 44 (ProRes mapping) and the CinemaDNG +//! specification's MXF mapping (frame-wrapped DNG pictures). //! //! [`open`] walks the file's KLV packets (keys and lengths only; essence payloads are never read //! while opening), parses the partition packs, the primer pack and header metadata of the most @@ -32,7 +33,7 @@ mod meta; pub mod write; pub use crate::source::{ByteSource, ReadSeekSource}; -pub use essence::{Codec, PictureInfo, SoundFormat, SoundInfo, aes3_element_samples, decode_pcm}; +pub use essence::{CINEMADNG_CONTAINER_PREFIX, CINEMADNG_ELEMENT_TYPE, Codec, PictureInfo, SoundFormat, SoundInfo, aes3_element_samples, decode_pcm}; pub use file::{Chunk, EssenceTrack, MxfFile, OperationalPattern, Partition, PartitionKind, Sample, Timecode, TrackKind, Wrapping, open}; pub use index::{IndexEntry, IndexSegment}; pub use klv::{Rational, Ul}; diff --git a/crates/media/src/mxf/write.rs b/crates/media/src/mxf/write.rs index 9fb6bfea2..42abf846f 100644 --- a/crates/media/src/mxf/write.rs +++ b/crates/media/src/mxf/write.rs @@ -107,6 +107,10 @@ pub enum PictureCoding { ProRes { profile: u8 }, /// H.264 / AVC Annex B byte stream; `profile_idc` from the SPS (66, 77, 100, …). Avc { profile_idc: u8, intra: bool }, + /// CinemaDNG: one DNG dataset per picture (CinemaDNG specification §7), frame-wrapped only, + /// with an RGBA descriptor of the decoded 16-bit RGB image. `linear_raw`: the pictures are + /// LinearRaw (PhotometricInterpretation 34892) rather than CFA (32803). + CinemaDng { linear_raw: bool }, } impl PictureCoding { @@ -116,9 +120,10 @@ impl PictureCoding { _ => true, } } - /// Picture essence coding label. - fn coding_label(&self) -> [u8; 16] { - match *self { + /// Picture essence coding label (None: CinemaDNG, which has no registered one, §7.2.2.2). + fn coding_label(&self) -> Option<[u8; 16]> { + Some(match *self { + PictureCoding::CinemaDng { .. } => return None, PictureCoding::Vc3 { cid } => { let code = if (1235..=1234 + 0x7F).contains(&cid) { (cid - 1234) as u8 } else { 0 }; [0x06, 0x0E, 0x2B, 0x34, 0x04, 0x01, 0x01, 0x0A, 0x04, 0x01, 0x02, 0x02, 0x71, code, 0x00, 0x00] @@ -145,7 +150,7 @@ impl PictureCoding { }; [0x06, 0x0E, 0x2B, 0x34, 0x04, 0x01, 0x01, 0x0A, 0x04, 0x01, 0x02, 0x02, 0x01, b13, b14, 0x01] } - } + }) } /// (essence container label, element type) for frame or clip wrapping. fn container(&self, clip: bool) -> ([u8; 16], u8) { @@ -154,6 +159,12 @@ impl PictureCoding { PictureCoding::Vc3 { .. } => (essence_container(0x0A, 0x11, w, 0x00), if clip { 0x0D } else { 0x0C }), PictureCoding::ProRes { .. } => (essence_container(0x0D, 0x1C, w, 0x00), if clip { 0x18 } else { 0x17 }), PictureCoding::Avc { .. } => (essence_container(0x0A, 0x10, 0x60, w), if clip { 0x16 } else { 0x15 }), + // CinemaDNG §7.2.1 (provisional label) and §7.1.1.2 (element type 42h); always frame + // wrapped (`MxfWriter::new` refuses clip wrapping) + PictureCoding::CinemaDng { linear_raw } => ( + [0x06, 0x0E, 0x2B, 0x34, 0x04, 0x01, 0x01, 0x01, 0x0F, 0x0A, 0x0D, 0x0B, 0x02, 0x7F, if *linear_raw { 0x02 } else { 0x01 }, 0x00], + 0x42, + ), } } } @@ -202,6 +213,8 @@ impl PictureDesc { if matches!(cid, 1235 | 1241 | 1250 | 1256 | 1270 | 1271) { 10 } else { 8 }, if matches!(cid, 1256 | 1270) { (1, 1) } else { (2, 1) }, ), + // §7.2.2.3: the decoded image is described as 16 bits per component + PictureCoding::CinemaDng { .. } => (16, (1, 1)), }; PictureDesc { coding, @@ -589,6 +602,19 @@ impl MxfWriter { return Err(io::Error::new(io::ErrorKind::InvalidInput, "invalid PCM sound description")); } let clip = cfg.pattern == Pattern::OpAtom; + if clip && matches!(cfg.picture.as_ref().map(|p| p.coding), Some(PictureCoding::CinemaDng { .. })) { + // CinemaDNG §7.1.1: a TIFF dataset has no length field, so DNG pictures can't be clip wrapped + return Err(io::Error::new( + io::ErrorKind::InvalidInput, + "CinemaDNG pictures are frame wrapped only (this writer's OP-Atom is clip wrapped)", + )); + } + if let Some(p) = &cfg.picture { + // CinemaDNG §7.2.2.3 fixes the decoded image at 16 bits per component + if matches!(p.coding, PictureCoding::CinemaDng { .. }) && p.depth != 16 { + return Err(io::Error::new(io::ErrorKind::InvalidInput, "CinemaDNG pictures are described at 16 bits per component")); + } + } let mut tracks = Vec::new(); let mut id = cfg.first_track_id.max(2); if let Some(p) = &cfg.picture { @@ -985,7 +1011,8 @@ impl MxfWriter { TrackKind::Picture => { // Picture tracks exist only with a picture configuration. let Some(p) = cfg.picture.as_ref() else { continue }; - let mut s = Set::new(0x28, uid); + let rgba = matches!(p.coding, PictureCoding::CinemaDng { .. }); + let mut s = Set::new(if rgba { 0x29 } else { 0x28 }, uid); if multiple { s.u32(0x3006, t.id); } @@ -1020,8 +1047,21 @@ impl MxfWriter { // field dominance: the temporally first field (1 top, 2 bottom) s.u8(0x3212, if p.bottom_field_first { 2 } else { 1 }); } - s.p(0x3201, &p.coding.coding_label()) - .p(0x3210, &tcl) + if let Some(label) = p.coding.coding_label() { + s.p(0x3201, &label); + } + if rgba { + // RGBA descriptor (ST 377-1 G.2.40): pixel layout R, G, B of 16 bits each and + // the component reference levels. CinemaDNG §7.2.2.3 fixes the decoded + // image at 16 bits; `MxfWriter::new` rejects any other `PictureDesc::depth`. + let mut layout = [0u8; 16]; + layout[..6].copy_from_slice(&[b'R', 16, b'G', 16, b'B', 16]); + let max = u32::from(u16::MAX); + s.p(0x3401, &layout).u32(0x3406, max).u32(0x3407, 0); + s.write(&mut out); + continue; + } + s.p(0x3210, &tcl) .p(0x321A, &ce) .p(0x3219, &cp) .u32(0x3301, depth) diff --git a/crates/media/src/mxf/write/labels.rs b/crates/media/src/mxf/write/labels.rs index eb8fbc040..12f62c7fd 100644 --- a/crates/media/src/mxf/write/labels.rs +++ b/crates/media/src/mxf/write/labels.rs @@ -120,6 +120,10 @@ pub const PRIMER: &[(u16, [u8; 16])] = &[ (0x3304, ul(&[0x01, 0x01, 0x01, 0x01, 0x04, 0x01, 0x05, 0x03, 0x03])), (0x3305, ul(&[0x01, 0x01, 0x01, 0x01, 0x04, 0x01, 0x05, 0x03, 0x04])), (0x3306, ul(&[0x01, 0x01, 0x01, 0x02, 0x04, 0x01, 0x05, 0x03, 0x05])), + // RGBA picture essence descriptor (ST 377-1; RP 210): PixelLayout, ComponentMaxRef, ComponentMinRef + (0x3401, ul(&[0x01, 0x01, 0x01, 0x02, 0x04, 0x01, 0x05, 0x03, 0x06])), + (0x3406, ul(&[0x01, 0x01, 0x01, 0x05, 0x04, 0x01, 0x05, 0x03, 0x0B])), + (0x3407, ul(&[0x01, 0x01, 0x01, 0x05, 0x04, 0x01, 0x05, 0x03, 0x0C])), (0x3D03, ul(&[0x01, 0x01, 0x01, 0x05, 0x04, 0x02, 0x03, 0x01, 0x01, 0x01])), (0x3D02, ul(&[0x01, 0x01, 0x01, 0x04, 0x04, 0x02, 0x03, 0x01, 0x04])), (0x3D07, ul(&[0x01, 0x01, 0x01, 0x05, 0x04, 0x02, 0x01, 0x01, 0x04])), diff --git a/crates/media/src/mxf/write/tests.rs b/crates/media/src/mxf/write/tests.rs index 91574fec7..bc60423e9 100644 --- a/crates/media/src/mxf/write/tests.rs +++ b/crates/media/src/mxf/write/tests.rs @@ -401,6 +401,60 @@ fn long_pcm_reads_backed_by_the_file_are_not_capped() { assert!(m.read_pcm(&f, 0, 0, frames + (1 << 24)).is_err()); } +/// A TIFF-headed stand-in for a DNG dataset (the demuxer only carries the bytes). +fn dng_like(i: usize) -> Vec { + let mut v = b"II*\0\x08\0\0\0".to_vec(); + v.extend(std::iter::repeat_n(i as u8 ^ 0x33, 300 + 11 * i)); + v +} + +#[test] +fn cinemadng_op1a_reads_back() { + for linear_raw in [false, true] { + let mut cfg = op1a_config(Rational::new(24, 1), PictureCoding::CinemaDng { linear_raw }); + cfg.sound.truncate(1); + let mut w = MxfWriter::new(Cursor::new(Vec::new()), cfg).unwrap(); + for i in 0..6 { + w.push_picture(&dng_like(i), FrameInfo::intra()).unwrap(); + w.push_sound(0, &pcm16(i * 2000, 2000)).unwrap(); + } + let f = w.finish().unwrap().into_inner(); + // CinemaDNG §7.1.1.2: picture element key 06.0E.2B.34.01.02.01.01.0D.01.03.01.15.01.42.01 + let key = [0x06, 0x0E, 0x2B, 0x34, 0x01, 0x02, 0x01, 0x01, 0x0D, 0x01, 0x03, 0x01, 0x15, 0x01, 0x42, 0x01]; + assert_eq!(f.windows(16).filter(|k| *k == key).count(), 6); + // §7.2.1: the essence container label of the content kind + let ec = [0x06, 0x0E, 0x2B, 0x34, 0x04, 0x01, 0x01, 0x01, 0x0F, 0x0A, 0x0D, 0x0B, 0x02, 0x7F, if linear_raw { 2 } else { 1 }, 0x00]; + assert!(f.windows(16).any(|k| k == ec)); + let m = open(&f).unwrap(); + assert!(m.warnings.is_empty(), "{:?}", m.warnings); + let v = m.track_of_kind(TrackKind::Picture).unwrap(); + let t = &m.tracks[v]; + assert_eq!(t.codec, Codec::CinemaDng); + assert_eq!(t.wrapping, Wrapping::Frame); + assert_eq!(t.samples.len(), 6); + let p = t.picture.as_ref().unwrap(); + assert!(p.rgba, "RGBA picture essence descriptor (§7.2.2)"); + assert_eq!(p.component_depth, 16); + assert_eq!(p.black_ref_level, Some(0)); + assert_eq!(p.white_ref_level, Some(65_535)); + assert!(p.full_range()); + assert_eq!((p.frame_width(), p.frame_height()), (64, 36)); + assert!(p.picture_coding.is_none()); + for i in 0..6 { + assert!(t.samples[i].key); + assert_eq!(m.read_sample(&f, v, i).unwrap(), dng_like(i)); + } + assert_eq!(m.timecode.unwrap().format(), "01:02:30:00"); + } +} + +#[test] +fn cinemadng_is_never_clip_wrapped() { + let mut cfg = WriterConfig::new(Pattern::OpAtom, Rational::new(24, 1), PackageIds::from_seed("t", "T")); + cfg.picture = Some(PictureDesc::new(PictureCoding::CinemaDng { linear_raw: false }, 64, 36)); + assert!(MxfWriter::new(Cursor::new(Vec::new()), cfg).is_err()); +} + /// One `push_sound` of a long clip must not cost more than the clip itself: draining the buffered /// PCM from the front for every edit unit made this quadratic (about 16 s for 10 minutes). #[test] diff --git a/crates/media/src/ogg/mod.rs b/crates/media/src/ogg/mod.rs index 387c3fc71..d487f9224 100644 --- a/crates/media/src/ogg/mod.rs +++ b/crates/media/src/ogg/mod.rs @@ -1,4 +1,4 @@ -//! Clean-room Ogg demuxer, written from RFC 3533 (the Ogg encapsulation format), RFC 7845 (Ogg +//! Clean-room Ogg demuxer and page writer, written from RFC 3533 (the Ogg encapsulation format), RFC 7845 (Ogg //! Opus) and the Vorbis I specification's Ogg mapping (section A). //! //! [`open`] reads every page (capture pattern, header, lacing values, CRC-32 checked; damaged @@ -12,6 +12,9 @@ //! TOC-derived durations, the pre-skip, and end trimming of the final page, so a player can seek //! to any 48 kHz sample. //! +//! [`PageWriter`] writes the pages of a logical bitstream, and [`opus_tags`] and +//! [`vorbis_comment`] build the comment headers of Ogg Opus and Ogg Vorbis. +//! //! Layer L0: no dependencies beyond `std`, no `unsafe`, builds for `wasm32-unknown-unknown`. //! //! From FilmCraft's `filmcraft-ogg` (storytold/filmcraft@8fda2f2). Public items keep their names, so @@ -25,6 +28,9 @@ use std::fmt; use std::io; pub use crate::source::{ByteSource, ReadSeekSource}; +pub use write::{PAGE_TARGET_BYTES, PAGE_TARGET_PACKETS, PageWriter, opus_tags, vorbis_comment}; + +mod write; /// Errors from reading an Ogg file. #[non_exhaustive] @@ -879,3 +885,5 @@ impl VorbisInfo { #[cfg(test)] mod tests; +#[cfg(test)] +mod write_tests; diff --git a/crates/media/src/ogg/write.rs b/crates/media/src/ogg/write.rs new file mode 100644 index 000000000..ff631c650 --- /dev/null +++ b/crates/media/src/ogg/write.rs @@ -0,0 +1,206 @@ +//! Ogg page writer (RFC 3533) and the comment-header packets of Ogg Opus (RFC 7845 §5.2) and Ogg +//! Vorbis (Vorbis I §5). +//! +//! [`PageWriter`] laces packets into pages of one logical bitstream: lacing values of 255 with a +//! shorter last value per packet, at most 255 per page, continued packets flagged on the next +//! page, the granule position of the last packet completed on each page (-1 on a page where no +//! packet completes), page sequence numbers, the CRC-32, and the beginning- and end-of-stream +//! flags. Header packets are flushed to pages of their own, as both mappings require. +//! +//! From FilmCraft's `filmcraft-oggmux` by @JordyUSA (storytold/filmcraft#4481). + +use super::crc32; + +/// Default size at which an audio page is closed before the next packet is added. +pub const PAGE_TARGET_BYTES: usize = 4096; +/// Default number of packets after which an audio page is closed (one second of 20 ms Opus). +pub const PAGE_TARGET_PACKETS: usize = 50; + +/// Writes the pages of one logical bitstream. +#[derive(Clone, Debug)] +pub struct PageWriter { + serial: u32, + seq: u32, + lacing: Vec, + body: Vec, + /// Granule position of the last packet completed on the pending page. + granule: Option, + /// The pending page starts with the continuation of a packet. + continued: bool, + packets: usize, + bos_written: bool, + finished: bool, + target_bytes: usize, + target_packets: usize, +} + +impl PageWriter { + /// A writer for the logical bitstream with serial number `serial`, closing audio pages at + /// [`PAGE_TARGET_BYTES`] or [`PAGE_TARGET_PACKETS`]. + pub fn new(serial: u32) -> PageWriter { + PageWriter { + serial, + seq: 0, + lacing: Vec::with_capacity(255), + body: Vec::new(), + granule: None, + continued: false, + packets: 0, + bos_written: false, + finished: false, + target_bytes: PAGE_TARGET_BYTES, + target_packets: PAGE_TARGET_PACKETS, + } + } + + /// Closes audio pages at `bytes` of payload or `packets` packets, whichever comes first. + pub fn with_page_target(mut self, bytes: usize, packets: usize) -> PageWriter { + self.target_bytes = bytes.max(1); + self.target_packets = packets.max(1); + self + } + + /// Pages written so far. + pub fn pages(&self) -> u32 { + self.seq + } + + /// A header packet on pages of its own (granule position 0); the first one opens the stream. + pub fn header(&mut self, data: &[u8]) -> Vec { + let mut out = self.flush(); + out.extend(self.add(data, 0)); + out.extend(self.flush()); + out + } + + /// Adds an audio packet that ends at granule position `granule`; returns the pages it + /// completed. The pending page is closed first once it is full enough, so the last packet + /// always stays pending for [`Self::finish`] to mark the end of the stream. + pub fn packet(&mut self, data: &[u8], granule: i64) -> Vec { + let mut out = Vec::new(); + if self.body.len() >= self.target_bytes || self.packets >= self.target_packets { + out = self.flush(); + } + out.extend(self.add(data, granule)); + out + } + + /// Closes the pending page now, if there is one. + pub fn flush(&mut self) -> Vec { + let mut out = Vec::new(); + if !self.lacing.is_empty() { + self.emit(&mut out, false); + } + out + } + + /// Ends the stream: the pending page (or an empty page when nothing is pending) with the + /// end-of-stream flag. `last_granule` replaces the granule position of the final page (Ogg + /// Opus end trimming) when a packet is pending; `None` keeps the last packet's. End trimming + /// applies only to packets completed on the last page (RFC 7845 §4.4), so do not + /// [`Self::flush`] the last packet before finishing. + pub fn finish(&mut self, last_granule: Option) -> Vec { + let mut out = Vec::new(); + if self.finished { + return out; + } + if let Some(g) = last_granule + && !self.lacing.is_empty() + { + self.granule = Some(g); + } + self.emit(&mut out, true); + self.finished = true; + out + } + + fn add(&mut self, data: &[u8], granule: i64) -> Vec { + let mut out = Vec::new(); + let mut rest = data; + let mut started = false; + loop { + if self.lacing.len() == 255 { + // A full page: the next one continues this packet, unless it has not begun yet. + self.emit(&mut out, false); + self.continued = started; + } + let n = rest.len().min(255); + self.lacing.push(n as u8); + started = true; + self.body.extend_from_slice(&rest[..n]); + rest = &rest[n..]; + if n < 255 { + break; + } + } + self.granule = Some(granule); + self.packets += 1; + out + } + + fn emit(&mut self, out: &mut Vec, eos: bool) { + let mut flags = 0u8; + if self.continued { + flags |= 0x01; + } + if !self.bos_written { + flags |= 0x02; + self.bos_written = true; + } + if eos { + flags |= 0x04; + } + let start = out.len(); + out.extend_from_slice(b"OggS"); + out.push(0); + out.push(flags); + out.extend_from_slice(&self.granule.unwrap_or(-1).to_le_bytes()); + out.extend_from_slice(&self.serial.to_le_bytes()); + out.extend_from_slice(&self.seq.to_le_bytes()); + out.extend_from_slice(&[0; 4]); + out.push(self.lacing.len() as u8); + out.extend_from_slice(&self.lacing); + out.extend_from_slice(&self.body); + let crc = crc32(&out[start..]); + out[start + 22..start + 26].copy_from_slice(&crc.to_le_bytes()); + self.seq = self.seq.wrapping_add(1); + self.lacing.clear(); + self.body.clear(); + self.granule = None; + self.continued = false; + self.packets = 0; + } +} + +/// A comment header body: vendor string and `KEY=value` comments, little-endian lengths. +fn comments(magic: &[u8], vendor: &str, tags: &[(&str, &str)]) -> Vec { + let mut v = magic.to_vec(); + let len = |n: usize| u32::try_from(n).unwrap_or(u32::MAX).to_le_bytes(); + v.extend(len(vendor.len())); + v.extend_from_slice(vendor.as_bytes()); + let tags: Vec = tags.iter().filter(|(k, _)| valid_key(k)).map(|(k, val)| format!("{k}={val}")).collect(); + v.extend(len(tags.len())); + for t in &tags { + v.extend(len(t.len())); + v.extend_from_slice(t.as_bytes()); + } + v +} + +/// A comment field name: printable ASCII 0x20–0x7D other than `=` (Vorbis I §5.2.1), not empty. +fn valid_key(k: &str) -> bool { + !k.is_empty() && k.bytes().all(|b| (0x20..=0x7D).contains(&b) && b != b'=') +} + +/// The `OpusTags` packet (RFC 7845 §5.2). Comments whose name is not a valid field name are +/// left out. +pub fn opus_tags(vendor: &str, tags: &[(&str, &str)]) -> Vec { + comments(b"OpusTags", vendor, tags) +} + +/// The Vorbis comment header packet (Vorbis I §5: type 3, `vorbis`, comments, framing bit). +pub fn vorbis_comment(vendor: &str, tags: &[(&str, &str)]) -> Vec { + let mut v = comments(b"\x03vorbis", vendor, tags); + v.push(1); + v +} diff --git a/crates/media/src/ogg/write_tests.rs b/crates/media/src/ogg/write_tests.rs new file mode 100644 index 000000000..eebf5732e --- /dev/null +++ b/crates/media/src/ogg/write_tests.rs @@ -0,0 +1,200 @@ +//! The page writer, read back by [`open`]. + +#![allow(clippy::unwrap_used)] + +use super::{Codec, OpusTiming, PageWriter, crc32, open, opus_tags, vorbis_comment}; + +/// A TOC-only 20 ms CELT packet of `len` bytes whose bytes encode `i` (so packets differ). +fn opus_packet(i: usize, len: usize) -> Vec { + let mut p = vec![0xF8u8]; + p.extend((0..len.saturating_sub(1)).map(|k| (i * 31 + k) as u8)); + p +} + +fn opus_head(pre_skip: u16) -> Vec { + let mut h = b"OpusHead".to_vec(); + h.extend([1, 2]); + h.extend(pre_skip.to_le_bytes()); + h.extend(48_000u32.to_le_bytes()); + h.extend([0, 0, 0]); + h +} + +/// An Ogg Opus file of `n` packets of `sizes(i)` bytes, `samples` playable samples after the pre-skip. +fn write_opus(n: usize, sizes: impl Fn(usize) -> usize, pre_skip: u16, samples: i64) -> (Vec, Vec>) { + let mut w = PageWriter::new(0x1234_5678); + let mut f = w.header(&opus_head(pre_skip)); + f.extend(w.header(&opus_tags("FilmCraft", &[("TITLE", "Test"), ("bad=key", "dropped"), ("", "dropped")]))); + let mut packets = Vec::new(); + for i in 0..n { + let p = opus_packet(i, sizes(i)); + f.extend(w.packet(&p, (i as i64 + 1) * 960)); + packets.push(p); + } + f.extend(w.finish(Some(i64::from(pre_skip) + samples))); + (f, packets) +} + +#[test] +fn written_pages_read_back_packet_for_packet() { + let (f, packets) = write_opus(230, |i| 3 + (i * 37) % 400, 120, 230 * 960 - 120 - 500); + let ogg = open(&f).unwrap(); + assert!(ogg.warnings.is_empty(), "{:?}", ogg.warnings); + assert_eq!(ogg.streams.len(), 1); + let s = &ogg.streams[0]; + assert_eq!((s.serial, s.codec, s.saw_eos), (0x1234_5678, Codec::Opus, true)); + assert_eq!(s.headers.len(), 2); + assert_eq!(s.headers[0], opus_head(120)); + assert_eq!(s.headers[1], opus_tags("FilmCraft", &[("TITLE", "Test")])); + assert_eq!(s.packets.len(), packets.len()); + for (k, p) in packets.iter().enumerate() { + assert_eq!(&ogg.read_packet(&f, 0, k).unwrap(), p, "packet {k}"); + } + assert!(s.packets.last().unwrap().eos); + assert_eq!(s.last_granule, Some(120 + 230 * 960 - 120 - 500)); + // RFC 7845 §4: the end trimming of the final page. + let t = OpusTiming::of(s, 120); + assert_eq!(t.total, 230 * 960 - 120 - 500); + assert_eq!(t.starts[0], -120); + assert!(t.durations.iter().all(|&d| d == 960)); +} + +#[test] +fn headers_have_pages_of_their_own_and_only_the_first_page_opens_the_stream() { + let (f, _) = write_opus(3, |_| 10, 0, 3 * 960); + let mut pos = 0; + let mut flags = Vec::new(); + let mut seqs = Vec::new(); + while pos + 27 <= f.len() { + assert_eq!(&f[pos..pos + 4], b"OggS"); + let mut page = f[pos..].to_vec(); + let nseg = page[26] as usize; + let len = 27 + nseg + page[27..27 + nseg].iter().map(|&l| l as usize).sum::(); + page.truncate(len); + let stored = u32::from_le_bytes(page[22..26].try_into().unwrap()); + page[22..26].fill(0); + assert_eq!(crc32(&page), stored, "page CRC"); + flags.push(f[pos + 5]); + seqs.push(u32::from_le_bytes(f[pos + 18..pos + 22].try_into().unwrap())); + pos += len; + } + assert_eq!(pos, f.len()); + assert_eq!(flags, [0x02, 0x00, 0x04], "BOS head page, tags page, one audio page with EOS"); + assert_eq!(seqs, [0, 1, 2]); +} + +#[test] +fn a_packet_longer_than_a_page_continues() { + let big = 2 * 255 * 255 + 7000; // spans three pages: the middle one completes no packet + let (f, packets) = write_opus(4, |i| if i == 1 { big } else { 50 }, 0, 4 * 960); + let ogg = open(&f).unwrap(); + assert!(ogg.warnings.is_empty()); + for (k, p) in packets.iter().enumerate() { + assert_eq!(&ogg.read_packet(&f, 0, k).unwrap(), p, "packet {k}"); + } + // a page on which no packet completes has granule position -1 + let mut pos = 0; + let mut saw_none = false; + while pos + 27 <= f.len() { + let nseg = f[pos + 26] as usize; + let lacing = &f[pos + 27..pos + 27 + nseg]; + let g = i64::from_le_bytes(f[pos + 6..pos + 14].try_into().unwrap()); + if lacing.iter().all(|&l| l == 255) { + assert_eq!(g, -1); + saw_none = true; + } + pos += 27 + nseg + lacing.iter().map(|&l| l as usize).sum::(); + } + assert!(saw_none); +} + +#[test] +fn page_targets_bound_page_size_and_finish_marks_the_end_once() { + let mut w = PageWriter::new(7).with_page_target(1000, 10); + let mut f = w.header(&opus_head(0)); + f.extend(w.header(&opus_tags("v", &[]))); + for i in 0..95 { + f.extend(w.packet(&opus_packet(i, 160), (i as i64 + 1) * 960)); + } + f.extend(w.finish(None)); + assert!(w.finish(None).is_empty()); + let ogg = open(&f).unwrap(); + let s = &ogg.streams[0]; + assert_eq!(s.packets.len(), 95); + assert_eq!(s.pages, 2 + 95_u32.div_ceil(7), "1000-byte pages hold 7 packets of 160 bytes"); + assert_eq!(s.last_granule, Some(95 * 960)); + // An empty stream still ends. + let mut w = PageWriter::new(1); + let mut f = w.header(&opus_head(0)); + f.extend(w.finish(Some(0))); + assert!(open(&f).unwrap().streams[0].saw_eos); + // A flushed last packet cannot be trimmed: the empty end page keeps its granule. + let mut w = PageWriter::new(1); + let mut f = w.header(&opus_head(312)); + f.extend(w.header(&opus_tags("t", &[]))); + f.extend(w.packet(&[0xFC; 40], 960)); + f.extend(w.flush()); + f.extend(w.finish(Some(500))); + let s = &open(&f).unwrap().streams[0]; + assert!(s.saw_eos); + assert_eq!(s.last_granule, Some(960)); +} + +#[test] +fn vorbis_comment_has_the_framing_bit() { + let c = vorbis_comment("FilmCraft", &[("ARTIST", "x")]); + assert_eq!(&c[..7], b"\x03vorbis"); + assert_eq!(*c.last().unwrap(), 1); + assert_eq!(u32::from_le_bytes(c[7..11].try_into().unwrap()), 9); +} + +/// Packets on and around the lacing boundaries (multiples of 255, where a packet ends with a +/// zero lacing value) and an empty packet read back exactly. +#[test] +fn lacing_boundaries_read_back() { + let sizes = [0usize, 1, 254, 255, 256, 509, 510, 511, 765, 65_024, 65_025, 65_026]; + let mut w = PageWriter::new(3); + let mut f = w.header(&opus_head(0)); + f.extend(w.header(&opus_tags("v", &[]))); + let packets: Vec> = sizes.iter().enumerate().map(|(i, &n)| (0..n).map(|k| (i * 7 + k) as u8).collect()).collect(); + for (i, p) in packets.iter().enumerate() { + f.extend(w.packet(p, (i as i64 + 1) * 960)); + } + f.extend(w.finish(None)); + let ogg = open(&f).unwrap(); + assert!(ogg.warnings.is_empty(), "{:?}", ogg.warnings); + assert_eq!(ogg.streams[0].packets.len(), packets.len()); + for (k, p) in packets.iter().enumerate() { + assert_eq!(ogg.read_packet(&f, 0, k).unwrap(), *p, "packet {k} of {} bytes", p.len()); + } +} + +/// A page exactly full of lacing values when a packet starts: the packet begins on the next page, +/// which must not be flagged as continuing one (the reader would drop its first packet). Review +/// of #4481; reachable with page targets larger than the defaults. +#[test] +fn a_packet_after_an_exactly_full_page_is_not_a_continuation() { + let mut w = PageWriter::new(5).with_page_target(1 << 20, 1000); + let mut f = w.header(&opus_head(0)); + f.extend(w.header(&opus_tags("v", &[]))); + let packets: Vec> = (0..300).map(|i| opus_packet(i, 1)).collect(); + for (i, p) in packets.iter().enumerate() { + f.extend(w.packet(p, (i as i64 + 1) * 960)); + } + f.extend(w.finish(None)); + let ogg = open(&f).unwrap(); + assert!(ogg.warnings.is_empty(), "{:?}", ogg.warnings); + assert_eq!(ogg.streams[0].packets.len(), 300, "every packet is read back"); + for (k, p) in packets.iter().enumerate() { + assert_eq!(&ogg.read_packet(&f, 0, k).unwrap(), p, "packet {k}"); + } + // the second audio page starts with packet 255 and is not a continuation + let mut pos = 0; + let mut flags = Vec::new(); + while pos + 27 <= f.len() { + let nseg = f[pos + 26] as usize; + flags.push(f[pos + 5]); + pos += 27 + nseg + f[pos + 27..pos + 27 + nseg].iter().map(|&l| l as usize).sum::(); + } + assert!(flags.iter().all(|&fl| fl & 0x01 == 0), "no page continues a packet: {flags:?}"); +} diff --git a/crates/mesh/src/fbx.rs b/crates/mesh/src/fbx.rs index acb654513..08ba73640 100644 --- a/crates/mesh/src/fbx.rs +++ b/crates/mesh/src/fbx.rs @@ -638,9 +638,9 @@ fn euler(d: [f64; 3], order: i64) -> Mat4 { } /// A Model's local transform: `T · Roff · Rp · Rpre · R · Rpostâ�»Â¹ · Rpâ�»Â¹ · Soff · Sp · S · Spâ�»Â¹`. -fn model_local(p: &Props) -> Mat4 { +fn model_local(p: &Props, apply_inactive_rotation: bool) -> Mat4 { let order = p.f("RotationOrder").map_or(0, |o| o as i64); - let active = p.f("RotationActive").is_some_and(|a| a != 0.0); + let active = apply_inactive_rotation || p.f("RotationActive").is_some_and(|a| a != 0.0); let (pre, post) = if active { (euler(p.v3("PreRotation", [0.0; 3]), 0), transpose(&euler(p.v3("PostRotation", [0.0; 3]), 0))) } else { (IDENTITY, IDENTITY) }; let rp = p.v3("RotationPivot", [0.0; 3]); @@ -671,6 +671,20 @@ fn geometric(p: &Props) -> Option { Some(chain(&[translate(t), euler(r, 0), scale(s)])) } +/// Optional FBX scene compatibility behavior. +#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)] +pub struct SceneOptions { + /// Apply pre/post rotations even when `RotationActive` is missing or zero. Some importers + /// retain these values regardless of the flag. + pub apply_inactive_rotation: bool, + /// Refuse malformed geometry arrays and accept an unterminated final polygon. + pub strict_geometry: bool, + /// If a model names several parents, use the last connection. + pub prefer_last_parent: bool, + /// Restrict model hierarchy, geometry and material links to object connections. + pub only_object_connections: bool, +} + impl Document { /// The top-level record of this name. pub fn top(&self, name: &str) -> Option<&Record> { @@ -720,6 +734,11 @@ impl Document { /// The scene the records describe. `resolve` is asked for texture image files (see the /// module documentation). pub fn scene(&self, limits: &Limits, resolve: Resolver) -> Result { + self.scene_with_options(limits, resolve, &SceneOptions::default()) + } + + /// Build a scene with explicitly selected compatibility behavior. + pub fn scene_with_options(&self, limits: &Limits, resolve: Resolver, options: &SceneOptions) -> Result { if self.version != 0 && self.version < 6000 { return Err(crate::Error::Unsupported(format!( "FBX {}.{} (only FBX 6.x and 7.x files are read)", @@ -730,10 +749,13 @@ impl Document { let objects = self.top("Objects").ok_or_else(|| malformed(KIND, "no Objects section"))?; let mut b = Build { limits, + strict_geometry: options.strict_geometry, + only_object_connections: options.only_object_connections, budget: Budget { left: Cell::new(self.left), limit: limits.max_data_bytes }, resolve, by_key: objects.children.iter().filter_map(|o| Some((key(o.props.first())?, o))).collect(), children: HashMap::new(), + object_children: HashMap::new(), scene: Scene::default(), material_of: HashMap::new(), texture_of: HashMap::new(), @@ -746,8 +768,14 @@ impl Document { let mut links: Vec<(Key, Key)> = vec![]; for e in self.top("Connections").into_iter().flat_map(|c| c.children.iter()).filter(|e| e.name == "C" || e.name == "Connect") { let (Some(child), Some(parent)) = (key(e.props.get(1)), key(e.props.get(2))) else { continue }; - links.push((child, parent)); + let object = e.str_at(0) == Some("OO"); + if !options.only_object_connections || object { + links.push((child, parent)); + } b.children.entry(parent).or_default().push((child, e.str_at(3))); + if options.only_object_connections && object { + b.object_children.entry(parent).or_default().push((child, e.str_at(3))); + } } let models: Vec<(Key, &Record)> = @@ -761,7 +789,7 @@ impl Document { node_of.insert(*id, b.scene.nodes.len()); b.scene.nodes.push(Node { name: short_name(o.first_str().unwrap_or("")).to_string(), - transform: model_local(&Props::of(o)), + transform: model_local(&Props::of(o), options.apply_inactive_rotation), ..Node::default() }); } @@ -772,7 +800,11 @@ impl Document { if let (Some(ci), Some(pi)) = (node_of.get(c), node_of.get(p)) && ci != pi { - parent.entry(*ci).or_insert(*pi); + if options.prefer_last_parent { + parent.insert(*ci, *pi); + } else { + parent.entry(*ci).or_insert(*pi); + } } } for i in 0..b.scene.nodes.len() { @@ -942,10 +974,13 @@ fn material(name: &str, record: &Record) -> Material { /// A scene under construction. struct Build<'a, 'r> { limits: &'a Limits, + strict_geometry: bool, + only_object_connections: bool, budget: Budget, resolve: Resolver<'r>, by_key: HashMap, &'a Record>, children: HashMap, Vec<(Key<'a>, Option<&'a str>)>>, + object_children: HashMap, Vec<(Key<'a>, Option<&'a str>)>>, scene: Scene, material_of: HashMap, usize>, texture_of: HashMap, Option>, @@ -960,7 +995,12 @@ struct Build<'a, 'r> { impl<'a> Build<'a, '_> { /// The objects of `kind` connected under `id`, with the property each connects to. fn children_of(&self, id: Key<'a>, kind: &str) -> Vec<(Key<'a>, Option<&'a str>)> { - let all = self.children.get(&id).map(Vec::as_slice).unwrap_or_default(); + let source = if self.only_object_connections && matches!(kind, "Geometry" | "Material" | "NodeAttribute") { + &self.object_children + } else { + &self.children + }; + let all = source.get(&id).map(Vec::as_slice).unwrap_or_default(); all.iter().filter(|l| self.by_key.get(&l.0).is_some_and(|o| o.name == kind)).copied().collect() } @@ -1074,6 +1114,32 @@ impl<'a> Build<'a, '_> { fn mesh(&mut self, name: &str, geom: &Record, materials: &[usize]) -> Result> { let limits = self.limits; let (Some(vertices), Some(pvi)) = (geom.child("Vertices"), geom.child("PolygonVertexIndex")) else { return Ok(None) }; + if self.strict_geometry { + let valid = |record: &Record, integers: bool| { + let values = record.child("a").unwrap_or(record); + values.props.iter().all(|p| match p { + Prop::Int(_) | Prop::Ints(_) => true, + Prop::Float(f) => f.is_finite() && (!integers || f.fract() == 0.0), + Prop::Floats(a) => a.iter().all(|f| f.is_finite() && (!integers || f.fract() == 0.0)), + _ => false, + }) + }; + if !valid(vertices, false) || !valid(pvi, true) { + return Err(malformed(KIND, "mesh array contains a non-numeric, non-finite or fractional index value")); + } + for (element, names) in [ + ("LayerElementNormal", &["Normals", "NormalsIndex"][..]), + ("LayerElementUV", &["UV", "UVIndex"][..]), + ("LayerElementMaterial", &["Materials"][..]), + ("LayerElementColor", &["Colors", "ColorIndex"][..]), + ] { + if let Some(layer) = geom.children_named(element).min_by_key(|e| e.props.first().and_then(Prop::as_i64) != Some(0)) + && names.iter().filter_map(|n| layer.child(n)).any(|r| !valid(r, false)) + { + return Err(malformed(KIND, "mesh array contains a non-numeric or non-finite value")); + } + } + } let pool = |what: &str, n: usize| { if n > limits.max_vertices { Err(too_large(KIND, format!("more than {} {what}", limits.max_vertices))) } else { Ok(()) } }; @@ -1091,15 +1157,18 @@ impl<'a> Build<'a, '_> { let mut part_of: HashMap, usize> = HashMap::new(); let mut face: Vec = vec![]; let (mut broken, mut poly) = (false, 0usize); - for (pv, raw) in pvi.ints().into_iter().enumerate() { + let polygon_indices = pvi.ints(); + let index_count = polygon_indices.len(); + for (pv, raw) in polygon_indices.into_iter().enumerate() { // The last corner of a polygon is stored as `-index - 1`. - let (v, last) = if raw < 0 { (!raw, true) } else { (raw, false) }; + let (v, last) = if raw < 0 { (!raw, true) } else { (raw, self.strict_geometry && pv + 1 == index_count) }; match u32::try_from(v).ok().filter(|v| (*v as usize) < positions.len()) { Some(p) => face.push(Corner { position: p, normal: Layer::at(normal_layer, pv, p as usize, poly), uv: Layer::at(uv_layer, pv, p as usize, poly), }), + None if self.strict_geometry => return Err(malformed(KIND, format!("polygon vertex index {raw} out of range"))), None => broken = true, } if face.len() > limits.max_face_corners { diff --git a/crates/mesh/src/obj.rs b/crates/mesh/src/obj.rs index 092a4e0ff..69428eb25 100644 --- a/crates/mesh/src/obj.rs +++ b/crates/mesh/src/obj.rs @@ -167,8 +167,22 @@ struct Object { slots: HashMap<(String, Option, u32), usize>, } +/// Options for callers that need strict parsing or original face-run order. +#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)] +pub struct ReadOptions { + /// Return an error when a vertex or face line cannot be read. + pub strict: bool, + /// Keep separate runs when groups, materials or smoothing change. + pub preserve_runs: bool, + /// Do not resolve or report material libraries; `usemtl` names remain available. + pub ignore_material_libraries: bool, + /// Treat each group declaration as an object boundary, retaining repeated names. + pub groups_as_objects: bool, +} + struct Reader<'a> { limits: &'a Limits, + options: ReadOptions, v: Vec<[f64; 3]>, colors: Vec<[f32; 4]>, vt: Vec<[f32; 2]>, @@ -314,7 +328,7 @@ impl Reader<'_> { let Some(object) = self.objects.last_mut() else { return Ok(None) }; let slot = match self.slot { Some(s) => s, - None => match object.slots.get(&(self.group.clone(), self.material, self.smoothing)) { + None => match (!self.options.preserve_runs).then(|| object.slots.get(&(self.group.clone(), self.material, self.smoothing))).flatten() { Some(s) => *s, None => { if self.parts >= self.limits.max_parts { @@ -331,15 +345,26 @@ impl Reader<'_> { Ok(object.parts.get_mut(slot)) } - fn face(&mut self, rest: &[&str]) -> Result<()> { - if rest.len() > self.limits.max_face_corners { - return Err(too_large(KIND, format!("a face of more than {} corners", self.limits.max_face_corners))); + fn face<'a>(&mut self, rest: impl Iterator) -> Result<()> { + let mut corners = Vec::new(); + let mut invalid = false; + for token in rest { + // Check before collecting tokens or corners: one line can hold millions of them. + if corners.len() >= self.limits.max_face_corners { + return Err(too_large(KIND, format!("a face of more than {} corners", self.limits.max_face_corners))); + } + match self.corner(token) { + Some(corner) => corners.push(corner), + None => { + invalid = true; + corners.push(Corner::at(0)); + } + } } - let corners: Option> = rest.iter().map(|t| self.corner(t)).collect(); - let Some(corners) = corners.filter(|c| c.len() >= 3) else { + if invalid || corners.len() < 3 { self.bad_lines += 1; return Ok(()); - }; + } self.triangles = self.triangles.saturating_add(corners.len() - 2); if self.triangles > self.limits.max_triangles { return Err(too_large(KIND, format!("more than {} triangles", self.limits.max_triangles))); @@ -361,7 +386,7 @@ impl Reader<'_> { let Some(key) = it.next() else { return Ok(()) }; let after = line.get(key.len()..).unwrap_or("").trim(); match key { - "v" => self.vertex(&it.collect::>())?, + "v" => self.vertex(&it.take(6).collect::>())?, "vt" => { self.pool_room(self.vt.len(), "texture coordinates")?; // `vt u` alone is a 1D coordinate; OBJ's v axis points up the image. @@ -375,6 +400,7 @@ impl Reader<'_> { self.vn.push(n.unwrap_or(NAN3)); self.bad_lines += usize::from(n.is_none()); } + "s" if self.options.groups_as_objects => {} "s" => { self.slot = None; self.smoothing = match it.next() { @@ -384,13 +410,17 @@ impl Reader<'_> { }; } "o" => self.start_object(after.to_string())?, + "g" if self.options.groups_as_objects => { + self.group.clear(); + self.start_object(after.to_string())?; + } "g" => { self.slot = None; self.group = after.to_string(); } "usemtl" => self.use_material(after)?, - "mtllib" => self.library(after, resolve)?, - "f" => self.face(&it.collect::>())?, + "mtllib" if !self.options.ignore_material_libraries => self.library(after, resolve)?, + "f" => self.face(it)?, _ => {} } Ok(()) @@ -445,9 +475,19 @@ impl Reader<'_> { /// writes; every name asked, and every image a library names, is listed in /// [`Scene::external`]. pub fn read(bytes: &[u8], limits: &Limits, resolve: Resolver) -> Result { + read_with_options(bytes, limits, resolve, &ReadOptions::default()) +} + +/// Read an OBJ using the selected compatibility behavior. +pub fn read_with_options(bytes: &[u8], limits: &Limits, resolve: Resolver, options: &ReadOptions) -> Result { + // Lossy UTF-8 conversion can allocate, so bound it before copying invalid input. + if std::str::from_utf8(bytes).is_err() && bytes.len() > limits.max_data_bytes { + return Err(too_large(KIND, "invalid UTF-8 text exceeds the memory limit")); + } let text = String::from_utf8_lossy(bytes); let mut r = Reader { limits, + options: *options, v: vec![], colors: vec![], vt: vec![], @@ -467,8 +507,12 @@ pub fn read(bytes: &[u8], limits: &Limits, resolve: Resolver) -> Result { }; // A line ending in `\` continues on the next. let mut joined = String::new(); - for raw in text.lines() { + for (i, raw) in text.lines().enumerate() { + let raw = if i == 0 { raw.trim_start_matches('\u{feff}') } else { raw }; let line = raw.split('#').next().unwrap_or(""); + if line.len() > limits.max_data_bytes || joined.len().checked_add(line.len()).is_none_or(|n| n > limits.max_data_bytes) { + return Err(too_large(KIND, "a continued line exceeds the memory limit")); + } if let Some(head) = line.trim_end().strip_suffix('\\') { joined.push_str(head); joined.push(' '); @@ -483,6 +527,9 @@ pub fn read(bytes: &[u8], limits: &Limits, resolve: Resolver) -> Result { } } r.line(joined.trim(), resolve)?; + if options.strict && r.bad_lines > 0 { + return Err(malformed(KIND, "invalid vertex or face line")); + } r.finish() } diff --git a/crates/mesh/src/tests/fbx.rs b/crates/mesh/src/tests/fbx.rs index a58ddeb19..679cb0124 100644 --- a/crates/mesh/src/tests/fbx.rs +++ b/crates/mesh/src/tests/fbx.rs @@ -627,3 +627,48 @@ fn fbx6_model_holds_its_mesh() { assert_eq!(d.unit_scale(), Some(100.0)); } } + +#[test] +fn optional_inactive_pre_rotation_keeps_legacy_placements() { + let text = "FBXHeaderExtension: { FBXVersion: 7400 } Objects: { Model: 1, \"Model::A\", \"Null\" { Properties70: { P: \"PreRotation\", \"\", \"\", \"\", 0,0,90 } } }"; + let doc = document(text.as_bytes()); + let normal = doc.scene(&Limits::default(), &mut no_external).unwrap(); + let options = crate::fbx::SceneOptions { apply_inactive_rotation: true, ..Default::default() }; + let legacy = doc.scene_with_options(&Limits::default(), &mut no_external, &options).unwrap(); + let p = [1.0, 0.0, 0.0]; + assert_eq!(transform_point(&normal.nodes[0].transform, p), p); + let turned = transform_point(&legacy.nodes[0].transform, p); + assert!((turned[0]).abs() < 1e-12 && (turned[1] - 1.0).abs() < 1e-12); +} + +#[test] +fn strict_geometry_rejects_bad_arrays_and_keeps_final_polygon() { + let options = crate::fbx::SceneOptions { strict_geometry: true, ..Default::default() }; + let read_strict = |s: &str| document(s.as_bytes()).scene_with_options(&Limits::default(), &mut no_external, &options); + let unterminated = ASCII.replace("a: 0,1,2,-4,0,1,-5", "a: 0,1,2,-4,0,1,4"); + assert_eq!(read_strict(&unterminated).unwrap().meshes[0].triangle_count(), 3); + for bad in ["a: 0,1,2,-4,0,1,200", "a: 0,1,2,-4,0,1,1.5", "a: 0,1,2,-4,0,1,NaN"] { + assert!(read_strict(&ASCII.replace("a: 0,1,2,-4,0,1,-5", bad)).is_err()); + } +} + +#[test] +fn scene_option_selects_the_last_parent_connection() { + let text = "FBXHeaderExtension: { FBXVersion: 7400 } Objects: { Model: 1, \"Model::A\", \"Null\" {} Model: 2, \"Model::B\", \"Null\" {} Model: 3, \"Model::Child\", \"Null\" {} } Connections: { C: \"OO\",3,1 C: \"OO\",3,2 }"; + let doc = document(text.as_bytes()); + let options = crate::fbx::SceneOptions { prefer_last_parent: true, ..Default::default() }; + let scene = doc.scene_with_options(&Limits::default(), &mut no_external, &options).unwrap(); + assert!(scene.nodes[1].children.contains(&2)); + let default = doc.scene(&Limits::default(), &mut no_external).unwrap(); + assert!(default.nodes[0].children.contains(&2)); +} + +#[test] +fn object_only_hierarchy_ignores_property_connections() { + let text = "FBXHeaderExtension: { FBXVersion: 7400 } Objects: { Model: 1, \"Model::A\", \"Null\" {} Model: 2, \"Model::B\", \"Null\" {} } Connections: { C: \"OP\",2,1,\"SomeProperty\" }"; + let doc = document(text.as_bytes()); + let options = crate::fbx::SceneOptions { only_object_connections: true, ..Default::default() }; + let scene = doc.scene_with_options(&Limits::default(), &mut no_external, &options).unwrap(); + assert_eq!(scene.roots.len(), 2); + assert_eq!(doc.scene(&Limits::default(), &mut no_external).unwrap().roots.len(), 1); +} diff --git a/crates/mesh/src/tests/obj.rs b/crates/mesh/src/tests/obj.rs index 9c61773cb..ed8a89755 100644 --- a/crates/mesh/src/tests/obj.rs +++ b/crates/mesh/src/tests/obj.rs @@ -341,3 +341,35 @@ fn a_library_named_again_is_read_once() { assert_eq!((s.materials.len(), s.images.len()), (1, 1)); assert_eq!(s.external, ["m.mtl", "red.png"]); } + +#[test] +fn face_tokens_and_continued_lines_are_bounded_before_collection() { + let limits = Limits { max_face_corners: 3, max_data_bytes: 1024, ..Limits::default() }; + let face = format!("v 0 0 0\nf {}\n", "1 ".repeat(100)); + assert!(matches!(read(face.as_bytes(), &limits, &mut no_external), Err(Error::TooLarge(_)))); + let long_line = format!("v {}\\\n{}\n", "1 ".repeat(300), "1 ".repeat(300)); + assert!(matches!(read(long_line.as_bytes(), &limits, &mut no_external), Err(Error::TooLarge(_)))); + let invalid = b"v 0 0 0\nf 1 1 1\n\xff"; + assert!(matches!(read(invalid, &Limits { max_data_bytes: 4, ..Limits::default() }, &mut no_external), Err(Error::TooLarge(_)))); +} + +#[test] +fn compatibility_options_keep_runs_and_skip_material_libraries() { + let input = b"\xef\xbb\xbf v 0 0 0\nv 1 0 0\nv 0 1 0\nmtllib missing.mtl\nusemtl Red\nf 1 2 3\nusemtl Blue\nf 1 2 3\nusemtl Red\nf 1 2 3\n"; + let options = crate::obj::ReadOptions { strict: true, preserve_runs: true, ignore_material_libraries: true, ..Default::default() }; + let scene = crate::obj::read_with_options(input, &Limits::default(), &mut no_external, &options).unwrap(); + let parts = &scene.meshes[0].parts; + assert_eq!(parts.len(), 3); + assert_eq!(parts[0].material, parts[2].material); + assert!(scene.warnings.is_empty()); + assert!(crate::obj::read_with_options(b"v NaN 0 0\nv 1 0 0\nv 0 1 0\nf 1 2 3\n", &Limits::default(), &mut no_external, &options).is_err()); +} + +#[test] +fn group_boundaries_keep_repeated_names_separate() { + let input = b"v 0 0 0\nv 1 0 0\nv 0 1 0\ng Same\nusemtl M\nf 1 2 3\ng Same\nf 1 3 2\n"; + let options = crate::obj::ReadOptions { groups_as_objects: true, ..Default::default() }; + let scene = crate::obj::read_with_options(input, &Limits::default(), &mut no_external, &options).unwrap(); + assert_eq!(scene.meshes.len(), 2); + assert!(scene.meshes.iter().all(|m| m.name == "Same")); +} diff --git a/crates/metadata/src/container.rs b/crates/metadata/src/container.rs index 842020683..65a47ba05 100644 --- a/crates/metadata/src/container.rs +++ b/crates/metadata/src/container.rs @@ -84,22 +84,83 @@ pub fn embedded(bytes: &[u8]) -> Embedded { const XMP_SIG: &[u8] = b"http://ns.adobe.com/xap/1.0/\0"; const XMP_EXT_SIG: &[u8] = b"http://ns.adobe.com/xmp/extension/\0"; +/// JPEG frame, marker and metadata details collected before the first scan. +#[derive(Debug, Default)] +pub struct JpegInspection { + /// Stored frame width. + pub width: u32, + /// Stored frame height. + pub height: u32, + /// Number of frame components. + pub components: u8, + /// Sample precision. + pub precision: u8, + /// Start-of-frame marker byte. + pub sof: u8, + /// Components in the first scan. + pub first_scan_components: u8, + /// Number of APP1 segments. + pub app1: u32, + /// First EXIF TIFF payload. + pub exif: Option>, + /// Offset of the EXIF TIFF payload. + pub exif_offset: usize, + /// First standard XMP packet. + pub xmp: Option, + /// Reassembled ICC profile. + pub icc: Option>, + /// ICC chunks were malformed or inconsistent. + pub icc_invalid: bool, + /// Adobe APP14 transform byte. + pub adobe_transform: Option, + /// MPF TIFF offset and payload. + pub mpf: Option<(usize, Vec)>, + /// ISO gain-map APP2 payload following its identifier. + pub iso_gainmap: Option>, + valid: bool, +} + +impl JpegInspection { + /// Whether metadata identifies this image as a gain map. + pub fn is_gain_map(&self) -> bool { + self.iso_gainmap.as_ref().is_some_and(|p| p.len() > 4) || self.xmp.as_deref().is_some_and(|x| x.contains("hdrgm:GainMapMax")) + } +} + +/// Inspect markers before the first scan, tolerating garbage between segments and truncated +/// final payloads. Returns `None` for a missing SOI or an invalid segment length. +pub fn inspect_jpeg(b: &[u8]) -> Option { + if !b.starts_with(&[0xff, 0xd8]) { + return None; + } + let (embedded, _, mut inspection) = jpeg_segments_inner(b, false, true); + if !inspection.valid { + return None; + } + inspection.exif = embedded.exif; + inspection.xmp = embedded.xmp.map(|x| x.trim_end_matches('\0').to_string()); + inspection.icc = embedded.icc; + Some(inspection) +} + /// Walk JPEG markers up to SOS and collect APP1 Exif / XMP, APP2 ICC and APP13 IPTC. pub fn jpeg_segments(b: &[u8]) -> Embedded { - jpeg_segments_inner(b, false).0 + jpeg_segments_inner(b, false, false).0 } /// Walk JPEG metadata and return every ordinary XMP APP1 packet in segment order, along with the /// usual [`Embedded`] projection (which retains only the first ordinary packet). pub fn jpeg_segments_with_all_xmp(b: &[u8]) -> (Embedded, Vec) { - jpeg_segments_inner(b, true) + let (embedded, packets, _) = jpeg_segments_inner(b, true, false); + (embedded, packets) } -fn jpeg_segments_inner(b: &[u8], collect_all_xmp: bool) -> (Embedded, Vec) { +fn jpeg_segments_inner(b: &[u8], collect_all_xmp: bool, inspect: bool) -> (Embedded, Vec, JpegInspection) { let mut out = Embedded::default(); let mut xmp_packets = Vec::new(); + let mut inspection = JpegInspection { valid: true, ..Default::default() }; if !b.starts_with(&[0xff, 0xd8]) { - return (out, xmp_packets); + return (out, xmp_packets, inspection); } let mut icc_parts: [Option<&[u8]>; 255] = [None; 255]; let mut icc_total = None; @@ -110,6 +171,10 @@ fn jpeg_segments_inner(b: &[u8], collect_all_xmp: bool) -> (Embedded, Vec (Embedded, Vec (Embedded, Vec out.exif = Some(payload.get(6..).unwrap_or(&[]).to_vec()), + 0xe1 if payload.starts_with(if inspect { b"Exif\0\0" } else { b"Exif\0" }) && out.exif.is_none() => { + out.exif = Some(payload.get(6..).unwrap_or(&[]).to_vec()) + } 0xe1 if payload.starts_with(XMP_SIG) && out.xmp.is_none() => { let packet = String::from_utf8_lossy(&payload[XMP_SIG.len()..]).into_owned(); if collect_all_xmp { @@ -145,7 +219,7 @@ fn jpeg_segments_inner(b: &[u8], collect_all_xmp: bool) -> (Embedded, Vec { xmp_packets.push(String::from_utf8_lossy(&payload[XMP_SIG.len()..]).into_owned()); } - 0xe1 if payload.starts_with(XMP_EXT_SIG) => { + 0xe1 if !inspect && payload.starts_with(XMP_EXT_SIG) => { // GUID (32) + full length (4) + offset (4) + data let p = &payload[XMP_EXT_SIG.len()..]; if p.len() >= 40 { @@ -156,35 +230,9 @@ fn jpeg_segments_inner(b: &[u8], collect_all_xmp: bool) -> (Embedded, Vec { - if !complete || payload.len() < 14 { - icc_invalid = true; - } else if let Some(header) = payload.get(12..14) { - let (Some(&sequence), Some(&total)) = (header.first(), header.get(1)) else { - icc_invalid = true; - i = end; - continue; - }; - if sequence == 0 || total == 0 || sequence > total || icc_total.is_some_and(|expected| expected != total) { - icc_invalid = true; - } else { - icc_total = Some(total); - if let Some(slot) = icc_parts.get_mut(usize::from(sequence - 1)) { - if slot.is_some() { - icc_invalid = true; - } else if let Some(data) = payload.get(14..) { - *slot = Some(data); - } else { - icc_invalid = true; - } - } else { - icc_invalid = true; - } - } - } else { - icc_invalid = true; - } + collect_icc_part(payload, complete, &mut icc_parts, &mut icc_total, &mut icc_invalid); } - 0xed if payload.starts_with(b"Photoshop 3.0\0") && out.iptc.is_none() => out.iptc = photoshop_iptc(&payload[14..]), + 0xed if !inspect && payload.starts_with(b"Photoshop 3.0\0") && out.iptc.is_none() => out.iptc = photoshop_iptc(&payload[14..]), _ => {} } i = end; @@ -214,8 +262,61 @@ fn jpeg_segments_inner(b: &[u8], collect_all_xmp: bool) -> (Embedded, Vec(payload: &'a [u8], complete: bool, parts: &mut [Option<&'a [u8]>; 255], total: &mut Option, invalid: &mut bool) { + if !complete || payload.len() < 14 { + *invalid = true; + return; + } + let (sequence, count) = (payload[12], payload[13]); + if sequence == 0 || count == 0 || sequence > count || total.is_some_and(|expected| expected != count) { + *invalid = true; + return; + } + *total = Some(count); + let Some(slot) = parts.get_mut(usize::from(sequence - 1)) else { + *invalid = true; + return; + }; + if slot.is_some() { + *invalid = true; + } else { + *slot = payload.get(14..); + } +} + +fn inspect_marker(info: &mut JpegInspection, marker: u8, payload: &[u8], start: usize, needs_exif: bool) { + if (0xc0..=0xcf).contains(&marker) && !matches!(marker, 0xc4 | 0xc8 | 0xcc) && payload.len() >= 6 && info.components == 0 { + info.precision = payload[0]; + info.sof = marker; + info.height = u32::from(u16::from_be_bytes([payload[1], payload[2]])); + info.width = u32::from(u16::from_be_bytes([payload[3], payload[4]])); + info.components = payload[5]; + } + if marker == 0xe1 { + info.app1 = info.app1.saturating_add(1); + } + if marker == 0xe1 && payload.starts_with(b"Exif\0\0") && needs_exif { + info.exif_offset = start + 6; + } + if marker == 0xe2 && payload.starts_with(b"MPF\0") && info.mpf.is_none() { + info.mpf = Some((start + 4, payload[4..].to_vec())); + } + if marker == 0xe2 + && info.iso_gainmap.is_none() + && let Some(data) = payload.strip_prefix(b"urn:iso:std:iso:ts:21496:-1\0") + { + info.iso_gainmap = Some(data.to_vec()); + } + if marker == 0xee && payload.starts_with(b"Adobe") && payload.len() >= 12 { + info.adobe_transform = Some(payload[11]); + } } /// Put the Extended XMP of JPEG `b` into `out`: when the standard packet links one, only that GUID's @@ -865,3 +966,45 @@ mod tests { } } } + +#[cfg(test)] +mod jpeg_inspection_tests { + use super::*; + fn segment(out: &mut Vec, marker: u8, data: &[u8]) { + out.extend_from_slice(&[0xff, marker]); + out.extend_from_slice(&((data.len() as u16) + 2).to_be_bytes()); + out.extend_from_slice(data); + } + #[test] + fn reports_mpf_gainmap_and_adobe_transform() { + let mut b = vec![0xff, 0xd8]; + segment(&mut b, 0xe2, b"MPF\0IIabc"); + segment(&mut b, 0xe2, b"urn:iso:std:iso:ts:21496:-1\0abcde"); + segment(&mut b, 0xee, b"Adobe\0\0\0\0\0\0\x02"); + b.extend_from_slice(&[0xff, 0xd9]); + let info = inspect_jpeg(&b).unwrap(); + let (offset, data) = info.mpf.unwrap(); + assert_eq!(&b[offset..offset + data.len()], data); + assert_eq!(info.iso_gainmap.as_deref(), Some(&b"abcde"[..])); + assert_eq!(info.adobe_transform, Some(2)); + } + + #[test] + fn tolerant_walk_keeps_offsets_frame_and_icc_chunks() { + let mut b = vec![0xff, 0xd8, 0x44]; + segment(&mut b, 0xe1, b"Exif\0\0II"); + segment(&mut b, 0xe2, b"ICC_PROFILE\0\x02\x02B"); + segment(&mut b, 0xe2, b"ICC_PROFILE\0\x01\x02A"); + segment(&mut b, 0xc0, &[8, 0, 2, 0, 3, 3]); + b.extend_from_slice(&[0xff, 0xda, 0, 3, 3]); + let info = inspect_jpeg(&b).unwrap(); + assert_eq!((info.width, info.height, info.components, info.first_scan_components), (3, 2, 3, 3)); + assert_eq!(info.app1, 1); + assert_eq!(&b[info.exif_offset..info.exif_offset + 2], b"II"); + assert_eq!(info.icc.as_deref(), Some(&b"AB"[..])); + assert!(!info.icc_invalid); + // The existing strict API still stops at the garbage byte. + assert!(jpeg_segments(&b).exif.is_none()); + assert!(inspect_jpeg(&[0xff, 0xd8, 0xff, 0xe1, 0, 1]).is_none()); + } +} diff --git a/crates/metadata/src/lib.rs b/crates/metadata/src/lib.rs index ff2c530fa..bf6700455 100644 --- a/crates/metadata/src/lib.rs +++ b/crates/metadata/src/lib.rs @@ -37,7 +37,7 @@ pub mod tags; mod xmp; mod xmp_merge; -pub use container::{Embedded, embedded, jpeg_segments, jpeg_segments_with_all_xmp, png_chunks, webp_chunks}; +pub use container::{Embedded, JpegInspection, embedded, inspect_jpeg, jpeg_segments, jpeg_segments_with_all_xmp, png_chunks, webp_chunks}; pub use datetime::{DatePrecision, DateTime}; pub use exif::{from_tiff, read_exif, strip_exif_header, try_read_exif, try_read_exif_with, write_exif}; pub use focus::FocusPoint; diff --git a/crates/metadata/src/mpf.rs b/crates/metadata/src/mpf.rs index 602c8f90d..9da7dc8c6 100644 --- a/crates/metadata/src/mpf.rs +++ b/crates/metadata/src/mpf.rs @@ -113,7 +113,40 @@ pub fn images(jpeg: &[u8]) -> Vec { }; // offsets count from the byte-order field, which follows the signature let Some(base) = segs.iter().find(|s| jpeg.get(s.1..s.2) == Some(payload)).map(|s| s.1 + SIGNATURE.len()) else { return Vec::new() }; - let Some(tiff) = payload.get(SIGNATURE.len()..).and_then(|t| Tiff::parse(t).ok()) else { return Vec::new() }; + let Some(index) = payload.get(SIGNATURE.len()..) else { return Vec::new() }; + parse_index(index, base) +} + +/// Read an already located MPF TIFF index. `base` is the index's offset in the containing file. +/// Entries are bounded to [`MAX_ENTRIES`] and returned without reading image payloads. +pub fn images_from_index(index: &[u8], base: usize) -> Vec { + // The bounded legacy reader requires a complete classic first IFD, while the generic TIFF + // parser can salvage partial directories. Preserve that refusal for located MPF indices. + let le = match index.get(..2) { + Some(b"II") => true, + Some(b"MM") => false, + _ => return Vec::new(), + }; + let read16 = + |at: usize| index.get(at..at.checked_add(2)?).map(|b| if le { u16::from_le_bytes([b[0], b[1]]) } else { u16::from_be_bytes([b[0], b[1]]) }); + let read32 = |at: usize| { + index + .get(at..at.checked_add(4)?) + .map(|b| if le { u32::from_le_bytes([b[0], b[1], b[2], b[3]]) } else { u32::from_be_bytes([b[0], b[1], b[2], b[3]]) }) + }; + if read16(2) != Some(42) { + return Vec::new(); + } + let Some(at) = read32(4).and_then(|v| usize::try_from(v).ok()) else { return Vec::new() }; + let Some(count) = read16(at).map(usize::from).filter(|n| *n <= 4096) else { return Vec::new() }; + if at.checked_add(2).and_then(|p| count.checked_mul(12).and_then(|n| p.checked_add(n))).is_none_or(|end| end > index.len()) { + return Vec::new(); + } + parse_index(index, base) +} + +fn parse_index(index: &[u8], base: usize) -> Vec { + let Some(tiff) = Tiff::parse(index).ok() else { return Vec::new() }; let Some(list) = tiff.ifds.first().and_then(|i| i.bytes(MP_ENTRY)) else { return Vec::new() }; let u32_at = |rec: &[u8], at: usize| rec.get(at..at + 4).and_then(|s| <[u8; 4]>::try_from(s).ok()).map(|s| tiff.order.u32(s) as usize); let mut out = Vec::new(); @@ -226,6 +259,7 @@ mod tests { let segs = segments(&file).unwrap(); assert_eq!(segs.iter().map(|s| s.0).collect::>(), [0xe0, 0xe1, 0xe2, 0xdb]); let (_, a, e) = segs[2]; + assert_eq!(images_from_index(&file[a + SIGNATURE.len()..e], a + SIGNATURE.len()), found); let mut without = file[..a - 4].to_vec(); without.extend_from_slice(&file[e..found[0].offset]); assert_eq!(without, primary); diff --git a/crates/plist/Cargo.toml b/crates/plist/Cargo.toml index 7ae95aae0..7ffaf3480 100644 --- a/crates/plist/Cargo.toml +++ b/crates/plist/Cargo.toml @@ -1,6 +1,6 @@ [package] name = "craft-plist" -description = "Bounded cross-platform binary Apple property-list reader" +description = "Bounded cross-platform binary and XML Apple property-list reader" version.workspace = true edition.workspace = true license.workspace = true diff --git a/crates/plist/README.md b/crates/plist/README.md index dc5c52007..a7adf6c79 100644 --- a/crates/plist/README.md +++ b/crates/plist/README.md @@ -4,8 +4,16 @@ A cross-platform bounded binary (`bplist00`) reader, extracted from Craft Anythi `parse_binary_with_limits` accepts explicit input, object, expanded-node, nesting and decoded-byte limits. Recursion always has a hard ceiling of 256 even if a caller supplies a larger depth. Counts are checked before container reservation; allocations can return a resource error. Limits are per parse, not a promise of a precise process RSS cap. -This first PR contains reading only. XML parsing/writing and keyed-archive interpretation are not implemented. JSON projection, date formatting, OS file access and preference/locale selection are not part of this crate. Existing updater/Photo streaming `plist` users remain unchanged. +This crate contains reading only. Writing and keyed-archive interpretation are not implemented. JSON projection, date formatting, OS file access and preference/locale selection are not part of this crate. Existing updater/Photo streaming `plist` users remain unchanged. -Initial consumers: PhotoCraft's AppleLanguages binary-reader path; later Craft Anything binary parsing and LightCraft XML presets (after XML support/parity). Photo must retain its 4 MiB input policy and first-language selection. +Initial consumers: PhotoCraft's AppleLanguages binary-reader path; later Craft Anything binary parsing and LightCraft XML presets (after consumer parity). Photo must retain its 4 MiB input policy and first-language selection. Source: `storytold/craft-anything` at `88d1d9cac852aec1c7fd97886db451e005e49d2a`, `crates/formats/src/plist.rs` and checked byte reader. Source license files retained. Imported changes retain Brandon Thomas, BFlat, RRFZ and Afloat as primary Git authors; imported regression tests are BFlat's. No external specification implementation copied. + +## XML reader + +`parse_xml` and `parse_xml_with_limits` now read UTF-8 and BOM-tagged UTF-16 XML into the same `Value` tree. XML input/decoded-text size, tokens, values, nesting and cumulative key/leaf text are bounded. The independent token ceiling is eight times the smaller node/object allowance plus eight; unusually fragmented text can exhaust it before the value limit. This is not a precise RSS limit. The decoded-XML limit also includes markup; `` consumes its encoded text allowance, including whitespace. XML `max_objects` and `max_nodes` both cap values. + +The attributed Craft Anything reader includes its complete-document, declaration-skipping (no external entity fetching), quoted-tag and normalized-line-ending behavior. Nonfinite reals are refused using contributor PR #1905. Unknown or malformed character references retain legacy literal behavior; scanning stops at the next non-entity byte, preserving LightCraft's valid zero-padded numeric references without suffix rescans. Base64 retains that reader's legacy optional-padding/ASCII-whitespace behavior, including unused trailing bits; it is not a signature/key-validation API. No writer, JSON projection, keyed-archive interpretation or app policy is added. + +XML source: Craft Anything `8f3b775157d331dd31cfa4da911e26d8856b8c77`, plus #1905 `d85e104f093a334961048881bf003351e1ac4992`. The import credits Brandon Thomas and the existing hardening contributors; #1905 is authored by gmrmad1 and the quoted-tag/line-ending tests by RRFZ. Consumer cutovers require their own policy/parity tests, especially LightCraft's preset mapping and Craft Anything's projections. diff --git a/crates/plist/src/lib.rs b/crates/plist/src/lib.rs index f3af79ddc..d84980398 100644 --- a/crates/plist/src/lib.rs +++ b/crates/plist/src/lib.rs @@ -1,10 +1,12 @@ -//! Bounded binary Apple property lists, independent of operating-system preferences. -//! XML and writing are intentionally not implemented in this first extraction. +//! Bounded binary and XML Apple property lists, independent of operating-system preferences. +//! Writing and application-specific projections remain with callers. //! File discovery, locale selection, bundle validation and application policy stay in callers. #![deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable)] mod binary; mod bytes; +mod xml; pub use binary::{DateValue, Value, first_string_in_array, parse_binary, parse_binary_with_limits}; +pub use xml::{XmlOptions, parse_xml, parse_xml_with_limits, parse_xml_with_options, parse_xml_with_token_limit}; /// Caller-selected resource limits. Zero values refuse the corresponding resource. #[derive(Clone, Copy, Debug)] diff --git a/crates/plist/src/xml.rs b/crates/plist/src/xml.rs new file mode 100644 index 000000000..f4eee291d --- /dev/null +++ b/crates/plist/src/xml.rs @@ -0,0 +1,561 @@ +//! XML property-list reader ported from Craft Anything at 8f3b7751. +use crate::{DateValue, FormatError, Limits, Value}; + +/// XML-specific representation policy, separate from resource allowances. +#[derive(Clone, Copy, Debug)] +pub struct XmlOptions { + /// Trim surrounding whitespace in date text, as the original Craft Anything reader does. + /// Disable for consumers that retain dates as opaque strings, such as Luminar presets. + pub trim_date_text: bool, +} +impl Default for XmlOptions { + fn default() -> Self { + Self { trim_date_text: true } + } +} +fn decode_base64(text: &str, url_safe: bool) -> Option> { + let mut out = Vec::with_capacity(text.len() / 4 * 3); + let mut acc = 0u32; + let mut bits = 0u32; + let mut digits = 0usize; + let mut padding = 0usize; + for c in text.bytes() { + match c { + b' ' | b'\t' | b'\r' | b'\n' => continue, + b'=' => { + padding += 1; + if padding > 2 { + return None; + } + continue; + } + _ if padding != 0 => return None, + _ => {} + } + 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'+' => 62, + b'/' => 63, + b'-' if url_safe => 62, + b'_' if url_safe => 63, + _ => return None, + }; + acc = (acc << 6) | u32::from(value); + bits += 6; + digits += 1; + if bits >= 8 { + bits -= 8; + out.push((acc >> bits) as u8); + acc &= (1 << bits) - 1; + } + } + let remainder = digits % 4; + if remainder == 1 || (padding != 0 && !matches!((remainder, padding), (2, 2) | (3, 1))) { + return None; + } + Some(out) +} + +fn base64_decode(s: &str) -> Option> { + decode_base64(s, false) +} +#[derive(Debug, Clone, PartialEq)] +enum Tok { + Open(String), + Close(String), + Empty(String), + Text(String), +} + +// Ported from LightCraft f786b875: scan one entity-name run, never the remaining document. +fn entity_end(rest: &str) -> Option { + let n = rest.bytes().skip(1).take_while(|b| b.is_ascii_alphanumeric() || *b == b'#').count(); + (rest.as_bytes().get(n + 1) == Some(&b';')).then_some(n + 1) +} + +fn decode_entities(s: &str) -> String { + if !s.contains('&') { + return s.to_string(); + } + let mut out = String::with_capacity(s.len()); + let mut rest = s; + while let Some(p) = rest.find('&') { + out.push_str(rest.get(..p).unwrap_or("")); + let after = rest.get(p..).unwrap_or(""); + let Some(end) = entity_end(after) else { + out.push('&'); + rest = after.get(1..).unwrap_or(""); + continue; + }; + let ent = after.get(1..end).unwrap_or(""); + let rep = match ent { + "lt" => Some('<'), + "gt" => Some('>'), + "amp" => Some('&'), + "quot" => Some('"'), + "apos" => Some('\''), + e if e.starts_with("#x") || e.starts_with("#X") => { + let d = e.get(2..).unwrap_or(""); + if d.bytes().all(|b| b.is_ascii_hexdigit()) { u32::from_str_radix(d, 16).ok().and_then(char::from_u32) } else { None } + } + e if e.starts_with('#') => { + let d = e.get(1..).unwrap_or(""); + if d.bytes().all(|b| b.is_ascii_digit()) { d.parse::().ok().and_then(char::from_u32) } else { None } + } + _ => None, + }; + match rep { + Some(c) => { + out.push(c); + rest = after.get(end + 1..).unwrap_or(""); + } + None => { + out.push('&'); + rest = after.get(1..).unwrap_or(""); + } + } + } + out.push_str(rest); + out +} + +/// Skip a declaration without interpreting or fetching any entity. A DOCTYPE's +/// internal subset can itself contain declarations, quoted delimiters, comments and PIs. +fn after_declaration(mut rest: &str) -> Result<&str, FormatError> { + let mut quote = None; + let mut brackets = 0usize; + while let Some(ch) = rest.chars().next() { + if quote.is_none() { + if let Some(comment) = rest.strip_prefix("").ok_or_else(|| FormatError::Truncated("comment in declaration".into()))?.1; + continue; + } + if let Some(pi) = rest.strip_prefix("").ok_or_else(|| FormatError::Truncated("processing instruction in declaration".into()))?.1; + continue; + } + } + rest = rest.get(ch.len_utf8()..).unwrap_or(""); + match (quote, ch) { + (Some(q), c) if q == c => quote = None, + (None, '\'' | '"') => quote = Some(ch), + (None, '[') => brackets += 1, + (None, ']') => brackets = brackets.saturating_sub(1), + (None, '>') if brackets == 0 => return Ok(rest), + _ => {} + } + } + Err(FormatError::Truncated("declaration".into())) +} + +fn tokenize(text: &str, max_tokens: usize) -> Result, FormatError> { + let mut toks = Vec::new(); + let mut rest = text; + while !rest.is_empty() { + if toks.len() >= max_tokens { + return Err(FormatError::TooLarge("too many XML tokens".into())); + } + match rest.find('<') { + None => { + if !rest.is_empty() { + toks.push(Tok::Text(decode_entities(rest))); + } + break; + } + Some(p) => { + let t = rest.get(..p).unwrap_or(""); + if !t.is_empty() { + toks.push(Tok::Text(decode_entities(t))); + } + rest = rest.get(p..).unwrap_or(""); + } + } + if let Some(r) = rest.strip_prefix("").ok_or_else(|| FormatError::Truncated("comment".into()))?; + rest = r.get(e + 3..).unwrap_or(""); + continue; + } + if let Some(r) = rest.strip_prefix("").ok_or_else(|| FormatError::Truncated("CDATA".into()))?; + toks.push(Tok::Text(r.get(..e).unwrap_or("").to_string())); + rest = r.get(e + 3..).unwrap_or(""); + continue; + } + if let Some(r) = rest.strip_prefix("").ok_or_else(|| FormatError::Truncated("processing instruction".into()))?; + rest = r.get(end + 2..).unwrap_or(""); + continue; + } + if let Some(declaration) = rest.strip_prefix(" { + quote = None; + false + } + (None, '\'' | '"') => { + quote = Some(ch); + false + } + (None, '>') => true, + _ => false, + }) + .ok_or_else(|| FormatError::Truncated("tag".into()))?; + let inner = rest.get(1..e).unwrap_or("").trim(); + rest = rest.get(e + 1..).unwrap_or(""); + if let Some(name) = inner.strip_prefix('/') { + toks.push(Tok::Close(name.trim().to_string())); + } else if let Some(body) = inner.strip_suffix('/') { + toks.push(Tok::Empty(body.split_whitespace().next().unwrap_or("").to_string())); + } else { + toks.push(Tok::Open(inner.split_whitespace().next().unwrap_or("").to_string())); + } + } + if toks.len() > max_tokens { + return Err(FormatError::TooLarge("too many XML tokens".into())); + } + Ok(toks) +} + +struct Xml { + toks: Vec, + pos: usize, + nodes: usize, + limits: Limits, + bytes_left: usize, + options: XmlOptions, +} + +impl Xml { + fn next(&mut self) -> Option { + let t = self.toks.get(self.pos).cloned(); + self.pos += 1; + t + } + + fn peek(&self) -> Option<&Tok> { + self.toks.get(self.pos) + } + + /// Text up to the closing tag `name`. + fn text_until(&mut self, name: &str) -> Result { + let mut s = String::new(); + loop { + match self.next() { + Some(Tok::Text(t)) => { + self.bytes_left = self.bytes_left.checked_sub(t.len()).ok_or_else(|| FormatError::TooLarge("XML text/data bytes".into()))?; + s.push_str(&t); + } + Some(Tok::Close(n)) if n == name => return Ok(s), + Some(other) => return Err(FormatError::Malformed(format!("unexpected {other:?} inside <{name}>"))), + None => return Err(FormatError::Truncated(format!("<{name}> is not closed"))), + } + } + } + + fn skip_space(&mut self) { + while matches!(self.peek(), Some(Tok::Text(t)) if t.bytes().all(|b| matches!(b, b' ' | b'\t' | b'\r' | b'\n'))) { + self.pos += 1; + } + } + + fn value(&mut self, depth: usize) -> Result { + self.nodes += 1; + if depth > self.limits.max_depth.min(256) || self.nodes > self.limits.max_nodes.min(self.limits.max_objects) { + return Err(FormatError::TooLarge("plist nesting or size limit".into())); + } + self.skip_space(); + match self.next() { + Some(Tok::Empty(n)) => Ok(match n.as_str() { + "true" => Value::Bool(true), + "false" => Value::Bool(false), + "dict" => Value::Dict(Vec::new()), + "array" => Value::Array(Vec::new()), + "string" => Value::String(String::new()), + "data" => Value::Data(Vec::new()), + other => return Err(FormatError::Malformed(format!("unexpected <{other}/>"))), + }), + Some(Tok::Open(n)) => match n.as_str() { + "string" => Ok(Value::String(self.text_until("string")?)), + "integer" => { + let t = self.text_until("integer")?; + let t = t.trim(); + let v = if let Some(h) = t.strip_prefix("0x").or_else(|| t.strip_prefix("0X")) { + i128::from_str_radix(h, 16).ok() + } else { + t.parse::().ok() + }; + v.map(Value::Integer).ok_or_else(|| FormatError::Malformed(format!("bad integer {t:?}"))) + } + "real" => { + let t = self.text_until("real")?; + t.trim() + .parse::() + .ok() + .filter(|v| v.is_finite()) + .map(Value::Real) + .ok_or_else(|| FormatError::Malformed(format!("bad real {t:?}"))) + } + "date" => { + let text = self.text_until("date")?; + Ok(Value::Date(DateValue::Text(if self.options.trim_date_text { text.trim().to_string() } else { text }))) + } + "data" => { + let t = self.text_until("data")?; + base64_decode(&t).map(Value::Data).ok_or_else(|| FormatError::Malformed("bad base64 in ".into())) + } + "true" | "false" => { + if !self.text_until(&n)?.is_empty() { + return Err(FormatError::Malformed(format!("<{n}> must be empty"))); + } + Ok(Value::Bool(n == "true")) + } + "array" => { + let mut out = Vec::new(); + loop { + self.skip_space(); + if matches!(self.peek(), Some(Tok::Close(c)) if c == "array") { + self.pos += 1; + return Ok(Value::Array(out)); + } + if self.peek().is_none() { + return Err(FormatError::Truncated(" is not closed".into())); + } + out.push(self.value(depth + 1)?); + } + } + "dict" => { + let mut out = Vec::new(); + loop { + self.skip_space(); + match self.next() { + Some(Tok::Close(c)) if c == "dict" => return Ok(Value::Dict(out)), + Some(Tok::Open(k)) if k == "key" => { + let key = self.text_until("key")?; + let v = self.value(depth + 1)?; + out.push((key, v)); + } + Some(Tok::Empty(k)) if k == "key" => { + let v = self.value(depth + 1)?; + out.push((String::new(), v)); + } + Some(other) => return Err(FormatError::Malformed(format!("expected in , found {other:?}"))), + None => return Err(FormatError::Truncated(" is not closed".into())), + } + } + } + other => Err(FormatError::Malformed(format!("unknown plist element <{other}>"))), + }, + Some(other) => Err(FormatError::Malformed(format!("expected a value, found {other:?}"))), + None => Err(FormatError::Truncated("expected a value".into())), + } + } +} + +/// Decode a UTF-16 XML property list after its byte-order mark into UTF-8 text. +/// Surrogate pairs join normally; a lone surrogate fails as malformed rather +/// than turning into replacement characters that tokenize into garbage element +/// names. +fn decode_utf16_xml(data: &[u8], big_endian: bool) -> Result { + if !data.len().is_multiple_of(2) { + return Err(FormatError::Truncated("UTF-16 code unit".into())); + } + let units: Vec = data.as_chunks::<2>().0.iter().map(|b| if big_endian { u16::from_be_bytes(*b) } else { u16::from_le_bytes(*b) }).collect(); + String::from_utf16(&units).map_err(|_| FormatError::Malformed("invalid UTF-16 in XML plist".into())) +} + +/// Parse an XML property list. +pub fn parse_xml(data: &[u8]) -> Result { + parse_xml_with_limits(data, Limits::default()) +} + +/// Parse UTF-8 or BOM-tagged UTF-16 XML with explicit resource allowances. +/// Input and decoded XML text are bounded before tokenization; token count, value count, +/// nesting and cumulative leaf/key text are bounded separately. `max_objects` and +/// `max_nodes` both cap values. Nesting never exceeds 256 even with a larger caller limit. +/// Data text consumes its encoded size, conservatively including whitespace. +pub fn parse_xml_with_limits(data: &[u8], limits: Limits) -> Result { + parse_xml_with_options(data, limits, XmlOptions::default()) +} + +/// Parse bounded XML using an explicit date-text representation policy. +pub fn parse_xml_with_options(data: &[u8], limits: Limits, options: XmlOptions) -> Result { + let max_tokens = limits.max_nodes.min(limits.max_objects).saturating_mul(8).saturating_add(8); + parse_xml_with_token_limit(data, limits, options, max_tokens) +} + +/// Parse XML with an independent token allowance, without reducing the value allowance. +/// This bounds token-heavy input (for example many CDATA chunks in one string). +/// Zero tokens refuses nonempty XML. Existing convenience APIs retain their default budget. +pub fn parse_xml_with_token_limit(data: &[u8], limits: Limits, options: XmlOptions, max_tokens: usize) -> Result { + if data.len() > limits.max_input_bytes { + return Err(FormatError::TooLarge("input bytes".into())); + } + // Apple's Core Foundation reader accepts XML plists in UTF-16 as well as + // UTF-8: it recognises both byte-order marks. `from_utf8_lossy` turns a + // UTF-16 byte stream into NUL-separated text, so the encoding is settled + // here, before tokenizing. + let text = match data { + [0xFF, 0xFE, rest @ ..] => decode_utf16_xml(rest, false)?, + [0xFE, 0xFF, rest @ ..] => decode_utf16_xml(rest, true)?, + _ => std::str::from_utf8(data.strip_prefix(b"\xef\xbb\xbf").unwrap_or(data)) + .map_err(|_| FormatError::Malformed("invalid UTF-8 in XML plist".into()))? + .to_owned(), + }; + if text.len() > limits.max_expanded_bytes { + return Err(FormatError::TooLarge("decoded XML bytes".into())); + } + // XML normalizes literal line endings before parsing; character references such as + // must still decode to a carriage return afterwards. + let text = if text.contains('\r') { text.replace("\r\n", "\n").replace('\r', "\n") } else { text }; + let toks = tokenize(&text, max_tokens)?; + let mut x = Xml { toks, pos: 0, nodes: 0, limits, bytes_left: limits.max_expanded_bytes, options }; + x.skip_space(); + let wrapped = matches!(x.peek(), Some(Tok::Open(n)) if n == "plist"); + if wrapped { + x.pos += 1; + } + let value = x.value(0)?; + x.skip_space(); + if wrapped { + if !matches!(x.next(), Some(Tok::Close(n)) if n == "plist") { + return Err(FormatError::Malformed("expected after its value".into())); + } + x.skip_space(); + } + if x.peek().is_some() { + return Err(FormatError::Malformed("unexpected content after the plist value".into())); + } + Ok(value) +} + +#[cfg(test)] +mod tests { + #![allow(clippy::unwrap_used, clippy::panic)] + use super::*; + #[test] + fn explicit_token_budget_is_independent_of_value_budget() { + let limits = Limits { max_nodes: 100, max_objects: 100, ..Limits::default() }; + let xml = b""; + assert!(matches!(parse_xml_with_token_limit(xml, limits, XmlOptions::default(), 3), Err(FormatError::TooLarge(_)))); + assert_eq!(parse_xml_with_token_limit(xml, limits, XmlOptions::default(), 4).unwrap(), Value::String("ab".into())); + assert!(matches!(parse_xml_with_token_limit(b"", limits, XmlOptions::default(), 0), Err(FormatError::TooLarge(_)))); + } + + #[test] + fn xml_has_one_complete_root_with_only_markup_space_around_values() { + for invalid in [ + "", + "", + "", + "", + "junk", + "text", + "junk", + "xgarbage", + "no", + "yes", + "", + ] { + assert!(parse_xml(invalid.as_bytes()).is_err(), "{invalid}"); + } + let xml = b" ?>\n "; + assert_eq!(parse_xml(xml).unwrap().get("").and_then(Value::as_str), Some(" ")); + assert_eq!(parse_xml(b" ").unwrap(), Value::Bool(true)); + } + + #[test] + fn xml_rejects_partial_code_units_and_preserves_the_utf8_bom() { + assert_eq!(parse_xml(b"\xef\xbb\xbfok").unwrap().as_str(), Some("ok")); + assert!(parse_xml(b"\xff").is_err()); + for be in [false, true] { + let mut bytes = if be { vec![0xfe, 0xff] } else { vec![0xff, 0xfe] }; + for unit in "".encode_utf16() { + bytes.extend(if be { unit.to_be_bytes() } else { unit.to_le_bytes() }); + } + bytes.push(0); + assert!(parse_xml(&bytes).is_err()); + } + } + + #[test] + fn a_doctype_internal_subset_is_skipped_as_one_declaration() { + let doc = br#"y"> ?>]>ok"#; + assert_eq!(parse_xml(doc).unwrap().as_str(), Some("ok")); + assert!(parse_xml(b"").is_err()); + } + + #[test] + fn xml_char_refs_require_plain_digits() { + let text = |s: &str| { + let doc = format!("{s}"); + parse_xml(doc.as_bytes()).unwrap().as_str().map(str::to_string) + }; + assert_eq!(text("a&#+65;b&#x+41;").as_deref(), Some("a&#+65;b&#x+41;")); + assert_eq!(text("AA").as_deref(), Some("AA")); + assert_eq!(text("&#;&#x;").as_deref(), Some("&#;&#x;")); + } + + #[test] + fn xml_real_must_be_finite() { + for bad in ["NaN", "1e999"] { + let doc = format!("k{bad}"); + assert!(parse_xml(doc.as_bytes()).is_err(), "{bad}"); + } + for (text, want) in [("1.5", 1.5), ("-2", -2.0), ("0", 0.0)] { + let doc = format!("k{text}"); + let v = parse_xml(doc.as_bytes()).unwrap_or_else(|e| panic!("{text} should parse: {e}")); + assert_eq!(v.get("k"), Some(&Value::Real(want)), "{text}"); + } + } + + #[test] + fn malformed_entity_scan_is_bounded_and_preserves_text() { + for n in [4096, 65536, 262144] { + let text = "&".repeat(n); + assert_eq!(decode_entities(&text), text); + assert_eq!(decode_entities(&(text.clone() + ";")), text + ";"); + } + assert_eq!(decode_entities("é&missing; & A"), "é&missing; & \r A"); + assert_eq!(decode_entities("Canon A"), "Canon A"); + } + #[test] + fn caller_limits_refuse_before_unbounded_tree_growth() { + let limits = Limits { max_input_bytes: 3, ..Limits::default() }; + assert!(matches!(parse_xml_with_limits(b"", limits), Err(FormatError::TooLarge(_)))); + let limits = Limits { max_nodes: 1, ..Limits::default() }; + assert!(matches!(parse_xml_with_limits(b"", limits), Err(FormatError::TooLarge(_)))); + let limits = Limits { max_depth: 0, ..Limits::default() }; + assert!(matches!(parse_xml_with_limits(b"", limits), Err(FormatError::TooLarge(_)))); + let limits = Limits { max_expanded_bytes: 8, ..Limits::default() }; + assert!(matches!(parse_xml_with_limits(b"abc", limits), Err(FormatError::TooLarge(_)))); + assert!(parse_xml_with_limits(b"", Limits::default()).is_ok()); + } + #[test] + fn data_preserves_existing_whitespace_padding_and_alphabet_rules() { + for (s, expected) in [(" Z g == ", Some(vec![102])), ("Zg", Some(vec![102])), ("Zg=", None), ("Zg==junk", None), ("_w==", None)] { + let xml = format!("{s}"); + assert_eq!(parse_xml(xml.as_bytes()).ok(), expected.map(Value::Data)); + } + } + #[test] + fn small_node_allowance_accepts_one_dictionary_entry() { + let limits = Limits { max_nodes: 2, max_objects: 2, ..Limits::default() }; + assert!(parse_xml_with_limits(b"kv", limits).is_ok()); + } + #[test] + fn callers_can_preserve_opaque_date_whitespace() { + let xml = b" 2026-01-01T00:00:00Z "; + assert_eq!(parse_xml(xml).unwrap(), Value::Date(DateValue::Text("2026-01-01T00:00:00Z".into()))); + assert_eq!( + parse_xml_with_options(xml, Limits::default(), XmlOptions { trim_date_text: false }).unwrap(), + Value::Date(DateValue::Text(" 2026-01-01T00:00:00Z ".into())) + ); + } +} diff --git a/crates/plist/tests/plist_line_endings.rs b/crates/plist/tests/plist_line_endings.rs new file mode 100644 index 000000000..3f8f82b50 --- /dev/null +++ b/crates/plist/tests/plist_line_endings.rs @@ -0,0 +1,17 @@ +use craft_plist::{self as plist, Value}; + +/// Literal line endings normalize before entities, including keys and CDATA. +#[test] +fn xml_line_endings_follow_xml_rules() { + let xml = "a\rba\r\nb\rc\nd e"; + for encoding in 0..3 { + let bytes = match encoding { + 1 => [vec![0xff, 0xfe], xml.encode_utf16().flat_map(u16::to_le_bytes).collect()].concat(), + 2 => [vec![0xfe, 0xff], xml.encode_utf16().flat_map(u16::to_be_bytes).collect()].concat(), + _ => xml.as_bytes().to_vec(), + }; + let value = plist::parse_xml(&bytes).unwrap(); + assert_eq!(value.get("a\nb").and_then(Value::as_str), Some("a\nb\nc\nd\ref\ng")); + } + assert!(plist::parse_xml(b"a\r\nb").is_err()); +} diff --git a/crates/plist/tests/plist_quoted_attributes.rs b/crates/plist/tests/plist_quoted_attributes.rs new file mode 100644 index 000000000..54cd042ef --- /dev/null +++ b/crates/plist/tests/plist_quoted_attributes.rs @@ -0,0 +1,22 @@ +use craft_plist::{self as plist, Value}; + +/// Quoted delimiters belong to attributes, including on empty elements. +#[test] +fn quoted_tag_delimiters_preserve_values() { + for quote in ['\'', '"'] { + let xml = format!( + "b{quote}>界{quote}>key/{quote}>valueempty/{quote}/>" + ); + let value = plist::parse_xml(xml.as_bytes()).unwrap(); + assert_eq!(value.get("key").and_then(Value::as_str), Some("value")); + assert_eq!(value.get("empty").and_then(Value::as_str), Some("")); + } +} + +/// A quote that never closes must not turn later markup into a value. +#[test] +fn unterminated_attribute_quotes_are_rejected() { + for xml in ["value", " i32 { + match self { + Self::Seed { status, .. } | Self::Apex { status, .. } => status, + } + } + + fn face(self) -> u32 { + match self { + Self::Seed { face, .. } | Self::Apex { face, .. } => face, + } + } +} + +#[derive(Clone)] +struct Encoded { + tolerance: f64, + origin: V, + points: Vec, + edge_status: Vec, + face_of: Vec, + is_reference: Vec, + references: Vec, + positions: Vec, + triangles: Vec<[u32; 3]>, +} + +impl Encoded { + fn arrays(&self, limit: usize) -> Arrays<'_> { + Arrays { + tolerance: self.tolerance, + origin: self.origin, + points: &self.points, + edge_status: &self.edge_status, + triangles: self.triangles.len(), + is_reference: &self.is_reference, + references: &self.references, + limit, + } + } + + #[allow(dead_code)] + fn faces(&self) -> usize { + self.face_of.iter().max().map_or(0, |&face| face as usize + 1) + } + + fn push_reference(&mut self, vertex: u32) -> u32 { + let index = usize::try_from(vertex).unwrap(); + assert!(index < self.positions.len(), "a recipe reference must name an earlier vertex"); + self.is_reference.push(true); + self.references.push(vertex); + vertex + } + + fn push_position(&mut self, position: V, point: [i64; 3]) -> u32 { + self.points.extend(point); + self.is_reference.push(false); + self.positions.push(position); + u32::try_from(self.positions.len() - 1).unwrap() + } + + fn push_seed_point(&mut self, index: usize, target: V, triangle: [u32; 3]) -> u32 { + let base = match index { + 0 => self.origin, + 1 => self.positions[usize::try_from(triangle[0]).unwrap()], + 2 => { + let first = self.positions[usize::try_from(triangle[0]).unwrap()]; + let second = self.positions[usize::try_from(triangle[1]).unwrap()]; + p_scale(p_add(first, second), 0.5) + } + _ => panic!("a seed has exactly three slots"), + }; + let point = quantize(p_sub(target, base), self.tolerance); + let position = p_add(base, scaled(point, self.tolerance)); + self.push_position(position, point) + } + + fn push_apex_point(&mut self, edge: (u32, u32), opposite: u32, target: V) -> u32 { + let (p, q) = edge; + let (lo, hi) = if p < q { (p, q) } else { (q, p) }; + let lo = self.positions[usize::try_from(lo).unwrap()]; + let hi = self.positions[usize::try_from(hi).unwrap()]; + let opposite = self.positions[usize::try_from(opposite).unwrap()]; + let (origin, x, y, z) = encoder_frame(lo, hi, opposite); + let offset = p_sub(target, origin); + let point = [ + quantize_component(p_dot(offset, x), self.tolerance), + quantize_component(-p_dot(offset, y), self.tolerance), + quantize_component(-p_dot(offset, z), self.tolerance), + ]; + let delta = scaled(point, self.tolerance); + let position = p_sub(p_sub(p_add(origin, p_scale(x, delta[0])), p_scale(y, delta[1])), p_scale(z, delta[2])); + self.push_position(position, point) + } +} + +fn encode(origin: V, tolerance: f64, recipes: &[Recipe]) -> Encoded { + assert!(tolerance.is_finite() && tolerance > 0.0, "a positive finite tolerance"); + let mut edge_status: Vec = recipes.iter().map(|recipe| recipe.status()).collect(); + edge_status.resize(recipes.len().saturating_mul(3), 0); + let mut encoded = Encoded { + tolerance, + origin, + points: Vec::new(), + edge_status, + face_of: Vec::new(), + is_reference: Vec::new(), + references: Vec::new(), + positions: Vec::new(), + triangles: Vec::new(), + }; + + for &recipe in recipes { + encoded.face_of.push(recipe.face()); + match recipe { + Recipe::Seed { slots, .. } => { + let mut triangle = [0; 3]; + for (index, slot) in slots.into_iter().enumerate() { + triangle[index] = match slot { + Slot::Point(target) => encoded.push_seed_point(index, target, triangle), + Slot::Reference(vertex) => encoded.push_reference(vertex), + }; + } + encoded.triangles.push(triangle); + } + Recipe::Apex { edge: (p, q), opposite, slot, .. } => { + let vertex = match slot { + Slot::Point(target) => encoded.push_apex_point((p, q), opposite, target), + Slot::Reference(vertex) => encoded.push_reference(vertex), + }; + encoded.triangles.push([p, q, vertex]); + } + } + } + encoded +} + +fn p_add(a: V, b: V) -> V { + [a[0] + b[0], a[1] + b[1], a[2] + b[2]] +} + +fn p_sub(a: V, b: V) -> V { + [a[0] - b[0], a[1] - b[1], a[2] - b[2]] +} + +fn p_scale(a: V, factor: f64) -> V { + a.map(|component| component * factor) +} + +fn p_dot(a: V, b: V) -> f64 { + a[0] * b[0] + a[1] * b[1] + a[2] * b[2] +} + +fn p_cross(a: V, b: V) -> V { + [a[1] * b[2] - a[2] * b[1], a[2] * b[0] - a[0] * b[2], a[0] * b[1] - a[1] * b[0]] +} + +fn quantize_component(value: f64, tolerance: f64) -> i64 { + let quantized = (value / tolerance).round(); + assert!(quantized.is_finite() && quantized >= i64::MIN as f64 && quantized <= i64::MAX as f64); + quantized as i64 +} + +fn quantize(value: V, tolerance: f64) -> [i64; 3] { + value.map(|component| quantize_component(component, tolerance)) +} + +fn scaled(point: [i64; 3], tolerance: f64) -> V { + point.map(|component| component as f64 * tolerance) +} + +fn encoder_length(value: V) -> Option { + let squared = (value[0] * value[0] + value[1] * value[1] + value[2] * value[2]) as f32; + let bits = squared.to_bits(); + let exponent = (bits >> 23) & 0xff; + let seed = if exponent > 127 { f32::from_bits((bits & !(0xff << 23)) | (((exponent - 127) / 2 + 127) << 23)) } else { squared }; + let squared = f64::from(squared); + let mut current = f64::from(seed); + for _ in 0..100 { + let next = 0.5 * (current + squared / current); + if next == current { + return Some(next); + } + current = next; + } + None +} + +fn encoder_unit(value: V) -> Option { + let length = encoder_length(value).filter(|&length| length >= f64::from(f32::EPSILON))?; + Some([value[0] / length, value[1] / length, value[2] / length]) +} + +fn encoder_ortho_z(x: V) -> Option { + let x = encoder_unit(x)?; + encoder_unit(p_cross(x, [0.0, 1.0, 0.0])).or_else(|| encoder_unit(p_cross(x, [1.0, 0.0, 0.0]))) +} + +fn encoder_frame(lo: V, hi: V, opposite: V) -> (V, V, V, V) { + let origin = p_scale(p_add(lo, hi), 0.5); + let difference = p_sub(lo, hi); + let x = encoder_unit(difference).unwrap_or(difference); + let residue = p_cross(p_sub(opposite, origin), x); + let (y, z) = encoder_unit(residue) + .and_then(|z| Some((encoder_unit(p_cross(z, x))?, z))) + .or_else(|| { + let z = encoder_ortho_z(x)?; + let y = encoder_unit(p_cross(z, x))?; + Some(if p_dot(residue, z) < 0.0 { (p_scale(y, -1.0), p_scale(z, -1.0)) } else { (y, z) }) + }) + .unwrap_or_default(); + (origin, x, y, z) +} + +fn assert_positions(actual: &[V], expected: &[V]) { + assert_eq!(actual.len(), expected.len(), "vertex counts differ"); + for (index, (actual, expected)) in actual.iter().zip(expected).enumerate() { + for component in 0..3 { + assert!( + (actual[component] - expected[component]).abs() < 1e-9, + "vertex {index}, component {component}: actual {actual:?}, expected {expected:?}", + ); + } + } +} + +#[allow(dead_code)] +fn assert_within_tolerance(actual: &[V], source: &[V], tolerance: f64) { + assert_eq!(actual.len(), source.len(), "vertex counts differ"); + let limit = tolerance * 3.0_f64.sqrt() / 2.0; + for (index, (actual, source)) in actual.iter().zip(source).enumerate() { + let difference = p_sub(*actual, *source); + let distance = p_dot(difference, difference).sqrt(); + assert!(distance <= limit + 1e-12, "vertex {index}: distance {distance} exceeds {limit}; actual {actual:?}, source {source:?}"); + } +} + +fn triangle_seed(x: f64, status: i32, face: u32) -> Recipe { + Recipe::Seed { slots: [Slot::Point([x, 0.0, 0.0]), Slot::Point([x + 2.0, 0.0, 0.0]), Slot::Point([x, 2.0, 0.0])], status, face } +} + +fn panel(target: V) -> Encoded { + encode( + [0.0; 3], + 0.5, + &[ + triangle_seed(0.0, 3, 0), + Recipe::Apex { edge: (2, 1), opposite: 0, slot: Slot::Point(target), status: 3, face: 0 }, + Recipe::Apex { edge: (2, 3), opposite: 1, slot: Slot::Reference(0), status: 2, face: 0 }, + Recipe::Apex { edge: (0, 3), opposite: 2, slot: Slot::Reference(1), status: 0, face: 0 }, + ], + ) +} + +#[test] +fn independent_seed_and_apex_match_a_hand_computed_stream() { + let encoded = encode( + [10.0, 0.0, 0.0], + 0.5, + &[ + Recipe::Seed { slots: [Slot::Point([10.0, 0.0, 0.0]), Slot::Point([12.0, 0.0, 0.0]), Slot::Point([10.0, 2.0, 0.0])], status: 2, face: 0 }, + Recipe::Apex { edge: (2, 1), opposite: 0, slot: Slot::Point([11.0 + 2.0 * S, 1.0 + S, -1.0]), status: 0, face: 0 }, + ], + ); + + assert_eq!(encoded.points, [0, 0, 0, 4, 0, 0, -2, 4, 0, 2, -6, 2]); + let rebuilt = rebuild(&encoded.arrays(usize::MAX)).unwrap(); + assert_eq!(rebuilt.triangles, [[0, 1, 2], [2, 1, 3]]); + let expected = [[10.0, 0.0, 0.0], [12.0, 0.0, 0.0], [10.0, 2.0, 0.0], [11.0 + 2.0 * S, 1.0 + S, -1.0]]; + assert_positions(&rebuilt.positions, &expected); +} + +#[test] +fn independent_components_and_a_reference_are_rebuilt() { + let encoded = encode( + [0.0; 3], + 0.5, + &[ + triangle_seed(0.0, 2, 0), + Recipe::Apex { edge: (2, 1), opposite: 0, slot: Slot::Reference(0), status: 0, face: 0 }, + triangle_seed(5.0, 0, 1), + ], + ); + + let rebuilt = rebuild(&encoded.arrays(usize::MAX)).unwrap(); + assert_eq!(rebuilt.triangles, [[0, 1, 2], [2, 1, 0], [3, 4, 5]]); + assert_eq!(encoded.references, [0]); + assert_eq!(encoded.face_of, [0, 0, 1]); + assert_eq!(rebuilt.positions.len(), 6); + let expected = [[0.0, 0.0, 0.0], [2.0, 0.0, 0.0], [0.0, 2.0, 0.0], [5.0, 0.0, 0.0], [7.0, 0.0, 0.0], [5.0, 2.0, 0.0]]; + assert_positions(&rebuilt.positions, &expected); +} + +#[test] +fn independent_fold_and_component_retry_rebuild_the_same_panel() { + let folded = [1.0 - 1.0 / S, 1.0 - S, 0.0]; + let unfolded = [1.0 + S, 1.0 + 1.0 / S, 0.0]; + for target in [folded, unfolded] { + let source = [[0.0, 0.0, 0.0], [2.0, 0.0, 0.0], [0.0, 2.0, 0.0], target]; + let encoded = panel(target); + assert_within_tolerance(&encoded.positions, &source, 0.5); + let rebuilt = rebuild(&encoded.arrays(usize::MAX)).unwrap(); + assert_eq!(rebuilt.triangles, [[0, 1, 2], [2, 1, 3], [2, 3, 0], [0, 3, 1]]); + assert_eq!(rebuilt.positions.len(), 4); + assert_within_tolerance(&rebuilt.positions, &source, 0.5); + } +} + +#[test] +fn independent_closed_continuation_uses_the_open_edge() { + let encoded = encode( + [0.0; 3], + 0.5, + &[ + triangle_seed(0.0, 2, 0), + Recipe::Apex { edge: (2, 1), opposite: 0, slot: Slot::Point([0.0, 0.0, 2.0]), status: 2, face: 0 }, + Recipe::Apex { edge: (3, 1), opposite: 2, slot: Slot::Reference(0), status: 2, face: 0 }, + Recipe::Apex { edge: (3, 0), opposite: 1, slot: Slot::Reference(2), status: 0, face: 0 }, + ], + ); + + let source = [[0.0, 0.0, 0.0], [2.0, 0.0, 0.0], [0.0, 2.0, 0.0], [0.0, 0.0, 2.0]]; + assert_within_tolerance(&encoded.positions, &source, 0.5); + let rebuilt = rebuild(&encoded.arrays(usize::MAX)).unwrap(); + assert_eq!(rebuilt.triangles, [[0, 1, 2], [2, 1, 3], [3, 1, 0], [3, 0, 2]]); + assert_within_tolerance(&rebuilt.positions, &source, 0.5); +} + +#[test] +fn independent_non_planar_strip_tracks_the_quantized_working_mesh() { + let source: Vec = (0..16).map(|index| [f64::from(index) * 0.05, f64::from(index % 2) * 2.0, (f64::from(index % 5) - 2.0) * 0.03]).collect(); + let mut recipes = vec![Recipe::Seed { slots: [Slot::Point(source[0]), Slot::Point(source[1]), Slot::Point(source[2])], status: 2, face: 0 }]; + let (mut p, mut q, mut opposite) = (2_u32, 1_u32, 0_u32); + + for index in 1..source.len() - 2 { + let vertex = u32::try_from(index + 2).unwrap(); + let status = if index % 2 == 1 { 1 } else { 2 }; + recipes.push(Recipe::Apex { edge: (p, q), opposite, slot: Slot::Point(source[index + 2]), status, face: 0 }); + if status == 1 { + (p, q, opposite) = (p, vertex, q); + } else { + (p, q, opposite) = (vertex, q, p); + } + } + + let encoded = encode([0.0; 3], 1e-4, &recipes); + assert_within_tolerance(&encoded.positions, &source, 1e-4); + let rebuilt = rebuild(&encoded.arrays(usize::MAX)).unwrap(); + assert_eq!( + rebuilt.triangles, + [ + [0, 1, 2], + [2, 1, 3], + [2, 3, 4], + [4, 3, 5], + [4, 5, 6], + [6, 5, 7], + [6, 7, 8], + [8, 7, 9], + [8, 9, 10], + [10, 9, 11], + [10, 11, 12], + [12, 11, 13], + [12, 13, 14], + [14, 13, 15], + ] + ); + assert_eq!(encoded.points.len(), source.len() * 3); + assert_within_tolerance(&rebuilt.positions, &source, 1e-4); +} + +#[test] +fn independent_near_degenerate_parent_uses_an_orthogonal_frame() { + let encoded = encode( + [0.0; 3], + 1e-8, + &[ + Recipe::Seed { + slots: [Slot::Point([1.0, 1.0 + 1e-8, 0.0]), Slot::Point([2.0, 0.0, 0.0]), Slot::Point([0.0, 2.0, 0.0])], + status: 2, + face: 0, + }, + Recipe::Apex { edge: (2, 1), opposite: 0, slot: Slot::Point([0.3, 0.7, 0.2]), status: 0, face: 0 }, + ], + ); + + let source = [[1.0, 1.0 + 1e-8, 0.0], [2.0, 0.0, 0.0], [0.0, 2.0, 0.0], [0.3, 0.7, 0.2]]; + assert_within_tolerance(&encoded.positions, &source, 1e-8); + let rebuilt = rebuild(&encoded.arrays(usize::MAX)).unwrap(); + assert_eq!(rebuilt.triangles, [[0, 1, 2], [2, 1, 3]]); + assert_within_tolerance(&rebuilt.positions, &source, 1e-8); +} + +#[test] +fn independent_rebuild_requires_exact_arrays_and_enough_budget() { + let encoded = encode([0.0; 3], 0.5, &[triangle_seed(0.0, 0, 0)]); + assert!(rebuild(&encoded.arrays(usize::MAX)).is_ok()); + + let mut extra_point = encoded.clone(); + extra_point.points.push(0); + assert_eq!(rebuild(&extra_point.arrays(usize::MAX)).unwrap_err(), "the decoded triangles do not use up the stored arrays",); + + let mut extra_slot = encoded.clone(); + extra_slot.is_reference.push(false); + assert_eq!(rebuild(&extra_slot.arrays(usize::MAX)).unwrap_err(), "the decoded triangles do not use up the stored arrays",); + + let mut extra_reference = encoded.clone(); + extra_reference.references.push(0); + assert_eq!(rebuild(&extra_reference.arrays(usize::MAX)).unwrap_err(), "the decoded triangles do not use up the stored arrays",); + + assert_eq!(rebuild(&encoded.arrays(0)).unwrap_err(), "the rebuild needs more than the decoded-element allowance left",); +} + +#[test] +fn independent_faces_group_walk_order_triangles() { + let encoded = encode([0.0; 3], 0.5, &[triangle_seed(0.0, 0, 1), triangle_seed(5.0, 0, 0)]); + let rebuilt = rebuild(&encoded.arrays(usize::MAX)).unwrap(); + let mesh = super::super::by_face(rebuilt, &encoded.face_of, encoded.faces(), &[]); + + assert_eq!(mesh.faces, [0..1, 1..2]); + assert_eq!(mesh.triangles, [[3, 4, 5], [0, 1, 2]]); + let expected = [[0.0, 0.0, 0.0], [2.0, 0.0, 0.0], [0.0, 2.0, 0.0], [5.0, 0.0, 0.0], [7.0, 0.0, 0.0], [5.0, 2.0, 0.0]]; + assert_positions(&mesh.positions, &expected); +} diff --git a/crates/ps/src/read/font.rs b/crates/ps/src/read/font.rs index 9a1af3fb6..ec47daf38 100644 --- a/crates/ps/src/read/font.rs +++ b/crates/ps/src/read/font.rs @@ -43,9 +43,27 @@ fn substitute_type42(ip: &mut Interp, font_name: &[u8], bytes: &[u8]) -> R Option { b"oslash" => 'ø', b"oe" => 'Å“', b"germandbls" => 'ß', - _ => return None, + _ => { + return crate::read::encodings::iso_latin1() + .iter() + .enumerate() + .skip(160) + .find(|(_, glyph)| glyph.as_bytes() == name) + .and_then(|(code, _)| char::from_u32(code as u32)); + } }) } @@ -429,7 +454,7 @@ type PerGlyph<'a> = dyn FnMut(&mut Interp, u32, usize) -> R<(f64, f64)> + 'a; /// Show (or, with `draw` false, measure) a string: the total advance in user space. /// `per_char` gives each glyph's extra displacement (or, with `replace`, its whole advance). -fn show_glyphs(ip: &mut Interp, s: &[u8], draw: bool, replace: bool, per_char: &mut PerGlyph<'_>) -> R<(f64, f64)> { +fn show_glyphs(ip: &mut Interp, s: &[u8], draw: bool, replace: bool, per_char: &mut PerGlyph<'_>, text: Option<&str>) -> R<(f64, f64)> { let font = match ip.gs.font.clone() { Some(f) => f, None => return err("invalidfont"), @@ -443,6 +468,12 @@ fn show_glyphs(ip: &mut Interp, s: &[u8], draw: bool, replace: bool, per_char: & let fm = font_matrix(&font); s.iter().map(|c| (u32::from(*c), font.clone(), fm, crate::read::composite::Key2::Code(*c))).collect() }; + // Glyph procedures can issue arbitrary paint and clip operations. The selected jobs, not + // the composite font's family, determine whether this show can have a synthetic wrapper. + let marked = text.filter(|_| !jobs.iter().any(|(_, leaf, _, _)| leaf.get_name(b"FontType").and_then(|v| v.num()) == Some(3.0))); + if let Some(text) = marked { + ip.out.begin_actual_text(text, &ip.gs.clip).map_err(PsErr::Fatal)?; + } let mut total = (0.0, 0.0); let mut text_paths: Vec<(Vec, crate::read::gstate::DevColor)> = Vec::new(); let mut text_segments = 0usize; @@ -547,11 +578,18 @@ fn show_glyphs(ip: &mut Interp, s: &[u8], draw: bool, replace: bool, per_char: & Ok(total) })(); let path_limit = matches!(&result, Err(PsErr::Fatal(message)) if message == "PostScript (path limit)"); - if (result.is_ok() || !path_limit) && !ip.gs.nulldev { - for (path, color) in text_paths { - ip.out.fill(&ip.gs.clip, &path, false, color).map_err(PsErr::Fatal)?; + let paint_result = (|| { + if (result.is_ok() || !path_limit) && !ip.gs.nulldev { + for (path, color) in text_paths { + ip.out.fill(&ip.gs.clip, &path, false, color).map_err(PsErr::Fatal)?; + } } + Ok(()) + })(); + if marked.is_some() { + ip.out.end_actual_text(); } + paint_result?; result } @@ -568,17 +606,8 @@ fn show_string(ip: &mut Interp, s: &[u8], draw: bool, per_char: &mut PerGlyph<'_ // Whitespace-only shows advance the text position but emit no painted glyphs; avoid bloating // the content stream with an empty semantic marked-content wrapper for them. let text = ip.gs.font.as_ref().and_then(|font| actual_text(font, s)); - if draw - && !ip.charpath - && let Some(text) = text.filter(|text| text.chars().any(|c| !c.is_whitespace())) - { - ip.out.begin_actual_text(&text); - let result = show_glyphs(ip, s, draw, false, per_char); - ip.out.end_actual_text(); - result - } else { - show_glyphs(ip, s, draw, false, per_char) - } + let text = text.filter(|text| draw && !ip.charpath && text.chars().any(|c| !c.is_whitespace())); + show_glyphs(ip, s, draw, false, per_char, text.as_deref()) } pub fn show(ip: &mut Interp) -> R { @@ -627,14 +656,21 @@ fn explicit_show(ip: &mut Interp, xs: bool, ys: bool) -> R { return err("nocurrentpoint"); } let step = if xs && ys { 2 } else { 1 }; - show_glyphs(ip, &s, true, true, &mut |_, _, i| { - let d = |k: usize| disp.get(i * step + k).copied().unwrap_or(0.0); - Ok(match (xs, ys) { - (true, true) => (d(0), d(1)), - (true, false) => (d(0), 0.0), - _ => (0.0, d(0)), - }) - }) + show_glyphs( + ip, + &s, + true, + true, + &mut |_, _, i| { + let d = |k: usize| disp.get(i * step + k).copied().unwrap_or(0.0); + Ok(match (xs, ys) { + (true, true) => (d(0), d(1)), + (true, false) => (d(0), 0.0), + _ => (0.0, d(0)), + }) + }, + None, + ) .map(|_| ()) } @@ -747,5 +783,43 @@ mod tests { assert_eq!(standard_glyph_char(b"quoteleft"), Some('‘')); assert_eq!(standard_glyph_char(b"quoteright"), Some('’')); assert_eq!(standard_glyph_char(b"unmapped.custom"), None); + assert_eq!(standard_glyph_char(b"eacute"), Some('é')); + assert_eq!(standard_glyph_char(b"Adieresis"), Some('Ä')); + assert_eq!(standard_glyph_char(b"space"), Some(' ')); + } + /// A tiny generated SFNT directory with named glyphs, not a copied font asset. + fn named_font() -> Vec { + let maxp = [0, 0, 0x50, 0, 0, 3]; + let mut post = vec![0u8; 32]; + post[..4].copy_from_slice(&[0, 2, 0, 0]); + post.extend_from_slice(&[0, 3, 0, 0, 1, 2, 1, 3]); + for n in [b"eacute".as_slice(), b"custom.accent"] { + post.push(n.len() as u8); + post.extend_from_slice(n); + } + let mut data = vec![0, 1, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0]; + for (tag, offset, len) in [(b"maxp", 44u32, maxp.len()), (b"post", 52u32, post.len())] { + data.extend_from_slice(tag); + data.extend_from_slice(&0u32.to_be_bytes()); + data.extend_from_slice(&offset.to_be_bytes()); + data.extend_from_slice(&(len as u32).to_be_bytes()); + } + data.extend_from_slice(&maxp); + data.resize(52, 0); + data.extend_from_slice(&post); + data + } + + #[test] + fn substitute_resolves_reencoded_names_and_charges_budget() -> Result<(), String> { + let bytes = named_font(); + let mut ip = Interp::new(vec![], (0.0, 0.0), Default::default()); + let font = substitute_type42(&mut ip, b"Fallback", &bytes).map_err(|e| format!("{e:?}"))?.ok_or("fixture font rejected")?; + assert_eq!(crate::read::type42::glyph_index(&font, b"eacute"), 1); + assert_eq!(crate::read::type42::glyph_index(&font, b"custom.accent"), 2); + assert_eq!(crate::read::type42::glyph_index(&font, b"unknown"), 0); + let mut limited = Interp::new(vec![], (0.0, 0.0), crate::read::Limits { steps: 0, ..Default::default() }); + assert!(substitute_type42(&mut limited, b"Fallback", &bytes).is_err()); + Ok(()) } } diff --git a/crates/ps/src/read/pdf.rs b/crates/ps/src/read/pdf.rs index 0faabe09a..3231dfaa4 100644 --- a/crates/ps/src/read/pdf.rs +++ b/crates/ps/src/read/pdf.rs @@ -153,8 +153,12 @@ impl Out { } /// Preserve the semantic text for consumers that group outlined glyphs. - pub fn begin_actual_text(&mut self, text: &str) { + pub fn begin_actual_text(&mut self, text: &str, clip: &Option>) -> Result<(), String> { + // A clip can outlive the show. Establish it before the marked span so EMC does not + // accidentally close the graphics state that later painting still needs. + self.sync_clip(clip)?; self.c.begin_marked_content_with_properties(Name(b"Span")).properties().actual_text(pdf_writer::TextStr(text)); + Ok(()) } pub fn end_actual_text(&mut self) { diff --git a/crates/ps/src/read/tests.rs b/crates/ps/src/read/tests.rs index 152041659..c694a7177 100644 --- a/crates/ps/src/read/tests.rs +++ b/crates/ps/src/read/tests.rs @@ -137,6 +137,7 @@ fn type3_fonts_draw_their_glyphs() { /T3 T3 definefont pop /T3 findfont 20 scalefont setfont 10 20 moveto (AA) show"); // Two glyphs, the second 12 points (600 units at 20 pt) to the right. assert_eq!(c.lines().filter(|l| *l == "f").count(), 2, "{c}"); + assert!(!c.contains("/ActualText"), "Type 3 procedures are not synthetic text spans: {c}"); assert!(c.contains("22 20 m"), "{c}"); } @@ -471,6 +472,28 @@ fn actual_text_uses_the_selected_encoding_and_omits_unknown_names() { assert!(!content.contains("/ActualText"), "{content}"); } +#[test] +fn composite_show_only_omits_text_for_selected_type3_glyphs() { + let setup = "/T3 8 dict dup begin /FontType 3 def /FontMatrix [0.001 0 0 0.001 0 0] def /FontBBox [0 0 1000 1000] def /Encoding 256 array def 0 1 255 { Encoding exch /.notdef put } for Encoding 65 /A put /BuildChar { pop pop 600 0 0 0 500 500 setcachedevice 0 0 500 500 rectfill } def end def /T3 T3 definefont pop /T0 8 dict dup begin /FontType 0 def /FMapType 2 def /FontMatrix [1 0 0 1 0 0] def /Encoding 256 array def 0 1 255 { Encoding exch /A put } for /FDepVector [ /Helvetica findfont /T3 findfont ] def end /T0 exch definefont pop /T0 findfont 20 scalefont setfont 10 20 moveto "; + let (standard, _) = run(&format!("{setup} <0041> show")); + assert!(standard.contains("/ActualText (AA)"), "{standard}"); + let (mixed, _) = run(&format!("{setup} <00410141> show")); + assert!(!mixed.contains("/ActualText"), "{mixed}"); + assert!(mixed.lines().any(|line| line == "f"), "the Type 3 glyph is still painted: {mixed}"); +} + +#[test] +fn actual_text_starts_after_persistent_clip() { + let (content, _) = run( + "newpath 0 0 moveto 50 0 lineto 50 50 lineto closepath clip newpath /Helvetica findfont 12 scalefont setfont 10 10 moveto (Hi) show 0 0 100 100 rectfill", + ); + let clip = content.find("\nW\n").unwrap(); + let span = content.find("/ActualText (Hi)").unwrap(); + let end = content.find("EMC").unwrap(); + let restore = content.rfind("\nQ").unwrap(); + assert!(clip < span && span < end && end < restore, "{content}"); +} + #[test] fn text_paint_keeps_completed_glyphs_when_a_later_glyph_hits_the_work_limit() { let ttf = tiny_ttf(); diff --git a/crates/psd/src/channel_resources.rs b/crates/psd/src/channel_resources.rs new file mode 100644 index 000000000..204d36eb0 --- /dev/null +++ b/crates/psd/src/channel_resources.rs @@ -0,0 +1,111 @@ +//! Raw PSD channel DisplayInfo (1077/1007) and Alternate Spot Colors (1067) records. +//! +//! Ported from PhotoCraft channel_map: Jason D's f7f8fdcd89 DisplayInfo framing and +//! Khawar Zamman Wani's 9a390661b5 alternate-spot reader. Colour interpretation, opacity +//! clamping, channel matching and legacy compatibility decisions belong to the caller. +//! Readers borrow their input and ignore a trailing partial record, preserving the +//! original tolerant import behavior. Version words are framing, not validation here. + +/// One 13-byte DisplayInfo record, without document-specific colour interpretation. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub struct DisplayInfo { + /// PSD colour-space identifier. + pub space: u16, + /// Four raw colour components. + pub components: [u16; 4], + /// Raw opacity (normally 0 through 100). + pub opacity: u16, + /// Raw channel kind (0 selected, 1 masked, 2 spot). + pub kind: u8, +} + +impl DisplayInfo { + /// Encode the version-1 resource's 13-byte record, without clamping fields. + pub fn to_bytes(self) -> [u8; 13] { + let mut out = [0; 13]; + let values = [self.space, self.components[0], self.components[1], self.components[2], self.components[3], self.opacity]; + for (slot, value) in out[..12].as_chunks_mut::<2>().0.iter_mut().zip(values) { + *slot = value.to_be_bytes(); + } + out[12] = self.kind; + out + } +} + +/// Iterate complete DisplayInfo records without allocating. +/// +/// `versioned` selects resource 1077 (skip four version bytes, 13-byte records), +/// otherwise resource 1007 (14-byte records, ignoring each padding byte). A short +/// version header or incomplete final record yields no corresponding record. +pub fn display_info(data: &[u8], versioned: bool) -> impl ExactSizeIterator + '_ { + let (body, size) = if versioned { (data.get(4..).unwrap_or(&[]), 13) } else { (data, 14) }; + body.chunks_exact(size).map(|rec| { + let u = |i: usize| u16::from_be_bytes([rec[i], rec[i + 1]]); + DisplayInfo { space: u(0), components: [u(2), u(4), u(6), u(8)], opacity: u(10), kind: rec[12] } + }) +} + +/// One Alternate Spot Colors record, retaining unknown colour spaces. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub struct AlternateSpotColor { + /// Channel identifier from resource 1053. + pub id: u32, + /// PSD colour-space identifier. + pub space: u16, + /// Four raw colour components. + pub components: [u16; 4], +} + +/// Iterate available complete 14-byte records up to the resource's declared u16 count. +/// +/// Short headers, absent declared records and trailing partial bytes are tolerated exactly +/// as in the original importer. No allocation is derived from an untrusted count. +pub fn alternate_spot_colors(data: &[u8]) -> impl ExactSizeIterator + '_ { + let count = data.get(2..4).map_or(0, |c| usize::from(u16::from_be_bytes([c[0], c[1]]))); + data.get(4..).unwrap_or_default().as_chunks::<14>().0.iter().take(count).map(|rec| { + let u = |i: usize| u16::from_be_bytes([rec[i], rec[i + 1]]); + AlternateSpotColor { id: u32::from_be_bytes([rec[0], rec[1], rec[2], rec[3]]), space: u(4), components: [u(6), u(8), u(10), u(12)] } + }) +} + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn display_round_trip_and_truncation() { + let record = DisplayInfo { space: 0x2331, components: [0, 1, 255, u16::MAX], opacity: 500, kind: 255 }; + let mut modern = 99u32.to_be_bytes().to_vec(); + modern.extend(record.to_bytes()); + assert_eq!(display_info(&modern, true).collect::>(), [record]); + for n in 0..modern.len() { + assert_eq!(display_info(&modern[..n], true).len(), 0); + } + modern.extend([1, 2, 3]); + assert_eq!(display_info(&modern, true).len(), 1); + let mut old = record.to_bytes().to_vec(); + old.push(42); + old.extend(record.to_bytes()); + old.push(99); + assert_eq!(display_info(&old, false).collect::>(), [record, record]); + assert_eq!(display_info(&old[..27], false).len(), 1); + } + + #[test] + fn alternate_count_is_bounded_by_available_records() { + let mut bytes = vec![0, 9, 255, 255]; + bytes.extend(123u32.to_be_bytes()); + for value in [7u16, 10000, 0xffff, 4, 5] { + bytes.extend(value.to_be_bytes()); + } + let want = AlternateSpotColor { id: 123, space: 7, components: [10000, 0xffff, 4, 5] }; + assert_eq!(alternate_spot_colors(&bytes).collect::>(), [want]); + for n in 0..bytes.len() { + assert_eq!(alternate_spot_colors(&bytes[..n]).len(), 0); + } + bytes.extend([1, 2, 3]); + assert_eq!(alternate_spot_colors(&bytes).len(), 1); + bytes[2..4].copy_from_slice(&0u16.to_be_bytes()); + assert_eq!(alternate_spot_colors(&bytes).len(), 0); + } +} diff --git a/crates/psd/src/duotone.rs b/crates/psd/src/duotone.rs new file mode 100644 index 000000000..6c9ae49e1 --- /dev/null +++ b/crates/psd/src/duotone.rs @@ -0,0 +1,169 @@ +//! Raw records in Photoshop's version 1 Duotone Color Mode Data. +//! +//! Layout extracted from PhotoCraft `duotone_map.rs`, originally by Gert Pellin +//! (772cf39c09930229e11b920196f486239bbbfab8), with colour-book integration by David Mann. +//! Colour conversion, curve sampling and fallback/display policy are caller-owned. +//! Parsing accepts the 412-byte record prefix as PhotoCraft did; dot gain and overprints +//! are optional trailing data, not prerequisites for interpreting the inks. + +/// Full block size emitted by Photoshop and PhotoCraft. +pub const BLOCK_LEN: usize = 524; +/// Alternate Duotone Colors image resource. +pub const ALTERNATE_DUOTONE_COLORS: u16 = 1066; +/// Raw colour space and four component words, without colour interpretation. +pub type ColorRecord = [u16; 5]; +/// Sentinel for an absent ink or computed overprint colour. +pub const NO_COLOR: ColorRecord = [0xffff, 0, 0, 0, 0]; + +/// Valid active ink counts for a serialized block. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +#[repr(u16)] +pub enum InkCount { + /// One active ink. + One = 1, + /// Two active inks. + Two = 2, + /// Three active inks. + Three = 3, + /// Four active inks. + Four = 4, +} + +/// One fixed-size ink slot, including uninterpreted reserved curve word. +#[derive(Clone, Debug, PartialEq, Eq)] +pub struct InkRecord { + /// Raw ink colour space and components. + pub color: ColorRecord, + /// Pascal name bytes, including its length byte; text encoding belongs to the caller. + pub name: [u8; 64], + /// Thirteen signed curve samples and the reserved final word. + pub curve: [i16; 14], +} + +fn word(data: &[u8], at: usize) -> Option { + let bytes = data.get(at..at.checked_add(2)?)?; + Some(u16::from_be_bytes([bytes[0], bytes[1]])) +} + +/// Decode a raw colour record at a byte offset, without guessing unknown colour spaces. +pub fn color_at(data: &[u8], at: usize) -> Option { + let bytes = data.get(at..at.checked_add(10)?)?; + Some(std::array::from_fn(|i| u16::from_be_bytes([bytes[i * 2], bytes[i * 2 + 1]]))) +} + +/// Borrowed, validated view over at most four active ink records. No input-sized allocation. +#[derive(Clone, Copy, Debug)] +pub struct DuotoneView<'a> { + data: &'a [u8], + count: usize, +} + +impl<'a> DuotoneView<'a> { + /// Accept version 1, one to four inks and the full 412-byte record prefix. + pub fn parse(data: &'a [u8]) -> Option { + if data.len() < 412 || word(data, 0)? != 1 { + return None; + } + let count = usize::from(word(data, 2)?); + (1..=4).contains(&count).then_some(Self { data, count }) + } + + /// Number of active ink records (one through four). + pub fn ink_count(self) -> usize { + self.count + } + + /// Copy one small fixed-size record. Inactive/out-of-range slots are not returned. + pub fn ink(self, index: usize) -> Option { + if index >= self.count { + return None; + } + let name_at = 44 + 64 * index; + let name = self.data.get(name_at..name_at + 64)?.try_into().ok()?; + let curve_at = 300 + 28 * index; + let curve_bytes = self.data.get(curve_at..curve_at + 28)?; + Some(InkRecord { + color: color_at(self.data, 4 + 10 * index)?, + name, + curve: std::array::from_fn(|i| i16::from_be_bytes([curve_bytes[i * 2], curve_bytes[i * 2 + 1]])), + }) + } +} + +/// Read one raw alternate colour when the resource version/count permits it. +pub fn alternate_color(data: &[u8], index: usize) -> Option { + if word(data, 0)? != 1 || index >= usize::from(word(data, 2)?) { + return None; + } + color_at(data, 4usize.checked_add(index.checked_mul(10)?)?) +} + +/// Serialize four raw slots and explicit trailing policy into a fixed-size block. +/// The typed count cannot be invalid. Reserved bytes and inactive slots are preserved as supplied. +pub fn write_block(count: InkCount, inks: &[InkRecord; 4], dot_gain: u16, overprints: &[ColorRecord; 11]) -> [u8; BLOCK_LEN] { + let mut out = [0u8; BLOCK_LEN]; + out[..2].copy_from_slice(&1u16.to_be_bytes()); + out[2..4].copy_from_slice(&(count as u16).to_be_bytes()); + for (i, ink) in inks.iter().enumerate() { + for (j, value) in ink.color.iter().enumerate() { + let at = 4 + 10 * i + 2 * j; + out[at..at + 2].copy_from_slice(&value.to_be_bytes()); + } + out[44 + 64 * i..44 + 64 * (i + 1)].copy_from_slice(&ink.name); + for (j, value) in ink.curve.iter().enumerate() { + let at = 300 + 28 * i + 2 * j; + out[at..at + 2].copy_from_slice(&value.to_be_bytes()); + } + } + out[412..414].copy_from_slice(&dot_gain.to_be_bytes()); + for (i, color) in overprints.iter().enumerate() { + for (j, value) in color.iter().enumerate() { + let at = 414 + 10 * i + 2 * j; + out[at..at + 2].copy_from_slice(&value.to_be_bytes()); + } + } + out +} + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn fixed_layout_roundtrip_preserves_raw_fields() { + let mut ink = InkRecord { color: [7, 1, 2, 3, 65535], name: [0; 64], curve: [-1; 14] }; + ink.name[..4].copy_from_slice(&[3, b'I', b'n', b'k']); + ink.curve[0] = 0; + ink.curve[12] = 1000; + let slots = std::array::from_fn(|_| ink.clone()); + let bytes = write_block(InkCount::Four, &slots, 20, &[NO_COLOR; 11]); + assert_eq!(&bytes[..4], &[0, 1, 0, 4]); + assert_eq!(&bytes[44..48], &[3, b'I', b'n', b'k']); + assert_eq!(&bytes[412..416], &[0, 20, 255, 255]); + let view = DuotoneView::parse(&bytes); + assert_eq!(view.map(DuotoneView::ink_count), Some(4)); + for i in 0..4 { + assert_eq!(view.and_then(|v| v.ink(i)), Some(ink.clone())); + } + assert!(view.and_then(|v| v.ink(usize::MAX)).is_none()); + for n in 0..412 { + assert!(DuotoneView::parse(&bytes[..n]).is_none()); + } + assert!(DuotoneView::parse(&bytes[..412]).is_some()); + } + + #[test] + fn rejects_hostile_offsets_versions_counts_and_alternate_truncation() { + let mut bytes = [0u8; BLOCK_LEN]; + bytes[..4].copy_from_slice(&[0, 1, 255, 255]); + assert!(DuotoneView::parse(&bytes).is_none()); + bytes[..4].copy_from_slice(&[0, 2, 0, 1]); + assert!(DuotoneView::parse(&bytes).is_none()); + assert!(color_at(&bytes, usize::MAX).is_none()); + let alt = [0, 1, 0, 1, 0, 7, 0, 1, 0, 2, 0, 3, 0, 4]; + assert_eq!(alternate_color(&alt, 0), Some([7, 1, 2, 3, 4])); + assert!(alternate_color(&alt[..13], 0).is_none()); + assert!(alternate_color(&alt, 1).is_none()); + assert!(alternate_color(&alt, usize::MAX).is_none()); + } +} diff --git a/crates/psd/src/guides.rs b/crates/psd/src/guides.rs new file mode 100644 index 000000000..5f100c480 --- /dev/null +++ b/crates/psd/src/guides.rs @@ -0,0 +1,103 @@ +//! PSD Grid and Guides version 2 (resource 1092) record decoding. +//! +//! Extracted from PhotoCraft `guides_map.rs` at f49564d281bf8892db58ed154767717223aff4f6 +//! by Bram Heerink. Document colour preference and resource-retention policy belong to callers. +//! Unknown direction values retain PhotoCraft's vertical interpretation. Grid cycles, reserved +//! bytes and trailing data are not interpreted. Resource 1097 remains opaque except its count. + +/// Grid and Guides version 2 resource ID. +pub const GUIDES_V2: u16 = 1092; +/// Associated guide identifiers resource ID. +pub const GUIDE_IDS: u16 = 1097; + +/// One version 2 guide record. +#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord)] +pub struct GuideV2 { + /// Signed position in units of 1/32 pixel. + pub location: i32, + /// True for vertical guides (direction byte other than 1). + pub vertical: bool, + /// Stored RGB colour; interpretation of preference colours is app policy. + pub rgb: [u8; 3], +} + +fn be_u32(data: &[u8], at: usize) -> Option { + let b = data.get(at..at.checked_add(4)?)?; + Some(u32::from_be_bytes([b[0], b[1], b[2], b[3]])) +} + +/// Validate the full declared record region before returning an allocation-free iterator. +/// Returns `None` for an unsupported version, short header or truncated records. +/// Record count cannot cause allocation or work beyond the supplied byte slice. +pub fn parse_v2(data: &[u8]) -> Option + '_> { + if be_u32(data, 0)? != 2 { + return None; + } + let count = usize::try_from(be_u32(data, 12)?).ok()?; + let end = 16usize.checked_add(count.checked_mul(9)?)?; + let records = data.get(16..end)?; + Some(records.as_chunks::<9>().0.iter().map(|g| GuideV2 { + location: i32::from_be_bytes([g[0], g[1], g[2], g[3]]), + vertical: g[4] != 1, + rgb: [g[6], g[7], g[8]], + })) +} + +/// Only the declared entry count in resource 1097; this does not validate its GUID payload. +pub fn ids_count(data: &[u8]) -> Option { + be_u32(data, 0) +} + +#[cfg(test)] +mod tests { + use super::*; + + fn fixture() -> Vec { + let mut v = Vec::new(); + for n in [2u32, 576, 576, 2] { + v.extend_from_slice(&n.to_be_bytes()); + } + for (position, direction, rgb) in [(512i32, 0u8, [74, 255, 255]), (-32, 1, [1, 2, 3])] { + v.extend_from_slice(&position.to_be_bytes()); + v.extend_from_slice(&[direction, 0]); + v.extend_from_slice(&rgb); + } + v + } + + #[test] + fn decodes_signed_positions_directions_and_colours() { + let data = fixture(); + let actual = parse_v2(&data).map(Iterator::collect::>); + assert_eq!( + actual, + Some(vec![GuideV2 { location: 512, vertical: true, rgb: [74, 255, 255] }, GuideV2 { location: -32, vertical: false, rgb: [1, 2, 3] },]) + ); + } + + #[test] + fn validates_every_truncation_and_hostile_count_before_iteration() { + let mut data = fixture(); + for n in 0..data.len() { + assert!(parse_v2(&data[..n]).is_none()); + } + data[12..16].copy_from_slice(&u32::MAX.to_be_bytes()); + assert!(parse_v2(&data).is_none()); + data[12..16].copy_from_slice(&0u32.to_be_bytes()); + assert_eq!(parse_v2(&data).map(|v| v.len()), Some(0)); + data[3] = 1; + assert!(parse_v2(&data).is_none()); + } + + #[test] + fn preserves_unknown_direction_reserved_bytes_and_trailing_data_policy() { + let mut data = fixture(); + data[20] = 255; + data[21] = 42; + data.extend_from_slice(&[9; 12]); + assert_eq!(parse_v2(&data).and_then(|mut v| v.next()).map(|g| g.vertical), Some(true)); + assert_eq!(parse_v2(&data).map(|v| v.len()), Some(2)); + assert_eq!(ids_count(&[0, 0, 0, 2, 255]), Some(2)); + assert_eq!(ids_count(&[0, 0, 0]), None); + } +} diff --git a/crates/psd/src/lib.rs b/crates/psd/src/lib.rs index 4c3222f46..efcc474b2 100644 --- a/crates/psd/src/lib.rs +++ b/crates/psd/src/lib.rs @@ -60,11 +60,13 @@ pub mod asl; pub mod atn; pub mod blend; pub mod builder; +pub mod channel_resources; mod codepage; pub mod compression; pub mod csh; pub mod descriptor; pub mod displace_map; +pub mod duotone; pub mod engine_data; pub mod error; pub mod file; @@ -74,6 +76,7 @@ pub mod fixtures; pub mod flat; pub mod float; pub mod grd; +pub mod guides; pub mod hdr; pub mod head; pub mod header; @@ -101,6 +104,7 @@ pub mod testgen; pub mod text; pub mod tiff; pub mod tree; +pub mod unicode_names; pub use blend::BlendMode; pub use builder::{GroupSpec, LayerSpec, MaskSpec, PixelData, PsdBuilder}; diff --git a/crates/psd/src/unicode_names.rs b/crates/psd/src/unicode_names.rs new file mode 100644 index 000000000..c4547243b --- /dev/null +++ b/crates/psd/src/unicode_names.rs @@ -0,0 +1,54 @@ +//! Counted UTF-16 name records, as used by PSD Unicode Alpha Names (resource 1045). +//! +//! Adapted from PhotoCraft's channel-name reader. Framing is shared; decoding malformed +//! UTF-16, removing terminators and assigning names to document channels remain caller policy. + +/// Iterate complete records as borrowed big-endian UTF-16 byte slices. +/// +/// A record starts with a big-endian `u32` code-unit count. Empty records are retained. +/// An incomplete count, truncated record or arithmetic overflow ends the list, preserving +/// preceding complete records. No allocation is driven by the declared count, and bytes +/// after a malformed record are not interpreted as new records. +pub fn records(mut data: &[u8]) -> impl Iterator { + std::iter::from_fn(move || { + let count = data.get(..4)?; + let count = u32::from_be_bytes([count[0], count[1], count[2], count[3]]); + let input = data; + // On any malformed length the iterator stays exhausted on subsequent calls. + data = &[]; + let len = usize::try_from(count).ok()?.checked_mul(2)?; + let end = 4usize.checked_add(len)?; + let record = input.get(4..end)?; + data = input.get(end..)?; + Some(record) + }) +} + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn borrows_empty_and_uninterpreted_code_units() { + let data = [0, 0, 0, 0, 0, 0, 0, 3, 0xd8, 0, 0, 65, 0, 0]; + let found: Vec<_> = records(&data).collect(); + assert_eq!(found, [&data[4..4], &data[8..14]]); + assert_eq!(found[1].as_ptr(), data[8..].as_ptr()); + } + + #[test] + fn partial_tail_and_hostile_count_stop_without_losing_prefix() { + let first = [0, 0, 0, 1, 0, 65]; + for tail in [&[0][..], &[0, 0, 0], &[0, 0, 0, 2, 0, 66], &[255, 255, 255, 255, 0, 66]] { + let mut bytes = first.to_vec(); + bytes.extend(tail); + let mut it = records(&bytes); + assert_eq!(it.next(), Some(&first[4..])); + assert_eq!(it.next(), None); + assert_eq!(it.next(), None); + } + for end in 0..first.len() { + assert_eq!(records(&first[..end]).count(), 0); + } + } +} diff --git a/crates/raw/src/dngwrite.rs b/crates/raw/src/dngwrite.rs index 8953f94ea..ace0c07f6 100644 --- a/crates/raw/src/dngwrite.rs +++ b/crates/raw/src/dngwrite.rs @@ -13,6 +13,7 @@ pub enum DngCompression { Uncompressed, /// Lossless JPEG tiles of `tile × tile` pixels (`tile` even). CFA tiles are encoded as two interleaved /// components of half the tile width (the common DNG layout), linear data with one component per sample. + /// Edges above 2048 (after even alignment) are rejected before padding the tile. Lj92 { tile: u32, }, @@ -35,6 +36,11 @@ const DEFLATE_LEVEL: u8 = 1; /// above the 256 default and every real tile this writer makes. const MAX_DEFLATE_TILE: u32 = 2048; +// Match the existing padded Deflate edge budget. JPEG's 16-bit dimension limit alone would permit +// a 65534²-sample allocation even for a tiny source image. At most four validated channels make +// this bound at most 32 MiB of u16 padding per tile, before encoder scratch/output. +const MAX_LJ92_TILE: usize = MAX_DEFLATE_TILE as usize; + /// `tile` aligned down to a multiple of 16 and clamped to [`MAX_DEFLATE_TILE`]. fn deflate_tile(tile: u32) -> usize { (tile.max(16) & !15).min(MAX_DEFLATE_TILE) as usize @@ -270,7 +276,7 @@ pub fn write_dng(raw: &RawImage, opts: &DngWriteOptions) -> Result> { } DngCompression::Lj92 { tile } => { let tile = (tile.max(2) & !1) as usize; - if tile > 65534 { + if tile > MAX_LJ92_TILE { return Err(RawError::Limit("tile too large")); } ifd.set(t::COMPRESSION, Value::Short(vec![compression::JPEG])); @@ -432,6 +438,58 @@ mod tests { use super::*; use crate::{BlackLevel, OpcodeLists, RawData, RawFormat, RawImage, Rect}; + fn tiny_raw() -> RawImage { + let (w, h) = (8usize, 8usize); + let full = Rect::new(0, 0, w, h); + RawImage { + format: RawFormat::Dng, + width: w, + height: h, + cpp: 1, + data: RawData::U16((0..64).map(|i| i as u16 * 129).collect()), + cfa: None, + bits: 14, + black: BlackLevel::uniform(0.0), + white: vec![16383.0], + active_area: full, + crop: full, + orientation: crate::Orientation::Normal, + color: crate::ColorData::default(), + wb_multipliers: None, + linearized: false, + pixel_aspect: 1.0, + opcodes: OpcodeLists::default(), + camera: Default::default(), + cinema: Default::default(), + lens_metadata: Default::default(), + level_sources: Default::default(), + notices: crate::RawNotices::default(), + } + } + + #[test] + fn oversized_lj92_padding_is_rejected() { + let raw = tiny_raw(); + // The near-boundary case previously completed with an oversized padded tile; keep it first + // so reverting the guard fails without trying the multi-gigabyte cases. + for tile in [2050, 65534, u32::MAX] { + let opts = DngWriteOptions { compression: DngCompression::Lj92 { tile }, ..Default::default() }; + assert!(matches!(write_dng(&raw, &opts), Err(RawError::Limit("tile too large")))); + } + } + + #[test] + fn default_lj92_tiles_preserve_samples() { + let raw = tiny_raw(); + let bytes = write_dng(&raw, &DngWriteOptions::default()).expect("normal tiles write"); + let back = crate::decode(&bytes).expect("normal tiles decode"); + assert_eq!((back.width, back.height), (raw.width, raw.height)); + match (&back.data, &raw.data) { + (RawData::U16(actual), RawData::U16(expected)) => assert_eq!(actual, expected), + _ => panic!("integer DNG changed sample type"), + } + } + /// Deflate tiles of `tile × tile` pad into a square buffer that grows with the square: a /// caller-supplied edge of `u32::MAX` would make the writer pad an 8 × 8 image into billions of /// edge-replicate iterations and a multi-gigabyte buffer. The clamp keeps the bounded shape — same diff --git a/crates/raw/src/lib.rs b/crates/raw/src/lib.rs index 602ec29b4..4fa225364 100644 --- a/crates/raw/src/lib.rs +++ b/crates/raw/src/lib.rs @@ -97,6 +97,8 @@ pub mod kelvin; mod kodak; pub mod lens_metadata; pub mod ljpeg; +/// Optimized lossless JPEG decoder for streams with line-aligned restarts. +pub mod ljpeg_lines; #[cfg(feature = "arw")] mod llvc; #[cfg(feature = "mef")] @@ -643,6 +645,7 @@ pub fn orientation(bytes: &[u8]) -> Option { } // A damaged/unsupported pixel IFD must not hide usable orientation metadata from preview // fallback. CR3 keeps IFD0 in its CMT1 box; TIFF-based raws keep it at the file start. + #[cfg(feature = "cr3")] if probe(bytes) == Some(RawFormat::Cr3) { return cr3::orientation(bytes); } diff --git a/crates/raw/src/ljpeg_lines.rs b/crates/raw/src/ljpeg_lines.rs new file mode 100644 index 000000000..e169fd893 --- /dev/null +++ b/crates/raw/src/ljpeg_lines.rs @@ -0,0 +1,543 @@ +//! Lossless JPEG decoding: ITU-T T.81 | ISO/IEC 10918-1 Annex H (process 14, Huffman coded, +//! "SOF3"), the compression DNG calls lossless JPEG (Compression = 7). +//! +//! - Frame header (B.2.2): 2–16-bit precision, 1–4 components, each sampled 1×1. +//! - Huffman tables (B.2.4.2, Annex C) decoded with a 9-bit lookup and the F.2.2.3 code-length +//! walk for longer codes. +//! - Scan header (B.2.3): one scan holding every component, predictor `Ss` 1–7 (Table H.1) and +//! point transform `Al`. +//! - Prediction (H.1.2.1): the first line (and the first line after a restart) predicts from the +//! left neighbour with 2^(P − Pt − 1) for its first sample; later lines predict their first +//! sample from the one above and the rest with the selected predictor. Differences are added +//! modulo 2^16 (H.1.2.2) and the SSSS = 16 category is 32768 with no extra bits. +//! - Restart intervals (DRI, RSTn) that span whole lines. + +/// Why a lossless JPEG stream cannot be decoded. +#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)] +pub enum Error { + /// A required byte is missing. + #[error("truncated lossless JPEG")] + Truncated, + /// The sample allocation exceeds the caller's bound. + #[error("lossless JPEG is too large")] + TooLarge, + /// Invalid stream structure. + #[error("{0}")] + Malformed(&'static str), + /// A valid variation outside the line-restart decoder's support. + #[error("{0}")] + Unsupported(String), +} +/// Result of the line-restart decoder. +pub type Result = std::result::Result; + +/// A decoded lossless JPEG image. +#[derive(Debug)] +pub struct Decoded { + pub width: usize, + pub height: usize, + pub components: usize, + pub precision: u8, + pub predictor: u8, + /// Samples, interleaved by component, line after line, shifted back up by the point transform. + pub samples: Vec, +} + +const LOOKUP_BITS: u32 = 9; +/// Bits of the combined code + difference lookup. +const FULL_BITS: u32 = 13; +/// A combined lookup entry: bit 8 set, the bits used in 0–7, the difference in 16–31. +const FULL_VALID: u32 = 0x100; + +/// A Huffman table for decoding (T.81 Annex C and F.2.2.3). +#[derive(Clone)] +struct Huffman { + /// For codes of up to LOOKUP_BITS bits: (length, value) indexed by the next LOOKUP_BITS bits; + /// length 0 means "longer code". + fast: Vec<(u8, u8)>, + maxcode: [i32; 17], + valptr: [i32; 17], + mincode: [i32; 17], + values: Vec, + /// For a code and its extra bits that fit in FULL_BITS together (difference categories + /// 0–15): `FULL_VALID | used bits | difference << 16`, indexed by the next FULL_BITS bits; + /// 0 means "decode the slow way". + full: Vec, +} + +impl Huffman { + fn new(bits: &[u8; 16], values: Vec) -> Result { + // C.2: code sizes and codes + let mut sizes = Vec::new(); + for (l, &n) in bits.iter().enumerate() { + for _ in 0..n { + sizes.push(l as u8 + 1); + } + } + if sizes.len() != values.len() || sizes.is_empty() { + return Err(Error::Malformed("lossless JPEG: Huffman table counts")); + } + let mut codes = Vec::with_capacity(sizes.len()); + let mut code: u32 = 0; + let mut si = sizes[0]; + let mut k = 0; + while k < sizes.len() { + while k < sizes.len() && sizes[k] == si { + codes.push(code); + code += 1; + k += 1; + } + // the codes of one length must fit in that length + if code > (1u32 << si) { + return Err(Error::Malformed("lossless JPEG: Huffman table over-subscribed")); + } + code <<= 1; + si += 1; + } + // F.2.2.3 decoder tables + let mut maxcode = [-1i32; 17]; + let mut valptr = [0i32; 17]; + let mut mincode = [0i32; 17]; + let mut j = 0usize; + for l in 1..=16usize { + let n = usize::from(bits[l - 1]); + if n == 0 { + maxcode[l] = -1; + } else { + valptr[l] = j as i32; + mincode[l] = codes[j] as i32; + j += n; + maxcode[l] = codes[j - 1] as i32; + } + } + let mut fast = vec![(0u8, 0u8); 1 << LOOKUP_BITS]; + for (i, (&c, &s)) in codes.iter().zip(&sizes).enumerate() { + let s = u32::from(s); + if s <= LOOKUP_BITS { + let shift = LOOKUP_BITS - s; + let base = (c << shift) as usize; + for f in fast.iter_mut().skip(base).take(1 << shift) { + *f = (s as u8, values[i]); + } + } + } + let mut full = vec![0u32; 1 << FULL_BITS]; + for (&c, (&s, &ssss)) in codes.iter().zip(sizes.iter().zip(&values)) { + let (s, ssss) = (u32::from(s), u32::from(ssss)); + if ssss >= 16 || s + ssss > FULL_BITS { + continue; + } + let total = s + ssss; + for v in 0..1u32 << ssss { + let diff = match ssss { + 0 => 0, + k if v < 1 << (k - 1) => v as i32 - (1 << k) + 1, + _ => v as i32, + }; + let entry = FULL_VALID | total | (u32::from(diff as i16 as u16) << 16); + let shift = FULL_BITS - total; + let base = (((c << ssss) | v) << shift) as usize; + for f in full.iter_mut().skip(base).take(1 << shift) { + *f = entry; + } + } + } + Ok(Huffman { fast, maxcode, valptr, mincode, values, full }) + } +} + +/// Entropy-coded data reader: removes stuffed zero bytes and stops at markers (feeding zeros). +struct Bits<'a> { + data: &'a [u8], + pos: usize, + acc: u64, + n: u32, + /// The marker that ended the data (after 0xFF), if one was met. + marker: Option, +} + +impl<'a> Bits<'a> { + fn new(data: &'a [u8], pos: usize) -> Self { + Bits { data, pos, acc: 0, n: 0, marker: None } + } + + fn fill(&mut self) { + // fast path: eight bytes without a 0xFF (no stuffing, no marker) go in at once + if self.marker.is_none() + && let Some(b) = self.data.get(self.pos..self.pos + 8) + { + let chunk = u64::from_be_bytes([b[0], b[1], b[2], b[3], b[4], b[5], b[6], b[7]]); + let inv = !chunk; + let has_ff = inv.wrapping_sub(0x0101_0101_0101_0101) & !inv & 0x8080_8080_8080_8080 != 0; + if !has_ff { + let take = (64 - self.n) / 8; + if take > 0 { + self.acc |= (chunk >> (64 - 8 * take)) << (64 - self.n - 8 * take); + self.pos += take as usize; + self.n += 8 * take; + return; + } + } + } + while self.n <= 56 { + let mut byte = 0u8; + if self.marker.is_none() { + match self.data.get(self.pos) { + Some(&0xFF) => match self.data.get(self.pos + 1) { + Some(&0x00) => { + byte = 0xFF; + self.pos += 2; + } + Some(&m) => self.marker = Some(m), + None => self.marker = Some(0xD9), + }, + Some(&b) => { + byte = b; + self.pos += 1; + } + None => self.marker = Some(0xD9), + } + } + self.acc |= u64::from(byte) << (56 - self.n); + self.n += 8; + } + } + + fn peek(&mut self, k: u32) -> u32 { + if self.n < k { + self.fill(); + } + (self.acc >> (64 - k)) as u32 + } + + fn skip(&mut self, k: u32) { + self.acc <<= k; + self.n -= k; + } + + fn bits(&mut self, k: u32) -> u32 { + if k == 0 { + return 0; + } + let v = self.peek(k); + self.skip(k); + v + } + + /// Decode one Huffman symbol (F.2.2.3). + fn decode(&mut self, h: &Huffman) -> Result { + let p = self.peek(LOOKUP_BITS) as usize; + if let Some(&(len, v)) = h.fast.get(p) + && len > 0 + { + self.skip(u32::from(len)); + return Ok(v); + } + let mut code = self.bits(1) as i32; + for l in 1..=16usize { + if code <= h.maxcode[l] { + let i = h.valptr[l] + code - h.mincode[l]; + return h.values.get(i as usize).copied().ok_or(Error::Malformed("lossless JPEG: bad Huffman code")); + } + code = (code << 1) | self.bits(1) as i32; + } + Err(Error::Malformed("lossless JPEG: bad Huffman code")) + } + + /// Skip to the byte after the next RSTn marker, starting a new bit stream there. + fn restart(&mut self) -> Result<()> { + // the decoder may have looked ahead into fill bits; find the marker in the bytes + let mut p = self.pos; + loop { + match (self.data.get(p), self.data.get(p + 1)) { + (Some(0xFF), Some(&m)) if (0xD0..=0xD7).contains(&m) => { + *self = Bits::new(self.data, p + 2); + return Ok(()); + } + (Some(_), Some(_)) => p += 1, + _ => return Err(Error::Malformed("lossless JPEG: missing restart marker")), + } + } + } +} + +struct Component { + id: u8, + table: usize, +} + +/// Decode a lossless JPEG stream. `max_samples` caps the picture (width × height × components). +pub fn decode(data: &[u8], max_samples: usize) -> Result { + if data.get(0..2) != Some(&[0xFF, 0xD8]) { + return Err(Error::Malformed("lossless JPEG: no SOI marker")); + } + let mut pos = 2usize; + let mut tables: [Option; 4] = [None, None, None, None]; + let mut frame: Option<(u8, usize, usize, Vec)> = None; + let mut restart_interval = 0usize; + loop { + // markers may be preceded by fill bytes 0xFF + while data.get(pos) == Some(&0xFF) && data.get(pos + 1) == Some(&0xFF) { + pos += 1; + } + let (Some(&0xFF), Some(&m)) = (data.get(pos), data.get(pos + 1)) else { + return Err(Error::Malformed("lossless JPEG: expected a marker")); + }; + pos += 2; + if m == 0xD9 { + return Err(Error::Malformed("lossless JPEG: no scan")); + } + if (0xD0..=0xD8).contains(&m) || m == 0x01 { + continue; // markers without a segment + } + let len_b = data.get(pos..pos + 2).ok_or(Error::Truncated)?; + let len = usize::from(u16::from_be_bytes([len_b[0], len_b[1]])); + if len < 2 { + return Err(Error::Malformed("lossless JPEG: segment length")); + } + let seg = data.get(pos + 2..pos + len).ok_or(Error::Truncated)?; + match m { + 0xC3 => { + let [p, y0, y1, x0, x1, nf, rest @ ..] = seg else { + return Err(Error::Malformed("lossless JPEG: short frame header")); + }; + let (precision, nf) = (*p, usize::from(*nf)); + let height = usize::from(u16::from_be_bytes([*y0, *y1])); + let width = usize::from(u16::from_be_bytes([*x0, *x1])); + if !(2..=16).contains(&precision) { + return Err(Error::Unsupported(format!("lossless JPEG precision {precision}"))); + } + if height == 0 { + return Err(Error::Unsupported("lossless JPEG with the height in a DNL marker".into())); + } + if width == 0 || !(1..=4).contains(&nf) || rest.len() < nf * 3 { + return Err(Error::Malformed("lossless JPEG: frame header")); + } + let mut comps = Vec::with_capacity(nf); + for c in rest.as_chunks::<3>().0.iter().take(nf) { + if c[1] != 0x11 { + return Err(Error::Unsupported("subsampled lossless JPEG components".into())); + } + comps.push(Component { id: c[0], table: 0 }); + } + frame = Some((precision, width, height, comps)); + } + 0xC4 => { + let mut s = seg; + while !s.is_empty() { + let [tcth, counts @ ..] = s else { break }; + let bits: [u8; 16] = counts.get(..16).and_then(|b| b.try_into().ok()).ok_or(Error::Truncated)?; + let n: usize = bits.iter().map(|&b| usize::from(b)).sum(); + let values = counts.get(16..16 + n).ok_or(Error::Truncated)?.to_vec(); + let th = usize::from(tcth & 0x0F); + if tcth >> 4 != 0 || th > 3 { + return Err(Error::Malformed("lossless JPEG: Huffman table class or id")); + } + if values.iter().any(|&v| v > 16) { + return Err(Error::Malformed("lossless JPEG: difference category above 16")); + } + tables[th] = Some(Huffman::new(&bits, values)?); + s = counts.get(16 + n..).unwrap_or(&[]); + } + } + 0xDD => { + let r = seg.get(0..2).ok_or(Error::Truncated)?; + restart_interval = usize::from(u16::from_be_bytes([r[0], r[1]])); + } + 0xDA => { + let Some((precision, width, height, mut comps)) = frame.take() else { + return Err(Error::Malformed("lossless JPEG: scan before the frame header")); + }; + let ns = usize::from(*seg.first().ok_or(Error::Truncated)?); + if ns != comps.len() { + return Err(Error::Unsupported("lossless JPEG with more than one scan".into())); + } + for k in 0..ns { + let c = seg.get(1 + 2 * k..3 + 2 * k).ok_or(Error::Truncated)?; + let comp = comps.iter_mut().find(|x| x.id == c[0]).ok_or(Error::Malformed("lossless JPEG: unknown scan component"))?; + comp.table = usize::from(c[1] >> 4); + if comp.table > 3 { + return Err(Error::Malformed("lossless JPEG: Huffman table id")); + } + } + let tail = seg.get(1 + 2 * ns..4 + 2 * ns).ok_or(Error::Truncated)?; + let (predictor, pt) = (tail[0], tail[2] & 0x0F); + if !(1..=7).contains(&predictor) { + return Err(Error::Unsupported(format!("lossless JPEG predictor {predictor}"))); + } + if pt >= precision { + return Err(Error::Malformed("lossless JPEG: point transform")); + } + let n = width.checked_mul(height).and_then(|v| v.checked_mul(comps.len())).ok_or(Error::TooLarge)?; + if n > max_samples { + return Err(Error::TooLarge); + } + let mut hufs = Vec::with_capacity(comps.len()); + for c in &comps { + hufs.push(tables.get(c.table).and_then(Option::as_ref).ok_or(Error::Malformed("lossless JPEG: missing Huffman table"))?); + } + // every sample costs at least one bit of entropy-coded data (a Huffman code is at + // least one bit long), so a frame bigger than the data cannot be real + if n.div_ceil(8) > data.len().saturating_sub(pos + len) { + return Err(Error::Truncated); + } + let params = ScanParams { width, height, precision, predictor, pt, restart: restart_interval }; + let samples = scan(data, pos + len, &hufs, ¶ms)?; + return Ok(Decoded { width, height, components: comps.len(), precision, predictor, samples }); + } + 0xC0..=0xC2 | 0xC5..=0xC7 | 0xC9..=0xCB | 0xCD..=0xCF => { + return Err(Error::Unsupported("lossy (DCT) JPEG in a DNG".into())); + } + _ => {} // APPn, COM, DQT … are skipped + } + pos += len; + } +} + +/// Frame and scan parameters needed to decode one scan. +struct ScanParams { + width: usize, + height: usize, + precision: u8, + predictor: u8, + pt: u8, + /// Restart interval in samples per component (0 = none). + restart: usize, +} + +/// The next difference (F.2.2.1 and H.1.2.2): category and extra bits from the combined table +/// when they fit, else symbol by symbol. +#[inline(always)] +fn difference(bits: &mut Bits, h: &Huffman) -> Result { + let e = h.full.get(bits.peek(FULL_BITS) as usize).copied().unwrap_or(0); + if e & FULL_VALID != 0 { + bits.skip(e & 0xFF); + return Ok(i32::from((e >> 16) as u16 as i16)); + } + Ok(match u32::from(bits.decode(h)?) { + 0 => 0, + 16 => 32768, + s => { + let v = bits.bits(s) as i32; + if v < (1 << (s - 1)) { v - (1 << s) + 1 } else { v } + } + }) +} + +/// Decode the entropy-coded data of one scan (H.1.2). +fn scan(data: &[u8], start: usize, hufs: &[&Huffman], params: &ScanParams) -> Result> { + let ScanParams { width, height, precision, predictor, pt, restart } = *params; + let nc = hufs.len(); + let row_len = width * nc; + if restart != 0 && !restart.is_multiple_of(width) { + return Err(Error::Unsupported("lossless JPEG restart interval that splits a line".into())); + } + let lines_per_interval = if restart == 0 { usize::MAX } else { restart / width }; + let mut out = vec![0u16; row_len * height]; + let mut bits = Bits::new(data, start); + let initial = 1i32 << (precision - pt - 1); + let mut line_in_interval = 0usize; + for y in 0..height { + if line_in_interval == lines_per_interval { + bits.restart()?; + line_in_interval = 0; + } + let (above, cur) = out.split_at_mut(y * row_len); + let cur = cur.get_mut(..row_len).ok_or(Error::TooLarge)?; + let above = if line_in_interval == 0 { None } else { above.get((y - 1) * row_len..) }; + match predictor { + 1 => line::<1>(&mut bits, hufs, cur, above, initial)?, + 2 => line::<2>(&mut bits, hufs, cur, above, initial)?, + 3 => line::<3>(&mut bits, hufs, cur, above, initial)?, + 4 => line::<4>(&mut bits, hufs, cur, above, initial)?, + 5 => line::<5>(&mut bits, hufs, cur, above, initial)?, + 6 => line::<6>(&mut bits, hufs, cur, above, initial)?, + _ => line::<7>(&mut bits, hufs, cur, above, initial)?, + } + line_in_interval += 1; + } + if pt > 0 { + for s in &mut out { + *s <<= pt; + } + } + Ok(out) +} + +/// One line of samples (H.1.2.1): the first line of a restart interval (`above` is `None`) +/// predicts from the left with `initial` for its first sample; other lines predict their first +/// sample from above and the rest with predictor `P`. +#[inline(always)] +fn line(bits: &mut Bits, hufs: &[&Huffman], cur: &mut [u16], above: Option<&[u16]>, initial: i32) -> Result<()> { + let nc = hufs.len(); + let row_len = cur.len(); + let above = match above { + Some(a) => Some(a.get(..row_len).ok_or(Error::TooLarge)?), + None => None, + }; + for (c, h) in hufs.iter().enumerate() { + let d = difference(bits, h)?; + let px = match above { + Some(a) => i32::from(a[c]), + None => initial, + }; + cur[c] = ((px + d) & 0xFFFF) as u16; + } + let mut i = nc; + match above { + None => { + while i < row_len { + for h in hufs { + let d = difference(bits, h)?; + cur[i] = ((i32::from(cur[i - nc]) + d) & 0xFFFF) as u16; + i += 1; + } + } + } + Some(a) => { + while i < row_len { + for h in hufs { + let d = difference(bits, h)?; + let px = predict(P, i32::from(cur[i - nc]), i32::from(a[i]), i32::from(a[i - nc])); + cur[i] = ((px + d) & 0xFFFF) as u16; + i += 1; + } + } + } + } + Ok(()) +} + +/// Table H.1 predictors (`ra` left, `rb` above, `rc` above left). +#[inline(always)] +pub fn predict(p: u8, ra: i32, rb: i32, rc: i32) -> i32 { + match p { + 1 => ra, + 2 => rb, + 3 => rc, + 4 => ra + rb - rc, + 5 => ra + ((rb - rc) >> 1), + 6 => rb + ((ra - rc) >> 1), + _ => (ra + rb) >> 1, + } +} + +#[cfg(test)] +mod tests { + use super::*; + #[test] + fn agrees_with_full_decoder_for_line_aligned_restarts() { + let pixels: Vec = (0..96).map(|i| ((i * 137) % 4096) as u16).collect(); + for predictor in 1..=7 { + for restart in [0, 12, 24] { + let bytes = crate::ljpeg::encode(&pixels, 12, 8, 1, 12, predictor, restart); + let fast = decode(&bytes, pixels.len()).unwrap(); + let full = crate::ljpeg::decode(&bytes, pixels.len()).unwrap(); + assert_eq!(fast.samples, pixels, "predictor {predictor}, restart {restart}"); + assert_eq!(fast.samples, full.data); + } + } + let bytes = crate::ljpeg::encode(&pixels, 12, 8, 1, 12, 1, 0); + assert_eq!(decode(&bytes, pixels.len() - 1).unwrap_err(), Error::TooLarge); + } +} diff --git a/crates/stream-codecs/Cargo.toml b/crates/stream-codecs/Cargo.toml index 8827f294e..14ce68603 100644 --- a/crates/stream-codecs/Cargo.toml +++ b/crates/stream-codecs/Cargo.toml @@ -15,8 +15,9 @@ workspace = true fax = "0.2.7" # CCITT fax decoding, the decoder the vendored PDF reader uses. MIT OR Apache-2.0. hayro-ccitt = { version = "0.3", default-features = false } -# JBIG2 decoding, the decoder the vendored PDF reader uses (the workspace patches it to `vendor/hayro-jbig2`). -hayro-jbig2 = { version = "=0.3.0", default-features = false, features = ["std"] } +# The bounded JBIG2 API is a local extension: consumers do not inherit workspace patches. +# Keep the vendored dependency explicit so Git consumers retain its work-budget safeguards. +hayro-jbig2 = { path = "../../vendor/hayro-jbig2", version = "=0.3.0", default-features = false, features = ["std"] } thiserror = { workspace = true } [dev-dependencies] diff --git a/crates/stream-codecs/src/jbig2/decode.rs b/crates/stream-codecs/src/jbig2/decode.rs index e5a49d359..0e04bd648 100644 --- a/crates/stream-codecs/src/jbig2/decode.rs +++ b/crates/stream-codecs/src/jbig2/decode.rs @@ -78,6 +78,15 @@ mod tests { use super::*; use crate::{encode_jbig2, encode_jbig2_generic}; + /// A registry fallback lacks the budget APIs and breaks consumers outside this workspace. + #[test] + fn bounded_decoder_dependency_is_explicit_for_external_consumers() { + let manifest = include_str!("../../Cargo.toml"); + let dependency = manifest.lines().find(|line| line.starts_with("hayro-jbig2 =")).unwrap(); + assert!(dependency.contains(r#"path = "../../vendor/hayro-jbig2""#)); + assert!(dependency.contains(r#"version = "=0.3.0""#)); + } + /// [`decode_jbig2_within`] with no limit on the work. fn decode_jbig2(data: &[u8], globals: Option<&[u8]>, columns: u32, rows: u32, max: usize) -> Result, CodecError> { let mut unbounded = u64::MAX; diff --git a/crates/svg/src/svgz.rs b/crates/svg/src/svgz.rs index de5232612..4cb466111 100644 --- a/crates/svg/src/svgz.rs +++ b/crates/svg/src/svgz.rs @@ -44,13 +44,15 @@ pub fn inflate(bytes: &[u8], limit: usize) -> Result, Error> { let kept = out.len(); let room = u64::try_from(limit.saturating_sub(kept)).unwrap_or(u64::MAX).saturating_add(1); let mut member = flate2::bufread::GzDecoder::new(rest); + // A valid header identifies a member even if its body or trailer is damaged. + let recognized = member.header().is_some(); let read = member.by_ref().take(room).read_to_end(&mut out); if out.len() > limit { return Err(Error::TooLarge(limit)); } match read { Ok(_) => {} - Err(e) if first => return Err(Error::Corrupt(e.to_string())), + Err(e) if first || recognized => return Err(Error::Corrupt(e.to_string())), // Trailing bytes that only start like a member: not part of the file. Err(_) => { out.truncate(kept); @@ -161,6 +163,7 @@ mod tests { let mut two = compress(a.as_bytes()); two.extend(compress(b.as_bytes())); assert_eq!(text(&two, MB).as_deref(), Ok(SVG)); + assert_eq!(text(&two, SVG.len()).as_deref(), Ok(SVG)); // The limit counts every member. assert_eq!(inflate(&two, SVG.len() - 1), Err(Error::TooLarge(SVG.len() - 1))); let mut padded = compress(SVG.as_bytes()); @@ -172,6 +175,26 @@ mod tests { assert_eq!(text(&junk, MB).as_deref(), Ok(SVG)); } + #[test] + fn damaged_later_members_are_errors() { + let mut first = compress(b""); + let second = compress(b""); + + let mut bad_crc = second.clone(); + let crc = bad_crc.len() - 8; + bad_crc[crc] ^= 1; + first.extend_from_slice(&bad_crc); + assert!(matches!(inflate(&first, MB), Err(Error::Corrupt(_)))); + assert!(matches!(inflate(&first, b"".len()), Err(Error::Corrupt(_)))); + + let mut truncated = compress(b""); + truncated.extend_from_slice(&second[..second.len() - 4]); + assert!(matches!(inflate(&truncated, MB), Err(Error::Corrupt(_)))); + let mut truncated_body = compress(b""); + truncated_body.extend_from_slice(&second[..10]); + assert!(matches!(inflate(&truncated_body, MB), Err(Error::Corrupt(_)))); + } + #[test] fn damaged_and_nested_streams_are_errors() { let z = compress(SVG.as_bytes()); diff --git a/crates/texture/Cargo.toml b/crates/texture/Cargo.toml index 78868dd87..173abb0b0 100644 --- a/crates/texture/Cargo.toml +++ b/crates/texture/Cargo.toml @@ -9,6 +9,7 @@ repository.workspace = true [features] default = [] +dds = ["dep:dds", "dep:bytemuck"] parallel = ["dep:rayon"] [lints] @@ -17,3 +18,11 @@ workspace = true [dependencies] rayon = { workspace = true, optional = true } thiserror = { workspace = true } + +# Same decoder used by MaterialCraft; no parallel backend enabled. +dds = { version = "0.2.0", default-features = false, optional = true } +bytemuck = { workspace = true, optional = true } + +[dev-dependencies] +dds = { version = "0.2.0", default-features = false } +bytemuck = { workspace = true } diff --git a/crates/texture/src/dds.rs b/crates/texture/src/dds.rs new file mode 100644 index 000000000..6971ad461 --- /dev/null +++ b/crates/texture/src/dds.rs @@ -0,0 +1,116 @@ +//! First-surface DDS decoding to unmodified RGBA floats. +//! +//! Adapted from MaterialCraft `dds_image.rs` (originally chasse20, cc052aeaa8c61fbe6975674b6d5e28bb3c597b9d). +//! Uses the same `dds` backend and first-surface semantics, including the first cube face or +//! volume slice. This is not a whole-resource decoder. Transfer functions and image policy +//! belong to callers. [`Limits`] bounds the output float plane, not backend scratch buffers. + +use crate::{ImageError, Kind, Limits}; + +/// A malformed DDS/backend failure or rejected output allocation. +#[derive(Debug, thiserror::Error)] +pub enum DecodeError { + /// Output size or allocation failed shared image limits. + #[error(transparent)] + Image(#[from] ImageError), + /// The decoder rejected input or output geometry. + #[error("DDS decode failed: {0}")] + Decode(String), +} + +/// Unmodified backend output. Alpha and out-of-range float values are not clamped. +#[derive(Debug)] +pub struct Surface { + /// Width of the first surface in pixels. + pub width: u32, + /// Height of the first surface in pixels. + pub height: u32, + /// Row-major RGBA float samples without transfer conversion. + pub rgba: Vec, + /// Exactly the backend's `Precision::F32` classification, as used by MaterialCraft. + pub source_precision_is_f32: bool, +} + +/// Decode the first surface after validating dimensions, pixels and output-byte limits. +/// The output reservation is fallible; no second full-sized float plane is created. +pub fn decode_first(bytes: &[u8], limits: &Limits) -> Result { + let error = |e: ::dds::DecodingError| DecodeError::Decode(e.to_string()); + let mut decoder = ::dds::Decoder::new(std::io::Cursor::new(bytes)).map_err(error)?; + let size = decoder.main_size(); + let len = limits.check(size.width, size.height, Kind::Color)?; + let mut rgba = Vec::new(); + rgba.try_reserve_exact(len).map_err(|_| ImageError::OutOfMemory(size.width, size.height))?; + rgba.resize(len, 0.0_f32); + let view = ::dds::ImageViewMut::new(bytemuck::cast_slice_mut(&mut rgba), size, ::dds::ColorFormat::RGBA_F32) + .ok_or_else(|| DecodeError::Decode("invalid DDS output dimensions".into()))?; + decoder.read_surface(view).map_err(error)?; + Ok(Surface { width: size.width, height: size.height, rgba, source_precision_is_f32: decoder.format().precision() == ::dds::Precision::F32 }) +} + +#[cfg(test)] +#[allow(clippy::unwrap_used)] +mod tests { + use super::*; + + fn bytes(format: ::dds::Format, payload: &[u8]) -> Vec { + let mut bytes = Vec::new(); + ::dds::header::Header::new_image(1, 1, format).write(&mut bytes).unwrap(); + bytes.extend_from_slice(payload); + bytes + } + + #[test] + fn rgba8_keeps_transfer_and_alpha_unmodified() { + let data = bytes(::dds::Format::R8G8B8A8_UNORM, &[128, 64, 255, 32]); + let image = decode_first(&data, &Limits::default()).unwrap(); + assert_eq!((image.width, image.height), (1, 1)); + assert!(!image.source_precision_is_f32); + for (actual, expected) in image.rgba.iter().zip([128.0 / 255.0, 64.0 / 255.0, 1.0, 32.0 / 255.0]) { + assert!((actual - expected).abs() < 1e-6); + } + } + + #[test] + fn float_hdr_and_negative_values_survive() { + let values = [-0.5_f32, 2.0, 0.25, 0.75]; + let payload: Vec = values.iter().flat_map(|v| v.to_le_bytes()).collect(); + let image = decode_first(&bytes(::dds::Format::R32G32B32A32_FLOAT, &payload), &Limits::default()).unwrap(); + assert!(image.source_precision_is_f32); + assert_eq!(image.rgba, values); + } + + #[test] + fn half_float_and_compressed_pixels_preserve_backend_values() { + let half: Vec = [0xb800u16, 0x4000, 0x3400, 0x3a00].iter().flat_map(|v| v.to_le_bytes()).collect(); + let image = decode_first(&bytes(::dds::Format::R16G16B16A16_FLOAT, &half), &Limits::default()).unwrap(); + assert_eq!(image.rgba, [-0.5, 2.0, 0.25, 0.75]); + assert_eq!(image.source_precision_is_f32, ::dds::Format::R16G16B16A16_FLOAT.precision() == ::dds::Precision::F32); + let mut compressed = Vec::new(); + ::dds::header::Header::new_image(4, 4, ::dds::Format::BC1_UNORM).write(&mut compressed).unwrap(); + compressed.extend_from_slice(&[0, 0xf8, 0, 0, 0, 0, 0, 0]); + let image = decode_first(&compressed, &Limits::default()).unwrap(); + assert!(!image.source_precision_is_f32); + assert!(image.rgba.as_chunks::<4>().0.iter().all(|p| *p == [1.0, 0.0, 0.0, 1.0])); + } + + #[test] + fn cube_keeps_first_surface_only_like_materialcraft() { + let mut data = Vec::new(); + ::dds::header::Header::new_cube_map(1, 1, ::dds::Format::R8G8B8A8_UNORM).write(&mut data).unwrap(); + data.extend_from_slice(&[255, 0, 0, 255]); + data.extend_from_slice(&[0; 20]); + let image = decode_first(&data, &Limits::default()).unwrap(); + assert_eq!(image.rgba, [1.0, 0.0, 0.0, 1.0]); + } + + #[test] + fn checks_output_budget_and_truncation() { + let data = bytes(::dds::Format::R8G8B8A8_UNORM, &[1, 2, 3, 4]); + let mut limits = Limits { max_bytes: 15, ..Limits::default() }; + assert!(matches!(decode_first(&data, &limits), Err(DecodeError::Image(ImageError::TooLarge(1, 1))))); + limits.max_bytes = 16; + assert!(decode_first(&data, &limits).is_ok()); + assert!(decode_first(&data[..data.len() - 1], &limits).is_err()); + assert!(decode_first(&[], &limits).is_err()); + } +} diff --git a/crates/texture/src/lib.rs b/crates/texture/src/lib.rs index 2aaef3e8b..e33b68d88 100644 --- a/crates/texture/src/lib.rs +++ b/crates/texture/src/lib.rs @@ -29,6 +29,9 @@ #![forbid(unsafe_code)] #![deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable)] +#[cfg(any(feature = "dds", test))] +pub mod dds; + use std::sync::Arc; #[cfg(feature = "parallel")] diff --git a/crates/timeline/src/fcp7.rs b/crates/timeline/src/fcp7.rs index 41243baf6..82ff410f9 100644 --- a/crates/timeline/src/fcp7.rs +++ b/crates/timeline/src/fcp7.rs @@ -117,6 +117,9 @@ struct PendingLink { refs: Vec, } +const MAX_NESTING_DEPTH: usize = 64; +const MAX_REIMPORTS: usize = 4096; + struct Imp<'a, 'c, 'r> { b: Builder<'c>, opts: &'a ImportOptions<'a>, @@ -126,6 +129,9 @@ struct Imp<'a, 'c, 'r> { file_items: HashMap, seq_items: HashMap, in_progress: HashSet, + depth_hits: usize, + truncated_at: HashMap, + reimports_left: usize, links: Vec, xml_clip_ids: HashMap<(ItemId, String), ClipId>, /// `(sequence key, file key)` pairs where some audio clip item of that file uses `sourcetrack/trackindex` > 1, @@ -177,6 +183,9 @@ pub fn import(text_in: &str, opts: &ImportOptions, report: &mut Report) -> Resul file_items: HashMap::new(), seq_items: HashMap::new(), in_progress: HashSet::new(), + depth_hits: 0, + truncated_at: HashMap::new(), + reimports_left: MAX_REIMPORTS, links: Vec::new(), xml_clip_ids: HashMap::new(), split_files: HashSet::new(), @@ -349,10 +358,27 @@ impl<'a> Imp<'a, '_, '_> { /// Sequence item for a `` element (definition or reference). fn sequence(&mut self, n: Node<'a>, bin: Option) -> Option { let key = node_key(n); + let depth = self.in_progress.len(); + let mut again = None; if let Some(id) = self.seq_items.get(&key) { - return Some(*id); + let shallower = self.truncated_at.get(&key).is_some_and(|d| depth < *d) && !self.in_progress.contains(&key); + if !shallower { + return Some(*id); + } + if self.reimports_left == 0 { + self.report.warn("truncated nested sequence kept incomplete after the re-import limit was reached"); + return Some(*id); + } + self.reimports_left -= 1; + again = Some(*id); } let def = if child(n, "media").is_some() { n } else { *self.seq_defs.get(&key)? }; + if depth >= MAX_NESTING_DEPTH { + self.depth_hits += 1; + self.report.warn("nested sequence is nested too deeply; imported as a gap"); + return None; + } + let hits_before = self.depth_hits; if !self.in_progress.insert(key.clone()) { self.report.warn("recursive nested sequence ignored"); return None; @@ -374,12 +400,18 @@ impl<'a> Imp<'a, '_, '_> { { settings.sample_rate = sr as u32; } - let id = self.b.reserve_sequence(&name, settings.clone(), bin); - if let Some(l) = path_text(def, &["labels", "label2"]).and_then(label_from) - && let Some(it) = self.b.p.item_mut(id) - { - it.label = l; - } + let id = match again { + Some(id) => id, + None => { + let id = self.b.reserve_sequence(&name, settings.clone(), bin); + if let Some(l) = path_text(def, &["labels", "label2"]).and_then(label_from) + && let Some(it) = self.b.p.item_mut(id) + { + it.label = l; + } + id + } + }; self.seq_items.insert(key.clone(), id); let mut seq = empty_sequence(settings); seq.start_timecode = tc.and_then(|t| child_i64(t, "frame")).unwrap_or(0); @@ -403,6 +435,11 @@ impl<'a> Imp<'a, '_, '_> { } self.b.put_sequence(id, seq); self.in_progress.remove(&key); + if self.depth_hits > hits_before { + self.truncated_at.insert(key, depth); + } else { + self.truncated_at.remove(&key); + } Some(id) } @@ -531,9 +568,19 @@ impl<'a> Imp<'a, '_, '_> { } } } else if let Some(s) = child(c, "sequence") { - self.sequence(s, None) + if let Some(id) = self.sequence(s, None) { + return Some(id); + } + self.report.warn(format!("clip \"{name}\" refers to a nested sequence that was left out; imported offline")); + Some(self.offline_media(kind, seq_rate, frame, name)) } else { self.report.warn(format!("clip \"{name}\" has no media reference; imported offline")); + Some(self.offline_media(kind, seq_rate, frame, name)) + } + } + + fn offline_media(&mut self, kind: TrackKind, seq_rate: FrameRate, frame: (u32, u32), name: &str) -> ItemId { + { let spec = MediaSpec { video: (kind == TrackKind::Video).then_some((frame.0, frame.1, seq_rate)), audio: (kind == TrackKind::Audio).then_some((48_000, 2)), @@ -545,7 +592,7 @@ impl<'a> Imp<'a, '_, '_> { if let Some(m) = self.b.p.item_mut(id).and_then(|i| i.as_media_mut()) { m.offline = true; } - Some(id) + id } } diff --git a/crates/timeline/tests/it/fcp7.rs b/crates/timeline/tests/it/fcp7.rs index 00d0704ec..4deb7a3bd 100644 --- a/crates/timeline/tests/it/fcp7.rs +++ b/crates/timeline/tests/it/fcp7.rs @@ -764,3 +764,24 @@ fn a_huge_link_group_does_not_write_quadratic_links() { assert_eq!(text.matches("").count(), 64 * 64); assert!(rep.has_warnings(), "{rep}"); } + +#[test] +fn nested_sequence_chain_is_bounded_and_reimported_at_a_shallower_depth() { + let mut doc = String::from(""); + for i in 0..66 { + let nested = if i < 65 { + format!( + "nest024024", + i + 1 + ) + } else { + String::new() + }; + doc.push_str(&format!("S{i}2424")); + } + doc.push_str(""); + let (imported, report) = fcp7_import(&doc, None); + assert!(report.mentions("nested too deeply")); + let nested = imported.project.items.values().find(|item| item.name == "S63").and_then(|item| imported.project.sequence(item.id)); + assert!(nested.is_some_and(|seq| seq.video_tracks.first().is_some_and(|track| track.items.len() == 1))); +} diff --git a/crates/usd/Cargo.toml b/crates/usd/Cargo.toml new file mode 100644 index 000000000..ac1c64d0c --- /dev/null +++ b/crates/usd/Cargo.toml @@ -0,0 +1,15 @@ +[package] +name = "craft-usd" +description = "Bounded USD Crate wire primitives, independent of scene and material policy" +version.workspace = true +edition.workspace = true +license.workspace = true +rust-version.workspace = true +repository.workspace = true + +[lints] +workspace = true + +[dependencies] +lz4_flex = { version = "0.11", default-features = false, features = ["safe-decode", "safe-encode"] } +thiserror.workspace = true diff --git a/crates/usd/README.md b/crates/usd/README.md new file mode 100644 index 000000000..9e9635185 --- /dev/null +++ b/crates/usd/README.md @@ -0,0 +1,7 @@ +# craft-usd + +Bounded wire primitives for USD Crate, independent of application scene and material models. + +Provides bounded integer-delta array decoding and Crate-framed LZ4 decompression: a single unprefixed block or up to 127 length-prefixed blocks, under a caller-provided output-byte cap with fallible output allocation. Trailing bytes after counted chunks retain the original reader's ignored-tail policy. This is not a complete USD reader/writer; scene interpretation, integer packing and export remain outside this extraction. Integer decoding uses explicit 32/64-bit widths, signed wrapping arithmetic, a caller output-byte cap and fallible allocation. + +Adapted from TextureCraft `crates/formats/src/mesh_import/usd/codec.rs`, introduced by bflatastic in `bdbb52c9b66b221a55b370105b42f485a76e6fc7`. The LZ4 backend is MIT-licensed `lz4_flex`. diff --git a/crates/usd/src/integers.rs b/crates/usd/src/integers.rs new file mode 100644 index 000000000..c50dd0e0f --- /dev/null +++ b/crates/usd/src/integers.rs @@ -0,0 +1,127 @@ +//! USD Crate integer delta decoding, adapted from TextureCraft's bflatastic implementation. + +/// Wire width of each integer and its common delta. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub enum Width { + /// Signed 32-bit wrapping arithmetic, sign-extended to i64 on output. + I32, + /// Signed 64-bit wrapping arithmetic. + I64, +} +impl Width { + fn bytes(self) -> usize { + match self { + Self::I32 => 4, + Self::I64 => 8, + } + } +} + +/// Invalid packed data or a caller output-budget/allocation failure. +#[derive(Debug, thiserror::Error)] +pub enum Error { + /// The declared count exceeds the caller's byte budget or addressable output. + #[error("USD integer array exceeds output budget")] + Limit, + /// A common delta, code table or selected delta is incomplete. + #[error("truncated USD integer array")] + Truncated, + /// The bounded output reservation failed. + #[error("cannot allocate USD integer array")] + Allocation, +} + +fn signed(bytes: &[u8], at: usize, width: usize) -> Option { + let part = bytes.get(at..at.checked_add(width)?)?; + Some(match width { + 1 => i64::from(i8::from_le_bytes(part.try_into().ok()?)), + 2 => i64::from(i16::from_le_bytes(part.try_into().ok()?)), + 4 => i64::from(i32::from_le_bytes(part.try_into().ok()?)), + 8 => i64::from_le_bytes(part.try_into().ok()?), + _ => return None, + }) +} + +/// Decode `count` values, charging eight output bytes per value against `max_output_bytes`. +/// +/// The common delta is followed by a packed two-bit code per value and variable-width +/// exceptional deltas. Arithmetic wraps at the specified wire width. Extra bytes after the +/// values are ignored, matching the original reader. Counts, the code table and the output +/// byte budget are checked before a fallible reservation. A zero count returns an empty vector. +pub fn decode(data: &[u8], count: usize, width: Width, max_output_bytes: usize) -> Result, Error> { + let output_bytes = count.checked_mul(8).filter(|n| *n <= max_output_bytes && *n <= isize::MAX as usize).ok_or(Error::Limit)?; + if output_bytes == 0 { + return Ok(Vec::new()); + } + let size = width.bytes(); + let common = signed(data, 0, size).ok_or(Error::Truncated)?; + let code_bytes = count.div_ceil(4); + let mut at = size.checked_add(code_bytes).ok_or(Error::Truncated)?; + let codes = data.get(size..at).ok_or(Error::Truncated)?; + let mut output = Vec::new(); + output.try_reserve_exact(count).map_err(|_| Error::Allocation)?; + let mut previous = 0i64; + for i in 0..count { + let code = (codes.get(i / 4).copied().ok_or(Error::Truncated)? >> (2 * (i % 4))) & 3; + let delta = if code == 0 { + common + } else { + let bytes = (size >> (3 - usize::from(code))).max(1); + let value = signed(data, at, bytes).ok_or(Error::Truncated)?; + at = at.checked_add(bytes).ok_or(Error::Truncated)?; + value + }; + previous = match width { + Width::I32 => i64::from((previous as i32).wrapping_add(delta as i32)), + Width::I64 => previous.wrapping_add(delta), + }; + output.push(previous); + } + Ok(output) +} + +#[cfg(test)] +#[allow(clippy::unwrap_used)] +mod tests { + use super::*; + + #[test] + fn common_deltas_and_output_budget() { + let mut data = 1i32.to_le_bytes().to_vec(); + data.push(0); + assert_eq!(decode(&data, 4, Width::I32, 32).unwrap(), [1, 2, 3, 4]); + assert!(matches!(decode(&data, 4, Width::I32, 31), Err(Error::Limit))); + assert!(matches!(decode(&data[..4], 4, Width::I32, 32), Err(Error::Truncated))); + assert!(matches!(decode(&[], usize::MAX, Width::I64, usize::MAX), Err(Error::Limit))); + assert!(decode(&[], 0, Width::I64, 0).unwrap().is_empty()); + } + + #[test] + fn narrow_exception_codes_sign_extend() { + for (width, size) in [(Width::I32, 4), (Width::I64, 8)] { + let mut data = vec![0; size]; + data.push(1 | (2 << 2)); + let first_size = (size >> 2).max(1); + let second_size = size >> 1; + data.extend_from_slice(&(-3i64).to_le_bytes()[..first_size]); + data.extend_from_slice(&7i64.to_le_bytes()[..second_size]); + assert_eq!(decode(&data, 2, width, 16).unwrap(), [-3, 4]); + } + } + + #[test] + fn exceptional_deltas_and_wrapping() { + for (width, size) in [(Width::I32, 4), (Width::I64, 8)] { + // Full-width positive maximum, followed by common +1, wraps to signed minimum. + let mut data = 1i64.to_le_bytes()[..size].to_vec(); + data.push(3); + let maximum = if width == Width::I32 { i64::from(i32::MAX) } else { i64::MAX }; + data.extend_from_slice(&maximum.to_le_bytes()[..size]); + let minimum = if width == Width::I32 { i64::from(i32::MIN) } else { i64::MIN }; + assert_eq!(decode(&data, 2, width, 16).unwrap(), [maximum, minimum]); + assert!(decode(&data[..data.len() - 1], 2, width, 16).is_err()); + data.extend(b"ignored tail"); + assert_eq!(decode(&data, 2, width, 16).unwrap(), [maximum, minimum]); + } + } +} diff --git a/crates/usd/src/lib.rs b/crates/usd/src/lib.rs new file mode 100644 index 000000000..edaeb89a2 --- /dev/null +++ b/crates/usd/src/lib.rs @@ -0,0 +1,98 @@ +//! Bounded USD Crate wire primitives. Not a complete USD scene reader or writer. +//! +//! The framing decoder is adapted from TextureCraft's `usd/codec.rs`, originally +//! bflatastic, commit bdbb52c9b66b221a55b370105b42f485a76e6fc7. +#![forbid(unsafe_code)] +#![deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable)] + +/// Bounded Crate integer delta arrays. +pub mod integers; + +/// Failure to read a bounded Crate LZ4 frame. +#[derive(Debug, thiserror::Error)] +pub enum Error { + /// Invalid or incomplete chunk framing. + #[error("invalid USD Crate LZ4 framing: {0}")] + Framing(&'static str), + /// Invalid compressed data or output exceeding the supplied capacity. + #[error("USD Crate LZ4 data: {0}")] + Lz4(#[from] lz4_flex::block::DecompressError), + /// The bounded output allocation could not be reserved. + #[error("cannot allocate USD Crate LZ4 output")] + Allocation, +} + +/// Decode Crate's chunk-count byte and LZ4 blocks to at most `max_output` bytes. +/// +/// A zero count means a single unprefixed block; 1..=127 means that many blocks, +/// each preceded by a signed little-endian 32-bit length. As in the original reader, +/// bytes following the declared chunks are ignored. The output allocation is fallible +/// and capped by both the caller's budget and the input's maximum LZ4 expansion. +/// This bounds the returned plane, not backend internal scratch memory. +pub fn decompress(input: &[u8], max_output: usize) -> Result, Error> { + let (&chunks, mut rest) = input.split_first().ok_or(Error::Framing("empty block"))?; + if chunks > 127 { + return Err(Error::Framing("chunk count exceeds 127")); + } + let capacity = max_output.min(input.len().saturating_mul(256).saturating_add(64)); + let mut output = Vec::new(); + output.try_reserve_exact(capacity).map_err(|_| Error::Allocation)?; + output.resize(capacity, 0); + let written = if chunks == 0 { + lz4_flex::block::decompress_into(rest, &mut output)? + } else { + let mut total = 0usize; + for _ in 0..chunks { + let length = rest.get(..4).and_then(|b| b.try_into().ok()).map(i32::from_le_bytes).ok_or(Error::Framing("missing chunk length"))?; + let length = usize::try_from(length).map_err(|_| Error::Framing("negative chunk length"))?; + let end = 4usize.checked_add(length).ok_or(Error::Framing("chunk length overflow"))?; + let chunk = rest.get(4..end).ok_or(Error::Framing("truncated chunk"))?; + rest = rest.get(end..).unwrap_or_default(); + let tail = output.get_mut(total..).ok_or(Error::Framing("output offset overflow"))?; + let count = lz4_flex::block::decompress_into(chunk, tail)?; + total = total.checked_add(count).ok_or(Error::Framing("output length overflow"))?; + } + total + }; + output.truncate(written); + Ok(output) +} + +#[cfg(test)] +#[allow(clippy::unwrap_used)] +mod tests { + use super::*; + + #[test] + fn single_block_matches_original_wire() { + for n in [0, 1, 15, 16, 255, 4096] { + let raw: Vec = (0..n).map(|i| (i % 251) as u8).collect(); + let mut frame = vec![0]; + frame.extend(lz4_flex::block::compress(&raw)); + assert_eq!(decompress(&frame, raw.len()).unwrap(), raw); + if n > 0 { + assert!(decompress(&frame, n - 1).is_err()); + } + } + } + + #[test] + fn counted_blocks_and_trailing_bytes_keep_existing_policy() { + let mut frame = vec![2]; + for raw in [b"first".as_slice(), b"second".as_slice()] { + let block = lz4_flex::block::compress(raw); + frame.extend((block.len() as i32).to_le_bytes()); + frame.extend(block); + } + frame.extend(b"ignored"); + assert_eq!(decompress(&frame, 11).unwrap(), b"firstsecond"); + assert!(decompress(&frame, 10).is_err()); + } + + #[test] + fn malformed_frames_fail_within_budget() { + for data in [vec![], vec![128], vec![1], vec![1, 255, 255, 255, 255], vec![1, 5, 0, 0, 0, 1]] { + assert!(decompress(&data, 32).is_err()); + } + } +} diff --git a/crates/video/src/av1/decode/decoder.rs b/crates/video/src/av1/decode/decoder.rs index 943923a3b..65c3de6ef 100644 --- a/crates/video/src/av1/decode/decoder.rs +++ b/crates/video/src/av1/decode/decoder.rs @@ -239,6 +239,9 @@ pub struct Decoder { /// Bytes of the pictures the chunk being parsed has shown so far, and whether one was refused for size. shown_bytes: usize, shown_overflow: bool, + /// Most bytes of pictures one chunk may show: [`MAX_SHOWN_BYTES_PER_CHUNK`] unless a caller + /// lowered it with [`Decoder::set_shown_budget`]. + shown_budget: usize, lf_deltas: LoopFilterDeltas, seg_features: SegmentationFeatures, prev_gm_params: [[i32; 6]; 8], @@ -301,8 +304,8 @@ fn shown_within_limit(n: usize) -> bool { /// copies of one 4K reference (about 6.4 GB) from a few hundred bytes of show_existing_frame headers. const MAX_SHOWN_BYTES_PER_CHUNK: usize = 1 << 30; -fn shown_bytes_within_limit(bytes: usize) -> bool { - bytes <= MAX_SHOWN_BYTES_PER_CHUNK +fn shown_bytes_within_limit(bytes: usize, budget: usize) -> bool { + bytes <= budget.min(MAX_SHOWN_BYTES_PER_CHUNK) } /// Bytes of the `u16` sample planes of a `w` × `h` picture. @@ -313,14 +316,14 @@ fn picture_bytes(w: usize, h: usize, c: &crate::av1::decode::header::ColorConfig } /// Add a shown picture to a chunk's outputs, refusing the chunk (and dropping everything it showed) -/// before the list can grow past [`MAX_SHOWN_PER_CHUNK`]. -fn push_shown(out: &mut Vec<(Output, usize, u32)>, shown_bytes: &mut usize, item: (Output, usize, u32)) -> Result<()> { +/// before the list can grow past [`MAX_SHOWN_PER_CHUNK`] or its pictures past `budget` bytes. +fn push_shown(out: &mut Vec<(Output, usize, u32)>, shown_bytes: &mut usize, budget: usize, item: (Output, usize, u32)) -> Result<()> { if !shown_within_limit(out.len() + 1) { out.clear(); return Err(Error::Invalid("too many pictures shown by one chunk")); } let bytes = shown_bytes.saturating_add(item.0.bytes()); - if !shown_bytes_within_limit(bytes) { + if !shown_bytes_within_limit(bytes, budget) { out.clear(); return Err(Error::Invalid("pictures shown by one chunk are too large")); } @@ -349,6 +352,14 @@ impl Decoder { self.memory_budget = bytes; } + /// Cap the bytes of the pictures one `decode` call may return together, below the built-in + /// 1 GiB (a larger value keeps that). The pictures are refused before they are built, so a + /// caller that keeps only one of them (an AVIF still) can hold a stream to its own memory + /// budget instead of collecting up to 256 pictures first. + pub fn set_shown_budget(&mut self, bytes: u64) { + self.shown_budget = usize::try_from(bytes).unwrap_or(usize::MAX).min(MAX_SHOWN_BYTES_PER_CHUNK); + } + /// A decoder using all available cores (with the `threads` feature). pub fn new() -> Decoder { Decoder::with_threads(default_threads()) @@ -369,6 +380,7 @@ impl Decoder { ref_color: Default::default(), shown_bytes: 0, shown_overflow: false, + shown_budget: MAX_SHOWN_BYTES_PER_CHUNK, lf_deltas: LoopFilterDeltas::defaults(), seg_features: SegmentationFeatures::default(), prev_gm_params: default_gm_params(), @@ -564,20 +576,20 @@ impl Decoder { let mut r = BitReader::new(payload); let shown = self.frame_header(&mut r, temporal_id, spatial_id)?; if let Some(o) = shown { - push_shown(out, &mut self.shown_bytes, (o, tu, spatial_id))?; + push_shown(out, &mut self.shown_bytes, self.shown_budget, (o, tu, spatial_id))?; continue; } if obu_type == OBU_FRAME_T { r.byte_align(); let rest = payload.get(r.byte_pos()..).ok_or(Error::Truncated)?; if let Some(o) = self.tile_group(rest)? { - push_shown(out, &mut self.shown_bytes, (o, tu, spatial_id))?; + push_shown(out, &mut self.shown_bytes, self.shown_budget, (o, tu, spatial_id))?; } } } OBU_TILE_GROUP_T => { if let Some(o) = self.tile_group(payload)? { - push_shown(out, &mut self.shown_bytes, (o, tu, spatial_id))?; + push_shown(out, &mut self.shown_bytes, self.shown_budget, (o, tu, spatial_id))?; } } OBU_METADATA_T | OBU_TILE_LIST_T => {} @@ -726,7 +738,7 @@ impl Decoder { // A picture past the chunk's byte budget is never built (decoding the frame as a reference // still happens, so the reference slots stay as the stream says), and the chunk is refused. let bytes = picture_bytes(job.fh.upscaled_width as usize, job.fh.frame_height as usize, &job.seq.color); - if job.picture.is_some() && !shown_bytes_within_limit(self.shown_bytes.saturating_add(bytes)) { + if job.picture.is_some() && !shown_bytes_within_limit(self.shown_bytes.saturating_add(bytes), self.shown_budget) { job.picture = None; self.shown_overflow = true; self.submit(job); @@ -1307,18 +1319,33 @@ mod shown_limit_tests { #[test] fn push_shown_refuses_a_chunk_past_either_limit_and_drops_what_it_showed() { let (mut out, mut bytes) = (Vec::new(), 0usize); - push_shown(&mut out, &mut bytes, shown(10)).unwrap(); + push_shown(&mut out, &mut bytes, MAX_SHOWN_BYTES_PER_CHUNK, shown(10)).unwrap(); assert_eq!((out.len(), bytes), (1, 10)); // past the byte budget - assert!(push_shown(&mut out, &mut bytes, shown(MAX_SHOWN_BYTES_PER_CHUNK)).is_err()); + assert!(push_shown(&mut out, &mut bytes, MAX_SHOWN_BYTES_PER_CHUNK, shown(MAX_SHOWN_BYTES_PER_CHUNK)).is_err()); assert!(out.is_empty(), "the refused chunk shows nothing"); // past the count limit let (mut out, mut bytes) = (Vec::new(), 0usize); for _ in 0..MAX_SHOWN_PER_CHUNK { - push_shown(&mut out, &mut bytes, shown(0)).unwrap(); + push_shown(&mut out, &mut bytes, MAX_SHOWN_BYTES_PER_CHUNK, shown(0)).unwrap(); } - assert!(push_shown(&mut out, &mut bytes, shown(0)).is_err()); + assert!(push_shown(&mut out, &mut bytes, MAX_SHOWN_BYTES_PER_CHUNK, shown(0)).is_err()); assert!(out.is_empty()); + // past a caller's lower budget + let (mut out, mut bytes) = (Vec::new(), 0usize); + push_shown(&mut out, &mut bytes, 100, shown(60)).unwrap(); + assert!(push_shown(&mut out, &mut bytes, 100, shown(60)).is_err()); + assert!(out.is_empty()); + } + + #[test] + fn a_shown_budget_can_lower_the_limit_but_not_raise_it() { + let mut d = Decoder::with_threads(1); + assert_eq!(d.shown_budget, MAX_SHOWN_BYTES_PER_CHUNK); + d.set_shown_budget(1 << 20); + assert_eq!(d.shown_budget, 1 << 20); + d.set_shown_budget(u64::MAX); + assert_eq!(d.shown_budget, MAX_SHOWN_BYTES_PER_CHUNK); } #[test] @@ -1333,8 +1360,8 @@ mod shown_limit_tests { // 4K 4:2:0, u16 samples: about 25 MB; 256 of them is far past the limit, a few are fine let one = picture_bytes(3840, 2160, &c); assert_eq!(one, (3840 * 2160 + 2 * 1920 * 1080) * 2); - assert!(shown_bytes_within_limit(4 * one)); - assert!(!shown_bytes_within_limit(MAX_SHOWN_PER_CHUNK * one)); + assert!(shown_bytes_within_limit(4 * one, MAX_SHOWN_BYTES_PER_CHUNK)); + assert!(!shown_bytes_within_limit(MAX_SHOWN_PER_CHUNK * one, MAX_SHOWN_BYTES_PER_CHUNK)); let mono = crate::av1::decode::header::ColorConfig { num_planes: 1, ..Default::default() }; assert_eq!(picture_bytes(10, 5, &mono), 100); assert_eq!(picture_bytes(usize::MAX, 2, &c), usize::MAX); diff --git a/crates/webp/src/frames.rs b/crates/webp/src/frames.rs index e97118482..4b5322d34 100644 --- a/crates/webp/src/frames.rs +++ b/crates/webp/src/frames.rs @@ -102,6 +102,15 @@ pub fn read_frames(bytes: &[u8], limits: &FrameLimits) -> Result limits.max_canvas_bytes { + return Err(FrameError::LimitExceeded(format!("a {width}x{height} canvas is over {} bytes", limits.max_canvas_bytes))); + } + if (width, height) != (info.width, info.height) { + return Err(FrameError::Malformed("conflicting canvas dimensions".into())); + } if !decoder.has_animation() { let mut rgba = vec![0u8; usize::try_from(canvas).map_err(|_| FrameError::LimitExceeded("canvas".into()))?]; let alpha = decoder.color_type().has_alpha(); @@ -244,6 +253,21 @@ mod tests { assert!(read_frames(&bytes, &FrameLimits { max_canvas_bytes: 24, ..FrameLimits::default() }).is_ok()); } + #[test] + fn conflicting_canvas_is_rejected_before_frame_decode() { + let mut bytes = anim(&[(1, 10), (2, 10)]); + // The decoder uses the first VP8X, while the scanner sees the last one. + // A tiny limit makes this an ordering test without allocating a huge frame. + let second = chunk(b"VP8X", &[0x12, 0, 0, 0, 0, 0, 0, 0, 0, 0]); + bytes.splice(30..30, second); + let size = (bytes.len() - 8) as u32; + bytes[4..8].copy_from_slice(&size.to_le_bytes()); + assert!(matches!(read_frames(&bytes, &FrameLimits { max_canvas_bytes: 4, ..FrameLimits::default() }), Err(FrameError::LimitExceeded(_)))); + assert!( + matches!(read_frames(&bytes, &FrameLimits::default()), Err(FrameError::Malformed(message)) if message.contains("conflicting canvas")) + ); + } + #[test] fn damage_and_hostile_input() { let bytes = anim(&[(1, 10), (2, 10), (3, 10)]); diff --git a/crates/zip/src/archive.rs b/crates/zip/src/archive.rs index 5739e830a..49da63585 100644 --- a/crates/zip/src/archive.rs +++ b/crates/zip/src/archive.rs @@ -2,6 +2,7 @@ use std::borrow::Cow; use std::io::{Cursor, Read}; +use std::sync::Arc; use crate::preflight::{ReadError, read_entry}; use crate::{Budget, Error, check_entry_count}; @@ -21,11 +22,26 @@ pub struct Limits { /// caller that reads through a clone keeps its own count. #[derive(Clone)] pub struct Archive<'a> { - zip: zip::ZipArchive>, + zip: zip::ZipArchive>>, budget: Budget, duplicates: bool, } +#[derive(Clone)] +enum Input<'a> { + Borrowed(&'a [u8]), + Owned(Arc>), +} + +impl AsRef<[u8]> for Input<'_> { + fn as_ref(&self) -> &[u8] { + match self { + Self::Borrowed(bytes) => bytes, + Self::Owned(bytes) => bytes.as_slice(), + } + } +} + const CENTRAL_HEADER: &[u8; 4] = b"PK\x01\x02"; const CENTRAL_HEADER_LEN: usize = 46; @@ -46,15 +62,20 @@ impl<'a> Archive<'a> { /// Indexes `bytes`. The declared entry count is checked against /// [`Limits::max_entries`] before `zip` parses the central directory, and again after. pub fn open(bytes: &'a [u8], limits: Limits) -> Result { - check_entry_count(bytes, limits.max_entries)?; - let zip = zip::ZipArchive::new(Cursor::new(bytes)).map_err(|e| Error::NotZip(e.to_string()))?; + Self::open_input(Input::Borrowed(bytes), limits) + } + + fn open_input(input: Input<'a>, limits: Limits) -> Result { + check_entry_count(input.as_ref(), limits.max_entries)?; + let bytes = input.clone(); + let zip = zip::ZipArchive::new(Cursor::new(input)).map_err(|e| Error::NotZip(e.to_string()))?; let found = zip.len() as u64; if found > limits.max_entries { return Err(Error::TooManyEntries { declared: found, limit: limits.max_entries }); } // `zip` keeps one entry per name, so a directory with more headers than `zip` has entries // repeats a name. - let duplicates = count_central_headers(bytes, zip.central_directory_start(), zip.len().saturating_add(1)) > zip.len(); + let duplicates = count_central_headers(bytes.as_ref(), zip.central_directory_start(), zip.len().saturating_add(1)) > zip.len(); Ok(Archive { zip, budget: Budget::new("archive", limits.max_total_bytes), duplicates }) } @@ -180,3 +201,28 @@ impl<'a> Archive<'a> { &mut self.budget } } + +impl Archive<'static> { + /// Indexes owned bytes without copying them. Clones share the backing storage. + pub fn open_owned(bytes: Vec, limits: Limits) -> Result { + Self::open_input(Input::Owned(Arc::new(bytes)), limits) + } +} + +#[cfg(test)] +mod ownership_tests { + use super::*; + + #[test] + fn owned_storage_lives_until_last_archive_clone_is_dropped() { + let bytes = crate::Writer::new().finish().unwrap(); + let bytes = Arc::new(bytes); + let weak = Arc::downgrade(&bytes); + let archive = Archive::open_input(Input::Owned(bytes), Limits { max_entries: 1, max_total_bytes: 1 }).unwrap(); + let clone = archive.clone(); + drop(archive); + assert!(weak.upgrade().is_some()); + drop(clone); + assert!(weak.upgrade().is_none()); + } +} diff --git a/crates/zip/tests/archive.rs b/crates/zip/tests/archive.rs index 2c4b84113..d3b135a9c 100644 --- a/crates/zip/tests/archive.rs +++ b/crates/zip/tests/archive.rs @@ -182,3 +182,26 @@ fn a_clone_keeps_its_own_count() { assert_eq!((c.budget().used(), z.budget().used()), (3, 0)); assert!(z.read("a", 10).unwrap().is_some()); } + +#[test] +fn owned_archive_reads_and_clones_with_independent_budgets() { + let bytes = archive(&[("a", b"abc", Compression::Deflated)]); + let mut z = Archive::open_owned(bytes, Limits { max_total_bytes: 3, ..ROOMY }).unwrap(); + let mut clone = z.clone(); + assert_eq!(z.read("a", 3).unwrap().as_deref(), Some(&b"abc"[..])); + assert!(matches!(z.read("a", 3), Err(Error::BudgetSpent { .. }))); + assert_eq!(clone.read("a", 3).unwrap().as_deref(), Some(&b"abc"[..])); + assert_eq!(clone.budget().used(), 3); +} + +#[test] +fn owned_archive_applies_directory_guards() { + let mut bytes = archive(&[("a", b"1", Compression::Stored), ("b", b"2", Compression::Stored)]); + assert!(matches!(Archive::open_owned(bytes.clone(), Limits { max_entries: 1, ..ROOMY }), Err(Error::TooManyEntries { .. }))); + let local = find(&bytes, b"PK\x03\x04", 4); + bytes[local + 30] = b'a'; + let central = find(&bytes, b"PK\x01\x02", find(&bytes, b"PK\x01\x02", 0) + 4); + bytes[central + 46] = b'a'; + assert!(Archive::open_owned(bytes, ROOMY).unwrap().has_duplicate_names()); + assert!(Archive::open_owned(b"invalid".to_vec(), ROOMY).is_err()); +}