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..95341b06d 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]] @@ -1456,6 +1451,14 @@ dependencies = [ "encoding_rs", ] +[[package]] +name = "craft-control-protocol" +version = "0.1.0" +dependencies = [ + "serde_json", + "subtle", +] + [[package]] name = "craft-crs" version = "0.1.0" @@ -1491,6 +1494,17 @@ 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" @@ -1671,6 +1685,15 @@ dependencies = [ "thiserror 2.0.21", ] +[[package]] +name = "craft-fcproj-format" +version = "0.1.0" +dependencies = [ + "serde", + "serde_json", + "thiserror 2.0.21", +] + [[package]] name = "craft-fig" version = "0.1.0" @@ -1762,6 +1785,7 @@ name = "craft-font-catalog" version = "0.1.0" dependencies = [ "fontique", + "kurbo 0.13.1", "skrifa 0.47.0", "thiserror 2.0.21", "windows-registry", @@ -1929,6 +1953,7 @@ dependencies = [ "craft-jpeg", "craft-media-testkit", "craft-metadata", + "craft-tiff", "craft-video", "crc32fast", "exr", @@ -2758,6 +2783,8 @@ dependencies = [ name = "craft-texture" version = "0.1.0" dependencies = [ + "bytemuck", + "dds", "rayon", "thiserror 2.0.21", ] @@ -2855,6 +2882,14 @@ 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 +3330,19 @@ 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 +4591,12 @@ 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 +6108,12 @@ 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 +6872,6 @@ 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 +8187,15 @@ 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 +10699,6 @@ 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..9558946f5 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" } @@ -68,6 +70,7 @@ craft-doc = { path = "crates/doc" } craft-docpath = { path = "crates/docpath" } craft-dpx = { path = "crates/dpx" } craft-dwg = { path = "crates/dwg" } +craft-fcproj-format = { path = "crates/fcproj-format" } craft-ecproj-format = { path = "crates/ecproj-format" } craft-figma-rest = { path = "crates/figma-rest" } craft-exr = { path = "crates/exr" } @@ -91,6 +94,7 @@ craft-indd = { path = "crates/indd" } craft-jmap = { path = "crates/jmap" } craft-ical = { path = "crates/ical" } craft-jpeg = { path = "crates/jpeg" } +craft-control-protocol = { path = "crates/control-protocol" } craft-jsonrpc = { path = "crates/jsonrpc" } craft-mcp = { path = "crates/mcp" } craft-jxl = { path = "crates/jxl" } 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/audio/Cargo.toml b/crates/audio/Cargo.toml index 645246130..fe28cec3f 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 @@ -29,10 +29,13 @@ encode = ["lossless-encode", "aac-encode", "alac-encode", "ac3-encode", "eac3-en lossless-encode = ["pcm-encode", "flac-encode"] # Encode WAV/BWF/RF64 and AIFF/AIFF-C files. pcm-encode = [] +# Decode FLAC files and frames natively (`flac::decode`, `flac::Decoder`). +flac-decode = [] # 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 +75,6 @@ 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 } @@ -88,6 +88,10 @@ required-features = ["decode", "lossless-encode", "peaks"] name = "codecs" required-features = ["decode", "encode"] +[[test]] +name = "flac_decode" +required-features = ["flac-decode", "flac-encode"] + [[test]] name = "qoa_oracle" required-features = ["qoa"] diff --git a/crates/audio/README.md b/crates/audio/README.md index 63251db8d..7469a9401 100644 --- a/crates/audio/README.md +++ b/crates/audio/README.md @@ -29,14 +29,16 @@ 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-decode` | `flac::decode` (whole `.flac` files) and `flac::Decoder` (single frames, for callers that demux FLAC themselves); native, CRC-checked, 4 to 32 bits. `decode` still uses symphonia for FLAC files | none | | `flac-encode` | `encode`, `FileEncoder` for FLAC; `flac::Encoder`, the packet encoder for callers that mux FLAC themselves | craft-media `bitstream` | | `aac-decode`, `aac-encode` | `aac::Decoder`, `aac::Encoder`; with `aac-decode`, `aac::latm_config` / `latm_split` for LOAS / LATM frames (AAC in MPEG transport streams); with `aac-encode`, `encode`, `FileEncoder` for AAC in M4A | craft-media `bitstream`; `aac-encode` also `isobmff` | | `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 +357,12 @@ 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 +374,27 @@ 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 +731,7 @@ 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/flac_decode.rs b/crates/audio/src/flac_decode.rs new file mode 100644 index 000000000..bb56e65c5 --- /dev/null +++ b/crates/audio/src/flac_decode.rs @@ -0,0 +1,603 @@ +//! A native FLAC decoder (RFC 9639): [`decode`] for whole `.flac` files and [`Decoder`] for single +//! frames (the packets of FLAC in MP4, Matroska or Ogg). +//! +//! Every frame is checked against its CRC-8 and CRC-16. Subframe types constant, verbatim, fixed +//! and LPC; Rice partitions with 4- and 5-bit parameters and escaped partitions; wasted bits; +//! left/side, side/right and mid/side stereo; fixed and variable block sizes; 4 to 32 bits per +//! sample. The STREAMINFO MD5 is not verified. +//! +//! Samples come back as `i32`, one `Vec` per channel, in the FLAC channel order. + +/// What went wrong decoding FLAC. +#[non_exhaustive] +#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)] +pub enum DecodeError { + /// The data ends early. + #[error("truncated FLAC data")] + Truncated, + /// The data is not valid FLAC. + #[error("invalid FLAC data: {0}")] + Invalid(&'static str), + /// A frame or header failed its CRC. + #[error("FLAC CRC mismatch in {0}")] + Crc(&'static str), + /// Valid FLAC that is not decoded. + #[error("unsupported FLAC feature: {0}")] + Unsupported(&'static str), + /// The stream is longer than the sample limit. + #[error("FLAC stream exceeds the sample limit")] + TooLarge, +} + +/// `Result` of this module. +pub type Result = std::result::Result; + +/// The STREAMINFO metadata block (RFC 9639 section 8.2). +#[derive(Debug, Clone, Copy, PartialEq, Eq)] +pub struct StreamInfo { + /// Smallest block size in samples. + pub min_block_size: u16, + /// Largest block size in samples. + pub max_block_size: u16, + /// Smallest frame in bytes, 0 if unknown. + pub min_frame_size: u32, + /// Largest frame in bytes, 0 if unknown. + pub max_frame_size: u32, + /// Sample rate in Hz. + pub sample_rate: u32, + /// Channel count, 1 to 8. + pub channels: u8, + /// Bits per sample, 4 to 32. + pub bits_per_sample: u8, + /// Samples per channel in the stream, 0 if unknown. + pub total_samples: u64, + /// MD5 of the unencoded audio, all zero if not set. + pub md5: [u8; 16], +} + +impl StreamInfo { + /// Parse the 34 bytes of a STREAMINFO block body. + pub fn parse(b: &[u8]) -> Result { + let b = b.get(..34).ok_or(DecodeError::Truncated)?; + let packed = u64::from_be_bytes([b[10], b[11], b[12], b[13], b[14], b[15], b[16], b[17]]); + let info = StreamInfo { + min_block_size: u16::from_be_bytes([b[0], b[1]]), + max_block_size: u16::from_be_bytes([b[2], b[3]]), + min_frame_size: u32::from_be_bytes([0, b[4], b[5], b[6]]), + max_frame_size: u32::from_be_bytes([0, b[7], b[8], b[9]]), + sample_rate: (packed >> 44) as u32, + channels: ((packed >> 41) & 7) as u8 + 1, + bits_per_sample: ((packed >> 36) & 31) as u8 + 1, + total_samples: packed & 0xF_FFFF_FFFF, + md5: b[18..34].try_into().map_err(|_| DecodeError::Truncated)?, + }; + if info.bits_per_sample < 4 { + return Err(DecodeError::Invalid("bits per sample below 4")); + } + if info.min_block_size != 0 + && (info.min_block_size < 16 && info.min_block_size != info.max_block_size || info.max_block_size < info.min_block_size) + { + return Err(DecodeError::Invalid("inconsistent block sizes")); + } + Ok(info) + } +} + +/// One decoded frame. +#[derive(Debug, Clone, PartialEq, Eq)] +pub struct Frame { + /// Sample rate in Hz. + pub sample_rate: u32, + /// Bits per sample. + pub bits_per_sample: u8, + /// `true` when `number` is the first sample's index (variable block size), `false` when it is + /// the frame's index (fixed block size). + pub variable_block_size: bool, + /// The coded number of the frame header. + pub number: u64, + /// Samples per channel in the frame. + pub block_size: usize, + /// One `Vec` per channel, after undoing any stereo decorrelation. + pub channels: Vec>, + /// Bytes the frame took, including the header and CRC footer. + pub size: usize, +} + +// ---- checksums ---- + +fn crc8(data: &[u8]) -> u8 { + let mut crc = 0u8; + for &b in data { + crc ^= b; + for _ in 0..8 { + crc = if crc & 0x80 != 0 { crc << 1 ^ 0x07 } else { crc << 1 }; + } + } + crc +} + +fn crc16(data: &[u8]) -> u16 { + let mut crc = 0u16; + for &b in data { + crc ^= u16::from(b) << 8; + for _ in 0..8 { + crc = if crc & 0x8000 != 0 { crc << 1 ^ 0x8005 } else { crc << 1 }; + } + } + crc +} + +// ---- bits ---- + +struct Bits<'a> { + data: &'a [u8], + /// Position in bits. + pos: usize, +} + +impl<'a> Bits<'a> { + fn new(data: &'a [u8]) -> Self { + Bits { data, pos: 0 } + } + + fn left(&self) -> usize { + self.data.len() * 8 - self.pos + } + + fn bit(&mut self) -> Result { + let byte = *self.data.get(self.pos / 8).ok_or(DecodeError::Truncated)?; + let v = (byte >> (7 - self.pos % 8)) & 1; + self.pos += 1; + Ok(u32::from(v)) + } + + /// `n` bits (0 to 32) as an unsigned number. + fn bits(&mut self, n: u32) -> Result { + if n as usize > self.left() { + return Err(DecodeError::Truncated); + } + let mut v = 0u64; + let mut n = n as usize; + while n > 0 { + let byte = u64::from(self.data[self.pos / 8]); + let avail = 8 - self.pos % 8; + let take = avail.min(n); + v = v << take | (byte >> (avail - take)) & ((1 << take) - 1); + self.pos += take; + n -= take; + } + Ok(v) + } + + /// `n` bits (0 to 33) as a two's complement number. + fn signed(&mut self, n: u32) -> Result { + if n == 0 { + return Ok(0); + } + let v = self.bits(n)? as i64; + Ok(if v >> (n - 1) & 1 == 1 { v - (1 << n) } else { v }) + } + + /// Zero bits up to the next one bit (which is consumed). + fn unary(&mut self) -> Result { + let mut n = 0u64; + loop { + // Skip whole zero bytes when aligned. + if self.pos.is_multiple_of(8) { + while self.data.get(self.pos / 8) == Some(&0) { + self.pos += 8; + n += 8; + } + } + if self.bit()? == 1 { + return Ok(n); + } + n += 1; + } + } + + fn align(&mut self) { + self.pos = self.pos.div_ceil(8) * 8; + } +} + +// ---- frame ---- + +const BLOCK_SIZES: [usize; 16] = [0, 192, 576, 1152, 2304, 4608, 0, 0, 256, 512, 1024, 2048, 4096, 8192, 16384, 32768]; + +/// A frame decoder. `info` supplies what a frame header leaves to the STREAMINFO. +#[derive(Debug, Clone, Copy, Default)] +pub struct Decoder { + info: Option, +} + +impl Decoder { + /// A decoder with the stream's STREAMINFO (needed for frames whose header defers the sample + /// rate or bit depth to it). + pub fn new(info: StreamInfo) -> Decoder { + Decoder { info: Some(info) } + } + + /// A decoder without STREAMINFO: frames that defer to it are an error. + pub fn without_streaminfo() -> Decoder { + Decoder { info: None } + } + + /// Decode the frame at the start of `data`. + pub fn decode_frame(&self, data: &[u8]) -> Result { + // StreamInfo has public fields: container callers can construct it without + // parse(), so validate before it can control shifts and sample reads. + if self.info.is_some_and(|i| !(4..=32).contains(&i.bits_per_sample) || !(1..=8).contains(&i.channels) || i.sample_rate == 0) { + return Err(DecodeError::Invalid("invalid decoder STREAMINFO")); + } + let mut r = Bits::new(data); + if r.bits(15)? != 0x7FFC { + return Err(DecodeError::Invalid("missing frame sync code")); + } + let variable = r.bit()? == 1; + let (bs_code, sr_code) = (r.bits(4)? as usize, r.bits(4)? as usize); + let (ch_code, bps_code) = (r.bits(4)? as usize, r.bits(3)? as usize); + if r.bit()? != 0 { + return Err(DecodeError::Invalid("reserved bit set")); + } + let number = coded_number(&mut r)?; + let block_size = match bs_code { + 0 => return Err(DecodeError::Invalid("reserved block size")), + 6 => r.bits(8)? as usize + 1, + 7 => r.bits(16)? as usize + 1, + c => BLOCK_SIZES[c], + }; + let sample_rate = match sr_code { + 0 => self.info.map(|i| i.sample_rate).ok_or(DecodeError::Invalid("sample rate only in STREAMINFO"))?, + 1 => 88_200, + 2 => 176_400, + 3 => 192_000, + 4 => 8_000, + 5 => 16_000, + 6 => 22_050, + 7 => 24_000, + 8 => 32_000, + 9 => 44_100, + 10 => 48_000, + 11 => 96_000, + 12 => r.bits(8)? as u32 * 1000, + 13 => r.bits(16)? as u32, + 14 => r.bits(16)? as u32 * 10, + _ => return Err(DecodeError::Invalid("forbidden sample rate code")), + }; + let header_len = r.pos.div_ceil(8); + let stored_crc = r.bits(8)? as u8; + if crc8(&data[..header_len]) != stored_crc { + return Err(DecodeError::Crc("a frame header")); + } + let bits = match bps_code { + 0 => self.info.map(|i| u32::from(i.bits_per_sample)).ok_or(DecodeError::Invalid("bit depth only in STREAMINFO"))?, + 1 => 8, + 2 => 12, + 4 => 16, + 5 => 20, + 6 => 24, + 7 => 32, + _ => return Err(DecodeError::Invalid("reserved bit depth")), + }; + let (channels, mode) = match ch_code { + 0..=7 => (ch_code + 1, Stereo::Independent), + 8 => (2, Stereo::LeftSide), + 9 => (2, Stereo::SideRight), + 10 => (2, Stereo::MidSide), + _ => return Err(DecodeError::Invalid("reserved channel assignment")), + }; + let mut out: Vec> = Vec::with_capacity(channels); + for c in 0..channels { + let extra = match (mode, c) { + (Stereo::LeftSide, 1) | (Stereo::SideRight, 0) | (Stereo::MidSide, 1) => 1, + _ => 0, + }; + out.push(subframe(&mut r, block_size, bits + extra)?); + } + r.align(); + let size = r.pos / 8; + let crc = r.bits(16)? as u16; + if crc16(&data[..size]) != crc { + return Err(DecodeError::Crc("a frame")); + } + let size = size + 2; + if mode != Stereo::Independent { + let (first, rest) = out.split_at_mut(1); + let (a, b) = (&mut first[0], &mut rest[0]); + for (x, y) in a.iter_mut().zip(b.iter_mut()) { + match mode { + Stereo::LeftSide => *y = x.wrapping_sub(*y), + Stereo::SideRight => *x = x.wrapping_add(*y), + _ => { + let m = x.wrapping_shl(1) | (*y & 1); + let side = *y; + *x = m.wrapping_add(side) >> 1; + *y = m.wrapping_sub(side) >> 1; + } + } + } + } + let mut chans = Vec::with_capacity(channels); + for c in out { + let v: Result> = c.into_iter().map(|s| checked_sample(s, bits)).collect(); + chans.push(v?); + } + Ok(Frame { sample_rate, bits_per_sample: bits as u8, variable_block_size: variable, number, block_size, channels: chans, size }) + } +} + +#[derive(Clone, Copy, PartialEq, Eq)] +enum Stereo { + Independent, + LeftSide, + SideRight, + MidSide, +} + +/// The UTF-8-like coded number of a frame header (up to 36 bits, 7 bytes). +fn coded_number(r: &mut Bits) -> Result { + let first = r.bits(8)? as u8; + let extra = match first { + 0x00..=0x7F => return Ok(u64::from(first)), + 0xC0..=0xDF => 1, + 0xE0..=0xEF => 2, + 0xF0..=0xF7 => 3, + 0xF8..=0xFB => 4, + 0xFC..=0xFD => 5, + 0xFE => 6, + _ => return Err(DecodeError::Invalid("bad coded number")), + }; + let mut v = if extra == 6 { 0 } else { u64::from(first & (0x3F >> extra)) }; + for _ in 0..extra { + let b = r.bits(8)?; + if b & 0xC0 != 0x80 { + return Err(DecodeError::Invalid("bad coded number")); + } + v = v << 6 | (b & 0x3F); + } + Ok(v) +} + +/// One subframe: `bits` is the frame bit depth (plus 1 for a side channel). +fn subframe(r: &mut Bits, n: usize, bits: u32) -> Result> { + if r.bit()? != 0 { + return Err(DecodeError::Invalid("subframe header padding bit set")); + } + let kind = r.bits(6)? as u32; + let wasted = if r.bit()? == 1 { r.unary()? as u32 + 1 } else { 0 }; + if wasted >= bits { + return Err(DecodeError::Invalid("wasted bits leave no sample bits")); + } + let bps = bits - wasted; + let mut s = vec![0i64; n]; + match kind { + 0 => { + let v = r.signed(bps)?; + s.fill(v); + } + 1 => { + for v in s.iter_mut() { + *v = r.signed(bps)?; + } + } + 8..=12 => { + let order = (kind - 8) as usize; + warmup(r, &mut s, order, bps)?; + residual(r, &mut s, order)?; + for i in order..n { + let at = |k: usize| s[i - k]; + // Wrapping: hostile residuals can overflow; such samples fail the range check later. + let p = match order { + 0 => 0, + 1 => at(1), + 2 => at(1).wrapping_mul(2).wrapping_sub(at(2)), + 3 => at(1).wrapping_mul(3).wrapping_sub(at(2).wrapping_mul(3)).wrapping_add(at(3)), + _ => at(1).wrapping_mul(4).wrapping_sub(at(2).wrapping_mul(6)).wrapping_add(at(3).wrapping_mul(4)).wrapping_sub(at(4)), + }; + s[i] = s[i].wrapping_add(p); + } + } + 32..=63 => { + let order = (kind - 31) as usize; + warmup(r, &mut s, order, bps)?; + let precision = r.bits(4)? as u32 + 1; + if precision == 16 { + return Err(DecodeError::Invalid("forbidden coefficient precision")); + } + let shift = r.signed(5)?; + if shift < 0 { + return Err(DecodeError::Invalid("negative prediction shift")); + } + let mut coefs = [0i64; 32]; + for c in coefs.iter_mut().take(order) { + *c = r.signed(precision)?; + } + residual(r, &mut s, order)?; + for i in order..n { + let mut sum = 0i64; + for (k, &c) in coefs.iter().take(order).enumerate() { + sum = sum.wrapping_add(c.wrapping_mul(s[i - 1 - k])); + } + s[i] = s[i].wrapping_add(sum >> shift); + } + } + _ => return Err(DecodeError::Invalid("reserved subframe type")), + } + if wasted > 0 { + for v in s.iter_mut() { + *v = v.wrapping_shl(wasted); + } + } + Ok(s) +} + +fn warmup(r: &mut Bits, s: &mut [i64], order: usize, bps: u32) -> Result<()> { + if order > s.len() { + return Err(DecodeError::Invalid("predictor order above the block size")); + } + for v in s.iter_mut().take(order) { + *v = r.signed(bps)?; + } + Ok(()) +} + +/// The coded residual (RFC 9639 section 9.2.7), into `s[order..]`. +fn residual(r: &mut Bits, s: &mut [i64], order: usize) -> Result<()> { + let method = r.bits(2)?; + let (param_bits, escape) = match method { + 0 => (4, 15), + 1 => (5, 31), + _ => return Err(DecodeError::Invalid("reserved residual coding method")), + }; + let part_order = r.bits(4)? as u32; + let n = s.len(); + let parts = 1usize << part_order; + if !n.is_multiple_of(parts) { + return Err(DecodeError::Invalid("partition order does not divide the block size")); + } + let per = n >> part_order; + if per < order { + return Err(DecodeError::Invalid("partition shorter than the predictor order")); + } + let mut i = order; + for p in 0..parts { + let count = if p == 0 { per - order } else { per }; + let k = r.bits(param_bits)? as u32; + if k == escape { + let width = r.bits(5)? as u32; + for v in s[i..i + count].iter_mut() { + *v = r.signed(width)?; + } + } else { + for v in s[i..i + count].iter_mut() { + let q = r.unary()?; + if q > (1 << 32) { + return Err(DecodeError::Invalid("Rice quotient too large")); + } + let low = r.bits(k)?; + let folded = q << k | low; + *v = if folded & 1 == 0 { (folded >> 1) as i64 } else { -((folded >> 1) as i64) - 1 }; + } + } + i += count; + } + Ok(()) +} + +// ---- files ---- + +/// A decoded file. +#[derive(Debug, Clone, PartialEq, Eq)] +pub struct Decoded { + /// The STREAMINFO block. + pub info: StreamInfo, + /// One `Vec` of samples per channel. + pub channels: Vec>, +} + +/// Most sample values (samples x channels) [`decode`] will produce. +pub const DEFAULT_MAX_VALUES: u64 = 1 << 28; + +/// Decode a whole `.flac` file (an ID3v2 tag in front is skipped). +pub fn decode(bytes: &[u8]) -> Result { + decode_with_limit(bytes, DEFAULT_MAX_VALUES) +} + +/// [`decode`] with a limit on the sample values (samples x channels) produced. +pub fn decode_with_limit(bytes: &[u8], max_values: u64) -> Result { + let mut pos = 0usize; + if bytes.starts_with(b"ID3") { + let size = bytes.get(6..10).ok_or(DecodeError::Truncated)?; + if size.iter().any(|&b| b & 0x80 != 0) { + return Err(DecodeError::Invalid("bad ID3v2 size")); + } + pos = 10 + size.iter().fold(0usize, |a, &b| a << 7 | usize::from(b)); + } + if bytes.get(pos..pos + 4) != Some(b"fLaC") { + return Err(DecodeError::Invalid("missing fLaC marker")); + } + pos += 4; + let mut info: Option = None; + loop { + let head = bytes.get(pos..pos + 4).ok_or(DecodeError::Truncated)?; + let last = head[0] & 0x80 != 0; + let kind = head[0] & 0x7F; + let len = usize::from(head[1]) << 16 | usize::from(head[2]) << 8 | usize::from(head[3]); + let body = bytes.get(pos + 4..pos + 4 + len).ok_or(DecodeError::Truncated)?; + match (kind, &info) { + (0, None) => info = Some(StreamInfo::parse(body)?), + (0, Some(_)) => return Err(DecodeError::Invalid("a second STREAMINFO block")), + (_, None) => return Err(DecodeError::Invalid("STREAMINFO is not the first block")), + (127, _) => return Err(DecodeError::Invalid("forbidden metadata block type")), + _ => {} + } + pos += 4 + len; + if last { + break; + } + } + let info = info.ok_or(DecodeError::Invalid("no STREAMINFO"))?; + let dec = Decoder::new(info); + let mut channels: Vec> = vec![Vec::new(); usize::from(info.channels)]; + let mut total = 0u64; + while pos + 2 <= bytes.len() && bytes[pos] == 0xFF && bytes[pos + 1] & 0xFE == 0xF8 { + let frame = dec.decode_frame(&bytes[pos..])?; + if frame.channels.len() != channels.len() { + return Err(DecodeError::Invalid("the channel count changes inside the stream")); + } + // One file has one sample rate and depth. Otherwise callers would play + // samples under incompatible STREAMINFO metadata. + if frame.sample_rate != info.sample_rate || frame.bits_per_sample != info.bits_per_sample { + return Err(DecodeError::Invalid("frame format differs from STREAMINFO")); + } + total += frame.block_size as u64; + if total.saturating_mul(channels.len() as u64) > max_values { + return Err(DecodeError::TooLarge); + } + for (dst, src) in channels.iter_mut().zip(frame.channels) { + dst.extend_from_slice(&src); + } + pos += frame.size; + } + // Unknown total_samples must not turn a corrupt/truncated next frame into + // a successfully decoded shorter recording. + if pos != bytes.len() { + return Err(DecodeError::Invalid("trailing data or missing frame sync")); + } + if info.total_samples != 0 && total != info.total_samples { + return Err(DecodeError::Truncated); + } + Ok(Decoded { info, channels }) +} + +// Validate the advertised precision, not just the storage type. Malformed +// predictors may produce an i32 that is still invalid for 4–24-bit audio. +fn checked_sample(sample: i64, bits: u32) -> Result { + if !(4..=32).contains(&bits) { + return Err(DecodeError::Invalid("invalid sample bit depth")); + } + let bound = 1i64 << (bits - 1); + if !(-bound..bound).contains(&sample) { + return Err(DecodeError::Invalid("sample outside the bit depth")); + } + i32::try_from(sample).map_err(|_| DecodeError::Invalid("sample outside the bit depth")) +} + +#[cfg(test)] +mod sample_bounds_tests { + use super::checked_sample; + + #[test] + fn samples_must_fit_declared_precision() { + for bits in 4..=32 { + let bound = 1i64 << (bits - 1); + assert_eq!(checked_sample(-bound, bits).map(i64::from), Ok(-bound)); + assert_eq!(checked_sample(bound - 1, bits).map(i64::from), Ok(bound - 1)); + assert!(checked_sample(-bound - 1, bits).is_err()); + assert!(checked_sample(bound, bits).is_err()); + } + assert!(checked_sample(0, 0).is_err()); + assert!(checked_sample(0, 255).is_err()); + } +} diff --git a/crates/audio/src/lib.rs b/crates/audio/src/lib.rs index be0678303..93fbec58d 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 @@ -27,8 +29,8 @@ //! - Codec modules for packet-level use, each with its own features: [`aac`] (`aac-decode`, //! `aac-encode`), [`alac`] (`alac-decode`, `alac-encode`), [`opus`] (`opus-decode`, and //! [`opus::encoder`] with `opus-encode`), [`ac3`] (`ac3-decode`), [`mpa`] (`mpa-decode`), [`mp3`] -//! (`mp3-decode`, `mp3-encode`), [`flac`] (`flac-encode`: the packet encoder), [`qoa`] -//! (`qoa-decode`, `qoa-encode`: whole QOA files). +//! (`mp3-decode`, `mp3-encode`), [`flac`] (`flac-decode`: the frame and file decoder; +//! `flac-encode`: the packet encoder), [`qoa`] (`qoa-decode`, `qoa-encode`: whole QOA files). //! //! The shared types below build with no features. @@ -56,6 +58,8 @@ 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. @@ -72,10 +76,12 @@ pub mod wav_markers; ))] 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 @@ -96,10 +102,15 @@ pub use write::{OGG_BITRATES, encode_ogg}; /// assert!(file.len() < 48_000 * 2 * 2 / 2); // well under half of 16-bit PCM /// # Ok::<(), craft_audio::flac::Error>(()) /// ``` -#[cfg(feature = "flac-encode")] +#[cfg(any(feature = "flac-encode", feature = "flac-decode"))] pub mod flac { + #[cfg(feature = "flac-decode")] + pub use crate::flac_decode::{DEFAULT_MAX_VALUES, DecodeError, Decoded, Decoder, Frame, StreamInfo, decode, decode_with_limit}; + #[cfg(feature = "flac-encode")] pub use crate::write::flac::encoder::{Encoder, EncoderConfig, Error, STREAMINFO_OFFSET}; } +#[cfg(feature = "flac-decode")] +mod flac_decode; pub use buffer::AudioBuffer; /// The craft-media this crate is built with, for its containers: using it rather than a separate @@ -278,7 +289,9 @@ 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, } @@ -379,7 +392,7 @@ 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"), } } @@ -400,9 +413,9 @@ 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, } } @@ -416,7 +429,10 @@ 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. @@ -458,10 +474,10 @@ 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 @@ -480,11 +496,12 @@ 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, π]`. +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 · π 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(2π/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^{-2πi 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/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", "Smörgåsbord ✓") + .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("Smörgåsbord ✓"), 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/audio/tests/flac_decode.rs b/crates/audio/tests/flac_decode.rs new file mode 100644 index 000000000..a1895789b --- /dev/null +++ b/crates/audio/tests/flac_decode.rs @@ -0,0 +1,211 @@ +//! The native FLAC decoder: round trips through this crate's encoder over bit depths, channel +//! counts, block sizes and effort levels; hostile input; and ffmpeg as the reference (oracle; +//! skipped when ffmpeg is absent, a failure with `CRAFT_MEDIA_REQUIRE_ORACLES=1`) for files +//! ffmpeg's encoder writes at every compression level. + +use craft_audio::flac::{DecodeError, Decoder, Encoder, EncoderConfig, STREAMINFO_OFFSET, StreamInfo, decode}; +use std::process::Command; + +fn signal(channels: usize, frames: usize, bits: u32, seed: u32) -> Vec> { + let max = if bits == 32 { i32::MAX as f64 } else { ((1i64 << (bits - 1)) - 1) as f64 }; + let mut s = seed; + (0..channels) + .map(|c| { + (0..frames) + .map(|i| { + s ^= s << 13; + s ^= s >> 17; + s ^= s << 5; + let t = i as f64; + let v = match (i / 3000) % 4 { + 0 => (t * (0.01 + 0.007 * c as f64)).sin() * 0.6 + (t * 0.31).sin() * 0.1, + 1 => (f64::from(s) / f64::from(u32::MAX) - 0.5) * 1.8, + 2 => 0.0, + _ => { + if (i / 40) % 2 == 0 { + 0.99 + } else { + -0.99 + } + } + }; + (v * max).round() as i32 + }) + .collect() + }) + .collect() +} + +fn encode(channels: &[Vec], rate: u32, bits: u8, block: u16, level: u8) -> Vec { + let mut cfg = EncoderConfig::new(rate, channels.len() as u8, bits); + cfg.block_size = block; + cfg.level = level; + let mut enc = Encoder::new(cfg).expect("encoder"); + let mut file = enc.header(&[]); + let refs: Vec<&[i32]> = channels.iter().map(Vec::as_slice).collect(); + file.extend(enc.encode(&refs).expect("encode")); + file.extend(enc.finish()); + let at = STREAMINFO_OFFSET as usize; + file[at..at + 34].copy_from_slice(&enc.streaminfo()); + file +} + +#[test] +fn round_trips_through_the_encoder() { + for (channels, bits) in [(1usize, 8u8), (2, 8), (1, 12), (2, 16), (2, 20), (3, 24), (6, 24), (8, 16), (2, 32), (1, 32)] { + for (block, level) in [(4096u16, 5u8), (16, 0), (1000, 8), (65535, 3), (576, 1)] { + let frames = 9000 + usize::from(block % 97); + let src = signal(channels, frames, u32::from(bits), 0x1234_5679 + u32::from(bits)); + let file = encode(&src, 44_100, bits, block, level); + let got = decode(&file).unwrap_or_else(|e| panic!("{channels}ch {bits}-bit block {block} level {level}: {e}")); + assert_eq!((got.info.channels as usize, got.info.bits_per_sample, got.info.sample_rate), (channels, bits, 44_100)); + assert_eq!(got.channels, src, "{channels}ch {bits}-bit block {block} level {level}"); + } + } +} + +#[test] +fn frames_decode_one_at_a_time() { + let src = signal(2, 10_000, 16, 7); + let file = encode(&src, 48_000, 16, 4096, 5); + let info = StreamInfo::parse(&file[STREAMINFO_OFFSET as usize..]).expect("streaminfo"); + let dec = Decoder::new(info); + // The first frame follows the metadata blocks; find the first position a frame decodes at. + let at = (42..file.len()).find(|&p| dec.decode_frame(&file[p..]).is_ok()).expect("a frame"); + let first = dec.decode_frame(&file[at..]).expect("frame"); + assert_eq!((first.block_size, first.number, first.sample_rate, first.bits_per_sample), (4096, 0, 48_000, 16)); + assert_eq!(first.channels[0], src[0][..4096]); + // Without the STREAMINFO a frame that spells out its own rate and depth still decodes. + let _ = Decoder::without_streaminfo().decode_frame(&file[at..]); +} + +#[test] +fn hostile_input_is_an_error_not_a_panic() { + let src = signal(2, 6000, 16, 11); + let file = encode(&src, 44_100, 16, 1152, 5); + for n in 0..file.len().min(3000) { + let _ = decode(&file[..n]); + } + // A cut file is truncated, not silently shorter. + assert!(decode(&file[..file.len() - 5]).is_err()); + // Every bit flip in the first frames is caught by a CRC or another check, never a panic. + for i in 0..file.len().min(2500) { + for bit in 0..8 { + let mut c = file.clone(); + c[i] ^= 1 << bit; + let _ = decode(&c); + } + } + // A sample limit stops a stream that claims too much. + assert_eq!(craft_audio::flac::decode_with_limit(&file, 1000), Err(DecodeError::TooLarge)); + assert!(decode(b"nope").is_err()); + assert!(decode(b"fLaC").is_err()); +} + +// ---- ffmpeg ---- + +/// Little-endian 16- or 32-bit samples of raw PCM. +fn pcm_values(raw: &[u8], bytes: usize) -> Vec { + if bytes == 2 { + raw.as_chunks::<2>().0.iter().map(|b| i32::from(i16::from_le_bytes(*b))).collect() + } else { + raw.as_chunks::<4>().0.iter().map(|b| i32::from_le_bytes(*b)).collect() + } +} + +/// Planar channels as interleaved samples. +fn interleave(channels: &[Vec]) -> Vec { + let frames = channels.first().map_or(0, Vec::len); + (0..frames).flat_map(|i| channels.iter().map(move |c| c[i])).collect() +} + +fn ffmpeg_run(ff: &std::path::Path, args: &[&str]) -> Vec { + let out = Command::new(ff).args(["-v", "error", "-nostdin", "-y"]).args(args).output().expect("run ffmpeg"); + assert!(out.status.success(), "ffmpeg {args:?}: {}", String::from_utf8_lossy(&out.stderr)); + out.stdout +} + +#[test] +fn ffmpeg_files_decode_to_ffmpegs_samples() { + let Some(ff) = craft_media_testkit::ffmpeg() else { return }; + let dir = craft_media_testkit::fixture_subdir("flac_decode"); + let tones = "aevalsrc=exprs=0.4*sin(2*PI*441*t)+0.2*sin(2*PI*3100*t)|0.3*sin(2*PI*660*t)*sin(2*PI*3*t):s=44100:d=1.3"; + let noise = "anoisesrc=d=1.3:c=white:r=44100:a=0.5:s=3"; + for (src, name) in [(tones, "tones"), (noise, "noise")] { + for channels in [1usize, 2, 6] { + for (fmt, bytes, bits) in [("s16", 2usize, 16u32), ("s32", 4, 24)] { + for level in [0, 3, 5, 8, 12] { + let file = dir.join(format!("ffmpeg_{name}_{channels}ch_{bits}_l{level}.flac")); + let f = file.to_str().expect("path"); + let (ch, lv) = (channels.to_string(), level.to_string()); + ffmpeg_run(&ff, &["-f", "lavfi", "-i", src, "-ac", &ch, "-sample_fmt", fmt, "-compression_level", &lv, "-c:a", "flac", f]); + let theirs = ffmpeg_run(&ff, &["-i", f, "-f", if bits == 16 { "s16le" } else { "s32le" }, "-"]); + let got = + decode(&std::fs::read(&file).expect("read")).unwrap_or_else(|e| panic!("{name} {channels}ch {bits}-bit level {level}: {e}")); + assert_eq!(got.info.channels as usize, channels); + let frames = got.channels[0].len(); + assert_eq!(theirs.len(), frames * channels * bytes, "{name} {channels}ch {bits} l{level}: length"); + // ffmpeg returns 24-bit audio left-aligned in 32 bits. + let shift = if bits == 24 { 8 } else { 0 }; + let mine: Vec = interleave(&got.channels).into_iter().map(|v| v << shift).collect(); + let ok = mine == pcm_values(&theirs, bytes); + assert!(ok, "{name} {channels}ch {bits}-bit level {level}: samples differ from ffmpeg's decode"); + } + } + } + } +} + +#[test] +fn ffmpeg_decodes_our_files_to_our_samples() { + let Some(ff) = craft_media_testkit::ffmpeg() else { return }; + let dir = craft_media_testkit::fixture_subdir("flac_decode"); + for (channels, bits) in [(1usize, 16u8), (2, 16), (2, 24), (6, 16)] { + let src = signal(channels, 20_000, u32::from(bits), 99); + let file = dir.join(format!("ours_{channels}ch_{bits}.flac")); + std::fs::write(&file, encode(&src, 44_100, bits, 4096, 5)).expect("write"); + let theirs = ffmpeg_run(&ff, &["-i", file.to_str().expect("path"), "-f", if bits == 16 { "s16le" } else { "s32le" }, "-"]); + let got = decode(&std::fs::read(&file).expect("read")).expect("decode"); + let bytes = if bits == 16 { 2 } else { 4 }; + let shift = if bytes == 4 { 8 } else { 0 }; + let want: Vec = pcm_values(&theirs, bytes).into_iter().map(|v| v >> shift).collect(); + assert!(interleave(&got.channels) == want, "{channels}ch {bits}-bit: samples differ from ffmpeg's decode"); + } +} + +#[test] +fn rejects_unknown_length_corruption_and_streaminfo_format_mismatch() { + let src = signal(1, 100, 16, 7); + let original = encode(&src, 44_100, 16, 4096, 5); + let at = STREAMINFO_OFFSET as usize; + // STREAMINFO is not CRC-covered. Clear only the 36-bit total sample count. + let mut unknown = original.clone(); + unknown[at + 13] &= 0xf0; + unknown[at + 14..at + 18].fill(0); + assert_eq!(decode(&unknown).unwrap().channels, src); + for tail in [vec![0xff], vec![0x00, 0x00]] { + let mut broken = unknown.clone(); + broken.extend(tail); + assert!(decode(&broken).is_err()); + } + // The encoder spells these in frame headers, independent of STREAMINFO. + let mut wrong_rate = original.clone(); + let mut packed = u64::from_be_bytes(wrong_rate[at + 10..at + 18].try_into().unwrap()); + packed = (packed & ((1u64 << 44) - 1)) | (48_000u64 << 44); + wrong_rate[at + 10..at + 18].copy_from_slice(&packed.to_be_bytes()); + assert!(decode(&wrong_rate).is_err()); + let mut wrong_depth = original; + wrong_depth[at + 12] = (wrong_depth[at + 12] & !1) | 1; + wrong_depth[at + 13] = (wrong_depth[at + 13] & 0x0f) | 0x70; + assert!(decode(&wrong_depth).is_err()); +} + +#[test] +fn caller_constructed_streaminfo_is_validated_before_decoding() { + let file = encode(&signal(1, 100, 16, 9), 44_100, 16, 4096, 5); + let info = StreamInfo::parse(&file[STREAMINFO_OFFSET as usize..]).unwrap(); + for bits in [0, 1, 3, 33, 64, 255] { + let invalid = StreamInfo { bits_per_sample: bits, ..info }; + assert!(matches!(Decoder::new(invalid).decode_frame(&[]), Err(DecodeError::Invalid(_)))); + } +} 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/control-protocol/Cargo.toml b/crates/control-protocol/Cargo.toml new file mode 100644 index 000000000..f00bbc1c6 --- /dev/null +++ b/crates/control-protocol/Cargo.toml @@ -0,0 +1,15 @@ +[package] +name = "craft-control-protocol" +description = "Authenticated bounded JSON-lines control protocol" +version.workspace = true +edition.workspace = true +license.workspace = true +rust-version.workspace = true +repository.workspace = true + +[lints] +workspace = true + +[dependencies] +serde_json = { workspace = true } +subtle = "2" diff --git a/crates/control-protocol/src/lib.rs b/crates/control-protocol/src/lib.rs new file mode 100644 index 000000000..8bdb08d66 --- /dev/null +++ b/crates/control-protocol/src/lib.rs @@ -0,0 +1,234 @@ +//! Authenticated JSON-lines request framing and dispatch for local control services. +#![cfg_attr(not(test), deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable))] +use serde_json::{Value, json}; +use std::io::{BufRead, BufReader, Read, Write}; +use subtle::ConstantTimeEq; + +/// Whether a session token has the control protocol's 64 hexadecimal character form. +pub fn valid_token(token: &str) -> bool { + token.len() == 64 && token.bytes().all(|b| b.is_ascii_hexdigit()) +} + +/// Encode one client request, including its terminating newline. +pub fn request_line(id: u64, token: Option<&str>, method: &str, params: &Value) -> String { + format!("{}\n", json!({"id": id, "token": token, "method": method, "params": params})) +} + +/// Extract the result or application error from a parsed control reply. +pub fn reply_result(reply: &Value) -> Result { + if reply.get("ok").and_then(Value::as_bool) == Some(true) { + Ok(reply.get("result").cloned().unwrap_or(Value::Null)) + } else { + Err(reply.get("error").and_then(Value::as_str).unwrap_or("error").to_owned()) + } +} + +/// One parsed line. +enum Line { + Request { + id: Value, + method: String, + params: Value, + }, + Blank, + /// Not a request: reply with this error and close the connection. + Reject(String), +} + +fn parse_line(bytes: &[u8], token: &str) -> Line { + let Ok(text) = std::str::from_utf8(bytes) else { + return Line::Reject("not a control request (invalid UTF-8); closing the connection".into()); + }; + let text = text.trim(); + if text.is_empty() { + return Line::Blank; + } + let msg = match serde_json::from_str::(text) { + Ok(v) => v, + Err(e) => return Line::Reject(format!("bad JSON ({e}); expected one JSON request object per line; closing the connection")), + }; + let Some(obj) = msg.as_object() else { + return Line::Reject("not a control request (expected a JSON object); closing the connection".into()); + }; + let Some(method) = obj.get("method").and_then(Value::as_str) else { + return Line::Reject("not a control request (missing string `method`); closing the connection".into()); + }; + // Authenticate every request before constructing or enqueueing a ControlRequest. A valid + // earlier request must not authorize a later one with a missing or incorrect credential. + let authenticated = obj.get("token").and_then(Value::as_str).is_some_and(|supplied| bool::from(supplied.as_bytes().ct_eq(token.as_bytes()))); + if !authenticated { + return Line::Reject("control authentication failed; closing the connection".into()); + } + Line::Request { + id: obj.get("id").cloned().unwrap_or(Value::Null), + method: method.to_string(), + params: obj.get("params").cloned().unwrap_or(json!({})), + } +} + +#[derive(Debug, PartialEq, Eq)] +pub enum ReadFailure { + Timeout, + Rejected(String), +} + +/// Read one `\n`-terminated line of at most `max` bytes into `buf` (newline stripped). +/// `Ok(false)` at end of stream; `Err` when the line is too long or the read fails. +pub fn read_bounded_line(reader: &mut impl BufRead, buf: &mut Vec, max: usize) -> Result { + buf.clear(); + // `take` bounds the bytes buffered for one line: a peer can't make us allocate more than `max`. + let limit = u64::try_from(max).unwrap_or(u64::MAX).saturating_add(1); + let n = reader.take(limit).read_until(b'\n', buf).map_err(|e| match e.kind() { + std::io::ErrorKind::TimedOut | std::io::ErrorKind::WouldBlock => ReadFailure::Timeout, + _ => ReadFailure::Rejected(e.to_string()), + })?; + if n == 0 { + return Ok(false); + } + if buf.last() == Some(&b'\n') { + buf.pop(); + if buf.last() == Some(&b'\r') { + buf.pop(); + } + return Ok(true); + } + if buf.len() > max { + return Err(ReadFailure::Rejected(format!("request line longer than {max} bytes; closing the connection"))); + } + // Last line without a newline, then end of stream. + Ok(true) +} + +/// Serve requests from `read`, answering on `out`, until the peer closes, sends something that is +/// not a request, or `handle` returns `None` (app gone). `handle` gets `(method, params)` and +/// returns the reply object (the request's `id` is added here). +pub fn serve_lines(read: R, mut out: W, token: &str, max_line: usize, mut handle: impl FnMut(String, Value) -> Option) { + let mut reader = BufReader::new(read); + let mut buf = Vec::new(); + loop { + let reject = match read_bounded_line(&mut reader, &mut buf, max_line) { + Ok(false) => return, + // Expiration is not a request: a pooled client must see EOF and reconnect, rather + // than read a stale error reply. A partial line must not be parsed on timeout. + Err(ReadFailure::Timeout) => return, + Err(ReadFailure::Rejected(e)) => e, + Ok(true) => match parse_line(&buf, token) { + Line::Blank => continue, + Line::Reject(e) => e, + Line::Request { id, method, params } => { + let Some(mut r) = handle(method, params) else { return }; + if let Some(o) = r.as_object_mut() { + o.insert("id".into(), id); + } + if writeln!(out, "{r}").and_then(|()| out.flush()).is_err() { + return; + } + continue; + } + }, + }; + let _ = writeln!(out, "{}", json!({"ok": false, "error": reject})).and_then(|()| out.flush()); + return; + } +} + +#[cfg(test)] +mod tests { + use super::*; + #[test] + fn rejects_bad_request_shapes_without_dispatching() { + for input in [&b"\xff\n"[..], &b"[]\n"[..], &b"{}\n"[..]] { + let mut out = Vec::new(); + serve_lines(input, &mut out, "secret", 1024, |_, _| panic!("must not dispatch")); + let reply: Value = serde_json::from_slice(&out).unwrap(); + assert_eq!(reply["ok"], false); + assert!(reply["error"].as_str().unwrap().contains("closing the connection")); + } + } + #[test] + fn timeout_after_partial_line_never_dispatches() { + struct Partial(bool); + impl Read for Partial { + fn read(&mut self, buf: &mut [u8]) -> std::io::Result { + if !self.0 { + self.0 = true; + buf[..2].copy_from_slice(b"{\""); + return Ok(2); + } + Err(std::io::ErrorKind::TimedOut.into()) + } + } + let mut out = Vec::new(); + serve_lines(Partial(false), &mut out, "secret", 1024, |_, _| panic!("must not dispatch")); + assert!(out.is_empty()); + } + #[test] + fn line_boundaries_and_closed_dispatch_are_preserved() { + let mut buf = Vec::new(); + assert_eq!(read_bounded_line(&mut &b"abc\r\n"[..], &mut buf, 4), Ok(true)); + assert_eq!(buf, b"abc"); + assert_eq!(read_bounded_line(&mut &b"abcd\n"[..], &mut buf, 4), Ok(true)); + assert_eq!(buf, b"abcd"); + let mut out = Vec::new(); + serve_lines(&b"{\"method\":\"x\",\"token\":\"s\"}\n"[..], &mut out, "s", 1024, |_, _| None); + assert!(out.is_empty()); + } + #[test] + fn output_failure_stops_following_requests() { + struct Closed; + impl Write for Closed { + fn write(&mut self, _: &[u8]) -> std::io::Result { + Err(std::io::ErrorKind::BrokenPipe.into()) + } + fn flush(&mut self) -> std::io::Result<()> { + Ok(()) + } + } + let input = b"{\"method\":\"x\",\"token\":\"s\"}\n{\"method\":\"y\",\"token\":\"s\"}\n"; + let mut calls = 0; + serve_lines(&input[..], Closed, "s", 1024, |_, _| { + calls += 1; + Some(json!({"ok": true})) + }); + assert_eq!(calls, 1); + } + #[test] + fn client_request_and_reply_keep_optional_fields() { + let line = request_line(7, None, "status", &json!({"a": 1})); + assert_eq!(serde_json::from_str::(&line).unwrap(), json!({"id": 7, "token": null, "method": "status", "params": {"a": 1}})); + assert_eq!(reply_result(&json!({"ok": true})).unwrap(), Value::Null); + assert_eq!(reply_result(&json!({"ok": false, "error": "busy"})).unwrap_err(), "busy"); + } + #[test] + fn validates_session_token_shape() { + assert!(valid_token(&"aB09".repeat(16))); + assert!(!valid_token(&"a".repeat(63))); + assert!(!valid_token(&"g".repeat(64))); + } + #[test] + fn authenticates_each_request_and_closes_after_failure() { + let input = + b"{\"id\":1,\"method\":\"first\",\"token\":\"secret\"}\n{\"id\":2,\"method\":\"second\"}\n{\"method\":\"third\",\"token\":\"secret\"}\n"; + let mut output = Vec::new(); + let mut calls = Vec::new(); + serve_lines(&input[..], &mut output, "secret", 4096, |method, _| { + calls.push(method); + Some(json!({"ok": true})) + }); + assert_eq!(calls, ["first"]); + let lines: Vec = std::str::from_utf8(&output).unwrap().lines().map(|s| serde_json::from_str(s).unwrap()).collect(); + assert_eq!(lines[0], json!({"ok": true, "id": 1})); + assert_eq!(lines[1]["error"], "control authentication failed; closing the connection"); + } + #[test] + fn bounds_requests_and_preserves_last_unterminated_line() { + let mut buf = Vec::new(); + assert_eq!(read_bounded_line(&mut &b"abc"[..], &mut buf, 3), Ok(true)); + assert_eq!(buf, b"abc"); + assert!(matches!(read_bounded_line(&mut &b"abcd"[..], &mut buf, 3), Err(ReadFailure::Rejected(_)))); + let mut out = Vec::new(); + serve_lines(&b"{broken}\n"[..], &mut out, "secret", 1024, |_, _| panic!("must not dispatch")); + let reply: Value = serde_json::from_slice(&out).unwrap(); + assert!(reply["error"].as_str().unwrap().starts_with("bad JSON (")); + } +} 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/docx/src/chart_spec.rs b/crates/docx/src/chart_spec.rs index e31363903..a090c5c96 100644 --- a/crates/docx/src/chart_spec.rs +++ b/crates/docx/src/chart_spec.rs @@ -6,7 +6,7 @@ //! [`spec_of`] turns a chart part back into the model only when writing that model again gives //! the same markup, so a chart with anything the model leaves out stays as it was read. -use craft_wordmodel::chart::{ChartSeries, ChartSpec, ChartType, LegendPos, MAX_POINTS, MAX_SERIES}; +use craft_wordmodel::chart::{ChartAxis, ChartSeries, ChartSpec, ChartType, LegendPos, MAX_POINTS, MAX_SERIES}; use craft_wordmodel::graphic::GraphicItem; use craft_wordmodel::props::Rgb; @@ -102,11 +102,24 @@ fn bounded(spec: &ChartSpec) -> ChartSpec { .series .iter() .take(MAX_SERIES) - .map(|s| ChartSeries { name: s.name.clone(), values: s.values.iter().take(MAX_POINTS).copied().collect() }) + .map(|s| ChartSeries { + name: s.name.clone(), + values: s.values.iter().take(MAX_POINTS).copied().collect(), + fill: s.fill, + no_fill: s.no_fill, + line_w: s.line_w, + no_line: s.no_line, + markers: s.markers, + smooth: s.smooth, + labels: s.labels, + fmt: s.fmt.clone(), + }) .collect(), title: spec.title.clone(), legend: spec.legend, data_labels: spec.data_labels, + cat_axis: spec.cat_axis.clone(), + val_axis: spec.val_axis.clone(), } .sanitized() } @@ -185,17 +198,19 @@ fn plot(w: &mut W, spec: &ChartSpec) { match k { ChartType::Pie | ChartType::Doughnut => {} ChartType::Scatter => { - value_axis(w, CAT_AX, VAL_AX, "b", false, "midCat"); - value_axis(w, VAL_AX, CAT_AX, "l", true, "midCat"); + value_axis(w, CAT_AX, VAL_AX, "b", "midCat", &spec.cat_axis); + value_axis(w, VAL_AX, CAT_AX, "l", "midCat", &spec.val_axis); } _ => { - category_axis(w, if horizontal { "l" } else { "b" }); - value_axis(w, VAL_AX, CAT_AX, if horizontal { "b" } else { "l" }, true, "between"); + category_axis(w, if horizontal { "l" } else { "b" }, &spec.cat_axis); + value_axis(w, VAL_AX, CAT_AX, if horizontal { "b" } else { "l" }, "between", &spec.val_axis); } } } -/// One `c:ser` (ECMA-376 element order: idx, order, tx, spPr, marker, cat/xVal, val/yVal, smooth). +/// One `c:ser` (ECMA-376 element order: idx, order, tx, spPr, marker, invertIfNegative, dLbls, +/// cat/xVal, val/yVal, smooth): the series' formatting (fill, line, markers, its own labels) sits +/// in `c:spPr`, `c:marker` and `c:dLbls`; the values carry their number format. fn series(w: &mut W, spec: &ChartSpec, i: usize, s: &ChartSeries) { let k = spec.kind; let i = i.to_string(); @@ -207,25 +222,26 @@ fn series(w: &mut W, spec: &ChartSpec, i: usize, s: &ChartSeries) { w.leaf("c:v", &[], &s.name); w.close("c:tx"); } - match k { - // Markers only: the series has no line. - ChartType::Scatter => { - w.open("c:spPr", &[]); - w.open("a:ln", &[("w", "19050")]); - w.empty("a:noFill", &[]); - w.close("a:ln"); - w.close("c:spPr"); - } - ChartType::Column | ChartType::ColumnStacked | ChartType::Bar | ChartType::BarStacked => { - w.empty("c:invertIfNegative", &[("val", "0")]); - } - ChartType::Line => { - w.open("c:marker", &[]); - w.empty("c:symbol", &[("val", "none")]); - w.close("c:marker"); - } - _ => {} + series_props(w, s); + series_marker(w, k, s); + if matches!(k, ChartType::Column | ChartType::ColumnStacked | ChartType::Bar | ChartType::BarStacked) { + w.empty("c:invertIfNegative", &[("val", "0")]); } + if let Some(on) = s.labels { + w.open("c:dLbls", &[]); + for (name, show) in [ + ("c:showLegendKey", false), + ("c:showVal", on), + ("c:showCatName", false), + ("c:showSerName", false), + ("c:showPercent", false), + ("c:showBubbleSize", false), + ] { + w.empty(name, &[("val", if show { "1" } else { "0" })]); + } + w.close("c:dLbls"); + } + let code = s.fmt.as_deref().unwrap_or("General"); if k == ChartType::Scatter { let xs: Option>> = spec .categories @@ -234,27 +250,85 @@ fn series(w: &mut W, spec: &ChartSpec, i: usize, s: &ChartSeries) { .collect(); w.open("c:xVal", &[]); match xs { - Some(xs) => num_lit(w, &xs), + Some(xs) => num_lit(w, &xs, "General"), None => str_lit(w, &spec.categories), } w.close("c:xVal"); w.open("c:yVal", &[]); - num_lit(w, &s.values); + num_lit(w, &s.values, code); w.close("c:yVal"); } else { w.open("c:cat", &[]); str_lit(w, &spec.categories); w.close("c:cat"); w.open("c:val", &[]); - num_lit(w, &s.values); + num_lit(w, &s.values, code); w.close("c:val"); } if matches!(k, ChartType::Line | ChartType::Scatter) { - w.empty("c:smooth", &[("val", "0")]); + w.empty("c:smooth", &[("val", if s.smooth { "1" } else { "0" })]); } w.close("c:ser"); } +/// `c:spPr`: the series' fill (`a:solidFill` / `a:noFill`) and line (`a:ln`), written when the +/// model sets them. `a:ln` without `w` is the default width, `w` a fixed one, and `a:noFill` +/// inside it "no line" (a scatter series keeps its markers with no line). +fn series_props(w: &mut W, s: &ChartSeries) { + let fill_set = s.no_fill || s.fill.is_some(); + let ln_set = s.no_line || s.line_w.is_some(); + if !(fill_set || ln_set) { + return; + } + w.open("c:spPr", &[]); + if fill_set { + // "No fill" wins over a stored colour, so the pair reads back as it was written. + match &s.fill { + Some(c) if !s.no_fill => { + w.open("a:solidFill", &[]); + w.empty("a:srgbClr", &[("val", &c.hex())]); + w.close("a:solidFill"); + } + _ => w.empty("a:noFill", &[]), + } + } + if ln_set { + let w_attr = s.line_w.map(emu).unwrap_or_default(); + let attrs = [("w", w_attr.as_str())]; + w.open("a:ln", if w_attr.is_empty() { &[] } else { &attrs }); + if s.no_line { + w.empty("a:noFill", &[]); + } + w.close("a:ln"); + } + w.close("c:spPr"); +} + +/// `c:marker`: line and scatter series say which symbol each point gets; other types have none. +fn series_marker(w: &mut W, k: ChartType, s: &ChartSeries) { + match k { + ChartType::Line => { + w.open("c:marker", &[]); + w.empty("c:symbol", &[("val", if s.markers == Some(true) { "circle" } else { "none" })]); + w.close("c:marker"); + } + // Unset: no `c:marker`, so Word picks its automatic symbol for each series. + ChartType::Scatter => { + if let Some(on) = s.markers { + w.open("c:marker", &[]); + w.empty("c:symbol", &[("val", if on { "circle" } else { "none" })]); + w.close("c:marker"); + } + } + _ => {} + } +} + +/// A line width in points as the `w` attribute DrawingML wants (EMU). +fn emu(pt: f32) -> String { + ((pt * 12700.0).round().max(0.0) as i64).to_string() +} + fn str_lit(w: &mut W, items: &[String]) { w.open("c:strLit", &[]); w.empty("c:ptCount", &[("val", &items.len().to_string())]); @@ -266,9 +340,9 @@ fn str_lit(w: &mut W, items: &[String]) { w.close("c:strLit"); } -fn num_lit(w: &mut W, values: &[Option]) { +fn num_lit(w: &mut W, values: &[Option], code: &str) { w.open("c:numLit", &[]); - w.leaf("c:formatCode", &[], "General"); + w.leaf("c:formatCode", &[], code); w.empty("c:ptCount", &[("val", &values.len().to_string())]); for (i, v) in values.iter().enumerate() { if let Some(v) = v.filter(|v| v.is_finite()) { @@ -285,17 +359,21 @@ fn num(v: f64) -> String { if v == 0.0 { "0".to_string() } else { format!("{v}") } } -fn category_axis(w: &mut W, pos: &str) { +fn category_axis(w: &mut W, pos: &str, ax: &ChartAxis) { w.open("c:catAx", &[]); w.empty("c:axId", &[("val", CAT_AX)]); - w.open("c:scaling", &[]); - w.empty("c:orientation", &[("val", "minMax")]); - w.close("c:scaling"); + scaling(w, ax); w.empty("c:delete", &[("val", "0")]); w.empty("c:axPos", &[("val", pos)]); + if ax.gridlines { + w.empty("c:majorGridlines", &[]); + } + axis_title(w, ax); + num_fmt(w, ax); w.empty("c:majorTickMark", &[("val", "none")]); w.empty("c:minorTickMark", &[("val", "none")]); - w.empty("c:tickLblPos", &[("val", "nextTo")]); + w.empty("c:tickLblPos", &[("val", if ax.labels { "nextTo" } else { "none" })]); + axis_line(w, ax); w.empty("c:crossAx", &[("val", VAL_AX)]); w.empty("c:crosses", &[("val", "autoZero")]); w.empty("c:auto", &[("val", "1")]); @@ -305,34 +383,81 @@ fn category_axis(w: &mut W, pos: &str) { w.close("c:catAx"); } -/// A value axis: with gridlines and no axis line (`grid`), else with its line and no gridlines. -fn value_axis(w: &mut W, id: &str, cross: &str, pos: &str, grid: bool, between: &str) { +/// A value axis (scatter x axes included): its scale, gridlines, title, labels and line, as the +/// model sets them. +fn value_axis(w: &mut W, id: &str, cross: &str, pos: &str, between: &str, ax: &ChartAxis) { w.open("c:valAx", &[]); w.empty("c:axId", &[("val", id)]); - w.open("c:scaling", &[]); - w.empty("c:orientation", &[("val", "minMax")]); - w.close("c:scaling"); + scaling(w, ax); w.empty("c:delete", &[("val", "0")]); w.empty("c:axPos", &[("val", pos)]); - if grid { + if ax.gridlines { w.empty("c:majorGridlines", &[]); } + axis_title(w, ax); + num_fmt(w, ax); w.empty("c:majorTickMark", &[("val", "none")]); w.empty("c:minorTickMark", &[("val", "none")]); - w.empty("c:tickLblPos", &[("val", "nextTo")]); - if grid { - w.open("c:spPr", &[]); - w.open("a:ln", &[]); - w.empty("a:noFill", &[]); - w.close("a:ln"); - w.close("c:spPr"); - } + w.empty("c:tickLblPos", &[("val", if ax.labels { "nextTo" } else { "none" })]); + axis_line(w, ax); w.empty("c:crossAx", &[("val", cross)]); w.empty("c:crosses", &[("val", "autoZero")]); w.empty("c:crossBetween", &[("val", between)]); w.close("c:valAx"); } +/// `c:scaling`: the orientation, then the fixed bounds when the model sets them. +fn scaling(w: &mut W, ax: &ChartAxis) { + w.open("c:scaling", &[]); + w.empty("c:orientation", &[("val", "minMax")]); + for (name, v) in [("c:max", ax.max), ("c:min", ax.min)] { + if let Some(v) = v.filter(|v| v.is_finite()) { + w.empty(name, &[("val", &num(v))]); + } + } + w.close("c:scaling"); +} + +/// The axis title, in the same rich-text shape the chart title uses. +fn axis_title(w: &mut W, ax: &ChartAxis) { + let Some(t) = &ax.title else { return }; + w.open("c:title", &[]); + w.open("c:tx", &[]); + w.open("c:rich", &[]); + w.empty("a:bodyPr", &[]); + w.empty("a:lstStyle", &[]); + w.open("a:p", &[]); + w.open("a:r", &[]); + w.leaf("a:t", &[], t); + w.close("a:r"); + w.close("a:p"); + w.close("c:rich"); + w.close("c:tx"); + w.empty("c:overlay", &[("val", "0")]); + w.close("c:title"); +} + +/// Tick-label number format: the model's code, unlinked from the source, else the default. +fn num_fmt(w: &mut W, ax: &ChartAxis) { + match &ax.fmt { + Some(code) => w.empty("c:numFmt", &[("formatCode", code), ("sourceLinked", "0")]), + None => w.empty("c:numFmt", &[("formatCode", "General"), ("sourceLinked", "1")]), + } +} + +/// `c:spPr`: an axis line only when the model keeps it (the default writes nothing, which reads +/// as a line); `a:noFill` when the axis line is off. +fn axis_line(w: &mut W, ax: &ChartAxis) { + if ax.line { + return; + } + w.open("c:spPr", &[]); + w.open("a:ln", &[]); + w.empty("a:noFill", &[]); + w.close("a:ln"); + w.close("c:spPr"); +} + /// The chart a chart part holds, when the part is exactly what [`chart_xml`] writes for it (so /// editing and saving it loses nothing). `None` for any other chart. pub(crate) fn spec_of(space: &El) -> Option { @@ -365,15 +490,40 @@ pub(crate) fn spec_of(space: &El) -> Option { let categories = sers.first().and_then(|s| s.child(if scatter { "c:xVal" } else { "c:cat" })).map(literal).unwrap_or_default(); let series = sers .iter() - .map(|s| ChartSeries { - name: s.child("c:tx").and_then(|t| t.child("c:v")).map(|v| v.text()).unwrap_or_default(), - values: s - .child(if scatter { "c:yVal" } else { "c:val" }) - .map(literal) - .unwrap_or_default() - .iter() - .map(|v| v.parse::().ok().filter(|v| v.is_finite())) - .collect(), + .map(|s| { + let ln = s.child("c:spPr").and_then(|p| p.child("a:ln")); + let symbol = s.child("c:marker").and_then(|m| m.child("c:symbol")).and_then(|m| m.attr("val")); + ChartSeries { + name: s.child("c:tx").and_then(|t| t.child("c:v")).map(|v| v.text()).unwrap_or_default(), + values: s + .child(if scatter { "c:yVal" } else { "c:val" }) + .map(literal) + .unwrap_or_default() + .iter() + .map(|v| v.parse::().ok().filter(|v| v.is_finite())) + .collect(), + fill: s + .child("c:spPr") + .and_then(|p| p.child("a:solidFill")) + .and_then(|f| f.child("a:srgbClr")) + .and_then(|c| c.attr("val")) + .and_then(Rgb::parse), + no_fill: s.child("c:spPr").and_then(|p| p.child("a:noFill")).is_some(), + line_w: ln.and_then(|l| l.attr("w")).and_then(|w| w.parse::().ok()).filter(|w| w.is_finite()).map(|w| (w / 12700.0) as f32), + no_line: ln.and_then(|l| l.child("a:noFill")).is_some(), + // Line series show no markers by default: only a setting that differs from that + // belongs in the model. A scatter series keeps what the file says, so a missing + // `c:marker` (Word's automatic symbols) stays unset and is not written back. + markers: symbol.map(|v| v != "none").filter(|m| scatter || *m), + smooth: s.child("c:smooth").and_then(|v| v.attr("val")).is_some_and(|v| v == "1"), + labels: s.child("c:dLbls").and_then(|d| d.child("c:showVal")).and_then(|v| v.attr("val")).map(|v| v == "1"), + fmt: s + .child(if scatter { "c:yVal" } else { "c:val" }) + .and_then(|v| v.els().find(|e| e.name == "c:numLit")) + .and_then(|l| l.child("c:formatCode")) + .map(|f| f.text()) + .filter(|f| *f != "General"), + } }) .collect(); let title = chart.child("c:title").and_then(|t| t.find("a:t")).map(|t| t.text()); @@ -384,12 +534,48 @@ pub(crate) fn spec_of(space: &El) -> Option { _ => LegendPos::Right, }); let data_labels = ty.child("c:dLbls").and_then(|d| d.child("c:showVal")).and_then(|v| v.attr("val")) == Some("1"); - let spec = ChartSpec { kind, categories, series, title, legend, data_labels }.sanitized(); - // Editable only when nothing would be lost: writing the model again gives the same part. - let again = xml::parse(chart_xml(&spec).as_bytes()).ok()?; + let (cat_a, val_a) = crate::read::chart::chart_axes(plot, scatter); + let axis = |a: crate::read::chart::Axis| ChartAxis { + title: a.title, + min: a.min, + max: a.max, + fmt: a.fmt, + gridlines: a.grid, + labels: a.labels, + line: a.line, + }; + let spec = ChartSpec { kind, categories, series, title, legend, data_labels, cat_axis: axis(cat_a), val_axis: axis(val_a) }.sanitized(); + // Editable only when nothing would be lost: writing the model again gives the same part, as + // this writer or the earlier one wrote it. + let mut again = xml::parse(chart_xml(&spec).as_bytes()).ok()?; + if same(space, &again, 0) { + return Some(spec); + } + as_previous_writer(&mut again); same(space, &again, 0).then_some(spec) } +/// `part` (what [`chart_xml`] writes) as the earlier WordCraft writer wrote the same model: it +/// had no axis `c:numFmt`. Only that element goes, and only with the value its absence reads as +/// (General, linked to the source), so a part that matches still holds nothing the model leaves out. +fn as_previous_writer(part: &mut El) { + fn els_mut(e: &mut El) -> impl Iterator { + e.kids.iter_mut().filter_map(|n| match n { + Node::El(e) => Some(e), + Node::Text(_) => None, + }) + } + fn drop_els(e: &mut El, drop: impl Fn(&El) -> bool) { + e.kids.retain(|n| !matches!(n, Node::El(c) if drop(c))); + } + let Some(plot) = els_mut(part).find(|e| e.name == "c:chart").and_then(|c| els_mut(c).find(|e| e.name == "c:plotArea")) else { return }; + for e in els_mut(plot) { + if matches!(e.name.as_str(), "c:catAx" | "c:valAx") { + drop_els(e, |c| c.name == "c:numFmt" && c.attr("sourceLinked") == Some("1")); + } + } +} + /// The points of a `c:strLit` / `c:numLit` (or a cache) as texts by index, `ptCount` long. fn literal(e: &El) -> Vec { let Some(lit) = e.els().find(|c| matches!(c.name.as_str(), "c:strLit" | "c:numLit")) else { return Vec::new() }; @@ -461,10 +647,11 @@ mod tests { let spec = ChartSpec { kind: ChartType::Column, categories: (0..4500).map(|i| i.to_string()).collect(), - series: (0..300).map(|i| ChartSeries { name: format!("S{i}"), values: vec![Some(1.0); 4500] }).collect(), + series: (0..300).map(|i| ChartSeries { name: format!("S{i}"), values: vec![Some(1.0); 4500], ..Default::default() }).collect(), title: None, legend: None, data_labels: false, + ..Default::default() }; let x = chart_xml(&spec); let series = x.matches("").count(); @@ -491,6 +678,33 @@ mod tests { assert_eq!(spec_of(&parse(&plain)), Some(plain)); } + #[test] + fn no_fill_wins_over_a_stored_colour() { + let mut spec = ChartSpec::sample(ChartType::Column); + spec.series[0].fill = Some(Rgb(255, 0, 0)); + spec.series[0].no_fill = true; + let x = chart_xml(&spec); + assert!(!x.contains("FF0000"), "colour written next to no_fill"); + let back = spec_of(&parse(&spec)).expect("stays editable"); + assert!(back.series[0].no_fill); + assert_eq!(back.series[0].fill, None); + } + + #[test] + fn an_axis_format_is_limited_in_characters_not_bytes() { + use craft_wordmodel::chart::MAX_FMT; + let cjk = format!("0\"{}\"", "\u{5143}".repeat(90)); + let ascii = format!("0\"{}\"", "x".repeat(90)); + for fmt in [cjk, ascii] { + assert!(fmt.chars().count() <= MAX_FMT); + let mut spec = ChartSpec::sample(ChartType::Column); + spec.val_axis.fmt = Some(fmt.clone()); + let spec = spec.sanitized(); + assert_eq!(spec.val_axis.fmt.as_deref(), Some(fmt.as_str())); + assert_eq!(spec_of(&parse(&spec)).as_ref(), Some(&spec), "{} chars", fmt.chars().count()); + } + } + #[test] fn a_chart_with_more_than_the_model_stays_read_only() { let spec = ChartSpec::sample(ChartType::Column); @@ -500,6 +714,61 @@ mod tests { assert_eq!(spec_of(&xml::parse(x.as_bytes()).unwrap()), None); } + /// Chart parts as the previous WordCraft chart writer saved them (its output for + /// `ChartSpec::sample`): no axis `c:numFmt`, and scatter series with no `c:marker`. + const PREVIOUS_COLUMN: &str = r#"Series 1Q1Q2Q3Q4General3.24.13.65Series 2Q1Q2Q3Q4General22.73.93.3Series 3Q1Q2Q3Q4General1.52.22.43.8"#; + const PREVIOUS_SCATTER: &str = r#"Series 1General1234General2.13.434.6Series 2General1234General1.21.92.83.1"#; + + #[test] + fn charts_saved_by_the_previous_writer_stay_editable() { + for (x, kind) in [(PREVIOUS_COLUMN, ChartType::Column), (PREVIOUS_SCATTER, ChartType::Scatter)] { + let spec = spec_of(&xml::parse(x.as_bytes()).unwrap()).unwrap_or_else(|| panic!("{kind:?}: the earlier part is read-only")); + assert_eq!((spec.kind, spec.series.len()), (kind, if kind == ChartType::Scatter { 2 } else { 3 })); + assert_eq!((spec.cat_axis.fmt.as_deref(), spec.val_axis.fmt.as_deref()), (None, None), "{kind:?}"); + assert!(spec.series.iter().all(|s| s.markers.is_none()), "{kind:?}: markers unset"); + // Saved again, it is in the current shape and still reads back as the same model. + assert_eq!(spec_of(&parse(&spec)).as_ref(), Some(&spec), "{kind:?}"); + } + // What neither writer writes still keeps such a part read-only: an axis format linked to + // the source that isn't General, a scatter marker other than the default, a marker size. + let edited = |x: &str, from: &str, to: &str| { + assert!(x.contains(from), "{from}"); + spec_of(&xml::parse(x.replacen(from, to, 1).as_bytes()).unwrap()) + }; + let ax = r#""#; + assert_eq!(edited(PREVIOUS_COLUMN, ax, &format!(r#"{ax}"#)), None); + let ln = r#""#; + assert_eq!(edited(PREVIOUS_SCATTER, ln, &format!(r#"{ln}"#)), None); + assert_eq!(edited(PREVIOUS_SCATTER, ln, &format!(r#"{ln}"#)), None); + assert_eq!(edited(PREVIOUS_COLUMN, "", ""), None); + } + + /// A scatter series with no `c:marker` (Word's automatic symbols) stays without one after an + /// unrelated edit and a save; an explicit circle or "none" stays explicit. + #[test] + fn scatter_series_without_a_marker_stay_without_after_a_save() { + let marker_of = |x: &El| -> Vec> { + let sc = x.child("c:chart").unwrap().child("c:plotArea").unwrap().child("c:scatterChart").unwrap(); + sc.children("c:ser").map(|s| s.child("c:marker").map(|m| m.child("c:symbol").unwrap().attr("val").unwrap().to_string())).collect() + }; + let mut spec = spec_of(&xml::parse(PREVIOUS_SCATTER.as_bytes()).unwrap()).unwrap(); + assert_eq!(marker_of(&xml::parse(PREVIOUS_SCATTER.as_bytes()).unwrap()), [None, None]); + spec.title = Some("Edited".into()); + spec.series[0].name = "Renamed".into(); + let saved = parse(&spec); + assert_eq!(marker_of(&saved), [None, None]); + let again = spec_of(&saved).unwrap(); + assert_eq!(again, spec); + assert!(again.series.iter().all(|s| s.markers.is_none())); + // Explicit choices are kept as written. + spec.series[0].markers = Some(true); + spec.series[1].markers = Some(false); + let saved = parse(&spec); + assert_eq!(marker_of(&saved), [Some("circle".to_string()), Some("none".to_string())]); + let again = spec_of(&saved).unwrap(); + assert_eq!((again.series[0].markers, again.series[1].markers), (Some(true), Some(false))); + } + #[test] fn elements_come_in_schema_order() { let mut spec = ChartSpec::sample(ChartType::ColumnStacked); @@ -517,6 +786,11 @@ mod tests { ["c:barDir", "c:grouping", "c:varyColors", "c:ser", "c:ser", "c:ser", "c:dLbls", "c:gapWidth", "c:overlap", "c:axId", "c:axId"] ); assert_eq!(names(bar.child("c:ser").unwrap()), ["c:idx", "c:order", "c:tx", "c:invertIfNegative", "c:cat", "c:val"]); + // A series with its own labels keeps `c:invertIfNegative` ahead of `c:dLbls` (CT_BarSer). + spec.series[0].labels = Some(true); + let labelled = parse(&spec); + let ser = labelled.child("c:chart").unwrap().child("c:plotArea").unwrap().child("c:barChart").unwrap().child("c:ser").unwrap(); + assert_eq!(names(ser), ["c:idx", "c:order", "c:tx", "c:invertIfNegative", "c:dLbls", "c:cat", "c:val"]); let val_ax = plot.child("c:valAx").unwrap(); assert_eq!( names(val_ax), @@ -526,6 +800,7 @@ mod tests { "c:delete", "c:axPos", "c:majorGridlines", + "c:numFmt", "c:majorTickMark", "c:minorTickMark", "c:tickLblPos", @@ -535,13 +810,35 @@ mod tests { "c:crossBetween" ] ); + let cat_ax = plot.child("c:catAx").unwrap(); + assert_eq!( + names(cat_ax), + [ + "c:axId", + "c:scaling", + "c:delete", + "c:axPos", + "c:numFmt", + "c:majorTickMark", + "c:minorTickMark", + "c:tickLblPos", + "c:crossAx", + "c:crosses", + "c:auto", + "c:lblAlgn", + "c:lblOffset", + "c:noMultiLvlLbl" + ] + ); let line = parse(&ChartSpec { data_labels: true, ..ChartSpec::sample(ChartType::Line) }); let lc = line.child("c:chart").unwrap().child("c:plotArea").unwrap().child("c:lineChart").unwrap(); assert_eq!(names(lc), ["c:grouping", "c:varyColors", "c:ser", "c:ser", "c:ser", "c:dLbls", "c:marker", "c:axId", "c:axId"]); assert_eq!(names(lc.child("c:ser").unwrap()), ["c:idx", "c:order", "c:tx", "c:marker", "c:cat", "c:val", "c:smooth"]); - let sc = parse(&ChartSpec::sample(ChartType::Scatter)); + let mut scatter = ChartSpec::sample(ChartType::Scatter); + scatter.series.iter_mut().for_each(|s| s.markers = Some(true)); + let sc = parse(&scatter); let sc = sc.child("c:chart").unwrap().child("c:plotArea").unwrap().child("c:scatterChart").unwrap(); - assert_eq!(names(sc.child("c:ser").unwrap()), ["c:idx", "c:order", "c:tx", "c:spPr", "c:xVal", "c:yVal", "c:smooth"]); + assert_eq!(names(sc.child("c:ser").unwrap()), ["c:idx", "c:order", "c:tx", "c:spPr", "c:marker", "c:xVal", "c:yVal", "c:smooth"]); let dn = parse(&ChartSpec::sample(ChartType::Doughnut)); let dn = dn.child("c:chart").unwrap().child("c:plotArea").unwrap(); assert_eq!(names(dn), ["c:layout", "c:doughnutChart"]); diff --git a/crates/docx/src/read/chart.rs b/crates/docx/src/read/chart.rs index a0087a3bb..5b02603d6 100644 --- a/crates/docx/src/read/chart.rs +++ b/crates/docx/src/read/chart.rs @@ -29,7 +29,9 @@ pub(super) use canvas::text_w; use canvas::{Rect, clip_text, legend, text}; #[cfg(test)] use format::{civil, excel_stamp, format_value}; -use model::{Axis, Grouping, Kind, Pos, Series}; +pub(crate) use model::Axis; +use model::{Grouping, Kind, Pos, Series}; +pub(crate) use parse::chart_axes; #[cfg(test)] use parse::{axis, category_text, labels, numbers}; use parse::{read_series, text_size, title_of}; diff --git a/crates/docx/src/read/chart/canvas.rs b/crates/docx/src/read/chart/canvas.rs index a8280e4ba..ab3fbf410 100644 --- a/crates/docx/src/read/chart/canvas.rs +++ b/crates/docx/src/read/chart/canvas.rs @@ -41,8 +41,13 @@ pub(super) fn widest(items: &[String], size: f32) -> f32 { items.iter().map(|s| text_w(s, size)).fold(0.0, f32::max) } +/// Text in `rect`; `angle` degrees clockwise turns it about the rect's centre (axis titles). pub(super) fn text(out: &mut Vec, rect: Rect, s: &str, size: f32, align: TextAlign, bold: bool) { - out.push(GraphicItem::Text { rect: rect.arr(), text: s.to_string(), size, color: TEXT, bold, align, font: None }); + text_angle(out, rect, s, size, align, bold, 0.0); +} + +pub(super) fn text_angle(out: &mut Vec, rect: Rect, s: &str, size: f32, align: TextAlign, bold: bool, angle: f32) { + out.push(GraphicItem::Text { rect: rect.arr(), text: s.to_string(), size, color: TEXT, bold, align, font: None, angle }); } /// Text centred on `cx` with its top at `y`. diff --git a/crates/docx/src/read/chart/model.rs b/crates/docx/src/read/chart/model.rs index 71ee7040f..13da1d654 100644 --- a/crates/docx/src/read/chart/model.rs +++ b/crates/docx/src/read/chart/model.rs @@ -7,6 +7,33 @@ pub(super) struct DataLabel { pub(super) y: f32, pub(super) text: String, pub(super) inside: bool, + /// Where a label that is not `inside` sits relative to (`x`, `y`). + pub(super) pos: LabelPos, + /// Distance from (`x`, `y`) to the label for `pos` other than `Center`, points (half the marker). + pub(super) off: f32, +} + +/// Where a data label sits beside its point (`c:dLblPos`: `t`, `b`, `l`, `r`, `ctr`). +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub(super) enum LabelPos { + Above, + Below, + Left, + Right, + Center, +} + +impl LabelPos { + pub(super) fn parse(v: &str) -> Option { + Some(match v.trim() { + "t" => LabelPos::Above, + "b" => LabelPos::Below, + "l" => LabelPos::Left, + "r" => LabelPos::Right, + "ctr" => LabelPos::Center, + _ => return None, + }) + } } #[derive(Clone, Copy, Debug, PartialEq, Eq)] @@ -52,16 +79,56 @@ pub(super) enum Pos { /// One value axis (or category axis) as the file states it. #[derive(Default)] -pub(super) struct Axis { - pub(super) deleted: bool, - pub(super) grid: bool, - pub(super) line: bool, +pub(crate) struct Axis { + pub(crate) deleted: bool, + pub(crate) grid: bool, + pub(crate) line: bool, /// `c:crossBetween val="midCat"`: categories sit on the tick marks, not between them. pub(super) mid: bool, - pub(super) min: Option, - pub(super) max: Option, + pub(crate) min: Option, + pub(crate) max: Option, /// Number format of the tick labels, when not linked to the source data. - pub(super) fmt: Option, + pub(crate) fmt: Option, + /// The axis title, when the file gives it one. + pub(crate) title: Option, + /// The title's `a:bodyPr/@rot` in degrees clockwise (`-5400000` is 90° counter-clockwise), + /// when the file gives one. + pub(crate) title_rot: Option, + /// Tick labels, as `c:tickLblPos` says (`none` hides them; a file that says nothing shows + /// them, as Word does). + pub(crate) labels: bool, +} + +/// The shape of a series' point markers (`ST_MarkerStyle`, ECMA-376 §21.2.3.28). +#[derive(Clone, Copy, Default, PartialEq, Eq, Debug)] +pub(super) enum Symbol { + #[default] + Circle, + Square, + Diamond, + Triangle, + X, + Plus, + Star, + Dash, + Dot, +} + +impl Symbol { + /// The symbol a `c:symbol` value names; anything unknown (`auto`, `picture`) is a circle. + pub(super) fn parse(name: &str) -> Symbol { + match name { + "square" => Symbol::Square, + "diamond" => Symbol::Diamond, + "triangle" => Symbol::Triangle, + "x" => Symbol::X, + "plus" => Symbol::Plus, + "star" => Symbol::Star, + "dash" => Symbol::Dash, + "dot" => Symbol::Dot, + _ => Symbol::Circle, + } + } } pub(super) struct Series { @@ -70,9 +137,21 @@ pub(super) struct Series { /// when missing or not a valid number. pub(super) idx: Option, pub(super) color: Option, - pub(super) line_w: f32, + /// Fixed line width in points, `None`: the chart's default. + pub(super) line_w: Option, /// Whether the series is drawn as a line (scatter series without one get markers). pub(super) line: bool, + /// The series' own fill: `Some` colour, `None` colour with `no_fill` set, else the chart's + /// accent for the series. + pub(super) no_fill: bool, + /// Markers on the series' points, when the file says. + pub(super) markers: Option, + /// The marker shape (`c:marker/c:symbol`; a circle when missing or unknown) and its size in + /// points (`c:marker/c:size`), when the file gives one. + pub(super) symbol: Symbol, + pub(super) marker_size: Option, + /// Draw the series' line as a smooth curve (`c:smooth`). + pub(super) smooth: bool, pub(super) values: Vec>, /// Scatter x values (index + 1 when the file has none). pub(super) xs: Vec>, @@ -83,6 +162,13 @@ pub(super) struct Series { pub(super) x_fmt: Option, /// Data labels on or off for this series (`c:dLbls` of its own), when it says. pub(super) labels: Option, + /// The labels' number format (`c:dLbls/c:numFmt`): `None` when not given (the chart type's + /// applies), `Some(None)` when linked to the source, else `Some(Some(code))`. + pub(super) label_fmt: Option>, + /// The data labels' `c:dLblPos`, when the file gives a known one. + pub(super) label_pos: Option, + /// The marker's own fill (`c:marker/c:spPr`); without it markers take the automatic colour. + pub(super) marker_color: Option, /// The values as the file has them, for data labels (`values` become shares in a 100% /// stacked chart); empty when the series shows no labels. pub(super) raw: Vec>, diff --git a/crates/docx/src/read/chart/parse.rs b/crates/docx/src/read/chart/parse.rs index 3c09fcfe9..522d50b3b 100644 --- a/crates/docx/src/read/chart/parse.rs +++ b/crates/docx/src/read/chart/parse.rs @@ -1,7 +1,7 @@ use super::format::{date_text, excel_stamp, format_value, is_date_code}; -use super::model::{Axis, Grouping, Kind, Pos, Series}; +use super::model::{Axis, Grouping, Kind, LabelPos, Pos, Series, Symbol}; use super::plot::label; -use super::{Chart, El, FONT, LINE_W, MAX_CELLS, MAX_POINTS, MAX_RADAR_SERIES, MAX_SERIES, MAX_SPOKES, Rgb, fill_colour, int, measure, slot}; +use super::{Chart, El, FONT, MAX_CELLS, MAX_POINTS, MAX_RADAR_SERIES, MAX_SERIES, MAX_SPOKES, Rgb, fill_colour, int, measure, slot}; impl Chart { pub(super) fn parse(space: &El, theme: &[Rgb]) -> Option { @@ -39,10 +39,15 @@ impl Chart { cats.resize(n, String::new()); let radar_style = ty.child("c:radarStyle").and_then(|r| r.attr("val")); let type_labels = show_values(ty.child("c:dLbls")); + let type_fmt = label_format(ty.child("c:dLbls")); for s in &mut series { s.values.resize(n, None); s.xs.resize(n, None); s.labels = Some(s.labels.or(type_labels).unwrap_or(false)); + // Absent (`None`) inherits the chart type's; explicitly source-linked (`Some(None)`) does not. + if s.label_fmt.is_none() { + s.label_fmt.clone_from(&type_fmt); + } if s.labels == Some(true) { s.raw = s.values.clone(); } @@ -98,6 +103,13 @@ impl Chart { s.and_then(|s| s.color).unwrap_or_else(|| self.accent(s.and_then(|s| s.idx).unwrap_or(j))) } + /// Colour of series `j`'s markers: the marker's own fill, else the automatic colour of the + /// series index (not the series' explicit line colour, as Word draws them). + pub(super) fn marker_colour(&self, j: usize) -> Rgb { + let s = self.series.get(j); + s.and_then(|s| s.marker_color).unwrap_or_else(|| self.accent(s.and_then(|s| s.idx).unwrap_or(j))) + } + /// Colour of point `i` of series `j`: its `c:dPt` override, else the varied accent (pie) or /// the series colour. pub(super) fn point_colour(&self, j: usize, i: usize) -> Rgb { @@ -149,7 +161,7 @@ fn to_percent(series: &mut [Series], n: usize) { } /// The category axis (scatter: the horizontal value axis) and the value axis of a plot area. -fn chart_axes(plot: &El, scatter: bool) -> (Axis, Axis) { +pub(crate) fn chart_axes(plot: &El, scatter: bool) -> (Axis, Axis) { let vals_ax: Vec<&El> = plot.children("c:valAx").collect(); let horiz_pos = |a: &&El| matches!(a.child("c:axPos").and_then(|p| p.attr("val")), Some("b" | "t")); let (val_el, x_el) = if scatter { @@ -180,6 +192,17 @@ fn show_values(d: Option<&El>) -> Option { d.child("c:showVal").map(|v| v.attr("val").is_none_or(truthy)) } +/// The number format of data labels (`c:dLbls/c:numFmt`): `None` when the labels say nothing, +/// `Some(None)` when they are linked to the source (so a chart-level format does not apply), +/// `Some(Some(code))` for their own format code. +fn label_format(d: Option<&El>) -> Option> { + let f = d?.child("c:numFmt")?; + if f.attr("sourceLinked").is_some_and(truthy) { + return Some(None); + } + f.attr("formatCode").filter(|f| f.chars().count() <= craft_wordmodel::chart::MAX_FMT).map(|f| Some(f.to_string())) +} + fn legend_pos(chart: &El) -> Option { chart.child("c:legend").map(|l| match l.child("c:legendPos").and_then(|p| p.attr("val")).unwrap_or("r") { "l" => Pos::Left, @@ -206,9 +229,24 @@ impl Series { let sp = ser.child("c:spPr"); let ln = sp.and_then(|p| p.child("a:ln")); let marker = ser.child("c:marker").and_then(|m| m.child("c:spPr")).and_then(|p| fill_colour(p, theme)); + let marker_color = marker; let color = sp.and_then(|p| fill_colour(p, theme)).or_else(|| ln.and_then(|l| fill_colour(l, theme))).or(marker); - let line_w = ln.and_then(|l| l.attr("w")).and_then(|v| measure(v, 12_700.0)).map_or(LINE_W, |w| w.clamp(0.0, 20.0)); + let no_fill = sp.and_then(|p| p.child("a:noFill")).is_some(); + let line_w = ln.and_then(|l| l.attr("w")).and_then(|v| measure(v, 12_700.0)).map(|w| w.clamp(0.0, 20.0)); let line = ln.is_none_or(|l| l.child("a:noFill").is_none()); + // `c:marker/c:symbol`: which symbol each point gets (`none` for no markers). Scatter + // series show markers unless the file says not; line series only when it says so. + let symbol = ser.child("c:marker").and_then(|m| m.child("c:symbol")).and_then(|s| s.attr("val")).map(str::to_string); + let markers = symbol.as_deref().map(|s| s != "none").or_else(|| scatter.then_some(true)); + let symbol = symbol.as_deref().map_or(Symbol::Circle, Symbol::parse); + let marker_size = ser + .child("c:marker") + .and_then(|m| m.child("c:size")) + .and_then(|s| s.attr("val")) + .and_then(|v| v.trim().parse::().ok()) + .filter(|v| v.is_finite()) + .map(|v| v.clamp(2.0, 72.0)); + let smooth = ser.child("c:smooth").and_then(|s| s.attr("val")).is_some_and(truthy); let (values, fmt) = numbers(ser.child(if scatter { "c:yVal" } else { "c:val" }), cap); let (mut xs, x_fmt) = if scatter { numbers(ser.child("c:xVal"), cap) } else { (Vec::new(), None) }; if scatter && xs.is_empty() { @@ -223,7 +261,30 @@ impl Series { } } let labels = show_values(ser.child("c:dLbls")); - Series { name, idx, color, line_w, line, values, xs, points, fmt, x_fmt, labels, raw: Vec::new() } + let label_fmt = label_format(ser.child("c:dLbls")); + let label_pos = ser.child("c:dLbls").and_then(|d| d.child("c:dLblPos")).and_then(|p| p.attr("val")).and_then(LabelPos::parse); + Series { + name, + idx, + color, + line_w, + line, + no_fill, + markers, + symbol, + marker_size, + smooth, + values, + xs, + points, + fmt, + x_fmt, + labels, + label_fmt, + label_pos, + marker_color, + raw: Vec::new(), + } } } @@ -290,7 +351,7 @@ fn pts(cache: &El, cap: usize) -> Vec> { } pub(super) fn axis(a: Option<&El>) -> Axis { - let Some(a) = a else { return Axis { deleted: true, ..Default::default() } }; + let Some(a) = a else { return Axis { deleted: true, labels: true, line: true, ..Default::default() } }; let scaling = a.child("c:scaling"); let num = |n: &str| scaling.and_then(|s| s.child(n)).and_then(|e| e.attr("val")).and_then(|v| v.trim().parse::().ok()).filter(|v| v.is_finite()); @@ -305,8 +366,16 @@ pub(super) fn axis(a: Option<&El>) -> Axis { .child("c:numFmt") .filter(|f| f.attr("sourceLinked") != Some("1")) .and_then(|f| f.attr("formatCode")) - .filter(|f| f.len() <= MAX_FORMAT_CODE) + .filter(|f| f.chars().count() <= craft_wordmodel::chart::MAX_FMT) .map(str::to_string), + title: a.child("c:title").and_then(title_text), + title_rot: a + .child("c:title") + .and_then(|t| t.find("a:bodyPr")) + .and_then(|b| b.attr("rot")) + .and_then(|v| v.trim().parse::().ok()) + .map(|v| (v as f64 / 60_000.0).clamp(-360.0, 360.0) as f32), + labels: a.child("c:tickLblPos").and_then(|p| p.attr("val")) != Some("none"), } } diff --git a/crates/docx/src/read/chart/plot.rs b/crates/docx/src/read/chart/plot.rs index 652ee79a8..22d05b554 100644 --- a/crates/docx/src/read/chart/plot.rs +++ b/crates/docx/src/read/chart/plot.rs @@ -1,9 +1,12 @@ -use super::canvas::{Rect, fill, rect_segs, runs, seg, stroke, text, text_at, text_w, widest, wrap}; +use super::canvas::{Rect, clip_text, fill, rect_segs, runs, seg, stroke, text, text_angle, text_at, text_w, widest, wrap}; use super::format::format_value; -use super::model::{DataLabel, Grouping, Kind, Series}; +use super::model::{DataLabel, Grouping, Kind, LabelPos, Series, Symbol}; use super::parse::category_text; use super::polar::{draw_data_labels, label_of}; -use super::{AXIS, Chart, El, GAP, GRID, GraphicItem, LINE, MAX_LABEL, MAX_TICKS, PathSeg, Rgb, ShapeKind, TextAlign, fill_colour, measure}; +use super::{AXIS, Chart, El, GAP, GRID, GraphicItem, LINE, LINE_W, MAX_LABEL, MAX_TICKS, PathSeg, Rgb, ShapeKind, TextAlign, fill_colour, measure}; + +/// Line width of the stroked marker symbols (x, plus, star), points. +const MARKER_STROKE: f32 = 1.5; pub(super) fn label(s: &str) -> String { s.chars().take(MAX_LABEL).collect() @@ -142,6 +145,8 @@ impl Axes { /// The room around the plot for axis labels, and the category labels as they'll be drawn. struct Margins { left: f32, + /// Part of `left` taken by the rotated category-axis title (bars): the labels start after it. + title: f32, top: f32, bottom: f32, /// Category labels in lines (none for scatter charts or a deleted category axis). @@ -151,22 +156,25 @@ struct Margins { } /// The margins the axis labels need around the plot in `area`. Category labels under a column -/// chart wrap to the band; labels that still don't fit show every k-th category. +/// chart wrap to the band; labels that still don't fit show every k-th category. Hidden tick +/// labels take no room; an axis title adds one line on its side. fn axis_margins(c: &Chart, ax: &Axes, area: Rect) -> Margins { let fs = c.text_size; let lh = fs * LINE; let scatter = c.kind == Kind::Scatter; let (cat_on, val_on) = (!c.cat.deleted, !c.val.deleted); + let (cat_labels, val_labels) = (cat_on && c.cat.labels, val_on && c.val.labels); + let title = title_left(c, cat_on, val_on, area.w); let left = if c.horizontal { - if cat_on { widest(&c.cats, fs).min(area.w * 0.4) + GAP } else { 0.0 } - } else if val_on { + if cat_labels { widest(&c.cats, fs).min(area.w * 0.4) + GAP } else { 0.0 } + } else if val_labels { ax.vticks.iter().map(|(_, s)| text_w(s, fs)).fold(0.0, f32::max) + GAP } else { 0.0 - }; + } + title; let band_w = (area.w - left).max(1.0) / c.n as f32; - let wrapped: Vec> = if cat_on && !scatter { - let width = if c.horizontal { left - GAP } else { band_w }; + let wrapped: Vec> = if cat_labels && !scatter { + let width = if c.horizontal { left - title - GAP } else { band_w }; c.cats.iter().map(|s| wrap(s, width, fs)).collect() } else { Vec::new() @@ -174,15 +182,48 @@ fn axis_margins(c: &Chart, ax: &Axes, area: Rect) -> Margins { let widest_line = wrapped.iter().flatten().map(|s| text_w(s, fs)).fold(0.0, f32::max); let skip = ((widest_line / band_w).ceil() as usize).max(1); let label_lines = wrapped.iter().map(Vec::len).max().unwrap_or(1) as f32; - let bottom = if cat_on && !c.horizontal { + let bottom = (if cat_labels && !c.horizontal { if scatter { lh + GAP } else { label_lines * lh + GAP } - } else if c.horizontal && val_on { + } else if c.horizontal && val_labels { lh + GAP } else { 0.0 - }; - let top = if !c.horizontal && val_on { lh / 2.0 } else { 0.0 }; - Margins { left, top, bottom, wrapped, skip } + }) + title_bottom(c, cat_on, val_on, lh); + let top = if !c.horizontal && val_labels { lh / 2.0 } else { 0.0 }; + Margins { left, title, top, bottom, wrapped, skip } +} + +/// Room for the vertical axis title on the plot's left: one line when the axis that sits there +/// (the value axis for columns, the category axis for bars) has a title. +/// A horizontal title (`rot="0"`) gets a strip as wide as its wrapped lines, at most a quarter of +/// `area_w`, so the plot keeps its room. +fn title_left(c: &Chart, cat_on: bool, val_on: bool, area_w: f32) -> f32 { + let side = if c.horizontal { cat_on && c.cat.title.is_some() } else { val_on && c.val.title.is_some() }; + if !side { + return 0.0; + } + match flat_title_lines(c, area_w) { + Some(lines) => widest(&lines, c.text_size).min((area_w * 0.25).max(0.0)) + GAP, + None => c.text_size * LINE, + } +} + +/// The wrapped lines of the left axis's title when the file keeps it horizontal (`rot="0"`). +fn flat_title_lines(c: &Chart, area_w: f32) -> Option> { + let left = if c.horizontal { &c.cat } else { &c.val }; + let r = left.title_rot?.rem_euclid(360.0); + if !(r < 1.0 || r > 359.0) { + return None; + } + let width = if area_w.is_finite() { (area_w * 0.18).max(1.0) } else { 1.0 }; + left.title.as_ref().map(|t| wrap(t, width, c.text_size)) +} + +/// Room for the horizontal axis title under the plot: one line when the axis that sits there +/// (the category axis for columns, the value axis for bars) has a title. +fn title_bottom(c: &Chart, cat_on: bool, val_on: bool, lh: f32) -> f32 { + let side = if c.horizontal { val_on && c.val.title.is_some() } else { cat_on && c.cat.title.is_some() }; + if side { lh } else { 0.0 } } /// A bar, column, line, area or scatter chart in `area`: gridlines, series, axes, then labels. @@ -200,6 +241,9 @@ pub(super) fn cartesian(out: &mut Vec, c: &Chart, area: Rect) { if c.val.grid { draw_grid(out, &g, &ax.vticks); } + if c.cat.grid { + draw_cat_grid(out, &g, &ax.xticks, c.kind); + } let mut labels = Vec::new(); match c.kind { Kind::Bar => bars(out, c, &g, &mut labels), @@ -210,12 +254,49 @@ pub(super) fn cartesian(out: &mut Vec, c: &Chart, area: Rect) { } draw_axes(out, c, &g); draw_data_labels(out, &labels, c.text_size, g.horiz && c.kind == Kind::Bar); - if !c.val.deleted { - draw_value_labels(out, c, &g, &ax.vticks, Rect { w: m.left, ..area }); + if !c.val.deleted && c.val.labels { + let room = title_left(c, !c.cat.deleted, true, area.w); + draw_value_labels(out, c, &g, &ax.vticks, Rect { x: area.x + room, w: m.left - room, ..area }); } - if !c.cat.deleted { + if !c.cat.deleted && c.cat.labels { draw_category_labels(out, c, &g, &ax.xticks, &m, area); } + draw_axis_titles(out, c, &g, area); +} + +/// The axis titles, one line: the title of the vertical axis (the value axis for columns, the +/// category axis for bars) is turned 90° on the plot's left; the horizontal axis's title sits +/// under the plot. +fn draw_axis_titles(out: &mut Vec, c: &Chart, g: &Frame, area: Rect) { + let lh = c.text_size * LINE; + // The axis drawn along the plot's left (the value axis for columns, the category axis for + // bars) gets its title rotated there; the other axis's title sits under the plot. + let (left, bottom) = if c.horizontal { (&c.cat, &c.val) } else { (&c.val, &c.cat) }; + // Word turns the vertical axis's title counter-clockwise (reading upwards) unless the file + // turns it the other way or keeps it horizontal (`rot="0"`, wrapped in its own strip); any + // other rotation would not fit the strip and is read as counter-clockwise. + let up = if left.title_rot.is_some_and(|r| (r.rem_euclid(360.0) - 90.0).abs() < 1.0) { 90.0 } else { -90.0 }; + if !left.deleted + && let Some(lines) = flat_title_lines(c, area.w) + { + let strip = (g.plot.x - area.x - GAP).max(1.0); + let top = g.plot.y + g.plot.h / 2.0 - lines.len() as f32 * lh / 2.0; + for (k, l) in lines.iter().enumerate() { + let l = clip_text(l, strip, c.text_size); + text_angle(out, Rect { x: area.x, y: top + k as f32 * lh, w: strip, h: lh }, &l, c.text_size, TextAlign::Center, false, 0.0); + } + } else if !left.deleted + && let Some(t) = &left.title + { + let t = clip_text(t, g.plot.h, c.text_size); + text_angle(out, Rect { x: area.x, y: g.plot.y, w: lh, h: g.plot.h }, &t, c.text_size, TextAlign::Center, false, up); + } + if !bottom.deleted + && let Some(t) = &bottom.title + { + let t = clip_text(t, g.plot.w, c.text_size); + text_angle(out, Rect { x: g.plot.x, y: area.bottom() - lh, w: g.plot.w, h: lh }, &t, c.text_size, TextAlign::Center, false, 0.0); + } } /// Major gridlines at the value ticks. @@ -228,6 +309,31 @@ fn draw_grid(out: &mut Vec, g: &Frame, ticks: &[(f64, String)]) { } } +/// Gridlines along the category axis: across the plot at each category's centre (where line and +/// area points sit, on the tick marks when the axis crosses "midCat"), at the x ticks of a +/// scatter chart. +fn draw_cat_grid(out: &mut Vec, g: &Frame, xticks: &[(f64, String)], kind: Kind) { + let plot = g.plot; + if kind == Kind::Scatter { + for (t, _) in xticks { + let p = g.scat_x(*t); + stroke(out, seg(p, plot.y, p, plot.bottom()), GRID, 0.75); + } + return; + } + for i in 0..g.n { + let segs = if g.horiz { + let p = g.band_start(i) + g.band() / 2.0; + seg(plot.x, p, plot.right(), p) + } else { + // Bars stay in their bands; only line and area points move onto the ticks. + let p = if matches!(kind, Kind::Line | Kind::Area) { g.cat_x(i) } else { g.band_start(i) + g.band() / 2.0 }; + seg(p, plot.y, p, plot.bottom()) + }; + stroke(out, segs, GRID, 0.75); + } +} + /// The category axis line, the plot area's outline and the value axis line, as the file shows them. fn draw_axes(out: &mut Vec, c: &Chart, g: &Frame) { let plot = g.plot; @@ -271,7 +377,7 @@ fn draw_category_labels(out: &mut Vec, c: &Chart, g: &Frame, xticks for (i, lines) in m.wrapped.iter().enumerate() { let top = g.band_start(i) + g.band() / 2.0 - lines.len() as f32 * lh / 2.0; for (k, s) in lines.iter().enumerate() { - text(out, Rect { x: area.x, y: top + k as f32 * lh, w: m.left - GAP, h: lh }, s, fs, TextAlign::Right, false); + text(out, Rect { x: area.x + m.title, y: top + k as f32 * lh, w: m.left - m.title - GAP, h: lh }, s, fs, TextAlign::Right, false); } } } else { @@ -361,7 +467,13 @@ fn bars(out: &mut Vec, c: &Chart, g: &Frame, labels: &mut Vec, c: &Chart, g: &Frame, labels: &mut Vec (if v >= 0.0 { p1 } else { p0 }, mid), (false, false) => (mid, if v >= 0.0 { p0 } else { p1 }), }; - labels.push(DataLabel { x, y, text, inside: stacked }); + labels.push(DataLabel { x, y, text, inside: stacked, pos: LabelPos::Above, off: 0.0 }); } } } @@ -401,13 +513,109 @@ fn lines(out: &mut Vec, c: &Chart, g: &Frame, labels: &mut Vec f32 { + if drawn { s.marker_size.filter(|v| v.is_finite()).unwrap_or(6.0).clamp(2.0, 72.0) / 2.0 } else { 0.0 } +} + +fn markers(out: &mut Vec, pts: impl Iterator, colour: Rgb, s: &Series) { + const MARKER: f32 = 6.0; + let size = s.marker_size.filter(|v| v.is_finite()).unwrap_or(MARKER).clamp(2.0, 72.0); + let h = size / 2.0; + let shape = |out: &mut Vec, rect: [f32; 4], kind: ShapeKind| { + out.push(GraphicItem::Shape { rect, kind, fill: Some(colour), stroke: None, stroke_width: 0.0 }); + }; + let mut strokes = Vec::new(); + for (x, y) in pts { + match s.symbol { + Symbol::Circle => shape(out, [x - h, y - h, size, size], ShapeKind::Ellipse), + Symbol::Square => shape(out, [x - h, y - h, size, size], ShapeKind::Rectangle), + Symbol::Dot => shape(out, [x - h / 2.0, y - h / 2.0, h, h], ShapeKind::Ellipse), + Symbol::Dash => shape(out, [x - h, y - size / 6.0, size, size / 3.0], ShapeKind::Rectangle), + Symbol::Diamond => fill( + out, + vec![PathSeg::Move(x, y - h), PathSeg::Line(x + h, y), PathSeg::Line(x, y + h), PathSeg::Line(x - h, y), PathSeg::Close], + colour, + ), + Symbol::Triangle => { + fill(out, vec![PathSeg::Move(x, y - h), PathSeg::Line(x + h, y + h), PathSeg::Line(x - h, y + h), PathSeg::Close], colour) + } + Symbol::X | Symbol::Plus | Symbol::Star => { + if s.symbol != Symbol::Plus { + strokes.extend(seg(x - h, y - h, x + h, y + h)); + strokes.extend(seg(x - h, y + h, x + h, y - h)); + } + if s.symbol != Symbol::X { + strokes.extend(seg(x - h, y, x + h, y)); + strokes.extend(seg(x, y - h, x, y + h)); + } + } + } + } + stroke(out, strokes, colour, MARKER_STROKE); +} + +/// A polyline through `pts`, a smooth curve (Catmull-Rom pieces) when `smooth` and there are at +/// least three points in a row; gaps (`None`) end a run. +fn curve(pts: impl Iterator>, smooth: bool) -> Vec { + if !smooth { + return runs(pts); + } + let pts: Vec> = pts.collect(); + let mut segs = Vec::new(); + let mut i = 0; + while i < pts.len() { + let Some(first) = pts[i] else { + i += 1; + continue; + }; + // Collect one run of points. + let mut run = vec![first]; + while let Some(p) = pts.get(i + 1).copied().flatten() { + run.push(p); + i += 1; + } + i += 1; + if run.len() < 3 { + segs.extend(runs(run.into_iter().map(Some))); + continue; + } + segs.push(PathSeg::Move(run[0].0, run[0].1)); + for k in 0..run.len() - 1 { + // Catmull-Rom: tangent at p1 from p0 to p2 (ends mirror their own tangent). + let p0 = run[k.max(1) - 1]; + let p1 = run[k]; + let p2 = run[k + 1]; + let p3 = run.get(k + 2).copied().unwrap_or(p2); + let c1 = (p1.0 + (p2.0 - p0.0) / 6.0, p1.1 + (p2.1 - p0.1) / 6.0); + let c2 = (p2.0 - (p3.0 - p1.0) / 6.0, p2.1 - (p3.1 - p1.1) / 6.0); + segs.push(PathSeg::Cubic(c1.0, c1.1, c2.0, c2.1, p2.0, p2.1)); + } + } + segs +} + fn areas(out: &mut Vec, c: &Chart, g: &Frame, labels: &mut Vec) { let mut base = vec![0.0f64; g.n]; for (j, s) in c.series.iter().enumerate().filter(|(_, s)| has_values(s)) { @@ -422,10 +630,12 @@ fn areas(out: &mut Vec, c: &Chart, g: &Frame, labels: &mut Vec, c: &Chart, g: &Frame, labels: &mut Vec, c: &Chart, g: &Frame, labels: &mut Vec) { - const MARKER: f32 = 6.0; for (j, s) in c.series.iter().enumerate() { let pts: Vec> = s.xs.iter() @@ -448,16 +657,21 @@ fn draw_scatter(out: &mut Vec, c: &Chart, g: &Frame, labels: &mut V .collect(); for (i, p) in pts.iter().enumerate() { if let (Some((x, y)), Some(text)) = (p, label_of(s, i)) { - labels.push(DataLabel { x: *x, y: *y - MARKER / 2.0, text, inside: false }); + labels.push(DataLabel { + x: *x, + y: *y, + text, + inside: false, + pos: s.label_pos.unwrap_or(LabelPos::Right), + off: marker_half(s, s.markers != Some(false)) + 2.0, + }); } } if s.line { - stroke(out, runs(pts.iter().copied()), c.series_colour(j), s.line_w); - continue; + stroke(out, curve(pts.iter().copied(), s.smooth), c.series_colour(j), s.line_w.unwrap_or(LINE_W)); } - for (x, y) in pts.iter().flatten() { - let rect = [x - MARKER / 2.0, y - MARKER / 2.0, MARKER, MARKER]; - out.push(GraphicItem::Shape { rect, kind: ShapeKind::Ellipse, fill: Some(c.series_colour(j)), stroke: None, stroke_width: 0.0 }); + if s.markers != Some(false) { + markers(out, pts.iter().flatten().copied(), c.marker_colour(j), s); } } } diff --git a/crates/docx/src/read/chart/polar.rs b/crates/docx/src/read/chart/polar.rs index 57ad10472..cf5e7452e 100644 --- a/crates/docx/src/read/chart/polar.rs +++ b/crates/docx/src/read/chart/polar.rs @@ -1,8 +1,8 @@ use super::canvas::{Rect, clip_text, fill, stroke, text, text_w, widest}; use super::format::format_value; -use super::model::{DataLabel, Kind, Series}; +use super::model::{DataLabel, Kind, LabelPos, Series}; use super::plot::{Axes, has_values}; -use super::{Chart, FRAC_PI_2, GAP, GRID, GraphicItem, LINE, PathSeg, Rgb, ShapeKind, TAU, TEXT, TextAlign}; +use super::{Chart, FRAC_PI_2, GAP, GRID, GraphicItem, LINE, LINE_W, PathSeg, Rgb, ShapeKind, TAU, TEXT, TextAlign}; pub(super) fn pie(out: &mut Vec, c: &Chart, area: Rect) { let Some(s) = c.series.first() else { return }; @@ -38,7 +38,7 @@ pub(super) fn pie(out: &mut Vec, c: &Chart, area: Rect) { // Halfway across the ring (pie: two thirds out from the centre). let at = if inner > 0.0 { (r + inner) / 2.0 } else { r * 0.65 }; let (x, y) = polar(cx, cy, at, (a + b) / 2.0); - labels.push(DataLabel { x: x as f32, y: y as f32, text, inside: true }); + labels.push(DataLabel { x: x as f32, y: y as f32, text, inside: true, pos: LabelPos::Above, off: 0.0 }); } a = b; } @@ -92,14 +92,14 @@ pub(super) fn radar(out: &mut Vec, c: &Chart, area: Rect) { if c.radar_fill { fill(out, segs.clone(), col); } - stroke(out, segs, col, s.line_w); + stroke(out, segs, col, s.line_w.unwrap_or(LINE_W)); for (i, p) in ring.iter().enumerate() { if c.radar_marks && s.values.get(i).copied().flatten().is_some() { let rect = [p.0 as f32 - MARKER / 2.0, p.1 as f32 - MARKER / 2.0, MARKER, MARKER]; out.push(GraphicItem::Shape { rect, kind: ShapeKind::Ellipse, fill: Some(col), stroke: None, stroke_width: 0.0 }); } if let Some(text) = label_of(s, i) { - labels.push(DataLabel { x: p.0 as f32, y: p.1 as f32, text, inside: false }); + labels.push(DataLabel { x: p.0 as f32, y: p.1 as f32, text, inside: false, pos: LabelPos::Above, off: 0.0 }); } } } @@ -134,13 +134,13 @@ pub(super) fn radar(out: &mut Vec, c: &Chart, area: Rect) { } /// The data label of point `i` of series `s`, when it shows one: the value as the file has it, -/// in the series' number format. +/// in the labels' own number format, else the series' one. pub(super) fn label_of(s: &Series, i: usize) -> Option { if s.labels != Some(true) { return None; } let v = s.raw.get(i).copied().flatten()?; - let fmt = s.fmt.as_deref().filter(|f| !f.eq_ignore_ascii_case("General")); + let fmt = s.label_fmt.as_ref().map_or(s.fmt.as_deref(), |f| f.as_deref().or(s.fmt.as_deref())).filter(|f| !f.eq_ignore_ascii_case("General")); Some(match fmt { Some(f) => format_value(v, f, 1.0), None => general_value(v), @@ -165,13 +165,22 @@ pub(super) fn draw_data_labels(out: &mut Vec, labels: &[DataLabel], let w = text_w(&l.text, fs) + 4.0; let (rect, align) = if l.inside { (Rect { x: l.x - w / 2.0, y: l.y - lh / 2.0, w, h: lh }, TextAlign::Center) + } else if l.off > 0.0 || l.pos != LabelPos::Above { + let o = l.off; + match l.pos { + LabelPos::Right => (Rect { x: l.x + o, y: l.y - lh / 2.0, w, h: lh }, TextAlign::Left), + LabelPos::Left => (Rect { x: l.x - o - w, y: l.y - lh / 2.0, w, h: lh }, TextAlign::Right), + LabelPos::Below => (Rect { x: l.x - w / 2.0, y: l.y + o, w, h: lh }, TextAlign::Center), + LabelPos::Center => (Rect { x: l.x - w / 2.0, y: l.y - lh / 2.0, w, h: lh }, TextAlign::Center), + LabelPos::Above => (Rect { x: l.x - w / 2.0, y: l.y - o - lh, w, h: lh }, TextAlign::Center), + } } else if beside_right { (Rect { x: l.x + GAP / 2.0, y: l.y - lh / 2.0, w, h: lh }, TextAlign::Left) } else { (Rect { x: l.x - w / 2.0, y: l.y - lh, w, h: lh }, TextAlign::Center) }; let color = if l.inside { Rgb::WHITE } else { TEXT }; - out.push(GraphicItem::Text { rect: rect.arr(), text: l.text.clone(), size: fs, color, bold: false, align, font: None }); + out.push(GraphicItem::Text { rect: rect.arr(), text: l.text.clone(), size: fs, color, bold: false, align, font: None, angle: 0.0 }); } } diff --git a/crates/docx/src/read/chart/tests.rs b/crates/docx/src/read/chart/tests.rs index 6e8bcfd82..ae80c028e 100644 --- a/crates/docx/src/read/chart/tests.rs +++ b/crates/docx/src/read/chart/tests.rs @@ -386,9 +386,13 @@ fn axis_format_code_length_is_capped() { let x = format!(r#""#); axis(Some(&crate::xml::parse(x.as_bytes()).unwrap())).fmt }; - let ok = format!("{}0", "x".repeat(254)); + use craft_wordmodel::chart::MAX_FMT; + // The limit is the model's, counted in characters (a CJK character is one). + let ok = format!("{}0", "x".repeat(MAX_FMT - 1)); assert_eq!(fmt(&ok), Some(ok.clone())); - assert_eq!(fmt(&format!("{}0", "x".repeat(300))), None); + let cjk = format!("{}0", "\u{5143}".repeat(MAX_FMT - 1)); + assert_eq!(fmt(&cjk), Some(cjk.clone())); + assert_eq!(fmt(&format!("{}0", "x".repeat(MAX_FMT))), None); } #[test] @@ -590,3 +594,308 @@ fn series_without_a_usable_c_idx_take_their_place() { let big = [renumbered(&a, &format!(r#""#, u32::MAX))]; assert_eq!(column_colours(&big).0.len(), 1); } + +#[test] +fn bar_category_labels_start_after_the_axis_title() { + let long = "A category label that is much too long for its column"; + let title = r#"Region"#; + let plot = format!( + r#"{}{title}"#, + ser(0, "S", Some("FF0000"), &[long, "Short"], &["4", "8"]), + ); + let items = chart_items(&space(&plot), &theme(), 300.0, 200.0); + let rect_of = |f: &dyn Fn(&str, &TextAlign) -> bool| -> Vec<[f32; 4]> { + items + .iter() + .filter_map(|it| if let GraphicItem::Text { rect, text, align, .. } = it { f(text, align).then_some(*rect) } else { None }) + .collect() + }; + let title_rect = rect_of(&|t, _| t == "Region"); + assert_eq!(title_rect.len(), 1); + let strip_end = title_rect[0][0] + title_rect[0][2]; + let labels = rect_of(&|t, a| matches!(a, TextAlign::Right) && t != "Region" && (long.contains(t) || t == "Short")); + assert!(!labels.is_empty()); + assert!(labels.iter().all(|r| r[0] >= strip_end - 1e-3), "labels {labels:?} overlap the title strip ending at {strip_end}"); + assert!(labels.iter().all(|r| r[2] <= 0.4 * 300.0), "{labels:?}"); +} + +#[test] +fn category_gridlines_follow_the_points_when_the_axis_crosses_mid() { + let line = |cross: &str| { + let plot = format!( + r#"{}"#, + ser(0, "L", Some("FF0000"), &["Q1", "Q2", "Q3"], &["1", "5", "3"]), + ); + let items = chart_items(&space(&plot), &theme(), 300.0, 200.0); + let grid: Vec = items + .iter() + .filter_map(|it| match it { + GraphicItem::Path { segs, stroke: Some(c), .. } if *c == GRID => match segs.first() { + Some(PathSeg::Move(x, _)) => Some(*x), + _ => None, + }, + _ => None, + }) + .collect(); + let pts: Vec = items + .iter() + .filter_map(|it| match it { + GraphicItem::Path { segs, stroke: Some(c), .. } if *c == Rgb(255, 0, 0) => Some( + segs.iter().filter_map(|s| if let PathSeg::Move(x, _) | PathSeg::Line(x, _) = s { Some(*x) } else { None }).collect::>(), + ), + _ => None, + }) + .next() + .unwrap_or_default(); + (grid, pts) + }; + for cross in ["midCat", "between"] { + let (grid, pts) = line(cross); + assert_eq!(grid.len(), 3, "{cross}"); + assert_eq!(pts.len(), 3, "{cross}"); + assert!(grid.iter().zip(&pts).all(|(g, p)| (g - p).abs() < 1e-3), "{cross}: grid {grid:?} points {pts:?}"); + } +} + +/// A line chart with one series; `extra` goes into the series (`c:marker`, `c:dLbls`), `axes` +/// into the value axis. +fn line_with(extra: &str, val_axis: &str) -> Vec { + let s = ser(0, "S", Some("FF0000"), &["Q1", "Q2", "Q3"], &["3.2", "4.1", "5"]).replace("", &format!("{extra}")); + let plot = format!( + r#"{s}{val_axis}"# + ); + chart_items(&space(&plot), &theme(), 300.0, 200.0) +} + +fn marker_xml(symbol: &str, size: u32) -> String { + format!(r#""#) +} + +#[test] +fn marker_symbols_are_drawn_as_their_shape() { + let shapes = |items: &[GraphicItem], kind: ShapeKind| { + items + .iter() + .filter(|it| matches!(it, GraphicItem::Shape { kind: k, rect, .. } if *k == kind && rect[2] > 3.0 && rect[2] < 20.0 && rect[3] > 1.0)) + .count() + }; + let paths = |items: &[GraphicItem], filled: bool| { + items + .iter() + .filter(|it| matches!(it, GraphicItem::Path { fill, segs, .. } if fill.is_some() == filled && segs.len() <= 12 && segs.len() >= 3)) + .count() + }; + // One marker per point, in the series' size. + let circle = line_with(&marker_xml("circle", 9), ""); + assert_eq!(shapes(&circle, ShapeKind::Ellipse), 3); + assert!(circle.iter().any(|it| matches!(it, GraphicItem::Shape { kind: ShapeKind::Ellipse, rect, .. } if (rect[2] - 9.0).abs() < 1e-3))); + let square = line_with(&marker_xml("square", 7), ""); + assert_eq!(shapes(&square, ShapeKind::Rectangle) - shapes(&circle, ShapeKind::Rectangle), 3, "a square per point"); + assert_eq!(shapes(&square, ShapeKind::Ellipse), 0, "no circles"); + // Diamonds and triangles are filled polygons; x, plus and star are strokes. + let base = paths(&line_with("", ""), true); + assert_eq!(paths(&line_with(&marker_xml("diamond", 7), ""), true) - base, 3); + assert_eq!(paths(&line_with(&marker_xml("triangle", 7), ""), true) - base, 3); + for symbol in ["x", "plus", "star"] { + let items = line_with(&marker_xml(symbol, 7), ""); + assert!(items.iter().any(|it| matches!(it, GraphicItem::Path { fill: None, stroke: Some(_), segs, .. } if segs.len() >= 6)), "{symbol}"); + assert_eq!(shapes(&items, ShapeKind::Ellipse), 0, "{symbol} is not a circle"); + } + // Dash and dot are small shapes; `none` and unknown names never crash (unknown: a circle). + assert_eq!(shapes(&line_with(&marker_xml("dash", 7), ""), ShapeKind::Rectangle) - shapes(&circle, ShapeKind::Rectangle), 3); + assert_eq!(shapes(&line_with(&marker_xml("dot", 7), ""), ShapeKind::Ellipse), 3); + assert_eq!(shapes(&line_with(&marker_xml("none", 7), ""), ShapeKind::Ellipse), 0); + assert_eq!(shapes(&line_with(&marker_xml("wibble", 7), ""), ShapeKind::Ellipse), 3); + // Hostile sizes are clamped. + for size in ["0", "99999999999", "-4", "x"] { + let items = line_with(&format!(r#""#), ""); + assert!(items.iter().all(|it| match it { + GraphicItem::Shape { rect, .. } => rect.iter().all(|v| v.is_finite()) && rect[2] <= 300.0, + _ => true, + })); + } +} + +#[test] +fn data_labels_use_their_own_format_code() { + let dl = |num: &str| format!(r#"{num}"#); + let shown = |items: &[GraphicItem]| -> Vec { + items.iter().filter_map(|it| if let GraphicItem::Text { text, .. } = it { Some(text.clone()) } else { None }).collect() + }; + let own = shown(&line_with(&dl(r#""#), "")); + assert!(own.contains(&"3.20".to_string()) && own.contains(&"5.00".to_string()), "{own:?}"); + // Linked to the source: the series' format (General here). + let linked = shown(&line_with(&dl(r#""#), "")); + assert!(linked.contains(&"3.2".to_string()) && linked.contains(&"5".to_string()), "{linked:?}"); + // A format code beyond the cap is ignored, not applied. + let long = "0".repeat(craft_wordmodel::chart::MAX_FMT + 1); + let capped = shown(&line_with(&dl(&format!(r#""#)), "")); + assert!(capped.contains(&"3.2".to_string()), "{capped:?}"); +} + +#[test] +fn value_axis_title_turns_counter_clockwise_like_word() { + let title = |body: &str| format!(r#"{body}Revenue"#); + let angle = |items: &[GraphicItem]| { + items.iter().find_map(|it| if let GraphicItem::Text { text, angle, .. } = it { (text == "Revenue").then_some(*angle) } else { None }) + }; + // The file's rotation (`-5400000` is 90° counter-clockwise), or Word's default when it says none. + assert_eq!(angle(&line_with("", &title(r#""#))), Some(-90.0)); + assert_eq!(angle(&line_with("", &title(""))), Some(-90.0)); + assert_eq!(angle(&line_with("", &title(r#""#))), Some(90.0)); + // A horizontal title (`rot="0"`) stays horizontal: the plot makes room for it (see below). + assert_eq!(angle(&line_with("", &title(r#""#))), Some(0.0)); + // Hostile values stay finite. + let odd = angle(&line_with("", &title(r#""#))); + assert!(odd.is_some_and(f32::is_finite)); +} + +fn first_ellipse(items: &[GraphicItem]) -> Option<([f32; 4], Rgb)> { + items.iter().find_map(|it| match it { + GraphicItem::Shape { kind: ShapeKind::Ellipse, rect, fill: Some(c), .. } if rect[2] > 3.0 && rect[2] < 20.0 => Some((*rect, *c)), + _ => None, + }) +} + +#[test] +fn markers_use_the_automatic_series_colour_not_the_line_colour() { + // The series has an explicit red line/fill but no marker fill: Word draws the markers in the + // palette colour of the series index (accent 1 here). + let items = line_with(&marker_xml("circle", 9), ""); + assert_eq!(first_ellipse(&items).map(|e| e.1), Some(theme()[4])); + // A marker's own fill wins. + let own = + r#""#; + assert_eq!(first_ellipse(&line_with(own, "")).map(|e| e.1), Some(Rgb(0, 0xAA, 0))); +} + +#[test] +fn line_data_labels_default_to_the_right_and_honour_dlblpos() { + let dl = |pos: &str| format!(r#"{}{pos}"#, marker_xml("circle", 9)); + let text_rect = |items: &[GraphicItem]| { + items.iter().find_map(|it| if let GraphicItem::Text { text, rect, .. } = it { (text == "3.2").then_some(*rect) } else { None }) + }; + let at = |pos: &str| { + let items = line_with(&dl(pos), ""); + let (m, _) = first_ellipse(&items).unwrap(); + let t = text_rect(&items).unwrap(); + // Centre of the marker and of the label box. + ((t[0] + t[2] / 2.0) - (m[0] + m[2] / 2.0), (t[1] + t[3] / 2.0) - (m[1] + m[3] / 2.0), t, m) + }; + let (dx, dy, t, m) = at(""); + assert!(t[0] >= m[0] + m[2] && dy.abs() < 1.0, "default is right of the marker: {dx} {dy}"); + let (dx, dy, t, m) = at(r#""#); + assert!(t[0] >= m[0] + m[2] && dy.abs() < 1.0, "{dx} {dy}"); + let (dx, dy, _, _) = at(r#""#); + assert!(dx.abs() < 1.0 && dy < -4.0, "{dx} {dy}"); + let (dx, dy, _, _) = at(r#""#); + assert!(dx.abs() < 1.0 && dy > 4.0, "{dx} {dy}"); + let (dx, dy, t, m) = at(r#""#); + assert!(t[0] + t[2] <= m[0] && dy.abs() < 1.0, "{dx} {dy}"); + let (dx, dy, _, _) = at(r#""#); + assert!(dx.abs() < 1.0 && dy.abs() < 1.0, "{dx} {dy}"); + // Unknown positions fall back to the default. + let (_, _, t, m) = at(r#""#); + assert!(t[0] >= m[0] + m[2]); +} + +#[test] +fn horizontal_value_axis_title_gets_its_own_strip() { + let title = |rot: &str, name: &str| { + format!(r#"{name}"#) + }; + let plot_left = |items: &[GraphicItem]| { + // The first gridline's start is the plot's left edge. + items.iter().find_map(|it| match it { + GraphicItem::Path { segs, stroke: Some(_), fill: None, .. } if segs.len() == 2 => match segs.first() { + Some(PathSeg::Move(x, _)) => Some(*x), + _ => None, + }, + _ => None, + }) + }; + let rect_of = |items: &[GraphicItem], name: &str| { + items.iter().find_map(|it| if let GraphicItem::Text { text, rect, .. } = it { text.starts_with(name).then_some(*rect) } else { None }) + }; + let turned = line_with("", &title(r#" rot="-5400000" vert="horz""#, "Revenue")); + let flat = line_with("", &title(r#" rot="0" vert="horz""#, "Revenue")); + let (a, b) = (plot_left(&turned).unwrap(), plot_left(&flat).unwrap()); + assert!(b > a + 10.0, "a horizontal title takes more room than a turned one: {a} {b}"); + // The title box lies left of the plot and inside the chart. + let r = rect_of(&flat, "Revenue").unwrap(); + assert!(r[0] >= 0.0 && r[0] + r[2] <= b + 0.5, "{r:?} {b}"); + // A hostile title never takes more than a quarter of the width and stays finite. + let long = "Wibble ".repeat(500); + let items = line_with("", &title(r#" rot="0""#, &long)); + let p = plot_left(&items).unwrap(); + assert!(p.is_finite() && p <= 300.0 * 0.3, "{p}"); +} + +/// A one-series chart whose series carries `sp` as its `c:spPr`. +fn chart_with_sp(kind: &str, sp: &str) -> Vec { + let s = ser(0, "S", None, &["Q1", "Q2"], &["4", "8"]).replace("", &format!("{sp}")); + let plot = match kind { + "area" => format!( + r#"{s}"# + ), + dir => format!( + r#"{s}"# + ), + }; + chart_items(&space(&plot), &theme(), 300.0, 200.0) +} + +/// The series' closed shapes: (filled, stroked with the given width). +fn series_shapes(items: &[GraphicItem], width: f32) -> (usize, usize) { + let closed = |segs: &[PathSeg]| segs.iter().any(|s| matches!(s, PathSeg::Close)); + let fills = items.iter().filter(|it| matches!(it, GraphicItem::Path { segs, fill: Some(_), .. } if closed(segs))).count(); + let strokes = items + .iter() + .filter(|it| matches!(it, GraphicItem::Path { segs, fill: None, stroke: Some(_), stroke_width, .. } if closed(segs) && (*stroke_width - width).abs() < 0.01)) + .count(); + (fills, strokes) +} + +#[test] +fn unfilled_bars_keep_their_explicit_outline() { + let outlined = r#""#; + let no_line = r#""#; + for dir in ["col", "bar"] { + // Two points, two outlines, no fill. + assert_eq!(series_shapes(&chart_with_sp(dir, outlined), 2.0), (0, 2), "{dir}"); + // Control: no fill and no line draws nothing. + assert_eq!(series_shapes(&chart_with_sp(dir, no_line), 2.0), (0, 0), "{dir}"); + } +} + +#[test] +fn no_fill_area_series_is_not_painted() { + let filled = r#""#; + let none = r#""#; + assert_eq!(series_shapes(&chart_with_sp("area", filled), 2.0).0, 1); + assert_eq!(series_shapes(&chart_with_sp("area", none), 2.0).0, 0); +} + +#[test] +fn source_linked_series_labels_do_not_inherit_the_chart_format() { + let dl = |fmt: &str| format!(r#""#); + let shown = |items: &[GraphicItem]| -> Vec { + items.iter().filter_map(|it| if let GraphicItem::Text { text, .. } = it { Some(text.clone()) } else { None }).collect() + }; + // Chart-level 0%; the series' own labels are `linked` (or have no labels element at all). + let run = |series_dl: &str| { + let s = ser(0, "S", Some("FF0000"), &["Q1", "Q2", "Q3"], &["3.2", "4.1", "5"]).replace("", &format!("{series_dl}")); + let chart_dl = dl("0%").replace("{}", "0"); + let plot = format!( + r#"{s}{chart_dl}"# + ); + shown(&chart_items(&space(&plot), &theme(), 300.0, 200.0)) + }; + for linked in ["1", "true"] { + let own = run(&dl("General").replace("{}", linked)); + assert!(own.contains(&"3.2".to_string()) && !own.contains(&"320%".to_string()), "sourceLinked={linked}: {own:?}"); + } + // Control: a series without a label format still takes the chart's. + let inherited = run(r#""#); + assert!(inherited.contains(&"320%".to_string()), "{inherited:?}"); +} diff --git a/crates/docx/src/read/diagram.rs b/crates/docx/src/read/diagram.rs index 7f03e8b51..7b7bbeebf 100644 --- a/crates/docx/src/read/diagram.rs +++ b/crates/docx/src/read/diagram.rs @@ -463,7 +463,16 @@ fn text(sp: &El, xf: Xf, b: [f64; 4], theme: &[Rgb], filled: bool, fit: f32, out let mut y = y0 + (th - total) / 2.0; for l in lines { let lh = l.size * LINE; - out.push(GraphicItem::Text { rect: [x0, y, tw, lh], text: l.text, size: l.size, color: l.color, bold: l.bold, align: l.align, font: l.font }); + out.push(GraphicItem::Text { + rect: [x0, y, tw, lh], + text: l.text, + size: l.size, + color: l.color, + bold: l.bold, + align: l.align, + font: l.font, + angle: 0.0, + }); y += lh; } } diff --git a/crates/docx/src/read/story/blocks.rs b/crates/docx/src/read/story/blocks.rs index 38c35a02b..878429454 100644 --- a/crates/docx/src/read/story/blocks.rs +++ b/crates/docx/src/read/story/blocks.rs @@ -249,10 +249,11 @@ impl Reader<'_> { .unwrap_or_default(); para.props.fmt_change = PropChange::boxed(self.revision(RevisionKind::Format, ch), old); } - if para.props.numbering.is_some() - && let Some(ch) = ppr.child("w:numPr").and_then(|n| n.child("w:numberingChange")) - { - let original = ch.attr("w:original").unwrap_or("").chars().take(256).collect(); + if let Some(ch) = ppr.child("w:numPr").and_then(|n| n.child("w:numberingChange")) { + // Multi-level numbering caches (`w:original`) routinely exceed 256 chars while staying + // bounded against hostile attributes. + const MAX_CHANGE_ORIGINAL: usize = 4096; + let original = ch.attr("w:original").unwrap_or("").chars().take(MAX_CHANGE_ORIGINAL).collect(); para.props.num_change = Some(Box::new(NumChange { rev: self.revision(RevisionKind::Format, ch), original })); } if let Some(s) = ppr.child("w:sectPr") { diff --git a/crates/docx/src/write/embed.rs b/crates/docx/src/write/embed.rs index 9a8051e09..7673a6f26 100644 --- a/crates/docx/src/write/embed.rs +++ b/crates/docx/src/write/embed.rs @@ -884,10 +884,11 @@ mod tests { let spec = ChartSpec { kind: ChartType::Column, categories: vec!["a".into(), "b".into()], - series: (0..series).map(|i| ChartSeries { name: format!("S{i}"), values: vec![Some(1.0); 2] }).collect(), + series: (0..series).map(|i| ChartSeries { name: format!("S{i}"), values: vec![Some(1.0); 2], ..Default::default() }).collect(), title: None, legend: None, data_labels: false, + ..Default::default() }; let graphic = Graphic { kind: GraphicKind::Chart, w: 100.0, h: 100.0, chart: Some(Arc::new(spec)), ..Default::default() }; InlineObject::Graphic { w: 100.0, h: 100.0, alt: String::new(), float: Default::default(), graphic: Arc::new(graphic) } diff --git a/crates/docx/src/write/props.rs b/crates/docx/src/write/props.rs index 0a64fc20a..0b92ce191 100644 --- a/crates/docx/src/write/props.rs +++ b/crates/docx/src/write/props.rs @@ -261,10 +261,12 @@ pub fn ppr_inner(w: &mut W, p: &ParaProps, framed: bool, num_change: Option<&Cha toggle(w, "w:pageBreakBefore", p.page_break_before); write_frame(w, p, framed); toggle(w, "w:widowControl", p.widow_control); - if let Some(nr) = p.numbering { + if p.numbering.is_some() || num_change.is_some() { w.open("w:numPr", &[]); - w.val("w:ilvl", &n(nr.level.min(8) as i64)); - w.val("w:numId", &n(nr.num as i64)); + if let Some(nr) = p.numbering { + w.val("w:ilvl", &n(nr.level.min(8) as i64)); + w.val("w:numId", &n(nr.num as i64)); + } if let Some(a) = num_change { let refs: Vec<(&str, &str)> = a.iter().map(|(k, v)| (*k, v.as_str())).collect(); w.empty("w:numberingChange", &refs); diff --git a/crates/docx/tests/it/fixtures.rs b/crates/docx/tests/it/fixtures.rs index 868e8cbb0..348d6239e 100644 --- a/crates/docx/tests/it/fixtures.rs +++ b/crates/docx/tests/it/fixtures.rs @@ -2197,6 +2197,50 @@ fn formatting_revisions_round_trip() { assert_eq!(again.revisions, d.revisions); } +#[test] +fn numbering_change_without_active_list_roundtrips() { + let body = r#"Removed"#; + let d = read_body(body); + let p = paras(&d); + assert_eq!(p.len(), 1); + assert!(p[0].props.numbering.is_none()); + let nc = p[0].props.num_change.as_deref().expect("numberingChange read back"); + assert_eq!(nc.original, "%1:1:0:."); + + let bytes = craft_docx::write(&d).unwrap(); + let mut zip = zip::ZipArchive::new(std::io::Cursor::new(&bytes)).unwrap(); + let mut xml = String::new(); + std::io::Read::read_to_string(&mut zip.by_name("word/document.xml").unwrap(), &mut xml).unwrap(); + assert!(xml.contains(r#""#), "{xml}"); + + let again = craft_docx::read(&bytes).unwrap(); + let ap = paras(&again); + assert!(ap[0].props.numbering.is_none()); + let anc = ap[0].props.num_change.as_deref().expect("numberingChange preserved after save"); + assert_eq!(anc.original, "%1:1:0:."); +} + +#[test] +fn numbering_change_original_preserves_long_values_within_bound() { + let long_orig = "%1:1:0:.".repeat(64); // 512 chars, well past the old 256-char cap + assert!(long_orig.len() > 256 && long_orig.len() < 4096); + let over_cap = "x".repeat(5000); + let body = format!( + r#"Long +Bounded"# + ); + let d = read_body(&body); + let p = paras(&d); + assert_eq!(p[0].props.num_change.as_deref().map(|n| n.original.as_str()), Some(long_orig.as_str())); + assert_eq!(p[1].props.num_change.as_deref().map(|n| n.original.chars().count()), Some(4096)); + + let bytes = craft_docx::write(&d).unwrap(); + let again = craft_docx::read(&bytes).unwrap(); + let ap = paras(&again); + assert_eq!(ap[0].props.num_change.as_deref().map(|n| n.original.as_str()), Some(long_orig.as_str())); +} + /// A package with `styles`, a theme (accent1 4472C4, accent2 ED7D31, major font "Major Face") and, /// when given, `settings`. fn themed_docx(body: &str, styles: &str, settings: Option<&str>) -> Document { 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/exr/src/lib.rs b/crates/exr/src/lib.rs index f1550f8de..78afd427e 100644 --- a/crates/exr/src/lib.rs +++ b/crates/exr/src/lib.rs @@ -18,6 +18,8 @@ //! - [`stream`]: one flat scan-line part read block by block, for pictures larger than memory. //! - [`cryptomatte_samples`]: a file's Cryptomatte layers and their per-pixel (ID, coverage) ranks. +/// Scanline OpenEXR writer using a caller-supplied linearization table. +pub mod rgba8; /// RFC 1950/1951 compressor for ZIP-compressed EXR chunks. pub mod zlib; diff --git a/crates/exr/src/rgba8.rs b/crates/exr/src/rgba8.rs new file mode 100644 index 000000000..6dc4bde4c --- /dev/null +++ b/crates/exr/src/rgba8.rs @@ -0,0 +1,235 @@ +//! Single-part scanline OpenEXR output from RGBA8 pixels. + +/// Scanline chunk compression. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub enum Compression { + /// Uncompressed. + None, + /// OpenEXR RLE. + Rle, + /// zlib per scanline. + Zips, + /// zlib per 16 scanlines. + Zip, +} +impl Compression { + fn code(self) -> u8 { + match self { + Self::None => 0, + Self::Rle => 1, + Self::Zips => 2, + Self::Zip => 3, + } + } + fn lines(self) -> usize { + if self == Self::Zip { 16 } else { 1 } + } +} +/// Why a frame cannot be encoded. +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub enum Error { + /// Empty, overflowing or mismatched RGBA dimensions. + FrameSize, + /// Dimensions exceed the OpenEXR signed-size fields. + TooLarge, +} +fn checked_len(rgba: &[u8], w: u32, h: u32) -> Result { + let n = (w as usize).checked_mul(h as usize).and_then(|n| n.checked_mul(4)).ok_or(Error::FrameSize)?; + if w == 0 || h == 0 || rgba.len() != n { + return Err(Error::FrameSize); + } + Ok(n) +} + +/// IEEE 754 binary16 bits of `v`, rounded to nearest even (overflow → infinity, NaN → quiet NaN). +pub fn half(v: f32) -> u16 { + let b = v.to_bits(); + let sign = ((b >> 16) & 0x8000) as u16; + let exp = ((b >> 23) & 0xFF) as i32; + let man = b & 0x7F_FFFF; + if exp == 0xFF { + return sign | 0x7C00 | if man != 0 { 0x200 } else { 0 }; + } + // rounds `m >> shift` to nearest, ties to even (`shift` is 1..=24) + let round = |m: u32, shift: u32| { + let (q, rem, mid) = (m >> shift, m & ((1 << shift) - 1), 1 << (shift - 1)); + if rem > mid || (rem == mid && q & 1 == 1) { q + 1 } else { q } + }; + let e = exp - 127 + 15; + if e >= 0x1F { + sign | 0x7C00 + } else if e <= 0 { + if e < -10 { + return sign; + } + // subnormal half: the 24-bit significand scaled by 2^(e - 14); rounding up may carry + // into the smallest normal, which is the right encoding + sign | round(man | 0x80_0000, (14 - e) as u32) as u16 + } else { + // a carry out of the mantissa bumps the exponent (up to infinity), as it should + sign | round(((e as u32) << 23) | man, 13) as u16 + } +} + +/// An OpenEXR file: half-float scene-linear Rec.709 RGB, or RGBA with `alpha` (premultiplied, as +/// EXR defines alpha; the frame's straight alpha is multiplied in after linearisation). +pub fn write_rgba8(rgba: &[u8], w: u32, h: u32, alpha: bool, compression: Compression, linear: &[f32; 256]) -> Result, Error> { + let n = checked_len(rgba, w, h)?; + let too_large = || Error::TooLarge; + let (Ok(x_max), Ok(y_max)) = (i32::try_from(w - 1), i32::try_from(h - 1)) else { + return Err(too_large()); + }; + // channels in the alphabetical order the file layout requires + let names: &[u8] = if alpha { b"ABGR" } else { b"BGR" }; + let line = (w as usize).checked_mul(names.len() * 2).ok_or(Error::FrameSize)?; + // a chunk's byte count is a signed 32-bit field + let lines = compression.lines(); + if line.checked_mul(lines).and_then(|c| i32::try_from(c).ok()).is_none() { + return Err(too_large()); + } + + let mut hd = Vec::with_capacity(512); + hd.extend_from_slice(&[0x76, 0x2F, 0x31, 0x01]); // magic number 20000630 + hd.extend_from_slice(&[2, 0, 0, 0]); // version 2, single-part scan-line file, short names + let attr = |hd: &mut Vec, name: &str, ty: &str, value: &[u8]| { + for s in [name, ty] { + hd.extend_from_slice(s.as_bytes()); + hd.push(0); + } + hd.extend_from_slice(&(value.len() as i32).to_le_bytes()); + hd.extend_from_slice(value); + }; + let le = |v: &[f32]| v.iter().flat_map(|f| f.to_le_bytes()).collect::>(); + let mut channels = Vec::new(); + for &c in names { + channels.extend_from_slice(&[c, 0]); + channels.extend_from_slice(&1_i32.to_le_bytes()); // pixel type HALF + channels.extend_from_slice(&[0, 0, 0, 0]); // pLinear 0, reserved + channels.extend_from_slice(&1_i32.to_le_bytes()); // x sampling + channels.extend_from_slice(&1_i32.to_le_bytes()); // y sampling + } + channels.push(0); + let window: Vec = [0, 0, x_max, y_max].iter().flat_map(|v| v.to_le_bytes()).collect(); + attr(&mut hd, "channels", "chlist", &channels); + attr(&mut hd, "chromaticities", "chromaticities", &le(&[0.64, 0.33, 0.30, 0.60, 0.15, 0.06, 0.3127, 0.3290])); + attr(&mut hd, "compression", "compression", &[compression.code()]); + attr(&mut hd, "dataWindow", "box2i", &window); + attr(&mut hd, "displayWindow", "box2i", &window); + attr(&mut hd, "lineOrder", "lineOrder", &[0]); // INCREASING_Y + attr(&mut hd, "pixelAspectRatio", "float", &le(&[1.0])); + attr(&mut hd, "screenWindowCenter", "v2f", &le(&[0.0, 0.0])); + attr(&mut hd, "screenWindowWidth", "float", &le(&[1.0])); + hd.push(0); // end of header + + // offset table (one entry per chunk), then the chunks: first y, data size, data + let chunks = (h as usize).div_ceil(lines); + let table = chunks.checked_mul(8); + let body = (h as usize).checked_mul(line + 8); + let size = table.zip(body).and_then(|(t, b)| hd.len().checked_add(t)?.checked_add(b)).ok_or(Error::FrameSize)?; + // the uncompressed size: compressed chunks only shrink it + let mut out = Vec::with_capacity(size); + out.extend_from_slice(&hd); + let table_at = out.len(); + out.resize(table_at + chunks * 8, 0); + let lin = linear; + let opaque: [u16; 256] = std::array::from_fn(|i| half(lin[i])); + let row = w as usize * 4; + let wn = w as usize; + let mut px = vec![0_u16; wn * names.len()]; + let mut raw = Vec::with_capacity(line * lines); + // the deflate hash tables, shared by the frame's chunks instead of zeroed for each one + let mut tables = crate::zlib::Tables::default(); + for (k, block) in rgba[..n].chunks(row * lines).enumerate() { + raw.clear(); + for src in block.chunks_exact(row) { + // planar per line: every pixel of the first channel, then the next channel … + for (x, &[r, g, b, a]) in src.as_chunks::<4>().0.iter().enumerate() { + let rgb = [b, g, r]; + if alpha { + let k = f32::from(a) / 255.0; + px[x] = half(k); + for (c, &v) in rgb.iter().enumerate() { + px[(c + 1) * wn + x] = if a == 255 { opaque[v as usize] } else { half(lin[v as usize] * k) }; + } + } else { + for (c, &v) in rgb.iter().enumerate() { + px[c * wn + x] = opaque[v as usize]; + } + } + } + raw.extend(px.iter().flat_map(|v| v.to_le_bytes())); + } + let packed = match compression { + Compression::None => None, + Compression::Rle => Some(rle(&predict(&raw))), + Compression::Zips | Compression::Zip => Some(crate::zlib::zlib(&predict(&raw), &mut tables)), + }; + // a chunk that does not shrink is stored as it is (readers compare the sizes) + let data = packed.as_deref().filter(|p| p.len() < raw.len()).unwrap_or(&raw); + let offset = (out.len() as u64).to_le_bytes(); + out[table_at + k * 8..table_at + k * 8 + 8].copy_from_slice(&offset); + out.extend_from_slice(&((k * lines) as i32).to_le_bytes()); + out.extend_from_slice(&(data.len() as i32).to_le_bytes()); + out.extend_from_slice(data); + } + Ok(out) +} + +/// The OpenEXR RLE / ZIP predictor: bytes split into even and odd halves, then each byte replaced +/// by its difference from the previous one, plus 128. +fn predict(raw: &[u8]) -> Vec { + let half = raw.len().div_ceil(2); + let mut t = vec![0_u8; raw.len()]; + for (i, &b) in raw.iter().enumerate() { + t[if i % 2 == 0 { i / 2 } else { half + i / 2 }] = b; + } + let mut p = t.first().copied().unwrap_or(0); + for v in t.iter_mut().skip(1) { + let d = v.wrapping_sub(p).wrapping_add(128); + p = *v; + *v = d; + } + t +} + +/// OpenEXR run-length coding: a signed count byte, then either `count + 1` copies of one byte +/// (count ≥ 0, runs of 3 to 128) or `-count` literal bytes (1 to 127). +pub fn rle(src: &[u8]) -> Vec { + let mut out = Vec::with_capacity(src.len() + src.len() / 127 + 1); + let mut i = 0; + while i < src.len() { + let b = src[i]; + let run = src[i..].iter().take(128).take_while(|&&x| x == b).count(); + if run >= 3 { + out.extend_from_slice(&[(run - 1) as u8, b]); + i += run; + continue; + } + // literals up to the next run of three (or 127 bytes) + let s = i; + while i < src.len() && i - s < 127 { + if src.get(i..i + 3).is_some_and(|t| t[0] == t[1] && t[1] == t[2]) { + break; + } + i += 1; + } + out.push(((i - s) as u8).wrapping_neg()); + out.extend_from_slice(&src[s..i]); + } + out +} + +#[cfg(test)] +mod tests { + use super::*; + #[test] + fn writes_all_scanline_compressions() { + let linear = std::array::from_fn(|i| i as f32 / 255.0); + for c in [Compression::None, Compression::Rle, Compression::Zips, Compression::Zip] { + let bytes = write_rgba8(&[128; 16 * 16 * 4], 16, 16, true, c, &linear).unwrap(); + assert_eq!(&bytes[..4], &[0x76, 0x2f, 0x31, 1]); + assert!(bytes.len() > 100); + } + assert_eq!(write_rgba8(&[], 1, 1, false, Compression::None, &linear), Err(Error::FrameSize)); + } +} diff --git a/crates/fcproj-format/Cargo.toml b/crates/fcproj-format/Cargo.toml new file mode 100644 index 000000000..aa210187d --- /dev/null +++ b/crates/fcproj-format/Cargo.toml @@ -0,0 +1,16 @@ +[package] +name = "craft-fcproj-format" +description = "FilmCraft project envelope and schema migrations" +version.workspace = true +edition.workspace = true +license.workspace = true +rust-version.workspace = true +repository.workspace = true + +[lints] +workspace = true + +[dependencies] +serde = { workspace = true } +serde_json = { workspace = true } +thiserror = { workspace = true } diff --git a/crates/fcproj-format/src/lib.rs b/crates/fcproj-format/src/lib.rs new file mode 100644 index 000000000..98e50cf59 --- /dev/null +++ b/crates/fcproj-format/src/lib.rs @@ -0,0 +1,297 @@ +//! FilmCraft `.fcproj` JSON envelope, version detection and historical migrations. +//! Project and editor-view types are supplied by the caller. +#![cfg_attr(not(test), deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable))] +use serde::{Deserialize, Serialize}; +use serde_json::Value; +/// The `format` tag written in every envelope. +pub const FORMAT_ID: &str = "filmcraft.project"; + +/// A migration upgrades a whole document from schema `n` to `n + 1`. +pub type Migration = fn(Value) -> Result; + +/// `MIGRATIONS[i]` upgrades schema `i + 1` to `i + 2`. Append one function per schema bump; never +/// edit a shipped one. +pub const MIGRATIONS: &[Migration] = &[ + v1_to_v2, v2_to_v3, v3_to_v4, v4_to_v5, v5_to_v6, v6_to_v7, v7_to_v8, v8_to_v9, v9_to_v10, v10_to_v11, v11_to_v12, v12_to_v13, v13_to_v14, + v14_to_v15, v15_to_v16, v16_to_v17, +]; + +/// The schema version this build writes (and the newest it reads). +pub const SCHEMA_VERSION: u32 = MIGRATIONS.len() as u32 + 1; + +/// Oldest schema this build can still upgrade. +pub const OLDEST_SCHEMA: u32 = 1; + +#[derive(Debug, thiserror::Error)] +pub enum FormatError { + #[error("not a FilmCraft project: {0}")] + NotAProject(String), + #[error( + "this project was saved by a newer version of FilmCraft (project schema v{found}); this build reads up to v{supported}. Update FilmCraft to open it." + )] + TooNew { found: u32, supported: u32 }, + #[error("project schema v{found} is too old for this build (oldest supported: v{oldest})")] + TooOld { found: u32, oldest: u32 }, + #[error("upgrading project from schema v{from} to v{}: {msg}", from + 1)] + Migration { from: u32, msg: String }, + #[error("project file is damaged: {0}")] + Corrupt(String), +} + +/// Serialized current-schema envelope. +#[derive(Serialize)] +pub struct EnvelopeOut<'a, P, V> { + pub format: &'static str, + pub schema_version: u32, + pub generator: String, + pub project: &'a P, + #[serde(skip_serializing_if = "Option::is_none")] + pub view: Option<&'a V>, +} +/// Parsed envelope. View parsing remains optional to preserve forward compatibility. +#[derive(Deserialize)] +pub struct EnvelopeIn

{ + #[serde(default)] + pub generator: Option, + pub project: P, + #[serde(default)] + pub view: Option, +} +/// Serialize the current envelope, with optional editor view and caller-provided generator. +pub fn encode(project: &P, view: Option<&V>, generator: String, pretty: bool) -> Vec { + let envelope = EnvelopeOut { format: FORMAT_ID, schema_version: SCHEMA_VERSION, generator, project, view }; + let result = if pretty { serde_json::to_vec_pretty(&envelope) } else { serde_json::to_vec(&envelope) }; + result.unwrap_or_default() +} + +/// Decode an envelope, migrating older documents while retaining the current-version fast path. +pub fn decode(bytes: &[u8]) -> Result<(EnvelopeIn

, u32), FormatError> { + let found = detect_bytes(bytes)?; + if found == SCHEMA_VERSION { + let env = serde_json::from_slice(bytes).map_err(|e| FormatError::Corrupt(e.to_string()))?; + return Ok((env, found)); + } + let doc: Value = serde_json::from_slice(bytes).map_err(|e| FormatError::Corrupt(e.to_string()))?; + let doc = migrate_with(MIGRATIONS, doc, found)?; + let env = serde_json::from_value(doc).map_err(|e| FormatError::Corrupt(format!("after upgrading from schema v{found}: {e}")))?; + Ok((env, found)) +} + +/// Probe and validate the schema directly from bytes without building a JSON tree. +pub fn detect_bytes(bytes: &[u8]) -> Result { + let probe: Probe = serde_json::from_slice(bytes) + .map_err(|e| if e.is_syntax() || e.is_eof() { FormatError::Corrupt(e.to_string()) } else { FormatError::NotAProject(e.to_string()) })?; + let found = probe_version(&probe)?; + check_supported(found)?; + Ok(found) +} +/// Just enough of a document to know which schema it is. +#[derive(Deserialize)] +struct Probe { + #[serde(default)] + format: Option, + #[serde(default)] + schema_version: Option, + /// Bare (pre-envelope) project: its `version` field. + #[serde(default)] + version: Option, + #[serde(default)] + items: Option, + #[serde(default)] + root: Option, +} + +/// Detect the schema version of a parsed document. +pub fn detect_version(doc: &Value) -> Result { + let probe: Probe = serde_json::from_value(doc.clone()).map_err(|e| FormatError::NotAProject(e.to_string()))?; + probe_version(&probe) +} + +fn probe_version(p: &Probe) -> Result { + if let Some(v) = p.schema_version { + if let Some(f) = &p.format + && f != FORMAT_ID + { + return Err(FormatError::NotAProject(format!("format is `{f}`, expected `{FORMAT_ID}`"))); + } + return Ok(v); + } + // Before the envelope (v1) a file was the bare Project object with `version`. + // Its `version` was 1, or 2 on builds that briefly bumped it when caption tracks were added + // (same shape; the new field loads with its default). Anything higher is from the future. + if p.items.is_some() && p.root.is_some() { + return Ok(match p.version.unwrap_or(1) { + 0..=2 => 1, + v => v, + }); + } + Err(FormatError::NotAProject("no `schema_version` and not a legacy project".into())) +} + +pub fn check_supported(found: u32) -> Result<(), FormatError> { + if found > SCHEMA_VERSION { + return Err(FormatError::TooNew { found, supported: SCHEMA_VERSION }); + } + if found < OLDEST_SCHEMA { + return Err(FormatError::TooOld { found, oldest: OLDEST_SCHEMA }); + } + Ok(()) +} + +/// Run `table` from schema `from` up to `table.len() + 1`, stamping `schema_version` after each step. +pub fn migrate_with(table: &[Migration], mut doc: Value, from: u32) -> Result { + let target = table.len() as u32 + 1; + if from > target { + return Err(FormatError::TooNew { found: from, supported: target }); + } + if from == 0 { + return Err(FormatError::TooOld { found: 0, oldest: 1 }); + } + for v in from..target { + let step = table[(v - 1) as usize]; + doc = step(doc).map_err(|msg| FormatError::Migration { from: v, msg })?; + match doc.as_object_mut() { + Some(o) => { + o.insert("schema_version".into(), Value::from(v + 1)); + } + None => return Err(FormatError::Migration { from: v, msg: "migration produced a non-object document".into() }), + } + } + Ok(doc) +} + +/// v1 → v2: wrap the bare project in the envelope and drop the per-project `version` field. +fn v1_to_v2(doc: Value) -> Result { + let Value::Object(mut project) = doc else { return Err("expected a JSON object".into()) }; + project.remove("version"); + Ok(serde_json::json!({ + "format": FORMAT_ID, + "schema_version": 2, + "generator": "FilmCraft (schema v1)", + "project": Value::Object(project), + })) +} + +/// v2 → v3: graphic clips added; existing data needs no change. +fn v2_to_v3(doc: Value) -> Result { + Ok(doc) +} + +/// v3 → v4: Essential Sound fields added; existing data needs no change. +fn v3_to_v4(doc: Value) -> Result { + Ok(doc) +} + +/// v4 → v5: colour-management fields added; existing data needs no change. +fn v4_to_v5(doc: Value) -> Result { + Ok(doc) +} + +/// v5 → v6: media identity and ingest settings added; existing data needs no change. +fn v5_to_v6(doc: Value) -> Result { + Ok(doc) +} + +/// v6 → v7: effect masks (Bézier mask paths as keyframable parameters); existing data needs no change. +fn v6_to_v7(doc: Value) -> Result { + Ok(doc) +} + +/// v7 → v8: multi-camera source sequences / clips and merged clips added; existing data needs no +/// change. +fn v7_to_v8(doc: Value) -> Result { + Ok(doc) +} + +/// v8 → v9: per-media-item transcripts (`project.transcripts`) added; existing data needs no change. +fn v8_to_v9(doc: Value) -> Result { + Ok(doc) +} + +/// v9 → v10: clip time interpolation, Hold Filters, Field Options and audio source channels; the +/// Modify â–¸ Audio Channels map; subclips' Restrict Trims flag. Existing data needs no change. +fn v9_to_v10(doc: Value) -> Result { + Ok(doc) +} + +/// v10 → v11: search bins (`project.search_bins`), Flash Cue markers, Project Settings safe areas, +/// capture format and scratch disks. Existing data needs no change. +fn v10_to_v11(doc: Value) -> Result { + Ok(doc) +} + +/// v11 → v12: graphics templates, rolls / crawls, responsive design and per-character text styles +/// on graphic clips, and source graphics. Existing data needs no change. +fn v11_to_v12(doc: Value) -> Result { + Ok(doc) +} + +/// v12 → v13: container audio streams and per-clip stream selection. `MediaInfo` reads the old +/// `audio` field as stream zero; `TrackItem::audio_stream` defaults to zero. +fn v12_to_v13(doc: Value) -> Result { + Ok(doc) +} + +/// v13 → v14: narrations (Text to Speech). Existing data needs no change. +fn v13_to_v14(doc: Value) -> Result { + Ok(doc) +} + +/// v14 → v15: project notes. Existing data needs no change. +fn v14_to_v15(doc: Value) -> Result { + Ok(doc) +} + +/// v15 → v16: existing colors remain presets; authors and the named custom palette default empty. +fn v15_to_v16(doc: Value) -> Result { + Ok(doc) +} + +/// v16 → v17: marker URLs default empty. Older builds must not silently drop saved URLs. +fn v16_to_v17(doc: Value) -> Result { + Ok(doc) +} + +#[cfg(test)] +mod tests { + use super::*; + use serde_json::json; + + #[test] + fn legacy_and_current_envelopes_preserve_values() { + let bare = json!({"version": 2, "items": {}, "root": {}, "extra": [1, 2]}); + assert_eq!(detect_version(&bare).unwrap(), 1); + let migrated = migrate_with(MIGRATIONS, bare, 1).unwrap(); + assert_eq!(migrated["schema_version"], 17); + assert_eq!(migrated["project"]["extra"], json!([1, 2])); + assert_eq!(migrated["generator"], "FilmCraft (schema v1)"); + assert_eq!(detect_bytes(serde_json::to_string(&migrated).unwrap().as_bytes()).unwrap(), 17); + let parsed: EnvelopeIn = serde_json::from_value(migrated).unwrap(); + assert!(parsed.view.is_none()); + assert_eq!(parsed.project["extra"], json!([1, 2])); + } + #[test] + fn typed_encode_and_decode_current_and_legacy() { + let project = json!({"items": {}, "root": {}, "name": "Cut"}); + let view = json!({"active": 3}); + for pretty in [false, true] { + let bytes = encode(&project, Some(&view), "FilmCraft test".into(), pretty); + let (file, version): (EnvelopeIn, _) = decode(&bytes).unwrap(); + assert_eq!(version, SCHEMA_VERSION); + assert_eq!(file.project, project); + assert_eq!(file.view, Some(view.clone())); + } + let old = br#"{"version":1,"items":{},"root":{},"name":"Cut"}"#; + let (file, version): (EnvelopeIn, _) = decode(old).unwrap(); + assert_eq!(version, 1); + assert_eq!(file.project["name"], "Cut"); + } + #[test] + fn rejects_wrong_future_and_damaged_envelopes() { + assert!(matches!(detect_bytes(br#"{"format":"other","schema_version":1}"#), Err(FormatError::NotAProject(_)))); + assert!(matches!(detect_bytes(br#"{"schema_version":18}"#), Err(FormatError::TooNew { found: 18, supported: 17 }))); + assert!(matches!(detect_bytes(br#"{"schema_version":0}"#), Err(FormatError::TooOld { found: 0, oldest: 1 }))); + assert!(matches!(detect_bytes(b"{"), Err(FormatError::Corrupt(_)))); + assert!(matches!(migrate_with(MIGRATIONS, json!([]), 1), Err(FormatError::Migration { from: 1, .. }))); + } +} diff --git a/crates/flow/src/from_doc.rs b/crates/flow/src/from_doc.rs index 341ded449..810e67d55 100644 --- a/crates/flow/src/from_doc.rs +++ b/crates/flow/src/from_doc.rs @@ -493,6 +493,8 @@ fn flow_paras_in(doc: &Document, p: &Paragraph, opts: FlowOpts) -> Vec { kind, list, align: p.props.align, + // Direction can come from the paragraph's style (`w:bidi` in the style's pPr). + rtl: doc.styles.resolve_para(&p.props).bidi, page_break: p.props.page_break_before.unwrap_or(false), keep_next: p.props.keep_next.unwrap_or(false), inlines: Vec::new(), @@ -536,6 +538,7 @@ fn flow_paras_in(doc: &Document, p: &Paragraph, opts: FlowOpts) -> Vec { kind, list: None, align: p.props.align, + rtl: out.rtl, page_break: true, keep_next: p.props.keep_next.unwrap_or(false), inlines: Vec::new(), @@ -658,7 +661,10 @@ fn flow_paras_in(doc: &Document, p: &Paragraph, opts: FlowOpts) -> Vec { { (li.list, li.label) = (0, None); } - let old = Para { kind, list, align: c.old.align, page_break: c.old.page_break_before.unwrap_or(false), ..Default::default() }; + // Direction resolved like the current paragraph's: the old properties' own `w:bidi`, else + // the style's. + let rtl = doc.styles.resolve_para(&c.old).bidi; + let old = Para { kind, list, align: c.old.align, rtl, page_break: c.old.page_break_before.unwrap_or(false), ..Default::default() }; Box::new(FmtChange { mark: rev_mark(doc, c.rev), old }) }); } @@ -749,7 +755,7 @@ fn graphic_group( let inline = Inline::Shape { kind: *kind, w: rect[2] * sx, h: rect[3] * sy, float: None, blocks: Vec::new(), look, path: None }; Some(Member { x: rect[0] * sx, y: rect[1] * sy, inline }) } - GraphicItem::Text { rect, text, size, color, bold, align, font } if finite(rect) && !text.trim().is_empty() => { + GraphicItem::Text { rect, text, size, color, bold, align, font, .. } if finite(rect) && !text.trim().is_empty() => { let size = if size.is_finite() { size.clamp(1.0, 400.0) } else { 11.0 }; // One line, centred vertically in its place as the graphic draws it. let line = size * 1.2; diff --git a/crates/flow/src/to_doc.rs b/crates/flow/src/to_doc.rs index c10be85c2..ce29a4448 100644 --- a/crates/flow/src/to_doc.rs +++ b/crates/flow/src/to_doc.rs @@ -159,6 +159,9 @@ fn para_props(p: &Para, list_num: Option) -> ParaProps { out.numbering = Some(NumRef { num, level: li.level.min(8) }); } out.align = p.align; + if p.rtl { + out.bidi = Some(true); + } let sp = p.spacing; out.space_before = sp.before; out.space_after = sp.after; diff --git a/crates/flow/src/types.rs b/crates/flow/src/types.rs index e897778f9..f1c281c62 100644 --- a/crates/flow/src/types.rs +++ b/crates/flow/src/types.rs @@ -676,6 +676,8 @@ pub struct Para { pub kind: Kind, pub list: Option, pub align: Option, + /// The paragraph reads right to left (OOXML `w:bidi`, RTF `\rtlpar`). + pub rtl: bool, pub page_break: bool, /// Keep the paragraph on the same page as the next one (ODT `fo:keep-with-next`). pub keep_next: bool, diff --git a/crates/font-catalog/Cargo.toml b/crates/font-catalog/Cargo.toml index a95e1c28f..1f0830a63 100644 --- a/crates/font-catalog/Cargo.toml +++ b/crates/font-catalog/Cargo.toml @@ -15,6 +15,7 @@ default = ["system"] system = ["dep:fontique"] [dependencies] +kurbo = { workspace = true } skrifa = { version = "0.47", default-features = false, features = ["std"] } thiserror = { workspace = true } 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..3f155d35d 100644 --- a/crates/font-catalog/src/lib.rs +++ b/crates/font-catalog/src/lib.rs @@ -21,16 +21,22 @@ //! ``` mod catalog; +mod embedding; mod env; mod fontconfig; +mod outline; 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 outline::{FontOutline, GlyphPen, OutlineError}; pub use scan::{FaceInfo, LocalizedName, Source, is_font_file, style_italic, style_weight, style_width}; #[cfg(test)] diff --git a/crates/font-catalog/src/outline.rs b/crates/font-catalog/src/outline.rs new file mode 100644 index 000000000..f48b61e4f --- /dev/null +++ b/crates/font-catalog/src/outline.rs @@ -0,0 +1,399 @@ +//! Unhinted glyph outlines and metrics for catalog faces. + +use kurbo::BezPath; +use skrifa::instance::{Location, Size}; +use skrifa::outline::{DrawSettings, OutlinePen}; +use skrifa::{FontRef, GlyphId, MetadataProvider, Tag}; + +use crate::{FontData, LoadedFace}; + +/// A parsed face and its variation location. Coordinates use font design units, with y up. +/// The bytes can come from a catalog or any other source, including a collection face. +pub struct FontOutline<'a> { + font: FontRef<'a>, + location: Location, +} + +/// Receives unhinted glyph commands in design units (y up). Implementations own their +/// storage and resource policy; drawing writes directly to the sink without an intermediate path. +pub trait GlyphPen { + /// Start a contour. + fn move_to(&mut self, x: f32, y: f32); + /// Add a straight segment. + fn line_to(&mut self, x: f32, y: f32); + /// Add a quadratic segment. + fn quad_to(&mut self, x1: f32, y1: f32, x: f32, y: f32); + /// Add a cubic segment. + fn curve_to(&mut self, x1: f32, y1: f32, x2: f32, y2: f32, x: f32, y: f32); + /// Close the current contour. + fn close(&mut self); +} + +struct Sink<'a>(&'a mut dyn GlyphPen); +impl OutlinePen for Sink<'_> { + fn move_to(&mut self, x: f32, y: f32) { + self.0.move_to(x, y); + } + fn line_to(&mut self, x: f32, y: f32) { + self.0.line_to(x, y); + } + fn quad_to(&mut self, x1: f32, y1: f32, x: f32, y: f32) { + self.0.quad_to(x1, y1, x, y); + } + fn curve_to(&mut self, x1: f32, y1: f32, x2: f32, y2: f32, x: f32, y: f32) { + self.0.curve_to(x1, y1, x2, y2, x, y); + } + fn close(&mut self) { + self.0.close(); + } +} + +/// Why a font or glyph outline could not be read. +#[derive(Debug, thiserror::Error)] +pub enum OutlineError { + /// The face index or font data is invalid. + #[error("invalid font face")] + InvalidFace, + /// The glyph has no outline in this face. + #[error("glyph has no outline")] + MissingGlyph, + /// The glyph's outline is malformed. + #[error("cannot draw glyph outline: {0}")] + Draw(#[from] skrifa::outline::DrawError), + /// The glyph emitted more path elements than the safety limit. + #[error("glyph outline exceeds the element limit")] + TooComplex, + /// The outline starts without a move command. + #[error("glyph outline has no initial move")] + InvalidPath, +} + +impl<'a> FontOutline<'a> { + /// Parse the given face and select axis values in user units. Unknown axes are ignored and + /// values are clamped to their axis ranges by the font parser. + pub fn new(data: &'a FontData, variations: &[([u8; 4], f32)]) -> Result { + Self::from_bytes(&data.bytes, data.index, variations) + } + + /// Parse borrowed bytes and a collection face index without copying the font. + pub fn from_bytes(bytes: &'a [u8], index: u32, variations: &[([u8; 4], f32)]) -> Result { + let font = FontRef::from_index(bytes, index).map_err(|_| OutlineError::InvalidFace)?; + let location = font.axes().location(variations.iter().map(|(tag, value)| (Tag::new(tag), *value))); + Ok(Self { font, location }) + } + + /// Parse a catalog face, retaining the axis settings of a named instance. + pub fn from_loaded(face: &'a LoadedFace) -> Result { + Self::new(&face.data, &face.info.variations) + } + + /// Font-wide metrics in unscaled design units at this variation location. + pub fn metrics(&self) -> skrifa::metrics::Metrics { + self.font.metrics(Size::unscaled(), &self.location) + } + + /// Resolve a Unicode scalar to a glyph identifier. + pub fn glyph_for(&self, character: char) -> Option { + self.font.charmap().map(character).map(|id| id.to_u32()) + } + + /// Stream an unhinted glyph into the caller's sink. On an error, commands already sent + /// remain in the sink. The sink owns any allocation and enforces its own resource limit. + pub fn draw_to(&self, glyph: u32, pen: &mut dyn GlyphPen) -> Result<(), OutlineError> { + let outline = self.font.outline_glyphs().get(GlyphId::new(glyph)).ok_or(OutlineError::MissingGlyph)?; + outline.draw(DrawSettings::unhinted(Size::unscaled(), &self.location), &mut Sink(pen))?; + Ok(()) + } + + /// Draw a glyph in design units. Empty glyphs such as spaces can produce an empty path. + /// A malformed or excessively complex glyph returns an error instead of partial geometry. + pub fn outline(&self, glyph: u32) -> Result { + let (path, status) = self.outline_recovering(glyph)?; + status.map(|()| path) + } + + /// Draw a glyph while retaining the bounded prefix produced before a draw error. This is + /// useful for consumers preserving existing rendering of damaged fonts. The inner result + /// reports whether the returned path is complete; the outer error means no glyph was found. + /// Paths stop growing at 100,000 elements, even when the font is untrusted. + pub fn outline_recovering(&self, glyph: u32) -> Result<(BezPath, Result<(), OutlineError>), OutlineError> { + let outline = self.font.outline_glyphs().get(GlyphId::new(glyph)).ok_or(OutlineError::MissingGlyph)?; + let mut pen = PathPen::new(); + let result = outline.draw(DrawSettings::unhinted(Size::unscaled(), &self.location), &mut pen); + let status = if pen.invalid { + Err(OutlineError::InvalidPath) + } else if pen.truncated { + Err(OutlineError::TooComplex) + } else { + result.map(|_| ()).map_err(OutlineError::Draw) + }; + Ok((pen.path, status)) + } +} + +const MAX_OUTLINE_ELEMENTS: usize = 100_000; +struct PathPen { + path: BezPath, + truncated: bool, + invalid: bool, +} +impl PathPen { + fn new() -> Self { + Self { path: BezPath::new(), truncated: false, invalid: false } + } + fn has_start(&mut self) -> bool { + if self.path.elements().is_empty() { + self.invalid = true; + false + } else { + true + } + } + fn room(&mut self) -> bool { + if self.path.elements().len() >= MAX_OUTLINE_ELEMENTS { + self.truncated = true; + false + } else { + true + } + } +} +impl OutlinePen for PathPen { + fn move_to(&mut self, x: f32, y: f32) { + if self.room() { + self.path.move_to((f64::from(x), f64::from(y))); + } + } + fn line_to(&mut self, x: f32, y: f32) { + if self.has_start() && self.room() { + self.path.line_to((f64::from(x), f64::from(y))); + } + } + fn quad_to(&mut self, x1: f32, y1: f32, x: f32, y: f32) { + if self.has_start() && self.room() { + self.path.quad_to((f64::from(x1), f64::from(y1)), (f64::from(x), f64::from(y))); + } + } + fn curve_to(&mut self, x1: f32, y1: f32, x2: f32, y2: f32, x: f32, y: f32) { + if self.has_start() && self.room() { + self.path.curve_to((f64::from(x1), f64::from(y1)), (f64::from(x2), f64::from(y2)), (f64::from(x), f64::from(y))); + } + } + fn close(&mut self) { + if self.has_start() && self.room() { + self.path.close_path(); + } + } +} + +#[cfg(test)] +mod tests { + use super::*; + use kurbo::PathEl; + use std::sync::Arc; + + impl GlyphPen for PathPen { + fn move_to(&mut self, x: f32, y: f32) { + OutlinePen::move_to(self, x, y); + } + fn line_to(&mut self, x: f32, y: f32) { + OutlinePen::line_to(self, x, y); + } + fn quad_to(&mut self, x1: f32, y1: f32, x: f32, y: f32) { + OutlinePen::quad_to(self, x1, y1, x, y); + } + fn curve_to(&mut self, x1: f32, y1: f32, x2: f32, y2: f32, x: f32, y: f32) { + OutlinePen::curve_to(self, x1, y1, x2, y2, x, y); + } + fn close(&mut self) { + OutlinePen::close(self); + } + } + + #[test] + fn rejects_invalid_face_and_collection_index() { + let data = FontData { bytes: Arc::from(&b"not a font"[..]), index: 0 }; + assert!(matches!(FontOutline::new(&data, &[]), Err(OutlineError::InvalidFace))); + } + + #[test] + fn pen_preserves_geometry_and_upward_y() { + let mut pen = PathPen::new(); + OutlinePen::move_to(&mut pen, 1.0, 2.0); + OutlinePen::line_to(&mut pen, 3.0, 4.0); + OutlinePen::quad_to(&mut pen, 5.0, 6.0, 7.0, 8.0); + OutlinePen::curve_to(&mut pen, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0); + OutlinePen::close(&mut pen); + assert_eq!( + pen.path.elements(), + &[ + PathEl::MoveTo((1.0, 2.0).into()), + PathEl::LineTo((3.0, 4.0).into()), + PathEl::QuadTo((5.0, 6.0).into(), (7.0, 8.0).into()), + PathEl::CurveTo((9.0, 10.0).into(), (11.0, 12.0).into(), (13.0, 14.0).into()), + PathEl::ClosePath, + ] + ); + } + + #[test] + fn streaming_adapter_forwards_each_command_and_rejects_an_invalid_prefix() { + let mut pen = PathPen::new(); + { + let mut sink = Sink(&mut pen); + OutlinePen::move_to(&mut sink, 1.0, 2.0); + OutlinePen::line_to(&mut sink, 3.0, 4.0); + OutlinePen::quad_to(&mut sink, 5.0, 6.0, 7.0, 8.0); + OutlinePen::curve_to(&mut sink, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0); + OutlinePen::close(&mut sink); + } + assert_eq!(pen.path.elements().len(), 5); + assert_eq!(pen.path.elements()[2], PathEl::QuadTo((5.0, 6.0).into(), (7.0, 8.0).into())); + let mut invalid = PathPen::new(); + OutlinePen::line_to(&mut invalid, 1.0, 2.0); + assert!(invalid.invalid); + assert!(invalid.path.elements().is_empty()); + } + + #[test] + fn pen_caps_path_growth() { + let mut pen = PathPen::new(); + OutlinePen::move_to(&mut pen, 0.0, 0.0); + for _ in 0..MAX_OUTLINE_ELEMENTS { + OutlinePen::line_to(&mut pen, 1.0, 2.0); + } + assert_eq!(pen.path.elements().len(), MAX_OUTLINE_ELEMENTS); + assert!(pen.truncated); + } + + #[test] + fn reads_craft_font_when_available() { + let Ok(dir) = std::env::var("CRAFT_FONTS_DIR") else { return }; + let bytes = std::fs::read(std::path::Path::new(&dir).join("fonts/noto-sans-hebrew/NotoSansHebrew-Regular.ttf")).unwrap(); + let data = FontData { bytes: bytes.into(), index: 0 }; + let face = FontOutline::new(&data, &[]).unwrap(); + assert!(face.metrics().units_per_em > 0); + let glyph = face.glyph_for('×�').unwrap(); + assert!(!face.outline(glyph).unwrap().elements().is_empty()); + assert!(matches!(face.outline(u32::MAX), Err(OutlineError::MissingGlyph))); + let invalid_index = FontData { bytes: data.bytes.clone(), index: u32::MAX }; + assert!(matches!(FontOutline::new(&invalid_index, &[]), Err(OutlineError::InvalidFace))); + } + + #[test] + fn variation_changes_glyph_geometry() { + let Ok(dir) = std::env::var("CRAFT_FONTS_DIR") else { return }; + let bytes = std::fs::read(std::path::Path::new(&dir).join("fonts/noto-sans-arabic/NotoSansArabic.ttf")).unwrap(); + let data = FontData { bytes: bytes.into(), index: 0 }; + let light = FontOutline::new(&data, &[(*b"wght", 100.0)]).unwrap(); + let bold = FontOutline::new(&data, &[(*b"wght", 900.0)]).unwrap(); + let glyph = light.glyph_for('ب').unwrap(); + let light_path = light.outline(glyph).unwrap(); + let bold_path = bold.outline(glyph).unwrap(); + assert!(!light_path.elements().is_empty()); + assert_ne!(light_path, bold_path); + } + + // Construct a minimal CFF font so CI can exercise the public API without installed fonts. + fn cff_font(program: &[u8]) -> Vec { + let mut cff = vec![1, 0, 4, 4, 0, 1, 1, 1, 2, b'X', 0, 1, 1, 1, 7, 29, 0, 0, 0, 25, 17, 0, 0, 0, 0]; + cff.extend_from_slice(&[0, 1, 4]); + cff.extend_from_slice(&1u32.to_be_bytes()); + cff.extend_from_slice(&((program.len() as u32) + 1).to_be_bytes()); + cff.extend_from_slice(program); + let mut builder = write_fonts::FontBuilder::new(); + let mut head = vec![0; 54]; + head[18..20].copy_from_slice(&1000u16.to_be_bytes()); + let mut hhea = vec![0; 36]; + hhea[34..36].copy_from_slice(&1u16.to_be_bytes()); + builder.add_raw(write_fonts::types::Tag::new(b"head"), head); + builder.add_raw(write_fonts::types::Tag::new(b"hhea"), hhea); + builder.add_raw(write_fonts::types::Tag::new(b"hmtx"), vec![0, 100, 0, 0]); + builder.add_raw(write_fonts::types::Tag::new(b"maxp"), vec![0, 0, 0x50, 0, 0, 1]); + builder.add_raw(write_fonts::types::Tag::new(b"CFF "), cff); + builder.build() + } + + #[test] + fn malformed_cff_recovers_the_same_prefix_as_direct_drawing() { + // Move and line, then callsubr without an operand: a prefix is emitted before failure. + let bytes = cff_font(&[139, 139, 21, 239, 139, 5, 10, 14]); + let gid = 0; + let f = FontRef::new(&bytes).unwrap(); + let mut old = PathPen::new(); + let result = + f.outline_glyphs().get(GlyphId::new(gid)).unwrap().draw(DrawSettings::unhinted(Size::unscaled(), &Location::default()), &mut old); + assert!(result.is_err()); + assert!(!old.path.elements().is_empty()); + let shared = FontOutline::from_bytes(&bytes, 0, &[]).unwrap(); + assert_eq!(shared.metrics().units_per_em, 1000); + assert_eq!(shared.glyph_for('A'), None); + let (prefix, status) = shared.outline_recovering(gid).unwrap(); + assert!(matches!(status, Err(OutlineError::Draw(_)))); + assert_eq!(prefix, old.path); + let mut streamed = PathPen::new(); + assert!(matches!(shared.draw_to(gid, &mut streamed), Err(OutlineError::Draw(_)))); + assert_eq!(streamed.path, old.path); + assert!(matches!(shared.outline(gid), Err(OutlineError::Draw(_)))); + } + + #[test] + fn streaming_preserves_a_large_valid_cff_outline() { + // Minimal CFF with more than 100,000 line segments. + let mut program = vec![139, 139, 21]; + for _ in 0..100_001 { + program.extend_from_slice(&[140, 6]); + } + program.push(14); + let bytes = cff_font(&program); + let shared = FontOutline::from_bytes(&bytes, 0, &[]).unwrap(); + struct Count(usize); + impl GlyphPen for Count { + fn move_to(&mut self, _: f32, _: f32) { + self.0 += 1; + } + fn line_to(&mut self, _: f32, _: f32) { + self.0 += 1; + } + fn quad_to(&mut self, _: f32, _: f32, _: f32, _: f32) { + self.0 += 1; + } + fn curve_to(&mut self, _: f32, _: f32, _: f32, _: f32, _: f32, _: f32) { + self.0 += 1; + } + fn close(&mut self) { + self.0 += 1; + } + } + impl OutlinePen for Count { + fn move_to(&mut self, x: f32, y: f32) { + GlyphPen::move_to(self, x, y); + } + fn line_to(&mut self, x: f32, y: f32) { + GlyphPen::line_to(self, x, y); + } + fn quad_to(&mut self, x1: f32, y1: f32, x: f32, y: f32) { + GlyphPen::quad_to(self, x1, y1, x, y); + } + fn curve_to(&mut self, x1: f32, y1: f32, x2: f32, y2: f32, x: f32, y: f32) { + GlyphPen::curve_to(self, x1, y1, x2, y2, x, y); + } + fn close(&mut self) { + GlyphPen::close(self); + } + } + let mut old = Count(0); + FontRef::new(&bytes) + .unwrap() + .outline_glyphs() + .get(GlyphId::new(0)) + .unwrap() + .draw(DrawSettings::unhinted(Size::unscaled(), &Location::default()), &mut old) + .unwrap(); + let mut count = Count(0); + shared.draw_to(0, &mut count).unwrap(); + assert_eq!(count.0, old.0); + assert!(count.0 > MAX_OUTLINE_ELEMENTS); + assert!(matches!(shared.outline(0), Err(OutlineError::TooComplex))); + } +} 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/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/lib.rs b/crates/image-codecs/src/lib.rs index 78df75333..5387a673f 100644 --- a/crates/image-codecs/src/lib.rs +++ b/crates/image-codecs/src/lib.rs @@ -33,8 +33,9 @@ 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}; @@ -110,7 +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) } @@ -167,10 +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), @@ -199,11 +205,12 @@ pub fn decode_png_rows>( } /// The upright pixel size [`decode`] would return, read from the file's header without decoding -/// pixels, for PNG, JPEG, TIFF (the page [`decode`] opens), WebP, OpenEXR (flat) 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 orientation of 5–8 (EXIF, or the TIFF page's -/// own tag), so wherever [`decode`] succeeds the two agree. Other formats (and deep EXR or AVIF) -/// return [`CodecError::Unsupported`]: decode those to learn their size. +/// pixels, for PNG, JPEG, TIFF (the page [`decode`] opens), 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 orientation of 5–8 (EXIF, +/// the TIFF page's own tag, or AVIF's container `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 { @@ -219,7 +226,11 @@ pub fn probe_size(bytes: &[u8]) -> Result<(u32, u32), CodecError> { } 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 6262976bb..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` diff --git a/crates/media/src/mpegts/tests.rs b/crates/media/src/mpegts/tests.rs index 0215680ad..b9cb1af3b 100644 --- a/crates/media/src/mpegts/tests.rs +++ b/crates/media/src/mpegts/tests.rs @@ -791,6 +791,23 @@ fn scrambled_packet_is_skipped_with_a_warning() { assert!(total(&f) < total(&base)); } +#[test] +fn repeated_pat_and_pmt_packets_are_dropped_before_the_section_parser() { + let clean = audio_ts(2); + let first = |pid: u16| { + (0..clean.len() / 188).map(|i| i * 188).find(|&o| (((clean[o + 1] & 0x1F) as u16) << 8 | clean[o + 2] as u16) == pid).expect("packet exists") + }; + let mut dup = clean.clone(); + for pid in [0x100, 0] { + let o = first(pid); + let copy = dup[o..o + 188].to_vec(); + dup.splice(o + 188..o + 188, copy); + } + let f = open(&dup).unwrap(); + assert!(f.warnings.iter().any(|w| w.contains("2 duplicate packet(s) dropped")), "{:?}", f.warnings); + assert!(!f.warnings.iter().any(|w| w.contains("continuity") || w.contains("CRC")), "{:?}", f.warnings); +} + #[test] fn duplicate_packet_is_dropped_but_a_different_payload_with_the_same_counter_is_kept() { let clean = audio_ts(6); diff --git a/crates/media/src/mpegts/ts.rs b/crates/media/src/mpegts/ts.rs index a4c01a986..1d28bb736 100644 --- a/crates/media/src/mpegts/ts.rs +++ b/crates/media/src/mpegts/ts.rs @@ -497,6 +497,7 @@ pub(crate) fn scan(src: &S, packet_size: usize, sync_off let mut warnings: Vec = Vec::new(); let mut pat = Psi::default(); let mut pmt = PmtState::default(); + let mut psi_dup: HashMap = HashMap::new(); let mut es: HashMap = HashMap::new(); let mut order: Vec = Vec::new(); // one timestamp unwrapper for the program: its streams share one clock, so a stream whose @@ -560,6 +561,13 @@ pub(crate) fn scan(src: &S, packet_size: usize, sync_off } continue; } + if pk.pid == 0 || Some(pk.pid) == pmt.pmt_pid { + // a repeated packet would append its payload to the section in progress + if psi_dup.entry(pk.pid).or_default().is_dup(&pk) { + duplicates += 1; + continue; + } + } if pk.pid == 0 { feed_pat(&pk, &mut pat, &mut pmt, &mut crc_errors); continue; 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/odf/src/text/chart.rs b/crates/odf/src/text/chart.rs index 9d073ecd2..2a79352a1 100644 --- a/crates/odf/src/text/chart.rs +++ b/crates/odf/src/text/chart.rs @@ -229,7 +229,7 @@ impl Reader { }; let name = s.label.as_deref().and_then(|a| cells_of(a).first().copied()).map(|at| label(self.cell(at))).unwrap_or_default(); let name = if name.is_empty() { format!("Series {}", i + 1) } else { name }; - ChartSeries { name, values } + ChartSeries { name, values, ..Default::default() } }) .collect(); // A scatter chart's x values: its first series' domain. @@ -248,7 +248,10 @@ impl Reader { _ => LegendPos::Right, }); let title = self.title.as_deref().map(str::trim).filter(|t| !t.is_empty()).map(str::to_string); - Some(ChartSpec { kind, categories, series, title, legend, data_labels }.sanitized()) + // ODF axis formatting is not read; the value axis keeps the look every chart had before + // axes were formattable (gridlines, no axis line), a round chart has none. + let val_axis = if kind.is_round() { Default::default() } else { craft_wordmodel::chart::value_axis() }; + Some(ChartSpec { kind, categories, series, title, legend, data_labels, val_axis, ..Default::default() }.sanitized()) } } diff --git a/crates/odf/src/text/tests_more/charts.rs b/crates/odf/src/text/tests_more/charts.rs index 9ac0b6b96..5aa892cf4 100644 --- a/crates/odf/src/text/tests_more/charts.rs +++ b/crates/odf/src/text/tests_more/charts.rs @@ -57,8 +57,8 @@ fn spec(chart: &str) -> Option { fn north_south() -> Vec { vec![ - ChartSeries { name: "North".into(), values: vec![Some(10.0), Some(14.0), Some(8.0), Some(16.0)] }, - ChartSeries { name: "South".into(), values: vec![Some(12.0), None, Some(11.0), Some(13.0)] }, + ChartSeries { name: "North".into(), values: vec![Some(10.0), Some(14.0), Some(8.0), Some(16.0)], ..Default::default() }, + ChartSeries { name: "South".into(), values: vec![Some(12.0), None, Some(11.0), Some(13.0)], ..Default::default() }, ] } @@ -78,6 +78,8 @@ fn a_column_chart_reads_its_data_title_and_legend() { title: Some("Sales by quarter".into()), legend: Some(LegendPos::Right), data_labels: false, + val_axis: craft_wordmodel::chart::value_axis(), + ..Default::default() } ); // The frame's size, in points. @@ -135,7 +137,7 @@ fn line_area_pie_doughnut_and_scatter_classes() { let s = spec(&chart_xml("chart:scatter", "", "", "", scatter)).unwrap(); assert_eq!(s.kind, ChartType::Scatter); assert_eq!(s.categories, vec!["10", "14", "8", "16"]); - assert_eq!(s.series, vec![ChartSeries { name: "South".into(), values: vec![Some(12.0), None, Some(11.0), Some(13.0)] }]); + assert_eq!(s.series, vec![ChartSeries { name: "South".into(), values: vec![Some(12.0), None, Some(11.0), Some(13.0)], ..Default::default() }]); } #[test] @@ -249,7 +251,15 @@ fn every_chart_type_is_saved_as_a_chart_object_and_reads_back() { let bytes = crate::text::export(&doc_with(spec.clone(), false)).unwrap(); let back = reread(&bytes); assert_eq!(back.len(), 1, "{kind:?}"); - assert_eq!(*back[0].spec, spec.clone().sanitized(), "{kind:?}"); + // ODF keeps no series or axis formatting here: only the data, the title and the legend + // come back; the value axis has the gridlines every chart had before axes were formattable. + let mut expected = spec.clone().sanitized(); + for s in &mut expected.series { + *s = ChartSeries { name: std::mem::take(&mut s.name), values: std::mem::take(&mut s.values), ..Default::default() }; + } + expected.cat_axis = Default::default(); + expected.val_axis = if kind.is_round() { Default::default() } else { craft_wordmodel::chart::value_axis() }; + assert_eq!(*back[0].spec, expected, "{kind:?}"); assert_eq!((back[0].w, back[0].h, back[0].alt.as_str()), (300.0, 180.0, "Sales chart"), "{kind:?}"); // The package lists the chart object. let z = crate::Package::open(&bytes, &crate::Limits::DEFAULT).unwrap(); diff --git a/crates/odf/src/text/tests_more/groups.rs b/crates/odf/src/text/tests_more/groups.rs index bf1b056a6..4fada68a9 100644 --- a/crates/odf/src/text/tests_more/groups.rs +++ b/crates/odf/src/text/tests_more/groups.rs @@ -130,6 +130,7 @@ fn a_diagram_is_saved_as_a_group_of_its_drawing() { bold: false, align: TextAlign::Center, font: None, + angle: 0.0, }, GraphicItem::Shape { rect: [90.0, 9.0, 18.0, 18.0], kind: ShapeKind::Arrow, fill: Some(blue), stroke: None, stroke_width: 0.0 }, ], 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/lib.rs b/crates/psd/src/lib.rs index 27797f1bc..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; @@ -102,6 +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/docs/cr3.md b/crates/raw/docs/cr3.md index 40e375b54..9c17c809c 100644 --- a/crates/raw/docs/cr3.md +++ b/crates/raw/docs/cr3.md @@ -187,7 +187,7 @@ the six verified fixtures are unchanged. Partial-subband and rounded-plane coding, vertical C-RAW tiling and adaptive low-frequency gains remain unsupported (see the variant table above; `cr3_variants.rs` pins each rejection). The decoder bounds image dimensions, sample count, tile count, marker headers and temporary allocations. The full sensor buffer is -allocated only after the first plane decodes, and plane buffers grow row by row, +allocated only after every plane has decoded, and plane buffers grow row by row, so a small crafted file fails before reserving memory for the image it claims. ## Further bodies: one CC0 file each diff --git a/crates/raw/src/crx.rs b/crates/raw/src/crx.rs index a8382f757..4d8f53ba8 100644 --- a/crates/raw/src/crx.rs +++ b/crates/raw/src/crx.rs @@ -95,60 +95,90 @@ pub(crate) fn decode(sample: &[u8], config: &Cr3Compression) -> Result> let count = image_size(config)?; let tiles = layout(sample, config)?; validate_planes(&tiles, config)?; - // Allocated once the first plane decodes: a tiny crafted file can claim a huge image. - let mut output = Vec::new(); + // A tile's QP map error is reported by its first plane, so errors keep tile, then plane order. + let qp: Vec>> = tiles.iter().map(|tile| tile_qp(sample, tile)).collect(); let midpoint = 1i32 << (config.bit_depth - 1); let maximum = (1i32 << config.bit_depth) - 1; - for tile in tiles { - let qp = tile_qp(sample, &tile)?; - // Each plane is its own entropy stream: decode a tile's planes in parallel. - let planes: Vec>> = tile - .planes - .par_iter() - .enumerate() - .map(|(p, plane)| { - if config.levels == 0 { - let band = plane.bands.first().ok_or_else(|| corrupt("missing lossless band"))?; - lossless_plane(band_bytes(sample, band)?, tile.width / 2, tile.height / 2) - } else { - let edges = TileEdges { left: tile.x > 0, right: tile.x + tile.width < width, top: false, bottom: false }; - wavelet_plane(sample, plane, tile.width / 2, tile.height / 2, edges, qp.as_ref()) - } - .map_err(|error| RawError::Corrupt(format!("Canon CRX tile ({}, {}), plane {p}: {error}", tile.x, tile.y))) - }) - .collect(); - for (p, decoded) in planes.into_iter().enumerate() { - let decoded = decoded?; - if output.is_empty() { - output.try_reserve_exact(count).map_err(|_| RawError::Limit("CRX sample allocation"))?; - output.resize(count, 0); - } - for (i, &value) in decoded.iter().enumerate() { - let sensor_value = sensor_sample(value, midpoint, maximum, config.levels != 0)?; - let (x, y) = (tile.x + 2 * (i % (tile.width / 2)) + p % 2, tile.y + 2 * (i / (tile.width / 2)) + p / 2); - let slot = y - .checked_mul(width) - .and_then(|n| n.checked_add(x)) - .and_then(|n| output.get_mut(n)) - .ok_or_else(|| corrupt("tile writes outside image"))?; - *slot = sensor_value; - } - } - } + // Every plane of every tile is its own entropy stream: decode them all in parallel, each straight + // to sensor samples, so a task holds only its own plane as i32. The first error in tile and plane + // order is reported, as when the tiles were decoded one after another. + let jobs: Vec<(&Tile, &Result>, usize, &Plane)> = + tiles.iter().zip(&qp).flat_map(|(tile, qp)| tile.planes.iter().enumerate().map(move |(p, plane)| (tile, qp, p, plane))).collect(); + let planes: Vec>> = jobs + .par_iter() + .map(|&(tile, qp, p, plane)| { + let qp = qp.as_ref().map_err(Clone::clone)?.as_ref(); + let decoded = decode_plane(sample, config, width, tile, plane, qp) + .map_err(|error| RawError::Corrupt(format!("Canon CRX tile ({}, {}), plane {p}: {error}", tile.x, tile.y)))?; + sensor_plane(&decoded, midpoint, maximum, config.levels != 0) + }) + .collect(); + let planes = planes.into_iter().collect::>>()?; + // Allocated once every plane has decoded: a tiny crafted file can claim a huge image. + let mut output = Vec::new(); + output.try_reserve_exact(count).map_err(|_| RawError::Limit("CRX sample allocation"))?; + output.resize(count, 0); + interleave(&mut output, width, &tiles, &planes)?; Ok(output) } -/// A decoded plane value as a sensor sample of the CMP1 bit depth. +fn decode_plane(sample: &[u8], config: &Cr3Compression, width: usize, tile: &Tile, plane: &Plane, qp: Option<&QpMap>) -> Result> { + if config.levels == 0 { + let band = plane.bands.first().ok_or_else(|| corrupt("missing lossless band"))?; + lossless_plane(band_bytes(sample, band)?, tile.width / 2, tile.height / 2) + } else { + let edges = TileEdges { left: tile.x > 0, right: tile.x + tile.width < width, top: false, bottom: false }; + wavelet_plane(sample, plane, tile.width / 2, tile.height / 2, edges, qp) + } +} + +/// Decoded plane values as sensor samples of the CMP1 bit depth. /// /// Lossless planes reproduce the sensor exactly, so a value outside the bit depth is corrupt data. /// Wavelet (C-RAW) planes are quantized: next to clipped highlights the reconstruction overshoots /// the sensor's range by up to a few quantization steps (measured on 17 CC0 files: at most 36 codes /// above 16383 at 14 bits, never below 0), while lossless files of the same bodies clip at exactly -/// 16383. Those samples are clamped to the range the sensor can record. -fn sensor_sample(value: i32, midpoint: i32, maximum: i32, quantized: bool) -> Result { - let sample = value.checked_add(midpoint).ok_or_else(|| corrupt("decoded sample outside its bit depth"))?; - let sample = if quantized { sample.clamp(0, maximum) } else { sample }; - u16::try_from(sample).ok().filter(|_| sample <= maximum).ok_or_else(|| corrupt("decoded sample outside its bit depth")) +/// 16383. Those samples are clamped to the range the sensor can record. The range is checked once, +/// on the plane's extremes; `midpoint` is positive, so only the largest sum can overflow. +fn sensor_plane(values: &[i32], midpoint: i32, maximum: i32, quantized: bool) -> Result> { + let outside = || corrupt("decoded sample outside its bit depth"); + let (Some(&low), Some(&high)) = (values.iter().min(), values.iter().max()) else { return Ok(Vec::new()) }; + let high = high.checked_add(midpoint).ok_or_else(outside)?; + if !quantized && (low.saturating_add(midpoint) < 0 || high > maximum) { + return Err(outside()); + } + let mut samples = Vec::new(); + samples.try_reserve_exact(values.len()).map_err(|_| RawError::Limit("CRX sample allocation"))?; + // Every sum fits (the largest did), and the clamp leaves in-range lossless samples unchanged. + samples.extend(values.iter().map(|&value| (value + midpoint).clamp(0, maximum) as u16)); + Ok(samples) +} + +/// Writes each tile's four planes (`planes`: four per tile, in tile order) into the Bayer mosaic. +/// Plane `p` holds the samples at row parity `p / 2` and column parity `p % 2`. Runs in parallel +/// over pairs of output rows. +fn interleave(output: &mut [u16], width: usize, tiles: &[Tile], planes: &[Vec]) -> Result<()> { + let outside = || corrupt("tile writes outside image"); + if width == 0 || planes.len() != tiles.len() * 4 { + return Err(outside()); + } + output.par_chunks_mut(2 * width).enumerate().try_for_each(|(pair, rows)| -> Result<()> { + let y = 2 * pair; + for (tile, planes) in tiles.iter().zip(planes.as_chunks::<4>().0).filter(|(tile, _)| (tile.y..tile.y + tile.height).contains(&y)) { + let half = tile.width / 2; + let start = (y - tile.y) / 2 * half; + for (row, [even, odd]) in rows.chunks_exact_mut(width).zip(planes.as_chunks::<2>().0) { + let target = row.get_mut(tile.x..tile.x + tile.width).ok_or_else(outside)?; + let even = even.get(start..start + half).ok_or_else(outside)?; + let odd = odd.get(start..start + half).ok_or_else(outside)?; + for ([a, b], (&left, &right)) in target.as_chunks_mut::<2>().0.iter_mut().zip(even.iter().zip(odd)) { + *a = left; + *b = right; + } + } + } + Ok(()) + }) } /// Header-only probing must reject the same unsupported coding variants as full decoding. @@ -413,24 +443,58 @@ fn wavelet_plane(sample: &[u8], plane: &Plane, width: usize, height: usize, edge for high in high_dimensions.iter().rev() { shapes.extend_from_slice(high); } - let mut bands = Vec::with_capacity(10); - for (i, (&(w, h), band)) in shapes.iter().zip(&plane.bands).enumerate() { - let bytes = band_bytes(sample, band)?; - let mut data = if i == 0 { lossless_plane(bytes, w, h)? } else { coded_high_frequency_band(bytes, w, coded_band_width(i, w, edges), h)? }; - for (position, value) in data.iter_mut().enumerate() { - let step = match band.quant { - Quantization::Uniform(value) => quant_step(value, MAX_UNIFORM_QUANT)?, - Quantization::Adaptive { base, gain } => { - adaptive_step(qp.ok_or_else(|| corrupt("missing QP map"))?, i, position % w, position / w, base, gain)? - } - }; - *value = value.checked_mul(step).ok_or_else(|| corrupt("dequantized coefficient overflow"))?; - } - bands.push(WaveletBand { width: w, height: h, data }); - } + // Each subband is its own entropy stream: decode and dequantize them in parallel. The first error + // in band order is reported, as when they were decoded one after another. + let bands: Vec> = shapes + .par_iter() + .zip(plane.bands.par_iter()) + .enumerate() + .map(|(i, (&(w, h), band))| { + let bytes = band_bytes(sample, band)?; + let mut data = if i == 0 { lossless_plane(bytes, w, h)? } else { coded_high_frequency_band(bytes, w, coded_band_width(i, w, edges), h)? }; + dequantize(&mut data, i, w, &band.quant, qp)?; + Ok(WaveletBand { width: w, height: h, data }) + }) + .collect(); + let bands = bands.into_iter().collect::>>()?; crx_wavelet::reconstruct(&bands, width, height, 3, edges) } +/// Multiplies the coefficients of subband `band` (`width` columns) by their quantization step. +/// +/// The step is constant per subband, except for adaptive quantization with a gain, where it is +/// constant per QP cell: 2 coefficients wide in bands 4-6 and 4 wide in bands 7-9 (see +/// [`adaptive_step`]), so it is looked up once per cell rather than per coefficient. +fn dequantize(data: &mut [i32], band: usize, width: usize, quant: &Quantization, qp: Option<&QpMap>) -> Result<()> { + let (map, base, gain) = match *quant { + Quantization::Uniform(value) => return scale(data, quant_step(value, MAX_UNIFORM_QUANT)?), + Quantization::Adaptive { base, gain } => (qp.ok_or_else(|| corrupt("missing QP map"))?, base, gain), + }; + if gain == 0 { + return scale(data, adaptive_step(map, band, 0, 0, base, gain)?); + } + if width == 0 { + return Err(corrupt("empty subband row")); + } + let cell = if (4..=6).contains(&band) { 2 } else { 4 }; + for (y, row) in data.chunks_exact_mut(width).enumerate() { + for (column, values) in row.chunks_mut(cell).enumerate() { + scale(values, adaptive_step(map, band, column * cell, y, base, gain)?)?; + } + } + Ok(()) +} + +fn scale(values: &mut [i32], step: i32) -> Result<()> { + let mut overflow = false; + for value in values.iter_mut() { + let (product, wrapped) = value.overflowing_mul(step); + *value = product; + overflow |= wrapped; + } + if overflow { Err(corrupt("dequantized coefficient overflow")) } else { Ok(()) } +} + fn adaptive_step(map: &QpMap, band: usize, x: usize, y: usize, base: u32, gain: u32) -> Result { let step = if gain == 0 { u64::from(base) @@ -502,17 +566,60 @@ fn be32(src: &[u8], at: usize) -> Result { Ok(u32::from_be_bytes(bytes)) } +/// An MSB-first bit reader over one entropy stream. +/// +/// Up to 64 stream bits are held in `cache`, left-aligned: the top `count` bits are the next unread +/// ones. A refill loads 8 bytes at once and keeps whole bytes only, so the bits below `count` are +/// either zero or the stream's own next bits at their final positions, which a later refill ORs in +/// again unchanged. Reads never trust those bits: they are checked against `count`. struct Bits<'a> { src: &'a [u8], - at: usize, + /// The next byte to load into `cache`. + next: usize, + cache: u64, + count: u32, } -impl Bits<'_> { - /// The next 32 bits, MSB first, when at least that many remain before the end of the stream. - fn peek32(&self) -> Option { - let bytes = self.src.get(self.at / 8..self.at / 8 + 5)?; - let word = u64::from_be_bytes([0, 0, 0, bytes[0], bytes[1], bytes[2], bytes[3], bytes[4]]); - Some((word >> (8 - self.at % 8)) as u32) +impl<'a> Bits<'a> { + fn new(src: &'a [u8]) -> Self { + Self { src, next: 0, cache: 0, count: 0 } + } + + /// The number of bits consumed so far. + fn position(&self) -> usize { + self.next * 8 - self.count as usize + } + + /// Loads whole bytes until more than 56 bits are cached or the stream ends. + fn refill(&mut self) { + if self.count > 56 { + return; + } + if let Some(bytes) = self.src.get(self.next..self.next + 8).and_then(|s| <[u8; 8]>::try_from(s).ok()) { + self.cache |= u64::from_be_bytes(bytes) >> self.count; + let loaded = (63 - self.count) / 8; + self.next += loaded as usize; + self.count += loaded * 8; + return; + } + // The last 7 bytes: one at a time. + while self.count <= 56 { + let Some(&byte) = self.src.get(self.next) else { break }; + self.cache |= u64::from(byte) << (56 - self.count); + self.next += 1; + self.count += 8; + } + } + + /// Drops `count` bits (at most `self.count`, below 64). + fn consume(&mut self, count: u32) { + self.cache <<= count; + self.count -= count; + } + + /// The top `count` cached bits (0 to 32) as a number; `>> 1 >> (63 - count)` is 0 for `count` 0. + fn top(&self, count: u32) -> u32 { + (self.cache >> 1 >> (63 - count)) as u32 } fn take(&mut self, count: u32) -> Result { @@ -522,27 +629,34 @@ impl Bits<'_> { if count == 0 { return Ok(0); } - if let Some(window) = self.peek32() { - self.at += count as usize; - return Ok(window >> (32 - count)); - } - // Near the end of the stream: bit by bit, so truncation is reported exactly. - let mut value = 0; - for _ in 0..count { - let byte = self.src.get(self.at / 8).ok_or_else(|| corrupt("truncated entropy stream"))?; - value = (value << 1) | u32::from((byte >> (7 - self.at % 8)) & 1); - self.at += 1; + if self.count < count { + self.refill(); + if self.count < count { + return Err(corrupt("truncated entropy stream")); + } } + let value = self.top(count); + self.consume(count); Ok(value) } fn rice(&mut self, k: u32) -> Result { - // Common case: the unary prefix and its terminating 1 lie within the next 32 bits. - if let Some(window) = self.peek32().filter(|&w| w != 0) { - let zeros = window.leading_zeros(); - self.at += zeros as usize + 1; - return if k <= 20 { Ok((zeros << k) | self.take(k)?) } else { Err(corrupt("Rice parameter exceeds sample precision")) }; + // Common case: the unary prefix (shorter than the 41-zero escape), its terminating 1 and the k + // low bits are all cached. + if k <= 20 { + let mut zeros = self.cache.leading_zeros(); + if zeros + 1 + k > self.count { + self.refill(); + zeros = self.cache.leading_zeros(); + } + if zeros < 41 && zeros + 1 + k <= self.count { + self.consume(zeros + 1); + let low = self.top(k); + self.consume(k); + return Ok((zeros << k) | low); + } } + // Escapes, k above 20 and the end of the stream: bit by bit, so truncation is reported exactly. let mut zeros = 0; while self.take(1)? == 0 { zeros += 1; @@ -580,11 +694,12 @@ impl Bits<'_> { fn finish(&self) -> Result<()> { let bits = self.src.len().checked_mul(8).ok_or(RawError::Limit("CRX entropy length"))?; - if bits.saturating_sub(self.at) > 7 { + let at = self.position(); + if bits.saturating_sub(at) > 7 { return Err(corrupt("unconsumed entropy data")); } if let Some(&last) = self.src.last() { - let remaining = (bits - self.at) as u32; + let remaining = (bits - at) as u32; if u32::from(last) & ((1 << remaining) - 1) != 0 { return Err(corrupt("nonzero entropy alignment bits")); } @@ -653,7 +768,7 @@ fn lossless_plane(src: &[u8], width: usize, height: usize) -> Result> { current.try_reserve_exact(width).map_err(|_| RawError::Limit("CRX row allocation"))?; current.resize(width, 0i32); let (mut k, mut run_index) = (0u32, 0usize); - let mut bits = Bits { src, at: 0 }; + let mut bits = Bits::new(src); for y in 0..height { let mut x = 0; let mut left = previous.first().copied().ok_or_else(|| corrupt("empty plane row"))?; @@ -716,7 +831,7 @@ fn high_frequency_band(src: &[u8], width: usize, height: usize) -> Result Result { Ok(values) }; let (mut previous, mut current) = (row()?, row()?); - let mut bits = Bits { src, at: 0 }; + let mut bits = Bits::new(src); let mut k = 0; for y in 0..height { let mut left = *previous.first().ok_or_else(|| corrupt("empty QP row"))?; @@ -879,11 +994,11 @@ mod tests { #[test] fn partial_final_run_segment_keeps_its_index() { - let mut bits = Bits { src: &[0xc0], at: 0 }; + let mut bits = Bits::new(&[0xc0]); let mut index = 11; assert_eq!(run(&mut bits, &mut index, 6).unwrap(), 6); assert_eq!(index, 12); - assert_eq!(bits.at, 2); + assert_eq!(bits.position(), 2); } #[test] @@ -1043,14 +1158,22 @@ mod tests { #[test] fn only_quantized_planes_clamp_out_of_range_samples() { - assert_eq!(sensor_sample(8200, 8192, 16383, true).unwrap(), 16383); - assert_eq!(sensor_sample(-8300, 8192, 16383, true).unwrap(), 0); - assert_eq!(sensor_sample(8191, 8192, 16383, false).unwrap(), 16383); - assert_eq!(sensor_sample(-8192, 8192, 16383, false).unwrap(), 0); + assert_eq!(sensor_plane(&[8200], 8192, 16383, true).unwrap(), [16383]); + assert_eq!(sensor_plane(&[-8300], 8192, 16383, true).unwrap(), [0]); + assert_eq!(sensor_plane(&[-8300, 0, 8200], 8192, 16383, true).unwrap(), [0, 8192, 16383]); + assert_eq!(sensor_plane(&[8191], 8192, 16383, false).unwrap(), [16383]); + assert_eq!(sensor_plane(&[-8192], 8192, 16383, false).unwrap(), [0]); + assert_eq!(sensor_plane(&[-8192, 5, 8191], 8192, 16383, false).unwrap(), [0, 8197, 16383]); // Lossless planes reproduce the sensor exactly: outside the bit depth is corrupt data. - assert!(matches!(sensor_sample(8192, 8192, 16383, false), Err(RawError::Corrupt(_)))); - assert!(matches!(sensor_sample(-8193, 8192, 16383, false), Err(RawError::Corrupt(_)))); - assert!(matches!(sensor_sample(i32::MAX, 8192, 16383, true), Err(RawError::Corrupt(_)))); + assert!(matches!(sensor_plane(&[8192], 8192, 16383, false), Err(RawError::Corrupt(_)))); + assert!(matches!(sensor_plane(&[-8193], 8192, 16383, false), Err(RawError::Corrupt(_)))); + assert!(matches!(sensor_plane(&[0, 8192, 0], 8192, 16383, false), Err(RawError::Corrupt(_)))); + assert!(matches!(sensor_plane(&[0, -8193, 0], 8192, 16383, false), Err(RawError::Corrupt(_)))); + assert!(matches!(sensor_plane(&[i32::MAX], 8192, 16383, true), Err(RawError::Corrupt(_)))); + assert!(matches!(sensor_plane(&[0, i32::MAX], 8192, 16383, true), Err(RawError::Corrupt(_)))); + // The lowest i32 cannot overflow below: it is clamped in a quantized plane and corrupt in a lossless one. + assert_eq!(sensor_plane(&[i32::MIN], 8192, 16383, true).unwrap(), [0]); + assert!(matches!(sensor_plane(&[i32::MIN], 8192, 16383, false), Err(RawError::Corrupt(_)))); } #[test] @@ -1081,14 +1204,14 @@ mod tests { #[test] fn qp_escape_and_average_precede_the_quantization_table() { // 23 zeros, the terminating 1, then the symbol in 8 bits. - let mut bits = Bits { src: &[0, 0, 1, 0x18], at: 0 }; + let mut bits = Bits::new(&[0, 0, 1, 0x18]); assert_eq!(bits.qp_rice(0).unwrap(), 24); - assert_eq!(bits.at, 32); + assert_eq!(bits.position(), 32); // 22 zeros are an ordinary Rice prefix. - let mut bits = Bits { src: &[0, 0, 2, 0], at: 0 }; + let mut bits = Bits::new(&[0, 0, 2, 0]); assert_eq!(bits.qp_rice(0).unwrap(), 22); - assert_eq!(bits.at, 23); - assert!(Bits { src: &[0, 0, 0, 0x80, 0], at: 0 }.qp_rice(0).is_err()); + assert_eq!(bits.position(), 23); + assert!(Bits::new(&[0, 0, 0, 0x80, 0]).qp_rice(0).is_err()); // The map's first symbol is a 32-bit literal (300 = QP 150 from prediction 0). let map = qp_map(&[0, 0, 1, 0x2c, 0x94], 1, 4).unwrap(); assert_eq!(map.data, [150; 4]); @@ -1187,11 +1310,17 @@ mod tests { for end in 0..sample.len() { assert!(decode(&sample[..end], &config).is_err()); } + // A corrupt QP map (nonzero alignment bits) is reported as such, not as a plane error. + let at = sample.windows(5).position(|bytes| bytes == [0, 0, 1, 0x2c, 0x94]).unwrap() + 4; + sample[at] ^= 1; + let error = decode(&sample, &config).unwrap_err().to_string(); + assert!(error.contains("alignment") && !error.contains("plane"), "{error}"); } #[test] fn optional_cc0_samples_match_external_oracle() { let Some(directory) = std::env::var_os("LIGHTCRAFT_CR3_CORPUS") else { return }; + let mut checked = 0; for (name, expected) in [ ("cr3-canon-r100-raw.cr3", 0xcbe5299ab9c52630u64), ("cr3-canon-m50-raw.cr3", 0x62261f0ba81cfcd2u64), @@ -1203,9 +1332,18 @@ mod tests { ("cr3-canon-rp-craw.cr3", 0xf48df06ca1d67382u64), ("cr3-canon-rp-raw-crop.cr3", 0x46808a9de3b3dcc7u64), ("cr3-canon-rp-craw-crop.cr3", 0x39a60c4efed4e1f4u64), + // EOS R5 Mark II (version 0x200 C-RAW and its lossless pair): pinned to this decoder's + // own output, see docs/cr3.md. + ("cr3-canon-r5m2-raw.cr3", 0xff716fb30248496fu64), + ("cr3-canon-r5m2-craw.cr3", 0xaf32b45357b534e5u64), + ("cr3-canon-r5m2-raw-crop.cr3", 0x31cf630afe868958u64), + ("cr3-canon-r5m2-craw-crop.cr3", 0x4721605b20339332u64), ] { let path = std::path::Path::new(&directory).join(name); - let bytes = std::fs::read(path).unwrap(); + let Ok(bytes) = std::fs::read(&path) else { + eprintln!("skipped {}: not in LIGHTCRAFT_CR3_CORPUS", path.display()); + continue; + }; let container = craft_metadata::cr3::parse_cr3(&bytes).unwrap(); let mut candidates = container .tracks @@ -1224,6 +1362,285 @@ mod tests { .flat_map(|value| value.to_le_bytes()) .fold(0xcbf29ce484222325u64, |hash, byte| (hash ^ u64::from(byte)).wrapping_mul(0x100000001b3)); assert_eq!(hash, expected, "{name}"); + checked += 1; + } + assert!(checked > 0, "LIGHTCRAFT_CR3_CORPUS holds none of the pinned files"); + } + + /// The bit reader before the 64-bit cache: a checked 5-byte window per read, bit by bit near the end. + struct ReferenceBits<'a> { + src: &'a [u8], + at: usize, + } + + impl ReferenceBits<'_> { + fn peek32(&self) -> Option { + let bytes = self.src.get(self.at / 8..self.at / 8 + 5)?; + let word = u64::from_be_bytes([0, 0, 0, bytes[0], bytes[1], bytes[2], bytes[3], bytes[4]]); + Some((word >> (8 - self.at % 8)) as u32) + } + + fn take(&mut self, count: u32) -> Result { + if count > 24 { + return Err(corrupt("invalid Rice parameter")); + } + if count == 0 { + return Ok(0); + } + if let Some(window) = self.peek32() { + self.at += count as usize; + return Ok(window >> (32 - count)); + } + let mut value = 0; + for _ in 0..count { + let byte = self.src.get(self.at / 8).ok_or_else(|| corrupt("truncated entropy stream"))?; + value = (value << 1) | u32::from((byte >> (7 - self.at % 8)) & 1); + self.at += 1; + } + Ok(value) + } + + fn rice(&mut self, k: u32) -> Result { + if let Some(window) = self.peek32().filter(|&w| w != 0) { + let zeros = window.leading_zeros(); + self.at += zeros as usize + 1; + return if k <= 20 { Ok((zeros << k) | self.take(k)?) } else { Err(corrupt("Rice parameter exceeds sample precision")) }; + } + let mut zeros = 0; + while self.take(1)? == 0 { + zeros += 1; + if zeros > 41 { + return Err(corrupt("Rice escape prefix exceeds 41 bits")); + } + } + if zeros == 41 { + self.take(21) + } else if k <= 20 { + Ok((zeros << k) | self.take(k)?) + } else { + Err(corrupt("Rice parameter exceeds sample precision")) + } + } + + fn qp_rice(&mut self, k: u32) -> Result { + let mut zeros = 0; + while self.take(1)? == 0 { + zeros += 1; + if zeros > QP_ESCAPE_ZEROS { + return Err(corrupt("QP Rice escape prefix exceeds 23 bits")); + } + } + if zeros == QP_ESCAPE_ZEROS { + self.take(8) + } else if k <= 20 { + Ok((zeros << k) | self.take(k)?) + } else { + Err(corrupt("QP Rice parameter exceeds precision")) + } + } + + fn finish(&self) -> bool { + let bits = self.src.len() * 8; + bits.saturating_sub(self.at) <= 7 && self.src.last().is_none_or(|&last| u32::from(last) & ((1u32 << (bits - self.at)) - 1) == 0) } } + + /// A xorshift generator: deterministic pseudo-random test input without a dependency. + fn xorshift(seed: &mut u64) -> u64 { + *seed ^= *seed << 13; + *seed ^= *seed >> 7; + *seed ^= *seed << 17; + *seed + } + + #[test] + fn cached_bit_reader_matches_the_bitwise_reference() { + let mut seed = 0x2545_f491_4f6c_dd1d; + for round in 0..400 { + let length = (xorshift(&mut seed) % 48) as usize; + // Mostly zero and small bytes, so long unary prefixes and both escapes occur. + let stream: Vec = (0..length) + .map(|_| match xorshift(&mut seed) % 4 { + 0 | 1 => 0, + 2 => (xorshift(&mut seed) & 0x0f) as u8, + _ => xorshift(&mut seed) as u8, + }) + .collect(); + let ops: Vec<(u64, u32)> = (0..48).map(|_| (xorshift(&mut seed) % 3, (xorshift(&mut seed) % 26) as u32)).collect(); + // Every truncation of the stream, so reads meet the end at every offset. + for end in 0..=stream.len() { + let src = &stream[..end]; + let (mut cached, mut reference) = (Bits::new(src), ReferenceBits { src, at: 0 }); + let mut complete = true; + for &(op, n) in &ops { + let (a, b) = match op { + 0 => (cached.take(n), reference.take(n)), + 1 => (cached.rice(n), reference.rice(n)), + _ => (cached.qp_rice(n % 22), reference.qp_rice(n % 22)), + }; + assert_eq!(a.as_ref().ok(), b.as_ref().ok(), "round {round}, end {end}, op {op} {n}, stream {stream:02x?}"); + assert_eq!( + a.as_ref().err().map(ToString::to_string), + b.as_ref().err().map(ToString::to_string), + "round {round}, end {end}, op {op} {n}" + ); + if a.is_err() { + complete = false; + break; + } + assert_eq!(cached.position(), reference.at, "round {round}, end {end}, op {op} {n}"); + } + if complete { + assert_eq!(cached.finish().is_ok(), reference.finish(), "round {round}, end {end}"); + } + } + } + } + + #[test] + fn rice_escape_is_an_error_when_cut_at_any_byte() { + // 41 zeros, the terminating 1 and the 21-bit literal 16400 (+8200): the whole stream is one escape. + let src = [0, 0, 0, 0, 0, 0x40, 0x80, 0x20]; + assert_eq!(lossless_plane(&src, 1, 1).unwrap(), [8200]); + for end in 0..src.len() { + assert!(lossless_plane(&src[..end], 1, 1).is_err(), "prefix {end}"); + } + // The same escape read from the cache, 9 bytes into a longer stream. + let mut long = vec![0xff; 9]; + long.extend(src); + let mut bits = Bits::new(&long); + for _ in 0..72 { + assert_eq!(bits.rice(0).unwrap(), 0); + } + assert_eq!(bits.rice(0).unwrap(), 16400); + bits.finish().unwrap(); + } + + #[test] + fn dequantization_per_cell_matches_a_step_per_coefficient() { + let mut seed = 0x9e37_79b9_7f4a_7c15; + // Odd map heights as on the EOS R7, and band sizes that fit the map and that overrun it by one. + for (map_width, map_height) in [(1, 1), (2, 3), (3, 4), (5, 7)] { + let map = QpMap { + width: map_width, + height: map_height, + data: (0..map_width * map_height).map(|_| 131 + (xorshift(&mut seed) % 45) as i32).collect(), + }; + for band in 0..10 { + for (width, height) in [ + (1, 1), + (3, 2), + (2 * map_width - 1, map_height.div_ceil(2)), + (2 * map_width + 1, map_height), + (4 * map_width, map_height), + (4 * map_width + 1, map_height + 1), + ] { + for quant in [ + Quantization::Uniform(4 + 4 * band as u8), + Quantization::Adaptive { base: 0, gain: 0 }, + Quantization::Adaptive { base: 3, gain: 0 }, + Quantization::Adaptive { base: 1, gain: 8 }, + Quantization::Adaptive { base: 0, gain: 16 }, + ] { + let data: Vec = (0..width * height).map(|_| (xorshift(&mut seed) % 2001) as i32 - 1000).collect(); + let mut cells = data.clone(); + let cells_result = dequantize(&mut cells, band, width, &quant, Some(&map)); + let mut reference = data; + let reference_result = reference.iter_mut().enumerate().try_for_each(|(position, value)| -> Result<()> { + let step = match quant { + Quantization::Uniform(uniform) => quant_step(uniform, MAX_UNIFORM_QUANT)?, + Quantization::Adaptive { base, gain } => adaptive_step(&map, band, position % width, position / width, base, gain)?, + }; + *value = value.checked_mul(step).ok_or_else(|| corrupt("dequantized coefficient overflow"))?; + Ok(()) + }); + let case = format!("map {map_width}x{map_height}, band {band}, {width}x{height}, {quant:?}"); + assert_eq!(cells_result.is_ok(), reference_result.is_ok(), "{case}"); + assert_eq!( + cells_result.as_ref().err().map(ToString::to_string), + reference_result.as_ref().err().map(ToString::to_string), + "{case}" + ); + if cells_result.is_ok() { + assert_eq!(cells, reference, "{case}"); + } + } + } + } + } + assert!(dequantize(&mut [1 << 30], 1, 1, &Quantization::Uniform(10), None).is_err()); + assert!(dequantize(&mut [1], 4, 1, &Quantization::Adaptive { base: 0, gain: 0 }, None).is_err()); + } + + #[test] + fn interleave_places_every_plane_of_every_tile_like_a_per_sample_scatter() { + // A 2 x 2 tile grid whose right column is narrower, as at the right edge of a real image. + let (width, height) = (12, 8); + let tile = |x, y, width, height| Tile { x, y, width, height, planes: Vec::new(), qp: None }; + let tiles = [tile(0, 0, 8, 4), tile(8, 0, 4, 4), tile(0, 4, 8, 4), tile(8, 4, 4, 4)]; + let planes: Vec> = tiles + .iter() + .enumerate() + .flat_map(|(t, tile)| { + (0..4).map(move |p| (0..tile.width * tile.height / 4).map(|i| ((t * 4 + p) * 100 + i) as u16).collect::>()) + }) + .collect(); + let mut expected = vec![0u16; width * height]; + for (t, tile) in tiles.iter().enumerate() { + for p in 0..4 { + for (i, &value) in planes[t * 4 + p].iter().enumerate() { + let (x, y) = (tile.x + 2 * (i % (tile.width / 2)) + p % 2, tile.y + 2 * (i / (tile.width / 2)) + p / 2); + expected[y * width + x] = value; + } + } + } + let mut output = vec![0u16; width * height]; + interleave(&mut output, width, &tiles, &planes).unwrap(); + assert_eq!(output, expected); + // A plane shorter than its tile is an error, not a partial image. + let mut short = planes.clone(); + short[5].pop(); + assert!(interleave(&mut output, width, &tiles, &short).is_err()); + assert!(interleave(&mut output, width, &tiles, &planes[..12]).is_err()); + } + + #[test] + fn lossless_tile_grid_decodes_each_plane_to_its_place() { + // Four 4 x 4 tiles in a 2 x 2 grid (horizontal and vertical tile boundaries). Each 2 x 2 plane + // is flat at value 1, 2 or 3, chosen per tile and plane. Bitstreams per value: row 0 has a + // zero-run flag 0, the Rice-coded first sample and a 0 symbol; row 1 is one run to the row end. + let stream = |value: usize| -> &'static [u8] { + match value { + 1 => &[0x1e], // 0, 001 (k 0), 1 (k 0), 1, 1 + 2 => &[0x06, 0xc0], // 0, 00001 (k 0), 10 (k 1), 1, 1 + _ => &[0x01, 0x98], // 0, 0000001 (k 0), 100 (k 2), 1, 1 + } + }; + let value = |t: usize, p: usize| 1 + (t + p) % 3; + let mut config = coding(); + config.width = 8; + config.height = 8; + config.header_size = 432; + let mut sample = Vec::new(); + let mut data = Vec::new(); + for t in 0..4 { + let sizes: Vec = (0..4).map(|p| stream(value(t, p)).len() as u32).collect(); + marker(&mut sample, 0xff01, sizes.iter().sum(), (t as u32) << 16); + for (p, &size) in sizes.iter().enumerate() { + marker(&mut sample, 0xff02, size, ((p as u32) << 28) | 0x0800_0000); + marker(&mut sample, 0xff03, size, 0x0020_0000); + data.extend_from_slice(stream(value(t, p))); + } + } + assert_eq!(sample.len(), 432); + sample.extend(data); + validate(&sample, &config).unwrap(); + let expected: Vec = + (0..64).map(|i| (i % 8, i / 8)).map(|(x, y)| 8192 + value((y / 4) * 2 + x / 4, (y % 2) * 2 + x % 2) as u16).collect(); + assert_eq!(decode(&sample, &config).unwrap(), expected); + // A corrupt plane in the last tile still fails the whole image. + let mut broken = sample.clone(); + *broken.last_mut().unwrap() ^= 1; + assert!(decode(&broken, &config).is_err()); + } } diff --git a/crates/raw/src/crx_wavelet.rs b/crates/raw/src/crx_wavelet.rs index ac5e51d1d..190840be7 100644 --- a/crates/raw/src/crx_wavelet.rs +++ b/crates/raw/src/crx_wavelet.rs @@ -6,6 +6,7 @@ //! neighbouring low/high coefficients; only outer image edges are symmetric. use crate::{RawError, Result}; +use rayon::prelude::*; #[derive(Clone, Debug)] pub(crate) struct Band { @@ -18,10 +19,6 @@ fn corrupt() -> RawError { RawError::Corrupt("invalid CRX wavelet subband geometry".into()) } -fn checked_sample(v: i64) -> Result { - i32::try_from(v).map_err(|_| RawError::Corrupt("CRX wavelet coefficient overflow".into())) -} - #[derive(Clone, Copy, Debug, Default)] pub(crate) struct TileEdges { pub left: bool, @@ -34,40 +31,112 @@ fn sizes(length: usize, before: bool, after: bool) -> (usize, usize) { (length.div_ceil(2) + usize::from(after && length.is_multiple_of(2)), length / 2 + usize::from(before) + usize::from(after)) } +/// Whether a lifted value still fits the i32 coefficient range. +fn fits(value: i64) -> bool { + i32::try_from(value).is_ok() +} + +fn overflow() -> RawError { + RawError::Corrupt("CRX wavelet coefficient overflow".into()) +} + fn inverse_line(low: &[i32], high: &[i32], out: &mut [i32], before: bool, after: bool) -> Result<()> { let (low_count, high_count) = sizes(out.len(), before, after); if out.is_empty() || low.len() != low_count || high.len() != high_count { return Err(corrupt()); } - if high.is_empty() { + let Some(&last_high) = high.last() else { let value = low.first().copied().ok_or_else(corrupt)?; *out.first_mut().ok_or_else(corrupt)? = value; return Ok(()); - } + }; // The high-pass storage starts at -1 when there is a preceding tile. // Only absent external neighbours are reflected at the image boundary. - let high_at = |index: usize| -> Result { high.get(index.min(high.len().saturating_sub(1))).copied().ok_or_else(corrupt) }; - let even = |i: usize| -> Result { - let left = high_at((i + usize::from(before)).saturating_sub(1))?; - let right = high_at(i + usize::from(before))?; - let update = (i64::from(left) + i64::from(right) + 2).div_euclid(4); - checked_sample(i64::from(*low.get(i).ok_or_else(corrupt)?) - update) + let shift = usize::from(before); + let high_at = |index: usize| i64::from(high.get(index).copied().unwrap_or(last_high)); + let even = |i: usize, value: i32| i64::from(value) - ((high_at((i + shift).saturating_sub(1)) + high_at(i + shift) + 2) >> 2); + let mut in_range = true; + for (i, (slot, &value)) in out.iter_mut().step_by(2).zip(low).enumerate() { + let lifted = even(i, value); + in_range &= fits(lifted); + *slot = lifted as i32; + } + // With a following tile, the last odd sample's right neighbour is the even sample past the end. + let length = out.len(); + let past_end = if after && length.is_multiple_of(2) { low.get(length / 2).map(|&value| even(length / 2, value)) } else { None }; + in_range &= past_end.is_none_or(fits); + for i in 0..length / 2 { + let left = out.get(2 * i).copied().map_or(0, i64::from); + let right = out.get(2 * i + 2).copied().map(i64::from).or(past_end).unwrap_or(left); + let lifted = high_at(i + shift) + ((left + right) >> 1); + in_range &= fits(lifted); + if let Some(slot) = out.get_mut(2 * i + 1) { + *slot = lifted as i32; + } + } + if in_range { Ok(()) } else { Err(overflow()) } +} + +/// `out = low - floor((left + right + 2) / 4)` per column: one even row of the vertical lifting. +fn even_row(out: &mut [i32], low: &[i32], left: &[i32], right: &[i32]) -> bool { + let mut in_range = true; + for (((slot, &l), &a), &b) in out.iter_mut().zip(low).zip(left).zip(right) { + let lifted = i64::from(l) - ((i64::from(a) + i64::from(b) + 2) >> 2); + in_range &= fits(lifted); + *slot = lifted as i32; + } + in_range +} + +/// `out = high + floor((above + below) / 2)` per column: one odd row of the vertical lifting. +fn odd_row(out: &mut [i32], high: &[i32], above: &[i32], below: &[i32]) -> bool { + let mut in_range = true; + for (((slot, &h), &a), &b) in out.iter_mut().zip(high).zip(above).zip(below) { + let lifted = i64::from(h) + ((i64::from(a) + i64::from(b)) >> 1); + in_range &= fits(lifted); + *slot = lifted as i32; + } + in_range +} + +/// [`inverse_line`] applied to every column of `out`, on whole rows instead of strided columns. +/// +/// `rows` holds the low-pass rows, then the high-pass rows, each `width` wide. Runs in parallel over +/// pairs of output rows: each pair lifts its even row, recomputes the next even row (the following +/// pair's, or the one a following tile supplies past the end) and lifts its odd row from both. +fn inverse_columns(rows: &[i32], out: &mut [i32], width: usize, before: bool, after: bool) -> Result<()> { + let height = out.len().checked_div(width).filter(|&h| h > 0 && h * width == out.len()).ok_or_else(corrupt)?; + let (low_count, high_count) = sizes(height, before, after); + if rows.len() != (low_count + high_count) * width { + return Err(corrupt()); + } + let (low, high) = (rows.get(..low_count * width).ok_or_else(corrupt)?, rows.get(low_count * width..).ok_or_else(corrupt)?); + let low_row = |i: usize| low.get(i * width..(i + 1) * width).ok_or_else(corrupt); + let Some(last_high) = high_count.checked_sub(1) else { + let first = out.get_mut(..width).ok_or_else(corrupt)?; + first.copy_from_slice(low_row(0)?); + return Ok(()); }; - for i in 0..out.len().div_ceil(2) { - *out.get_mut(i * 2).ok_or_else(corrupt)? = even(i)?; - } - for i in 0..out.len() / 2 { - let left = out.get(i * 2).copied().ok_or_else(corrupt)?; - let right = match out.get(i * 2 + 2).copied() { - Some(value) => value, - None if after => even(i + 1)?, - None => left, - }; - let prediction = (i64::from(left) + i64::from(right)).div_euclid(2); - let value = high_at(i + usize::from(before))?; - *out.get_mut(i * 2 + 1).ok_or_else(corrupt)? = checked_sample(i64::from(value) + prediction)?; - } - Ok(()) + let high_row = |i: usize| { + let i = i.min(last_high); + high.get(i * width..(i + 1) * width).ok_or_else(corrupt) + }; + let shift = usize::from(before); + out.par_chunks_mut(2 * width).enumerate().try_for_each_init(Vec::new, |next, (i, pair)| -> Result<()> { + let (even, odd) = pair.split_at_mut(width); + let mut in_range = even_row(even, low_row(i)?, high_row((i + shift).saturating_sub(1))?, high_row(i + shift)?); + if !odd.is_empty() { + let below: &[i32] = if 2 * i + 2 < height || after { + next.resize(width, 0); + in_range &= even_row(next, low_row(i + 1)?, high_row(i + shift)?, high_row(i + shift + 1)?); + next.as_slice() + } else { + &*even + }; + in_range &= odd_row(odd, high_row(i + shift)?, even, below); + } + if in_range { Ok(()) } else { Err(overflow()) } + }) } fn valid_band(band: &Band, width: usize, height: usize) -> bool { @@ -97,32 +166,18 @@ pub(crate) fn inverse_level(ll: &Band, hl: &Band, lh: &Band, hh: &Band, width: u let mut temporary = zeroed(intermediate_count)?; // CRX analysis is vertical then horizontal. Integer rounding means that // synthesis must reverse that order, including for quantized subbands. - for y in 0..lhgt + hhgt { + temporary.par_chunks_mut(width).enumerate().try_for_each(|(y, row)| { let (lo, hi, by) = if y < lhgt { (ll, hl, y) } else { (lh, hh, y - lhgt) }; inverse_line( lo.data.get(by * lw..(by + 1) * lw).ok_or_else(corrupt)?, hi.data.get(by * hw..(by + 1) * hw).ok_or_else(corrupt)?, - temporary.get_mut(y * width..(y + 1) * width).ok_or_else(corrupt)?, + row, edges.left, edges.right, - )?; - } + ) + })?; let mut data = zeroed(count)?; - let mut low = zeroed(lhgt)?; - let mut high = zeroed(hhgt)?; - let mut column = zeroed(height)?; - for x in 0..width { - for (y, value) in low.iter_mut().enumerate() { - *value = temporary.get(y * width + x).copied().ok_or_else(corrupt)?; - } - for (y, value) in high.iter_mut().enumerate() { - *value = temporary.get((y + lhgt) * width + x).copied().ok_or_else(corrupt)?; - } - inverse_line(&low, &high, &mut column, edges.top, edges.bottom)?; - for (y, &value) in column.iter().enumerate() { - *data.get_mut(y * width + x).ok_or_else(corrupt)? = value; - } - } + inverse_columns(&temporary, &mut data, width, edges.top, edges.bottom)?; Ok(Band { width, height, data }) } @@ -335,4 +390,139 @@ mod tests { assert!(reconstruct(&bands, 20, 9, 3, TileEdges::default()).is_err()); } } + + /// The lifting before the row-wise rewrite: a checked closure per sample, as a reference. + fn reference_line(low: &[i32], high: &[i32], out: &mut [i32], before: bool, after: bool) -> Result<()> { + let (low_count, high_count) = sizes(out.len(), before, after); + if out.is_empty() || low.len() != low_count || high.len() != high_count { + return Err(corrupt()); + } + if high.is_empty() { + out[0] = low[0]; + return Ok(()); + } + let checked = |v: i64| i32::try_from(v).map_err(|_| overflow()); + let high_at = |index: usize| high[index.min(high.len() - 1)]; + let even = |i: usize| -> Result { + let left = high_at((i + usize::from(before)).saturating_sub(1)); + let right = high_at(i + usize::from(before)); + checked(i64::from(low[i]) - (i64::from(left) + i64::from(right) + 2).div_euclid(4)) + }; + for i in 0..out.len().div_ceil(2) { + out[i * 2] = even(i)?; + } + for i in 0..out.len() / 2 { + let left = out[i * 2]; + let right = match out.get(i * 2 + 2).copied() { + Some(value) => value, + None if after => even(i + 1)?, + None => left, + }; + out[i * 2 + 1] = checked(i64::from(high_at(i + usize::from(before))) + (i64::from(left) + i64::from(right)).div_euclid(2))?; + } + Ok(()) + } + + /// One level before the row-wise rewrite: horizontal lines, then strided columns. + fn reference_level(bands: &[Band; 4], width: usize, height: usize, edges: TileEdges) -> Result> { + let [ll, hl, lh, hh] = bands; + let (lhgt, hhgt) = sizes(height, edges.top, edges.bottom); + let mut temporary = vec![0; width * (lhgt + hhgt)]; + for y in 0..lhgt + hhgt { + let (lo, hi, by) = if y < lhgt { (ll, hl, y) } else { (lh, hh, y - lhgt) }; + reference_line( + &lo.data[by * lo.width..(by + 1) * lo.width], + &hi.data[by * hi.width..(by + 1) * hi.width], + &mut temporary[y * width..(y + 1) * width], + edges.left, + edges.right, + )?; + } + let mut data = vec![0; width * height]; + let mut column = vec![0; height]; + for x in 0..width { + let low: Vec = (0..lhgt).map(|y| temporary[y * width + x]).collect(); + let high: Vec = (lhgt..lhgt + hhgt).map(|y| temporary[y * width + x]).collect(); + reference_line(&low, &high, &mut column, edges.top, edges.bottom)?; + for (y, &value) in column.iter().enumerate() { + data[y * width + x] = value; + } + } + Ok(data) + } + + fn xorshift(seed: &mut u64) -> u64 { + *seed ^= *seed << 13; + *seed ^= *seed >> 7; + *seed ^= *seed << 17; + *seed + } + + /// Small coefficients, or (with `extreme`) values near the i32 limits so lifting can overflow. + fn coefficients(seed: &mut u64, count: usize, extreme: bool) -> Vec { + (0..count) + .map(|_| { + let r = xorshift(seed); + if extreme && r.is_multiple_of(5) { + if r & 32 == 0 { i32::MAX - (r >> 40) as i32 % 4 } else { i32::MIN + (r >> 40) as i32 % 4 } + } else { + (r % 2001) as i32 - 1000 + } + }) + .collect() + } + + const EDGE_CASES: [(bool, bool); 4] = [(false, false), (true, false), (false, true), (true, true)]; + + #[test] + fn line_lifting_matches_the_reference_for_every_length_and_edge() { + let mut seed = 0x0123_4567_89ab_cdef; + for length in 1..=40 { + for (before, after) in EDGE_CASES { + let (low_count, high_count) = sizes(length, before, after); + for extreme in [false, true] { + let low = coefficients(&mut seed, low_count, extreme); + let high = coefficients(&mut seed, high_count, extreme); + let (mut fast, mut reference) = (vec![0; length], vec![0; length]); + let fast_result = inverse_line(&low, &high, &mut fast, before, after); + let reference_result = reference_line(&low, &high, &mut reference, before, after); + let case = format!("length {length}, before {before}, after {after}, extreme {extreme}"); + assert_eq!(fast_result.is_ok(), reference_result.is_ok(), "{case}"); + assert_eq!( + fast_result.as_ref().err().map(ToString::to_string), + reference_result.as_ref().err().map(ToString::to_string), + "{case}" + ); + if fast_result.is_ok() { + assert_eq!(fast, reference, "{case}"); + } + } + } + } + } + + #[test] + fn row_wise_level_matches_column_lifting_for_every_size_and_edge() { + let mut seed = 0xfeed_beef_dead_cafe; + for width in 1..=17 { + for height in 1..=17 { + for (left, right) in EDGE_CASES { + for (top, bottom) in EDGE_CASES { + let edges = TileEdges { left, right, top, bottom }; + let (lw, hw) = sizes(width, left, right); + let (lhgt, hhgt) = sizes(height, top, bottom); + // Overflow cases on the smaller sizes only; they fail early and are cheap. + let extreme = width * height < 40 && xorshift(&mut seed).is_multiple_of(2); + let mut band = |w, h| Band { width: w, height: h, data: coefficients(&mut seed, w * h, extreme) }; + let bands = [band(lw, lhgt), band(hw, lhgt), band(lw, hhgt), band(hw, hhgt)]; + let fast = inverse_level(&bands[0], &bands[1], &bands[2], &bands[3], width, height, edges); + let reference = reference_level(&bands, width, height, edges); + let case = format!("{width}x{height}, {edges:?}, extreme {extreme}"); + assert_eq!(fast.as_ref().ok().map(|band| &band.data), reference.as_ref().ok(), "{case}"); + assert_eq!(fast.as_ref().err().map(ToString::to_string), reference.as_ref().err().map(ToString::to_string), "{case}"); + } + } + } + } + } } 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/rafc.rs b/crates/raw/src/rafc.rs index 2b3aa9a34..c044323e5 100644 --- a/crates/raw/src/rafc.rs +++ b/crates/raw/src/rafc.rs @@ -183,14 +183,15 @@ struct State { } impl State { + /// The smallest `k` with `count << k >= sum`. `count << k0` has the bit length of `sum` (for `k0` the bit-length + /// difference), so it is at most one step short. + #[inline] fn k(self) -> u32 { - let mut k = 0; - while self.count << k < self.sum { - k += 1; - } - k + let k = self.count.leading_zeros().saturating_sub(self.sum.leading_zeros()); + k + u32::from(self.count << k < self.sum) } + #[inline] fn update(&mut self, delta: i32) { self.sum += delta.unsigned_abs(); if self.count == 64 { @@ -202,8 +203,7 @@ impl State { } #[inline] -fn quant(v: i32, near: u32) -> i32 { - let t = [18 + 3 * near, 67 + 5 * near, 276 + 7 * near]; +fn quant(v: i32, near: u32, t: &[u32; 3]) -> i32 { let a = v.unsigned_abs(); let bucket = if a <= near { 0 @@ -232,124 +232,169 @@ fn even_prediction(nw: i32, n: i32, ne: i32, nn: i32) -> i32 { (pair + 2 * n) >> 2 } -#[inline] -fn sample(rows: &[Vec], y: usize, x: usize) -> Result { - rows.get(y).and_then(|r| r.get(x)).copied().ok_or_else(|| corrupt("invalid predictor position")) +/// Residual coding parameters of one precision (`2 * precision + 1` is the quantizer step). +#[derive(Clone, Copy)] +struct Coding { + range: u32, + bits: u32, + limit: u32, + step: i32, } -fn decode_stripe(stripe: &Stripe<'_>, height: usize, bits: u32, xtrans: bool) -> Result> { - let size = if xtrans { BLOCK * 2 / 3 } else { BLOCK / 2 }; - let mut previous = vec![vec![vec![0; size + 2]; 2]; 3]; - let max = (1u32 << bits) - 1; - let range = |near: u32| (max + 2 * near) / (2 * near + 1) + 1; - let seed = |near: u32| State { sum: ((range(near) + 32) / 64).max(2), count: 1 }; - let mut states = [[[seed(0); 41]; 2]; 3]; - // Smooth regions retain finer precision and independent state across block quantizer changes. - let mut fine = [[[[seed(0); 5]; 2]; 3], [[[seed(1); 5]; 2]; 3], [[[seed(2); 5]; 2]; 3]]; - let mut last_near = None; - let mut reader = Bits::new(stripe.bytes); - let n = height.checked_mul(stripe.width).filter(|&n| n <= crate::MAX_SAMPLES).ok_or(RawError::Limit("image too large"))?; - // a capacity hint only: the stripe's declared height is the file's word, so the reservation is capped at a sample - // per bit of the stripe's real bytes (all stripes reserve in parallel; a one-byte stripe must not take 2 GiB) - let mut output = Vec::with_capacity(n.min(stripe.bytes.len().saturating_mul(8))); - for group in 0..height / 6 { - let near = match stripe.quantizers { - Some(q) => u32::from(*q.get(group).ok_or_else(|| corrupt("truncated quantizer table"))?), - None => 0, +impl Coding { + fn new(max: u32, sample_bits: u32, precision: u32) -> Self { + let range = (max + 2 * precision) / (2 * precision + 1) + 1; + let bits = 32 - (range - 1).leading_zeros(); + Self { range, bits, limit: 4 * sample_bits - bits - 1, step: (2 * precision + 1) as i32 } + } +} + +/// Everything that depends only on a six-row group's quantizer, computed when the quantizer changes. +struct Group { + near: u32, + thresholds: [u32; 3], + main: Coding, + /// Precisions 0, 1 and 2 for the low-gradient sites. + fine: [Coding; 3], +} + +impl Group { + fn new(near: u32, max: u32, sample_bits: u32) -> Self { + Self { + near, + thresholds: [18 + 3 * near, 67 + 5 * near, 276 + 7 * near], + main: Coding::new(max, sample_bits, near), + fine: [0, 1, 2].map(|p| Coding::new(max, sample_bits, p)), + } + } +} + +fn seed(max: u32, near: u32) -> State { + let range = (max + 2 * near) / (2 * near + 1) + 1; + State { sum: ((range + 32) / 64).max(2), count: 1 } +} + +/// Colour vectors of one stripe, kept for the rows of a six-row group plus the two rows above it. Red and blue get +/// three rows per group, green six; they are stored one after the other in one buffer of rows `stride` wide. +const ROWS: [usize; 3] = [5, 8, 5]; +const BASE: [usize; 3] = [0, 5, 13]; +const TOTAL_ROWS: usize = 18; + +/// The three lines a sample is predicted from: the two above and the one being decoded. Each line has one edge +/// cell on both sides, which the new line takes from the line above. +struct Lines<'a> { + above2: &'a [i32], + above: &'a [i32], + current: &'a mut [i32], +} + +impl<'a> Lines<'a> { + fn new(plane: &'a mut [i32], y: usize, stride: usize) -> Self { + let (done, rest) = plane.split_at_mut(y * stride); + let done: &[i32] = done; + let current = &mut rest[..stride]; + let above = &done[(y - 1) * stride..]; + let above2 = &done[(y - 2) * stride..(y - 1) * stride]; + current[0] = above[1]; + current[stride - 1] = above[stride - 2]; + Self { above2, above, current } + } +} + +struct Coder<'a> { + reader: Bits<'a>, + states: [[[State; 41]; 2]; 3], + /// Smooth regions retain finer precision and independent state across block quantizer changes. + fine: [[[[State; 5]; 2]; 3]; 3], + max: u32, + xtrans: bool, +} + +impl Coder<'_> { + /// Decodes one sample's residual with the context its gradients select and applies it to the prediction. + #[inline] + fn residual(&mut self, g: &Group, pair: usize, odd: usize, (d1, d2): (i32, i32), predict: i32) -> Result { + let near = g.near; + let contrast = d1.unsigned_abs() + d2.unsigned_abs(); + let precision = match contrast { + 0..=5 => 0, + 6 => near.min(1), + 7 => near.min(2), + _ => near, }; - if last_near != Some(near) { - states = [[[seed(near); 41]; 2]; 3]; - last_near = Some(near); + let (context, state, coding) = if precision < near { + let sign = |v: i32| if v.unsigned_abs() <= precision { 0 } else { v.signum() }; + let context = 3 * sign(d1) + sign(d2); + (context, &mut self.fine[precision as usize][pair % 3][odd][context.unsigned_abs() as usize], g.fine[precision as usize]) + } else { + let context = 9 * quant(d1, near, &g.thresholds) + quant(d2, near, &g.thresholds); + (context, &mut self.states[pair % 3][odd][context.unsigned_abs() as usize], g.main) + }; + let mut delta = self.reader.residual(state.k(), coding.bits, coding.range, coding.limit)?; + state.update(delta); + if context < 0 { + delta = -delta; } - let mut rows = previous; + let mut value = predict + delta * coding.step; + if value < -(precision as i32) { + value += coding.range as i32 * coding.step; + } else if value > (self.max + precision) as i32 { + value -= coding.range as i32 * coding.step; + } + Ok(value.clamp(0, self.max as i32)) + } + + /// One position of the two colour lines of a pair: the even column (`odd == 0`) or the one seven behind it. + #[inline] + fn position(&mut self, g: &Group, pair: usize, odd: usize, pos: usize, lines: &mut [Lines<'_>; 2]) -> Result<()> { + let x = pos + 1; + for (which, l) in lines.iter_mut().enumerate() { + let (nw, north, ne) = (l.above[x - 1], l.above[x], l.above[x + 1]); + let real = !self.xtrans || (MASKS[pair][which] >> (pos % 4)) & 1 != 0; + l.current[x] = if odd == 0 { + let nn = l.above2[x]; + let predict = even_prediction(nw, north, ne, nn); + if real { self.residual(g, pair, 0, (north - nn, nw - north), predict)? } else { predict } + } else { + let (west, east) = (l.current[x - 1], l.current[x + 1]); + let predict = if (nw.min(ne)..=nw.max(ne)).contains(&north) { (west + east) >> 1 } else { (west + east + 2 * north) >> 2 }; + if real { self.residual(g, pair, 1, (north - nw, nw - west), predict)? } else { predict } + }; + } + Ok(()) + } + + /// Decodes the six colour-vector pairs of one group into `buf`. + fn group(&mut self, g: &Group, buf: &mut [i32], size: usize) -> Result<()> { + let stride = size + 2; + let (red, rest) = buf.split_at_mut(ROWS[0] * stride); + let (green, blue) = rest.split_at_mut(ROWS[1] * stride); for pair in 0..6 { - let colours = if pair % 2 == 0 { [0, 1] } else { [1, 2] }; - for &colour in &colours { - let lines = rows.get_mut(colour).ok_or_else(|| corrupt("invalid colour vector"))?; - let mut row = vec![0; size + 2]; - let prev = lines.len() - 1; - let left = sample(lines, prev, 1)?; - let right = sample(lines, prev, size)?; - if let Some(v) = row.first_mut() { - *v = left; - } - if let Some(v) = row.last_mut() { - *v = right; - } - lines.push(row); - } + let mut lines = if pair % 2 == 0 { + [Lines::new(&mut *red, 2 + pair / 2, stride), Lines::new(&mut *green, 2 + pair, stride)] + } else { + [Lines::new(&mut *green, 2 + pair, stride), Lines::new(&mut *blue, 2 + pair / 2, stride)] + }; for even in (0..size + 8).step_by(2) { - for (odd, position) in [(0, Some(even)), (1, even.checked_sub(7))] { - let Some(pos) = position.filter(|&x| x < size) else { continue }; - let x = pos + 1; - for (which, &colour) in colours.iter().enumerate() { - let lines = rows.get_mut(colour).ok_or_else(|| corrupt("invalid colour vector"))?; - let y = lines.len() - 1; - let (nw, north, ne, nn) = - (sample(lines, y - 1, x - 1)?, sample(lines, y - 1, x)?, sample(lines, y - 1, x + 1)?, sample(lines, y - 2, x)?); - let west = sample(lines, y, x - 1)?; - let east = sample(lines, y, x + 1)?; - let (predict, d1, d2) = if odd == 0 { - (even_prediction(nw, north, ne, nn), north - nn, nw - north) - } else { - let p = if (nw.min(ne)..=nw.max(ne)).contains(&north) { (west + east) >> 1 } else { (west + east + 2 * north) >> 2 }; - (p, north - nw, nw - west) - }; - let real = !xtrans || MASKS.get(pair).and_then(|m| m.get(which)).is_some_and(|m| (m >> (pos % 4)) & 1 != 0); - let value = if real { - let contrast = d1.unsigned_abs() + d2.unsigned_abs(); - let precision = match contrast { - 0..=5 => 0, - 6 => near.min(1), - 7 => near.min(2), - _ => near, - }; - let (context, state) = if precision < near { - let sign = |v: i32| if v.unsigned_abs() <= precision { 0 } else { v.signum() }; - let context = 3 * sign(d1) + sign(d2); - let state = fine - .get_mut(precision as usize) - .and_then(|f| f.get_mut(pair % 3)) - .and_then(|s| s.get_mut(odd)) - .and_then(|s| s.get_mut(context.unsigned_abs() as usize)) - .ok_or_else(|| corrupt("invalid fine coding context"))?; - (context, state) - } else { - let context = 9 * quant(d1, near) + quant(d2, near); - let state = states - .get_mut(pair % 3) - .and_then(|s| s.get_mut(odd)) - .and_then(|s| s.get_mut(context.unsigned_abs() as usize)) - .ok_or_else(|| corrupt("invalid coding context"))?; - (context, state) - }; - let coded_range = range(precision); - let coded_bits = 32 - (coded_range - 1).leading_zeros(); - let mut delta = reader.residual(state.k(), coded_bits, coded_range, 4 * bits - coded_bits - 1)?; - state.update(delta); - if context < 0 { - delta = -delta; - } - let step = (2 * precision + 1) as i32; - let mut value = predict + delta * step; - if value < -(precision as i32) { - value += coded_range as i32 * step; - } else if value > (max + precision) as i32 { - value -= coded_range as i32 * step; - } - value.clamp(0, max as i32) - } else { - predict - }; - *lines.get_mut(y).and_then(|r| r.get_mut(x)).ok_or_else(|| corrupt("invalid predictor position"))? = value; - } + if even < size { + self.position(g, pair, 0, even, &mut lines)?; + } + if even >= 7 && even - 7 < size { + self.position(g, pair, 1, even - 7, &mut lines)?; } } } - for y in 0..6 { - for x in 0..stripe.width { + Ok(()) + } +} + +/// For each of the six rows of a group, where each output sample sits in the colour-vector buffer. +fn placement(xtrans: bool, width: usize, stride: usize) -> [Vec; 6] { + std::array::from_fn(|y| { + (0..width) + .map(|x| { let colour = if xtrans { - XTRANS.get(y).and_then(|r| r.get(x % 6)).copied().ok_or_else(|| corrupt("invalid CFA position"))? + XTRANS[y][x % 6] } else { match (y % 2, x % 2) { (0, 0) => 0, @@ -359,10 +404,49 @@ fn decode_stripe(stripe: &Stripe<'_>, height: usize, bits: u32, xtrans: bool) -> }; let row = 2 + if colour == 1 { y } else { y / 2 }; let col = 1 + if xtrans { 2 * (x / 3) + (x % 3).min(1) } else { x / 2 }; - output.push(sample(rows.get(colour).ok_or_else(|| corrupt("invalid colour vector"))?, row, col)? as u16); - } + (BASE[colour] + row) * stride + col + }) + .collect() + }) +} + +fn decode_stripe(stripe: &Stripe<'_>, height: usize, bits: u32, xtrans: bool) -> Result> { + let size = if xtrans { BLOCK * 2 / 3 } else { BLOCK / 2 }; + let stride = size + 2; + let max = (1u32 << bits) - 1; + let mut coder = Coder { + reader: Bits::new(stripe.bytes), + states: [[[seed(max, 0); 41]; 2]; 3], + fine: [[[[seed(max, 0); 5]; 2]; 3], [[[seed(max, 1); 5]; 2]; 3], [[[seed(max, 2); 5]; 2]; 3]], + max, + xtrans, + }; + let mut last_near = None; + let mut group = Group::new(0, max, bits); + let n = height.checked_mul(stripe.width).filter(|&n| n <= crate::MAX_SAMPLES).ok_or(RawError::Limit("image too large"))?; + // a capacity hint only: the stripe's declared height is the file's word, so the reservation is capped at a sample + // per bit of the stripe's real bytes (all stripes reserve in parallel; a one-byte stripe must not take 2 GiB) + let mut output = Vec::with_capacity(n.min(stripe.bytes.len().saturating_mul(8))); + let place = placement(xtrans, stripe.width, stride); + let mut buf = vec![0i32; TOTAL_ROWS * stride]; + for index in 0..height / 6 { + let near = match stripe.quantizers { + Some(q) => u32::from(*q.get(index).ok_or_else(|| corrupt("truncated quantizer table"))?), + None => 0, + }; + if last_near != Some(near) { + coder.states = [[[seed(max, near); 41]; 2]; 3]; + group = Group::new(near, max, bits); + last_near = Some(near); + } + coder.group(&group, &mut buf, size)?; + for row in &place { + output.extend(row.iter().map(|&at| buf[at] as u16)); + } + // the last two rows of each colour are the next group's rows above + for (base, rows) in BASE.iter().zip(ROWS) { + buf.copy_within((base + rows - 2) * stride..(base + rows) * stride, base * stride); } - previous = rows.into_iter().map(|mut r| r.split_off(r.len() - 2)).collect(); } Ok(output) } @@ -590,3 +674,260 @@ mod tests { } } } + +/// The straightforward decoder this module started from (shift loop for `k`, an `Option`/`Result` neighbour read +/// per access, a `Vec` per colour line, one pixel placed at a time), kept as the oracle for the optimised one. +#[cfg(test)] +mod differential { + use super::*; + use sha2::{Digest, Sha256}; + + fn k_by_shifting(state: State) -> u32 { + let mut k = 0; + while state.count << k < state.sum { + k += 1; + } + k + } + + fn quant_by_table(v: i32, near: u32) -> i32 { + let t = [18 + 3 * near, 67 + 5 * near, 276 + 7 * near]; + quant(v, near, &t) + } + + fn sample(rows: &[Vec], y: usize, x: usize) -> Result { + rows.get(y).and_then(|r| r.get(x)).copied().ok_or_else(|| corrupt("invalid predictor position")) + } + + fn reference_stripe(stripe: &Stripe<'_>, height: usize, bits: u32, xtrans: bool) -> Result> { + let size = if xtrans { BLOCK * 2 / 3 } else { BLOCK / 2 }; + let mut previous = vec![vec![vec![0; size + 2]; 2]; 3]; + let max = (1u32 << bits) - 1; + let range = |near: u32| (max + 2 * near) / (2 * near + 1) + 1; + let seed = |near: u32| State { sum: ((range(near) + 32) / 64).max(2), count: 1 }; + let mut states = [[[seed(0); 41]; 2]; 3]; + let mut fine = [[[[seed(0); 5]; 2]; 3], [[[seed(1); 5]; 2]; 3], [[[seed(2); 5]; 2]; 3]]; + let mut last_near = None; + let mut reader = Bits::new(stripe.bytes); + let n = height.checked_mul(stripe.width).filter(|&n| n <= crate::MAX_SAMPLES).ok_or(RawError::Limit("image too large"))?; + let mut output = Vec::with_capacity(n.min(stripe.bytes.len().saturating_mul(8))); + for group in 0..height / 6 { + let near = match stripe.quantizers { + Some(q) => u32::from(*q.get(group).ok_or_else(|| corrupt("truncated quantizer table"))?), + None => 0, + }; + if last_near != Some(near) { + states = [[[seed(near); 41]; 2]; 3]; + last_near = Some(near); + } + let mut rows = previous; + for pair in 0..6 { + let colours = if pair % 2 == 0 { [0, 1] } else { [1, 2] }; + for &colour in &colours { + let lines = rows.get_mut(colour).ok_or_else(|| corrupt("invalid colour vector"))?; + let mut row = vec![0; size + 2]; + let prev = lines.len() - 1; + let left = sample(lines, prev, 1)?; + let right = sample(lines, prev, size)?; + if let Some(v) = row.first_mut() { + *v = left; + } + if let Some(v) = row.last_mut() { + *v = right; + } + lines.push(row); + } + for even in (0..size + 8).step_by(2) { + for (odd, position) in [(0, Some(even)), (1, even.checked_sub(7))] { + let Some(pos) = position.filter(|&x| x < size) else { continue }; + let x = pos + 1; + for (which, &colour) in colours.iter().enumerate() { + let lines = rows.get_mut(colour).ok_or_else(|| corrupt("invalid colour vector"))?; + let y = lines.len() - 1; + let (nw, north, ne, nn) = + (sample(lines, y - 1, x - 1)?, sample(lines, y - 1, x)?, sample(lines, y - 1, x + 1)?, sample(lines, y - 2, x)?); + let west = sample(lines, y, x - 1)?; + let east = sample(lines, y, x + 1)?; + let (predict, d1, d2) = if odd == 0 { + (even_prediction(nw, north, ne, nn), north - nn, nw - north) + } else { + let p = if (nw.min(ne)..=nw.max(ne)).contains(&north) { (west + east) >> 1 } else { (west + east + 2 * north) >> 2 }; + (p, north - nw, nw - west) + }; + let real = !xtrans || MASKS.get(pair).and_then(|m| m.get(which)).is_some_and(|m| (m >> (pos % 4)) & 1 != 0); + let value = if real { + let contrast = d1.unsigned_abs() + d2.unsigned_abs(); + let precision = match contrast { + 0..=5 => 0, + 6 => near.min(1), + 7 => near.min(2), + _ => near, + }; + let (context, state) = if precision < near { + let sign = |v: i32| if v.unsigned_abs() <= precision { 0 } else { v.signum() }; + let context = 3 * sign(d1) + sign(d2); + let state = fine + .get_mut(precision as usize) + .and_then(|f| f.get_mut(pair % 3)) + .and_then(|s| s.get_mut(odd)) + .and_then(|s| s.get_mut(context.unsigned_abs() as usize)) + .ok_or_else(|| corrupt("invalid fine coding context"))?; + (context, state) + } else { + let context = 9 * quant_by_table(d1, near) + quant_by_table(d2, near); + let state = states + .get_mut(pair % 3) + .and_then(|s| s.get_mut(odd)) + .and_then(|s| s.get_mut(context.unsigned_abs() as usize)) + .ok_or_else(|| corrupt("invalid coding context"))?; + (context, state) + }; + let coded_range = range(precision); + let coded_bits = 32 - (coded_range - 1).leading_zeros(); + let mut delta = reader.residual(k_by_shifting(*state), coded_bits, coded_range, 4 * bits - coded_bits - 1)?; + state.update(delta); + if context < 0 { + delta = -delta; + } + let step = (2 * precision + 1) as i32; + let mut value = predict + delta * step; + if value < -(precision as i32) { + value += coded_range as i32 * step; + } else if value > (max + precision) as i32 { + value -= coded_range as i32 * step; + } + value.clamp(0, max as i32) + } else { + predict + }; + *lines.get_mut(y).and_then(|r| r.get_mut(x)).ok_or_else(|| corrupt("invalid predictor position"))? = value; + } + } + } + } + for y in 0..6 { + for x in 0..stripe.width { + let colour = if xtrans { + XTRANS.get(y).and_then(|r| r.get(x % 6)).copied().ok_or_else(|| corrupt("invalid CFA position"))? + } else { + match (y % 2, x % 2) { + (0, 0) => 0, + (1, 1) => 2, + _ => 1, + } + }; + let row = 2 + if colour == 1 { y } else { y / 2 }; + let col = 1 + if xtrans { 2 * (x / 3) + (x % 3).min(1) } else { x / 2 }; + output.push(sample(rows.get(colour).ok_or_else(|| corrupt("invalid colour vector"))?, row, col)? as u16); + } + } + previous = rows.into_iter().map(|mut r| r.split_off(r.len() - 2)).collect(); + } + Ok(output) + } + + /// Deterministic bytes: uniform (0), mostly zero (1: long unary prefixes and escapes) or one-heavy (2: short + /// prefixes and small residuals, which stay inside the range of every quantizer so the stream decodes to the end). + fn bytes(seed: u64, len: usize, kind: u8) -> Vec { + let mut x = seed | 1; + (0..len) + .map(|_| { + x ^= x << 13; + x ^= x >> 7; + x ^= x << 17; + let b = (x >> 24) as u8; + match kind { + 1 if x & 0b11 != 0 => 0, + 2 => b | (x >> 40) as u8, + _ => b, + } + }) + .collect() + } + + #[test] + fn k_without_the_shift_loop_matches_it() { + let mut sums: Vec = (0..=4200).collect(); + for i in 0..=30 { + sums.extend([(1u32 << i) - 1, 1 << i, (1 << i) + 1]); + } + for count in 1..=64 { + for &sum in &sums { + let state = State { sum, count }; + assert_eq!(state.k(), k_by_shifting(state), "count {count}, sum {sum}"); + } + } + } + + /// Every variant the stripe decoder handles (lossless and lossy, Bayer and X-Trans, 12/14/16 bits, full and + /// partial stripes, quantizer changes, fine contexts) gives the oracle's samples, or its error, on + /// uniform pseudo-random entropy streams that reach all contexts, one-heavy ones that every quantizer table decodes + /// to the end, and sparse ones that reach the escapes. Each quantizer table must have streams that finish. + #[test] + fn stripe_decoder_matches_the_straightforward_one() { + let quantizers: [Option<&[u8]>; 4] = + [None, Some([0, 1, 2, 3].as_slice()), Some([2, 2, 7, 0, 40, 255].as_slice()), Some([1, 1, 0, 5, 5, 9, 2].as_slice())]; + // uniform and one-heavy streams that decode to the end, per quantizer table + let mut finished = [0usize; 4]; + for bits in [12, 14, 16] { + for xtrans in [false, true] { + for (table, q) in quantizers.into_iter().enumerate() { + for (seed, width, kind) in [(1, BLOCK, 0), (2, 301, 0), (3, 12, 0), (4, BLOCK, 1), (5, BLOCK, 2), (6, 301, 2)] { + let data = bytes(seed * 31 + u64::from(bits), 32 << 10, kind); + let groups = q.map_or(3, <[u8]>::len); + let stripe = Stripe { bytes: &data, quantizers: q, width }; + let want = reference_stripe(&stripe, groups * 6, bits, xtrans).map_err(|e| e.to_string()); + let got = decode_stripe(&stripe, groups * 6, bits, xtrans).map_err(|e| e.to_string()); + assert_eq!(got, want, "bits {bits}, xtrans {xtrans}, quantizers {q:?}, width {width}, stream kind {kind}"); + finished[table] += usize::from(want.is_ok() && kind != 1); + } + } + } + } + eprintln!("dense streams decoded to the end per quantizer table (of 30): {finished:?}"); + assert!(finished.iter().all(|&n| n > 0), "some quantizer table never decodes to the end: {finished:?}"); + } + + /// Optional: the CC0 files of docs/raf-corpus.sha256 in the directory named by `LIGHTCRAFT_RAF_CORPUS` (file + /// names as listed there, e.g. `raf-fuji-xt5-6122.raf`). Each present file's SHA-256 is checked, then the + /// decoded sensor samples are compared with the hash recorded from the decoder before the optimisation. + #[test] + fn optional_cc0_samples_decode_as_recorded() { + let Some(directory) = std::env::var_os("LIGHTCRAFT_RAF_CORPUS") else { return }; + let list = include_str!("../docs/raf-corpus.sha256"); + for line in list.lines().filter(|l| !l.starts_with('#') && !l.trim().is_empty()) { + let mut columns = line.split_whitespace(); + let (Some(digest), Some(path)) = (columns.next(), columns.next()) else { panic!("bad corpus row {line:?}") }; + let name = path.rsplit('/').next().unwrap(); + let Ok(bytes) = std::fs::read(std::path::Path::new(&directory).join(name)) else { continue }; + let sha = Sha256::digest(&bytes).iter().map(|b| format!("{b:02x}")).collect::(); + assert_eq!(sha, digest, "{name}: not the published file"); + let Some(&(_, expected)) = RECORDED.iter().find(|(n, _)| *n == name) else { panic!("{name}: no recorded hash") }; + let image = crate::decode(&bytes).unwrap(); + let crate::RawData::U16(data) = &image.data else { panic!("{name}: not 16-bit samples") }; + let hash = data.iter().flat_map(|v| v.to_le_bytes()).fold(0xcbf29ce484222325u64, |h, b| (h ^ u64::from(b)).wrapping_mul(0x100000001b3)); + assert_eq!(hash, expected, "{name}"); + } + } + + /// FNV-1a over the little-endian decoded samples, recorded from `upstream/main` before the change. + const RECORDED: &[(&str, u64)] = &[ + ("raf-fuji-gfx100-3773.raf", 0x1195d3ab70da7cb4), + ("raf-fuji-gfx100-3775.raf", 0xb7c7c1a7717ef93c), + ("raf-fuji-gfx100rf-8091.raf", 0x49110f1b80eefe2e), + ("raf-fuji-gfx100s-4495.raf", 0x84ab46122f72796c), + ("raf-fuji-gfx100s-4503.raf", 0x63debf0afde73e2e), + ("raf-fuji-gfx50s-1435.raf", 0xb3e570a6c88455f5), + ("raf-fuji-xe5-8509.raf", 0x6a7d5cde2832d4f3), + ("raf-fuji-xh2-6001.raf", 0x466a5707ea484f9c), + ("raf-fuji-xh2-6002.raf", 0xbbf043676476925a), + ("raf-fuji-xm5-7748.raf", 0xedd17960a117a553), + ("raf-fuji-xt2-865.raf", 0xe020816ec114436f), + ("raf-fuji-xt20-compressed.raf", 0x85014025edc2cebd), + ("raf-fuji-xt4-3914.raf", 0x7c98e527dbe01f23), + ("raf-fuji-xt4-3918.raf", 0xfb4ba0df34c1c904), + ("raf-fuji-xt5-6122.raf", 0x8f76abd57de79a94), + ("raf-fuji-xt5-6123.raf", 0x0b2784853c31278d), + ("raf-fuji-xt50-7807.raf", 0xa8e2be9cac387ebc), + ]; +} diff --git a/crates/raw/src/tiffraw.rs b/crates/raw/src/tiffraw.rs index 9a5649364..e4c9a26b4 100644 --- a/crates/raw/src/tiffraw.rs +++ b/crates/raw/src/tiffraw.rs @@ -2,7 +2,7 @@ #[cfg(feature = "dng")] use crate::unpack::inflate; -use crate::unpack::{f16_to_f32, f24_to_f32, read_u16s, undo_diff_u16, undo_float_predictor, unpack_lsb, unpack_msb, unpack_words32_msb}; +use crate::unpack::{f16_to_f32, f24_to_f32, read_u16s, undo_diff_u16, undo_float_predictor, unpack_lsb, unpack_msb, unpack_words32_msb_with}; use crate::{Limits, RawData, RawError, Result, ljpeg}; use craft_tiff::image::{Chunk, ImageInfo, chunk_bytes}; use craft_tiff::{ByteOrder, tags::compression as comp}; @@ -18,8 +18,8 @@ pub enum Packing { Lsb, /// Each sample stored in 16 bits in file byte order, regardless of `bits`. Word16, - /// MSB-first bit stream inside 32-bit words stored in file byte order ([`unpack_words32_msb`]); rows start on a - /// word boundary. + /// MSB-first bit stream inside 32-bit words stored in file byte order ([`unpack_words32_msb`](crate::unpack::unpack_words32_msb)); + /// rows start on a word boundary. Words32Msb, } @@ -184,46 +184,72 @@ pub(crate) fn read_image(data: &[u8], info: &ImageInfo, order: ByteOrder, packin .par_iter() .map(|c| decode_chunk(data, info, c, order, packing, ccpp, bits, float, limits).map(|p| (*c, p))) .collect(); - for r in decoded { - // a chunk that cannot be decoded is not a partial image: the file stays preview-only - let (c, px) = r?; - let (cw, ch) = (c.width as usize, c.height as usize); - let (x0, y0) = (c.x as usize, c.y as usize); - if x0 >= w || y0 >= h { - continue; - } - let copy_w = cw.min(w - x0); - let copy_h = ch.min(h - y0); - for y in 0..copy_h { - for x in 0..copy_w { - for s in 0..ccpp { - let src = (y * cw + x) * ccpp + s; - let dst_s = if planar { c.plane as usize } else { s }; - if dst_s >= cpp { - continue; - } - let dst = ((y0 + y) * w + x0 + x) * cpp + dst_s; - match &px { - ChunkPx::U16(v) => { - if let (Some(o), Some(&s)) = (out_u16.get_mut(dst), v.get(src)) { - *o = s; - } - } - ChunkPx::F32(v) => { - if let (Some(o), Some(&s)) = (out_f32.get_mut(dst), v.get(src)) { - *o = s; - } - } - } + // a chunk that cannot be decoded is not a partial image: the file stays preview-only + let decoded = decoded.into_iter().collect::>>()?; + // The copy runs in parallel over bands of output rows: each band takes its rows of every chunk of the + // group that reaches it, in chunk order (so overlapping chunks resolve as a serial copy would). + let row_samples = w * cpp; + let threads = rayon::current_num_threads().max(1); + let band_rows = h.div_ceil(threads * 4).max(1); + if float { + out_f32.par_chunks_mut(band_rows * row_samples).enumerate().for_each(|(bi, band)| { + for (c, px) in &decoded { + if let ChunkPx::F32(v) = px { + place_chunk(band, bi * band_rows, c, v, w, h, cpp, planar); } } - } + }); + } else { + out_u16.par_chunks_mut(band_rows * row_samples).enumerate().for_each(|(bi, band)| { + for (c, px) in &decoded { + if let ChunkPx::U16(v) = px { + place_chunk(band, bi * band_rows, c, v, w, h, cpp, planar); + } + } + }); } decode_from = to; } Ok(if float { RawData::F32(out_f32) } else { RawData::U16(out_u16) }) } +/// Copy the rows of chunk `c` (samples `src`, `c.width` per row, `ccpp` samples per pixel) that fall into the output +/// band `band`, which holds the image rows from `band_y0` on (`w` pixels of `cpp` samples each, `h` rows in the whole +/// image). Pixels past the image edge and samples `src` does not hold are skipped; a planar chunk (`planar`) writes +/// its one sample per pixel into the channel `c.plane`, a chunk of a plane the image does not have is skipped. +#[allow(clippy::too_many_arguments)] +fn place_chunk(band: &mut [T], band_y0: usize, c: &Chunk, src: &[T], w: usize, h: usize, cpp: usize, planar: bool) { + let (cw, ch) = (c.width as usize, c.height as usize); + let (x0, y0) = (c.x as usize, c.y as usize); + if x0 >= w || y0 >= h || cpp == 0 { + return; + } + let plane = c.plane as usize; + if planar && plane >= cpp { + return; + } + let ccpp = if planar { 1 } else { cpp }; + let copy_w = cw.min(w - x0); + let copy_h = ch.min(h - y0); + let band_rows = band.len() / (w * cpp); + let first = y0.max(band_y0); + let last = (y0 + copy_h).min(band_y0 + band_rows); + for y in first..last { + let s0 = (y - y0) * cw * ccpp; + let Some(avail) = src.get(s0..) else { break }; + let d0 = ((y - band_y0) * w + x0) * cpp; + if planar { + let row = &mut band[d0..d0 + copy_w * cpp]; + for (o, &v) in row[plane..].iter_mut().step_by(cpp).zip(&avail[..copy_w.min(avail.len())]) { + *o = v; + } + } else { + let n = (copy_w * cpp).min(avail.len()); + band[d0..d0 + n].copy_from_slice(&avail[..n]); + } + } +} + /// Framing precedes allocation even in header-only probes: missing source rows may not /// be reported as a decodable image and later silently filled with zero samples. fn check_uncompressed_chunks(data: &[u8], info: &ImageInfo, chunks: &[Chunk], packing: Packing) -> Result<()> { @@ -512,13 +538,14 @@ fn unpack_chunk( let row_bytes = row_n * bp; let mut out = vec![0f32; row_n * ch]; let mut rowbuf = vec![0u8; row_bytes]; + let mut scratch = Vec::new(); for y in 0..ch { let s = src.get(y * row_bytes..).unwrap_or(&[]); let len = s.len().min(row_bytes); rowbuf[..len].copy_from_slice(&s[..len]); rowbuf[len..].fill(0); let big = if fp { - undo_float_predictor(&mut rowbuf, row_n, bp, cpp * factor); + undo_float_predictor(&mut rowbuf, row_n, bp, cpp * factor, &mut scratch); true } else { if factor > 0 { @@ -526,16 +553,15 @@ fn unpack_chunk( } order == ByteOrder::Big }; - for (i, o) in out[y * row_n..(y + 1) * row_n].iter_mut().enumerate() { - let b = &rowbuf[i * bp..(i + 1) * bp]; - *o = match (bp, big) { - (2, true) => f16_to_f32(u16::from_be_bytes([b[0], b[1]])), - (2, false) => f16_to_f32(u16::from_le_bytes([b[0], b[1]])), - (3, true) => f24_to_f32(u32::from_be_bytes([0, b[0], b[1], b[2]])), - (3, false) => f24_to_f32(u32::from_le_bytes([b[0], b[1], b[2], 0])), - (4, true) => f32::from_be_bytes([b[0], b[1], b[2], b[3]]), - _ => f32::from_le_bytes([b[0], b[1], b[2], b[3]]), - }; + let dst = &mut out[y * row_n..(y + 1) * row_n]; + let samples = rowbuf.chunks_exact(bp); + match (bp, big) { + (2, true) => dst.iter_mut().zip(samples).for_each(|(o, b)| *o = f16_to_f32(u16::from_be_bytes([b[0], b[1]]))), + (2, false) => dst.iter_mut().zip(samples).for_each(|(o, b)| *o = f16_to_f32(u16::from_le_bytes([b[0], b[1]]))), + (3, true) => dst.iter_mut().zip(samples).for_each(|(o, b)| *o = f24_to_f32(u32::from_be_bytes([0, b[0], b[1], b[2]]))), + (3, false) => dst.iter_mut().zip(samples).for_each(|(o, b)| *o = f24_to_f32(u32::from_le_bytes([b[0], b[1], b[2], 0]))), + (4, true) => dst.iter_mut().zip(samples).for_each(|(o, b)| *o = f32::from_be_bytes([b[0], b[1], b[2], b[3]])), + _ => dst.iter_mut().zip(samples).for_each(|(o, b)| *o = f32::from_le_bytes([b[0], b[1], b[2], b[3]])), } } return Ok(ChunkPx::F32(out)); @@ -544,6 +570,7 @@ fn unpack_chunk( return Err(RawError::Unsupported("floating-point predictor on integer data".into())); } let mut out = vec![0u16; row_n * ch]; + let mut words = Vec::new(); let (row_bytes, kind) = match (bits, packing) { (8, _) => (row_n, 0), (16, _) | (_, Packing::Word16) => (row_n * 2, 1), @@ -559,7 +586,7 @@ fn unpack_chunk( 0 => row.iter_mut().zip(s).for_each(|(o, &b)| *o = b as u16), 1 => read_u16s(s, order, row), 2 => unpack_msb(s, bits, row), - 4 => unpack_words32_msb(s, order, bits, row), + 4 => unpack_words32_msb_with(s, order, bits, row, &mut words), _ => unpack_lsb(s, bits, row), } if factor > 0 { @@ -581,6 +608,86 @@ fn unpack_chunk( Ok(ChunkPx::U16(out)) } +#[cfg(test)] +mod place_tests { + use super::*; + + /// The serial per-sample copy `place_chunk` replaced: the oracle. + #[allow(clippy::too_many_arguments)] + fn reference(out: &mut [u16], c: &Chunk, v: &[u16], w: usize, h: usize, cpp: usize, planar: bool) { + let ccpp = if planar { 1 } else { cpp }; + let (cw, ch) = (c.width as usize, c.height as usize); + let (x0, y0) = (c.x as usize, c.y as usize); + if x0 >= w || y0 >= h { + return; + } + for y in 0..ch.min(h - y0) { + for x in 0..cw.min(w - x0) { + for s in 0..ccpp { + let src = (y * cw + x) * ccpp + s; + let dst_s = if planar { c.plane as usize } else { s }; + if dst_s >= cpp { + continue; + } + let dst = ((y0 + y) * w + x0 + x) * cpp + dst_s; + if let (Some(o), Some(&s)) = (out.get_mut(dst), v.get(src)) { + *o = s; + } + } + } + } + } + + fn chunk(x: u32, y: u32, width: u32, height: u32, plane: u16) -> Chunk { + Chunk { index: 0, x, y, width, height, plane, offset: 0, len: 0 } + } + + /// Chunks that overlap (the later one wins), run past the image edge, hold fewer samples than their size, or name a + /// plane the image does not have, placed through bands of several heights, equal the serial copy. + #[test] + fn place_chunk_matches_the_serial_copy_for_odd_chunks() { + let (w, h) = (11, 9); + for cpp in [1usize, 3] { + for planar in [false, true] { + let ccpp = if planar { 1 } else { cpp }; + let specs = [ + (0, 0, 8, 6, 0, 1u16), + (4, 3, 8, 8, 0, 2), // overlaps the first and runs past both edges + (4, 3, 8, 8, 0, 3), // the same place again, later wins + (0, 5, 5, 4, 1, 4), + (2, 2, 6, 6, 2, 5), + (0, 0, 4, 4, 3, 6), // plane 3: not an image plane for cpp <= 3 + (9, 0, 6, 12, 0, 7), // only two columns inside + (20, 0, 4, 4, 0, 8), // right of the image + (0, 0, 11, 9, 0, 9), // whole image, short source below + ]; + let chunks: Vec<(Chunk, Vec)> = specs + .iter() + .map(|&(x, y, cw, ch, plane, val)| { + let n = (cw * ch) as usize * ccpp; + // the last chunk holds only 40 samples + let n = if val == 9 { 40 } else { n }; + (chunk(x, y, cw, ch, plane), (0..n).map(|i| val * 100 + i as u16).collect()) + }) + .collect(); + let mut want = vec![0u16; w * h * cpp]; + for (c, v) in &chunks { + reference(&mut want, c, v, w, h, cpp, planar); + } + for band_rows in [1usize, 2, 4, 9, 20] { + let mut got = vec![0u16; w * h * cpp]; + for (bi, band) in got.chunks_mut(band_rows * w * cpp).enumerate() { + for (c, v) in &chunks { + place_chunk(band, bi * band_rows, c, v, w, h, cpp, planar); + } + } + assert_eq!(got, want, "cpp {cpp} planar {planar} band rows {band_rows}"); + } + } + } + } +} + #[cfg(test)] #[cfg(feature = "dng")] mod lossy_tests { diff --git a/crates/raw/src/unpack.rs b/crates/raw/src/unpack.rs index fd749c386..27d77b55e 100644 --- a/crates/raw/src/unpack.rs +++ b/crates/raw/src/unpack.rs @@ -46,12 +46,18 @@ pub fn unpack_lsb(src: &[u8], bits: u32, out: &mut [u16]) { /// data, which is not the byte-wise MSB stream of [`unpack_msb`] unless `order` is big-endian. A short last word /// is padded with zeros. pub fn unpack_words32_msb(src: &[u8], order: ByteOrder, bits: u32, out: &mut [u16]) { - let words = src.chunks(4).flat_map(|c| { + unpack_words32_msb_with(src, order, bits, out, &mut Vec::new()); +} + +/// [`unpack_words32_msb`] with the byte-swapped copy of `src` kept in `scratch` (reused across rows). +pub fn unpack_words32_msb_with(src: &[u8], order: ByteOrder, bits: u32, out: &mut [u16], scratch: &mut Vec) { + scratch.clear(); + scratch.extend(src.chunks(4).flat_map(|c| { let mut w = [0u8; 4]; w[..c.len()].copy_from_slice(c); order.u32(w).to_be_bytes() - }); - unpack_msb(&words.collect::>(), bits, out); + })); + unpack_msb(scratch, bits, out); } pub fn read_u16s(src: &[u8], order: ByteOrder, out: &mut [u16]) { @@ -126,10 +132,12 @@ pub fn undo_diff_u8(row: &mut [u8], stride: usize) { /// Undo the floating-point predictor on one row: `row` holds `n` samples of `bytes_per` bytes, byte-differenced /// with stride `stride` samples, then stored as byte planes (most significant first). Returns the samples as -/// big-endian byte groups in place. -pub fn undo_float_predictor(row: &mut [u8], n: usize, bytes_per: usize, stride: usize) { +/// big-endian byte groups in place. `scratch` holds the byte planes (reused across rows). +pub fn undo_float_predictor(row: &mut [u8], n: usize, bytes_per: usize, stride: usize, scratch: &mut Vec) { undo_diff_u8(row, stride); - let tmp = row.to_vec(); + scratch.clear(); + scratch.extend_from_slice(row); + let tmp = &scratch[..]; for i in 0..n { for b in 0..bytes_per { if let (Some(dst), Some(&s)) = (row.get_mut(i * bytes_per + b), tmp.get(b * n + i)) { @@ -167,6 +175,22 @@ mod tests { assert_eq!(o, [0x000, 0x012]); // the word 0x00001234, top bits first } + #[test] + fn words32_scratch_is_reusable() { + let long: Vec = (0u8..24).map(|i| i.wrapping_mul(37).wrapping_add(5)).collect(); + let short = &long[3..10]; + let mut scratch = Vec::new(); + for order in [ByteOrder::Little, ByteOrder::Big] { + for (src, n) in [(&long[..], 16usize), (short, 4)] { + let mut want = vec![0u16; n]; + unpack_words32_msb(src, order, 12, &mut want); + let mut got = vec![0u16; n]; + unpack_words32_msb_with(src, order, 12, &mut got, &mut scratch); + assert_eq!(got, want); + } + } + } + #[test] fn packing() { // 12-bit MSB: 0xABC, 0xDEF -> AB CD EF @@ -210,7 +234,7 @@ mod tests { for i in (1..row.len()).rev() { row[i] = row[i].wrapping_sub(row[i - 1]); } - undo_float_predictor(&mut row, 2, 4, 1); + undo_float_predictor(&mut row, 2, 4, 1, &mut Vec::new()); assert_eq!(f32::from_be_bytes([row[0], row[1], row[2], row[3]]), 1.0); assert_eq!(f32::from_be_bytes([row[4], row[5], row[6], row[7]]), 2.0); } diff --git a/crates/raw/tests/it/dng.rs b/crates/raw/tests/it/dng.rs index 216937862..43b0d8a94 100644 --- a/crates/raw/tests/it/dng.rs +++ b/crates/raw/tests/it/dng.rs @@ -403,6 +403,172 @@ fn floating_point_dngs() { assert_eq!(decode(&dng(raw, ByteOrder::Big)).unwrap().data, RawData::F32(vals)); } +/// The chunks (tiles of `tw` x `th`, or strips when `tw == w`) of an image of `cpp` interleaved samples per pixel, each +/// nominal-sized (strips: clipped to the image height) with zero padding past the image edge: one chunk per tile, and for `planar` one set of chunks per +/// plane (one sample per pixel), plane after plane. +fn chunk_samples(px: &[T], w: usize, h: usize, cpp: usize, tw: usize, th: usize, planar: bool) -> Vec> { + let planes = if planar { cpp } else { 1 }; + let ccpp = if planar { 1 } else { cpp }; + let mut out = Vec::new(); + for plane in 0..planes { + for ty in 0..h.div_ceil(th) { + for tx in 0..w.div_ceil(tw) { + // a strip holds only the rows inside the image, a tile its nominal size + let rows = if tw == w { th.min(h - ty * th) } else { th }; + let mut c = vec![T::default(); tw * rows * ccpp]; + for y in 0..th.min(h - ty * th) { + for x in 0..tw.min(w - tx * tw) { + for s in 0..ccpp { + let src = ((ty * th + y) * w + tx * tw + x) * cpp + if planar { plane } else { s }; + c[(y * tw + x) * ccpp + s] = px[src]; + } + } + } + out.push(c); + } + } + } + out +} + +fn layout_image(raw: &mut IfdBuilder, chunks: Vec>, w: usize, tile: (usize, usize)) { + raw.set_image(if tile.0 == w { + ImageData::Strips { rows_per_strip: tile.1 as u32, strips: chunks } + } else { + ImageData::Tiles { tile_width: tile.0 as u32, tile_height: tile.1 as u32, tiles: chunks } + }); +} + +/// Uncompressed integer DNGs in every chunk layout (chunky and planar, tiles and strips with edge chunks, several chunks +/// per image) at several bit depths and byte orders decode to the samples they were built from. +#[test] +fn uncompressed_chunk_layouts_place_every_sample() { + let (w, h) = (37, 23); + for cpp in [1usize, 3] { + for bits in [8u32, 12, 16] { + let px = pattern(w, h, cpp, bits); + // tiles overhanging both edges, strips with a shorter last strip + for tile in [(16, 16), (w, 5), (w, h), (16, 32)] { + for planar in [false, true] { + if planar && cpp == 1 { + continue; + } + for order in [ByteOrder::Little, ByteOrder::Big] { + let ccpp = if planar { 1 } else { cpp }; + let chunks: Vec> = chunk_samples(&px, w, h, cpp, tile.0, tile.1, planar) + .iter() + .map(|c| match bits { + 8 => c.iter().map(|&v| v as u8).collect(), + 16 => u16_bytes(c, order), + _ => pack_msb(c, bits, tile.0 * ccpp), + }) + .collect(); + let mut raw = base_ifd(w, h, bits as u16, cpp, cpp == 1); + raw.set(t::COMPRESSION, Value::Short(vec![1])); + if planar { + raw.set(t::PLANAR_CONFIGURATION, Value::Short(vec![2])); + } + layout_image(&mut raw, chunks, w, tile); + let img = decode(&dng(raw, order)).unwrap(); + assert_eq!(img.cpp, cpp); + assert_eq!(img.data, RawData::U16(px.clone()), "cpp {cpp} bits {bits} tile {tile:?} planar {planar} {order:?}"); + } + } + } + } + } +} + +/// Deflate with the integer and floating-point predictors, chunky and planar tiles and strips. +#[test] +fn deflate_chunk_layouts_place_every_sample() { + let (w, h) = (37, 23); + for planar in [false, true] { + let cpp = 3; + let ccpp = if planar { 1 } else { cpp }; + let px = pattern(w, h, cpp, 16); + for tile in [(16, 16), (w, 5)] { + // integer predictor 2 on 16-bit samples + let chunks: Vec> = chunk_samples(&px, w, h, cpp, tile.0, tile.1, planar) + .into_iter() + .map(|mut c| { + for row in c.chunks_mut(tile.0 * ccpp) { + for i in (ccpp..row.len()).rev() { + row[i] = row[i].wrapping_sub(row[i - ccpp]); + } + } + miniz_oxide::deflate::compress_to_vec_zlib(&u16_bytes(&c, ByteOrder::Little), 6) + }) + .collect(); + let mut raw = base_ifd(w, h, 16, cpp, false); + raw.set(t::COMPRESSION, Value::Short(vec![compression::ADOBE_DEFLATE])); + raw.set(t::PREDICTOR, Value::Short(vec![2])); + if planar { + raw.set(t::PLANAR_CONFIGURATION, Value::Short(vec![2])); + } + layout_image(&mut raw, chunks, w, tile); + let img = decode(&dng(raw, ByteOrder::Little)).unwrap(); + assert_eq!(img.data, RawData::U16(px.clone()), "integer predictor, tile {tile:?} planar {planar}"); + } + } +} + +/// Floating-point DNGs (uncompressed and with the floating-point predictor) in tiled and strip, chunky and planar +/// layouts with several chunks and edge chunks. +#[test] +fn float_chunk_layouts_place_every_sample() { + let (w, h) = (37, 23); + for cpp in [1usize, 3] { + let vals: Vec = (0..w * h * cpp).map(|i| ((i * 37) % 200) as f32 / 64.0).collect(); + for tile in [(16, 16), (w, 5)] { + for planar in [false, true] { + if planar && cpp == 1 { + continue; + } + let ccpp = if planar { 1 } else { cpp }; + let parts = chunk_samples(&vals, w, h, cpp, tile.0, tile.1, planar); + for bits in [16u16, 24, 32] { + let bp = bits as usize / 8; + for (pred, order) in [(1u16, ByteOrder::Little), (1, ByteOrder::Big), (3, ByteOrder::Little)] { + if pred == 1 && bits != 16 && bits != 32 { + continue; // uncompressed 24-bit floats are not written by this test + } + let chunks: Vec> = parts + .iter() + .map(|c| { + if pred == 3 { + let enc = fp_encode(c, bp, ccpp, tile.0 * ccpp); + miniz_oxide::deflate::compress_to_vec_zlib(&enc, 6) + } else if bits == 16 { + c.iter() + .flat_map( + |&v| if order == ByteOrder::Big { f32_to_f16(v).to_be_bytes() } else { f32_to_f16(v).to_le_bytes() }, + ) + .collect() + } else { + c.iter().flat_map(|&v| if order == ByteOrder::Big { v.to_be_bytes() } else { v.to_le_bytes() }).collect() + } + }) + .collect(); + let mut raw = base_ifd(w, h, bits, cpp, cpp == 1); + raw.set(t::SAMPLE_FORMAT, Value::Short(vec![3; cpp])); + raw.set(t::COMPRESSION, Value::Short(vec![if pred == 3 { compression::DEFLATE } else { 1 }])); + if pred == 3 { + raw.set(t::PREDICTOR, Value::Short(vec![3])); + } + if planar { + raw.set(t::PLANAR_CONFIGURATION, Value::Short(vec![2])); + } + layout_image(&mut raw, chunks, w, tile); + let img = decode(&dng(raw, order)).unwrap(); + assert_eq!(img.data, RawData::F32(vals.clone()), "cpp {cpp} {bits}-bit pred {pred} tile {tile:?} planar {planar} {order:?}"); + } + } + } + } + } +} + #[test] fn non_finite_float_samples_decode_as_levels() { // NaN samples in a float CFA DNG spread through any arithmetic downstream: the decoder hands on numbers only diff --git a/crates/rtf/src/doc/export.rs b/crates/rtf/src/doc/export.rs index daa3afbda..0cd34498c 100644 --- a/crates/rtf/src/doc/export.rs +++ b/crates/rtf/src/doc/export.rs @@ -659,6 +659,10 @@ impl Writer { Some(Align::Distribute) => out.push_str("\\qd"), _ => out.push_str("\\ql"), } + // The paragraph's text direction, as Word writes it (`\rtlpar`; `\ltrpar` is the default). + if p.rtl { + out.push_str("\\rtlpar"); + } if p.page_break { out.push_str("\\pagebb"); } diff --git a/crates/rtf/src/doc/state.rs b/crates/rtf/src/doc/state.rs index 486b94a32..c27758d05 100644 --- a/crates/rtf/src/doc/state.rs +++ b/crates/rtf/src/doc/state.rs @@ -148,6 +148,8 @@ pub(super) struct St { pub(super) hidden: bool, // Paragraph properties. pub(super) align: Option, + /// `\rtlpar`: the paragraph reads right to left. + pub(super) rtldir: Option, pub(super) style: i32, pub(super) intbl: bool, pub(super) ls: Option, @@ -185,6 +187,7 @@ impl Default for St { uc: 1, hidden: false, align: None, + rtldir: None, style: 0, intbl: false, ls: None, diff --git a/crates/rtf/src/doc/styles.rs b/crates/rtf/src/doc/styles.rs index 014a105b4..6047bf3c4 100644 --- a/crates/rtf/src/doc/styles.rs +++ b/crates/rtf/src/doc/styles.rs @@ -30,6 +30,7 @@ impl Reader { out.italic = out.italic.or(d.italic); out.size = out.size.or(d.size); out.font = out.font.or(d.font); + out.rtl = out.rtl.or(d.rtl); cur = d.based_on; } out @@ -87,6 +88,8 @@ impl Reader { pub(super) fn para_props(&self, st: &St) -> Para { let mut p = Para::new(self.para_kind(st)); p.align = st.align; + // Direct `\rtlpar`/`\ltrpar` wins over the style's direction. + p.rtl = st.rtldir.or_else(|| self.file_style(st.style).rtl).unwrap_or(false); p.page_break = st.pagebb; p.spacing = spacing(st); p.borders = st.borders.any_visible().then_some(st.borders); @@ -135,6 +138,8 @@ pub(super) struct StyleDef { pub(super) size: Option, /// A `\fonttbl` index, named when it's used (the table may follow). pub(super) font: Option, + /// `\rtlpar` (`Some(true)`) or `\ltrpar`: the paragraph direction the style sets. + pub(super) rtl: Option, } /// Clear from `f` the formatting a paragraph of kind `kind` gets from its style, whose @@ -203,6 +208,10 @@ pub(super) fn is_format_word(w: &str) -> bool { | "s" | "outlinelevel" | "pagebb" + | "rtlpar" + | "ltrpar" + | "rtldir" + | "ltrdir" | "intbl" | "ls" | "ilvl" diff --git a/crates/rtf/src/doc/tests.rs b/crates/rtf/src/doc/tests.rs index 003b32872..2a4103eb1 100644 --- a/crates/rtf/src/doc/tests.rs +++ b/crates/rtf/src/doc/tests.rs @@ -38,6 +38,86 @@ fn basic_reader() { assert!(matches!(&l.inlines[0], Inline::Text(_, f) if f.link.as_deref() == Some("http://a.b"))); } +#[test] +fn rtl_paragraphs_read_and_write_their_direction() { + // Word writes an RTL paragraph as `\rtlpar` (`\ltrpar` for LTR); reading it keeps the + // direction as w:bidi on the document paragraph, and export writes it back. + let rtf = br#"{\rtf1\ansi\pard\rtlpar RTL\par\pard\ltrpar LTR\par\pard plain\par}"#; + let flow = parse(rtf).unwrap(); + let paras: Vec<&Para> = flow.blocks.iter().filter_map(|b| if let FBlock::Para(p) = b { Some(p) } else { None }).collect(); + assert!(paras[0].rtl, "rtlpar sets the direction"); + assert!(!paras[1].rtl && !paras[2].rtl, "ltrpar and no word: not RTL"); + // \rtldir and \ltrdir, which some writers used for paragraphs, are read too. + let old = parse(br#"{\rtf1\ansi\pard\rtldir RTL\par\pard\ltrdir LTR\par}"#).unwrap(); + let old: Vec<&Para> = old.blocks.iter().filter_map(|b| if let FBlock::Para(p) = b { Some(p) } else { None }).collect(); + assert!(old[0].rtl && !old[1].rtl, "rtldir and ltrdir are read"); + // \pard resets the direction. + let reset = parse(br#"{\rtf1\ansi\pard\rtlpar A\par\pard B\par}"#).unwrap(); + let reset: Vec<&Para> = reset.blocks.iter().filter_map(|b| if let FBlock::Para(p) = b { Some(p) } else { None }).collect(); + assert!(reset[0].rtl && !reset[1].rtl, "pard resets the direction"); + // Import keeps it as bidi on the document paragraph. + let bidi = |d: &Document| { + d.body.iter().filter_map(|b| if let Block::Para(p) = &**b { Some(p.props.bidi.unwrap_or(false)) } else { None }).collect::>() + }; + let doc = import(rtf).unwrap(); + assert_eq!(bidi(&doc), [true, false, false]); + // Export writes \rtlpar once, and not the \rtldir. + let out = export(&doc); + assert_eq!(out.matches("\\rtlpar").count(), 1, "{out}"); + assert!(!out.contains("rtldir"), "{out}"); + assert_eq!(bidi(&import(out.as_bytes()).unwrap()), [true, false, false]); +} + +#[test] +fn rtl_paragraph_styles_carry_their_direction() { + // Word often writes the direction in the style: `{\s1\rtlpar RTL;}` then `\pard\s1`. + let rtf = br#"{\rtf1\ansi{\stylesheet{\s0 Normal;}{\s1\rtlpar RTL;}{\s2\sbasedon1 Child;}{\s3\sbasedon1\ltrpar Back;}} +\pard\s1 A\par\pard\s2 B\par\pard\s3 C\par\pard\s1\ltrpar D\par\pard\s0\rtlpar E\par\pard F\par}"#; + let flow = parse(rtf).unwrap(); + let rtl: Vec = flow.blocks.iter().filter_map(|b| if let FBlock::Para(p) = b { Some(p.rtl) } else { None }).collect(); + // A: style. B: inherited from the base style. C: the style overrides its base. D: direct + // \ltrpar beats the style. E: direct \rtlpar on Normal. F: \pard drops direct formatting. + assert_eq!(rtl, [true, true, false, false, true, false]); +} + +#[test] +fn oldpprops_keep_the_old_paragraph_direction() { + // A tracked change of direction: now LTR, before RTL (`\rtlpar` in the old properties). + let rtf = br"{\rtf1\ansi{\*\revtbl {Unknown;}{Alice;}} +\pard\ltrpar\prauth1\prdate-2014675576{\*\oldpprops \pard \rtlpar}A\par}"; + let flow = parse(rtf).unwrap(); + let Some(FBlock::Para(p)) = flow.blocks.first() else { panic!("no paragraph") }; + assert!(!p.rtl); + assert!(p.fmt_change.as_deref().expect("paragraph change").old.rtl, "old direction is RTL"); + let doc = import(rtf).unwrap(); + let c = doc_paras(&doc)[0].props.fmt_change.as_deref().expect("change"); + assert_eq!(c.old.bidi, Some(true)); +} + +#[test] +fn tracked_rtl_to_ltr_change_from_a_document_keeps_the_old_direction() { + use craft_wordmodel::props::{CharProps, ParaProps, PropChange}; + use craft_wordmodel::{Revision, RevisionKind}; + // Document origin (not RTF): now LTR, before RTL. + let mut p = craft_wordmodel::Paragraph::with_text("A", CharProps::default()); + p.props.bidi = Some(false); + p.props.fmt_change = PropChange::boxed(0, ParaProps { bidi: Some(true), ..Default::default() }); + let doc = doc_with(vec![p], vec![Revision::new(RevisionKind::Format, "Bob", "2025-01-02T08:05:00Z")]); + let flow = model::from_doc_tracked(&doc); + let Some(FBlock::Para(fp)) = flow.blocks.first() else { panic!("no paragraph") }; + assert!(!fp.rtl); + assert!(fp.fmt_change.as_deref().expect("change").old.rtl, "snapshot keeps RTL"); + let rtf = export(&doc); + assert!(rtf.contains("{\\*\\oldpprops"), "{rtf}"); + let back = import(rtf.as_bytes()).unwrap(); + let p = doc_paras(&back)[0]; + assert_ne!(p.props.bidi, Some(true), "current direction is LTR"); + let change = p.props.fmt_change.as_deref().expect("change read back"); + assert_eq!(change.old.bidi, Some(true)); + // Reject restores RTL. + assert_eq!(change.old.formatting().bidi, Some(true)); +} + #[test] fn stylesheet_names_take_unicode_escapes() { // \u116?=t \u105?=i \u116?=t \u108?=l \u101?=e — how Word writes a non-Latin style name. @@ -1863,3 +1943,46 @@ fn compound_borders_round_trip() { let back = import(export(&d).as_bytes()).unwrap(); assert_eq!(tops(&back), want); } + +fn para_aligns(flow: &Flow) -> Vec<(bool, Option)> { + flow.blocks.iter().filter_map(|b| if let FBlock::Para(p) = b { Some((p.rtl, p.align)) } else { None }).collect() +} + +#[test] +fn alignment_words_after_rtlpar_mean_start_and_end() { + // Word reads `\ql`/`\qr` that follow `\rtlpar` as start/end (the model's alignment is + // logical), so `\rtlpar\qr` sits at the right edge and `\rtlpar\ql` at the left edge. + // Words written before `\rtlpar` keep their plain meaning (checked in Word by saving to DOCX: + // `\qr\rtlpar` gives jc=right, i.e. end, in the bidi paragraph; `\ql\rtlpar` gives no jc). + let rtf = br#"{\rtf1\ansi\deff0{\fonttbl{\f0 Arial;}}\pard LTR one.\par \pard\rtlpar\qr A.\par \pard\rtlpar B.\par \pard\ltrpar Back.\par \pard\rtlpar\ql C.\par \pard\rtlpar\qc D.\par \pard\ql\rtlpar E.\par \pard\qr\rtlpar F.\par \pard\rtlpar\ltrpar\qr G.\par}"#; + let got = para_aligns(&parse(rtf).unwrap()); + assert_eq!( + got, + [ + (false, None), + (true, Some(Align::Left)), + (true, None), + (false, None), + (true, Some(Align::Right)), + (true, Some(Align::Center)), + (true, Some(Align::Left)), + (true, Some(Align::Right)), + (false, Some(Align::Right)), + ] + ); +} + +#[test] +fn alignment_survives_an_export_import_round_trip() { + let rtf = br#"{\rtf1\ansi\pard\ql L.\par \pard\qr R.\par \pard\qc C.\par \pard\rtlpar\qr start.\par \pard\rtlpar\ql end.\par \pard\rtlpar\qc mid.\par \pard\rtlpar plain.\par}"#; + let first = para_aligns(&parse(rtf).unwrap()); + let doc = import(rtf).unwrap(); + let out = export(&doc); + let second = para_aligns(&parse(out.as_bytes()).unwrap()); + let norm = |v: Vec<(bool, Option)>| v.into_iter().map(|(r, a)| (r, a.filter(|a| *a != Align::Left))).collect::>(); + assert_eq!(norm(first), norm(second.clone()), "{out}"); + // `\rtlpar\qr` is start (no explicit alignment in the model) and `\rtlpar\ql` is end. + assert_eq!(norm(second)[3..5], [(true, None), (true, Some(Align::Right))], "{out}"); + // The export writes the alignment before the direction, which Word reads as logical start/end. + assert!(out.contains("\\qr\\rtlpar"), "{out}"); +} diff --git a/crates/rtf/src/doc/word.rs b/crates/rtf/src/doc/word.rs index 28360c616..2eb86692f 100644 --- a/crates/rtf/src/doc/word.rs +++ b/crates/rtf/src/doc/word.rs @@ -3,6 +3,18 @@ use super::styles::{StyleDef, is_char_format_word, is_format_word}; use super::{Align, Fmt, Reader, VAlign, model}; impl Reader { + /// The logical alignment for a side word (`Left` for `\ql`, `Right` for `\qr`). The model's + /// alignment is logical (start/end), and Word reads a side word that comes after `\rtlpar` as + /// start/end: the sides swap in a paragraph already marked right to left. Before `\rtlpar` + /// the side keeps its plain meaning, which is why the writer puts the alignment first. + fn side_align(&self, side: Align) -> Align { + match (self.st.rtldir, side) { + (Some(true), Align::Left) => Align::Right, + (Some(true), Align::Right) => Align::Left, + _ => side, + } + } + pub(super) fn word(&mut self, w: &str, p: Option, starred: bool, first_in_group: bool) { // Every control word ends a multibyte run: the code page it is read in may change here // (`\f`, `\fcharset`, `\cpg`, `\ansicpg`, `\pard`), and `\par` must see the text before it. @@ -166,11 +178,15 @@ impl Reader { "cf" => self.st.fmt.color = self.colors.get(pv.max(0) as usize).copied().flatten(), "highlight" | "chcbpat" | "cb" => self.st.fmt.background = self.colors.get(pv.max(0) as usize).copied().flatten(), "v" => self.st.hidden = on, - "ql" => self.st.align = Some(Align::Left), + "ql" => self.st.align = Some(self.side_align(Align::Left)), "qc" => self.st.align = Some(Align::Center), - "qr" => self.st.align = Some(Align::Right), + "qr" => self.st.align = Some(self.side_align(Align::Right)), "qj" => self.st.align = Some(Align::Justify), "qd" => self.st.align = Some(Align::Distribute), + // `\rtldir`/`\ltrdir` are not paragraph words (they set a table's or section's + // direction in RTF); some writers used them for paragraphs, so they are read too. + "rtlpar" | "rtldir" => self.st.rtldir = Some(true), + "ltrpar" | "ltrdir" => self.st.rtldir = Some(false), "s" => self.st.style = pv, "outlinelevel" => self.st.outline = Some(pv.clamp(0, 9) as u8), "pagebb" => self.st.pagebb = on, @@ -417,6 +433,8 @@ impl Reader { "i" => def.italic = Some(on), "fs" => def.size = (value > 0).then(|| (value as f32 / 2.0).clamp(1.0, 1638.0)), "f" => def.font = Some(value), + "rtlpar" | "rtldir" => def.rtl = Some(true), + "ltrpar" | "ltrdir" => def.rtl = Some(false), _ => {} } true 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/tiff/src/classic.rs b/crates/tiff/src/classic.rs new file mode 100644 index 000000000..c1ab6f64a --- /dev/null +++ b/crates/tiff/src/classic.rs @@ -0,0 +1,470 @@ +//! The TIFF container (TIFF 6.0 §2, as DNG uses it): byte order, image file directories (IFDs), +//! their entries and values, and the SubIFD tree. +//! +//! Every offset and count comes from the file and is checked before use: an entry whose value lies +//! outside the file is an error when it is read, the walk visits at most [`MAX_IFDS`] directories, +//! never the same one twice (cyclic chains end), and goes at most [`MAX_DEPTH`] SubIFD levels deep. + +use std::collections::HashSet; + +/// Why a classic TIFF directory could not be read. +#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)] +pub enum Error { + /// The input has no supported classic TIFF header. + #[error("not a classic TIFF")] + NotTiff, + /// The directory or a referenced value is invalid. + #[error("{0}")] + Malformed(&'static str), +} +/// Result of the bounded classic TIFF reader. +pub type Result = std::result::Result; + +/// Most IFDs one file may have (main chain and SubIFDs together). +pub const MAX_IFDS: usize = 64; +/// Deepest SubIFD nesting followed. +pub const MAX_DEPTH: usize = 4; +/// Invalid SubIFD edges count too: otherwise a file can make failed offset probes quadratic. +const MAX_IFD_ATTEMPTS: usize = MAX_IFDS * 16; +/// SubIFD references (zero, repeated and invalid ones included) accepted over the whole walk, +/// counted before a list is read: many directories can name the same long SubIFD table. +const MAX_SUB_IFD_REFS: usize = MAX_IFD_ATTEMPTS; + +/// TIFF field types (TIFF 6.0 §2, plus type 13 IFD from TIFF Technical Note 1). +#[derive(Clone, Copy, Debug, PartialEq, Eq)] +pub enum Type { + Byte, + Ascii, + Short, + Long, + Rational, + SByte, + Undefined, + SShort, + SLong, + SRational, + Float, + Double, + Ifd, +} + +impl Type { + fn from_code(c: u16) -> Option { + Some(match c { + 1 => Type::Byte, + 2 => Type::Ascii, + 3 => Type::Short, + 4 => Type::Long, + 5 => Type::Rational, + 6 => Type::SByte, + 7 => Type::Undefined, + 8 => Type::SShort, + 9 => Type::SLong, + 10 => Type::SRational, + 11 => Type::Float, + 12 => Type::Double, + 13 => Type::Ifd, + _ => return None, + }) + } + + /// Bytes per value. + pub fn size(self) -> usize { + match self { + Type::Byte | Type::Ascii | Type::SByte | Type::Undefined => 1, + Type::Short | Type::SShort => 2, + Type::Long | Type::SLong | Type::Float | Type::Ifd => 4, + Type::Rational | Type::SRational | Type::Double => 8, + } + } +} + +/// One directory entry. +#[derive(Clone, Copy, Debug)] +pub struct Entry { + pub tag: u16, + pub ty: Type, + pub count: u32, + /// Where the value bytes start: inside the entry when they fit in four bytes, else the offset + /// the entry holds (not yet checked against the file length). + value_at: u64, +} + +/// One image file directory. +#[derive(Clone, Debug, Default)] +pub struct Ifd { + /// File offset of the directory. + pub offset: u64, + pub entries: Vec, + /// Nesting: 0 for the main chain, 1 for its SubIFDs, … + pub depth: usize, +} + +impl Ifd { + pub fn get(&self, tag: u16) -> Option<&Entry> { + self.entries.iter().find(|e| e.tag == tag) + } +} + +/// A TIFF file in memory. +#[derive(Clone, Copy)] +pub struct Tiff<'a> { + pub data: &'a [u8], + pub big_endian: bool, +} + +impl<'a> Tiff<'a> { + /// Read the header. Classic TIFF only (BigTIFF, version 43, is not used by DNG). + pub fn new(data: &'a [u8]) -> Result<(Tiff<'a>, u64)> { + let big_endian = match data.get(0..4) { + Some(b"II*\0") => false, + Some(b"MM\0*") => true, + _ => return Err(Error::NotTiff), + }; + let t = Tiff { data, big_endian }; + let first = u64::from(t.u32_at(4)?); + Ok((t, first)) + } + + fn bytes_at(&self, off: u64, len: usize) -> Result<&'a [u8]> { + let start = usize::try_from(off).map_err(|_| Error::Malformed("offset outside the file"))?; + let end = start.checked_add(len).ok_or(Error::Malformed("offset outside the file"))?; + self.data.get(start..end).ok_or(Error::Malformed("offset outside the file")) + } + + pub fn u16_of(&self, b: [u8; 2]) -> u16 { + if self.big_endian { u16::from_be_bytes(b) } else { u16::from_le_bytes(b) } + } + + pub fn u32_of(&self, b: [u8; 4]) -> u32 { + if self.big_endian { u32::from_be_bytes(b) } else { u32::from_le_bytes(b) } + } + + fn u16_at(&self, off: u64) -> Result { + let b = self.bytes_at(off, 2)?; + Ok(self.u16_of([b[0], b[1]])) + } + + fn u32_at(&self, off: u64) -> Result { + let b = self.bytes_at(off, 4)?; + Ok(self.u32_of([b[0], b[1], b[2], b[3]])) + } + + /// Read the IFD at `off`: its entries and the offset of the next IFD (0 at the end of a chain). + pub fn read_ifd(&self, off: u64, depth: usize) -> Result<(Ifd, u64)> { + let n = self.u16_at(off)?; + let mut entries = Vec::with_capacity(usize::from(n)); + for i in 0..u64::from(n) { + let at = off + 2 + i * 12; + let e = self.bytes_at(at, 12)?; + let tag = self.u16_of([e[0], e[1]]); + let code = self.u16_of([e[2], e[3]]); + let count = self.u32_of([e[4], e[5], e[6], e[7]]); + // unknown field types are skipped (TIFF 6.0: readers ignore them) + let Some(ty) = Type::from_code(code) else { continue }; + let len = u64::from(count).saturating_mul(ty.size() as u64); + let value_at = if len <= 4 { at + 8 } else { u64::from(self.u32_of([e[8], e[9], e[10], e[11]])) }; + entries.push(Entry { tag, ty, count, value_at }); + } + let next = u64::from(self.u32_at(off + 2 + u64::from(n) * 12)?); + Ok((Ifd { offset: off, entries, depth }, next)) + } + + /// Every IFD of the file: the main chain from `first` and, depth first, the SubIFDs (tag 330) of + /// each. Bounded by [`MAX_IFDS`] and [`MAX_DEPTH`]; a directory reached twice ends that branch. + pub fn all_ifds(&self, first: u64) -> Result> { + let mut out: Vec = Vec::new(); + let mut seen = HashSet::new(); + let mut attempts = 0usize; + let mut sub_refs = 0usize; + let mut stack: Vec<(u64, usize)> = vec![(first, 0)]; + while let Some((off, depth)) = stack.pop() { + let mut next = off; + while next != 0 { + if !seen.insert(next) { + break; + } + attempts += 1; + if attempts > MAX_IFD_ATTEMPTS { + return Err(Error::Malformed("too many IFD offsets")); + } + if out.len() >= MAX_IFDS { + return Err(Error::Malformed("too many IFDs")); + } + let (ifd, after) = match self.read_ifd(next, depth) { + Ok(x) => x, + // a broken directory ends the chain; the first one must be readable + Err(e) if out.is_empty() => return Err(e), + Err(_) => break, + }; + if depth < MAX_DEPTH + && let Some(sub) = ifd.get(tags::SUB_IFDS) + { + sub_refs = sub_refs.saturating_add(usize::try_from(sub.count).unwrap_or(usize::MAX)); + if sub_refs > MAX_SUB_IFD_REFS { + return Err(Error::Malformed("too many SubIFD references")); + } + if let Ok(offs) = self.uints(sub) { + for o in offs.into_iter().rev().filter(|o| *o != 0 && !seen.contains(o)) { + stack.push((o, depth + 1)); + } + } + } + out.push(ifd); + next = after; + } + } + Ok(out) + } + + /// The value bytes of an entry. + pub fn value_bytes(&self, e: &Entry) -> Result<&'a [u8]> { + let len = (e.count as usize).checked_mul(e.ty.size()).ok_or(Error::Malformed("entry too large"))?; + self.bytes_at(e.value_at, len) + } + + /// The `i`-th value of a numeric entry as f64 (rationals divided; a zero denominator is an + /// error). + fn number(&self, e: &Entry, b: &[u8], i: usize) -> Result { + let s = e.ty.size(); + let v = b.get(i * s..i * s + s).ok_or(Error::Malformed("entry too short"))?; + let u16v = |o: usize| self.u16_of([v[o], v[o + 1]]); + let u32v = |o: usize| self.u32_of([v[o], v[o + 1], v[o + 2], v[o + 3]]); + Ok(match e.ty { + Type::Byte | Type::Undefined => f64::from(v[0]), + Type::SByte => f64::from(v[0] as i8), + Type::Short => f64::from(u16v(0)), + Type::SShort => f64::from(u16v(0) as i16), + Type::Long | Type::Ifd => f64::from(u32v(0)), + Type::SLong => f64::from(u32v(0) as i32), + Type::Rational => { + let d = u32v(4); + if d == 0 { + return Err(Error::Malformed("rational with a zero denominator")); + } + f64::from(u32v(0)) / f64::from(d) + } + Type::SRational => { + let d = u32v(4) as i32; + if d == 0 { + return Err(Error::Malformed("rational with a zero denominator")); + } + f64::from(u32v(0) as i32) / f64::from(d) + } + Type::Float => f64::from(f32::from_bits(u32v(0))), + Type::Double => { + let mut a = [0u8; 8]; + a.copy_from_slice(&v[..8]); + if self.big_endian { f64::from_be_bytes(a) } else { f64::from_le_bytes(a) } + } + Type::Ascii => return Err(Error::Malformed("text where a number belongs")), + }) + } + + /// All values of a numeric entry as f64. Non-finite values are an error. + pub fn floats(&self, e: &Entry) -> Result> { + let b = self.value_bytes(e)?; + let mut out = Vec::with_capacity(e.count as usize); + for i in 0..e.count as usize { + let v = self.number(e, b, i)?; + if !v.is_finite() { + return Err(Error::Malformed("non-finite number")); + } + out.push(v); + } + Ok(out) + } + + /// All values of an unsigned integer entry (BYTE, SHORT, LONG, IFD). + pub fn uints(&self, e: &Entry) -> Result> { + if !matches!(e.ty, Type::Byte | Type::Undefined | Type::Short | Type::Long | Type::Ifd) { + return Err(Error::Malformed("expected an unsigned integer field")); + } + let b = self.value_bytes(e)?; + let s = e.ty.size(); + Ok(b.chunks_exact(s) + .map(|v| match s { + 1 => u64::from(v[0]), + 2 => u64::from(self.u16_of([v[0], v[1]])), + _ => u64::from(self.u32_of([v[0], v[1], v[2], v[3]])), + }) + .collect()) + } + + /// The first value of an unsigned integer entry. + pub fn uint(&self, e: &Entry) -> Result { + self.uints(e)?.first().copied().ok_or(Error::Malformed("empty field")) + } + + /// An ASCII entry up to its first NUL, lossily decoded. + pub fn ascii(&self, e: &Entry) -> Result { + let b = self.value_bytes(e)?; + let end = b.iter().position(|&c| c == 0).unwrap_or(b.len()); + Ok(String::from_utf8_lossy(&b[..end]).trim().to_string()) + } +} + +/// Tag numbers (TIFF 6.0, TIFF/EP, DNG 1.7, CinemaDNG). +pub mod tags { + pub const NEW_SUBFILE_TYPE: u16 = 254; + pub const IMAGE_WIDTH: u16 = 256; + pub const IMAGE_LENGTH: u16 = 257; + pub const BITS_PER_SAMPLE: u16 = 258; + pub const COMPRESSION: u16 = 259; + pub const PHOTOMETRIC: u16 = 262; + pub const MAKE: u16 = 271; + pub const MODEL: u16 = 272; + pub const STRIP_OFFSETS: u16 = 273; + pub const SAMPLES_PER_PIXEL: u16 = 277; + pub const ROWS_PER_STRIP: u16 = 278; + pub const STRIP_BYTE_COUNTS: u16 = 279; + pub const PLANAR_CONFIGURATION: u16 = 284; + pub const PREDICTOR: u16 = 317; + pub const TILE_WIDTH: u16 = 322; + pub const TILE_LENGTH: u16 = 323; + pub const TILE_OFFSETS: u16 = 324; + pub const TILE_BYTE_COUNTS: u16 = 325; + pub const SUB_IFDS: u16 = 330; + pub const SAMPLE_FORMAT: u16 = 339; + pub const CFA_REPEAT_PATTERN_DIM: u16 = 33421; + pub const CFA_PATTERN: u16 = 33422; + pub const DNG_VERSION: u16 = 50706; + pub const UNIQUE_CAMERA_MODEL: u16 = 50708; + pub const CFA_PLANE_COLOR: u16 = 50710; + pub const CFA_LAYOUT: u16 = 50711; + pub const LINEARIZATION_TABLE: u16 = 50712; + pub const BLACK_LEVEL_REPEAT_DIM: u16 = 50713; + pub const BLACK_LEVEL: u16 = 50714; + pub const BLACK_LEVEL_DELTA_H: u16 = 50715; + pub const BLACK_LEVEL_DELTA_V: u16 = 50716; + pub const WHITE_LEVEL: u16 = 50717; + pub const DEFAULT_CROP_ORIGIN: u16 = 50719; + pub const DEFAULT_CROP_SIZE: u16 = 50720; + pub const COLOR_MATRIX_1: u16 = 50721; + pub const COLOR_MATRIX_2: u16 = 50722; + pub const CAMERA_CALIBRATION_1: u16 = 50723; + pub const CAMERA_CALIBRATION_2: u16 = 50724; + pub const ANALOG_BALANCE: u16 = 50727; + pub const AS_SHOT_NEUTRAL: u16 = 50728; + pub const AS_SHOT_WHITE_XY: u16 = 50729; + pub const BASELINE_EXPOSURE: u16 = 50730; + pub const CALIBRATION_ILLUMINANT_1: u16 = 50778; + pub const CALIBRATION_ILLUMINANT_2: u16 = 50779; + pub const ACTIVE_AREA: u16 = 50829; + pub const OPCODE_LIST_1: u16 = 51008; + pub const OPCODE_LIST_2: u16 = 51009; + pub const OPCODE_LIST_3: u16 = 51022; + pub const FORWARD_MATRIX_1: u16 = 50964; + pub const FORWARD_MATRIX_2: u16 = 50965; + /// CinemaDNG: SMPTE ST 12-1 time codes, 8 bytes each. + pub const TIME_CODES: u16 = 51043; + /// CinemaDNG: frames per second (SRATIONAL). + pub const FRAME_RATE: u16 = 51044; + /// CinemaDNG: the camera's reel name. + pub const REEL_NAME: u16 = 51081; +} + +#[cfg(test)] +mod tests { + use super::*; + + /// A little-endian TIFF: a chain of `ifds` directories, each with one SubIFDs entry of + /// `count` LONGs, all naming the same table (of zeros, or of `target`). + fn chain(ifds: usize, count: u32, target: u32) -> Vec { + let mut b = b"II*\0".to_vec(); + b.extend_from_slice(&8u32.to_le_bytes()); + let table = 8 + ifds * 18; + for i in 0..ifds { + let next = if i + 1 < ifds { 8 + (i + 1) * 18 } else { 0 }; + b.extend_from_slice(&1u16.to_le_bytes()); + b.extend_from_slice(&tags::SUB_IFDS.to_le_bytes()); + b.extend_from_slice(&4u16.to_le_bytes()); + b.extend_from_slice(&count.to_le_bytes()); + b.extend_from_slice(&(table as u32).to_le_bytes()); + b.extend_from_slice(&(next as u32).to_le_bytes()); + } + for _ in 0..count.max(2) { + b.extend_from_slice(&target.to_le_bytes()); + } + b + } + + #[test] + fn shared_sub_ifd_tables_are_budgeted_before_they_are_queued() { + // 60 directories each naming one 100 000-entry table: 6 million queued edges before + let b = chain(60, 100_000, 0); + let (t, first) = Tiff::new(&b).unwrap(); + assert!(matches!(t.all_ifds(first), Err(Error::Malformed("too many SubIFD references")))); + } + + #[test] + fn matches_general_reader_for_a_simple_ifd_chain() { + let bytes = chain(3, 0, 0); + let (reader, first) = Tiff::new(&bytes).unwrap(); + let classic = reader.all_ifds(first).unwrap(); + let general = crate::Tiff::parse(&bytes).unwrap(); + assert_eq!(classic.len(), general.ifds.len()); + assert!(classic.iter().all(|ifd| ifd.get(tags::SUB_IFDS).is_some())); + } + + #[test] + fn ordinary_sub_ifds_are_still_walked() { + // one directory whose SubIFD is the same directory: visited once + let b = chain(1, 1, 8); + let (t, first) = Tiff::new(&b).unwrap(); + assert_eq!(t.all_ifds(first).unwrap().len(), 1); + // a short chain with a few zero SubIFD offsets + let b = chain(3, 4, 0); + let (t, first) = Tiff::new(&b).unwrap(); + assert_eq!(t.all_ifds(first).unwrap().len(), 3); + } + #[test] + fn numeric_field_types_and_malformed_values() { + fn field(code: u16, payload: &[u8]) -> Vec { + let mut b = b"II*\0".to_vec(); + b.extend_from_slice(&8u32.to_le_bytes()); + b.extend_from_slice(&1u16.to_le_bytes()); + b.extend_from_slice(&256u16.to_le_bytes()); + b.extend_from_slice(&code.to_le_bytes()); + b.extend_from_slice(&1u32.to_le_bytes()); + if payload.len() <= 4 { + let mut inline = [0; 4]; + inline[..payload.len()].copy_from_slice(payload); + b.extend_from_slice(&inline); + } else { + b.extend_from_slice(&26u32.to_le_bytes()); + } + b.extend_from_slice(&0u32.to_le_bytes()); + if payload.len() > 4 { + b.extend_from_slice(payload); + } + b + } + let cases: &[(u16, &[u8], f64)] = &[ + (1, &[7], 7.0), + (6, &[255], -1.0), + (3, &[2, 0], 2.0), + (8, &[255, 255], -1.0), + (4, &[3, 0, 0, 0], 3.0), + (9, &[255, 255, 255, 255], -1.0), + (5, &[3, 0, 0, 0, 2, 0, 0, 0], 1.5), + (10, &[253, 255, 255, 255, 2, 0, 0, 0], -1.5), + (11, &1.5_f32.to_le_bytes(), 1.5), + (12, &1.5_f64.to_le_bytes(), 1.5), + ]; + for &(code, data, expected) in cases { + let bytes = field(code, data); + let (t, at) = Tiff::new(&bytes).unwrap(); + let (ifd, _) = t.read_ifd(at, 0).unwrap(); + assert_eq!(t.floats(&ifd.entries[0]).unwrap(), [expected], "type {code}"); + } + let bytes = field(5, &[1, 0, 0, 0, 0, 0, 0, 0]); + let (t, at) = Tiff::new(&bytes).unwrap(); + let (ifd, _) = t.read_ifd(at, 0).unwrap(); + assert!(t.floats(&ifd.entries[0]).is_err()); + let bytes = field(11, &f32::NAN.to_le_bytes()); + let (t, at) = Tiff::new(&bytes).unwrap(); + let (ifd, _) = t.read_ifd(at, 0).unwrap(); + assert!(t.floats(&ifd.entries[0]).is_err()); + } +} diff --git a/crates/tiff/src/lib.rs b/crates/tiff/src/lib.rs index bd095f4f0..0e9d595c7 100644 --- a/crates/tiff/src/lib.rs +++ b/crates/tiff/src/lib.rs @@ -22,6 +22,9 @@ #![forbid(unsafe_code)] #![deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable)] +/// Bounded classic TIFF directory reader with lazy value access. +pub mod classic; + pub mod image; pub mod makernote; pub mod packbits; 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/README.md b/crates/video/src/av1/decode/README.md index 1518df92c..e1b8e4732 100644 --- a/crates/video/src/av1/decode/README.md +++ b/crates/video/src/av1/decode/README.md @@ -73,11 +73,17 @@ with libdav1d: **bit-exact** on all 6 fixtures. `Decoder::new()` uses all cores (`Decoder::with_threads(1)` decodes on the calling thread; on wasm32, or without the default `threads` feature, everything runs on the calling thread). Headers are parsed in order; each frame is -then decoded by a frame worker once the reference frames it reads are finished, so independent -frames (hierarchical GOP layers) decode concurrently; tiles of a frame decode in parallel into -private buffers that are merged afterwards; the in-loop filters (deblocking, CDEF, loop -restoration) each run in parallel bands of 64 rows (`postfilter.rs` explains why the bands are -independent). Results never depend on scheduling (the oracle tests compare 1 thread, many threads +then decoded by a frame worker. With worker threads it first reads the symbols of all its tiles, +which needs only the parse results of its references (CDFs, segment ids, saved motion vectors), +publishes its own, and then reconstructs from the recorded blocks. Frames with several tiles +read the symbols of their tiles in parallel (one after another with intra block copy). A frame reconstructs superblock row by superblock +row (that row of every tile), each row waiting only for the reference rows its motion vectors +reach (row wavefront), and the in-loop filters (deblocking, CDEF, loop restoration) follow a few +64-row bands behind and publish each final band for the frames predicting from it +(`decoder.rs` and `postfilter.rs` give the dependency rules). Super-resolution frames wait for +their whole references and filter the whole frame. There is one frame worker per thread (at most +16), and the frames in flight are bounded by their size. Independent frames (hierarchical GOP +layers) decode concurrently. Results never depend on scheduling (the oracle tests compare 1 thread, many threads and libdav1d). With frame threads, pictures may be returned by a later `decode_pts` call or by `flush`; each carries the `pts` of the temporal unit that showed it. @@ -130,6 +136,44 @@ motion vectors nor CDFs reach another frame — skip deblocking, CDEF and loop r (super-resolution still runs, it sets the output size) and are flagged `Picture::draft`. In SVT-AV1's hierarchical GOPs that is the top layer, half of the frames. +Frame threading on a Ryzen 7800X3D (8 cores / 16 threads, release `examples/av1_av1dec`, median of 7 +alternating runs against Filmcraft `6909e461`, 2026-11-08), wall time for the whole file: + +| Fixture | Threads | Before | After | Speed-up | +|---|---|---|---|---| +| `perf_1080p_gop` (60 frames) | 1 | 69.2 fps | 69.1 fps | 1.00x | +| `perf_1080p_gop` | 16 | 159.6 fps | 227.3 fps | 1.42x | +| `perf_2160p_gop` (30 frames) | 1 | 17.9 fps | 17.8 fps | 1.00x | +| `perf_2160p_gop` | 16 | 40.2 fps | 57.7 fps | 1.43x | + +Peak RSS at 16 threads goes from 278 to 365 MB (1080p) and from 1031 to 1217 MB (2160p, +18%): +every frame in flight holds its parse record and its final samples are kept in 64-row bands. + +With one frame worker per thread, the row wavefront for frames with several tiles and parse +waits limited to what parsing reads (same machine, median of 5 alternating runs, 2026-11-08): + +| Fixture | Threads | Before (`6909e461`) | After | Speed-up | Peak RSS before / after | +|---|---|---|---|---|---| +| `perf_1080p_gop` | 4 | 95.7 fps | 185.2 fps | 1.94x | 148 / 236 MB | +| `perf_1080p_gop` | 16 | 155.0 fps | 259.7 fps | 1.68x | 274 / 560 MB | +| `perf_2160p_gop` | 4 | 23.6 fps | 47.1 fps | 2.00x | 512 / 798 MB | +| `perf_2160p_gop` | 16 | 39.9 fps | 65.8 fps | 1.65x | 994 / 1353 MB | +| `perf_1080p_gop_tiles` (4x2 tiles) | 16 | 186.3 fps | 237.2 fps | 1.27x | 328 / 617 MB | +| SVT-AV1 2160p, 4x2 tiles (30 frames) | 16 | 57.1 fps | 70.8 fps | 1.24x | 1197 / 1502 MB | + +Known limitations of the frame threading (the output is the same in every case, only the +overlap is smaller): + +- A frame whose parse record would pass 32 bytes per luma sample of a tile (pathological streams: + tiny blocks, large palettes, every coefficient coded), or that cannot reserve its record from + the decoder-wide budget of 1 GiB shared by the frames in flight, is decoded again in one pass. +- A superblock row using local warped motion, global motion beyond translation or a scaled + reference waits for the whole reference instead of the rows its motion vectors reach. +- Super-resolution frames do not use the row wavefront: they wait for their whole references + and filter the whole frame. +- The tile columns of a superblock row are reconstructed one after another on the frame's worker + (their symbols are still read in parallel). + ## Limits Frames beyond the AV1 level 6.x limits (Annex A: at most 16384 wide, 8704 high and 35 651 584 luma samples diff --git a/crates/video/src/av1/decode/cdef.rs b/crates/video/src/av1/decode/cdef.rs index edd119292..05eb0a2a7 100644 --- a/crates/video/src/av1/decode/cdef.rs +++ b/crates/video/src/av1/decode/cdef.rs @@ -10,19 +10,33 @@ use crate::av1::decode::state::FrameState; /// equals CurrFrame). Bands of 64 luma rows (whole 64x64 filter blocks) are filtered in /// parallel: every block reads only CurrFrame and writes only its own samples. pub(crate) fn apply(fs: &FrameState, pool: &Pool) -> Option { + if !active(fs) { + return None; + } + let mut out = fs.cur.clone(); + let ranges = bands(pool, fs.fh.mi_rows as usize * 4); + filter_bands(fs, pool, out.row_bands(&ranges), &ranges); + Some(out) +} + +/// Whether CDEF filters any block of the frame (it reads only parse results). +pub(crate) fn active(fs: &FrameState) -> bool { let fh = &fs.fh; if fh.coded_lossless || fh.allow_intrabc || !fs.seq.enable_cdef { - return None; + return false; } let (mi_rows, mi_cols) = (fh.mi_rows as usize, fh.mi_cols as usize); let fb_rows = mi_rows.div_ceil(16); let fb_cols = mi_cols.div_ceil(16); - if !(0..fb_rows).any(|fr| (0..fb_cols).any(|fc| fb_skip_mask(fs, fr * 16, fc * 16).is_some())) { - return None; - } - let mut out = fs.cur.clone(); - let ranges = bands(pool, mi_rows * 4); - let mut planes = out.row_bands(&ranges); + (0..fb_rows).any(|fr| (0..fb_cols).any(|fc| fb_skip_mask(fs, fr * 16, fc * 16).is_some())) +} + +/// Filter the bands `planes` covering luma rows `ranges` (whole 64x64 filter block rows), which +/// hold CurrFrame; reads CurrFrame up to 2 rows beyond each band. +pub(crate) fn filter_bands(fs: &FrameState, pool: &Pool, mut planes: Vec<[RowBand; 3]>, ranges: &[(usize, usize)]) { + let (mi_rows, mi_cols) = (fs.fh.mi_rows as usize, fs.fh.mi_cols as usize); + let fb_rows = mi_rows.div_ceil(16); + let fb_cols = mi_cols.div_ceil(16); pool.map_mut(&mut planes, |b, band| { let (start, end) = ranges[b]; let mut buf = [0i16; BUF_LEN]; @@ -34,8 +48,6 @@ pub(crate) fn apply(fs: &FrameState, pool: &Pool) -> Option { } } }); - drop(planes); - Some(out) } /// The filtered 8x8 blocks of the 64x64 filter block at (r, c): bit `(i << 3) | j` is set for diff --git a/crates/video/src/av1/decode/decoder.rs b/crates/video/src/av1/decode/decoder.rs index 943923a3b..4a922cce1 100644 --- a/crates/video/src/av1/decode/decoder.rs +++ b/crates/video/src/av1/decode/decoder.rs @@ -2,31 +2,37 @@ //! (7.20), show-existing-frame and output (7.18). //! //! Threading: headers are parsed in order on the calling thread; each frame then becomes a job -//! that waits for the reference frames it uses and decodes on a frame worker thread, so frames -//! that do not depend on each other decode concurrently. Within a frame, tiles and the -//! post-filter row bands run on a shared worker pool. Every result depends only on finished -//! reference frames, so the output is identical whatever the scheduling. Shown pictures are -//! returned in order; with frame threads they can come out of a later `decode` call (or -//! `flush`), each carrying the `pts` of the temporal unit that showed it. +//! on a frame worker thread. A job first reads the symbols of its tiles (parse), which needs +//! only the parse results of its references (CDFs, segment ids, saved motion vectors), and +//! publishes its own; it then reconstructs. A frame is reconstructed superblock row by +//! superblock row (that row of every tile), each row waiting only for the reference rows its +//! motion vectors reach, with the in-loop filters following a few bands behind and every final +//! band published for the frames that predict from it (row wavefront); super-resolution frames +//! wait for their whole references. So a frame can parse while its references are still being reconstructed, and reconstruct +//! while their bottom is still being filtered. Within a frame, tiles and the post-filter row +//! bands run on a shared worker pool; waits happen only on frame worker threads. Every result +//! depends only on final reference data, so the output is identical whatever the scheduling. +//! Shown pictures are returned in order; with frame threads they can come out of a later +//! `decode` call (or `flush`), each carrying the `pts` of the temporal unit that showed it. use std::collections::VecDeque; use std::sync::{Arc, Condvar, Mutex}; use crate::av1::decode::bits::{BitReader, leb128}; use crate::av1::decode::cdf::Cdfs; -use crate::av1::decode::frame::FrameBuf; +use crate::av1::decode::frame::{BandFrame, FrameBuf, PlaneRef}; use crate::av1::decode::header::{ ColorConfig, FilmGrainParams, FrameHeader, HeaderState, LoopFilterDeltas, RefInfo, SegmentationFeatures, SequenceHeader, default_gm_params, }; use crate::av1::decode::par::Pool; use crate::av1::decode::spec_tables::{ - ALTREF_FRAME, ALTREF2_FRAME, BWDREF_FRAME, DEFAULT_INTRA_FRAME_Y_MODE_CDF, DIV_MULT, GOLDEN_FRAME, INTRA_FRAME, INTRA_ONLY_FRAME, KEY_FRAME, - LAST_FRAME, LAST2_FRAME, MAX_FRAME_DISTANCE, MAX_OFFSET_HEIGHT, MAX_OFFSET_WIDTH, MFMV_STACK_SIZE, NUM_REF_FRAMES, PRIMARY_REF_NONE, - REFMVS_LIMIT, REFS_PER_FRAME, + ALTREF_FRAME, ALTREF2_FRAME, BWDREF_FRAME, DEFAULT_INTRA_FRAME_Y_MODE_CDF, DIV_MULT, GLOBAL_GLOBALMV, GLOBALMV, GOLDEN_FRAME, INTRA_FRAME, + INTRA_ONLY_FRAME, KEY_FRAME, LAST_FRAME, LAST2_FRAME, LOCALWARP, MAX_FRAME_DISTANCE, MAX_OFFSET_HEIGHT, MAX_OFFSET_WIDTH, MFMV_STACK_SIZE, + NUM_REF_FRAMES, PRIMARY_REF_NONE, REFMVS_LIMIT, REFS_PER_FRAME, TRANSLATION, }; use crate::av1::decode::state::{FrameShared, FrameState, TileState}; use crate::av1::decode::stats::{DecodeStats, Stage, Timer}; -use crate::av1::decode::tile::TileDecoder; +use crate::av1::decode::tile::{Pass, TileDecoder, TileRecord}; use crate::av1::decode::{Error, Picture, Result}; /// Largest accepted frame: the AV1 level 6.x limits (Annex A: MaxHSize 16384, MaxVSize 8704, MaxPicSize @@ -49,22 +55,156 @@ const OBU_FRAME_T: u32 = 6; const OBU_REDUNDANT_FRAME_HEADER_T: u32 = 7; const OBU_TILE_LIST_T: u32 = 8; -/// A stored reference frame (FrameStore plus the saved per-frame state). +/// The samples of a stored reference frame (FrameStore), after every in-loop filter: a whole +/// buffer (decoding on one thread) or the bands published while decoding (with worker threads). pub(crate) struct RefFrame { pub buf: FrameBuf, - pub cdfs: Box, + pub bands: Option>, + pub num_planes: usize, pub bit_depth: u8, pub subsampling_x: u8, pub subsampling_y: u8, + pub color: (u8, u8, u8, bool), +} + +impl RefFrame { + pub fn plane(&self, p: usize) -> PlaneRef<'_> { + match &self.bands { + Some(b) => PlaneRef::Bands(&b.planes[p]), + None => PlaneRef::Whole(&self.buf.planes[p]), + } + } +} + +/// The saved per-frame state that later frames read while parsing (CDFs, segment ids and the +/// motion vectors for motion field estimation). It depends only on the frame's symbols, so it +/// is published as soon as the tiles are decoded, before the in-loop filters run. +pub(crate) struct Parsed { + pub cdfs: Box, pub segment_ids: Vec, pub mi_cols: usize, pub mi_rows: usize, - pub color: (u8, u8, u8, bool), - /// SavedRefFrames / SavedMvs (MfRefFrames / MfMvs of 7.19), mi units. + /// SavedRefFrames / SavedMvs (MfRefFrames / MfMvs of 7.19) at the odd row and column of + /// every 8x8 block, the only positions the projection reads. pub saved_ref_frames: Vec, pub saved_mvs: Vec<[i32; 2]>, } +/// How many luma rows of a frame are final (every in-loop filter applied), or that the frame +/// failed. Rows only ever grow; a failure wakes every waiter with an error. +pub(crate) struct Progress { + state: Mutex, + cv: Condvar, +} + +#[derive(Clone, Copy, Default)] +struct ProgressState { + rows: usize, + failed: bool, +} + +impl Progress { + fn new() -> Progress { + Progress { state: Mutex::new(ProgressState::default()), cv: Condvar::new() } + } + + /// Rows [0, rows) are final. A smaller value than already published is ignored. + pub fn publish(&self, rows: usize) { + #[cfg(test)] + tests::jitter(); + let mut g = self.state.lock().unwrap_or_else(|e| e.into_inner()); + if rows > g.rows { + g.rows = rows; + self.cv.notify_all(); + } + } + + /// The frame will publish no more rows. + pub fn fail(&self) { + let mut g = self.state.lock().unwrap_or_else(|e| e.into_inner()); + g.failed = true; + self.cv.notify_all(); + } + + /// Rows published so far. + #[cfg(test)] + fn rows(&self) -> usize { + self.state.lock().map(|g| g.rows).unwrap_or(0) + } + + /// Blocks until rows [0, rows) are final; an error when the frame failed first. + pub fn wait(&self, rows: usize) -> Result<()> { + #[cfg(test)] + tests::jitter(); + let mut g = self.state.lock().unwrap_or_else(|e| e.into_inner()); + loop { + if g.rows >= rows { + return Ok(()); + } + if g.failed { + return Err(Error::Invalid("reference frame failed to decode")); + } + g = self.cv.wait(g).unwrap_or_else(|e| e.into_inner()); + } + } +} + +/// Everything a decoding frame publishes for the frames that reference it and for the output: +/// its parse results, its final rows (bands, then the row progress) and the finished frame. The +/// parse results and the frame are set exactly once and the progress ends complete or failed, +/// also when decoding fails, so no waiter blocks forever. +pub(crate) struct FrameSlot { + parsed: Slot>, + pub progress: Progress, + /// The final samples, published band by band before the progress covers them (with + /// worker threads). + pub bands: Arc, + frame: Slot>, +} + +impl FrameSlot { + fn new(bands: BandFrame) -> Arc { + Arc::new(FrameSlot { parsed: Slot::new_plain(), progress: Progress::new(), bands: Arc::new(bands), frame: Slot::new_plain() }) + } + + /// The finished frame. + fn wait(&self) -> Result> { + self.frame.wait() + } + + fn is_set(&self) -> bool { + self.frame.is_set() + } + + /// The frame's parse results (available before its samples are final). + fn wait_parsed(&self) -> Result> { + self.parsed.wait() + } + + /// Publish the parse results (ignored when already set). + fn set_parsed(&self, v: Result>) { + self.parsed.set_once(v); + } + + /// Publish the finished frame: every row is final. On an error, the parse results (when + /// not yet published) and the progress fail too. + fn finish(&self, v: Result>, rows: usize) { + match &v { + Ok(_) => {} + Err(e) => { + self.parsed.set_once(Err(e.clone())); + } + } + let ok = v.is_ok(); + self.frame.set(v); + if ok { + self.progress.publish(rows); + } else { + self.progress.fail(); + } + } +} + /// A value produced by a decoding job, waited for by later jobs and by the output. pub(crate) struct Slot { value: Mutex>>, @@ -73,7 +213,11 @@ pub(crate) struct Slot { impl Slot { fn new() -> Arc> { - Arc::new(Slot { value: Mutex::new(None), cv: Condvar::new() }) + Arc::new(Slot::new_plain()) + } + + fn new_plain() -> Slot { + Slot { value: Mutex::new(None), cv: Condvar::new() } } fn set(&self, v: Result) { @@ -82,6 +226,15 @@ impl Slot { self.cv.notify_all(); } + /// [`Slot::set`] unless a value is already there. + fn set_once(&self, v: Result) { + let mut g = self.value.lock().unwrap_or_else(|e| e.into_inner()); + if g.is_none() { + *g = Some(v); + self.cv.notify_all(); + } + } + fn is_set(&self) -> bool { self.value.lock().map(|g| g.is_some()).unwrap_or(true) } @@ -108,8 +261,6 @@ impl Slot { } } -type FrameSlot = Slot>; - /// Everything needed to decode one frame, captured when its header is parsed. struct FrameJob { seq: SequenceHeader, @@ -130,6 +281,15 @@ struct FrameJob { draft: bool, #[cfg(test)] force_panic: bool, + /// Panic in the row wavefront at the first superblock row from this one after which rows + /// are published (tests). + #[cfg(test)] + panic_at_row: Option, + /// Overrides [`RECORD_BYTES_PER_SAMPLE`] (tests). + #[cfg(test)] + record_limit: Option, + /// The decoder's parse record budget. + record_budget: Arc, } /// A shown picture in output order. @@ -239,6 +399,9 @@ 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], @@ -251,6 +414,8 @@ pub struct Decoder { draft: bool, stats: Arc>, pool: Pool, + /// Parse record bytes all frames in flight may hold together. + record_budget: Arc, #[cfg(all(feature = "threads", not(target_arch = "wasm32")))] workers: Option, /// Frames submitted to the workers and not known to be finished (oldest first). @@ -265,6 +430,14 @@ pub struct Decoder { /// Make every frame job panic (tests of the worker panic handling). #[cfg(test)] force_panic: bool, + /// (frame number, superblock row): that frame panics in the row wavefront after that row, + /// once it has published rows that later frames may already be using (tests). + #[cfg(test)] + panic_at_row: Option<(usize, usize)>, + #[cfg(test)] + record_limit: Option, + #[cfg(test)] + frame_number: usize, } fn default_threads() -> usize { @@ -278,10 +451,25 @@ fn default_threads() -> usize { } } -/// Frames decoded concurrently with `threads` threads (each in-flight 1080p frame holds a few -/// tens of MB of state). +/// Frame worker threads with `threads` threads. A frame worker spends most of its time waiting +/// for the parse results and rows of its references, so there is one per thread (at least two, +/// at most [`MAX_FRAME_THREADS`]); the memory is bounded by `in_flight_limit` instead. fn frame_threads(threads: usize) -> usize { - if threads <= 1 { 1 } else { threads.div_ceil(2).clamp(2, 8) } + if threads <= 1 { 1 } else { threads.clamp(2, MAX_FRAME_THREADS) } +} + +const MAX_FRAME_THREADS: usize = 16; + +/// Luma samples of the frames in flight (each in-flight frame holds its state, parse record and +/// samples, roughly 10 bytes per luma sample): ten 2160p frames. +#[cfg(all(feature = "threads", not(target_arch = "wasm32")))] +const IN_FLIGHT_SAMPLES: usize = 10 * 3840 * 2160; + +/// Frames submitted to the frame workers that may be unfinished: two more than the workers, +/// fewer for large frames (at least four). +#[cfg(all(feature = "threads", not(target_arch = "wasm32")))] +fn in_flight_limit(frame_threads: usize, luma_samples: usize) -> usize { + (frame_threads + 2).min((IN_FLIGHT_SAMPLES / luma_samples.max(1)).max(4)) } impl Default for Decoder { @@ -301,8 +489,8 @@ 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 +501,14 @@ 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 +537,14 @@ 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 +565,7 @@ 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(), @@ -379,6 +576,7 @@ impl Decoder { draft: false, stats, pool, + record_budget: Arc::new(crate::av1::decode::tile::RecordBudget::new(RECORD_BUDGET_BYTES)), #[cfg(all(feature = "threads", not(target_arch = "wasm32")))] workers, in_flight: VecDeque::new(), @@ -388,6 +586,12 @@ impl Decoder { pts: 0, #[cfg(test)] force_panic: false, + #[cfg(test)] + panic_at_row: None, + #[cfg(test)] + record_limit: None, + #[cfg(test)] + frame_number: 0, } } @@ -564,20 +768,20 @@ 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 => {} @@ -655,6 +859,13 @@ impl Decoder { let picture = fh.show_frame.then(Slot::new); let draft = self.draft && fh.show_frame && fh.refresh_frame_flags == 0; let tile_count = fh.tile_info.cols.checked_mul(fh.tile_info.rows).ok_or(Error::Invalid("tile count overflows"))?; + let slot = FrameSlot::new(BandFrame::new( + fh.upscaled_width as usize, + fh.frame_height as usize, + seq.color.num_planes, + seq.color.subsampling_x as usize, + seq.color.subsampling_y as usize, + )); self.pending = Some(Box::new(FrameJob { seq, fh, @@ -662,14 +873,23 @@ impl Decoder { ref_info: self.refs.clone(), tiles: Vec::new(), tile_covered: vec![false; tile_count], - slot: Slot::new(), + slot, picture, pts: self.pts, apply_film_grain: self.apply_film_grain, draft, #[cfg(test)] force_panic: self.force_panic, + #[cfg(test)] + panic_at_row: self.panic_at_row.filter(|&(f, _)| f == self.frame_number).map(|(_, r)| r), + #[cfg(test)] + record_limit: self.record_limit, + record_budget: self.record_budget.clone(), })); + #[cfg(test)] + { + self.frame_number += 1; + } Ok(None) } @@ -726,7 +946,7 @@ 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); @@ -767,6 +987,8 @@ impl Decoder { .clone() .map(|p| Output::Decoded(p, picture_bytes(job.fh.upscaled_width as usize, job.fh.frame_height as usize, &job.seq.color))); #[cfg(all(feature = "threads", not(target_arch = "wasm32")))] + let fh_samples = (fh.upscaled_width as usize).max(fh.frame_width as usize) * fh.frame_height as usize; + #[cfg(all(feature = "threads", not(target_arch = "wasm32")))] if let Some(w) = &self.workers && let Some(tx) = &w.tx { @@ -775,7 +997,8 @@ impl Decoder { Ok(()) => { self.in_flight.push_back(slot); // Bound the frames in flight (memory): wait for the oldest. - while self.in_flight.len() > self.max_in_flight { + let limit = in_flight_limit(w.handles.len(), fh_samples); + while self.in_flight.len() > limit { let Some(s) = self.in_flight.pop_front() else { break }; let _ = s.wait(); } @@ -796,18 +1019,19 @@ impl Decoder { fn run_job(job: FrameJob, pool: &Pool, stats: &Mutex) { let mut st = DecodeStats::default(); let res = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| decode_frame(&job, pool, &mut st))); + let rows = job.fh.frame_height as usize; match res.unwrap_or(Err(Error::Invalid("decoder panicked"))) { Ok((rf, pic)) => { if let Some(ps) = &job.picture { ps.set(pic.ok_or(Error::Invalid("shown frame without picture"))); } - job.slot.set(Ok(rf)); + job.slot.finish(Ok(rf), rows); } Err(e) => { if let Some(ps) = &job.picture { ps.set(Err(e.clone())); } - job.slot.set(Err(e)); + job.slot.finish(Err(e), rows); } } if let Ok(mut s) = stats.lock() { @@ -832,71 +1056,191 @@ fn decode_frame(job: &FrameJob, pool: &Pool, stats: &mut DecodeStats) -> Result< if fh.primary_ref_frame != PRIMARY_REF_NONE { used[fh.ref_frame_idx[fh.primary_ref_frame]] = true; } - let mut refs: [Option>; NUM_REF_FRAMES] = Default::default(); - for i in 0..NUM_REF_FRAMES { - if used[i] { - let slot = job.refs[i].as_ref().ok_or(Error::Invalid("missing reference frame"))?; - refs[i] = Some(slot.wait()?); + // Prediction reads the reference planes with this frame's chroma format. + if !fh.frame_is_intra { + for i in 0..NUM_REF_FRAMES { + let Some(slot) = job.refs[i].as_ref().filter(|_| used[i]) else { continue }; + let b = &slot.bands; + let c = &seq.color; + if b.num_planes != c.num_planes || b.subsampling_x != c.subsampling_x as usize || b.subsampling_y != c.subsampling_y as usize { + return Err(Error::Invalid("reference frame has another chroma format")); + } } } - let t = Timer::start(); - let mut fs = FrameState::new(seq, fh); - fs.refs = refs; - fs.ref_info = job.ref_info.clone(); + // Parse dependencies first, then the samples. Parsing reads the CDFs and segment ids of + // the primary reference frame, and the saved motion vectors of the inter references only + // with use_ref_frame_mvs; the parse results of the other references are not waited for + // (their failures still fail this frame through the waits for their samples below). + let mut parse_deps = [false; NUM_REF_FRAMES]; if fh.primary_ref_frame != PRIMARY_REF_NONE { - // load_previous_segment_ids( ) - let (mi_cols, mi_rows) = (fs.mi.cols, fs.mi.rows); - let sh = &mut fs.sh; - if let Some(rf) = &sh.refs[fh.ref_frame_idx[fh.primary_ref_frame]] - && fh.seg.enabled - && rf.mi_cols == mi_cols - && rf.mi_rows == mi_rows - { - sh.prev_segment_ids.copy_from_slice(&rf.segment_ids); - } + parse_deps[fh.ref_frame_idx[fh.primary_ref_frame]] = true; } if fh.use_ref_frame_mvs { - motion_field_estimation(&mut fs); + for i in 0..REFS_PER_FRAME { + parse_deps[fh.ref_frame_idx[i]] = true; + } } - let mut cdfs = if fh.primary_ref_frame == PRIMARY_REF_NONE { - Cdfs::new(fh.quant.base_q_idx) - } else { - let rf = fs.refs[fh.ref_frame_idx[fh.primary_ref_frame]].as_ref().ok_or(Error::Invalid("primary reference frame missing"))?; - let mut c = rf.cdfs.clone(); - c.clear_counts(); - c + let mut parsed: [Option>; NUM_REF_FRAMES] = Default::default(); + for i in 0..NUM_REF_FRAMES { + if parse_deps[i] { + let slot = job.refs[i].as_ref().ok_or(Error::Invalid("missing reference frame"))?; + parsed[i] = Some(slot.wait_parsed()?); + } + } + // The samples of the references predicted from (inter frames only). + let wait_samples = || -> Result<[Option>; NUM_REF_FRAMES]> { + let mut refs: [Option>; NUM_REF_FRAMES] = Default::default(); + if !fh.frame_is_intra { + for i in 0..NUM_REF_FRAMES { + if used[i] { + let slot = job.refs[i].as_ref().ok_or(Error::Invalid("missing reference frame"))?; + slot.progress.wait(job.ref_info[i].frame_height as usize)?; + refs[i] = Some(slot.wait()?); + } + } + } + Ok(refs) }; + // With worker threads, the symbols of every tile are read first, so the parse results + // (which the next frames need before they can read their own symbols) are published + // before reconstruction, and parsing does not wait for the reference samples; on one + // thread each block is reconstructed as it is read. + let split = pool.is_parallel(); + // The frame state before tile decoding (with the reference samples when `with_refs`). + let setup = |with_refs: bool| -> Result<(FrameState, Box)> { + let mut fs = FrameState::new(seq, fh); + if with_refs { + fs.refs = wait_samples()?; + } + fs.parsed = parsed.clone(); + fs.ref_info = job.ref_info.clone(); + if fh.primary_ref_frame != PRIMARY_REF_NONE { + // load_previous_segment_ids( ) + let (mi_cols, mi_rows) = (fs.mi.cols, fs.mi.rows); + let sh = &mut fs.sh; + if let Some(rf) = &sh.parsed[fh.ref_frame_idx[fh.primary_ref_frame]] + && fh.seg.enabled + && rf.mi_cols == mi_cols + && rf.mi_rows == mi_rows + { + sh.prev_segment_ids.copy_from_slice(&rf.segment_ids); + } + } + if fh.use_ref_frame_mvs { + motion_field_estimation(&mut fs); + } + let cdfs = if fh.primary_ref_frame == PRIMARY_REF_NONE { + Cdfs::new(fh.quant.base_q_idx) + } else { + let rf = fs.parsed[fh.ref_frame_idx[fh.primary_ref_frame]].as_ref().ok_or(Error::Invalid("primary reference frame missing"))?; + let mut c = rf.cdfs.clone(); + c.clear_counts(); + c + }; + Ok((fs, cdfs)) + }; + let t = Timer::start(); + let (mut fs, mut cdfs) = setup(!split)?; stats.add(Stage::Setup, t.secs()); let t = Timer::start(); - let saved = decode_tiles(pool, &mut fs, &cdfs, &job.tiles)?; + // Time spent waiting for reference samples is not decoding time. + let mut waited = 0.0; + #[cfg(not(test))] + let record_limit = RECORD_BYTES_PER_SAMPLE; + #[cfg(test)] + let record_limit = job.record_limit.unwrap_or(RECORD_BYTES_PER_SAMPLE); + let mut tiles = parse_tiles(pool, &mut fs, &cdfs, &job.tiles, if split { Pass::Parse } else { Pass::Full }, record_limit, &job.record_budget)?; + // A frame whose record would take too much memory (pathological streams) is decoded again + // in one pass, as on one thread. + let one_pass = split && tiles.records.iter().any(|(_, r)| r.over_limit()); + if one_pass { + // The fallback starts again from the tile data: release the records first. + drop(tiles); + let wait = Timer::start(); + (fs, cdfs) = setup(true)?; + waited = wait.secs(); + tiles = parse_tiles(pool, &mut fs, &cdfs, &job.tiles, Pass::Full, record_limit, &job.record_budget)?; + } stats.tiles += job.tiles.len() as u64; - stats.add(Stage::Tiles, t.secs()); if !fh.disable_frame_end_update_cdf - && let Some(s) = saved + && let Some(s) = tiles.saved.take() { cdfs = s; } - crate::av1::decode::postfilter::apply(&mut fs, stats, pool, job.draft); + if split && !one_pass { + if tiles.whole { + if let Some(state) = tiles.states.pop() { + fs.put_tile_state(state); + job.slot.set_parsed(Ok(Arc::new(parse_results(&fs, cdfs)))); + tiles.states.push(fs.take_tile_state()); + } + } else { + for state in &tiles.states { + fs.merge_tile_modes(state); + } + job.slot.set_parsed(Ok(Arc::new(parse_results(&fs, cdfs)))); + } + if !fh.use_superres { + if !tiles.whole { + // The tiles were parsed into private buffers, merged into the frame above: + // reconstruct them in the frame's own buffers. + tiles.states.clear(); + tiles.states.push(fs.take_tile_state()); + tiles.whole = true; + } + // Reconstruct superblock row by superblock row (every tile of the row), each + // waiting only for the reference rows it reads, with the in-loop filters following + // behind. + waited = reconstruct_wavefront(job, pool, &mut fs, &mut tiles, stats)?; + } else { + let wait = Timer::start(); + fs.refs = wait_samples()?; + waited = wait.secs(); + reconstruct_tiles(pool, &fs.sh, &mut tiles)?; + tiles.finish(&mut fs); + let f = Timer::start(); + if fh.use_superres { + crate::av1::decode::postfilter::apply(&mut fs, stats, pool, job.draft); + publish_bands(&job.slot, &fs.cur, 0, usize::MAX, fh.frame_height as usize); + } else { + let mut pipe = crate::av1::decode::postfilter::Pipeline::new(&fs, job.draft); + let done = pipe.advance(&mut fs, stats, pool, usize::MAX); + publish_final(&job.slot, done, fh.frame_height as usize); + } + waited += f.secs(); + } + } else { + tiles.finish(&mut fs); + job.slot.set_parsed(Ok(Arc::new(parse_results(&fs, cdfs)))); + } + stats.add(Stage::Tiles, t.secs() - waited); + if !split || one_pass { + crate::av1::decode::postfilter::apply(&mut fs, stats, pool, job.draft); + } + if one_pass { + publish_bands(&job.slot, &fs.cur, 0, usize::MAX, fh.frame_height as usize); + } + if split { + // A frame using a reference only for its parse results (an intra frame's CDFs) still + // fails with it, as it does on one thread. + for i in 0..NUM_REF_FRAMES { + if used[i] { + job.refs[i].as_ref().ok_or(Error::Invalid("missing reference frame"))?.wait()?; + } + } + } stats.draft_frames += job.draft as u64; let t = Timer::start(); stats.frames += 1; - if fh.seg.enabled && !fh.seg.update_map { - let prev = std::mem::take(&mut fs.prev_segment_ids); - fs.mi.segment_id.copy_from_slice(&prev); - } - let (mf_refs, mf_mvs) = motion_vector_storage(&fs); let rf = Arc::new(RefFrame { - buf: std::mem::take(&mut fs.cur), - cdfs, + // With worker threads the samples are in the published bands. + buf: if split { FrameBuf::default() } else { std::mem::take(&mut fs.cur) }, + bands: split.then(|| job.slot.bands.clone()), + num_planes: seq.color.num_planes, bit_depth: seq.color.bit_depth, subsampling_x: seq.color.subsampling_x, subsampling_y: seq.color.subsampling_y, - segment_ids: std::mem::take(&mut fs.mi.segment_id), - mi_cols: fs.mi.cols, - mi_rows: fs.mi.rows, color: (seq.color.color_primaries, seq.color.transfer_characteristics, seq.color.matrix_coefficients, seq.color.color_range), - saved_ref_frames: mf_refs, - saved_mvs: mf_mvs, }); // job.picture is None for a shown frame refused for its size (see tile_group) let pic = job.picture.is_some().then(|| { @@ -908,6 +1252,232 @@ fn decode_frame(job: &FrameJob, pool: &Pool, stats: &mut DecodeStats) -> Result< Ok((rf, pic)) } +/// Publish bands [from, to) copied from the final samples `src`, and the row progress they make +/// final. +fn publish_bands(slot: &FrameSlot, src: &FrameBuf, from: usize, to: usize, height: usize) { + let to = to.min(slot.bands.bands()); + for b in from..to { + slot.bands.publish(b, src); + } + slot.progress.publish((to << crate::av1::decode::frame::BAND_LOG2).min(height)); +} + +/// Publish the bands that became final (in order, following those published before) and the +/// row progress they make final. +fn publish_final(slot: &FrameSlot, bands: crate::av1::decode::postfilter::FinalBands, height: usize) { + let mut end = None; + for (b, planes) in bands { + slot.bands.publish_owned(b, planes); + end = Some(b + 1); + } + if let Some(end) = end { + slot.progress.publish((end << crate::av1::decode::frame::BAND_LOG2).min(height)); + } +} + +/// Extra luma rows waited for below what a motion vector reaches: the 8-tap filters read 4 +/// rows below a block, chroma positions round, and the vertical sub-sample phase adds one. +const ROW_WAIT_MARGIN: usize = 16; + +/// The rows of each reference slot that reconstructing the blocks `blocks` (a superblock row +/// and the row above it, whose motion overlapped block motion compensation can reuse) reads, +/// for blocks ending at luma row `bottom`: their reach plus a margin, or everything +/// (`usize::MAX`) for warped motion and scaled references, whose reach is not bounded by the +/// motion vector. +fn rows_needed( + fh: &FrameHeader, + ref_info: &[RefInfo; NUM_REF_FRAMES], + blocks: impl Iterator, + bottom: usize, +) -> [usize; NUM_REF_FRAMES] { + let mut need = [0usize; NUM_REF_FRAMES]; + let mut used = [false; NUM_REF_FRAMES]; + let mut warp = false; + let global = (LAST_FRAME..=ALTREF_FRAME).any(|r| fh.gm_type[r] as usize > TRANSLATION); + for b in blocks { + if !b.inter { + continue; + } + if b.motion_mode as usize == LOCALWARP || (global && (b.y_mode == GLOBALMV || b.y_mode == GLOBAL_GLOBALMV)) { + warp = true; + } + let bottom = bottom.max((b.mi_row + b.bh4) * 4); + for list in 0..2 { + let rf = b.ref_frame[list]; + if rf <= INTRA_FRAME as i8 { + continue; + } + let Some(&i) = fh.ref_frame_idx.get(rf as usize - LAST_FRAME) else { continue }; + used[i] = true; + let ri = &ref_info[i]; + if ri.upscaled_width != fh.frame_width || ri.frame_height != fh.frame_height { + need[i] = usize::MAX; + continue; + } + let down = (b.mv[list][0].max(0) as usize).div_ceil(8); + need[i] = need[i].max(bottom.saturating_add(down).saturating_add(ROW_WAIT_MARGIN)); + } + } + if warp { + for i in 0..NUM_REF_FRAMES { + if used[i] { + need[i] = usize::MAX; + } + } + } + need +} + +/// Reconstructed luma rows between two runs of the in-loop filter pipeline in +/// [`reconstruct_wavefront`] (finer steps publish rows sooner, coarser ones filter more bands in +/// parallel). +const FILTER_STEP_ROWS: usize = 128; + +/// Reconstruct a frame superblock row by superblock row (row wavefront): before each row, wait +/// only for the reference rows it reads; then reconstruct that row in every tile it crosses +/// (tiles share no prediction context, so the order of the tiles within a row does not matter) +/// and run the in-loop filters on what that allows, publishing the bands that became final. +/// `tiles` holds the whole frame's buffers. Returns the seconds spent waiting for references +/// and filtering (not tile decoding). +fn reconstruct_wavefront(job: &FrameJob, pool: &Pool, fs: &mut FrameState, tiles: &mut FrameTiles, stats: &mut DecodeStats) -> Result { + let fh = &job.fh; + let mut other = 0.0; + let ti = fh.tile_info.clone(); + let Some(state) = tiles.states.pop() else { + return Err(Error::Invalid("no tile to reconstruct")); + }; + // The record of every tile, by tile number. + let mut recs: Vec> = (0..ti.rows * ti.cols).map(|_| None).collect(); + for (num, rec) in std::mem::take(&mut tiles.records) { + *recs.get_mut(num).ok_or(Error::Invalid("tile number out of range"))? = Some(rec); + } + let mut used = [false; NUM_REF_FRAMES]; + if !fh.frame_is_intra { + for i in 0..REFS_PER_FRAME { + used[fh.ref_frame_idx[i]] = true; + } + } + for i in 0..NUM_REF_FRAMES { + if used[i] { + let slot = job.refs[i].as_ref().ok_or(Error::Invalid("missing reference frame"))?; + fs.ref_rows[i] = Some(slot.bands.clone()); + } + } + fs.put_tile_state(state); + let mut pipe = crate::av1::decode::postfilter::Pipeline::new(fs, job.draft); + let sb4 = if job.seq.use_128x128_superblock { 32 } else { 16 }; + let mi_rows = fh.mi_rows as usize; + let sb_rows = mi_rows.div_ceil(sb4); + let mut bufs = (crate::av1::decode::tile::Buffers::new(), Cdfs::new(fh.quant.base_q_idx)); + let mut filtered = 0; + let mut tile_row = 0; + for row in 0..sb_rows { + // Tile rows start on superblock rows. + while tile_row + 1 < ti.rows && row * sb4 >= ti.mi_row_starts[tile_row + 1] as usize { + tile_row += 1; + } + let local = (row * sb4).saturating_sub(ti.mi_row_starts[tile_row] as usize) / sb4; + let w = Timer::start(); + if !fh.frame_is_intra { + let mut need = [0usize; NUM_REF_FRAMES]; + for col in 0..ti.cols { + let Some(rec) = &recs[tile_row * ti.cols + col] else { continue }; + let width = (ti.mi_col_starts[col + 1] as usize).min(fh.mi_cols as usize).saturating_sub(ti.mi_col_starts[col] as usize); + let n = rows_needed(fh, &job.ref_info, rec.row_blocks(width.div_ceil(sb4), local.saturating_sub(1), local + 1), (row + 1) * sb4 * 4); + for i in 0..NUM_REF_FRAMES { + need[i] = need[i].max(n[i]); + } + } + for i in 0..NUM_REF_FRAMES { + if used[i] && need[i] > 0 { + let slot = job.refs[i].as_ref().ok_or(Error::Invalid("missing reference frame"))?; + slot.progress.wait(need[i].min(job.ref_info[i].frame_height as usize))?; + } + } + } + other += w.secs(); + let mut t = fs.take_tile_state(); + let mut res = Ok(()); + for col in 0..ti.cols { + let Some(rec) = recs[tile_row * ti.cols + col].as_mut() else { continue }; + let (b, cdf) = bufs; + let mut td = TileDecoder::with_buffers(&fs.sh, &mut t, &[], cdf, tile_row, col, b); + td.pass = Pass::Reconstruct; + td.record = std::mem::take(rec); + res = td.reconstruct_sb_rows(1); + *rec = std::mem::take(&mut td.record); + bufs = td.into_buffers(); + if res.is_err() { + break; + } + } + fs.put_tile_state(t); + res?; + let f = Timer::start(); + let recon_rows = ((row + 1) * sb4 * 4).min(mi_rows * 4); + // Filter in steps of a few bands, so the bands of a step run in parallel. + if recon_rows >= filtered + FILTER_STEP_ROWS || recon_rows == mi_rows * 4 { + filtered = recon_rows; + let done = pipe.advance(fs, stats, pool, recon_rows); + publish_final(&job.slot, done, fh.frame_height as usize); + } + other += f.secs(); + #[cfg(test)] + assert!( + !(job.panic_at_row.is_some_and(|r| row >= r) && job.slot.progress.rows() > 0), + "forced panic in the row wavefront after publishing rows" + ); + } + // A reference that failed after publishing the rows used here still fails this frame, as + // it does when frames are decoded one after another. + for i in 0..NUM_REF_FRAMES { + if used[i] { + let slot = job.refs[i].as_ref().ok_or(Error::Invalid("missing reference frame"))?; + slot.wait()?; + } + } + Ok(other) +} + +/// Most bytes a tile's parse record may take per luma sample of the tile. Typical streams need +/// a fraction of this; past it (pathological streams: tiny blocks, large palettes, every +/// coefficient coded) the frame is decoded in one pass instead, so memory stays bounded. +const RECORD_BYTES_PER_SAMPLE: usize = 32; + +/// Most parse record bytes all frames of a decoder hold at once. Past it, the frames that +/// cannot reserve more are decoded in one pass, so the frames in flight together stay bounded +/// too (typical streams need a few bytes per luma sample: under 100 MB for a 2160p frame). +const RECORD_BUDGET_BYTES: usize = 1 << 30; + +/// The tiles of a frame between parsing and reconstruction. +struct FrameTiles { + /// One state per tile (decoded in parallel into private buffers), or one for the whole + /// frame (tiles one after another). + states: Vec, + /// Per tile: (tile number, record of [`Pass::Parse`]). + records: Vec<(usize, TileRecord)>, + /// The final CDFs of tile context_update_tile_id. + saved: Option>, + /// `states` holds the frame's own buffers (taken with [`FrameState::take_tile_state`]). + whole: bool, +} + +impl FrameTiles { + /// Move the tile buffers (mode info, filter parameters and, after reconstruction, the + /// samples) into the frame. + fn finish(self, fs: &mut FrameState) { + if self.whole { + if let Some(t) = self.states.into_iter().next() { + fs.put_tile_state(t); + } + } else { + for t in &self.states { + fs.merge_tile_state(t); + } + } + } +} + /// Bytes all parallel tile states may hold of whole-frame CDEF and restoration metadata; past it /// the tiles decode one after another on the frame's own metadata. const TILE_METADATA_BUDGET: usize = 256 << 20; @@ -916,16 +1486,30 @@ fn tile_metadata_fits(tiles: usize, per_tile: usize) -> bool { tiles.checked_mul(per_tile).is_some_and(|n| n <= TILE_METADATA_BUDGET) } -/// Decode the tiles of a frame (tile number, data), in parallel when several; returns the final -/// CDFs of tile context_update_tile_id. -fn decode_tiles(pool: &Pool, fs: &mut FrameState, cdfs: &Cdfs, tiles: &[(usize, Vec)]) -> Result>> { +/// Run `pass` (Full or Parse) over the tiles of a frame. +/// `record_limit`: most record bytes per luma sample of a tile (see [`RECORD_BYTES_PER_SAMPLE`]). +/// `budget`: the decoder-wide budget the records are reserved from. +fn parse_tiles( + pool: &Pool, + fs: &mut FrameState, + cdfs: &Cdfs, + tiles: &[(usize, Vec)], + pass: Pass, + record_limit: usize, + budget: &Arc, +) -> Result { let ti = fs.fh.tile_info.clone(); let ctx_tile = ti.context_update_tile_id as usize; - let run = |sh: &FrameShared, t: &mut TileState, num: usize, data: &[u8]| -> Result>> { + let run = |sh: &FrameShared, t: &mut TileState, num: usize, data: &[u8]| -> Result<(Option>, TileRecord)> { t.intra_frame_y_mode_cdf = DEFAULT_INTRA_FRAME_Y_MODE_CDF; let mut td = TileDecoder::new(sh, t, data, Box::new(cdfs.clone()), num / ti.cols, num % ti.cols); + td.pass = pass; + let (rows, cols) = (td.mi_row_end.saturating_sub(td.mi_row_start), td.mi_col_end.saturating_sub(td.mi_col_start)); + td.record.limit = rows.saturating_mul(cols).saturating_mul(16 * record_limit); + td.record.budget = Some(budget.clone()); td.decode_tile()?; - Ok((num == ctx_tile).then_some(td.cdf)) + let rec = std::mem::take(&mut td.record); + Ok(((num == ctx_tile).then_some(td.cdf), rec)) }; // Intra block copy reads the current frame through the tile's own buffer: keep whole-frame // buffers (tiles one after another) for it. @@ -941,42 +1525,75 @@ fn decode_tiles(pool: &Pool, fs: &mut FrameState, cdfs: &Cdfs, tiles: &[(usize, .collect(); let sh = &fs.sh; let results = pool.map_mut(&mut states, |i, t| run(sh, t, tiles[i].0, &tiles[i].1)); - for t in &states { - fs.merge_tile_state(t); - } let mut saved = None; - for r in results { - if let Some(c) = r? { - saved = Some(c); + let mut records = Vec::with_capacity(tiles.len()); + for (r, (num, _)) in results.into_iter().zip(tiles) { + let (c, rec) = r?; + if c.is_some() { + saved = c; } + records.push((*num, rec)); } - return Ok(saved); + return Ok(FrameTiles { states, records, saved, whole: false }); } let mut t = fs.take_tile_state(); let mut saved = None; + let mut records = Vec::with_capacity(tiles.len()); let mut res = Ok(()); for (num, data) in tiles { match run(&fs.sh, &mut t, *num, data) { - Ok(Some(c)) => saved = Some(c), - Ok(None) => {} + Ok((c, rec)) => { + if c.is_some() { + saved = c; + } + records.push((*num, rec)); + } Err(e) => { res = Err(e); break; } } } - fs.put_tile_state(t); - res.map(|_| saved) + if let Err(e) = res { + fs.put_tile_state(t); + return Err(e); + } + Ok(FrameTiles { states: vec![t], records, saved, whole: true }) +} + +/// Reconstruct parsed tiles (in parallel when each has its own state). +fn reconstruct_tiles(pool: &Pool, fs: &FrameShared, tiles: &mut FrameTiles) -> Result<()> { + let ti = fs.fh.tile_info.clone(); + let run = |t: &mut TileState, num: usize, rec: TileRecord| -> Result<()> { + let mut td = TileDecoder::new(fs, t, &[], Cdfs::new(fs.fh.quant.base_q_idx), num / ti.cols, num % ti.cols); + td.pass = Pass::Reconstruct; + td.record = rec; + td.decode_tile() + }; + let records = std::mem::take(&mut tiles.records); + if !tiles.whole { + let mut items: Vec<(&mut TileState, Option<(usize, TileRecord)>)> = tiles.states.iter_mut().zip(records.into_iter().map(Some)).collect(); + let results = pool.map_mut(&mut items, |_, (t, rec)| match rec.take() { + Some((num, rec)) => run(t, num, rec), + None => Ok(()), + }); + return results.into_iter().collect(); + } + let Some(t) = tiles.states.first_mut() else { return Ok(()) }; + for (num, rec) in records { + run(t, num, rec)?; + } + Ok(()) } fn output_picture(seq: &SequenceHeader, rf: &RefFrame, w: usize, h: usize, grain: &FilmGrainParams, apply_film_grain: bool, pts: i64) -> Picture { let ssx = rf.subsampling_x as usize; let ssy = rf.subsampling_y as usize; - let num_planes = rf.buf.num_planes; + let num_planes = rf.num_planes; let mut planes: [Vec; 3] = Default::default(); for p in 0..num_planes { let (pw, ph) = if p == 0 { (w, h) } else { ((w + ssx) >> ssx, (h + ssy) >> ssy) }; - let src = &rf.buf.planes[p]; + let src = rf.plane(p); let mut v = Vec::with_capacity(pw * ph); for y in 0..ph { v.extend_from_slice(&src.row(y)[..pw]); @@ -1004,25 +1621,40 @@ fn output_picture(seq: &SequenceHeader, rf: &RefFrame, w: usize, h: usize, grain pic } -/// Motion field motion vector storage process (7.19): (MfRefFrames, MfMvs). +/// What later frames read from this one while parsing: the final CDFs, the segment ids (the +/// previous frame's when the map is not updated) and the stored motion vectors (7.19). The +/// in-loop filters read the decoded segment ids, so the frame's own map is left alone. +fn parse_results(fs: &FrameState, cdfs: Box) -> Parsed { + let fh = &fs.fh; + let segment_ids = if fh.seg.enabled && !fh.seg.update_map { fs.prev_segment_ids.clone() } else { fs.mi.segment_id.clone() }; + let (saved_ref_frames, saved_mvs) = motion_vector_storage(fs); + Parsed { cdfs, segment_ids, mi_cols: fs.mi.cols, mi_rows: fs.mi.rows, saved_ref_frames, saved_mvs } +} + +/// Motion field motion vector storage process (7.19): (MfRefFrames, MfMvs), kept only where +/// the projection process (7.9.2) reads them, the odd row and column of every 8x8 block +/// (index `y8 * (MiCols >> 1) + x8`). fn motion_vector_storage(fs: &FrameState) -> (Vec, Vec<[i32; 2]>) { let mi = &fs.mi; - let n = mi.rows * mi.cols; - let mut refs = vec![-1i8; n]; - let mut mvs = vec![[0i32; 2]; n]; let fh = &fs.fh; - for i in 0..n { - for list in 0..2 { - let r = mi.ref_frame[i][list]; - if r > INTRA_FRAME as i8 { - let ref_idx = fh.ref_frame_idx[r as usize - LAST_FRAME]; - let dist = crate::av1::decode::header::relative_dist(&fs.seq, fs.ref_info[ref_idx].order_hint, fh.order_hint); - if dist < 0 { - let m = mi.mv[i][list]; - let (row, col) = (m.row as i32, m.col as i32); - if row.abs() <= REFMVS_LIMIT as i32 && col.abs() <= REFMVS_LIMIT as i32 { - refs[i] = r; - mvs[i] = [row, col]; + let (mi_cols, w8, h8) = (fh.mi_cols as usize, fh.mi_cols as usize >> 1, fh.mi_rows as usize >> 1); + let mut refs = vec![-1i8; w8 * h8]; + let mut mvs = vec![[0i32; 2]; w8 * h8]; + for y8 in 0..h8 { + for x8 in 0..w8 { + let (i, o) = ((2 * y8 + 1) * mi_cols + 2 * x8 + 1, y8 * w8 + x8); + for list in 0..2 { + let r = mi.ref_frame[i][list]; + if r > INTRA_FRAME as i8 { + let ref_idx = fh.ref_frame_idx[r as usize - LAST_FRAME]; + let dist = crate::av1::decode::header::relative_dist(&fs.seq, fs.ref_info[ref_idx].order_hint, fh.order_hint); + if dist < 0 { + let m = mi.mv[i][list]; + let (row, col) = (m.row as i32, m.col as i32); + if row.abs() <= REFMVS_LIMIT as i32 && col.abs() <= REFMVS_LIMIT as i32 { + refs[o] = r; + mvs[o] = [row, col]; + } } } } @@ -1080,7 +1712,7 @@ fn project_pos(v8: i32, delta: i32, dst_sign: i32, max8: i32, max_off8: i32) -> /// Projection process (7.9.2). fn project(fs: &mut FrameState, src: usize, dst_sign: i32) -> bool { - let FrameShared { seq, fh, refs, ref_info, motion_field, .. } = &mut fs.sh; + let FrameShared { seq, fh, parsed, ref_info, motion_field, .. } = &mut fs.sh; let src_idx = fh.ref_frame_idx[src - LAST_FRAME]; let w8 = fh.mi_cols as i32 >> 1; let h8 = fh.mi_rows as i32 >> 1; @@ -1088,20 +1720,18 @@ fn project(fs: &mut FrameState, src: usize, dst_sign: i32) -> bool { if ri.mi_rows != fh.mi_rows || ri.mi_cols != fh.mi_cols || ri.frame_type == INTRA_ONLY_FRAME as u8 || ri.frame_type == KEY_FRAME as u8 { return false; } - let Some(rf) = &refs[src_idx] else { return false }; + let Some(rf) = &parsed[src_idx] else { return false }; let rd = |a: u32, b: u32| crate::av1::decode::header::relative_dist(seq, a, b); - let mi_cols = fh.mi_cols as usize; for y8 in 0..h8 { for x8 in 0..w8 { - let row = (2 * y8 + 1) as usize; - let col = (2 * x8 + 1) as usize; - let src_ref = rf.saved_ref_frames[row * mi_cols + col]; + let at = (y8 * w8 + x8) as usize; + let src_ref = rf.saved_ref_frames[at]; if src_ref > INTRA_FRAME as i8 { let ref_to_cur = rd(fh.order_hints[src], fh.order_hint); let ref_offset = rd(fh.order_hints[src], ri.saved_order_hints[src_ref as usize]); let pos_valid = ref_to_cur.abs() <= MAX_FRAME_DISTANCE as i32 && ref_offset.abs() <= MAX_FRAME_DISTANCE as i32 && ref_offset > 0; if pos_valid { - let mv = rf.saved_mvs[row * mi_cols + col]; + let mv = rf.saved_mvs[at]; let proj = get_mv_projection(mv, ref_to_cur * dst_sign, ref_offset); let py = project_pos(y8, proj[0], dst_sign, h8, MAX_OFFSET_HEIGHT as i32); let px = project_pos(x8, proj[1], dst_sign, w8, MAX_OFFSET_WIDTH as i32); @@ -1133,13 +1763,18 @@ mod tests { ref_info: Default::default(), tiles: Vec::new(), tile_covered: vec![false; 3], - slot: Slot::new(), + slot: FrameSlot::new(BandFrame::new(0, 0, 1, 0, 0)), picture: None, pts: 0, apply_film_grain: false, draft: false, #[cfg(test)] force_panic: false, + #[cfg(test)] + panic_at_row: None, + #[cfg(test)] + record_limit: None, + record_budget: decoder.record_budget.clone(), })); decoder } @@ -1286,6 +1921,317 @@ mod tests { assert!(frame_size_allowed(1, 1)); } + /// Random short sleeps in row progress publishing and waiting, to shake out scheduling + /// orders a quiet machine never produces (enabled by [`JitterGuard`]). + static JITTER: std::sync::atomic::AtomicUsize = std::sync::atomic::AtomicUsize::new(0); + + pub(super) fn jitter() { + use std::cell::Cell; + use std::sync::atomic::Ordering; + if JITTER.load(Ordering::Relaxed) == 0 { + return; + } + static SEED: std::sync::atomic::AtomicU64 = std::sync::atomic::AtomicU64::new(0x2545_f491_4f6c_dd1d); + thread_local!(static RNG: Cell = Cell::new(SEED.fetch_add(0x9e37_79b9_7f4a_7c15, Ordering::Relaxed) | 1)); + let r = RNG.with(|c| { + let mut x = c.get(); + x ^= x << 13; + x ^= x >> 7; + x ^= x << 17; + c.set(x); + x + }); + match r % 4 { + 0 => std::thread::yield_now(), + 1 => std::thread::sleep(std::time::Duration::from_micros(r % 300)), + _ => {} + } + } + + struct JitterGuard; + impl JitterGuard { + fn on() -> JitterGuard { + JITTER.fetch_add(1, std::sync::atomic::Ordering::SeqCst); + JitterGuard + } + } + impl Drop for JitterGuard { + fn drop(&mut self) { + JITTER.fetch_sub(1, std::sync::atomic::Ordering::SeqCst); + } + } + + /// Inter test streams made with ffmpeg / SVT-AV1 (None, reported, without them): several + /// superblock rows and hierarchical references; odd size; super-resolution; 2x2 tiles. + const STREAMS: [(&str, &str, &str); 4] = [ + ("wavefront_unit_352x288", "testsrc2=size=352x288:rate=25", "-preset 8 -crf 35"), + ("wavefront_unit_odd", "testsrc2=size=203x117:rate=25,noise=alls=15:allf=t", "-preset 8 -crf 40"), + ("wavefront_unit_superres", "testsrc2=size=416x240:rate=25", "-preset 8 -crf 35 -svtav1-params superres-mode=1:superres-denom=12"), + ("wavefront_unit_tiles", "mandelbrot=size=650x370:rate=25", "-preset 8 -crf 32 -svtav1-params tile-columns=1:tile-rows=1"), + ]; + + fn inter_stream((name, src, args): (&str, &str, &str)) -> Option>> { + let ff = craft_media_testkit::ffmpeg_or_skip("av1 wavefront")?; + if !craft_media_testkit::ffmpeg_has_encoder(&ff, "libsvtav1") { + craft_media_testkit::skip_test("av1 wavefront", "this ffmpeg has no libsvtav1 encoder"); + return None; + } + let out = craft_media_testkit::fixtures_dir("av1").join(format!("{name}.ivf")); + let path = craft_media_testkit::fixtures::generate(&out, |tmp| { + std::process::Command::new(&ff) + .args([ + "-hide_banner", + "-loglevel", + "error", + "-y", + "-f", + "lavfi", + "-i", + src, + "-frames:v", + "16", + "-pix_fmt", + "yuv420p", + "-c:v", + "libsvtav1", + ]) + .args(args.split_whitespace()) + .args(["-f", "ivf"]) + .arg(tmp) + .stdin(std::process::Stdio::null()) + .status() + .is_ok_and(|s| s.success()) + }); + let path = path.unwrap_or_else(|| panic!("ffmpeg has libsvtav1 but could not encode {name}")); + let d = std::fs::read(&path).unwrap_or_else(|e| panic!("read {}: {e}", path.display())); + let mut pos = usize::from(u16::from_le_bytes([d[6], d[7]])); + let mut out = Vec::new(); + while let Some(h) = d.get(pos..pos + 12) { + let n = u32::from_le_bytes([h[0], h[1], h[2], h[3]]) as usize; + pos += 12; + out.push(d[pos..pos + n].to_vec()); + pos += n; + } + assert_eq!(out.len(), 16, "{name}: temporal units"); + Some(out) + } + + /// Decode `tus` with `dec`: (pictures, errors). + fn decode_tus(mut dec: Decoder, tus: &[Vec]) -> (Vec, usize) { + let (mut pics, mut errors) = (Vec::new(), 0); + for (i, tu) in tus.iter().enumerate() { + match dec.decode_pts(tu, i as i64) { + Ok(p) => pics.extend(p), + Err(_) => errors += 1, + } + } + match dec.flush_result() { + Ok(p) => pics.extend(p), + Err(_) => errors += 1, + } + (pics, errors) + } + + #[test] + fn jittered_wavefront_gives_the_single_thread_pictures() { + for stream in STREAMS { + let Some(tus) = inter_stream(stream) else { return }; + let name = stream.0; + let (want, errors) = decode_tus(Decoder::with_threads(1), &tus); + assert_eq!(errors, 0, "{name}"); + let _jitter = JitterGuard::on(); + for threads in [2, 3, 5, 8] { + for round in 0..2 { + let tus = tus.clone(); + let (got, errors) = within_a_minute(move || decode_tus(Decoder::with_threads(threads), &tus)); + assert_eq!(errors, 0, "{name}: threads {threads} round {round}"); + assert!(got == want, "{name}: threads {threads} round {round}: pictures differ"); + } + } + } + } + + #[test] + fn frames_over_the_record_limit_decode_in_one_pass() { + for stream in [STREAMS[0], STREAMS[3]] { + let Some(tus) = inter_stream(stream) else { return }; + let name = stream.0; + let (want, _) = decode_tus(Decoder::with_threads(1), &tus); + for (threads, limit) in [(3, 0), (8, 0), (8, 1)] { + let tus = tus.clone(); + let (got, errors) = within_a_minute(move || { + let mut dec = Decoder::with_threads(threads); + dec.record_limit = Some(limit); + decode_tus(dec, &tus) + }); + assert_eq!(errors, 0, "{name}: threads {threads}, limit {limit}"); + assert!(got == want, "{name}: threads {threads}, limit {limit}: pictures differ"); + } + } + } + + #[test] + fn the_parse_records_of_all_frames_stay_within_the_decoder_budget() { + for stream in [STREAMS[0], STREAMS[3]] { + let Some(tus) = inter_stream(stream) else { return }; + let name = stream.0; + let (want, _) = decode_tus(Decoder::with_threads(1), &tus); + // Room for about one frame's records: the other frames in flight decode in one pass. + for (threads, budget) in [(3, 0), (8, 3 << 20), (16, 1 << 20)] { + let tus = tus.clone(); + let b = Arc::new(crate::av1::decode::tile::RecordBudget::new(budget)); + let mut dec = Decoder::with_threads(threads); + dec.record_budget = b.clone(); + let (got, errors) = within_a_minute(move || decode_tus(dec, &tus)); + assert_eq!(errors, 0, "{name}: threads {threads}, budget {budget}"); + assert!(got == want, "{name}: threads {threads}, budget {budget}: pictures differ"); + assert!(b.peak.load(std::sync::atomic::Ordering::Acquire) <= budget, "{name}: threads {threads}: over the budget"); + assert_eq!(b.used(), 0, "{name}: threads {threads}: records not released"); + } + } + } + + #[cfg(all(feature = "threads", not(target_arch = "wasm32")))] + #[test] + fn a_reference_panicking_after_publishing_rows_fails_without_a_hang() { + let Some(tus) = inter_stream(STREAMS[0]) else { return }; + let _jitter = JitterGuard::on(); + // Frame 1 is the first inter frame; later frames predict from it. + for threads in [2, 3, 8] { + for frame in [0, 1, 2] { + let tus = tus.clone(); + let (pics, errors) = within_a_minute(move || { + let mut dec = Decoder::with_threads(threads); + dec.panic_at_row = Some((frame, 1)); + decode_tus(dec, &tus) + }); + assert!(errors > 0, "threads {threads}, frame {frame}: the panic was not reported ({} pictures)", pics.len()); + } + } + } + + #[test] + fn dropping_a_busy_decoder_does_not_hang() { + let Some(tus) = inter_stream(STREAMS[0]) else { return }; + for threads in [2, 3, 8, 16] { + for cut in [1, 3, tus.len() / 2] { + let tus = tus.clone(); + within_a_minute(move || { + let mut dec = Decoder::with_threads(threads); + for (i, tu) in tus.iter().take(cut).enumerate() { + let _ = dec.decode_pts(tu, i as i64); + } + drop(dec); + }); + } + } + } + + #[test] + fn a_flush_mid_stream_keeps_the_pictures() { + let Some(tus) = inter_stream(STREAMS[0]) else { return }; + let (want, _) = decode_tus(Decoder::with_threads(1), &tus); + for threads in [2, 8] { + for cut in [2, 5, tus.len() / 2] { + let tus = tus.clone(); + let got = within_a_minute(move || { + let mut dec = Decoder::with_threads(threads); + let mut pics = Vec::new(); + for (i, tu) in tus.iter().enumerate() { + if i == cut { + pics.extend(dec.flush_result().expect("flush")); + } + pics.extend(dec.decode_pts(tu, i as i64).expect("decode")); + } + pics.extend(dec.flush_result().expect("flush")); + pics + }); + assert!(got == want, "threads {threads}, flush before unit {cut}"); + } + } + } + + #[test] + fn progress_wakes_every_waiter_with_the_rows_or_the_failure() { + let p = Arc::new(Progress::new()); + let waiters: Vec<_> = (1..=8) + .map(|k| { + let p = p.clone(); + std::thread::spawn(move || p.wait(k * 10)) + }) + .collect(); + p.publish(40); + p.publish(20); // never goes back + assert_eq!(p.wait(40), Ok(())); + p.fail(); + let res: Vec<_> = waiters.into_iter().map(|w| w.join().expect("waiter")).collect(); + assert!(res[..4].iter().all(|r| r.is_ok()), "{res:?}"); + assert!(res[4..].iter().all(|r| r.is_err()), "{res:?}"); + assert_eq!(p.wait(40), Ok(()), "rows published before a failure stay readable"); + } + + #[test] + fn a_failed_frame_keeps_its_published_parse_results() { + let slot = FrameSlot::new(BandFrame::new(16, 16, 1, 0, 0)); + let parsed = Parsed { cdfs: Cdfs::new(0), segment_ids: vec![], mi_cols: 4, mi_rows: 4, saved_ref_frames: vec![], saved_mvs: vec![] }; + slot.set_parsed(Ok(Arc::new(parsed))); + slot.finish(Err(Error::Invalid("x")), 16); + assert!(slot.wait_parsed().is_ok()); + assert!(slot.wait().is_err()); + assert!(slot.progress.wait(1).is_err()); + // A frame failing before its parse results: every waiter gets the error. + let slot = FrameSlot::new(BandFrame::new(16, 16, 1, 0, 0)); + slot.finish(Err(Error::Invalid("y")), 16); + assert!(slot.wait_parsed().is_err()); + } + + fn motion(mi_row: usize, bh4: usize, ref_frame: [i8; 2], mv: [[i32; 2]; 2]) -> crate::av1::decode::tile::BlockMotion { + crate::av1::decode::tile::BlockMotion { + mi_row, + bh4, + inter: true, + ref_frame, + mv, + motion_mode: 0, + y_mode: crate::av1::decode::spec_tables::NEARESTMV, + } + } + + #[test] + fn rows_needed_follow_the_motion_vectors() { + let mut fh = FrameHeader { frame_width: 64, upscaled_width: 64, frame_height: 256, ..Default::default() }; + fh.ref_frame_idx = [0, 1, 2, 3, 4, 5, 6]; + let mut ri: [RefInfo; NUM_REF_FRAMES] = Default::default(); + for r in ri.iter_mut() { + (r.upscaled_width, r.frame_height) = (64, 256); + } + let none = [[0; 2]; 2]; + // 41/8 rows down rounds up to 6; upward motion needs only the block's own rows. + let need = rows_needed(&fh, &ri, [motion(0, 16, [1, -1], [[41, 0], [0, 0]]), motion(0, 16, [2, 3], [[-80, 0], [7, 0]])].into_iter(), 64); + assert_eq!(need[0], 64 + 6 + ROW_WAIT_MARGIN); + assert_eq!(need[1], 64 + ROW_WAIT_MARGIN); + assert_eq!(need[2], 64 + 1 + ROW_WAIT_MARGIN); + assert_eq!(need[3], 0); + // Intra blocks and intra block copy read no reference. + let intra = crate::av1::decode::tile::BlockMotion { inter: false, ..motion(0, 16, [1, -1], none) }; + assert_eq!(rows_needed(&fh, &ri, [intra].into_iter(), 64), [0; NUM_REF_FRAMES]); + // Blocks reaching below the superblock row count from their own bottom. + assert_eq!(rows_needed(&fh, &ri, [motion(12, 8, [1, -1], none)].into_iter(), 64)[0], 80 + ROW_WAIT_MARGIN); + // Local warp: the whole of every reference the rows use. + let warp = + crate::av1::decode::tile::BlockMotion { motion_mode: crate::av1::decode::spec_tables::LOCALWARP as u8, ..motion(0, 16, [1, -1], none) }; + let need = rows_needed(&fh, &ri, [warp, motion(0, 16, [2, -1], none)].into_iter(), 64); + assert_eq!((need[0], need[1], need[2]), (usize::MAX, usize::MAX, 0)); + // Global motion beyond translation for GLOBALMV blocks. + fh.gm_type[3] = (TRANSLATION + 1) as u8; + let global = crate::av1::decode::tile::BlockMotion { y_mode: GLOBALMV, ..motion(0, 16, [1, -1], none) }; + assert_eq!(rows_needed(&fh, &ri, [global].into_iter(), 64)[0], usize::MAX); + fh.gm_type[3] = 0; + // A scaled reference. + ri[0].frame_height = 128; + assert_eq!(rows_needed(&fh, &ri, [motion(0, 16, [1, -1], none)].into_iter(), 64)[0], usize::MAX); + } + #[test] fn frames_beyond_the_limits_are_refused() { assert!(!frame_size_allowed(16384, 2177)); // one row over MaxPicSize @@ -1307,18 +2253,33 @@ 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 +2294,8 @@ 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/video/src/av1/decode/frame.rs b/crates/video/src/av1/decode/frame.rs index d0432e2a8..e28ca89e9 100644 --- a/crates/video/src/av1/decode/frame.rs +++ b/crates/video/src/av1/decode/frame.rs @@ -197,6 +197,50 @@ impl FrameBuf { } } + /// Copy luma rows [y0, y1) (chroma rows scaled down; clipped to the stored rows) of `src`, a + /// frame of the same geometry. + pub fn copy_rows_from(&mut self, src: &FrameBuf, y0: usize, y1: usize) { + for p in 0..self.num_planes { + let sy = if p == 0 { 0 } else { self.subsampling_y }; + let (d, s) = (&mut self.planes[p], &src.planes[p]); + if d.stride != s.stride { + continue; + } + let a = (y0 >> sy).min(d.rows).min(s.rows) * d.stride; + let b = (y1 >> sy).min(d.rows).min(s.rows) * d.stride; + if let (Some(dst), Some(src)) = (d.data.get_mut(a..b), s.data.get(a..b)) { + dst.copy_from_slice(src); + } + } + } + + /// Copies of luma rows [y0, y1) (chroma rows scaled down; clipped to the stored rows), one + /// plane region each. + pub fn rows_copy(&self, y0: usize, y1: usize) -> [Plane; 3] { + std::array::from_fn(|p| { + if p >= self.num_planes { + return Plane::default(); + } + let sy = if p == 0 { 0 } else { self.subsampling_y }; + let s = &self.planes[p]; + let (a, b) = ((y0 >> sy).min(s.rows), (y1 >> sy).min(s.rows)); + let mut out = Plane::region(0, a, s.stride, b - a); + if let Some(src) = s.data.get(a * s.stride..b * s.stride) { + out.data.copy_from_slice(src); + } + out + }) + } + + /// A frame of this geometry (samples not copied: zero). + pub fn same_geometry(&self) -> FrameBuf { + let planes = std::array::from_fn(|p| { + let pl = &self.planes[p]; + Plane { data: zeroed(pl.data.len()), stride: pl.stride, rows: pl.rows, ox: pl.ox, oy: pl.oy } + }); + FrameBuf { planes, ..*self } + } + /// [`Plane::row_bands`] of all three planes for ranges of luma rows (chroma rows scaled /// down): element `i` holds the bands of range `i`. pub fn row_bands(&mut self, luma: &[(usize, usize)]) -> Vec<[RowBand<'_>; 3]> { @@ -208,6 +252,128 @@ impl FrameBuf { } } +/// Luma rows per published band of a [`BandFrame`] (log2). +pub const BAND_LOG2: u32 = 6; + +/// One plane of a frame's final samples, published band by band while the rest of the frame +/// is still being decoded (see [`BandFrame`]). A band is written once and then only read. +pub struct BandPlane { + bands: Vec>, + stride: usize, + rows: usize, + log2: u32, + /// Returned for rows of a band that is not published: zeros instead of a panic. Readers wait + /// for the row progress first and references of another chroma format are refused, so + /// this only happens through a scheduling bug. + missing: Box<[u16]>, +} + +impl BandPlane { + fn new(stride: usize, rows: usize, log2: u32) -> BandPlane { + let n = rows.div_ceil(1 << log2); + BandPlane { bands: (0..n).map(|_| std::sync::OnceLock::new()).collect(), stride, rows, log2, missing: vec![0; stride].into_boxed_slice() } + } + + /// Row `y` (`stride` samples) of a published band. + #[inline(always)] + pub fn row(&self, y: usize) -> &[u16] { + let o = (y & ((1 << self.log2) - 1)) * self.stride; + match self.bands.get(y >> self.log2).and_then(|b| b.get()).and_then(|b| b.data.get(o..o + self.stride)) { + Some(r) => r, + None => { + debug_assert!(false, "row {y} read before its band was published"); + &self.missing + } + } + } + + /// Publish band `band` copied from the rows of the whole plane `src` (same geometry); a + /// band is published once. + fn publish(&self, band: usize, src: &Plane) { + let y0 = band << self.log2; + let y1 = ((band + 1) << self.log2).min(self.rows); + let mut p = Plane::region(0, y0, self.stride, y1.saturating_sub(y0)); + let w = self.stride.min(src.stride); + for (y, out) in (y0..y1).zip(p.data.chunks_exact_mut(self.stride.max(1))) { + if let Some(r) = src.data.get(y * src.stride..y * src.stride + w) { + out[..w].copy_from_slice(r); + } + } + self.publish_owned(band, p); + } + + /// Publish band `band` as `p`, whose rows start at the band's first row (stride `stride`; + /// rows past the band are ignored). + fn publish_owned(&self, band: usize, p: Plane) { + if let Some(slot) = self.bands.get(band) + && p.stride == self.stride + { + let _ = slot.set(p); + } + } +} + +/// A frame's final samples (every in-loop filter applied) in bands of 64 luma rows, so later +/// frames can predict from the top of a reference while its bottom is still being decoded. +/// Each band is published once, before the frame's row progress says it is final. +pub struct BandFrame { + pub planes: [BandPlane; 3], + pub num_planes: usize, + pub subsampling_x: usize, + pub subsampling_y: usize, +} + +impl BandFrame { + /// Bands for the visible rows of a `width` x `height` frame stored like [`FrameBuf::new`]. + pub fn new(width: usize, height: usize, num_planes: usize, subsampling_x: usize, subsampling_y: usize) -> BandFrame { + let aw = width.div_ceil(128) * 128 + 160; + let planes = std::array::from_fn(|p| { + if p >= num_planes { + return BandPlane::new(0, 0, BAND_LOG2); + } + let (sx, sy) = if p == 0 { (0, 0) } else { (subsampling_x, subsampling_y) }; + BandPlane::new(aw >> sx, (height + sy) >> sy, BAND_LOG2 - sy as u32) + }); + BandFrame { planes, num_planes, subsampling_x, subsampling_y } + } + + /// Number of bands (of 64 luma rows). + pub fn bands(&self) -> usize { + self.planes[0].bands.len() + } + + /// Publish luma band `band` (and the chroma rows it covers) copied from `src`. + pub fn publish(&self, band: usize, src: &FrameBuf) { + for p in 0..self.num_planes { + self.planes[p].publish(band, &src.planes[p]); + } + } + + /// Publish luma band `band` from per-plane buffers starting at the band's first row. + pub fn publish_owned(&self, band: usize, planes: [Plane; 3]) { + for (p, pl) in planes.into_iter().enumerate().take(self.num_planes) { + self.planes[p].publish_owned(band, pl); + } + } +} + +/// A plane to predict from: a whole stored plane or the published bands of a reference. +#[derive(Clone, Copy)] +pub enum PlaneRef<'a> { + Whole(&'a Plane), + Bands(&'a BandPlane), +} + +impl<'a> PlaneRef<'a> { + #[inline(always)] + pub fn row(&self, y: usize) -> &'a [u16] { + match *self { + PlaneRef::Whole(p) => p.row(y), + PlaneRef::Bands(b) => b.row(y), + } + } +} + /// A motion vector (row, col) in 1/8 sample units. #[derive(Clone, Copy, Debug, Default, PartialEq, Eq)] pub struct Mv { diff --git a/crates/video/src/av1/decode/inter.rs b/crates/video/src/av1/decode/inter.rs index 8b40717a3..8d80e96b0 100644 --- a/crates/video/src/av1/decode/inter.rs +++ b/crates/video/src/av1/decode/inter.rs @@ -3,7 +3,7 @@ use std::sync::OnceLock; -use crate::av1::decode::frame::Plane; +use crate::av1::decode::frame::{Plane, PlaneRef}; use crate::av1::decode::mvpred::round2_signed64; use crate::av1::decode::spec_tables::{ BLOCK_8X8, BLOCK_SIZES, COMPOUND_AVERAGE, COMPOUND_DIFFWTD, COMPOUND_DISTANCE, COMPOUND_INTRA, COMPOUND_WEDGE, DIV_LUT, DIV_LUT_BITS, @@ -227,12 +227,9 @@ impl TileDecoder<'_, '_> { (start_x as i32, start_y as i32, step_x as i32, step_y as i32) } - fn ref_plane(&self, ref_idx: i32, plane: usize) -> &Plane { + fn ref_plane(&self, ref_idx: i32, plane: usize) -> PlaneRef<'_> { // A missing reference (rejected earlier for valid streams) predicts from the current frame. - match usize::try_from(ref_idx).ok().and_then(|i| self.fs.refs.get(i)).and_then(|r| r.as_ref()) { - Some(r) => &r.buf.planes[plane], - None => &self.t.cur.planes[plane], - } + self.fs.ref_samples(ref_idx, plane).unwrap_or(PlaneRef::Whole(&self.t.cur.planes[plane])) } /// Block inter prediction process (7.11.3.4) into `pred` (stride PRED_STRIDE); `tmp` holds @@ -512,12 +509,12 @@ impl TileDecoder<'_, '_> { return false; } let fs = self.fs; - let Some(r) = usize::try_from(ref_idx).ok().and_then(|i| fs.refs.get(i)).and_then(|r| r.as_ref()) else { + let Some(refp) = fs.ref_samples(ref_idx, plane) else { return false; }; let dims = self.ref_plane_dims(ref_idx, plane); let max = ((1u32 << fs.bit_depth) - 1) as u16; - copy_fullpel(&r.buf.planes[plane], dims, McBlock { x: sx, y: sy, w, h }, &mut self.t.cur.planes[plane], (x, y), max); + copy_fullpel(refp, dims, McBlock { x: sx, y: sy, w, h }, &mut self.t.cur.planes[plane], (x, y), max); true } @@ -763,7 +760,7 @@ impl TileDecoder<'_, '_> { /// Unscaled single-reference prediction at a whole-sample position: both filter phases are the /// identity and Round2(128 * Round2(128 * v, InterRound0), InterRound1) == v for every bit /// depth, so the prediction is the (clamped) reference sample. -fn copy_fullpel(refp: &Plane, (last_x, last_y): (i32, i32), src_blk: McBlock, dst: &mut Plane, (dx, dy): (usize, usize), max: u16) { +fn copy_fullpel(refp: PlaneRef, (last_x, last_y): (i32, i32), src_blk: McBlock, dst: &mut Plane, (dx, dy): (usize, usize), max: u16) { let McBlock { x, y, w, h } = src_blk; let x0 = x >> 10; let y0 = y >> 10; @@ -831,7 +828,7 @@ fn vpass(rows: &[i32], vf: &[i16; 8], out: &mut [i32], half: i32 /// Block inter prediction without reference scaling: every column / row uses the same filter /// phase, so the filters run over whole rows (bit-identical to the per-sample process). fn mc_unscaled( - refp: &Plane, + refp: PlaneRef, (last_x, last_y): (i32, i32), blk: McBlock, (hf, vf): (&[i16; 8], &[i16; 8]), @@ -982,9 +979,17 @@ mod tests { [(0, 0), (-5, -3), (7, 9), (last_x - w as i32 / 2, last_y - h as i32 / 2), (last_x + 3, last_y + 2), (-(w as i32) - 4, 3)] { let (x, y) = ((px << 10) | 32, (py << 10) | 32); - mc_unscaled(&refp, (last_x, last_y), McBlock { x, y, w, h }, (&IDENTITY_TAPS, &IDENTITY_TAPS), rnd, &mut pred, &mut tmp); + mc_unscaled( + PlaneRef::Whole(&refp), + (last_x, last_y), + McBlock { x, y, w, h }, + (&IDENTITY_TAPS, &IDENTITY_TAPS), + rnd, + &mut pred, + &mut tmp, + ); let mut dst = Plane::new(w + 3, h); - copy_fullpel(&refp, (last_x, last_y), McBlock { x, y, w, h }, &mut dst, (3, 0), max); + copy_fullpel(PlaneRef::Whole(&refp), (last_x, last_y), McBlock { x, y, w, h }, &mut dst, (3, 0), max); for i in 0..h { for j in 0..w { let want = spec_phase0(&refp, (last_x, last_y), (px, py), (j, i), (w * 3 + h) % SUBPEL_FILTERS.len(), rnd) diff --git a/crates/video/src/av1/decode/postfilter.rs b/crates/video/src/av1/decode/postfilter.rs index 6304fdfb6..aa9c2e3ea 100644 --- a/crates/video/src/av1/decode/postfilter.rs +++ b/crates/video/src/av1/decode/postfilter.rs @@ -1,7 +1,7 @@ //! In-loop post filters run by decode_frame_wrapup: loop filter (7.14), CDEF (7.15), //! super-resolution upscaling (7.16) and loop restoration (7.17). -use crate::av1::decode::frame::RowBand; +use crate::av1::decode::frame::{FrameBuf, Plane, RowBand}; use crate::av1::decode::spec_tables::{ GLOBAL_GLOBALMV, GLOBALMV, INTRA_FRAME, MAX_LOOP_FILTER, NEARESTMV, NUM_4X4_BLOCKS_HIGH, NUM_4X4_BLOCKS_WIDE, SEG_LVL_ALT_LF_Y_V, TX_HEIGHT, TX_WIDTH, @@ -48,6 +48,121 @@ pub(crate) fn apply(fs: &mut FrameState, stats: &mut DecodeStats, pool: &crate:: stats.add(Stage::Restoration, t.lap()); } +/// The in-loop filters run band by band behind reconstruction (row wavefront): each stage of a +/// band starts once the stage before has finished the band below it, which bounds what the +/// stage reads beyond the band (the loop filter changes up to 7 rows above a band and intra +/// prediction of the next superblock row reads the unfiltered row above it; CDEF reads 2 rows +/// and loop restoration 3 rows beyond a band). Bands are 64 luma rows. The result is the same +/// as running each filter over the whole frame. Super-resolution frames use [`apply`]. +/// +/// The last active stage writes each band into its own buffers, which are handed out as the +/// band's final samples without a copy. +pub(crate) struct Pipeline { + bands: Vec<(usize, usize)>, + /// Luma rows of the coded frame (mi_rows * 4). + rows: usize, + lf_on: bool, + cdef_on: bool, + lr_on: bool, + /// Bands finished by each stage. + lf: usize, + cdef: usize, + lr: usize, + pending: [Vec; 3], + /// CdefFrame when loop restoration reads it. + cdef_out: Option, +} + +/// Final bands from [`Pipeline::advance`]: (band, its samples starting at the band's first row). +pub(crate) type FinalBands = Vec<(usize, [Plane; 3])>; + +impl Pipeline { + pub fn new(fs: &FrameState, draft: bool) -> Pipeline { + let rows = fs.fh.mi_rows as usize * 4; + let n = rows.div_ceil(BAND_ROWS).max(1); + let bands = (0..n).map(|b| (b * BAND_ROWS, if b + 1 == n { usize::MAX } else { (b + 1) * BAND_ROWS })).collect(); + let lvl = fs.fh.lf.level; + let cdef_on = !draft && crate::av1::decode::cdef::active(fs); + let lr_on = fs.fh.lr.uses_lr && !draft; + let cdef_out = (cdef_on && lr_on).then(|| fs.cur.same_geometry()); + Pipeline { + bands, + rows, + lf_on: (lvl[0] != 0 || lvl[1] != 0) && !draft, + cdef_on, + lr_on, + lf: 0, + cdef: 0, + lr: 0, + pending: Default::default(), + cdef_out, + } + } + + /// Filter what luma rows [0, `recon_rows`) of reconstructed samples allow (everything once + /// they cover the frame); returns the bands whose samples became final. + pub fn advance(&mut self, fs: &mut FrameState, stats: &mut DecodeStats, pool: &crate::av1::decode::par::Pool, recon_rows: usize) -> FinalBands { + let n = self.bands.len(); + let all = recon_rows >= self.rows; + let mut out = FinalBands::new(); + let mut t = Timer::start(); + // The loop filter of a band waits for the band below it to be reconstructed. + let lf_to = if all { n } else { (recon_rows / BAND_ROWS).saturating_sub(1) }; + if lf_to > self.lf { + if self.lf_on { + loop_filter_bands(fs, pool, &self.bands, self.lf..lf_to, &mut self.pending); + } + self.lf = lf_to; + } + stats.add(Stage::LoopFilter, t.lap()); + let cdef_to = if all { n } else { self.lf.saturating_sub(1) }; + if cdef_to > self.cdef { + let r = &self.bands[self.cdef..cdef_to]; + if self.cdef_on { + if let Some(cdef) = &mut self.cdef_out { + cdef.copy_rows_from(&fs.cur, r[0].0, r[r.len() - 1].1); + crate::av1::decode::cdef::filter_bands(fs, pool, cdef.row_bands(r), r); + } else { + let mut own: Vec<[Plane; 3]> = r.iter().map(|&(a, b)| fs.cur.rows_copy(a, b)).collect(); + crate::av1::decode::cdef::filter_bands(fs, pool, own_bands(&mut own), r); + out.extend((self.cdef..cdef_to).zip(own)); + } + } else if !self.lr_on { + // The loop filter is the last stage: the band is final once the band below it + // is filtered. + out.extend((self.cdef..cdef_to).zip(r.iter().map(|&(a, b)| fs.cur.rows_copy(a, b)))); + } + self.cdef = cdef_to; + } + stats.add(Stage::Cdef, t.lap()); + let lr_to = if all { n } else { self.cdef.saturating_sub(1) }; + if self.lr_on && lr_to > self.lr { + let r = &self.bands[self.lr..lr_to]; + let cdef = self.cdef_out.as_ref().unwrap_or(&fs.cur); + let mut own: Vec<[Plane; 3]> = r.iter().map(|&(a, b)| cdef.rows_copy(a, b)).collect(); + crate::av1::decode::restoration::restore_bands(fs, &fs.cur, cdef, own_bands(&mut own), pool, r); + out.extend((self.lr..lr_to).zip(own)); + self.lr = lr_to; + } + stats.add(Stage::Restoration, t.lap()); + out + } +} + +/// Row bands over per-band buffers made by [`FrameBuf::rows_copy`]. +fn own_bands(own: &mut [[Plane; 3]]) -> Vec<[RowBand<'_>; 3]> { + fn band(p: &mut Plane) -> RowBand<'_> { + let (stride, y0) = (p.stride.max(1), p.oy); + RowBand { data: p.data.as_mut_slice(), stride, y0 } + } + own.iter_mut() + .map(|planes| { + let [a, b, c] = planes; + [band(a), band(b), band(c)] + }) + .collect() +} + // --------------------------------------------------------------------------------------------- // Loop filter (7.14) @@ -69,23 +184,40 @@ const LF_MARGIN: usize = 8; /// written back; the band above never touches them. fn loop_filter(fs: &mut FrameState, pool: &crate::av1::decode::par::Pool) { let luma = bands(pool, fs.fh.mi_rows as usize * 4); + loop_filter_bands(fs, pool, &luma, 0..luma.len(), &mut Default::default()); +} + +/// The loop filter on bands `range` of `all` (the frame's bands). The bands above the range +/// are filtered already: `pending` holds the unfiltered rows above the first band of the range +/// (empty for band 0), saved before the band above was filtered, and is set to those of the +/// band after the range. +fn loop_filter_bands( + fs: &mut FrameState, + pool: &crate::av1::decode::par::Pool, + all: &[(usize, usize)], + range: std::ops::Range, + pending: &mut [Vec; 3], +) { + let Some(luma) = all.get(range.clone()) else { return }; let planes: Vec = (0..fs.num_planes).filter(|&p| p == 0 || fs.fh.lf.level[1 + p] != 0).collect(); let mut cur = std::mem::take(&mut fs.cur); let f: &FrameState = fs; - // Unfiltered copies of the rows above each band but the first. - let margins: Vec<[Vec; 3]> = luma - .iter() - .map(|&(start, _)| { - std::array::from_fn(|p| { - let sub_y = if p == 0 { 0 } else { f.ssy }; - let pl = &cur.planes[p]; - let (y0, y1) = ((start.saturating_sub(LF_MARGIN)) >> sub_y, start >> sub_y); - let rows = pl.data.get(y0 * pl.stride..y1 * pl.stride); - if start == 0 || !planes.contains(&p) { Vec::new() } else { rows.map(<[u16]>::to_vec).unwrap_or_default() } - }) + // Unfiltered copies of the rows above a band (none above the first). + let margin = |cur: &crate::av1::decode::frame::FrameBuf, start: usize| -> [Vec; 3] { + std::array::from_fn(|p| { + let sub_y = if p == 0 { 0 } else { f.ssy }; + let pl = &cur.planes[p]; + let (y0, y1) = ((start.saturating_sub(LF_MARGIN)) >> sub_y, start >> sub_y); + let rows = pl.data.get(y0 * pl.stride..y1 * pl.stride); + if start == 0 || !planes.contains(&p) { Vec::new() } else { rows.map(<[u16]>::to_vec).unwrap_or_default() } }) - .collect(); - let mut bands: Vec<([RowBand; 3], [Vec; 3])> = cur.row_bands(&luma).into_iter().zip(margins).collect(); + }; + let margins: Vec<[Vec; 3]> = + luma.iter().enumerate().map(|(i, &(start, _))| if i == 0 { std::mem::take(pending) } else { margin(&cur, start) }).collect(); + if let Some(&(next, _)) = all.get(range.end) { + *pending = margin(&cur, next); + } + let mut bands: Vec<([RowBand; 3], [Vec; 3])> = cur.row_bands(luma).into_iter().zip(margins).collect(); let changed = pool.map_mut(&mut bands, |b, (band, margin)| { let mut changed: [Vec<(usize, u16)>; 3] = Default::default(); for &p in &planes { diff --git a/crates/video/src/av1/decode/restoration.rs b/crates/video/src/av1/decode/restoration.rs index efc5a751f..d252f7b06 100644 --- a/crates/video/src/av1/decode/restoration.rs +++ b/crates/video/src/av1/decode/restoration.rs @@ -70,10 +70,16 @@ struct LrPlane<'a> { /// Bands of 64 luma rows are restored in parallel: every unit reads only the two input frames /// and writes only its own samples. pub(crate) fn loop_restoration(fs: &FrameState, cur: &FrameBuf, cdef: &FrameBuf, pool: &Pool) -> FrameBuf { - let fh = &fs.fh; let mut lr = cdef.clone(); - let ranges = bands(pool, fh.frame_height as usize); - let mut planes = lr.row_bands(&ranges); + let ranges = bands(pool, fs.fh.frame_height as usize); + restore_bands(fs, cur, cdef, lr.row_bands(&ranges), pool, &ranges); + lr +} + +/// Restore the bands `planes` covering luma rows `ranges`, which hold UpscaledCdefFrame; reads +/// both inputs up to 3 rows beyond each band. +pub(crate) fn restore_bands(fs: &FrameState, cur: &FrameBuf, cdef: &FrameBuf, mut planes: Vec<[RowBand; 3]>, pool: &Pool, ranges: &[(usize, usize)]) { + let fh = &fs.fh; pool.map_mut(&mut planes, |b, band| { let (start, end) = ranges[b]; let mut scratch = Scratch::default(); @@ -100,8 +106,6 @@ pub(crate) fn loop_restoration(fs: &FrameState, cur: &FrameBuf, cdef: &FrameBuf, p.restore(out, start >> sub_y, end >> sub_y, &mut scratch); } }); - drop(planes); - lr } /// Row buffers reused across stripes. diff --git a/crates/video/src/av1/decode/state.rs b/crates/video/src/av1/decode/state.rs index 4c4806f16..f498a7f05 100644 --- a/crates/video/src/av1/decode/state.rs +++ b/crates/video/src/av1/decode/state.rs @@ -6,8 +6,8 @@ //! with several, each tile decodes into private buffers covering its area (tiles never read //! outside themselves), which are then copied into the frame. -use crate::av1::decode::decoder::RefFrame; -use crate::av1::decode::frame::{FrameBuf, MiInfo}; +use crate::av1::decode::decoder::{Parsed, RefFrame}; +use crate::av1::decode::frame::{BandFrame, FrameBuf, MiInfo, PlaneRef}; use crate::av1::decode::header::{FrameHeader, RefInfo, SequenceHeader}; use crate::av1::decode::spec_tables::{DEFAULT_INTRA_FRAME_Y_MODE_CDF, NUM_REF_FRAMES, RESTORE_NONE, SUBSAMPLED_SIZE}; use std::ops::{Deref, DerefMut}; @@ -26,6 +26,11 @@ pub(crate) struct FrameShared { pub cdef_cols: usize, /// Reference slots (FrameStore and saved state) used by this frame. pub refs: [Option>; NUM_REF_FRAMES], + /// Parse results (CDFs, segment ids, saved motion vectors) of the same slots. + pub parsed: [Option>; NUM_REF_FRAMES], + /// The published bands of references that are still being decoded (row wavefront): read + /// only below the row progress that was waited for. + pub ref_rows: [Option>; NUM_REF_FRAMES], pub ref_info: [RefInfo; NUM_REF_FRAMES], /// MotionFieldMvs[ ref ][ y8 * w8 + x8 ] (7.9); empty when unused. pub motion_field: [Vec<[i32; 2]>; 8], @@ -34,6 +39,16 @@ pub(crate) struct FrameShared { } impl FrameShared { + /// The samples of reference slot `ref_idx` to predict from (None for a missing reference). + #[inline] + pub fn ref_samples(&self, ref_idx: i32, plane: usize) -> Option> { + let i = usize::try_from(ref_idx).ok()?; + if let Some(b) = self.ref_rows.get(i)?.as_ref() { + return b.planes.get(plane).map(PlaneRef::Bands); + } + self.refs.get(i)?.as_ref().map(|r| r.plane(plane)) + } + pub fn plane_residual_size(&self, subsize: usize, plane: usize) -> usize { let sx = if plane > 0 { self.ssx } else { 0 }; let sy = if plane > 0 { self.ssy } else { 0 }; @@ -210,6 +225,8 @@ impl FrameState { lr_unit_cols, cdef_cols, refs: Default::default(), + parsed: Default::default(), + ref_rows: Default::default(), ref_info: Default::default(), motion_field: Default::default(), prev_segment_ids: vec![0; mi_cols * mi_rows], @@ -283,6 +300,14 @@ impl FrameState { pub fn merge_tile_state(&mut self, t: &TileState) { for p in 0..self.num_planes { t.cur.planes[p].copy_into(&mut self.cur.planes[p]); + } + self.merge_tile_modes(t); + } + + /// Copy a tile's private buffers except the samples (after parsing, before reconstruction) + /// into the frame. + pub fn merge_tile_modes(&mut self, t: &TileState) { + for p in 0..self.num_planes { t.lf_tx_size[p].copy_into(&mut self.lf_tx_size[p]); for &u in &t.lr_written[p] { self.lr.lr_type[p][u] = t.lr.lr_type[p][u]; diff --git a/crates/video/src/av1/decode/tile.rs b/crates/video/src/av1/decode/tile.rs index 16d59ed3a..2572f9d21 100644 --- a/crates/video/src/av1/decode/tile.rs +++ b/crates/video/src/av1/decode/tile.rs @@ -94,6 +94,191 @@ pub(crate) struct Block { pub local_warp_params: [i32; 6], } +/// What a tile decoder does with the blocks it visits. +#[derive(Clone, Copy, PartialEq, Eq, Debug)] +pub(crate) enum Pass { + /// Read the symbols and reconstruct each block right away (single-threaded decoding). + Full, + /// Read the symbols only, recording what reconstruction needs in a [`TileRecord`]. + Parse, + /// Reconstruct from a [`TileRecord`], reading no symbols. + Reconstruct, +} + +/// The per-block values reconstruction needs beyond the mode info (which is final after +/// parsing): the block variables as parsed, its quantizer index and the warp samples / palette +/// color maps when used. +#[derive(Clone)] +pub(crate) struct BlockRecord { + b: Block, + current_q_index: i32, + warp_samples: Option>, + color_maps: Option, Vec)>>, +} + +/// A parsed tile: its blocks in decoding order, how many of them each superblock holds and the +/// coefficients of every coded transform block as a stream of +/// `eob, tx_type, count, (position, value) * count` (only non-zero values). +#[derive(Default)] +pub(crate) struct TileRecord { + blocks: Vec, + sb_blocks: Vec, + coeffs: Vec, + /// Read positions while reconstructing. + next_block: usize, + next_coeff: usize, + next_sb: usize, + next_sb_row: usize, + /// Bytes recorded and the most allowed. + pub bytes: usize, + pub limit: usize, + /// The decoder-wide budget the recorded bytes are reserved from, and how much is reserved. + pub budget: Option>, + reserved: usize, +} + +impl Drop for TileRecord { + fn drop(&mut self) { + if let Some(b) = &self.budget { + b.release(self.reserved); + } + } +} + +/// The parse record bytes all frames of a decoder may hold at once (frames in flight decode in +/// one pass once it is used up). +pub(crate) struct RecordBudget { + used: std::sync::atomic::AtomicUsize, + max: usize, + /// Most bytes ever reserved at once (tests). + #[cfg(test)] + pub peak: std::sync::atomic::AtomicUsize, +} + +/// Bytes reserved from a [`RecordBudget`] at a time. +const RESERVE_STEP: usize = 1 << 20; + +impl RecordBudget { + pub fn new(max: usize) -> RecordBudget { + RecordBudget { + used: Default::default(), + max, + #[cfg(test)] + peak: Default::default(), + } + } + + /// Reserve `n` bytes; false (nothing reserved) if that would go over the budget. + fn reserve(&self, n: usize) -> bool { + use std::sync::atomic::Ordering; + let ok = self.used.try_update(Ordering::AcqRel, Ordering::Acquire, |u| u.checked_add(n).filter(|&v| v <= self.max)).is_ok(); + #[cfg(test)] + if ok { + self.peak.fetch_max(self.used.load(Ordering::Acquire), Ordering::AcqRel); + } + ok + } + + fn release(&self, n: usize) { + self.used.fetch_sub(n, std::sync::atomic::Ordering::AcqRel); + } + + /// Bytes reserved now (tests). + #[cfg(test)] + pub fn used(&self) -> usize { + self.used.load(std::sync::atomic::Ordering::Acquire) + } +} + +/// The motion of a recorded block (what a reference row wait depends on). +pub(crate) struct BlockMotion { + pub mi_row: usize, + pub bh4: usize, + /// Inter prediction from reference frames (not intra block copy). + pub inter: bool, + pub ref_frame: [i8; 2], + /// Mv[ list ] (row, col) in 1/8 samples. + pub mv: [[i32; 2]; 2], + pub motion_mode: u8, + pub y_mode: usize, +} + +impl TileRecord { + /// Whether recording stopped at [`TileRecord::limit`] (the record is unusable). + pub fn over_limit(&self) -> bool { + self.bytes > self.limit + } + + /// Account for `n` more recorded bytes; false once over the limit (nothing more is kept). + fn grow(&mut self, n: usize) -> bool { + self.bytes = self.bytes.saturating_add(n); + if self.bytes > self.limit { + return false; + } + if self.bytes > self.reserved + && let Some(b) = &self.budget + { + let step = (self.bytes - self.reserved).max(RESERVE_STEP); + if !b.reserve(step) { + // The decoder-wide budget is used up: this frame is decoded in one pass. + self.limit = 0; + return false; + } + self.reserved += step; + } + true + } + /// The recorded blocks of superblock rows [first, end), for a tile `sbs_per_row` superblocks wide. + pub fn row_blocks(&self, sbs_per_row: usize, first: usize, end: usize) -> impl Iterator + '_ { + let start: usize = self.sb_blocks.iter().take(first * sbs_per_row).map(|&n| n as usize).sum(); + let count: usize = self.sb_blocks.iter().skip(first * sbs_per_row).take((end - first) * sbs_per_row).map(|&n| n as usize).sum(); + self.blocks.iter().skip(start).take(count).map(|r| BlockMotion { + mi_row: r.b.mi_row, + bh4: r.b.bh4, + inter: r.b.is_inter && !r.b.use_intrabc, + ref_frame: r.b.ref_frame, + mv: r.b.mv_i, + motion_mode: r.b.motion_mode, + y_mode: r.b.y_mode, + }) + } + + fn next_coeff(&mut self) -> Result { + let v = *self.coeffs.get(self.next_coeff).ok_or(crate::av1::decode::Error::Invalid("coefficient record exhausted"))?; + self.next_coeff += 1; + Ok(v) + } +} + +/// The work buffers of a [`TileDecoder`] (`quant` all zero between transform blocks). +pub(crate) struct Buffers { + quant: Vec, + dequant: Vec, + residual: Vec, + color_map_y: Vec, + color_map_uv: Vec, + pred_buf: [Vec; 2], + mask: Vec, + mc_tmp: Vec, + levels: Vec, +} + +impl Buffers { + pub fn new() -> Buffers { + Buffers { + quant: vec![0; 1024], + dequant: vec![0; 64 * 64], + residual: vec![0; 64 * 64], + color_map_y: vec![0; 64 * 64], + color_map_uv: vec![0; 64 * 64], + pred_buf: [vec![0; 128 * 128], vec![0; 128 * 128]], + mask: vec![0; 128 * 128], + mc_tmp: vec![0; (2 * 128 + 8) * 128], + levels: vec![0; (32 + TX_PAD) * (32 + TX_PAD)], + } + } +} + /// Decoder for one tile. pub(crate) struct TileDecoder<'a, 'f> { pub fs: &'f FrameShared, @@ -128,6 +313,9 @@ pub(crate) struct TileDecoder<'a, 'f> { pub mc_tmp: Vec, /// Coefficient levels of the current transform block, padded (see TX_PAD). levels: Vec, + pub pass: Pass, + /// Written by [`Pass::Parse`], read by [`Pass::Reconstruct`]. + pub record: TileRecord, } const SB_MAX: usize = 32; // superblock size in 4x4 units (128 / 4) @@ -141,6 +329,21 @@ fn block_height(bs: usize) -> usize { impl<'a, 'f> TileDecoder<'a, 'f> { pub fn new(fs: &'f FrameShared, t: &'f mut TileState, data: &'a [u8], cdf: Box, tile_row: usize, tile_col: usize) -> Self { + Self::with_buffers(fs, t, data, cdf, tile_row, tile_col, Buffers::new()) + } + + /// [`TileDecoder::new`] reusing the work buffers of an earlier decoder of the same frame + /// (see [`TileDecoder::into_buffers`]). + pub fn with_buffers( + fs: &'f FrameShared, + t: &'f mut TileState, + data: &'a [u8], + cdf: Box, + tile_row: usize, + tile_col: usize, + bufs: Buffers, + ) -> Self { + let Buffers { quant, dequant, residual, color_map_y, color_map_uv, pred_buf, mask, mc_tmp, levels } = bufs; let ti = &fs.fh.tile_info; let (rs, re) = (ti.mi_row_starts[tile_row] as usize, ti.mi_row_starts[tile_row + 1] as usize); let (cs, ce) = (ti.mi_col_starts[tile_col] as usize, ti.mi_col_starts[tile_col + 1] as usize); @@ -164,19 +367,27 @@ impl<'a, 'f> TileDecoder<'a, 'f> { b: Block::default(), mvs: Default::default(), nb: Default::default(), - quant: vec![0; 1024], - dequant: vec![0; 64 * 64], - residual: vec![0; 64 * 64], - color_map_y: vec![0; 64 * 64], - color_map_uv: vec![0; 64 * 64], + quant, + dequant, + residual, + color_map_y, + color_map_uv, plane_tx_type: 0, - pred_buf: [vec![0; 128 * 128], vec![0; 128 * 128]], - mask: vec![0; 128 * 128], - mc_tmp: vec![0; (2 * 128 + 8) * 128], - levels: vec![0; (32 + TX_PAD) * (32 + TX_PAD)], + pred_buf, + mask, + mc_tmp, + levels, + pass: Pass::Full, + record: TileRecord::default(), } } + /// The work buffers and CDFs, for [`TileDecoder::with_buffers`]. + pub fn into_buffers(self) -> (Buffers, Box) { + let TileDecoder { quant, dequant, residual, color_map_y, color_map_uv, pred_buf, mask, mc_tmp, levels, cdf, .. } = self; + (Buffers { quant, dequant, residual, color_map_y, color_map_uv, pred_buf, mask, mc_tmp, levels }, cdf) + } + fn seq(&self) -> &SequenceHeader { &self.fs.seq } @@ -189,8 +400,11 @@ impl<'a, 'f> TileDecoder<'a, 'f> { c >= self.mi_col_start as isize && c < self.mi_col_end as isize && r >= self.mi_row_start as isize && r < self.mi_row_end as isize } - /// decode_tile( ) + /// decode_tile( ), or its reconstruction from [`TileDecoder::record`] in [`Pass::Reconstruct`]. pub fn decode_tile(&mut self) -> Result<()> { + if self.pass == Pass::Reconstruct { + return self.reconstruct_tile(); + } // clear_above_context( ) let t = &mut *self.t; for p in 0..3 { @@ -223,7 +437,11 @@ impl<'a, 'f> TileDecoder<'a, 'f> { self.clear_cdef(r, c); self.clear_block_decoded_flags(r, c, sb4); self.read_lr(r, c, sb_size); + let before = self.record.blocks.len(); self.decode_partition(r, c, sb_size)?; + if self.pass == Pass::Parse { + self.record.sb_blocks.push((self.record.blocks.len() - before) as u32); + } c += sb4; } r += sb4; @@ -231,6 +449,76 @@ impl<'a, 'f> TileDecoder<'a, 'f> { Ok(()) } + /// The reconstruction half of decode_tile( ): the superblocks in the same order, each block + /// predicted and reconstructed from its record. + fn reconstruct_tile(&mut self) -> Result<()> { + self.reconstruct_sb_rows(usize::MAX) + } + + /// Reconstruct the next `rows` superblock rows of the tile (the record keeps the position). + pub fn reconstruct_sb_rows(&mut self, rows: usize) -> Result<()> { + let sb4 = if self.seq().use_128x128_superblock { 32 } else { 16 }; + let mut r = self.mi_row_start + self.record.next_sb_row * sb4; + let mut done = 0; + while r < self.mi_row_end && done < rows { + let mut c = self.mi_col_start; + while c < self.mi_col_end { + self.clear_block_decoded_flags(r, c, sb4); + let n = *self.record.sb_blocks.get(self.record.next_sb).ok_or(crate::av1::decode::Error::Invalid("block record exhausted"))?; + self.record.next_sb += 1; + for _ in 0..n { + self.reconstruct_block()?; + } + c += sb4; + } + r += sb4; + done += 1; + self.record.next_sb_row += 1; + } + Ok(()) + } + + fn reconstruct_block(&mut self) -> Result<()> { + let rec = self.record.blocks.get(self.record.next_block).ok_or(crate::av1::decode::Error::Invalid("block record exhausted"))?; + self.record.next_block += 1; + self.b = rec.b.clone(); + self.current_q_index = rec.current_q_index; + if let Some(w) = &rec.warp_samples { + self.mvs.num_samples = w.0; + self.mvs.cand_list = w.1; + } + if let Some(maps) = &rec.color_maps { + if let Some(d) = self.color_map_y.get_mut(..maps.0.len()) { + d.copy_from_slice(&maps.0); + } + if let Some(d) = self.color_map_uv.get_mut(..maps.1.len()) { + d.copy_from_slice(&maps.1); + } + } + self.compute_prediction()?; + self.residual() + } + + /// Records the block for [`Pass::Reconstruct`] (its variables as parsed). + fn record_block(&mut self) { + let b = &self.b; + // The rows of the color maps (stride 64) the block's palette prediction reads. + let map_rows = |n: usize, h: usize| if n > 0 { (h + 2).min(64) * 64 } else { 0 }; + let rows_y = map_rows(b.palette_size_y, block_height(b.mi_size)); + let rows_uv = map_rows(b.palette_size_uv, block_height(b.mi_size) >> self.fs.ssy); + let size = std::mem::size_of::() + + rows_y + + rows_uv + + if b.motion_mode as usize == crate::av1::decode::spec_tables::LOCALWARP { 160 } else { 0 }; + if !self.record.grow(size) { + return; + } + let warp_samples = + (b.motion_mode as usize == crate::av1::decode::spec_tables::LOCALWARP).then(|| Box::new((self.mvs.num_samples, self.mvs.cand_list))); + let color_maps = (rows_y + rows_uv > 0).then(|| Box::new((self.color_map_y[..rows_y].to_vec(), self.color_map_uv[..rows_uv].to_vec()))); + self.record.blocks.push(BlockRecord { b: b.clone(), current_q_index: self.current_q_index, warp_samples, color_maps }); + } + fn clear_cdef(&mut self, r: usize, c: usize) { let t = &mut *self.t; t.set_cdef_idx(r, c, -1); @@ -461,7 +749,11 @@ impl<'a, 'f> TileDecoder<'a, 'f> { } } } - self.compute_prediction()?; + if self.pass == Pass::Parse { + self.record_block(); + } else { + self.compute_prediction()?; + } self.residual()?; let b = &self.b; let delta_lf = self.delta_lf; diff --git a/crates/video/src/av1/decode/tile/coeffs.rs b/crates/video/src/av1/decode/tile/coeffs.rs index 6a704f8ac..df358e68d 100644 --- a/crates/video/src/av1/decode/tile/coeffs.rs +++ b/crates/video/src/av1/decode/tile/coeffs.rs @@ -459,6 +459,58 @@ impl<'a, 'f> TileDecoder<'a, 'f> { } } + /// Moves the coefficients `coeffs` left in `quant` into the record (leaving `quant` zero, as + /// reconstruction does). + pub(super) fn record_coeffs(&mut self, eob: usize, tx_sz: usize) { + let scan = self.get_scan(tx_sz); + let scan = scan.get(..eob).unwrap_or(scan); + if !self.record.grow(4 * (3 + 2 * scan.len())) { + // Over the limit: the record is dropped (the frame is decoded again in one pass); + // `quant` must still be left zero. + for &pos in scan { + if let Some(q) = self.quant.get_mut(pos as usize) { + *q = 0; + } + } + return; + } + let rec = &mut self.record.coeffs; + rec.push(eob as i32); + if eob == 0 { + return; + } + rec.push(self.plane_tx_type as i32); + let count_at = rec.len(); + rec.push(0); + let mut count = 0; + for &pos in scan { + let v = std::mem::take(&mut self.quant[pos as usize]); + if v != 0 { + rec.push(pos as i32); + rec.push(v); + count += 1; + } + } + rec[count_at] = count; + } + + /// The next transform block's coefficients from the record into `quant` (and its type); + /// returns its eob. + pub(super) fn replay_coeffs(&mut self) -> crate::av1::decode::Result { + let eob = self.record.next_coeff()?; + if eob <= 0 { + return Ok(0); + } + self.plane_tx_type = usize::try_from(self.record.next_coeff()?).map_err(|_| crate::av1::decode::Error::Invalid("coefficient record"))?; + let count = self.record.next_coeff()?; + for _ in 0..count { + let pos = self.record.next_coeff()? as usize; + let v = self.record.next_coeff()?; + *self.quant.get_mut(pos).ok_or(crate::av1::decode::Error::Invalid("coefficient record"))? = v; + } + Ok(eob as usize) + } + /// Reconstruct process (7.12.3). pub(super) fn reconstruct(&mut self, plane: usize, x: usize, y: usize, tx_sz: usize, dc_only: bool) { // dqDenom (7.12.3) is 1, 2 or 4 and the dividend is non-negative. diff --git a/crates/video/src/av1/decode/tile/predict.rs b/crates/video/src/av1/decode/tile/predict.rs index b1a53f399..244be62c6 100644 --- a/crates/video/src/av1/decode/tile/predict.rs +++ b/crates/video/src/av1/decode/tile/predict.rs @@ -213,7 +213,8 @@ impl<'a, 'f> TileDecoder<'a, 'f> { if start_x >= max_x || start_y >= max_y { return Ok(()); } - if !self.b.is_inter { + let parse_only = self.pass == super::Pass::Parse; + if !self.b.is_inter && !parse_only { let palette = if plane == 0 { self.b.palette_size_y } else { self.b.palette_size_uv }; if palette > 0 { self.predict_palette(plane, start_x, start_y, x, y, tx_sz); @@ -266,9 +267,23 @@ impl<'a, 'f> TileDecoder<'a, 'f> { } } if !self.b.skip { - let eob = self.coeffs(plane, start_x, start_y, tx_sz); - if eob > 0 { - self.reconstruct(plane, start_x, start_y, tx_sz, eob == 1); + match self.pass { + super::Pass::Full => { + let eob = self.coeffs(plane, start_x, start_y, tx_sz); + if eob > 0 { + self.reconstruct(plane, start_x, start_y, tx_sz, eob == 1); + } + } + super::Pass::Parse => { + let eob = self.coeffs(plane, start_x, start_y, tx_sz); + self.record_coeffs(eob, tx_sz); + } + super::Pass::Reconstruct => { + let eob = self.replay_coeffs()?; + if eob > 0 { + self.reconstruct(plane, start_x, start_y, tx_sz, eob == 1); + } + } } } for i in 0..step_y { diff --git a/crates/video/tests/av1_decode/fuzz_hostile.rs b/crates/video/tests/av1_decode/fuzz_hostile.rs index 05dbbe126..31cb66073 100644 --- a/crates/video/tests/av1_decode/fuzz_hostile.rs +++ b/crates/video/tests/av1_decode/fuzz_hostile.rs @@ -1,7 +1,6 @@ //! Hostile-input robustness: truncation, header-region sweeps and random mutation of small //! ffmpeg-generated AV1 streams (frame by frame) must never panic or run away. Skips without ffmpeg. -use std::panic::{AssertUnwindSafe, catch_unwind}; use std::process::{Command, Stdio}; use std::sync::atomic::{AtomicUsize, Ordering}; @@ -20,8 +19,15 @@ impl Rng { } } -/// (name, encoder args, pix_fmt) -const SEEDS: &[(&str, &str, &str)] = &[("av1_fz_default", "-c:v libsvtav1 -g 4", "yuv420p"), ("av1_fz_10bit", "-c:v libsvtav1 -g 6", "yuv420p10le")]; +// Seeds are at least 64x64: SVT-AV1 before 3.0 (the Ubuntu runners have 2.3) refuses a source under 64 lines. +/// (name, encoder args, pix_fmt, size, decoder thread counts) +const SEEDS: &[(&str, &str, &str, &str, &[usize])] = &[ + ("av1_fz_default", "-c:v libsvtav1 -g 4", "yuv420p", "64x64", &[1, 3]), + ("av1_fz_10bit", "-c:v libsvtav1 -g 6", "yuv420p10le", "64x64", &[1, 3]), + // Several superblock rows, so corrupt references meet frames already predicting from their + // top rows (row wavefront with frame threads). + ("av1_fz_rows", "-c:v libsvtav1 -g 6", "yuv420p", "64x136", &[1, 8]), +]; fn ivf_frames(data: &[u8]) -> Vec> { let mut out = Vec::new(); @@ -37,14 +43,13 @@ fn ivf_frames(data: &[u8]) -> Vec> { out } -// Seeds are 64x64: SVT-AV1 before 3.0 (the Ubuntu runners have 2.3) refuses a source under 64 lines. -fn make(ff: &std::path::Path, name: &str, args: &str, pix: &str) -> Option>> { +fn make(ff: &std::path::Path, name: &str, args: &str, pix: &str, size: &str) -> Option>> { let out = craft_media_testkit::fixtures_dir("av1").join(format!("{name}.ivf")); let p = craft_media_testkit::fixtures::generate(&out, |tmp| { let a: Vec<&str> = args.split_whitespace().collect(); matches!( Command::new(ff) - .args(["-hide_banner", "-loglevel", "error", "-y", "-f", "lavfi", "-i", "testsrc2=size=64x64:rate=25", "-frames:v", "6", "-pix_fmt", pix]) + .args(["-hide_banner", "-loglevel", "error", "-y", "-f", "lavfi", "-i", &format!("testsrc2=size={size}:rate=25"), "-frames:v", "6", "-pix_fmt", pix]) .args(a) .args(["-f", "ivf"]) .arg(tmp) @@ -69,20 +74,29 @@ fn count_panics() { }); } -fn run_all(frames: &[Vec], label: &str) { +fn run_all(frames: &[Vec], threads: &[usize], label: &str) { craft_media_testkit::set_progress(label); count_panics(); let before = PANICS.load(Ordering::SeqCst); - let res = catch_unwind(AssertUnwindSafe(|| { - for threads in [1, 3] { - let mut dec = Decoder::with_threads(threads); - for (i, f) in frames.iter().enumerate() { - let _ = dec.decode_pts(f, i as i64); - } - let _ = dec.flush_result(); - } - })); - assert!(res.is_ok() && PANICS.load(Ordering::SeqCst) == before, "panic: {label}"); + // Every thread count decodes the stream at the same time, each on a thread of its own. + let ok = std::thread::scope(|s| { + let runs: Vec<_> = threads + .iter() + .map(|&threads| { + s.spawn(move || { + let mut dec = Decoder::with_threads(threads); + for (i, f) in frames.iter().enumerate() { + let _ = dec.decode_pts(f, i as i64); + } + let _ = dec.flush_result(); + }) + }) + .collect(); + // Join every run (no short cut), then check them all. + let joined: Vec = runs.into_iter().map(|r| r.join().is_ok()).collect(); + joined.into_iter().all(|ok| ok) + }); + assert!(ok && PANICS.load(Ordering::SeqCst) == before, "panic: {label}"); } /// Random-mutation rounds per seed (raise with FILMCRAFT_FUZZ_ITERS for a long hunt). @@ -95,17 +109,23 @@ const DEADLINE_SECS: u64 = 600; #[test] fn hostile_streams_never_panic() { - craft_media_testkit::with_deadline(DEADLINE_SECS, sweep); + craft_media_testkit::with_deadline(DEADLINE_SECS, || sweep(&SEEDS[..2])); } -fn sweep() { +/// The seed with several superblock rows, in a test of its own so it runs beside the others. +#[test] +fn hostile_streams_with_superblock_rows_never_panic() { + craft_media_testkit::with_deadline(DEADLINE_SECS, || sweep(&SEEDS[2..])); +} + +fn sweep(seeds: &[(&str, &str, &str, &str, &[usize])]) { let Some(ff) = craft_media_testkit::ffmpeg_or_skip("AV1 fuzz") else { return }; - for &(name, args, pix) in SEEDS { + for &(name, args, pix, size, threads) in seeds { if !craft_media_testkit::ffmpeg_has_encoder(&ff, "libsvtav1") { craft_media_testkit::skip_test("AV1 fuzz", &format!("this ffmpeg has no libsvtav1 encoder (seed {name})")); continue; } - let Some(seed) = make(&ff, name, args, pix) else { + let Some(seed) = make(&ff, name, args, pix, size) else { panic!("ffmpeg has libsvtav1 but could not encode the {name} seed"); }; if seed.is_empty() { @@ -118,7 +138,7 @@ fn sweep() { for n in (0..len).step_by((len / 200).max(1)) { let mut d = seed.clone(); d[fi].truncate(n); - run_all(&d, &format!("{name} frame {fi} truncate {n}")); + run_all(&d, threads, &format!("{name} frame {fi} truncate {n}")); } } // header region of every frame: every byte to a few values, every bit @@ -127,12 +147,12 @@ fn sweep() { for v in [0u8, 0xff, 0x80, 0x01, 0x7f] { let mut d = seed.clone(); d[fi][p] = v; - run_all(&d, &format!("{name} frame {fi} byte {p}={v:#x}")); + run_all(&d, threads, &format!("{name} frame {fi} byte {p}={v:#x}")); } for b in 0..8 { let mut d = seed.clone(); d[fi][p] ^= 1 << b; - run_all(&d, &format!("{name} frame {fi} flip {p}.{b}")); + run_all(&d, threads, &format!("{name} frame {fi} flip {p}.{b}")); } } } @@ -151,7 +171,7 @@ fn sweep() { d[fi][p] = rng.next() as u8; } } - run_all(&d, &format!("{name} random {i}")); + run_all(&d, threads, &format!("{name} random {i}")); } } } diff --git a/crates/video/tests/av1_decode/main.rs b/crates/video/tests/av1_decode/main.rs index 50c3939f6..76a5adfcf 100644 --- a/crates/video/tests/av1_decode/main.rs +++ b/crates/video/tests/av1_decode/main.rs @@ -9,3 +9,4 @@ mod hostile_dims; mod oracle_inter; mod oracle_intra; mod perf; +mod wavefront; diff --git a/crates/video/tests/av1_decode/wavefront.rs b/crates/video/tests/av1_decode/wavefront.rs new file mode 100644 index 000000000..725b02544 --- /dev/null +++ b/crates/video/tests/av1_decode/wavefront.rs @@ -0,0 +1,93 @@ +//! Stress tests of the frame threading (parse / reconstruct split and the row wavefront across +//! frames): every thread count gives the single-thread pictures, run after run, and dropping +//! or flushing a busy decoder never hangs. A hang fails the test through the deadline. + +use crate::common; +use common::*; +use craft_video::av1::decode::{Decoder, Picture}; + +/// A GOP with many superblock rows (the wavefront has rows to overlap), hierarchical +/// references and motion reaching across rows. +fn spec() -> Spec { + Spec::new("wavefront_640x480", "testsrc2=s=640x480:r=25,rotate=a=t*0.5:c=black", 24, "yuv420p", &["-preset", "8", "-crf", "38"]) +} + +/// Seconds a test may take before it counts as a hang (override: FILMCRAFT_HOSTILE_TIMEOUT_SECS). +const DEADLINE_SECS: u64 = 300; + +fn decode(tus: &[Vec], threads: usize) -> Vec { + let mut dec = Decoder::with_threads(threads); + let mut pics = Vec::new(); + for (i, tu) in tus.iter().enumerate() { + pics.extend(dec.decode_pts(tu, i as i64).unwrap_or_else(|e| panic!("threads {threads}, unit {i}: {e}"))); + } + pics.extend(dec.flush_result().unwrap_or_else(|e| panic!("threads {threads}, flush: {e}"))); + pics +} + +#[test] +fn every_thread_count_gives_the_single_thread_pictures_repeatedly() { + let Some(ff) = ffmpeg() else { return }; + let tus = ivf_frames(&make(&ff, &spec())); + craft_media_testkit::with_deadline(DEADLINE_SECS, move || { + let want = decode(&tus, 1); + assert_eq!(want.len(), 24); + for threads in [2, 3, 4, 5, 8, 16] { + for round in 0..4 { + assert!(decode(&tus, threads) == want, "threads {threads}, round {round}: pictures differ"); + } + } + }); +} + +/// 128x128 superblocks in two tile columns (libaom), with motion field estimation: the +/// wavefront over the superblock rows of every tile column. +fn spec_sb128_tiles() -> Spec { + let enc = ["-cpu-used", "8", "-crf", "40", "-tile-columns", "1", "-lag-in-frames", "16", "-aom-params", "sb-size=128:enable-ref-frame-mvs=1"]; + Spec::new("wavefront_sb128_tiles", "testsrc2=s=640x400:r=25,rotate=a=t*0.5:c=black", 16, "yuv420p", &enc).encoder("libaom-av1") +} + +#[test] +fn superblocks_of_128_in_tile_columns_give_the_single_thread_pictures() { + let Some(ff) = ffmpeg_libaom() else { return }; + let tus = ivf_frames(&make(&ff, &spec_sb128_tiles())); + craft_media_testkit::with_deadline(DEADLINE_SECS, move || { + let want = decode(&tus, 1); + assert_eq!(want.len(), 16); + for threads in [2, 3, 4, 8, 16] { + for round in 0..2 { + assert!(decode(&tus, threads) == want, "threads {threads}, round {round}: pictures differ"); + } + } + }); +} + +#[test] +fn dropping_or_flushing_a_busy_decoder_never_hangs() { + let Some(ff) = ffmpeg() else { return }; + let tus = ivf_frames(&make(&ff, &spec())); + craft_media_testkit::with_deadline(DEADLINE_SECS, move || { + let want = decode(&tus, 1); + for threads in [2, 3, 8, 16] { + for cut in [1, 2, 7, 12, 23] { + // Dropped with frames in flight (no flush). + let mut dec = Decoder::with_threads(threads); + for (i, tu) in tus.iter().take(cut).enumerate() { + dec.decode_pts(tu, i as i64).expect("decode"); + } + drop(dec); + // Flushed mid-stream (as before a seek), then continued: the same pictures. + let mut dec = Decoder::with_threads(threads); + let mut pics = Vec::new(); + for (i, tu) in tus.iter().enumerate() { + if i == cut { + pics.extend(dec.flush_result().expect("flush")); + } + pics.extend(dec.decode_pts(tu, i as i64).expect("decode")); + } + pics.extend(dec.flush_result().expect("flush")); + assert!(pics == want, "threads {threads}, flush before unit {cut}: pictures differ"); + } + } + }); +} 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/wordmodel/src/chart.rs b/crates/wordmodel/src/chart.rs index b0cb6ad60..6c7296647 100644 --- a/crates/wordmodel/src/chart.rs +++ b/crates/wordmodel/src/chart.rs @@ -11,6 +11,8 @@ use serde::{Deserialize, Serialize}; +use crate::props::Rgb; + /// Most series in a chart. pub const MAX_SERIES: usize = 255; /// Most categories (points per series) in a chart. @@ -123,17 +125,111 @@ impl LegendPos { } } -/// One series: its name and a value per category (`None` = no value, a gap). +/// One series: its name, a value per category (`None` = no value, a gap) and how it is drawn +/// (Chart Design â–¸ Format Series): fill, line, markers, smooth lines, its own data labels and +/// number format. Unset fields follow the chart's defaults. #[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)] #[serde(rename_all = "camelCase")] pub struct ChartSeries { pub name: String, #[serde(default)] pub values: Vec>, + /// Fill colour; `None`: the chart's accent colour for the series. + #[serde(default)] + pub fill: Option, + /// "None" fill: the series draws no fill. + #[serde(default)] + pub no_fill: bool, + /// Line width in points; `None`: the chart's default. + #[serde(default)] + pub line_w: Option, + /// "None" line: the series draws no line. + #[serde(default)] + pub no_line: bool, + /// Markers on points; `None`: unset (scatter charts draw them; a saved scatter series then has no `c:marker`, so Word picks its own symbols). + #[serde(default)] + pub markers: Option, + /// Draw the line as a smooth curve. + #[serde(default)] + pub smooth: bool, + /// Data labels for this series; `None`: follow the chart's setting. + #[serde(default)] + pub labels: Option, + /// Number format of this series's labels and values; `None`: linked to the source. + #[serde(default)] + pub fmt: Option, +} + +/// One chart axis as Word formats it (Format Axis): title, scale bounds, tick-label number +/// format, gridlines, tick labels and the axis line. +#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)] +#[serde(rename_all = "camelCase")] +pub struct ChartAxis { + /// `None`: no axis title; text cut to [`MAX_TEXT`] characters. + #[serde(default)] + pub title: Option, + /// Fixed lower bound; `None`: automatic. + #[serde(default)] + pub min: Option, + /// Fixed upper bound; `None`: automatic. + #[serde(default)] + pub max: Option, + /// Tick-label number format (`#,##0`, `0.00`, `0%`…); `None`: linked to the source data. + #[serde(default)] + pub fmt: Option, + /// Major gridlines drawn along this axis. + #[serde(default)] + pub gridlines: bool, + /// Tick labels; `false` hides the axis the way Word's Axis Labels â–¸ None does. + #[serde(default = "the_true")] + pub labels: bool, + /// The axis line. + #[serde(default = "the_true")] + pub line: bool, +} + +fn the_true() -> bool { + true +} + +impl Default for ChartAxis { + fn default() -> Self { + ChartAxis { title: None, min: None, max: None, fmt: None, gridlines: false, labels: true, line: true } + } +} + +/// The value axis as a new chart has it: gridlines, labels, no axis line. +pub fn value_axis() -> ChartAxis { + ChartAxis { gridlines: true, line: false, ..ChartAxis::default() } +} + +impl ChartAxis { + /// Make the axis safe to store: texts cleaned, bounds finite, `min` under `max` (an + /// impossible pair drops both), the format code cut to [`MAX_FMT`] characters. + pub fn sanitize(&mut self) { + self.title = self.title.as_deref().map(clean_text).filter(|t| !t.is_empty()); + self.min = self.min.filter(|v| v.is_finite()); + self.max = self.max.filter(|v| v.is_finite()); + if matches!((self.min, self.max), (Some(lo), Some(hi)) if lo >= hi) { + self.min = None; + self.max = None; + } + self.fmt = self.fmt.as_deref().map(clean_fmt).filter(|f| !f.is_empty()); + } +} + +/// Longest axis number format kept, characters. +pub const MAX_FMT: usize = 128; + +/// A number format code: control characters as spaces, ends trimmed, at most [`MAX_FMT`] +/// characters. +pub fn clean_fmt(s: &str) -> String { + let s: String = s.chars().take(MAX_FMT * 4).map(|c| if c.is_control() { ' ' } else { c }).collect(); + s.trim().chars().take(MAX_FMT).collect::().trim_end().to_string() } /// A chart the app can edit. -#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize)] +#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)] #[serde(rename_all = "camelCase")] pub struct ChartSpec { #[serde(rename = "type")] @@ -151,6 +247,29 @@ pub struct ChartSpec { /// Show each point's value next to it. #[serde(default)] pub data_labels: bool, + /// The category axis (scatter charts: the x axis). + #[serde(default)] + pub cat_axis: ChartAxis, + /// The value axis (scatter charts: the y axis); gridlines on, no axis line, as a new chart has them. + #[serde(default = "value_axis")] + pub val_axis: ChartAxis, +} + +impl Default for ChartSpec { + /// An empty chart whose value axis is the one a new chart (and a missing JSON field) has: + /// gridlines on, no axis line. + fn default() -> Self { + ChartSpec { + kind: ChartType::default(), + categories: Vec::new(), + series: Vec::new(), + title: None, + legend: None, + data_labels: false, + cat_axis: ChartAxis::default(), + val_axis: value_axis(), + } + } } impl ChartSpec { @@ -168,10 +287,23 @@ impl ChartSpec { ChartSpec { kind, categories: categories.into_iter().map(str::to_string).collect(), - series: series.into_iter().map(|(n, v)| ChartSeries { name: n.to_string(), values: v.into_iter().map(Some).collect() }).collect(), + series: series + .into_iter() + .map(|(n, v)| ChartSeries { + name: n.to_string(), + values: v.into_iter().map(Some).collect(), + // Scatter series are drawn as markers without a line. + no_line: kind == ChartType::Scatter, + ..Default::default() + }) + .collect(), title: None, legend: Some(LegendPos::Bottom), data_labels: false, + cat_axis: ChartAxis::default(), + // Pie and doughnut charts have no axes: their axes stay plain, so the chart reads back + // as the same model. + val_axis: if kind.is_round() { ChartAxis::default() } else { value_axis() }, } } @@ -203,8 +335,12 @@ impl ChartSpec { *v = v.filter(|x| x.is_finite()); } s.values.resize(n, None); + s.fmt = s.fmt.as_deref().map(clean_fmt).filter(|f| !f.is_empty()); + s.line_w = s.line_w.filter(|w| w.is_finite()).map(|w| w.clamp(0.0, 20.0)); } self.title = self.title.as_deref().map(clean_text).filter(|t| !t.is_empty()); + self.cat_axis.sanitize(); + self.val_axis.sanitize(); } /// The chart after `sanitize`. @@ -229,7 +365,11 @@ mod tests { let mut c = ChartSpec { categories: vec!["a".into(); MAX_POINTS + 10], series: (0..MAX_SERIES + 5) - .map(|i| ChartSeries { name: format!("{}\n{}", "x".repeat(400), i), values: vec![Some(f64::NAN), Some(f64::INFINITY), Some(1.0)] }) + .map(|i| ChartSeries { + name: format!("{}\n{}", "x".repeat(400), i), + values: vec![Some(f64::NAN), Some(f64::INFINITY), Some(1.0)], + ..Default::default() + }) .collect(), title: Some(" ".into()), ..Default::default() @@ -250,7 +390,7 @@ mod tests { fn short_categories_are_padded_to_the_longest_series() { let c = ChartSpec { categories: vec!["a".into()], - series: vec![ChartSeries { name: "s".into(), values: vec![Some(1.0), Some(2.0)] }], + series: vec![ChartSeries { name: "s".into(), values: vec![Some(1.0), Some(2.0)], ..Default::default() }], ..Default::default() } .sanitized(); @@ -258,4 +398,53 @@ mod tests { assert_eq!(ChartType::from_id("BARSTACKED"), Some(ChartType::BarStacked)); assert_eq!(LegendPos::from_id("top"), Some(LegendPos::Top)); } + + #[test] + fn default_and_missing_json_axis_fields_agree() { + let from_json: ChartSpec = serde_json::from_str(r#"{"type":"column"}"#).unwrap(); + let d = ChartSpec::default(); + assert_eq!((d.cat_axis.clone(), d.val_axis.clone()), (from_json.cat_axis, from_json.val_axis)); + assert!(d.val_axis.gridlines && !d.val_axis.line); + } + + #[test] + fn axis_bounds_and_formats_are_cleaned() { + let mut a = ChartAxis { + title: Some(" Sales\n".into()), + min: Some(f64::NAN), + max: Some(50.0), + fmt: Some("x".repeat(MAX_FMT + 40)), + gridlines: true, + labels: false, + line: true, + }; + a.sanitize(); + assert_eq!(a.title.as_deref(), Some("Sales")); + assert_eq!(a.min, None); + assert_eq!(a.max, Some(50.0)); + assert_eq!(a.fmt.as_deref().map(|f| f.chars().count()), Some(MAX_FMT)); + + // min not under max drops the pair; a format of only spaces is no format. + let mut b = ChartAxis { min: Some(50.0), max: Some(50.0), fmt: Some(" ".into()), ..Default::default() }; + b.sanitize(); + assert_eq!((b.min, b.max), (None, None)); + assert_eq!(b.fmt, None); + + let mut spec = ChartSpec { + series: vec![ChartSeries { name: "s".into(), values: vec![Some(1.0)], ..Default::default() }], + categories: vec!["a".into()], + ..Default::default() + }; + spec.series[0].line_w = Some(f32::NAN); + spec.series[0].fmt = Some("0.00".into()); + spec.sanitize(); + assert_eq!(spec.series[0].line_w, None); + assert_eq!(spec.series[0].fmt.as_deref(), Some("0.00")); + let mut spec = ChartSpec { + series: vec![ChartSeries { name: "s".into(), values: vec![], line_w: Some(100.0), ..Default::default() }], + ..Default::default() + }; + spec.sanitize(); + assert_eq!(spec.series[0].line_w, Some(20.0), "a wild width clamps to 20 pt"); + } } diff --git a/crates/wordmodel/src/graphic.rs b/crates/wordmodel/src/graphic.rs index c77c0e5e7..cba7e766a 100644 --- a/crates/wordmodel/src/graphic.rs +++ b/crates/wordmodel/src/graphic.rs @@ -115,7 +115,9 @@ pub enum GraphicItem { #[serde(default)] stroke_width: f32, }, - /// One line of text in `rect`, aligned horizontally by `align`, centred vertically. + /// One line of text in `rect`, aligned horizontally by `align`, centred vertically; `angle` + /// degrees clockwise turns the text about `rect`'s centre as it is drawn (chart axis titles + /// turn the vertical axis's title -90°, reading upwards, as Word does). Text { rect: [f32; 4], text: String, @@ -127,6 +129,8 @@ pub enum GraphicItem { align: TextAlign, #[serde(default)] font: Option, + #[serde(default)] + angle: f32, }, /// A picture (`Document::media` key) stretched into `rect`. Image { rect: [f32; 4], media: String }, 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()); +}