From 9399c25341289c1d523e64c92a1b95735c81933f Mon Sep 17 00:00:00 2001 From: FredApps <153173854+FredApps@users.noreply.github.com> Date: Sun, 15 Nov 2026 16:10:22 +0100 Subject: [PATCH 01/20] image-codecs: decode AVIF pixels with craft-video's pure-Rust AV1 decoder New optional `avif-decode` feature (native targets) reads still AVIF files: the HEIF/ISOBMFF item graph, grids, clap/irot/imir, alpha auxiliary images (including premultiplied), 8/10/12-bit 4:0:0, 4:2:0, 4:2:2 and 4:4:4, nclx/ICC colour, Exif and XMP, and PQ/HLG to linear F32 through cicp::linearize as the HEIF path does. The container walk follows LightCraft #3017 by @emmertex. Every size is checked against the caller's Limits before the AV1 decoder runs, the decoder's memory budget is capped by max_alloc, and the decode is wrapped in catch_unwind so a decoder panic becomes an error. Fixtures are tiny ffmpeg/libaom files; the 4:2:0 planes match libdav1d bit for bit. Adds a decode_avif fuzz target. Part of storytold/craft-libs#778. --- crates/image-codecs/Cargo.toml | 4 + crates/image-codecs/README.md | 27 +- crates/image-codecs/fuzz/Cargo.toml | 9 +- .../fuzz/fuzz_targets/decode_avif.rs | 14 + crates/image-codecs/src/cicp.rs | 2 +- crates/image-codecs/src/codecs.rs | 2 + crates/image-codecs/src/codecs/avif_decode.rs | 1239 +++++++++++++++++ crates/image-codecs/src/format.rs | 13 +- crates/image-codecs/src/lib.rs | 24 +- .../tests/fixtures/avif/README.md | 12 + .../tests/fixtures/avif/quad-420-bt709.avif | Bin 0 -> 369 bytes .../fixtures/avif/quad-420-bt709.dav1d.yuv | 1 + .../fixtures/avif/quad-444-gbr-lossless.avif | Bin 0 -> 345 bytes .../tests/fixtures/avif/quad-source.rgb | 1 + crates/image-codecs/tests/it/detect_caps.rs | 11 +- 15 files changed, 1335 insertions(+), 24 deletions(-) create mode 100644 crates/image-codecs/fuzz/fuzz_targets/decode_avif.rs create mode 100644 crates/image-codecs/src/codecs/avif_decode.rs create mode 100644 crates/image-codecs/tests/fixtures/avif/README.md create mode 100644 crates/image-codecs/tests/fixtures/avif/quad-420-bt709.avif create mode 100644 crates/image-codecs/tests/fixtures/avif/quad-420-bt709.dav1d.yuv create mode 100644 crates/image-codecs/tests/fixtures/avif/quad-444-gbr-lossless.avif create mode 100644 crates/image-codecs/tests/fixtures/avif/quad-source.rgb diff --git a/crates/image-codecs/Cargo.toml b/crates/image-codecs/Cargo.toml index df71d9860..1b422f912 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" @@ -41,6 +44,7 @@ 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 6e9e11617..1728d8308 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: @@ -56,7 +57,7 @@ craft-image-codecs = { git = "https://github.com/storytold/craft-libs", rev = "< | Cineon | yes | yes | U16 (10-bit code values read and written exactly through U16) | Gray, RGB | no | no | no | no | no | no | no | built in | | DICOM | yes | **no** | U8, U16 | Gray, RGB | no | no | no | no | no | no | n/a | built in | | Amiga IFF (ILBM/PBM) | yes | yes (24-bit ILBM) | U8 | RGB, RGBA | yes (1-bit mask plane on write) | no | no | no | no | no | no | built in | -| 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 @@ -177,10 +178,20 @@ 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/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 0288ae22b..71db85f18 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..302433661 --- /dev/null +++ b/crates/image-codecs/src/codecs/avif_decode.rs @@ -0,0 +1,1239 @@ +//! 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 is capped by `max_alloc` too; 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)] +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" => { + let mut q = || -> Result { + let n = r.u32()? as i32; + let d = r.u32()? as i32; + if d == 0 { + return Err(bad("clap with a zero denominator")); + } + Ok(f64::from(n) / f64::from(d)) + }; + let (w, h, dx, dy) = (q()?, q()?, q()?, q()?); + 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 }) + } + fn transforms(&self, id: u32) -> Vec { + self.props_of(id).filter_map(|p| if let Prop::T(t) = p { Some(*t) } else { None }).collect() + } + 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)); + 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 | 10 => 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 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, &c.transforms(id)); + } + 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())); + } + + #[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, 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/format.rs b/crates/image-codecs/src/format.rs index ecaf945cd..fbdb412b1 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, @@ -89,8 +90,10 @@ 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."), ( @@ -298,7 +301,9 @@ 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 6516b9ff2..a4480b2b8 100644 --- a/crates/image-codecs/src/lib.rs +++ b/crates/image-codecs/src/lib.rs @@ -106,7 +106,11 @@ 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) } @@ -163,10 +167,11 @@ 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), @@ -195,10 +200,11 @@ pub fn decode_png_rows>( } /// The upright pixel size [`decode`] would return, read from the file's header without decoding -/// pixels, for WebP, OpenEXR (flat) and the GIF/BMP/TGA/ICO/QOI/HDR family. It runs the same +/// pixels, for WebP, OpenEXR (flat), AVIF (with `avif-decode`) and the GIF/BMP/TGA/ICO/QOI/HDR family. It runs the same /// header reads and size checks as the decoder (default [`Limits`]) and swaps the size for an -/// EXIF orientation of 5–8, so wherever [`decode`] succeeds the two agree. Other formats (and -/// deep EXR or AVIF) return [`CodecError::Unsupported`]: decode those to learn their size. +/// EXIF orientation of 5–8 (AVIF: its container's `clap` and `irot`), so wherever [`decode`] +/// succeeds the two agree. Other formats (and deep EXR) return [`CodecError::Unsupported`]: decode +/// those to learn their size. pub fn probe_size(bytes: &[u8]) -> Result<(u32, u32), CodecError> { let format = detect(bytes).ok_or(CodecError::UnknownFormat)?; if !caps(format).read { @@ -208,7 +214,11 @@ pub fn probe_size(bytes: &[u8]) -> Result<(u32, u32), CodecError> { let (w, h, exif) = match format { Format::WebP => webp::probe(bytes, l)?, Format::OpenExr => exr::probe(bytes, l).map(|(w, h)| (w, h, None))?, - Format::Gif | Format::Bmp | Format::Tga | Format::Ico | Format::Qoi | Format::Hdr | Format::Avif => via_image::probe(format, bytes, l)?, + #[cfg(all(feature = "avif-decode", not(target_arch = "wasm32")))] + Format::Avif => codecs::avif_decode::probe(bytes, l).map(|(w, h)| (w, h, None))?, + #[cfg(not(all(feature = "avif-decode", not(target_arch = "wasm32"))))] + Format::Avif => via_image::probe(format, bytes, l)?, + Format::Gif | Format::Bmp | Format::Tga | Format::Ico | Format::Qoi | Format::Hdr => via_image::probe(format, bytes, l)?, _ => return Err(CodecError::unsupported(format, "no header-only size probe")), }; let o = exif.as_deref().map_or(1, exif_orientation); diff --git a/crates/image-codecs/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 0000000000000000000000000000000000000000..419196cb178743782cf147ea50d111e2b8642d42 GIT binary patch literal 369 zcmZQzU{FXasVqn=%S>Yc0uY^>nP!-qnV9D5Xy^nK`k9+rk_eIm0{x7XoFWL5fuSHX zxdg@r(E&hyX;G?2VqQv4Dp(a?L1sxVSP3JLl*`P?PX_T^7#J9V5JWOCDgx;MAlAst z%mX_CEF%Tvq%r}8LA=B=14HNhoP4mES(ybz1we7{%!1^6AT5$vT#yP<2m~Or1c6wf zAhRM9B*4tU!2zUr63Yyo8yR>&Lc+=UIYoKNITZ{{KnOIQfk8a8AU6@j00N-HSQwd_ zS%A!l+?2!;E=~zih2|6v4i*6;XO{y&vw`U4k2c-I-AWygkL2ynp7_}AP4*4enEOQm q)z5b1%rN?KQ|H$?;i_Mo{+`I#w8y3Pf9fWWdacYc0uY^>nP!-qnV9D5Xy^zO`k9+rk_eIm0{x7XoFWL5fuSHX zxdg@r(E&hyX;G?2VqQv4Dp(a?L1sxVSP3JLl*`P?PX_T^7#J9V5JWOCDgtQ>AlAst z%mX_CEF%Tvq%r}8LA=B=14HNhoP4mES(ybz1we7{%!1^6AT5$vT#yP<2m~Or1c6wf zAhRM9B*4tU!2zUr63Yyo8x?>o1_t5e{G6h^jRAvYzlgp0#MRG~S8BZJXM2k7n(sn>EClKWl Date: Sun, 15 Nov 2026 18:08:34 +0100 Subject: [PATCH 02/20] craft-raw: one lookup per Huffman symbol and a fixed predictor per run in LJ92 and NEF/RW2 decoding Co-Authored-By: Claude Sonnet 5.5 --- crates/raw/src/huffman.rs | 107 ++++++++++++++++++--- crates/raw/src/ljpeg.rs | 169 ++++++++++++++++++++++++++-------- crates/raw/src/ljpeg/tests.rs | 88 ++++++++++++++++++ 3 files changed, 311 insertions(+), 53 deletions(-) diff --git a/crates/raw/src/huffman.rs b/crates/raw/src/huffman.rs index cce27237b..85b242be1 100644 --- a/crates/raw/src/huffman.rs +++ b/crates/raw/src/huffman.rs @@ -4,16 +4,19 @@ use crate::{RawError, Result}; -/// A canonical-by-table Huffman decoder with a 12-bit lookup. +/// A canonical-by-table Huffman decoder with a 12-bit lookup that also holds the difference when code and +/// additional bits fit in 12 bits. pub(crate) struct Huffman { - /// Indexed by the next 12 bits: (symbol, length); length 0 = invalid. - lut: Vec<(u8, u8)>, + /// Indexed by the next 12 bits. Per entry: bits 0-7 the symbol, bits 8-11 the code length (0 = invalid), bits + /// 12-15 the length of code plus additional bits when both fit in 12 bits (0 = they do not), bits 16-31 the + /// signed difference in that case. + lut: Box<[u32; 4096]>, } impl Huffman { /// From `(left-aligned 12-bit code, length)` per symbol. pub fn new(codes: &[(u16, u8)]) -> Result { - let mut lut = vec![(0u8, 0u8); 4096]; + let mut lut = Box::new([0u32; 4096]); for (sym, &(code, len)) in codes.iter().enumerate() { if len == 0 { continue; @@ -21,10 +24,19 @@ impl Huffman { if len > 12 || sym > 16 { return Err(RawError::Unsupported(format!("PEF Huffman code of length {len}"))); } + let (len, s) = (len as u32, sym as u32); let first = (code >> (12 - len) << (12 - len)) as usize; - for e in &mut lut[first..first + (1 << (12 - len))] { - if e.1 == 0 { - *e = (sym as u8, len); + for (i, e) in lut[first..first + (1 << (12 - len))].iter_mut().enumerate() { + if (*e >> 8) & 15 == 0 { + let mut entry = (len << 8) | s; + if s == 0 { + entry |= len << 12; + } else if len + s <= 12 { + let v = (((first + i) >> (12 - len - s)) & ((1 << s) - 1)) as i32; + let d = if v < 1 << (s - 1) { v - (1 << s) + 1 } else { v }; + entry |= ((len + s) << 12) | ((d as i16 as u16 as u32) << 16); + } + *e = entry; } } } @@ -85,21 +97,90 @@ impl<'a> Bits<'a> { } } -/// Decode one difference: Huffman category then T.81 additional bits. +/// Decode one difference: Huffman category then T.81 additional bits, in one lookup when both fit in 12 bits. #[inline] pub(crate) fn diff(bits: &mut Bits, h: &Huffman) -> Option { - let (sym, len) = h.lut[bits.peek(12) as usize]; + let e = h.lut[(bits.peek(12) & 0xfff) as usize]; + let len = (e >> 8) & 15; if len == 0 { return None; } - bits.skip(len as u32); - let s = sym as u32; - if s == 0 { - return Some(0); + let total = (e >> 12) & 15; + if total > 0 { + bits.skip(total); + return Some((e as i32) >> 16); } + bits.skip(len); + let s = e & 0xff; if s > 16 { return None; } let v = bits.get(s) as i32; Some(if v < 1 << (s - 1) { v - (1 << s) + 1 } else { v }) } + +#[cfg(test)] +mod tests { + use super::*; + + /// Canonical left-aligned 12-bit codes for symbols 0..lens.len() (shorter codes first, ties by symbol). + fn canonical(lens: &[u8]) -> Vec<(u16, u8)> { + let mut order: Vec = (0..lens.len()).collect(); + order.sort_by_key(|&s| (lens[s], s)); + let mut codes = vec![(0u16, 0u8); lens.len()]; + let (mut code, mut prev) = (0u32, 0u8); + for s in order { + code <<= lens[s] - prev; + prev = lens[s]; + codes[s] = ((code << (12 - lens[s])) as u16, lens[s]); + code += 1; + } + codes + } + + fn bit(src: &[u8], i: usize) -> u32 { + src.get(i / 8).map_or(0, |b| ((b >> (7 - i % 8)) & 1) as u32) + } + + /// Reference: match the code bit by bit, then read the category's additional bits (T.81 F.1.2.1). + fn reference(src: &[u8], pos: &mut usize, codes: &[(u16, u8)]) -> Option { + let window: u32 = (0..12).fold(0, |a, k| (a << 1) | bit(src, *pos + k)); + let (sym, len) = codes.iter().enumerate().find_map(|(s, &(c, l))| { + (l > 0 && (c as u32 >> (12 - l)) == window >> (12 - l)).then_some((s as u32, l as usize)) + })?; + *pos += len; + if sym == 0 { + return Some(0); + } + let v = (0..sym as usize).fold(0u32, |a, k| (a << 1) | bit(src, *pos + k)) as i32; + *pos += sym as usize; + Some(if v < 1 << (sym - 1) { v - (1 << sym) + 1 } else { v }) + } + + #[test] + fn diff_matches_bitwise_reference_for_every_category() { + // categories 0..=16, code lengths up to 12 so that short codes combine with their bits and long ones do not + let lens = [2u8, 2, 3, 3, 4, 4, 5, 6, 7, 8, 9, 10, 11, 12, 12, 12, 12]; + let codes = canonical(&lens); + let h = Huffman::new(&codes).unwrap(); + let mut s = 0x1234_5678_9abc_def0u64; + for _ in 0..40 { + let src: Vec = (0..400) + .map(|_| { + s = s.wrapping_mul(6364136223846793005).wrapping_add(1442695040888963407); + (s >> 40) as u8 + }) + .collect(); + let mut bits = Bits::new(&src); + let mut pos = 0; + while pos < src.len() * 8 - 64 { + let want = reference(&src, &mut pos, &codes); + assert_eq!(diff(&mut bits, &h), want); + if want.is_none() { + break; + } + assert_eq!(bits.consumed_bits(), pos); + } + } + } +} diff --git a/crates/raw/src/ljpeg.rs b/crates/raw/src/ljpeg.rs index 56b07a533..ab27ed57f 100644 --- a/crates/raw/src/ljpeg.rs +++ b/crates/raw/src/ljpeg.rs @@ -32,15 +32,17 @@ fn err(s: &str) -> RawError { /// Huffman decoding table (T.81 Annex C "BITS"/"HUFFVAL"). #[derive(Clone, Debug)] pub struct Huffman { - /// For peeks of `LOOKUP` bits: (code length, symbol); length 0 = slow path. - fast: Vec<(u8, u8)>, + /// Indexed by the next `LOOKUP` bits; length 0 = slow path. Per entry: bits 0-7 the symbol, bits 8-11 the code + /// length, bits 12-15 the length of code plus extra bits when both fit in the lookup (0 = they do not), and + /// bits 16-31 the signed difference in that case. + fast: Vec, maxcode: [i32; 18], valptr: [i32; 17], mincode: [i32; 17], values: Vec, } -const LOOKUP: u32 = 10; +const LOOKUP: u32 = 12; impl Huffman { /// Build from the 16 code-length counts and the symbol list. @@ -52,7 +54,7 @@ impl Huffman { let mut maxcode = [-1i32; 18]; let mut valptr = [0i32; 17]; let mut mincode = [0i32; 17]; - let mut fast = vec![(0u8, 0u8); 1 << LOOKUP]; + let mut fast = vec![0u32; 1 << LOOKUP]; let mut code = 0i32; let mut k = 0usize; for len in 1..=16usize { @@ -63,9 +65,10 @@ impl Huffman { if len as u32 <= LOOKUP { let shift = LOOKUP - len as u32; let base = (code as usize) << shift; + let ssss = values[k] as u32; for i in 0..(1usize << shift) { if let Some(slot) = fast.get_mut(base + i) { - *slot = (len as u8, values[k]); + *slot = fast_entry(len as u32, ssss, base + i); } } } @@ -84,12 +87,41 @@ impl Huffman { #[inline] pub fn decode(&self, br: &mut B) -> Result { - let peek = br.peek(LOOKUP); - let (len, sym) = self.fast[peek as usize]; + let e = self.fast[br.peek(LOOKUP) as usize]; + let len = (e >> 8) & 15; if len > 0 { - br.consume(len as u32); - return Ok(sym); + br.consume(len); + return Ok(e as u8); } + self.decode_slow(br) + } + + /// One symbol and its T.81 difference in a single lookup when code and extra bits fit in `LOOKUP` bits; the same + /// as [`Huffman::decode`] followed by [`diff_value`]. `strict`: a category above 16 is an error (subsampled + /// scans) instead of reading 16 bits. + #[inline] + pub fn decode_diff(&self, br: &mut B, strict: bool) -> Result { + let e = self.fast[br.peek(LOOKUP) as usize]; + let len = (e >> 8) & 15; + let total = (e >> 12) & 15; + if total > 0 { + br.consume(total); + return Ok((e as i32) >> 16); + } + let ssss = if len > 0 { + br.consume(len); + e as u8 + } else { + self.decode_slow(br)? + }; + if strict && ssss > 16 { + return Err(err("invalid difference category")); + } + Ok(diff_value(br, ssss)) + } + + #[cold] + fn decode_slow(&self, br: &mut B) -> Result { let mut code = br.peek(LOOKUP + 1) as i32; let mut len = LOOKUP as usize + 1; while len <= 16 && code > self.maxcode[len] { @@ -105,6 +137,21 @@ impl Huffman { } } +/// Lookup entry for a code of `len` bits (`len <= LOOKUP`) naming category `ssss`, at window position `index` +/// (the code is the top `len` bits of the `LOOKUP`-bit window). See [`Huffman::fast`]. +fn fast_entry(len: u32, ssss: u32, index: usize) -> u32 { + let base = (len << 8) | ssss; + if ssss == 0 { + return base | (len << 12); + } + if ssss > 11 || len + ssss > LOOKUP { + return base; + } + let v = ((index >> (LOOKUP - len - ssss)) & ((1 << ssss) - 1)) as i32; + let d = if v < (1 << (ssss - 1)) { v - (1 << ssss) + 1 } else { v }; + base | ((len + ssss) << 12) | ((d as i16 as u16 as u32) << 16) +} + /// A source of entropy-coded bits, first bit most significant. pub trait Bits { /// The next `n` (1..=32) bits without consuming them; zeros past the end. @@ -614,11 +661,7 @@ pub fn decode_subsampled(d: &[u8], max_samples: usize, prediction: Prediction) - Prediction::Sequential if mx == 0 && dx == 0 && dy == 0 => row_first[c], Prediction::Sequential => last[c], }; - let ssss = tables[c].decode(&mut br)?; - if ssss > 16 { - return Err(err("invalid difference category")); - } - let value = ((pred + diff_value(&mut br, ssss)) & 0xffff) as u16; + let value = ((pred + tables[c].decode_diff(&mut br, true)?) & 0xffff) as u16; plane[i] = value; last[c] = i32::from(value); if mx == 0 && dx == 0 && dy == 0 { @@ -642,6 +685,44 @@ pub fn decode_subsampled(d: &[u8], max_samples: usize, prediction: Prediction) - Ok(FrameSubsampled { width: h.width, height: h.height, vertical_subsampling: vertical, planes }) } +/// T.81 Table H.1 predictor `pred` (1..=7; 0 = none) from the left, above and above-left neighbours. +#[inline(always)] +fn predict(pred: u8, ra: i32, rb: i32, rc: i32) -> i32 { + match pred { + 0 => 0, + 1 => ra, + 2 => rb, + 3 => rc, + 4 => ra + rb - rc, + 5 => ra + ((rb - rc) >> 1), + 6 => rb + ((ra - rc) >> 1), + _ => (ra + rb) >> 1, + } +} + +/// Decode the pixels `x0..x1` (`x0 >= 1`) of one row with the fixed predictor `P`. `prev` is empty on the first +/// line, which only predicts from the left (`P = 1`). +#[inline(always)] +fn decode_run( + tables: &[&Huffman], + br: &mut BitReader<'_>, + cur: &mut [u16], + prev: &[u16], + nc: usize, + x0: usize, + x1: usize, +) -> Result<(), RawError> { + for x in x0..x1 { + for (c, table) in tables.iter().enumerate() { + let i = x * nc + c; + let ra = cur[i - nc] as i32; + let p = if P <= 1 { predict(P, ra, 0, 0) } else { predict(P, ra, prev[i] as i32, prev[i - nc] as i32) }; + cur[i] = ((p + table.decode_diff(br, false)?) & 0xffff) as u16; + } + } + Ok(()) +} + /// Decode a lossless JPEG stream. `max_samples` bounds the allocation (hostile headers). pub fn decode(d: &[u8], max_samples: usize) -> Result { let h = parse_header(d)?; @@ -680,19 +761,23 @@ pub fn decode(d: &[u8], max_samples: usize) -> Result { let (prev_rows, cur_rows) = out.split_at_mut(y * row_len); let cur = &mut cur_rows[..row_len]; let prev = if y > 0 { &prev_rows[(y - 1) * row_len..] } else { &[][..] }; - for x in 0..w { - if h.restart > 0 { - if restart_left == 0 { - if !br.restart() { - return Err(err("missing restart marker")); - } - restart_left = h.restart; - first_line = true; - reset_pending = true; + let mut x = 0; + while x < w { + if h.restart > 0 && restart_left == 0 { + if !br.restart() { + return Err(err("missing restart marker")); } - restart_left -= 1; + restart_left = h.restart; + first_line = true; + reset_pending = true; } - #[allow(clippy::needless_range_loop)] // `c` indexes the tables and the sample position alike + // pixels up to the next restart marker or the end of the row + let mut end = w; + if h.restart > 0 { + end = w.min(x + restart_left); + restart_left -= end - x; + } + // the first pixel of the run has the special predictions (start of the interval, first column) for c in 0..nc { let i = x * nc + c; let p = if reset_pending { @@ -702,24 +787,28 @@ pub fn decode(d: &[u8], max_samples: usize) -> Result { } else if x == 0 || pred == 0 { if pred == 0 { 0 } else { prev[i] as i32 } } else { - let ra = cur[i - nc] as i32; - let rb = prev[i] as i32; - let rc = prev[i - nc] as i32; - match pred { - 1 => ra, - 2 => rb, - 3 => rc, - 4 => ra + rb - rc, - 5 => ra + ((rb - rc) >> 1), - 6 => rb + ((ra - rc) >> 1), - _ => (ra + rb) >> 1, - } + predict(pred, cur[i - nc] as i32, prev[i] as i32, prev[i - nc] as i32) }; - let ssss = tables[c].decode(&mut br)?; - let diff = diff_value(&mut br, ssss); - cur[i] = ((p + diff) & mask) as u16; + cur[i] = ((p + tables[c].decode_diff(&mut br, false)?) & mask) as u16; } reset_pending = false; + // the rest of the run: the predictor is fixed, so it is chosen once here, not per sample + if first_line { + // the first line of an interval predicts from the left neighbour + decode_run::<1>(&tables, &mut br, cur, prev, nc, x + 1, end)?; + } else { + match pred { + 0 => decode_run::<0>(&tables, &mut br, cur, prev, nc, x + 1, end)?, + 1 => decode_run::<1>(&tables, &mut br, cur, prev, nc, x + 1, end)?, + 2 => decode_run::<2>(&tables, &mut br, cur, prev, nc, x + 1, end)?, + 3 => decode_run::<3>(&tables, &mut br, cur, prev, nc, x + 1, end)?, + 4 => decode_run::<4>(&tables, &mut br, cur, prev, nc, x + 1, end)?, + 5 => decode_run::<5>(&tables, &mut br, cur, prev, nc, x + 1, end)?, + 6 => decode_run::<6>(&tables, &mut br, cur, prev, nc, x + 1, end)?, + _ => decode_run::<7>(&tables, &mut br, cur, prev, nc, x + 1, end)?, + } + } + x = end; } first_line = false; if br.overrun > 64 * 8 { diff --git a/crates/raw/src/ljpeg/tests.rs b/crates/raw/src/ljpeg/tests.rs index c5c6ccf25..17791ea8b 100644 --- a/crates/raw/src/ljpeg/tests.rs +++ b/crates/raw/src/ljpeg/tests.rs @@ -125,6 +125,94 @@ fn roundtrip_extremes_and_restarts() { } } +/// Re-label an encoded stream as `precision` bits with point transform `pt`: the encoder codes the samples at +/// `precision - pt` bits, which is what a transformed scan holds. +fn with_point_transform(mut enc: Vec, precision: u8, pt: u8) -> Vec { + enc[6] = precision; // SOF3 precision follows SOI, marker and length + let sos = enc.windows(2).position(|w| w == [0xff, 0xda]).unwrap(); + let seg_len = u16::from_be_bytes([enc[sos + 2], enc[sos + 3]]) as usize; + enc[sos + 2 + seg_len - 1] = pt; + enc +} + +#[test] +fn roundtrip_point_transform_all_predictors() { + for pred in 1..=7u8 { + for comps in [1usize, 2, 3] { + for (bits, pt) in [(12u8, 1u8), (14, 2), (16, 4), (8, 3)] { + let (w, h) = (11, 6); + let data = noise(w * h * comps, (bits - pt) as u32, pred as u64 * 17 + comps as u64); + let enc = with_point_transform(encode(&data, w, h, comps, bits - pt, pred, 0), bits, pt); + let f = decode(&enc, usize::MAX).unwrap(); + assert_eq!((f.precision, f.point_transform, f.predictor), (bits, pt, pred)); + let want: Vec = data.iter().map(|&v| v << pt).collect(); + assert_eq!(f.data, want, "pred {pred} comps {comps} bits {bits} pt {pt}"); + } + } + } + // with restart intervals that end mid-row + let data = noise(20 * 9 * 2, 10, 9); + for rst in [3usize, 7, 20] { + let enc = with_point_transform(encode(&data, 20, 9, 2, 10, 4, rst), 12, 2); + let want: Vec = data.iter().map(|&v| v << 2).collect(); + assert_eq!(decode(&enc, usize::MAX).unwrap().data, want, "restart {rst}"); + } +} + +#[test] +fn roundtrip_skewed_categories_with_long_codes() { + // category k with probability 2^-(k+1): the table gets codes longer than the lookup, rare 16-bit differences + let mut s = 0x9e37_79b9_7f4a_7c15u64; + let mut last = 0u32; + let data: Vec = (0..256 * 256) + .map(|_| { + s = s.wrapping_mul(6364136223846793005).wrapping_add(1442695040888963407); + let r = (s >> 32) as u32; + let cat = r.trailing_zeros().min(16); + let mag = (((s >> 8) as u32) & ((1u32 << cat) - 1)) | (1 << cat) >> 1; + let d = if s & 1 == 0 { mag } else { 0u32.wrapping_sub(mag) }; + last = last.wrapping_add(d) & 0xffff; + last as u16 + }) + .collect(); + for pred in [1u8, 4, 7] { + let enc = encode(&data, 256, 256, 1, 16, pred, 0); + assert_eq!(decode(&enc, usize::MAX).unwrap().data, data, "pred {pred}"); + } +} + +/// `decode_diff` is `decode` followed by `diff_value`, for tables with codes inside and beyond the lookup. +#[test] +fn decode_diff_matches_decode_and_diff_value() { + for lens in [[2u8, 2, 3, 3, 4, 4, 5, 6, 7, 8, 9, 10, 11, 12, 12, 12, 12], [2, 2, 3, 3, 4, 4, 5, 6, 7, 8, 9, 10, 13, 14, 15, 16, 16]] { + let mut counts = [0u8; 16]; + for &l in &lens { + counts[l as usize - 1] += 1; + } + let values: Vec = (0..17).collect(); // lengths are already ascending + let t = Huffman::new(&counts, &values).unwrap(); + let mut s = 77u64; + let src: Vec = (0..6000) + .map(|_| { + s = s.wrapping_mul(6364136223846793005).wrapping_add(1442695040888963407); + ((s >> 40) as u8).min(0xfe) + }) + .collect(); + let (mut a, mut b) = (BitReader::new(&src), BitReader::new(&src)); + for n in 0..8000 { + let want = t.decode(&mut a).map(|ssss| diff_value(&mut a, ssss)); + let got = t.decode_diff(&mut b, false); + assert_eq!(got.ok(), want.ok(), "symbol {n}"); + assert_eq!(a.peek(32), b.peek(32), "symbol {n}"); + if want.is_err() { + // step past the invalid code so the walk goes on + a.consume(1); + b.consume(1); + } + } + } +} + #[test] fn constant_image_single_symbol_table() { let data = vec![1234u16; 16 * 4]; From df89f6b6b7534c717debfbcb19b287a14aba16fb Mon Sep 17 00:00:00 2001 From: vangel76 Date: Sun, 15 Nov 2026 18:13:19 +0100 Subject: [PATCH 03/20] craft-raw: test tables where code plus extra bits fill the lookup, canonical reference, clippy allow Co-Authored-By: Claude Sonnet 5.5 --- crates/raw/src/huffman.rs | 50 ++++++++++++++++++++++++++++------- crates/raw/src/ljpeg.rs | 1 + crates/raw/src/ljpeg/tests.rs | 41 +++++++++++++++++++++++----- 3 files changed, 76 insertions(+), 16 deletions(-) diff --git a/crates/raw/src/huffman.rs b/crates/raw/src/huffman.rs index 85b242be1..4f35f64bb 100644 --- a/crates/raw/src/huffman.rs +++ b/crates/raw/src/huffman.rs @@ -145,9 +145,10 @@ mod tests { /// Reference: match the code bit by bit, then read the category's additional bits (T.81 F.1.2.1). fn reference(src: &[u8], pos: &mut usize, codes: &[(u16, u8)]) -> Option { let window: u32 = (0..12).fold(0, |a, k| (a << 1) | bit(src, *pos + k)); - let (sym, len) = codes.iter().enumerate().find_map(|(s, &(c, l))| { - (l > 0 && (c as u32 >> (12 - l)) == window >> (12 - l)).then_some((s as u32, l as usize)) - })?; + let (sym, len) = codes + .iter() + .enumerate() + .find_map(|(s, &(c, l))| (l > 0 && (c as u32 >> (12 - l)) == window >> (12 - l)).then_some((s as u32, l as usize)))?; *pos += len; if sym == 0 { return Some(0); @@ -157,13 +158,25 @@ mod tests { Some(if v < 1 << (sym - 1) { v - (1 << sym) + 1 } else { v }) } - #[test] - fn diff_matches_bitwise_reference_for_every_category() { - // categories 0..=16, code lengths up to 12 so that short codes combine with their bits and long ones do not - let lens = [2u8, 2, 3, 3, 4, 4, 5, 6, 7, 8, 9, 10, 11, 12, 12, 12, 12]; - let codes = canonical(&lens); + /// `diff` against the reference on random streams (invalid codes are stepped over) and on every (code, extra + /// bits) pair of every symbol. + fn check_table(lens: &[u8]) { + let codes = canonical(lens); let h = Huffman::new(&codes).unwrap(); + for (sym, &(code, len)) in codes.iter().enumerate() { + for v in 0..1u64 << sym { + let (len, n) = (len as u32, sym as u32); + let word = (((code >> (12 - len)) as u64) << n | v) << (64 - len - n); + let src = word.to_be_bytes(); + let mut pos = 0; + let want = reference(&src, &mut pos, &codes); + let mut bits = Bits::new(&src); + assert_eq!(diff(&mut bits, &h), want, "symbol {sym} bits {v:#b}"); + assert_eq!(bits.consumed_bits(), pos); + } + } let mut s = 0x1234_5678_9abc_def0u64; + let (mut invalid, mut long) = (0, 0); for _ in 0..40 { let src: Vec = (0..400) .map(|_| { @@ -174,13 +187,32 @@ mod tests { let mut bits = Bits::new(&src); let mut pos = 0; while pos < src.len() * 8 - 64 { + let before = pos; let want = reference(&src, &mut pos, &codes); assert_eq!(diff(&mut bits, &h), want); if want.is_none() { - break; + // step over the invalid code + invalid += 1; + pos += 1; + bits.skip(1); + } else { + long += (pos - before > 12) as u32; } assert_eq!(bits.consumed_bits(), pos); } } + assert!(long > 0, "{invalid} invalid codes, no symbol longer than the lookup"); + } + + #[test] + fn diff_matches_bitwise_reference_for_every_category() { + // categories 0..=16, code lengths up to 12 so that short codes combine with their bits and long ones do not + check_table(&[2, 2, 3, 3, 4, 4, 5, 6, 7, 8, 9, 10, 11, 12, 12, 12, 12]); + } + + #[test] + fn diff_matches_reference_when_code_and_bits_fill_the_lookup_exactly() { + // category 8 with a 4-bit code and category 7 with a 5-bit code: len + category == 12 + check_table(&[2, 3, 3, 5, 5, 5, 6, 5, 4, 7, 8, 9, 10, 11, 12, 12, 12]); } } diff --git a/crates/raw/src/ljpeg.rs b/crates/raw/src/ljpeg.rs index ab27ed57f..498a8302f 100644 --- a/crates/raw/src/ljpeg.rs +++ b/crates/raw/src/ljpeg.rs @@ -778,6 +778,7 @@ pub fn decode(d: &[u8], max_samples: usize) -> Result { restart_left -= end - x; } // the first pixel of the run has the special predictions (start of the interval, first column) + #[allow(clippy::needless_range_loop)] // `c` indexes the tables and the sample position alike for c in 0..nc { let i = x * nc + c; let p = if reset_pending { diff --git a/crates/raw/src/ljpeg/tests.rs b/crates/raw/src/ljpeg/tests.rs index 17791ea8b..a77425d40 100644 --- a/crates/raw/src/ljpeg/tests.rs +++ b/crates/raw/src/ljpeg/tests.rs @@ -181,15 +181,42 @@ fn roundtrip_skewed_categories_with_long_codes() { } } -/// `decode_diff` is `decode` followed by `diff_value`, for tables with codes inside and beyond the lookup. +/// Canonical decoding from the code lengths alone, bit by bit (T.81 Annex F, Figure F.16): the reference for the +/// lookup tables. +fn canonical_symbol(br: &mut BitReader, lens: &[u8]) -> Option { + let window = br.peek(16); + let (mut first, mut index) = (0u32, 0usize); + for len in 1..=16u8 { + let code = window >> (16 - len); + let n = lens.iter().filter(|&&l| l == len).count() as u32; + if code < first + n { + br.consume(len as u32); + return Some((index + (code - first) as usize) as u8); + } + index += n as usize; + first = (first + n) << 1; + } + None +} + +/// `decode_diff` against canonical decoding and `diff_value`, for tables with codes inside and beyond the lookup +/// and with `len + category` equal to the lookup width. #[test] -fn decode_diff_matches_decode_and_diff_value() { - for lens in [[2u8, 2, 3, 3, 4, 4, 5, 6, 7, 8, 9, 10, 11, 12, 12, 12, 12], [2, 2, 3, 3, 4, 4, 5, 6, 7, 8, 9, 10, 13, 14, 15, 16, 16]] { +fn decode_diff_matches_canonical_decoding() { + let tables: [[u8; 17]; 3] = [ + [2, 2, 3, 3, 4, 4, 5, 6, 7, 8, 9, 10, 11, 12, 12, 12, 12], + [2, 2, 3, 3, 4, 4, 5, 6, 7, 8, 9, 10, 13, 14, 15, 16, 16], + [2, 3, 3, 5, 5, 5, 6, 5, 4, 7, 8, 9, 10, 11, 12, 12, 12], + ]; + for lens in tables { let mut counts = [0u8; 16]; for &l in &lens { counts[l as usize - 1] += 1; } - let values: Vec = (0..17).collect(); // lengths are already ascending + // symbols in code order: by length, ties by symbol + let mut values: Vec = (0..17).collect(); + values.sort_by_key(|&v| (lens[v as usize], v)); + let sorted: Vec = values.iter().map(|&v| lens[v as usize]).collect(); let t = Huffman::new(&counts, &values).unwrap(); let mut s = 77u64; let src: Vec = (0..6000) @@ -200,11 +227,11 @@ fn decode_diff_matches_decode_and_diff_value() { .collect(); let (mut a, mut b) = (BitReader::new(&src), BitReader::new(&src)); for n in 0..8000 { - let want = t.decode(&mut a).map(|ssss| diff_value(&mut a, ssss)); + let want = canonical_symbol(&mut a, &sorted).map(|i| diff_value(&mut a, values[i as usize])); let got = t.decode_diff(&mut b, false); - assert_eq!(got.ok(), want.ok(), "symbol {n}"); + assert_eq!(got.ok(), want, "symbol {n}"); assert_eq!(a.peek(32), b.peek(32), "symbol {n}"); - if want.is_err() { + if want.is_none() { // step past the invalid code so the walk goes on a.consume(1); b.consume(1); From a05083e2f02ff88f0729f3ef61ee2e6eaef24d04 Mon Sep 17 00:00:00 2001 From: "Dylan M. Taylor" Date: Sun, 15 Nov 2026 17:37:40 +0000 Subject: [PATCH 04/20] Expose original AVIF sample depth during decode --- crates/image-codecs/src/codecs/avif_decode.rs | 8 ++++++-- crates/image-codecs/src/lib.rs | 8 ++++++++ 2 files changed, 14 insertions(+), 2 deletions(-) diff --git a/crates/image-codecs/src/codecs/avif_decode.rs b/crates/image-codecs/src/codecs/avif_decode.rs index 302433661..4dd15fd7d 100644 --- a/crates/image-codecs/src/codecs/avif_decode.rs +++ b/crates/image-codecs/src/codecs/avif_decode.rs @@ -822,6 +822,10 @@ fn read_metadata(c: &Container) -> (Option>, Option) { /// 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 { + decode_with_depth(bytes, limits, keep_orientation).map(|(image, _)| image) +} + +pub(crate) fn decode_with_depth(bytes: &[u8], limits: &Limits, keep_orientation: bool) -> Result<(Image, u8)> { let c = parse(bytes)?; let id = c.primary; let (w, h) = item_size(&c, id)?; @@ -860,9 +864,9 @@ pub(crate) fn decode(bytes: &[u8], limits: &Limits, keep_orientation: bool) -> R // 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); + return crate::cicp::linearize(img, cicp).map(|image| (image, bits)); } - Ok(img) + Ok((img, bits)) } #[cfg(test)] diff --git a/crates/image-codecs/src/lib.rs b/crates/image-codecs/src/lib.rs index a4480b2b8..86c4290b8 100644 --- a/crates/image-codecs/src/lib.rs +++ b/crates/image-codecs/src/lib.rs @@ -115,6 +115,14 @@ pub fn decode_as_with(format: Format, bytes: &[u8], opts: &DecodeOptions) -> Res finish_format(format, bytes, img, opts) } +/// Decode AVIF pixels and report their original AV1 bit depth (8, 10 or 12). +/// The returned image may use a wider sample type, particularly for linear HDR. +#[cfg(all(feature = "avif-decode", not(target_arch = "wasm32")))] +pub fn decode_avif_with_depth(bytes: &[u8], opts: &DecodeOptions) -> Result<(Image, u8), CodecError> { + let (image, depth) = codecs::avif_decode::decode_with_depth(bytes, &opts.limits, opts.keep_orientation)?; + Ok((finish_format(Format::Avif, bytes, image, opts)?, depth)) +} + /// What [`decode_as_with`] does with a PNG decoded in steps by [`PngDecodeRows`] (`img` is its /// output): the resolution from the metadata, the upright orientation and the final limit check, /// so the result equals [`decode_as_with`]'s on the same file. From 6325cd286bd87effa3020b01d6b9d289d4f7c055 Mon Sep 17 00:00:00 2001 From: "Dylan M. Taylor" Date: Sun, 15 Nov 2026 17:37:54 +0000 Subject: [PATCH 05/20] Test AVIF source depth and decoded-image parity --- crates/image-codecs/src/codecs/avif_decode.rs | 3 +++ 1 file changed, 3 insertions(+) diff --git a/crates/image-codecs/src/codecs/avif_decode.rs b/crates/image-codecs/src/codecs/avif_decode.rs index 4dd15fd7d..eaeafb65d 100644 --- a/crates/image-codecs/src/codecs/avif_decode.rs +++ b/crates/image-codecs/src/codecs/avif_decode.rs @@ -1194,6 +1194,9 @@ mod tests { 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(); + let (with_depth, depth) = crate::decode_avif_with_depth(&file, &crate::DecodeOptions::default()).unwrap(); + assert_eq!(depth, 10); + assert_eq!(with_depth, img); 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}"); From 9a9b89e9e8f057705e56398069c6f288a254b077 Mon Sep 17 00:00:00 2001 From: "Dylan M. Taylor" Date: Sun, 15 Nov 2026 17:43:57 +0000 Subject: [PATCH 06/20] Reject duplicate attributes in strict XML reads --- crates/xml/README.md | 6 +++--- crates/xml/src/read.rs | 25 +++++++++++++++++++++++- crates/xml/tests/hostile.rs | 38 +++++++++++++++++++++++++++++++++++++ 3 files changed, 65 insertions(+), 4 deletions(-) diff --git a/crates/xml/README.md b/crates/xml/README.md index adb178159..a61a46832 100644 --- a/crates/xml/README.md +++ b/crates/xml/README.md @@ -46,7 +46,7 @@ assert_eq!((rows, sheet.map(|s| s.children.len())), (3, Ok(0))); | `max_input` | 256 MiB | `TooLarge` | | `max_depth` | 256 levels (the root is 1) | `TooDeep` | | `max_nodes` | 2,000,000 nodes kept in the tree: elements, text stretches, kept CDATA, instructions and comments | `TooManyNodes` | -| `max_attrs` | 1024 on one start tag, namespace declarations and malformed ones included; with canonical names, also the declarations in scope at once | `TooManyAttrs` | +| `max_attrs` | 1024 on one start tag, namespace declarations and malformed ones included; with canonical names or strict reads, also the declarations in scope at once | `TooManyAttrs` | | `max_attrs_total` | 4,000,000 attributes kept in the tree, over all its elements | `TooManyAttrsTotal` | A limit exceeded is an error, never a panic and never a truncated tree. `parse_streaming_partial` @@ -73,9 +73,9 @@ A lenient read keeps what a damaged document had: - reading stops when the root element ends: nothing after it is looked at; - a malformed attribute is skipped; bytes that are not UTF-8 become U+FFFD. -A strict read makes each of those an error, and also a second root element and text outside the +A strict read rejects duplicate raw or expanded attribute names (while allowing undeclared prefixes in fragments). It makes each of those an error, and also a second root element and text outside the root, in the pull reader as in the tree. Markup the tokenizer cannot read (` Events<'o, 'i> { _ => None, }; let scope = self.scope.len(); - if canonical.is_some() { + if canonical.is_some() || self.opts.strict { self.declare(e)?; } let raw = e.name(); @@ -707,6 +708,8 @@ impl<'o, 'i> Events<'o, 'i> { let mut out = Vec::new(); let mut attrs = e.attributes(); attrs.with_checks(false); + let mut raw_names = BTreeSet::new(); + let mut expanded_names = BTreeSet::new(); for (seen, a) in attrs.enumerate() { if seen >= limits.max_attrs { return Err(Error::TooManyAttrs(limits.max_attrs)); @@ -718,6 +721,26 @@ impl<'o, 'i> Events<'o, 'i> { }; let key = a.key.as_ref(); let decl = is_ns_decl(key); + if self.opts.strict { + let raw_key = self.decode(key)?; + if !raw_names.insert(raw_key.to_string()) { + return self.syntax(format!("duplicate attribute `{raw_key}`")); + } + if !decl { + // Default namespaces do not apply to attributes. Keep undeclared prefixes + // distinct so standalone fragments retain the reader's existing behaviour. + let expanded = match raw_key.split_once(':') { + Some(("xml", local)) => (XML_NS, local), + Some((prefix, local)) => { + self.scope.iter().rev().find(|(p, _)| p == prefix).map_or(("", raw_key.as_ref()), |(_, uri)| (uri.as_str(), local)) + } + None => ("", raw_key.as_ref()), + }; + if !expanded_names.insert((expanded.0.to_string(), expanded.1.to_string())) { + return self.syntax(format!("duplicate expanded attribute `{raw_key}`")); + } + } + } if decl && !self.opts.ns_decls { continue; } diff --git a/crates/xml/tests/hostile.rs b/crates/xml/tests/hostile.rs index 2d387519b..6a846d3a7 100644 --- a/crates/xml/tests/hostile.rs +++ b/crates/xml/tests/hostile.rs @@ -524,3 +524,41 @@ fn a_strict_pull_reader_refuses_what_is_outside_the_root() { // white space, comments and instructions may stand outside the root assert!(read(strict, b"\r\n \tx\n ").is_ok()); } + +#[test] +fn strict_attributes_are_unique_before_name_rewriting() { + use craft_xml::{Names, Options}; + for names in [Names::AsWritten, Names::Local] { + let strict = Options { names, ..Options::default() }.strict(); + for xml in [ + "", + "", + "", + "", + ] { + assert!(strict.parse(xml.as_bytes()).is_err(), "{xml}"); + let mut events = strict.events(xml.as_bytes()); + assert!( + loop { + match events.next() { + Ok(Some(_)) => {} + Ok(None) => break false, + Err(_) => break true, + } + }, + "{xml}" + ); + } + for xml in [ + "", + "", + "", + ] { + assert!(strict.parse(xml.as_bytes()).is_ok(), "{xml}"); + } + assert!(Options { names, ..Options::default() }.parse(b"").is_ok()); + } + let strict = Options { limits: craft_xml::Limits { max_attrs: 2, ..craft_xml::Limits::DEFAULT }, ..Options::default() }.strict(); + assert!(strict.parse(b"").is_err()); + assert!(strict.parse(b"").is_err()); +} From 99007f1915d8b3495effffdfb8550a94a4d0bf87 Mon Sep 17 00:00:00 2001 From: "Dylan M. Taylor" Date: Sun, 15 Nov 2026 17:45:11 +0000 Subject: [PATCH 07/20] Allow strict duplicate error in attribute flood test --- crates/xml/tests/hostile.rs | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/crates/xml/tests/hostile.rs b/crates/xml/tests/hostile.rs index 6a846d3a7..37684f059 100644 --- a/crates/xml/tests/hostile.rs +++ b/crates/xml/tests/hostile.rs @@ -117,7 +117,7 @@ fn an_attribute_flood_is_refused_at_the_limit() { let opts = opts.with_limits(limits); for (k, kind) in kinds.iter().enumerate() { let _ = opts.parse(tag(16, kind).as_bytes()); - assert_eq!(opts.parse(tag(17, kind).as_bytes()).err().map(|e| e.is_limit() || (opts.strict && k == 3)), Some(true), "kind {k}"); + assert_eq!(opts.parse(tag(17, kind).as_bytes()).err().map(|e| e.is_limit() || (opts.strict && (k == 1 || k == 3))), Some(true), "kind {k}"); assert!(opts.first_start(tag(17, kind).as_bytes(), "a").is_none()); } assert_eq!(opts.parse(tag(17, kinds[0]).as_bytes()), Err(Error::TooManyAttrs(16))); From 24c496e95af81484812d62c72a6dcf8f119679ed Mon Sep 17 00:00:00 2001 From: "Dylan M. Taylor" Date: Sun, 15 Nov 2026 18:00:22 +0000 Subject: [PATCH 08/20] Format strict XML attribute flood regression --- crates/xml/tests/hostile.rs | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/crates/xml/tests/hostile.rs b/crates/xml/tests/hostile.rs index 37684f059..4ef151914 100644 --- a/crates/xml/tests/hostile.rs +++ b/crates/xml/tests/hostile.rs @@ -117,7 +117,11 @@ fn an_attribute_flood_is_refused_at_the_limit() { let opts = opts.with_limits(limits); for (k, kind) in kinds.iter().enumerate() { let _ = opts.parse(tag(16, kind).as_bytes()); - assert_eq!(opts.parse(tag(17, kind).as_bytes()).err().map(|e| e.is_limit() || (opts.strict && (k == 1 || k == 3))), Some(true), "kind {k}"); + assert_eq!( + opts.parse(tag(17, kind).as_bytes()).err().map(|e| e.is_limit() || (opts.strict && (k == 1 || k == 3))), + Some(true), + "kind {k}" + ); assert!(opts.first_start(tag(17, kind).as_bytes(), "a").is_none()); } assert_eq!(opts.parse(tag(17, kinds[0]).as_bytes()), Err(Error::TooManyAttrs(16))); From 02224df6cb51563b4cade129399dc2b197a4bee9 Mon Sep 17 00:00:00 2001 From: Daniel Voshart <205798369+voshart@users.noreply.github.com> Date: Sun, 15 Nov 2026 16:48:09 -0500 Subject: [PATCH 09/20] Import LightCraft HTTP URL parser --- crates/http-url/src/lib.rs | 126 +++++++++++++++++++++++++++++++++++++ 1 file changed, 126 insertions(+) create mode 100644 crates/http-url/src/lib.rs diff --git a/crates/http-url/src/lib.rs b/crates/http-url/src/lib.rs new file mode 100644 index 000000000..3babf1b13 --- /dev/null +++ b/crates/http-url/src/lib.rs @@ -0,0 +1,126 @@ +/// A parsed `http://` or `https://` URL. +#[derive(Debug, Clone, PartialEq, Eq)] +pub struct Url { + pub tls: bool, + pub host: String, + pub port: u16, + /// Path and query, starting with `/`. + pub path: String, +} + +impl Url { + pub fn parse(s: &str) -> Result { + let bad = |why: &str| HttpError::BadUrl(format!("{s}: {why}")); + let s = s.trim(); + if s.chars().any(|c| c.is_control() || c == ' ') { + return Err(bad("contains spaces or control characters")); + } + let (tls, rest) = if let Some(r) = s.strip_prefix("https://") { + (true, r) + } else if let Some(r) = s.strip_prefix("http://") { + (false, r) + } else { + return Err(bad("only http:// and https:// are supported")); + }; + let rest = rest.split('#').next().unwrap_or_default(); + let (authority, path) = match rest.find(['/', '?']) { + Some(i) => (rest.get(..i).unwrap_or_default(), rest.get(i..).unwrap_or_default()), + None => (rest, ""), + }; + let path = if path.starts_with('/') { path.to_string() } else { format!("/{path}") }; + if authority.contains('@') { + return Err(bad("user names and passwords in URLs are not supported")); + } + let (host, port) = if let Some(v6) = authority.strip_prefix('[') { + let end = v6.find(']').ok_or_else(|| bad("unterminated IPv6 address"))?; + let host = v6.get(..end).unwrap_or_default(); + let after = v6.get(end + 1..).unwrap_or_default(); + (host.to_string(), after.strip_prefix(':')) + } else { + match authority.rsplit_once(':') { + Some((h, p)) => (h.to_string(), Some(p)), + None => (authority.to_string(), None), + } + }; + let port = match port { + Some(p) => p.parse::().ok().filter(|p| *p != 0).ok_or_else(|| bad("bad port"))?, + None if tls => 443, + None => 80, + }; + if host.is_empty() || !host.chars().all(|c| c.is_ascii_alphanumeric() || matches!(c, '-' | '.' | ':' | '_')) { + return Err(bad("bad host name")); + } + Ok(Url { tls, host: host.to_ascii_lowercase(), port, path }) + } + + /// `Location` of a redirect, resolved against this URL. + pub fn join(&self, location: &str) -> Result { + let l = location.trim(); + if l.starts_with("http://") || l.starts_with("https://") { + return Url::parse(l); + } + // another scheme (`ftp:`, `file:`, `data:`…) + if let Some((scheme, _)) = l.split_once(':') + && !scheme.is_empty() + && scheme.chars().all(|c| c.is_ascii_alphanumeric() || matches!(c, '+' | '-' | '.')) + { + return Err(HttpError::BadUrl(format!("redirect to an unsupported location `{scheme}:`"))); + } + let scheme = if self.tls { "https" } else { "http" }; + let host = if self.host.contains(':') { format!("[{}]", self.host) } else { self.host.clone() }; + let base = format!("{scheme}://{host}:{}", self.port); + if let Some(rest) = l.strip_prefix("//") { + return Url::parse(&format!("{scheme}://{rest}")); + } + if l.starts_with('/') { + return Url::parse(&format!("{base}{l}")); + } + // relative to the current directory + let dir = self.path.split('?').next().unwrap_or("/"); + let dir = dir.rsplit_once('/').map_or("", |(d, _)| d); + Url::parse(&format!("{base}{dir}/{l}")) + } + + fn host_header(&self) -> String { + let host = if self.host.contains(':') { format!("[{}]", self.host) } else { self.host.clone() }; + let default = if self.tls { 443 } else { 80 }; + if self.port == default { host } else { format!("{host}:{}", self.port) } + } + + /// The URL without its query (for messages: a query may hold a signed token). + pub fn display(&self) -> String { + let scheme = if self.tls { "https" } else { "http" }; + format!("{scheme}://{}{}", self.host_header(), self.path.split('?').next().unwrap_or("/")) + } +} + +#[cfg(test)] +mod tests { + use super::*; + + #[test] + fn parses_urls() { + let u = Url::parse("https://cdn.example.com/models/sam3/model.safetensors?sig=1#x").unwrap(); + assert_eq!((u.tls, u.host.as_str(), u.port, u.path.as_str()), (true, "cdn.example.com", 443, "/models/sam3/model.safetensors?sig=1")); + assert_eq!(u.display(), "https://cdn.example.com/models/sam3/model.safetensors"); + let u = Url::parse("http://127.0.0.1:8080").unwrap(); + assert_eq!((u.tls, u.port, u.path.as_str()), (false, 8080, "/")); + let u = Url::parse("http://[::1]:9/a").unwrap(); + assert_eq!((u.host.as_str(), u.port), ("::1", 9)); + for bad in ["ftp://x/y", "https://", "https://a b/", "https://u:p@h/", "http://h:0/", "http://h:99999/", "https://h/\r\nX: y", "file:///etc"] + { + assert!(Url::parse(bad).is_err(), "{bad}"); + } + } + + #[test] + fn resolves_redirects() { + let u = Url::parse("https://a.example/x/y/file?q").unwrap(); + assert_eq!(u.join("https://b.example/z").unwrap().host, "b.example"); + assert_eq!(u.join("/root").unwrap().path, "/root"); + assert_eq!(u.join("other").unwrap().path, "/x/y/other"); + assert_eq!(u.join("//c.example/p").unwrap().host, "c.example"); + assert!(u.join("ftp://nope").is_err()); + } + +} From d577b23c9c5b796161d6eec275dd9afd8163a24d Mon Sep 17 00:00:00 2001 From: Daniel Voshart <205798369+voshart@users.noreply.github.com> Date: Sun, 15 Nov 2026 16:53:07 -0500 Subject: [PATCH 10/20] Preserve bounded URLs and query-safe error messages --- crates/http-url/src/lib.rs | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-) diff --git a/crates/http-url/src/lib.rs b/crates/http-url/src/lib.rs index 3babf1b13..bae96da5b 100644 --- a/crates/http-url/src/lib.rs +++ b/crates/http-url/src/lib.rs @@ -10,7 +10,10 @@ pub struct Url { impl Url { pub fn parse(s: &str) -> Result { - let bad = |why: &str| HttpError::BadUrl(format!("{s}: {why}")); + let bad = |why: &str| HttpError::BadUrl(why.to_string()); + if s.len() > 16 * 1024 { + return Err(bad("URL exceeds 16 KiB")); + } let s = s.trim(); if s.chars().any(|c| c.is_control() || c == ' ') { return Err(bad("contains spaces or control characters")); @@ -98,6 +101,12 @@ impl Url { mod tests { use super::*; + #[test] + fn rejects_oversized_and_control_character_urls() { + assert!(Url::parse(&format!("https://example.com/{}", "x".repeat(16 * 1024))).is_err()); + assert!(Url::parse("https://example.com/x?secret=abc\r\nInjected: 1").is_err()); + } + #[test] fn parses_urls() { let u = Url::parse("https://cdn.example.com/models/sam3/model.safetensors?sig=1#x").unwrap(); From 07fa210956f57ae31c09515c49fc3e59120a03b6 Mon Sep 17 00:00:00 2001 From: MarekSpodBiedry <73122968+MarekSpodBiedry@users.noreply.github.com> Date: Sun, 15 Nov 2026 16:53:07 -0500 Subject: [PATCH 11/20] Preserve HTTPS redirect downgrade rejection --- crates/http-url/src/lib.rs | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-) diff --git a/crates/http-url/src/lib.rs b/crates/http-url/src/lib.rs index bae96da5b..5c80f3c4c 100644 --- a/crates/http-url/src/lib.rs +++ b/crates/http-url/src/lib.rs @@ -60,7 +60,11 @@ impl Url { pub fn join(&self, location: &str) -> Result { let l = location.trim(); if l.starts_with("http://") || l.starts_with("https://") { - return Url::parse(l); + let url = Url::parse(l)?; + if self.tls && !url.tls { + return Err(HttpError::BadUrl("redirect from https to http refused".into())); + } + return Ok(url); } // another scheme (`ftp:`, `file:`, `data:`…) if let Some((scheme, _)) = l.split_once(':') @@ -130,6 +134,11 @@ mod tests { assert_eq!(u.join("other").unwrap().path, "/x/y/other"); assert_eq!(u.join("//c.example/p").unwrap().host, "c.example"); assert!(u.join("ftp://nope").is_err()); + // no downgrade from https to http, however the redirect is written + assert!(u.join("http://a.example/z").is_err()); + let plain = Url::parse("http://127.0.0.1:8080/a").unwrap(); + assert!(!plain.join("http://127.0.0.1:9/b").unwrap().tls); + assert!(plain.join("https://b.example/z").unwrap().tls); } } From 837e4d365a3ce56ac21b2fc95ebbaeeee8629b35 Mon Sep 17 00:00:00 2001 From: NetterSilver <133535268+Silvertree2010@users.noreply.github.com> Date: Sun, 15 Nov 2026 16:48:09 -0500 Subject: [PATCH 12/20] Preserve URL normalization and malformed authority regressions --- crates/http-url/src/lib.rs | 112 +++++++++++++++++++++++++++++++++++-- 1 file changed, 108 insertions(+), 4 deletions(-) diff --git a/crates/http-url/src/lib.rs b/crates/http-url/src/lib.rs index 5c80f3c4c..5bde118a3 100644 --- a/crates/http-url/src/lib.rs +++ b/crates/http-url/src/lib.rs @@ -8,6 +8,43 @@ pub struct Url { pub path: String, } +/// Remove `.` and `..` segments from the path part of `path` (RFC 3986 section 5.2.4); the query is kept as is. +fn remove_dot_segments(path: &str) -> String { + let (p, query) = match path.find('?') { + Some(i) => (path.get(..i).unwrap_or_default(), path.get(i..).unwrap_or_default()), + None => (path, ""), + }; + let mut out: Vec<&str> = Vec::new(); + let mut segments = p.split('/').skip(1).peekable(); + let mut trailing_slash = false; + while let Some(seg) = segments.next() { + let last = segments.peek().is_none(); + match seg { + "." => trailing_slash = last, + ".." => { + out.pop(); + trailing_slash = last; + } + _ => { + out.push(seg); + trailing_slash = false; + } + } + } + let mut result = format!("/{}", out.join("/")); + if trailing_slash && !result.ends_with('/') { + result.push('/'); + } + result.push_str(query); + result +} + +/// `s` without the leading `prefix`, matched ASCII case-insensitively (RFC 3986 §3.1). +fn strip_scheme<'a>(s: &'a str, prefix: &str) -> Option<&'a str> { + let head = s.get(..prefix.len())?; + head.eq_ignore_ascii_case(prefix).then(|| s.get(prefix.len()..).unwrap_or_default()) +} + impl Url { pub fn parse(s: &str) -> Result { let bad = |why: &str| HttpError::BadUrl(why.to_string()); @@ -18,9 +55,9 @@ impl Url { if s.chars().any(|c| c.is_control() || c == ' ') { return Err(bad("contains spaces or control characters")); } - let (tls, rest) = if let Some(r) = s.strip_prefix("https://") { + let (tls, rest) = if let Some(r) = strip_scheme(s, "https://") { (true, r) - } else if let Some(r) = s.strip_prefix("http://") { + } else if let Some(r) = strip_scheme(s, "http://") { (false, r) } else { return Err(bad("only http:// and https:// are supported")); @@ -31,6 +68,7 @@ impl Url { None => (rest, ""), }; let path = if path.starts_with('/') { path.to_string() } else { format!("/{path}") }; + let path = remove_dot_segments(&path); if authority.contains('@') { return Err(bad("user names and passwords in URLs are not supported")); } @@ -38,6 +76,9 @@ impl Url { let end = v6.find(']').ok_or_else(|| bad("unterminated IPv6 address"))?; let host = v6.get(..end).unwrap_or_default(); let after = v6.get(end + 1..).unwrap_or_default(); + if !after.is_empty() && !after.starts_with(':') { + return Err(bad("unexpected text after IPv6 address")); + } (host.to_string(), after.strip_prefix(':')) } else { match authority.rsplit_once(':') { @@ -46,7 +87,8 @@ impl Url { } }; let port = match port { - Some(p) => p.parse::().ok().filter(|p| *p != 0).ok_or_else(|| bad("bad port"))?, + Some(p) if p.bytes().all(|b| b.is_ascii_digit()) => p.parse::().ok().filter(|p| *p != 0).ok_or_else(|| bad("bad port"))?, + Some(_) => return Err(bad("bad port")), None if tls => 443, None => 80, }; @@ -59,7 +101,7 @@ impl Url { /// `Location` of a redirect, resolved against this URL. pub fn join(&self, location: &str) -> Result { let l = location.trim(); - if l.starts_with("http://") || l.starts_with("https://") { + if strip_scheme(l, "http://").is_some() || strip_scheme(l, "https://").is_some() { let url = Url::parse(l)?; if self.tls && !url.tls { return Err(HttpError::BadUrl("redirect from https to http refused".into())); @@ -82,8 +124,16 @@ impl Url { if l.starts_with('/') { return Url::parse(&format!("{base}{l}")); } + // empty or fragment-only: the same document + if l.is_empty() || l.starts_with('#') { + return Ok(self.clone()); + } // relative to the current directory let dir = self.path.split('?').next().unwrap_or("/"); + // a query-only reference keeps the whole path, file name included (RFC 3986 §5.2.2) + if l.starts_with('?') { + return Url::parse(&format!("{base}{dir}{l}")); + } let dir = dir.rsplit_once('/').map_or("", |(d, _)| d); Url::parse(&format!("{base}{dir}/{l}")) } @@ -105,12 +155,46 @@ impl Url { mod tests { use super::*; + #[test] + fn fragment_only_or_empty_location_keeps_file_name_and_query() { + let base = Url::parse("https://a.example/x/y/file?q").unwrap(); + for l in ["#top", "", " "] { + let u = base.join(l).unwrap(); + assert_eq!((u.tls, u.host.as_str(), u.port, u.path.as_str()), (true, "a.example", 443, "/x/y/file?q"), "{l:?}"); + } + } + #[test] fn rejects_oversized_and_control_character_urls() { assert!(Url::parse(&format!("https://example.com/{}", "x".repeat(16 * 1024))).is_err()); assert!(Url::parse("https://example.com/x?secret=abc\r\nInjected: 1").is_err()); } + #[test] + fn join_query_only_keeps_file_name() { + let u = Url::parse("https://a.example/x/y/file?q=1").unwrap(); + assert_eq!(u.join("?page=2").unwrap().path, "/x/y/file?page=2"); + } + + #[test] + fn join_removes_dot_segments() { + let base = Url::parse("https://a.example/x/y/file").unwrap(); + for (loc, want) in + [("../z", "/x/z"), ("./z", "/x/y/z"), ("../../../z", "/z"), ("/a/./b/../c", "/a/c"), ("../z?q=1", "/x/z?q=1"), ("..", "/x/")] + { + assert_eq!(base.join(loc).unwrap().path, want, "{loc}"); + } + } + + #[test] + fn url_scheme_is_matched_case_insensitively() { + let u = Url::parse("HTTPS://A.example/z").unwrap(); + assert_eq!((u.tls, u.host.as_str(), u.port, u.path.as_str()), (true, "a.example", 443, "/z")); + assert!(Url::parse("Http://a.example/").is_ok()); + let j = Url::parse("https://a.example/start").unwrap().join("HTTPS://cdn.example/file").unwrap(); + assert_eq!((j.tls, j.host.as_str(), j.path.as_str()), (true, "cdn.example", "/file")); + } + #[test] fn parses_urls() { let u = Url::parse("https://cdn.example.com/models/sam3/model.safetensors?sig=1#x").unwrap(); @@ -126,6 +210,26 @@ mod tests { } } + #[test] + fn rejects_signed_ports() { + for bad in ["http://example.com:+80/x", "http://example.com:-80/x", "http://[::1]:+80/x"] { + assert!(Url::parse(bad).is_err(), "{bad}"); + } + assert!(Url::parse("http://a.test/").unwrap().join("http://example.com:+80/x").is_err()); + assert_eq!(Url::parse("http://example.com:80/x").unwrap().port, 80); + } + + #[test] + fn rejects_text_after_a_bracketed_ipv6_host() { + for bad in ["http://[::1]junk/x", "https://[::1]junk/x", "http://[::1]8080/x", "http://[::1]8080"] { + assert!(matches!(Url::parse(bad), Err(HttpError::BadUrl(_))), "{bad}"); + } + let u = Url::parse("http://[::1]:8080/x").unwrap(); + assert_eq!((u.host.as_str(), u.port), ("::1", 8080)); + let u = Url::parse("http://[::1]/x").unwrap(); + assert_eq!((u.host.as_str(), u.port), ("::1", 80)); + } + #[test] fn resolves_redirects() { let u = Url::parse("https://a.example/x/y/file?q").unwrap(); From c247a085be0b2db5456f1f9178024d6a12e4cc57 Mon Sep 17 00:00:00 2001 From: "Dylan M. Taylor" Date: Sun, 15 Nov 2026 16:50:07 -0500 Subject: [PATCH 13/20] Expose dependency-free HTTP URL syntax with bounded redirect input --- Cargo.lock | 4 ++++ Cargo.toml | 1 + crates/http-url/Cargo.toml | 11 +++++++++ crates/http-url/README.md | 25 ++++++++++++++++++++ crates/http-url/src/lib.rs | 47 ++++++++++++++++++++++++++++++-------- 5 files changed, 79 insertions(+), 9 deletions(-) create mode 100644 crates/http-url/Cargo.toml create mode 100644 crates/http-url/README.md diff --git a/Cargo.lock b/Cargo.lock index b3325f790..b36240d6e 100644 --- a/Cargo.lock +++ b/Cargo.lock @@ -1879,6 +1879,10 @@ dependencies = [ "proptest", ] +[[package]] +name = "craft-http-url" +version = "0.1.0" + [[package]] name = "craft-i18n" version = "0.1.0" diff --git a/Cargo.toml b/Cargo.toml index 0c9729e00..1b409247e 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-http-url = { path = "crates/http-url" } craft-smime = { path = "crates/smime" } craft-ber = { path = "crates/ber" } craft-proto = { path = "crates/proto" } diff --git a/crates/http-url/Cargo.toml b/crates/http-url/Cargo.toml new file mode 100644 index 000000000..f23d9bbc7 --- /dev/null +++ b/crates/http-url/Cargo.toml @@ -0,0 +1,11 @@ +[package] +name = "craft-http-url" +version.workspace = true +edition.workspace = true +rust-version.workspace = true +license = "Apache-2.0" +repository.workspace = true +description = "Bounded HTTP URL parsing and redirect resolution without network access" + +[lints] +workspace = true diff --git a/crates/http-url/README.md b/crates/http-url/README.md new file mode 100644 index 000000000..cc3dc1bf5 --- /dev/null +++ b/crates/http-url/README.md @@ -0,0 +1,25 @@ +# craft-http-url + +HTTP/HTTPS URL parsing, dot-segment removal, relative redirects, HTTPS downgrade rejection, +and query-redacted display. No dependencies or network operations. Parsing rejects inputs +larger than 16 KiB and redirects are bounded before concatenation. + +This is not an SSRF filter, DNS resolver or HTTP client. Local/private hosts are allowed. +Applications keep destination permission, authentication, TLS, request/response limits and +redirect-count policy. Public fields preserve the original API; manually constructed values +must not be treated as validated URLs. + +## Attribution + +Extracted from LightCraft `crates/fetch/src/http.rs` at +`2dcb258ba40738230a07484a95eb27dc84ed66fe` (Apache-2.0). +Original parser: Daniel Voshart, `7fa7506db46d69d609e85e39b338f33786ed0293`. +HTTPS downgrade protection: MarekSpodBiedry, `4a91a1c218bf3d0ac49cc73ebb60621a21ae55e2`. +URL normalization, scheme, fragment and authority fixes: NetterSilver, including +`edc369565eef92161b56b8f1e39ef53629c05489`, +`78f0e6e650444886069f4d4ae4e6593f221b98d4`, +`cee93496587d31c514aaeced3e6b8027cfdb41a0`, +`dee5f7910ad8c24cf28bab528f8d007550af2b23`, +`c3a5d9b7f1f837dc14c01942e3d0293ecbf36345` and +`13849307a9343504339d68b086f1a32cdcbbb457`. +Source-author commits are retained separately in this extraction. diff --git a/crates/http-url/src/lib.rs b/crates/http-url/src/lib.rs index 5bde118a3..1e3e5875f 100644 --- a/crates/http-url/src/lib.rs +++ b/crates/http-url/src/lib.rs @@ -1,3 +1,24 @@ +//! Bounded HTTP/HTTPS URL parsing and redirect resolution, extracted from LightCraft. +//! +//! This is syntax handling, not a network permission or SSRF policy. It permits local/private +//! hosts and performs no DNS, socket or TLS operations. Callers own destination allowlists, +//! credential forwarding, redirect counts, connection deadlines and response limits. +//! Only parsed values have validated fields; constructing public fields bypasses parsing. + +#![forbid(unsafe_code)] +#![deny(clippy::unwrap_used, clippy::expect_used, clippy::panic, clippy::unimplemented, clippy::todo, clippy::unreachable)] + +/// A URL or redirect cannot be represented by this bounded HTTP URL syntax. +#[derive(Debug, Clone, PartialEq, Eq)] +pub struct UrlError(pub String); + +impl std::fmt::Display for UrlError { + fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { + f.write_str(&self.0) + } +} +impl std::error::Error for UrlError {} + /// A parsed `http://` or `https://` URL. #[derive(Debug, Clone, PartialEq, Eq)] pub struct Url { @@ -46,8 +67,8 @@ fn strip_scheme<'a>(s: &'a str, prefix: &str) -> Option<&'a str> { } impl Url { - pub fn parse(s: &str) -> Result { - let bad = |why: &str| HttpError::BadUrl(why.to_string()); + pub fn parse(s: &str) -> Result { + let bad = |why: &str| UrlError(why.to_string()); if s.len() > 16 * 1024 { return Err(bad("URL exceeds 16 KiB")); } @@ -99,12 +120,16 @@ impl Url { } /// `Location` of a redirect, resolved against this URL. - pub fn join(&self, location: &str) -> Result { + pub fn join(&self, location: &str) -> Result { + // Reject before concatenating a hostile Location with the base URL. + if location.len() > 16 * 1024 || self.host.len().saturating_add(self.path.len()) > 16 * 1024 { + return Err(UrlError("URL exceeds 16 KiB".into())); + } let l = location.trim(); if strip_scheme(l, "http://").is_some() || strip_scheme(l, "https://").is_some() { let url = Url::parse(l)?; if self.tls && !url.tls { - return Err(HttpError::BadUrl("redirect from https to http refused".into())); + return Err(UrlError("redirect from https to http refused".into())); } return Ok(url); } @@ -113,7 +138,7 @@ impl Url { && !scheme.is_empty() && scheme.chars().all(|c| c.is_ascii_alphanumeric() || matches!(c, '+' | '-' | '.')) { - return Err(HttpError::BadUrl(format!("redirect to an unsupported location `{scheme}:`"))); + return Err(UrlError(format!("redirect to an unsupported location `{scheme}:`"))); } let scheme = if self.tls { "https" } else { "http" }; let host = if self.host.contains(':') { format!("[{}]", self.host) } else { self.host.clone() }; @@ -138,7 +163,8 @@ impl Url { Url::parse(&format!("{base}{dir}/{l}")) } - fn host_header(&self) -> String { + /// Authority formatted for an HTTP Host field. + pub fn host_header(&self) -> String { let host = if self.host.contains(':') { format!("[{}]", self.host) } else { self.host.clone() }; let default = if self.tls { 443 } else { 80 }; if self.port == default { host } else { format!("{host}:{}", self.port) } @@ -153,7 +179,8 @@ impl Url { #[cfg(test)] mod tests { - use super::*; + #![allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)] + use super::{Url, UrlError}; #[test] fn fragment_only_or_empty_location_keeps_file_name_and_query() { @@ -166,6 +193,9 @@ mod tests { #[test] fn rejects_oversized_and_control_character_urls() { + let base = Url::parse("https://example.com/path").unwrap(); + assert!(base.join(&format!("#{}", "x".repeat(16 * 1024))).is_err()); + assert!(base.join(&"x".repeat(16 * 1024 + 1)).is_err()); assert!(Url::parse(&format!("https://example.com/{}", "x".repeat(16 * 1024))).is_err()); assert!(Url::parse("https://example.com/x?secret=abc\r\nInjected: 1").is_err()); } @@ -222,7 +252,7 @@ mod tests { #[test] fn rejects_text_after_a_bracketed_ipv6_host() { for bad in ["http://[::1]junk/x", "https://[::1]junk/x", "http://[::1]8080/x", "http://[::1]8080"] { - assert!(matches!(Url::parse(bad), Err(HttpError::BadUrl(_))), "{bad}"); + assert!(matches!(Url::parse(bad), Err(UrlError(_))), "{bad}"); } let u = Url::parse("http://[::1]:8080/x").unwrap(); assert_eq!((u.host.as_str(), u.port), ("::1", 8080)); @@ -244,5 +274,4 @@ mod tests { assert!(!plain.join("http://127.0.0.1:9/b").unwrap().tls); assert!(plain.join("https://b.example/z").unwrap().tls); } - } From 17db8defc24dc0ae1811f8700a58bb474dd270a9 Mon Sep 17 00:00:00 2001 From: NetterSilver <133535268+Silvertree2010@users.noreply.github.com> Date: Sun, 15 Nov 2026 17:31:18 -0500 Subject: [PATCH 14/20] cms: preserve every numeric token when reading Cube domains Adapt the strict-token correction from storytold/lightcraft#3464, original commit 77bb537552d59142814eaa65cfea4626c33d6dfc, to the existing shared Cube reader. --- crates/cms/src/lutfile.rs | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/crates/cms/src/lutfile.rs b/crates/cms/src/lutfile.rs index b7cc99bf3..8b3a8495f 100644 --- a/crates/cms/src/lutfile.rs +++ b/crates/cms/src/lutfile.rs @@ -176,7 +176,7 @@ pub fn parse_cube(text: &str) -> Result { let mut it = line.split_whitespace(); let key = it.next().unwrap_or(""); let nums = |it: std::str::SplitWhitespace<'_>| -> Result<[f32; 3], LutError> { - let v: Vec = it.filter_map(|s| s.parse().ok()).collect(); + let v: Vec = it.map(|s| s.parse().ok()).collect::>().ok_or_else(|| LutError(format!("bad line `{line}`")))?; if v.len() == 3 { Ok([v[0], v[1], v[2]]) } else { err(format!("bad line `{line}`")) } }; match key { From 33852afad44d968f2c78a0fef7f4011e76aa7044 Mon Sep 17 00:00:00 2001 From: Brandon Thomas Date: Sun, 15 Nov 2026 17:31:18 -0500 Subject: [PATCH 15/20] cms: share bounded three-dimensional Cube profile parsing Move LightCraft profile-reader semantics (origin 4a490db2) into the existing shared Cube engine. Preserve optional titles, finite endpoints and caller-selected limits; leave application interpolation and registry policies in the consumer. Integration and resource-bound adaptations by Dylan M. Taylor. --- crates/cms/src/lutfile.rs | 164 ++++++++++++++++++++++++++++---------- 1 file changed, 122 insertions(+), 42 deletions(-) diff --git a/crates/cms/src/lutfile.rs b/crates/cms/src/lutfile.rs index 8b3a8495f..1dff864c1 100644 --- a/crates/cms/src/lutfile.rs +++ b/crates/cms/src/lutfile.rs @@ -150,62 +150,142 @@ fn decode_text(bytes: &[u8]) -> std::borrow::Cow<'_, str> { } } -/// Adobe / Resolve `.cube`. 1D LUTs (`LUT_1D_SIZE`) are expanded to a 33³ table. -pub fn parse_cube(text: &str) -> Result { - let mut title = String::new(); - let (mut size3, mut size1) = (0usize, 0usize); - let (mut dmin, mut dmax) = ([0.0f32; 3], [1.0f32; 3]); - let mut rows: Vec = Vec::new(); - for line in text.lines() { - // Only a quote at the beginning of a TITLE value protects '#'. Quotes inside an - // unquoted title's comment must not turn that comment into part of the title. - if let Some(quoted) = line - .trim_start() - .strip_prefix("TITLE") - .filter(|rest| rest.starts_with(char::is_whitespace)) - .and_then(|rest| rest.trim_start().strip_prefix('"')) +/// A bounded, three-dimensional Cube profile with RGB rows in file order. +/// Unlike [`parse_cube`], this preserves the optional title and does not resample 1D curves. +#[derive(Clone, Debug, PartialEq)] +pub struct Cube3d { + pub title: Option, + pub size: usize, + pub data: Vec<[f32; 3]>, + pub domain_min: [f32; 3], + pub domain_max: [f32; 3], +} + +/// Read the simple 3D profile dialect used by camera-profile consumers. +/// +/// `max_size` must be in 2..=256; the row budget is checked before each allocation. +/// Whole-line comments and literal titles are preserved. Unknown alphabetic directives +/// are ignored; 1D tables are refused. Finite domains are returned unchanged so the +/// consumer retains its existing policy for equal or reversed endpoints. +/// Numeric fields must consist of exactly three finite numbers (never silently skip tokens). +pub fn parse_cube_3d(text: &str, max_size: usize) -> Result { + let p = read_cube(text, Some(max_size))?; + if p.size3 == 0 { + return err("not a 3D .cube LUT (no LUT_3D_SIZE)"); + } + let expected = p.size3.checked_pow(3).ok_or_else(|| LutError("LUT size overflow".into()))?; + if p.rows.len() != expected { + return err(format!("expected {expected} entries, found {}", p.rows.len())); + } + Ok(Cube3d { title: p.title, size: p.size3, data: p.rows, domain_min: p.min, domain_max: p.max }) +} + +struct CubeRows { + title: Option, + size3: usize, + size1: usize, + min: [f32; 3], + max: [f32; 3], + rows: Vec<[f32; 3]>, +} + +fn cube_triplet(text: &str) -> Result<[f32; 3], LutError> { + let mut it = text.split_whitespace(); + let mut row = [0.0; 3]; + for value in &mut row { + *value = it.next().and_then(|s| s.parse::().ok()).filter(|v| v.is_finite()).ok_or_else(|| LutError(format!("bad line `{text}`")))?; + } + if it.next().is_some() { + return err(format!("bad line `{text}`")); + } + Ok(row) +} + +fn read_cube(text: &str, profile_max: Option) -> Result { + if profile_max.is_some_and(|n| !(2..=256).contains(&n)) { + return err("3D profile size limit must be in 2..=256"); + } + let mut p = CubeRows { title: None, size3: 0, size1: 0, min: [0.0; 3], max: [1.0; 3], rows: Vec::new() }; + let row_limit = profile_max.unwrap_or(256).checked_pow(3).ok_or_else(|| LutError("LUT size overflow".into()))?; + for raw in text.lines() { + let profile = profile_max.is_some(); + let line = if profile { raw.trim() } else { raw.split('#').next().unwrap_or("").trim() }; + // The standard dialect protects '#' inside quoted titles; profile titles are literal. + if !profile + && let Some(quoted) = + raw.trim_start().strip_prefix("TITLE").filter(|r| r.starts_with(char::is_whitespace)).and_then(|r| r.trim_start().strip_prefix('"')) { - let Some((value, _)) = quoted.split_once('"') else { return err("unterminated quoted TITLE") }; - title = value.to_string(); + let Some((title, _)) = quoted.split_once('"') else { return err("unterminated quoted TITLE") }; + p.title = Some(title.to_string()); continue; } - let line = line.split('#').next().unwrap_or("").trim(); - if line.is_empty() { + if line.is_empty() || line.starts_with('#') { continue; } - let mut it = line.split_whitespace(); - let key = it.next().unwrap_or(""); - let nums = |it: std::str::SplitWhitespace<'_>| -> Result<[f32; 3], LutError> { - let v: Vec = it.map(|s| s.parse().ok()).collect::>().ok_or_else(|| LutError(format!("bad line `{line}`")))?; - if v.len() == 3 { Ok([v[0], v[1], v[2]]) } else { err(format!("bad line `{line}`")) } + let (key, rest) = if profile { + ["TITLE", "LUT_3D_SIZE", "LUT_1D_SIZE", "DOMAIN_MIN", "DOMAIN_MAX"] + .into_iter() + .find_map(|key| line.strip_prefix(key).map(|rest| (key, rest.trim()))) + .unwrap_or_else(|| line.split_once(char::is_whitespace).unwrap_or((line, ""))) + } else { + line.split_once(char::is_whitespace).map_or((line, ""), |(key, rest)| (key, rest.trim())) }; match key { - "TITLE" => title = line[5..].trim().trim_matches('"').to_string(), - "LUT_3D_SIZE" => size3 = it.next().and_then(|s| s.parse().ok()).ok_or_else(|| LutError("bad LUT_3D_SIZE".into()))?, - "LUT_1D_SIZE" => size1 = it.next().and_then(|s| s.parse().ok()).ok_or_else(|| LutError("bad LUT_1D_SIZE".into()))?, - "DOMAIN_MIN" => dmin = nums(it)?, - "DOMAIN_MAX" => dmax = nums(it)?, - "LUT_3D_INPUT_RANGE" | "LUT_1D_INPUT_RANGE" => { - let v: Vec = it.filter_map(|s| s.parse().ok()).collect(); - if v.len() == 2 { - dmin = [v[0]; 3]; - dmax = [v[1]; 3]; + "TITLE" => p.title = Some(rest.trim_matches('"').to_string()), + "LUT_3D_SIZE" => { + let value = if profile { rest } else { rest.split_whitespace().next().unwrap_or("") }; + p.size3 = value.parse().map_err(|_| LutError(format!("bad LUT_3D_SIZE `{value}`")))?; + let limit = profile_max.unwrap_or(MAX_SIZE); + if !(2..=limit).contains(&p.size3) { + return err(format!("LUT_3D_SIZE {} out of range", p.size3)); } } - k if k.starts_with(|c: char| c.is_ascii_digit() || c == '-' || c == '+' || c == '.') => { - let mut v = vec![k.parse::().map_err(|_| LutError(format!("bad number in `{line}`")))?]; - for value in it { - v.push(value.parse::().map_err(|_| LutError(format!("bad number in `{line}`")))?); + "LUT_1D_SIZE" => { + if profile { + return err("1D LUTs are not supported (only 3D .cube files)"); } - if v.len() != 3 { - return err(format!("bad data line `{line}`")); + p.size1 = rest.split_whitespace().next().and_then(|s| s.parse().ok()).ok_or_else(|| LutError("bad LUT_1D_SIZE".into()))?; + if p.size1 > row_limit { + return err("1D LUT exceeds row budget"); + } + } + "DOMAIN_MIN" => p.min = cube_triplet(rest).map_err(|_| LutError("bad DOMAIN_MIN".into()))?, + "DOMAIN_MAX" => p.max = cube_triplet(rest).map_err(|_| LutError("bad DOMAIN_MAX".into()))?, + "LUT_3D_INPUT_RANGE" | "LUT_1D_INPUT_RANGE" if !profile => { + let mut it = rest.split_whitespace(); + let min = it.next().and_then(|s| s.parse::().ok()); + let max = it.next().and_then(|s| s.parse::().ok()); + if let (Some(min), Some(max), None) = (min, max, it.next()) { + p.min = [min; 3]; + p.max = [max; 3]; + } + } + _ => { + let numeric = key.starts_with(|c: char| c.is_ascii_digit() || c == '-' || c == '+' || c == '.'); + let nonfinite = key.parse::().is_ok(); + if numeric || nonfinite || (profile && !key.starts_with(|c: char| c.is_ascii_alphabetic())) { + let row = cube_triplet(line)?; + if p.rows.len() >= row_limit { + return err("LUT exceeds row budget"); + } + // Explicit fallible growth prevents capacity growth past the configured budget. + if p.rows.len() == p.rows.capacity() { + let grow = (p.rows.capacity().max(1)).min(row_limit - p.rows.len()); + p.rows.try_reserve_exact(grow).map_err(|_| LutError("LUT allocation failed".into()))?; + } + p.rows.push(row); } - rows.extend(v); } - // Unknown keywords (e.g. Resolve's LUT_IN_VIDEO_RANGE) are ignored, as the spec allows. - _ => {} } } + Ok(p) +} + +/// Adobe / Resolve `.cube`. 1D LUTs (`LUT_1D_SIZE`) are expanded to a 33³ table. +pub fn parse_cube(text: &str) -> Result { + let p = read_cube(text, None)?; + let (title, size3, size1, dmin, dmax) = (p.title.unwrap_or_default(), p.size3, p.size1, p.min, p.max); + let rows: Vec = p.rows.into_iter().flatten().collect(); let domain_ok = (0..3).all(|k| dmin[k] < dmax[k] && dmin[k].is_finite() && dmax[k].is_finite()); if (size3 > 0 || size1 >= 2) && !domain_ok { return err(format!("DOMAIN_MIN {:?} must lie below DOMAIN_MAX {:?}", dmin, dmax)); From 753c676f7bc66e9665727712999a6962a81f5c02 Mon Sep 17 00:00:00 2001 From: "Dylan M. Taylor" Date: Sun, 15 Nov 2026 17:31:18 -0500 Subject: [PATCH 16/20] test(cms): cover bounded Cube profile compatibility and token rejection --- crates/cms/src/lutfile.rs | 49 +++++++++++++++++++++++++++++++++++++++ 1 file changed, 49 insertions(+) diff --git a/crates/cms/src/lutfile.rs b/crates/cms/src/lutfile.rs index 1dff864c1..e2e0a5980 100644 --- a/crates/cms/src/lutfile.rs +++ b/crates/cms/src/lutfile.rs @@ -1466,3 +1466,52 @@ LUT_3D_SIZE 2 assert_eq!(back.data, [0.0, 0.25, 1.0].repeat(8)); } } + +#[cfg(test)] +mod profile_cube_tests { + use super::*; + + fn cube(prefix: &str) -> String { + format!("{prefix}\nLUT_3D_SIZE 2\n{}", "0.25 0.5 0.75\n".repeat(8)) + } + + #[test] + fn profile_preserves_optional_literal_title_and_domains() { + let no_title = parse_cube_3d(&cube(""), 256).unwrap(); + assert_eq!(no_title.title, None); + let value = parse_cube_3d(&cube("TITLE \"look # literal\"\nDOMAIN_MIN 1 0 2\nDOMAIN_MAX 0 0 1"), 256).unwrap(); + assert_eq!(value.title.as_deref(), Some("look # literal")); + assert_eq!(value.domain_min, [1.0, 0.0, 2.0]); + assert_eq!(value.domain_max, [0.0, 0.0, 1.0]); + assert_eq!(value.data, vec![[0.25, 0.5, 0.75]; 8]); + assert_eq!(parse_cube_3d(&cube("TITLE \"\""), 256).unwrap().title.as_deref(), Some("")); + assert!(parse_cube(&cube("DOMAIN_MIN 1 0 2\nDOMAIN_MAX 0 0 1")).is_err()); + } + + #[test] + fn profile_limits_are_checked_without_large_allocations() { + assert!(parse_cube_3d("LUT_3D_SIZE 256", 256).unwrap_err().0.contains("16777216")); + assert!(parse_cube_3d("LUT_3D_SIZE 257", 256).unwrap_err().0.contains("out of range")); + assert!(parse_cube_3d("LUT_3D_SIZE 256", 129).unwrap_err().0.contains("out of range")); + assert!(parse_cube_3d(&cube(""), usize::MAX).is_err()); + assert!(parse_cube_3d(&cube(""), 1).is_err()); + assert!(parse_cube_3d(&format!("{}0 0 0\n", cube("")), 2).unwrap_err().0.contains("budget")); + assert!(parse_cube_3d("LUT_1D_SIZE 2\n0 0 0\n1 1 1", 256).is_err()); + } + + #[test] + fn every_numeric_token_must_be_valid_and_finite() { + for row in ["0 ignored 0 0", "0 0 0 ignored", "NaN 0 0", "inf 0 0", "0 0 -inf", "1e39 0 0", "0 0", "0 0 0 0"] { + assert!(parse_cube_3d(&format!("LUT_3D_SIZE 2\n{row}\n{}", "0 0 0\n".repeat(7)), 256).is_err(), "{row}"); + assert!(parse_cube_3d(&cube(&format!("DOMAIN_MIN {row}")), 256).is_err(), "{row}"); + } + } + + #[test] + fn standard_reader_keeps_quoted_comments_and_1d_expansion() { + assert_eq!(parse_cube(&cube("TITLE \"look # preserved\" # comment")).unwrap().title, "look # preserved"); + let value = parse_cube("LUT_1D_SIZE 2\n0 0 0\n1 1 1").unwrap(); + assert_eq!(value.size, 33); + assert!(parse_cube_3d(&cube("DOMAIN_MIN 0 0 0 # not a whole-line comment"), 256).is_err()); + } +} From f0fa27110d43519832df9a7a60962bef25daa273 Mon Sep 17 00:00:00 2001 From: Brandon Thomas Date: Sun, 15 Nov 2026 17:44:34 -0500 Subject: [PATCH 17/20] Preserve attributed print raster writer as shared extension source --- crates/pdf/src/write/raster_print.rs | 520 +++++++++++++++++++++++++++ 1 file changed, 520 insertions(+) create mode 100644 crates/pdf/src/write/raster_print.rs diff --git a/crates/pdf/src/write/raster_print.rs b/crates/pdf/src/write/raster_print.rs new file mode 100644 index 000000000..abce4061e --- /dev/null +++ b/crates/pdf/src/write/raster_print.rs @@ -0,0 +1,520 @@ +//! A small PDF writer for print output: pages of a given size in points, each with raster images +//! placed on it, tagged with their ICC profile (`ICCBased`) so the numbers mean the same colour in +//! every reader and RIP, Flate-compressed (lossless) at 8 or 16 bits per component, or carried as +//! JPEG (`DCTDecode`) where small files matter more than exact samples. Optionally an output intent: +//! the printing condition (the printer / paper profile) the file was made for, as PDF/X readers +//! expect it in the catalog. Profiles shared by several images are written once. +//! +//! The file is PDF 1.6 (16-bit images need 1.5): classic cross-reference table, no object streams, +//! deterministic (no dates or IDs), so the same pages always give the same bytes. +//! +//! Sources: ISO 32000-1:2008 (§7.3 objects, §7.5 file structure, §8.6.5.5 ICCBased colour spaces, +//! §8.9 images, §14.11.5 output intents). Own design. + +use std::collections::HashMap; +use std::sync::Arc; + +/// Pages per document: a print job beyond this is a mistake, not a book. +pub const MAX_PAGES: usize = 1000; +/// Images per page. +pub const MAX_IMAGES_PER_PAGE: usize = 256; +/// Largest page side in points (200 inches): more than any printer's roll. +pub const MAX_PAGE_POINTS: f64 = 14_400.0; +/// Largest image, in pixels per side. +pub const MAX_IMAGE_SIDE: u32 = 65_535; + +/// How an image's samples are stored. +#[derive(Clone, Debug, PartialEq)] +pub enum ImageData { + /// Interleaved samples, `bits` (8 or 16) per component, 16-bit values big-endian: compressed + /// losslessly in the file. + Raw { bits: u8, samples: Vec }, + /// A baseline JPEG stream, copied into the file as it is. + Jpeg(Vec), + /// Samples already compressed with zlib (see [`PdfImage::deflated`]), `bits` per component. + Flate { bits: u8, data: Vec }, +} + +/// A raster image and the colour space its samples are in. +#[derive(Clone, Debug, PartialEq)] +pub struct PdfImage { + pub width: u32, + pub height: u32, + /// Components per pixel: 1 (grey), 3 (RGB) or 4 (CMYK). + pub channels: u8, + pub data: ImageData, + /// The ICC profile the samples are in; without one the image is device colour. + pub icc: Option>>, +} + +impl PdfImage { + /// An image from interleaved 16-bit samples. + pub fn from_u16(width: u32, height: u32, channels: u8, samples: &[u16], icc: Option>>) -> PdfImage { + let bytes = samples.iter().flat_map(|v| v.to_be_bytes()).collect(); + PdfImage { width, height, channels, data: ImageData::Raw { bits: 16, samples: bytes }, icc } + } + + /// The same image with its samples compressed now (checked first), so a long print doesn't + /// hold every page's uncompressed samples until the file is written. + pub fn deflated(self) -> Result { + self.check()?; + let data = match self.data { + ImageData::Raw { bits, samples } => ImageData::Flate { bits, data: deflate(&samples) }, + other => other, + }; + Ok(PdfImage { data, ..self }) + } + + fn check(&self) -> Result<(), String> { + if self.width == 0 || self.height == 0 || self.width > MAX_IMAGE_SIDE || self.height > MAX_IMAGE_SIDE { + return Err(format!("an image of {} × {} pixels can't be written", self.width, self.height)); + } + if !matches!(self.channels, 1 | 3 | 4) { + return Err(format!("images have 1, 3 or 4 components, not {}", self.channels)); + } + match &self.data { + ImageData::Raw { bits, samples } => { + if !matches!(bits, 8 | 16) { + return Err(format!("images have 8 or 16 bits per component, not {bits}")); + } + let want = u64::from(self.width) * u64::from(self.height) * u64::from(self.channels) * u64::from(bits / 8); + if samples.len() as u64 != want { + return Err(format!("the image needs {want} bytes of samples, not {}", samples.len())); + } + } + ImageData::Jpeg(data) => { + if !data.starts_with(&[0xFF, 0xD8]) { + return Err("the image isn't a JPEG stream".into()); + } + } + ImageData::Flate { bits, data } => { + if !matches!(bits, 8 | 16) || data.is_empty() { + return Err("a compressed image has 8 or 16 bits per component and some data".into()); + } + } + } + Ok(()) + } +} + +/// Where an image goes on its page, in points from the page's bottom-left corner. +#[derive(Clone, Copy, Debug, PartialEq)] +pub struct Placement { + pub x: f64, + pub y: f64, + pub width: f64, + pub height: f64, +} + +/// One page: its size in points and the images on it, drawn in order (later ones on top). +#[derive(Clone, Debug, PartialEq)] +pub struct PdfPage { + pub width: f64, + pub height: f64, + pub images: Vec<(PdfImage, Placement)>, +} + +/// The printing condition the document is made for (PDF/X-style output intent). +#[derive(Clone, Debug, PartialEq)] +pub struct OutputIntent { + /// The printer / paper profile. + pub icc: Arc>, + /// Its components: 1, 3 or 4. + pub channels: u8, + /// A name for the condition (the profile's description). + pub condition: String, +} + +/// Document-level settings. +#[derive(Clone, Debug, Default, PartialEq)] +pub struct PdfInfo { + pub title: Option, + pub output_intent: Option, +} + +fn finite(v: f64) -> bool { + v.is_finite() +} + +/// A number as PDF writes it: up to four decimals, no exponent, no trailing zeros. +fn num(v: f64) -> String { + let s = format!("{:.4}", if finite(v) { v } else { 0.0 }); + let s = s.trim_end_matches('0').trim_end_matches('.'); + if s == "-0" || s.is_empty() { "0".into() } else { s.into() } +} + +/// A text string: plain ASCII as a literal string, anything else as UTF-16BE with a byte order mark. +fn text(s: &str) -> String { + if s.bytes().all(|b| (0x20..0x7f).contains(&b)) { + let mut out = String::with_capacity(s.len() + 2); + out.push('('); + for c in s.chars() { + if matches!(c, '(' | ')' | '\\') { + out.push('\\'); + } + out.push(c); + } + out.push(')'); + out + } else { + let mut out = String::from("'); + out + } +} + +struct Writer { + out: Vec, + offsets: Vec, +} + +impl Writer { + /// Reserve an object number (filled later by [`Writer::object`]). + fn reserve(&mut self) -> usize { + self.offsets.push(0); + self.offsets.len() + } + + fn object(&mut self, id: usize, dict: &str, stream: Option<&[u8]>) { + if let Some(slot) = self.offsets.get_mut(id.wrapping_sub(1)) { + *slot = self.out.len(); + } + self.out.extend_from_slice(format!("{id} 0 obj\n{dict}").as_bytes()); + if let Some(data) = stream { + self.out.extend_from_slice(b"\nstream\n"); + self.out.extend_from_slice(data); + self.out.extend_from_slice(b"\nendstream"); + } + self.out.extend_from_slice(b"\nendobj\n"); + } +} + +fn deflate(data: &[u8]) -> Vec { + miniz_oxide::deflate::compress_to_vec_zlib(data, 6) +} + +/// Write a document. Errors name the first page or image that can't be written. +pub fn write(pages: &[PdfPage], info: &PdfInfo) -> Result, String> { + if pages.is_empty() { + return Err("there are no pages to write".into()); + } + if pages.len() > MAX_PAGES { + return Err(format!("a document holds at most {MAX_PAGES} pages")); + } + for (n, page) in pages.iter().enumerate() { + let n = n + 1; + if !(finite(page.width) && finite(page.height)) + || page.width <= 0.0 + || page.height <= 0.0 + || page.width > MAX_PAGE_POINTS + || page.height > MAX_PAGE_POINTS + { + return Err(format!("page {n} has no usable size")); + } + if page.images.len() > MAX_IMAGES_PER_PAGE { + return Err(format!("page {n} has more than {MAX_IMAGES_PER_PAGE} images")); + } + for (image, at) in &page.images { + image.check().map_err(|e| format!("page {n}: {e}"))?; + if ![at.x, at.y, at.width, at.height].into_iter().all(finite) || at.width <= 0.0 || at.height <= 0.0 { + return Err(format!("page {n}: an image has no usable placement")); + } + } + } + if let Some(intent) = &info.output_intent + && !matches!(intent.channels, 1 | 3 | 4) + { + return Err("the output intent's profile has 1, 3 or 4 components".into()); + } + + let mut w = Writer { out: Vec::new(), offsets: Vec::new() }; + w.out.extend_from_slice(b"%PDF-1.6\n%\xE2\xE3\xCF\xD3\n"); + let catalog = w.reserve(); + let pages_id = w.reserve(); + let info_id = w.reserve(); + + // ICC profiles, each written once (keyed by content) with its component count + let mut profiles: HashMap, usize> = HashMap::new(); + let mut profile = |w: &mut Writer, icc: &Arc>, n: u8| -> usize { + if let Some(&id) = profiles.get(icc.as_slice()) { + return id; + } + let id = w.reserve(); + let data = deflate(icc); + let alternate = match n { + 1 => "/DeviceGray", + 4 => "/DeviceCMYK", + _ => "/DeviceRGB", + }; + w.object(id, &format!("<< /N {n} /Alternate {alternate} /Filter /FlateDecode /Length {} >>", data.len()), Some(&data)); + profiles.insert(icc.to_vec(), id); + id + }; + + let intent_id = info.output_intent.as_ref().map(|intent| profile(&mut w, &intent.icc, intent.channels)); + + let mut kids = Vec::with_capacity(pages.len()); + for page in pages { + let page_id = w.reserve(); + let mut xobjects = String::new(); + let mut content = String::new(); + for (k, (image, at)) in page.images.iter().enumerate() { + let space = match &image.icc { + Some(icc) => format!("[/ICCBased {} 0 R]", profile(&mut w, icc, image.channels)), + None => match image.channels { + 1 => "/DeviceGray".into(), + 4 => "/DeviceCMYK".into(), + _ => "/DeviceRGB".into(), + }, + }; + let id = w.reserve(); + let (bits, filter, data): (u8, &str, std::borrow::Cow<[u8]>) = match &image.data { + ImageData::Raw { bits, samples } => (*bits, "/FlateDecode", deflate(samples).into()), + ImageData::Flate { bits, data } => (*bits, "/FlateDecode", data.as_slice().into()), + ImageData::Jpeg(jpeg) => (8, "/DCTDecode", jpeg.as_slice().into()), + }; + w.object( + id, + &format!( + "<< /Type /XObject /Subtype /Image /Width {} /Height {} /ColorSpace {space} /BitsPerComponent {bits} /Filter {filter} /Length {} >>", + image.width, + image.height, + data.len() + ), + Some(&data), + ); + xobjects.push_str(&format!("/Im{k} {id} 0 R ")); + content.push_str(&format!("q {} 0 0 {} {} {} cm /Im{k} Do Q\n", num(at.width), num(at.height), num(at.x), num(at.y))); + } + let content_id = w.reserve(); + let stream = deflate(content.as_bytes()); + w.object(content_id, &format!("<< /Filter /FlateDecode /Length {} >>", stream.len()), Some(&stream)); + w.object( + page_id, + &format!( + "<< /Type /Page /Parent {pages_id} 0 R /MediaBox [0 0 {} {}] /Resources << /XObject << {xobjects}>> >> /Contents {content_id} 0 R >>", + num(page.width), + num(page.height) + ), + None, + ); + kids.push(format!("{page_id} 0 R")); + } + w.object(pages_id, &format!("<< /Type /Pages /Count {} /Kids [{}] >>", kids.len(), kids.join(" ")), None); + + let intents = match (&info.output_intent, intent_id) { + (Some(intent), Some(id)) => format!( + " /OutputIntents [<< /Type /OutputIntent /S /GTS_PDFX /OutputConditionIdentifier (Custom) /Info {} /DestOutputProfile {id} 0 R >>]", + text(&intent.condition) + ), + _ => String::new(), + }; + w.object(catalog, &format!("<< /Type /Catalog /Pages {pages_id} 0 R{intents} >>"), None); + let title = info.title.as_deref().map(|t| format!(" /Title {}", text(t))).unwrap_or_default(); + w.object(info_id, &format!("<< /Producer (LightCraft){title} >>"), None); + + let xref = w.out.len(); + let mut table = format!("xref\n0 {}\n0000000000 65535 f \n", w.offsets.len() + 1); + for offset in &w.offsets { + table.push_str(&format!("{offset:010} 00000 n \n")); + } + table.push_str(&format!("trailer\n<< /Size {} /Root {catalog} 0 R /Info {info_id} 0 R >>\nstartxref\n{xref}\n%%EOF\n", w.offsets.len() + 1)); + w.out.extend_from_slice(table.as_bytes()); + Ok(w.out) +} + +#[cfg(test)] +mod tests { + use super::*; + use lightcraft_codecs::NamedSpace; + + /// The body of object `id` (dictionary and raw stream bytes), found through the xref table. + fn object(pdf: &[u8], id: usize) -> (String, Vec) { + // byte offsets: the streams are binary, so search the bytes, not a lossy string + let start = pdf.windows(10).rposition(|w| w == b"startxref\n").unwrap() + 10; + let tail = std::str::from_utf8(&pdf[start..]).unwrap(); + let xref: usize = tail.lines().next().unwrap().trim().parse().unwrap(); + let table = std::str::from_utf8(&pdf[xref..]).unwrap(); + let line = table.lines().nth(2 + id).unwrap(); + let offset: usize = line[..10].parse().unwrap(); + let head = format!("{id} 0 obj\n"); + assert_eq!(&pdf[offset..offset + head.len()], head.as_bytes(), "xref entry for {id}"); + let rest = &pdf[offset + head.len()..]; + let end = rest.windows(7).position(|w| w == b"\nendobj").unwrap(); + let body = &rest[..end]; + match body.windows(8).position(|w| w == b"\nstream\n") { + Some(p) => { + let data = &body[p + 8..body.len() - "\nendstream".len()]; + (String::from_utf8_lossy(&body[..p]).into_owned(), data.to_vec()) + } + None => (String::from_utf8_lossy(body).into_owned(), vec![]), + } + } + + fn inflate(data: &[u8]) -> Vec { + miniz_oxide::inflate::decompress_to_vec_zlib(data).unwrap() + } + + /// Object ids of every image XObject, in file order. + fn images(pdf: &[u8]) -> Vec { + let s = String::from_utf8_lossy(pdf); + s.match_indices(" 0 obj\n<< /Type /XObject /Subtype /Image").map(|(i, _)| s[..i].rsplit('\n').next().unwrap().parse().unwrap()).collect() + } + + fn reference(dict: &str, key: &str) -> usize { + let at = dict.find(key).unwrap() + key.len(); + dict[at..].trim_start().split(' ').next().unwrap().parse().unwrap() + } + + fn srgb() -> Arc> { + Arc::new(lightcraft_codecs::icc::write_named(NamedSpace::Srgb)) + } + + fn page(images: Vec<(PdfImage, Placement)>) -> PdfPage { + PdfPage { width: 612.0, height: 792.0, images } + } + + fn full(w: f64, h: f64) -> Placement { + Placement { x: 0.0, y: 0.0, width: w, height: h } + } + + #[test] + fn a_16_bit_icc_image_round_trips() { + let samples: Vec = (0..2 * 3 * 3).map(|i| (i * 3851) as u16).collect(); + let img = PdfImage::from_u16(3, 2, 3, &samples, Some(srgb())); + let pdf = write(&[page(vec![(img, full(612.0, 792.0))])], &PdfInfo::default()).unwrap(); + assert!(pdf.starts_with(b"%PDF-1.6\n")); + assert!(pdf.ends_with(b"%%EOF\n")); + let ids = images(&pdf); + assert_eq!(ids.len(), 1); + let (dict, data) = object(&pdf, ids[0]); + assert!(dict.contains("/Width 3 /Height 2") && dict.contains("/BitsPerComponent 16") && dict.contains("/FlateDecode"), "{dict}"); + let back: Vec = inflate(&data).chunks(2).map(|b| u16::from_be_bytes([b[0], b[1]])).collect(); + assert_eq!(back, samples); + let (profile, icc) = object(&pdf, reference(&dict, "/ICCBased")); + assert!(profile.contains("/N 3") && profile.contains("/Alternate /DeviceRGB"), "{profile}"); + assert_eq!(inflate(&icc), *srgb()); + assert_eq!(lightcraft_codecs::icc::parse(&inflate(&icc)).and_then(|i| i.named), Some(NamedSpace::Srgb)); + } + + #[test] + fn pages_sizes_placements_and_a_shared_profile() { + let px = |v: u16| PdfImage::from_u16(1, 1, 3, &[v, v, v], Some(srgb())); + let at = Placement { x: 36.0, y: 72.5, width: 288.0, height: 216.0 }; + let pages = [page(vec![(px(1), at), (px(2), full(612.0, 792.0))]), PdfPage { width: 841.89, height: 595.28, images: vec![(px(3), at)] }]; + let pdf = write(&pages, &PdfInfo::default()).unwrap(); + let s = String::from_utf8_lossy(&pdf); + assert!(s.contains("/Type /Pages /Count 2"), "two pages"); + assert!(s.contains("/MediaBox [0 0 612 792]") && s.contains("/MediaBox [0 0 841.89 595.28]")); + // the sRGB profile is written once for all three images + assert_eq!(s.matches("/N 3 /Alternate").count(), 1); + let first: Vec<_> = images(&pdf).iter().map(|&id| reference(&object(&pdf, id).0, "/ICCBased")).collect(); + assert!(first.windows(2).all(|w| w[0] == w[1]), "{first:?}"); + // placements reach the content streams + let contents: Vec = s + .match_indices("/Contents ") + .map(|(i, _)| s[i + 10..].split(' ').next().unwrap().parse::().unwrap()) + .map(|id| String::from_utf8(inflate(&object(&pdf, id).1)).unwrap()) + .collect(); + assert_eq!(contents[0], "q 288 0 0 216 36 72.5 cm /Im0 Do Q\nq 612 0 0 792 0 0 cm /Im1 Do Q\n"); + assert_eq!(write(&pages, &PdfInfo::default()).unwrap(), pdf, "deterministic"); + } + + #[test] + fn output_intent_and_title() { + let cmyk_like = srgb(); + let info = PdfInfo { + title: Some("Garden – prints (2)".into()), + output_intent: Some(OutputIntent { icc: cmyk_like.clone(), channels: 3, condition: "Glossy (Photo) \\ test".into() }), + }; + let img = PdfImage::from_u16(1, 1, 3, &[0, 0, 0], Some(cmyk_like)); + let pdf = write(&[page(vec![(img, full(10.0, 10.0))])], &info).unwrap(); + let (catalog, _) = object(&pdf, 1); + assert!(catalog.contains("/OutputIntents [<< /Type /OutputIntent /S /GTS_PDFX"), "{catalog}"); + assert!(catalog.contains("/Info (Glossy \\(Photo\\) \\\\ test)"), "{catalog}"); + // the intent and the image share one profile object + let dest = reference(&catalog, "/DestOutputProfile"); + assert_eq!(reference(&object(&pdf, images(&pdf)[0]).0, "/ICCBased"), dest); + let (info_dict, _) = object(&pdf, 3); + assert!(info_dict.contains("/Title = ids.iter().map(|&id| object(&pdf, id)).collect(); + assert!(dicts[0].0.contains("/ColorSpace /DeviceCMYK") && dicts[0].0.contains("/BitsPerComponent 8")); + assert!(dicts[1].0.contains("/ColorSpace /DeviceGray")); + assert!(dicts[2].0.contains("/DCTDecode") && dicts[2].0.contains("/DeviceRGB")); + assert_eq!(dicts[2].1, jpeg, "JPEG streams are copied as they are"); + } + + #[test] + fn compressing_early_writes_the_same_file() { + let samples: Vec = (0..4 * 3 * 3).map(|i| (i * 977) as u16).collect(); + let raw = PdfImage::from_u16(4, 3, 3, &samples, Some(srgb())); + let early = raw.clone().deflated().unwrap(); + assert!(matches!(early.data, ImageData::Flate { bits: 16, .. })); + let at = full(100.0, 75.0); + let a = write(&[page(vec![(raw, at)])], &PdfInfo::default()).unwrap(); + let b = write(&[page(vec![(early, at)])], &PdfInfo::default()).unwrap(); + assert_eq!(a, b); + let mut short = PdfImage::from_u16(4, 3, 3, &samples, None); + short.width = 5; + assert!(short.deflated().is_err(), "checked before compressing"); + } + + #[test] + fn hostile_documents_are_errors() { + let ok = || PdfImage::from_u16(1, 1, 3, &[1, 2, 3], None); + let at = full(10.0, 10.0); + let bad = |pages: Vec| write(&pages, &PdfInfo::default()).is_err(); + assert!(bad(vec![])); + assert!(bad(vec![PdfPage { width: f64::NAN, height: 10.0, images: vec![] }])); + assert!(bad(vec![PdfPage { width: -1.0, height: 10.0, images: vec![] }])); + assert!(bad(vec![PdfPage { width: 1e9, height: 10.0, images: vec![] }])); + assert!(bad(vec![page(vec![(ok(), Placement { x: f64::INFINITY, ..at })])])); + assert!(bad(vec![page(vec![(ok(), Placement { width: 0.0, ..at })])])); + let mut short = ok(); + short.data = ImageData::Raw { bits: 16, samples: vec![0; 5] }; + assert!(bad(vec![page(vec![(short, at)])])); + let mut empty = ok(); + empty.width = 0; + assert!(bad(vec![page(vec![(empty, at)])])); + let mut odd = ok(); + odd.channels = 2; + assert!(bad(vec![page(vec![(odd, at)])])); + let mut not_jpeg = ok(); + not_jpeg.data = ImageData::Jpeg(vec![1, 2, 3]); + assert!(bad(vec![page(vec![(not_jpeg, at)])])); + let too_many = (0..MAX_PAGES + 1).map(|_| page(vec![])).collect(); + assert!(bad(too_many)); + let info = PdfInfo { title: None, output_intent: Some(OutputIntent { icc: srgb(), channels: 2, condition: String::new() }) }; + assert!(write(&[page(vec![])], &info).is_err()); + } + + #[test] + fn numbers_and_strings() { + assert_eq!(num(612.0), "612"); + assert_eq!(num(72.5), "72.5"); + assert_eq!(num(-0.00001), "0"); + assert_eq!(num(f64::NAN), "0"); + assert_eq!(num(1.0 / 3.0), "0.3333"); + assert_eq!(text("a(b)c\\"), "(a\\(b\\)c\\\\)"); + assert_eq!(text("é"), ""); + } +} From 2651dbeb2b13f621db23f65347380e7f26e316f0 Mon Sep 17 00:00:00 2001 From: "Dylan M. Taylor" Date: Sun, 15 Nov 2026 17:48:04 -0500 Subject: [PATCH 18/20] Expose bounded multi-image print PDF serialization with per-image ICC --- crates/pdf/src/write/raster.rs | 4 + crates/pdf/src/write/raster_print.rs | 252 +++++++++++++++++++++------ 2 files changed, 199 insertions(+), 57 deletions(-) diff --git a/crates/pdf/src/write/raster.rs b/crates/pdf/src/write/raster.rs index 1aa64ec3c..bcc32c23f 100644 --- a/crates/pdf/src/write/raster.rs +++ b/crates/pdf/src/write/raster.rs @@ -3,6 +3,10 @@ //! Callers map their page and color models to these neutral records. This module owns PDF object //! layout, image streams, ICC/Separation resources, page labels, marks, xref and trailer syntax. +/// Multi-image 8/16-bit print pages with per-image ICC resources and explicit byte budgets. +#[path = "raster_print.rs"] +pub mod print; + use std::collections::BTreeSet; use std::fmt::Write as _; use std::io::Write as _; diff --git a/crates/pdf/src/write/raster_print.rs b/crates/pdf/src/write/raster_print.rs index abce4061e..9866d73e1 100644 --- a/crates/pdf/src/write/raster_print.rs +++ b/crates/pdf/src/write/raster_print.rs @@ -8,10 +8,17 @@ //! The file is PDF 1.6 (16-bit images need 1.5): classic cross-reference table, no object streams, //! deterministic (no dates or IDs), so the same pages always give the same bytes. //! +//! Encoded JPEG and Flate payloads are copied, not decoded or verified against declared +//! dimensions. Callers own their encoding validity and ICC/profile-channel compatibility. +//! An output intent describes a printing condition; this API does not certify PDF/X. +//! Byte limits bound accounted input and emitted output, not peak process memory or +//! caller-owned image construction. This is not the streaming slideshow writer. +//! //! Sources: ISO 32000-1:2008 (§7.3 objects, §7.5 file structure, §8.6.5.5 ICCBased colour spaces, //! §8.9 images, §14.11.5 output intents). Own design. use std::collections::HashMap; +use std::io::Write as _; use std::sync::Arc; /// Pages per document: a print job beyond this is a mistake, not a book. @@ -23,6 +30,19 @@ pub const MAX_PAGE_POINTS: f64 = 14_400.0; /// Largest image, in pixels per side. pub const MAX_IMAGE_SIDE: u32 = 65_535; +/// Serialization budgets; input records already belong to the caller. These are not a +/// process-wide peak-memory guarantee. Repeated references count toward the input budget. +#[derive(Clone, Copy, Debug)] +pub struct Limits { + pub input_bytes: usize, + pub output_bytes: usize, +} +impl Default for Limits { + fn default() -> Self { + Self { input_bytes: 512 * 1024 * 1024, output_bytes: 512 * 1024 * 1024 } + } +} + /// How an image's samples are stored. #[derive(Clone, Debug, PartialEq)] pub enum ImageData { @@ -58,8 +78,11 @@ impl PdfImage { /// hold every page's uncompressed samples until the file is written. pub fn deflated(self) -> Result { self.check()?; + if matches!(&self.data, ImageData::Raw { samples, .. } if samples.len() > Limits::default().input_bytes) { + return Err("PDF input byte limit exceeded".into()); + } let data = match self.data { - ImageData::Raw { bits, samples } => ImageData::Flate { bits, data: deflate(&samples) }, + ImageData::Raw { bits, samples } => ImageData::Flate { bits, data: deflate(&samples, Limits::default().output_bytes)? }, other => other, }; Ok(PdfImage { data, ..self }) @@ -130,6 +153,8 @@ pub struct OutputIntent { pub struct PdfInfo { pub title: Option, pub output_intent: Option, + /// Application name, not a PDF/X conformance claim. + pub producer: String, } fn finite(v: f64) -> bool { @@ -157,47 +182,89 @@ fn text(s: &str) -> String { out.push(')'); out } else { - let mut out = String::from("'); - out + super::pdf_unicode_string(s) } } struct Writer { out: Vec, offsets: Vec, + limit: usize, } - impl Writer { - /// Reserve an object number (filled later by [`Writer::object`]). - fn reserve(&mut self) -> usize { - self.offsets.push(0); - self.offsets.len() + fn put(&mut self, bytes: &[u8]) -> Result<(), String> { + let end = self.out.len().checked_add(bytes.len()).ok_or("PDF output size overflow")?; + if end > self.limit { + return Err("PDF output byte limit exceeded".into()); + } + self.out.try_reserve(bytes.len()).map_err(|_| "PDF output allocation failed")?; + self.out.extend_from_slice(bytes); + Ok(()) } - - fn object(&mut self, id: usize, dict: &str, stream: Option<&[u8]>) { - if let Some(slot) = self.offsets.get_mut(id.wrapping_sub(1)) { - *slot = self.out.len(); - } - self.out.extend_from_slice(format!("{id} 0 obj\n{dict}").as_bytes()); + fn reserve(&mut self) -> Result { + self.offsets.try_reserve(1).map_err(|_| "PDF object allocation failed")?; + self.offsets.push(0); + Ok(self.offsets.len()) + } + fn object(&mut self, id: usize, dict: &str, stream: Option<&[u8]>) -> Result<(), String> { + let pos = self.out.len(); + *self.offsets.get_mut(id.checked_sub(1).ok_or("invalid PDF object")?).ok_or("invalid PDF object")? = pos; + self.put(format!("{id} 0 obj\n{dict}").as_bytes())?; if let Some(data) = stream { - self.out.extend_from_slice(b"\nstream\n"); - self.out.extend_from_slice(data); - self.out.extend_from_slice(b"\nendstream"); + self.put(b"\nstream\n")?; + self.put(data)?; + self.put(b"\nendstream")?; } - self.out.extend_from_slice(b"\nendobj\n"); + self.put(b"\nendobj\n") } } -fn deflate(data: &[u8]) -> Vec { - miniz_oxide::deflate::compress_to_vec_zlib(data, 6) +fn deflate(data: &[u8], max: usize) -> Result, String> { + struct Sink { + bytes: Vec, + max: usize, + } + impl std::io::Write for Sink { + fn write(&mut self, bytes: &[u8]) -> std::io::Result { + if bytes.len() > self.max.saturating_sub(self.bytes.len()) { + return Err(std::io::Error::other("PDF compressed byte limit exceeded")); + } + self.bytes.try_reserve(bytes.len()).map_err(std::io::Error::other)?; + self.bytes.extend_from_slice(bytes); + Ok(bytes.len()) + } + fn flush(&mut self) -> std::io::Result<()> { + Ok(()) + } + } + let mut encoder = flate2::write::ZlibEncoder::new(Sink { bytes: Vec::new(), max }, flate2::Compression::new(6)); + encoder.write_all(data).map_err(|e| e.to_string())?; + Ok(encoder.finish().map_err(|e| e.to_string())?.bytes) } /// Write a document. Errors name the first page or image that can't be written. pub fn write(pages: &[PdfPage], info: &PdfInfo) -> Result, String> { + write_with_limits(pages, info, Limits::default()) +} + +/// Write with explicit total input and serialized output budgets. ICC bytes are opaque: +/// callers validate their profile semantics; resources are keyed by bytes AND channels. +pub fn write_with_limits(pages: &[PdfPage], info: &PdfInfo, limits: Limits) -> Result, String> { + let mut input = 0usize; + let mut account = |n: usize| -> Result<(), String> { + input = input.checked_add(n).ok_or("PDF input size overflow")?; + if input > limits.input_bytes { + return Err("PDF input byte limit exceeded".into()); + } + Ok(()) + }; + account(info.producer.len())?; + account(info.title.as_ref().map_or(0, String::len))?; + if let Some(intent) = &info.output_intent { + account(intent.icc.len())?; + account(intent.condition.len())?; + } + if pages.is_empty() { return Err("there are no pages to write".into()); } @@ -218,6 +285,13 @@ pub fn write(pages: &[PdfPage], info: &PdfInfo) -> Result, String> { return Err(format!("page {n} has more than {MAX_IMAGES_PER_PAGE} images")); } for (image, at) in &page.images { + account(match &image.data { + ImageData::Raw { samples, .. } => samples.len(), + ImageData::Jpeg(data) | ImageData::Flate { data, .. } => data.len(), + })?; + if let Some(icc) = &image.icc { + account(icc.len())?; + } image.check().map_err(|e| format!("page {n}: {e}"))?; if ![at.x, at.y, at.width, at.height].into_iter().all(finite) || at.width <= 0.0 || at.height <= 0.0 { return Err(format!("page {n}: an image has no usable placement")); @@ -230,49 +304,50 @@ pub fn write(pages: &[PdfPage], info: &PdfInfo) -> Result, String> { return Err("the output intent's profile has 1, 3 or 4 components".into()); } - let mut w = Writer { out: Vec::new(), offsets: Vec::new() }; - w.out.extend_from_slice(b"%PDF-1.6\n%\xE2\xE3\xCF\xD3\n"); - let catalog = w.reserve(); - let pages_id = w.reserve(); - let info_id = w.reserve(); + let mut w = Writer { out: Vec::new(), offsets: Vec::new(), limit: limits.output_bytes }; + w.put(b"%PDF-1.6\n%\xE2\xE3\xCF\xD3\n")?; + let catalog = w.reserve()?; + let pages_id = w.reserve()?; + let info_id = w.reserve()?; // ICC profiles, each written once (keyed by content) with its component count - let mut profiles: HashMap, usize> = HashMap::new(); - let mut profile = |w: &mut Writer, icc: &Arc>, n: u8| -> usize { - if let Some(&id) = profiles.get(icc.as_slice()) { - return id; + let mut profiles: HashMap<(Arc>, u8), usize> = HashMap::new(); + let mut profile = |w: &mut Writer, icc: &Arc>, n: u8| -> Result { + if let Some(&id) = profiles.get(&(Arc::clone(icc), n)) { + return Ok(id); } - let id = w.reserve(); - let data = deflate(icc); + let id = w.reserve()?; + let data = deflate(icc, limits.output_bytes)?; let alternate = match n { 1 => "/DeviceGray", 4 => "/DeviceCMYK", _ => "/DeviceRGB", }; - w.object(id, &format!("<< /N {n} /Alternate {alternate} /Filter /FlateDecode /Length {} >>", data.len()), Some(&data)); - profiles.insert(icc.to_vec(), id); - id + w.object(id, &format!("<< /N {n} /Alternate {alternate} /Filter /FlateDecode /Length {} >>", data.len()), Some(&data))?; + profiles.try_reserve(1).map_err(|_| "PDF profile allocation failed")?; + profiles.insert((Arc::clone(icc), n), id); + Ok(id) }; - let intent_id = info.output_intent.as_ref().map(|intent| profile(&mut w, &intent.icc, intent.channels)); + let intent_id = info.output_intent.as_ref().map(|intent| profile(&mut w, &intent.icc, intent.channels)).transpose()?; let mut kids = Vec::with_capacity(pages.len()); for page in pages { - let page_id = w.reserve(); + let page_id = w.reserve()?; let mut xobjects = String::new(); let mut content = String::new(); for (k, (image, at)) in page.images.iter().enumerate() { let space = match &image.icc { - Some(icc) => format!("[/ICCBased {} 0 R]", profile(&mut w, icc, image.channels)), + Some(icc) => format!("[/ICCBased {} 0 R]", profile(&mut w, icc, image.channels)?), None => match image.channels { 1 => "/DeviceGray".into(), 4 => "/DeviceCMYK".into(), _ => "/DeviceRGB".into(), }, }; - let id = w.reserve(); + let id = w.reserve()?; let (bits, filter, data): (u8, &str, std::borrow::Cow<[u8]>) = match &image.data { - ImageData::Raw { bits, samples } => (*bits, "/FlateDecode", deflate(samples).into()), + ImageData::Raw { bits, samples } => (*bits, "/FlateDecode", deflate(samples, limits.output_bytes)?.into()), ImageData::Flate { bits, data } => (*bits, "/FlateDecode", data.as_slice().into()), ImageData::Jpeg(jpeg) => (8, "/DCTDecode", jpeg.as_slice().into()), }; @@ -285,13 +360,13 @@ pub fn write(pages: &[PdfPage], info: &PdfInfo) -> Result, String> { data.len() ), Some(&data), - ); + )?; xobjects.push_str(&format!("/Im{k} {id} 0 R ")); content.push_str(&format!("q {} 0 0 {} {} {} cm /Im{k} Do Q\n", num(at.width), num(at.height), num(at.x), num(at.y))); } - let content_id = w.reserve(); - let stream = deflate(content.as_bytes()); - w.object(content_id, &format!("<< /Filter /FlateDecode /Length {} >>", stream.len()), Some(&stream)); + let content_id = w.reserve()?; + let stream = deflate(content.as_bytes(), limits.output_bytes)?; + w.object(content_id, &format!("<< /Filter /FlateDecode /Length {} >>", stream.len()), Some(&stream))?; w.object( page_id, &format!( @@ -300,10 +375,10 @@ pub fn write(pages: &[PdfPage], info: &PdfInfo) -> Result, String> { num(page.height) ), None, - ); + )?; kids.push(format!("{page_id} 0 R")); } - w.object(pages_id, &format!("<< /Type /Pages /Count {} /Kids [{}] >>", kids.len(), kids.join(" ")), None); + w.object(pages_id, &format!("<< /Type /Pages /Count {} /Kids [{}] >>", kids.len(), kids.join(" ")), None)?; let intents = match (&info.output_intent, intent_id) { (Some(intent), Some(id)) => format!( @@ -312,9 +387,9 @@ pub fn write(pages: &[PdfPage], info: &PdfInfo) -> Result, String> { ), _ => String::new(), }; - w.object(catalog, &format!("<< /Type /Catalog /Pages {pages_id} 0 R{intents} >>"), None); + w.object(catalog, &format!("<< /Type /Catalog /Pages {pages_id} 0 R{intents} >>"), None)?; let title = info.title.as_deref().map(|t| format!(" /Title {}", text(t))).unwrap_or_default(); - w.object(info_id, &format!("<< /Producer (LightCraft){title} >>"), None); + w.object(info_id, &format!("<< /Producer {}{title} >>", text(&info.producer)), None)?; let xref = w.out.len(); let mut table = format!("xref\n0 {}\n0000000000 65535 f \n", w.offsets.len() + 1); @@ -322,14 +397,14 @@ pub fn write(pages: &[PdfPage], info: &PdfInfo) -> Result, String> { table.push_str(&format!("{offset:010} 00000 n \n")); } table.push_str(&format!("trailer\n<< /Size {} /Root {catalog} 0 R /Info {info_id} 0 R >>\nstartxref\n{xref}\n%%EOF\n", w.offsets.len() + 1)); - w.out.extend_from_slice(table.as_bytes()); + w.put(table.as_bytes())?; Ok(w.out) } #[cfg(test)] +#[allow(clippy::unwrap_used)] mod tests { use super::*; - use lightcraft_codecs::NamedSpace; /// The body of object `id` (dictionary and raw stream bytes), found through the xref table. fn object(pdf: &[u8], id: usize) -> (String, Vec) { @@ -355,7 +430,12 @@ mod tests { } fn inflate(data: &[u8]) -> Vec { - miniz_oxide::inflate::decompress_to_vec_zlib(data).unwrap() + { + use std::io::Read; + let mut out = Vec::new(); + flate2::read::ZlibDecoder::new(data).read_to_end(&mut out).unwrap(); + out + } } /// Object ids of every image XObject, in file order. @@ -370,7 +450,7 @@ mod tests { } fn srgb() -> Arc> { - Arc::new(lightcraft_codecs::icc::write_named(NamedSpace::Srgb)) + Arc::new(b"opaque RGB profile fixture".to_vec()) } fn page(images: Vec<(PdfImage, Placement)>) -> PdfPage { @@ -397,7 +477,6 @@ mod tests { let (profile, icc) = object(&pdf, reference(&dict, "/ICCBased")); assert!(profile.contains("/N 3") && profile.contains("/Alternate /DeviceRGB"), "{profile}"); assert_eq!(inflate(&icc), *srgb()); - assert_eq!(lightcraft_codecs::icc::parse(&inflate(&icc)).and_then(|i| i.named), Some(NamedSpace::Srgb)); } #[test] @@ -428,6 +507,7 @@ mod tests { let cmyk_like = srgb(); let info = PdfInfo { title: Some("Garden – prints (2)".into()), + producer: "Fixture".into(), output_intent: Some(OutputIntent { icc: cmyk_like.clone(), channels: 3, condition: "Glossy (Photo) \\ test".into() }), }; let img = PdfImage::from_u16(1, 1, 3, &[0, 0, 0], Some(cmyk_like)); @@ -503,7 +583,11 @@ mod tests { assert!(bad(vec![page(vec![(not_jpeg, at)])])); let too_many = (0..MAX_PAGES + 1).map(|_| page(vec![])).collect(); assert!(bad(too_many)); - let info = PdfInfo { title: None, output_intent: Some(OutputIntent { icc: srgb(), channels: 2, condition: String::new() }) }; + let info = PdfInfo { + producer: String::new(), + title: None, + output_intent: Some(OutputIntent { icc: srgb(), channels: 2, condition: String::new() }), + }; assert!(write(&[page(vec![])], &info).is_err()); } @@ -517,4 +601,58 @@ mod tests { assert_eq!(text("a(b)c\\"), "(a\\(b\\)c\\\\)"); assert_eq!(text("é"), ""); } + #[test] + fn distinct_profiles_and_component_counts_are_not_conflated() { + let common = Arc::new(b"opaque profile A".to_vec()); + let different = Arc::new(b"opaque profile B".to_vec()); + let image = |channels, profile| PdfImage { + width: 1, + height: 1, + channels, + data: ImageData::Raw { bits: 8, samples: vec![7; channels as usize] }, + icc: Some(profile), + }; + let pages = [page(vec![ + (image(3, common.clone()), full(1.0, 1.0)), + (image(4, common.clone()), full(1.0, 1.0)), + (image(3, different.clone()), full(1.0, 1.0)), + (image(3, common.clone()), full(1.0, 1.0)), + ])]; + let pdf = write(&pages, &PdfInfo::default()).unwrap(); + let refs: Vec<_> = images(&pdf).iter().map(|&id| reference(&object(&pdf, id).0, "/ICCBased")).collect(); + assert_ne!(refs[0], refs[1]); + assert_ne!(refs[0], refs[2]); + assert_eq!(refs[0], refs[3]); + for (i, channels, bytes) in [(0, 3, common.as_slice()), (1, 4, common.as_slice()), (2, 3, different.as_slice())] { + let (dict, data) = object(&pdf, refs[i]); + assert!(dict.contains(&format!("/N {channels}"))); + assert_eq!(inflate(&data), bytes); + } + } + + #[test] + fn gray_rgb_cmyk_raw_depths_preserve_samples() { + for channels in [1, 3, 4] { + for bits in [8, 16] { + let bytes: Vec = (0..channels * (bits / 8)).collect(); + let image = PdfImage { width: 1, height: 1, channels, data: ImageData::Raw { bits, samples: bytes.clone() }, icc: None }; + let pdf = write(&[page(vec![(image, full(1.0, 1.0))])], &PdfInfo::default()).unwrap(); + let (dict, data) = object(&pdf, images(&pdf)[0]); + assert!(dict.contains(&format!("/BitsPerComponent {bits}"))); + assert_eq!(inflate(&data), bytes); + } + } + } + + #[test] + fn input_and_output_budgets_fail_without_partial_success() { + let image = PdfImage::from_u16(1, 1, 3, &[1, 2, 3], None); + let pages = [page(vec![(image, full(1.0, 1.0))])]; + let info = PdfInfo::default(); + let expected = write(&pages, &info).unwrap(); + assert_eq!(write_with_limits(&pages, &info, Limits { input_bytes: 6, output_bytes: expected.len() }).unwrap(), expected); + assert!(write_with_limits(&pages, &info, Limits { input_bytes: 5, output_bytes: expected.len() }).is_err()); + assert!(write_with_limits(&pages, &info, Limits { input_bytes: 6, output_bytes: expected.len() - 1 }).is_err()); + assert!(write_with_limits(&pages, &info, Limits { input_bytes: 6, output_bytes: 0 }).is_err()); + } } From 17f402798e5aad206d0d29b88a09346f100f7176 Mon Sep 17 00:00:00 2001 From: FredApps <153173854+FredApps@users.noreply.github.com> Date: Mon, 16 Nov 2026 00:00:18 +0100 Subject: [PATCH 19/20] xml: compare strict expanded attribute names by namespace id, not a URI copy A declaration's URI was copied once per attribute that used it, so one long URI and many attributes multiplied memory. Give each binding an id (shared by equal URIs, found once at declaration) and key expanded names by that id. --- crates/xml/src/read.rs | 42 +++++++++++++++++++++++++++++++++++------- 1 file changed, 35 insertions(+), 7 deletions(-) diff --git a/crates/xml/src/read.rs b/crates/xml/src/read.rs index a5d9b58ee..aaea862fc 100644 --- a/crates/xml/src/read.rs +++ b/crates/xml/src/read.rs @@ -13,6 +13,10 @@ use crate::write::is_xml_char; /// The namespace the `xml` prefix is always bound to. pub const XML_NS: &str = "http://www.w3.org/XML/1998/namespace"; +/// [`Events::scope_ids`] id of the XML namespace, bound to `xml:` without a declaration. +const XML_NS_ID: usize = usize::MAX; +/// Expanded-name id of an attribute in no namespace (or with an undeclared prefix). +const NO_NS_ID: usize = usize::MAX - 1; /// What the reader refuses to exceed. #[derive(Clone, Copy, Debug, PartialEq, Eq)] @@ -470,6 +474,10 @@ pub struct Events<'o, 'i> { end_names: bool, open: Vec, scope: Vec<(String, String)>, + /// For each binding in `scope`, an id shared by every binding of an equal URI (the index of + /// the first such binding, or [`XML_NS_ID`]), with the URI's hash. Strict reads compare + /// expanded attribute names by this id instead of copying the URI for each attribute. + scope_ids: Vec<(u64, usize)>, /// The end of an empty element, due after its start. pending: bool, /// Whether a root element has started. @@ -501,7 +509,7 @@ impl<'o, 'i> Events<'o, 'i> { cfg.trim_text_start = false; cfg.trim_text_end = false; cfg.expand_empty_elements = false; - Events { reader, opts, end_names, open: Vec::new(), scope: Vec::new(), pending: false, rooted: false, span: 0..0 } + Events { reader, opts, end_names, open: Vec::new(), scope: Vec::new(), scope_ids: Vec::new(), pending: false, rooted: false, span: 0..0 } } /// The bytes of the input the last event was read from. Both events of an empty element @@ -615,6 +623,7 @@ impl<'o, 'i> Events<'o, 'i> { fn end(&mut self) -> Event { let (scope, name) = self.open.pop().map_or((0, String::new()), |o| (o.scope, o.name)); self.scope.truncate(scope); + self.scope_ids.truncate(scope); Event::End { name } } @@ -670,6 +679,7 @@ impl<'o, 'i> Events<'o, 'i> { if is_ns_decl(key) { let prefix = self.decode(key.get(6..).unwrap_or_default())?.into_owned(); let uri = self.attr_value(&a.value)?; + self.scope_ids.push(self.uri_id(&uri)); self.scope.push((prefix, uri)); } } @@ -681,6 +691,20 @@ impl<'o, 'i> Events<'o, 'i> { Ok(()) } + /// The id an expanded name uses for `uri`: one hash and at most one full comparison per + /// declaration, so a long URI costs its length once however many attributes use it. + fn uri_id(&self, uri: &str) -> (u64, usize) { + use std::hash::{Hash, Hasher}; + let mut h = std::collections::hash_map::DefaultHasher::new(); + uri.hash(&mut h); + let hash = h.finish(); + if uri == XML_NS { + return (hash, XML_NS_ID); + } + let first = self.scope_ids.iter().zip(&self.scope).position(|((h, _), (_, u))| *h == hash && u == uri); + (hash, first.map_or(self.scope.len(), |i| self.scope_ids.get(i).map_or(i, |(_, id)| *id))) + } + /// In a strict read: is the reader before or after the root element? fn outside_root(&self) -> bool { self.opts.strict && self.open.is_empty() @@ -730,13 +754,17 @@ impl<'o, 'i> Events<'o, 'i> { // Default namespaces do not apply to attributes. Keep undeclared prefixes // distinct so standalone fragments retain the reader's existing behaviour. let expanded = match raw_key.split_once(':') { - Some(("xml", local)) => (XML_NS, local), - Some((prefix, local)) => { - self.scope.iter().rev().find(|(p, _)| p == prefix).map_or(("", raw_key.as_ref()), |(_, uri)| (uri.as_str(), local)) - } - None => ("", raw_key.as_ref()), + Some(("xml", local)) => (XML_NS_ID, local), + Some((prefix, local)) => self + .scope + .iter() + .zip(&self.scope_ids) + .rev() + .find(|((p, _), _)| p == prefix) + .map_or((NO_NS_ID, raw_key.as_ref()), |(_, (_, id))| (*id, local)), + None => (NO_NS_ID, raw_key.as_ref()), }; - if !expanded_names.insert((expanded.0.to_string(), expanded.1.to_string())) { + if !expanded_names.insert((expanded.0, expanded.1.to_string())) { return self.syntax(format!("duplicate expanded attribute `{raw_key}`")); } } From 920c745f4217c04c5d881d2c721d537d26239408 Mon Sep 17 00:00:00 2001 From: FredApps <153173854+FredApps@users.noreply.github.com> Date: Mon, 16 Nov 2026 01:13:47 +0100 Subject: [PATCH 20/20] raw: initialize linearization_table in the dngwrite test image --- crates/raw/src/dngwrite.rs | 1 + 1 file changed, 1 insertion(+) diff --git a/crates/raw/src/dngwrite.rs b/crates/raw/src/dngwrite.rs index 3a715085a..384152e66 100644 --- a/crates/raw/src/dngwrite.rs +++ b/crates/raw/src/dngwrite.rs @@ -465,6 +465,7 @@ mod tests { color: crate::ColorData::default(), wb_multipliers: None, linearized: false, + linearization_table: None, pixel_aspect: 1.0, opcodes: OpcodeLists::default(), camera: Default::default(),