Files
pdfcraft/crates/engine
d415749556 Heavy pages: cull off-tile images, cache decodes, preview backdrops, calmer queues (#1117)
* Heavy pages: cull off-tile images, cache decodes, preview backdrops, calmer queues

A large CAD sheet (page 24 of the benchmark file: ~5050 pt wide, 2104
images, 9 MB of content) cost a full page interpret per tile: 4.3 s for a
1024 px tile at 1000%, 10.5 s for the whole page, and a storm of renders on
every pan frame. Tiles now cost 0.6 s, the sheet is covered by a handful of
large tiles, and gestures no longer queue work faster than workers finish it.

- Vendored hayro patch (10): skip images whose stream-size box misses the
  canvas before decoding them, so a tile no longer decodes, resamples and
  rasterizes every image on its page. Pixel-identical; the skip regression
  test now covers images too.
- Vendored hayro-interpret patch (23): decoded image pixels shared across
  renders through the interpreter cache, keyed by stream and decode target
  size within a 64 MiB budget.
- Preview renders (hayro patch 11, RenderRequest::preview through to the
  device): backdrops, thumbnails, grid cells and signature-drag backgrounds
  skip images under 16 device px. Tiles, whole pages, print and export always
  draw everything. Previews may differ by those pixels; finals don't.
- Viewer: zoom/pan settle holds apply only to pages whose raster proved slow
  (slow_pages, 500 ms), so fast pages render with no added latency; tile sides
  grow with zoom and page size (giant sheets need ~8 tiles, not ~28), queued
  center-first and keyed by exact coordinates; Edit-text outlines capped.
- vendor/README.md records the new patches with their tests.

Co-Authored-By: Muse Spark <noreply@meta.com>

* Decoded-image cache: key every decode input, survive a poisoned lock

The key was the dictionary's bytes and the target size, so two images with byte-identical
dictionaries (uncompressed, one size) shared pixels, and an image drawn under a transfer
function reused pixels decoded without it. DecodedImageKey now holds the dictionary's place
in the document (its data follows it), its bytes and object, mask/stencil/raster output,
interpolation, the target size and preview or final quality; images under a transfer
function and inline images (named colour spaces, dictionaries in content streams) are not
cached, and an image over the budget is never kept. A poisoned lock decodes uncached
instead of panicking. Soft masks: the per-renderer cache can't carry a preview mask into a
final render (documented). Regression tests for identical dictionaries, preview then final
through one cache, and transfer functions.

---------

Co-authored-by: Muse Spark <noreply@meta.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
2026-10-11 17:42:48 -04:00
..

pdfcraft-engine

The façade every frontend talks to. It holds the open documents, their edit history and the tool catalogue; frontends never reach past it into the parsing, rendering or editing crates.

  • Layer: L6. It depends on most of L1–L4 and on nothing above itself. No UI toolkit — egui, winit, eframe and rfd are forbidden below L7, and xtask layers enforces it.
  • The crate denies unwrap/expect/panic!/todo!. Entry points that run untrusted-document code keep a last-resort catch_unwind guard so an escaped panic becomes an error and the user keeps their document.

Two frontends exist on top of it today: pdfcraft-ui-egui (desktop and web) and pdfcraft-automation (headless tools, the CLI and MCP). Anything one can do, the other must be able to do — see AGENTS.md §3.

The editing model

Each open document holds a pdfcraft_cos::Document: the object graph plus a copy-on-write overlay.

  1. An edit runs on a clone. On success the previous state is pushed onto the undo stack. Clones share all unchanged data, so a snapshot is cheap — this is what makes 100 levels of undo affordable on a 200 MB file.
  2. After every edit the working file is produced by an incremental write: the original bytes plus one appended revision.
  3. The view is refreshed from those bytes, so what you see is exactly what Save will write.
  4. Saving rebases onto the written bytes, so the next save appends only new edits.

Consequences worth preserving: the original bytes of an opened file are never rewritten in place; a full rewrite happens only where it must (redaction, PDF/A conversion, print imposition, a changed password); and unknown data survives a round trip because nothing is reconstructed that wasn't understood.

API

let mut s = Session::new();
let id = s.open("report.pdf", Some(path), bytes, None)?;   // -> DocId

s.apply(id, Edit::RotatePages { pages: vec![0], degrees: 90 })?;
let label = s.undo(id)?;                                    // "Rotate pages"
s.redo(id)?;

let doc = s.get(id).ok_or("gone")?;                         // &Document: queries, never edits
doc.can_undo(); doc.editable(); doc.allows_printing();
doc.text_blocks(page); doc.page_boxes(page); doc.security_summary();

let bytes = s.save_bytes(id)?;        // incremental: keeps signatures over earlier revisions valid
let bytes = s.save_full_bytes(id)?;   // full rewrite, garbage-collected
s.mark_saved(id, bytes.clone(), path)?;

Edit is one enum of 75 variants — every undoable change in the product, from RotatePages to AddAnnotation, SetFieldValue, ApplyRedactions and AddWatermark. Adding a feature means adding a variant and handling it in apply, which is what keeps undo, the dirty flag, autosave and the automation tools consistent for free.

Session also owns the operations that produce a new document rather than editing one: create_blank, create_from_images, create_from_text, combine, extract, split, split_by_size, reduced_bytes, optimized_bytes, print_pdf, export_data, and the signing family (sign, sign_with_timestamp, timestamp_document, embed_ltv).

Modules

Module What it holds
commands The command registry: id, label, menu, shortcut, and a Needs precondition encoding document security. Menus, the ⌘K palette, shortcuts and the control channel all read it, so every surface agrees on what exists and when it is enabled
catalog Acrobat's "All tools" information architecture as data — groups, sections, icons, hues, and an Availability of Ready, Planned(milestone) or Provider. The UI renders what is not built yet, honestly
actions PDF action dispatch (go-to, URI, named, set-layer-state, JavaScript)
links Link creation, properties, and finding URLs in page text
js The bridge to pdfcraft-js. forms never depends on js directly; the engine injects it, so a build without JavaScript is possible
xfa, ocr, compare, export Façades over the L3/L4 crates of the same names

Not done here

  • Rendering is still borrowed. pdfcraft-render wraps hayro and inspects through lopdf (ADR-0004). The engine only sees inspect() -> DocInfo, RenderPool and text::PageText, and that API stays frozen while M2 replaces what is behind it.
  • Performance budgets (misc.performance-budgets) are not set or enforced.
  • Some shipped features are still reachable only from the UI, not through an automation tool; cargo xtask parity lists them; closing that gap is tracked in the execution plan (D4).