The cell pixmap was round(scale x step) pixels and repeated, but the cell
was drawn, and the paint transform mapped back, at the unrounded scale.
So the pixmap repeated a fraction of a pixel too far at every step (31 px
for a 30.56 px step at 110 dpi), and cells drifted further from their
place with every step from the pattern's origin.
Vendored hayro patch (25), `tiling_cell_pixels`: the cell is drawn at
round(scale x step) / step, so one step is exactly the pixmap's side, and
the paint transform uses that same scale. The side stays within 1 and the
existing 3000 px cap; a non-finite product keeps the scale and a bounded
side.
Tests: crates/render/tests/it/tiling_cell_placement.rs (fill, turned and
scaled matrix, stroke, all at 110 dpi far from the origin; the helper's
bounds).
Claude-Session: https://claude.ai/code/session_0175FWj4xFhMhSH3UpMU8DSC
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
crates/render/src/raster.rs: main's tests plus the PR's #2970 test after the knockout test
vendor/README.md: main's rows with the PR's rewording of hayro patch 18
crates/render/src/raster.rs: both new tests (#2975 stencil masks, #2971 knockout groups)
vendor/hayro/src/renderer.rs: set_paint returns main's PaintSetup; the PR's knockout erase early return wraps its clip in PaintSetup::Clip
vendor/README.md: main's rows plus the PR's hayro-interpret patch 39 and hayro patch numbered 22 (19 is #3296's), with tests
vendor/hayro/src/renderer.rs: main's PaintSetup::Clip match with the PR's fill paint (is_stroke false) and comment
vendor/README.md: main's row plus the PR's hayro-interpret patch and test, numbered 46 (35 is #3296's)
Review on #2974: the interpreter decided with a full-resolution decode
(`decoded_raster(None)`) before the renderer decoded at its draw size, a
second decode of every downscaled image under another cache key; and an
image a preview skips for its size still cleared what was under it.
`Device::knockout_image` now lets the renderer decide, with the same
eligibility and decode size as `draw_image` (shared `image_draw_size`), so
the draw reuses that decode and a skipped image clears nothing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gv2WwdgjmBz33GWkwgNn3Q
Review on #2974. In the erase pass `set_paint` returned no clip, so a
shading fill cleared its whole path while the redraw is clipped to the
shading's /BBox (ISO 32000-2 Table 77): red outside the box stayed erased.
It now returns the shading's clip there too. And an unmasked image cleared
its placement before decoding, so an image that doesn't decode (over
MAX_IMAGE_PIXELS, damaged data) erased the group's earlier artwork and drew
nothing; the clear now happens only when the image decodes, asked only in a
knockout group (Device::in_knockout_group). The stencil case in the test now
also checks that the stencil draws.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gv2WwdgjmBz33GWkwgNn3Q
Since #1557 the page background was painted inside the page
transparency group, as its initial backdrop, so that a Multiply
highlight under reading colours blends with the dark background. With
document colours that made every object blend with the white paper:
a green /BM /Screen fill over an empty part of the page came out
white, and Difference or Exclusion magenta. Only Normal, Multiply,
Darken and HardLight, for which blending with white gives back the
source colour, looked right. 0.8.0 was correct.
The page's objects composite with an initially transparent backdrop
and the page group is then put on the paper (ISO 32000-2, page group).
Vendored hayro patch (18) now paints the background inside the group
only under reading colours, where #1548 needs it, and outside it
otherwise, as before #1557.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RUNQSPVXx3dDBabRZHAnjz
After rebasing onto main: #3356's one-cell pattern renderer (merged) builds
a Renderer without the knockout fields this PR adds. Like the tiled cell's
renderer, a pattern cell is painted as its own artwork, so it starts with
no knockout group open.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gv2WwdgjmBz33GWkwgNn3Q
An image or a nested group is drawn into a layer of its own, which
covers the whole canvas (transparent where nothing was drawn), so it
can't be composited with Compose::Copy. What is under its shape is
now cleared first and the layer then composites over it:
- an unmasked image's shape is its placement rectangle: the
interpreter hands it to the device (`Device::knockout_shape`), which
clears it with a transparent Copy fill;
- a nested group's shape is the union of its elements' shapes: the
interpreter draws its content once in the device's erase mode
(`knockout_erase_wanted`, `begin_knockout_erase`,
`end_knockout_erase`), where every paint clears under its shape and
nothing is layered or drawn as an image, and then draws the group,
always in a layer of its own (drawn straight onto the knockout
group, its own elements knocked each other out).
A masked image's shape is its mask; it composites as before, which
matches MuPDF where the mask is opaque or transparent.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RUNQSPVXx3dDBabRZHAnjz
In a knockout group (/K true) each element composites with the group's
initial backdrop, not with the elements drawn before it (ISO 32000-2,
knockout groups). The renderer read /K (hayro-interpret patch 28) but
drew such a group like any other, so overlapping semi-transparent
shapes in it blended with each other.
hayro-interpret patch (39) tells the device a group is a knockout
group (`Device::set_knockout`, default no-op). hayro patch (19) then
draws that group's fills, strokes, glyph outlines and shadings with
Compose::Copy, which vello_cpu weights by coverage, so each replaces
what the group drew under its shape. Images and nested groups are
drawn into a group layer of their own and still composite normally:
copying a whole layer would erase everything outside it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RUNQSPVXx3dDBabRZHAnjz
* M2: a tiling pattern drawn once shows its cell at full scale (#3349)
PowerPoint writes a picture fill as a tiling pattern with /XStep 2147480100.
Fitting that step into the 3000 px cell pixmap shrank the picture below a
pixel. When what is painted can only show the first cell, the cell's /BBox is
drawn at full scale with a transparent border and padded instead of tiled.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gv2WwdgjmBz33GWkwgNn3Q
* M2: a tiling cell whose scale isn't finite keeps the old path (#3349)
Review follow-up on #3356: a /BBox 1e-40 wide passes the zero-area check and
gives an infinite cell scale, which would send non-finite geometry to the cell
renderer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gv2WwdgjmBz33GWkwgNn3Q
* M2: a stroke's miter counts in the one-cell check (#3349)
Review follow-up on #3356: shows_one_cell took the line width as a stroke's
reach; a miter join reaches half the width times the miter limit. The reach
formula patch (9) uses is now one function, stroke_reach, used by both.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gv2WwdgjmBz33GWkwgNn3Q
* M2: a skewed stroke's reach uses the transform's largest stretch (#3349)
Review on #3356: under a skew, a direction between the axes stretches
further than either axis, so the larger axis stretch could place a miter
tip short of the next cell when it reached it, and the one-cell path left
that cell unpainted. The reach now uses the spectral norm.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Gv2WwdgjmBz33GWkwgNn3Q
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Review follow-up. A pattern paint keeps the fill or stroke opacity it was
set with (`scn`), so resetting the graphics state's fill alpha did not stop
a pattern-filled stencil mask from being drawn at the opacity squared:
`Pattern::without_opacity` now drops it for the mask's paint (patch 35).
The renderer also painted a pattern-filled stencil mask with the stroke
paint, so an uncoloured tiling pattern drew in its stroke colour (black by
default); it now uses the fill paint.
The stencil test covers a colour fill, an uncoloured tiling pattern and a
shading pattern at ca 1, 0.5 and 0.25; each of the three changes, removed,
makes it fail. Rebased onto main; the vendored hayro-interpret patch is now
number 35 (36 was taken on main).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RUNQSPVXx3dDBabRZHAnjz
A BricsCAD plan page (3370 × 1524 pt, 251,000 single-segment strokes and 310
one-triangle type 4 shadings painted with `sh` at 50% alpha) took 2 s per
1024 px tile at every zoom, and 7.9 s for a fit-to-width render; on a busy
machine the 20 s watchdog gave the page up ("page 57 took longer than 20 s and
was skipped"). A 16 px tile of the same page took 65 ms, so neither parsing nor
the strokes were the cost: `sh` fills the whole canvas with the shading, and
`set_paint` sampled the mesh texture over the whole path, one hash lookup and
one texel per canvas pixel, 310 times per tile.
hayro-interpret patch (35): `EncodedShadingPattern::extent`, the device-space
pixels a mesh's triangles touched when it was encoded (`None` for shadings that
paint everywhere, and for a mesh with a `/Background`). hayro patch (19):
`set_paint` returns `PaintSetup::Nothing` for a mesh with no triangle on the
canvas (an empty extent, tested before it is intersected: `Rect::ZERO` sits at
the canvas origin), which the fill, stroke, rect and stencil callers skip, and
otherwise intersects the texture with the extent and pads instead of repeating it.
Measured with `page_timing` on the page: fit 7864 → 880 ms, a 1024 px tile
2030 → 85 ms at 100% and 2030 → 70 ms at 400% (the 16 px tile's 65 ms is the
page's re-interpretation, paid per tile), with byte-identical output hashes.
The three `mesh_extent_tests` pin what the texture must and must not cover.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* fix(render): keep annotations readable under reading colours (#1548)
Reading colours remapped annotation paints as line art: the yellow
highlight fill is luminance-light, so it was mapped toward the dark
background, and Multiply then darkened the remapped-white words under
it into invisibility.
Two changes in hayro:
- The /Annots loop in interpret_page is bracketed by a new
begin_annotations/end_annotations Device signal (patch 26), and the
renderer's reading-colours mapping leaves annotation paints alone.
Reading colours replace document colours, never the colours a
comment was given — the Acrobat model.
- The page background is painted inside the page transparency group
(patch 27), as the initial backdrop, so a Multiply highlight blends
with the background too. With the background outside the group the
band kept its own flat colour and the words vanished into it. White
pages are unaffected: multiplying with white is the identity.
Regression test reading_colors_keep_highlight_annotations_readable
covers both annotation paths (/AP and the no-/AP /QuadPoints
fallback) under WHITE_ON_BLACK; parity entry a11y.replace-page-colors
updated with the behaviour and evidence.
Co-Authored-By: GLM <noreply@z.ai>
* contributors: credit @edurbs as Eduardo Soares
The contributor's own entry, chosen by them (real and display name).
Co-Authored-By: GLM <noreply@z.ai>
* #1557: record the vendored hayro patches in vendor/README.md
Main already uses hayro-interpret patches (26)-(28), so the annotation hooks become
hayro-interpret (29) and the renderer's parts hayro (17) and (18).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: GLM <noreply@z.ai>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* M2: snap shrunk interpolated images to whole pixels so scans stay sharp
A scanned page (one 300 dpi image with /Interpolate true) was shrunk to a
ceil-sized raster and then sampled a second time at a fractional scale and
offset, which softened every edge next to Acrobat. An axis-aligned shrunk
interpolated image is now shrunk once to its rounded device rectangle and
drawn 1:1 at the rounded corner. Rotated, skewed, flipped, Type 3 and
non-interpolated images are unchanged.
On a synthetic scan page the mean error against an area-average reference
falls from 3.4 to 1.6 grey levels at 96 dpi (3.0 to 1.5 at 144 dpi).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
* vendor: list the hayro snap patch as (14) in the Patches and Tests columns; correct gaps.md patch counts
The patch description had been appended to the Tests column of one of the
two hayro rows. gaps.md now counts the distinct patches vendor/README.md
lists (26 hayro-interpret, 15 hayro, 3 hayro-jbig2, 17 hayro-syntax).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* test: give faux-redaction audit its own workdir (parallel-test flake)
* hayro: ColorReplacement type and luminance mapping (reading colours, #897)
* fmt: engine dates (upstream Greek localization miss)
* hayro: apply ColorReplacement to solid text and line-art paints (#897)
* render: ReadingColors view setting flows into RenderSettings (#897)
* engine: per-document reading colours state (#897)
* automation: view_reading_colors tool; page_render honours it (#897)
* ui: Reading colours preference (Replace Document Colors, #897)
* ui: seed reading colours on revision and signed-version opens (#897)
* parity: a11y.replace-page-colors shipped (#897)
* render: map stencil-mask paints through reading colours (#897)
An ImageMask stencil's solid paint was baked into the raster bytes raw,
bypassing the reading-colours hook every other solid fill goes through:
under white-on-black, CCITT scans, JBIG2 text and stencil logos drew
black ink on the black mapped background and vanished. Route it through
replaced(paint, false) — stencils are line art in disguise — and cover
it with an ImageMask fixture test sampling mapped ink on a mapped
background.
* render: print previews render with reading colours off (#897)
Print previews went through the document's render pool, whose config now
carries the reading colours: a white-on-black user saw black sheets in
the Print dialog while the actual print stayed uncoloured. The request
tag is the per-request channel that already crosses the render boundary,
so PRINT_TAG now lives in pdfcraft_render and the renderer honours it by
rendering previews with reading colours off — a preview's job is
fidelity to the printout. The main view stays recoloured. Covered on
both render paths (inline PageRenderer and pool worker).
* ui: seed reading colours on created-document opens (#897)
open_created (comment summaries) and open_created_bytes (images, text,
blank pages, clipboard) pushed new DocViews without seeding the reading
colours, unlike every other open path. A created document now starts
with the app-wide preference too; the seed-on-open test drives both
paths.
* docs: reading-colours flag echo and display-fidelity parity notes (#897)
The view_reading_colors description now says a flag that differs from
the mode's preset default makes the echoed mode read custom, and the
a11y.replace-page-colors parity notes record that print previews and
view snapshots render with the document's true colours: reading colours
apply to the reading view only.
* Merge follow-ups: reading colours beside preview renders and the new catalogs
- Renderer::new carries both color_replacement and the new preview flag
- our raster tests construct RenderRequest with preview: false
- the preference dialog grew past the window: scroll the combo into view
and give the fixture a valid xref so no repair banner covers the dialog
- checkboxes use the crate's square-box widget (clippy disallows egui's)
- Kazakh and Turkish catalogs gain the ten reading-colours strings
* Reading colours: record the hayro patch in vendor/README.md; add missing translations
vendor/README.md gains hayro patch (13): `RenderSettings::color_replacement` /
`ColorReplacement`, the `render_pdf` signature change and the removed
`[dev-dependencies.log]`, with its tests. Adds the ten reading-colours labels
to el, es, pl, te and zh-hant.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* 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>
* Fix#624: preserve hard edges for non-interpolated images
Select nearest-neighbour for Hayro's image pre-minification when /Interpolate is absent or false, retaining Catmull-Rom for true. Add a synthetic regression covering default, false, and true behaviour and record the vendored patch.
Co-authored-by: Codex <noreply@openai.com>
* M2: test non-interpolated minification with exact colours
The #624 regression used alternating black and white columns. Nearest 8→4 sampling reads only white columns from that pattern, so an image skipped entirely still passed, and the check only asked for near-black or near-white pixels.
Replace it with 2×2 colour blocks whose exact device output is asserted, with /Interpolate omitted equal to false, and /Interpolate true still blending. Add exact cases for gray, a non-integer 10→6 minification, and a matched soft mask. All pass on this head. On origin/main the same block image renders with blended mid-tones (for example 12eb12ff where pure green is expected), so these tests fail without the #624 change.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
* M2: plan nearest-neighbour pre-resize without filter limits
Non-interpolated images are pre-resized with nearest-neighbour, which copies source pixels and needs no filter tables, crossed intermediate or scratch. The preflight still applied the Catmull-Rom checks to it: a source wider than 2^20 pixels, or a filter plan over 128 MiB, was refused, and once such a source was wider than the backend's u16 limit the image was skipped.
resampling_byte_len and image_resampling_size take the image's interpolation flag. Nearest keeps the source and target buffer bounds and refuses any scratch it reports. Filtered images keep their limits unchanged, and the existing image_resampling_size tests pass true.
Measured: a 1.5M-pixel-wide RGB row drawn at 750×4 device pixels renders 0 non-white pixels on origin/main (the image is skipped) and all 3000 pixels with this commit. The 1.1M-pixel gray case is covered by wide_non_interpolated_sources_draw_after_shrinking.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
* M2: keep non-interpolated Type 3 glyph images nearest
Minified images inside Type 3 glyphs were drawn with ImageQuality::High, a bicubic sampler, whatever their /Interpolate. A non-interpolated image had already been pre-resized with nearest-neighbour, and bicubic sampling of that result blended it again. High is now used only for interpolated images, as elsewhere.
Before this commit the new test saw an interior glyph pixel of [28, 28, 27] (not a palette colour) for a 4.6 × 2.3 px glyph of a palette image; after it, every interior pixel is a palette colour. /Interpolate true still blends.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
* M2: nearest-sample masked minified images straight into premultiplied RGBA
A non-interpolated image with a matching soft mask used to build a full-resolution RGBA copy, nearest-resize it, then premultiply the target. For a minified image most of that copy was discarded.
nearest_premultiplied_rgba reads colour and mask from the same source pixel, selected with PicScale's nearest index (same fixed-point centre sampling as ResampleNearestPlan) and premultiplies only the target pixels. It falls back to the previous path if its target buffer cannot be allocated. Interpolated images, unmasked images and magnification are unchanged.
New test: a ten-column gray ramp with a soft mask drawn six pixels wide must pick the same columns for colour and mask.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
* M2: record the #624 image patch in vendor/README.md
The hayro row listed the #624 regression test but not the patch itself. Add patch item (10) covering the nearest pre-resize and its planning, the Type 3 bicubic restriction and the premultiplied soft-mask path, and list every regression test that covers them.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Claude Haiku 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
InDesign writes fades as luminosity soft masks whose group is one
extended axial `sh` over the whole page. hayro 0.7 sampled every such
shading pixel by pixel through the shading function, so a report page
with 40 of them took 10-15 s at 144 dpi.
Backport hayro 0.8.0's `gradient` module: an axial or radial shading
that a native gradient draws exactly becomes sRGB stops (within 0.01
per channel, a stop at every stitching bound), painted as a vello_cpu
gradient. Upstream's half-pixel shift is left out: it suits vello_cpu
0.3 but moves 0.0.8's gradients off pixel centres.
Page 4 of the reported file: 10.8 s -> 1.0 s.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A tile of a large page renders the whole page with an offset, and hayro
stroked and filled every path on it before vello discarded what fell
outside the tile. Vendored hayro patch (9), `may_paint_canvas`, skips a
fill or stroke whose box misses the canvas: the box of the path's points
and control points (a curve stays inside their hull), grown by the
stroke's reach (half the width times the larger of the miter limit and
sqrt 2), by 0.5 user units for vello's and kurbo's approximation
tolerances, and by 2 device pixels for antialiasing. Empty paths and
non-finite bounds or miter limits are always drawn.
Renders are byte-identical with and without skipping
(`skipping_paths_off_the_canvas_changes_no_pixel`; the hidden per-thread
`set_skip_offscreen_paths` exists for that test). On #623's sample at
scale 2, a 1024 px tile takes about 38% less CPU time; whole pages are
unchanged.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Check source, destination, convolution and filter-plan sizes before allocation. Preserve original image geometry when a resize exceeds the policy, while allowing valid wide sources to shrink.
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* Never crash: mesh shadings can't exhaust memory or run for minutes
`cargo xtask fuzz` (with the dash, JBIG2 and decompression fixes in
place) found a page whose Coons patch shading (type 6), a 147-byte
stream with 1-bit coordinates decoded onto +-40000, took the renderer
past 3 GB. hayro-interpret samples mesh shadings (types 4-7) pixel by
pixel over each triangle's device-space bounding box into a map, with
nothing but the u16 cast bounding it: up to 65536 x 65536 entries for
one triangle. A synthetic single patch passed 6 GB in six seconds; ten
thousand triangles that each cover half a 1000 x 1000 page ran for
minutes. Two rounds of review of earlier versions found more: every
patch was cut into 722 triangles (about 150 KB) before anything limited
them; a lattice mesh with `/VerticesPerRow 0` pushed empty rows forever,
and one with 4294967295 read the whole stream into one row; a mesh
painted a thousand times was triangulated a thousand times; and the
first caps cut content without a word.
Vendored hayro-interpret patch (18):
- `ShadingPattern::encode_within(area)`: the vendored hayro renderer
passes the area it will actually draw (the path's bounding box on the
canvas, which it already computes) plus a pixel for texels at a
fractional edge, and triangles are sampled only inside it, so the map
holds at most that area's pixels.
- `TriangleSampler`: at most 8 visits per pixel of that area plus 4 per
triangle in it. A triangle's bounding box is about twice its area and a
sub-pixel triangle's is at most 2 x 2 pixels, so tiling meshes and
small dense ones (a patch in a few pixels is 722 triangles) still
shade completely; overlapping ones stop.
- Coons and tensor-product patches are triangulated one at a time.
- Meshes hold at most `MAX_MESH_PATCHES` (2^16; a 200 x 200 gradient
mesh is 40 thousand) patches and `MAX_MESH_TRIANGLES` (2^20)
triangles, and a lattice needs two (ISO 32000-2 8.7.4.5.6) to 2^19
vertices per row.
- A page samples at most `MAX_PAGE_MESH_TRIANGLES` (2^25; the 200 x 200
mesh is 29 million) mesh triangles in all, reset per page like the
existing paint budget.
- Each of these cuts logs a warning.
`encode()` keeps its signature; with no area it samples within a fixed
2^24 visits.
Synthetic regression tests in render:
- `mesh_shadings_far_off_the_page_render`: a Coons patch and a triangle
mesh decoded onto +-40000 render and the rest of the page draws; a
patch on the page, the part of the huge one that covers it, and a
patch filling a 4 x 4 pt rectangle as a pattern are all shaded.
- `many_overlapping_mesh_triangles_finish`: 10,000 triangles each
covering half of a 1000 x 1000 page finish, and are drawn.
- `mesh_shadings_with_many_patches_or_no_rows_finish`: 40,000 patches
beside the page, and a lattice with `/VerticesPerRow 0`.
Red on main (out of memory, killed at 3.9 GB). Each part is pinned by a
mutation: no visit budget stalls past 60 s; no clip leaves the covering
patch unshaded; no per-triangle allowance leaves the small patch
unshaded; triangulating all patches first runs out of memory (killed at
5.8 GB); no rows check runs out of memory.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Never crash: decompression bombs can't exhaust the renderer's memory
The renderer's parser (vendored hayro-syntax) decoded Flate, LZW and
RunLength streams with no output limit. A 2 MB page whose content
stream inflates to 2 GB, or whose 8 x 8 image does, took `pdfcraft-cli
render` and `text` past 3 GB in under a second, and a few more MB of
file scale it further. On the desktop that is the first-page render of
an emailed PDF, and the allocation-failure abort takes every open
document with it. Our own layers already bound single streams (lopdf's
load-time decoding at 256 MiB, cos at 1 GiB), so `info` stays at a few
MB; the renderer path didn't. Random mutation almost never makes a
high-ratio stream, so `cargo xtask fuzz` doesn't find this; found by
hand while triaging the fuzz run.
Vendored hayro-syntax patch (6):
- `decode_limit` gives each filter an output limit, and Flate (both
flate2 and the fallback inflater), LZW and RunLength never write past
it (the buffer can still grow to under twice it):
`MAX_DECODED_STREAM`, 256 MiB, lopdf's figure; or, for the last filter
of an image whose size, bits per component and components are known,
twice its pixel bytes (a PNG predictor tag byte per byte at most),
capped at 1 GiB. Earlier filters in a chain produce compressed data
and get the general limit.
- `Stream::decoded_within(max)` for callers that budget.
- A page's `/Contents` array shares one `MAX_DECODED_STREAM` (naming one
bomb twenty times decoded 5 GB), reserving exactly what it appends.
Vendored hayro-interpret patch (18):
- The content a page holds at once (its contents, and the forms, tiling
patterns and Type 3 glyph procedures being painted) stays within
`MAX_PAGE_CONTENT`, 512 MiB, reset per page like the existing paint
budget. Forms nest fifty deep and each level kept its content while
the next ran: a form painting itself around a 256 MiB stream would
have held over 12 GB. Each level gives its bytes back when it is
dropped, so content painted again and again keeps drawing, and a
stream without filters (borrowed from the file) costs nothing.
- Type 0 sampled functions read every sample into a 4-byte value and
then a table entry: 256 MiB of 1-bit samples asked for over 8 GB. They
now take at most 2^21 table points and 2^23 values (the largest real
tables, DeviceN tint transforms such as 33^4 points x 3 outputs, are
1.2 million and 3.6 million), counted with checked arithmetic, and
read no further than `/Size` asks. `/Size 0` no longer underflows
building the default `/Encode`.
Three rounds of independent review found the aggregate cases, buffer
doubling past the limit, a chain/predictor shape the first image limit
would have truncated, the Type 0 amplifier, and a first, cumulative
budget that would have dropped legitimately repeated content, and a
first Type 0 cap low enough to refuse real tint transforms; all fixed
here.
Measured in a release build (render; OS peak commit):
- content stream inflating to 2 GB: 389 MB, 0.3 s
- 8 x 8 image inflating to 2 GB: 4 MB, 0.2 s
- /Contents naming it twenty times: 516 MB, 0.3 s
- a form painting itself around it: 644 MB, 0.4 s
Synthetic regression tests in render (flate2 is a new dev-dependency
there, already in the workspace):
- `decompression_bombs_stop_at_the_stream_limit`: 288 MiB bombs through
flate2, the fallback inflater (a stored block with a wrong length
check, which flate2 refuses and the fallback accepts), LZW and
RunLength each decode exactly 256 MiB, and at most 128 bytes as an
8 x 8 image.
- `page_contents_and_nested_forms_share_one_budget`: three /Contents
references decode one stream's worth; a self-painting form finishes.
- `content_painted_again_and_again_keeps_drawing`: a 2 MiB form painted
300 times draws the 300th time.
- `sampled_functions_with_billions_of_samples_are_refused`.
Ten mutations each turn them red: no limit at all; no limit in flate2,
the fallback, LZW, RunLength or for images; an unshared /Contents limit;
no page budget; no refund on drop; no Type 0 sample cap. (Taking a form
out of the budget no longer compiles: its content is the budget guard.)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* M2: normalize JPEG2000 samples without overflowing
Use bounded integer sample scaling for 1-32-bit JPX images and reject empty dimensions before chunking. Cover public decoding and continued page rendering with synthetic regressions.
Co-Authored-By: OpenAI Codex <codex@openai.com>
* Never crash: decompression cut at its limit is logged, not silent; vendor README renumbered after #575
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Streetblock <github@streetblock.de>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* M3: rasters reach egui without a copy on the UI thread
Every page, tile and thumbnail raster was copied pixel by pixel on the UI
thread (`ColorImage::from_rgba_premultiplied`), because egui's `Color32` is
4-byte aligned and a `Vec<u8>` isn't. That is 0.5 ms per 1024² tile, 5.4 ms
for a 33 MB page and 36 ms for a 197 MB page, against a 16.7 ms frame.
hayro, vendored patch 7: `render_size` and `render_into` render into a buffer
the caller allocates. `render` is `render_into` on a new pixmap, so its output
is unchanged. pdfcraft-render renders straight into `Pixels`, the same
premultiplied RGBA bytes kept in one 4-byte word per pixel. `Pixels`
dereferences to `[u8]`, so every reader of `RenderedPage::rgba` compiles
unchanged. The canvas reinterprets the words as `Color32`s and hands them to
egui as they are.
Pixels are identical: all 33 renders of #260's test file hash the same. The
UI-thread copy is gone, and rendering a 6182×8000 page takes 67 ms instead of
71–76 ms.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* M3: render_into leaves a buffer of the wrong length untouched instead of reaching vello's assert
* M3: regression test for render_into with a wrong-length buffer
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
`cargo xtask fuzz` found a 392-byte page whose ExtGState sets
`/D [[11] 0]` on a line that starts at x = 92234775807. Stroke expansion
keeps every dash, so it cut the line into billions of them and allocated
more than 5 GB before the run was stopped. Triage found the same class
in ordinary-looking geometry, and a second bug next to it:
- A tiny dash on an ordinary line (`[0.000001] 0 d` on 20 pt) passed
10 GB in under two seconds; so did `[1] 0 d` under a shrinking matrix,
and a long pattern on many short subpaths, because kurbo restarts the
pattern at every subpath. The render watchdog can skip a stuck page but
can't stop it allocating, so on a desktop this ends in an abort that
takes every open document with it.
- A dash array whose entries sum to less than zero (`[-1 -1] 0 d`) made
kurbo's search for the starting dash loop forever. ISO 32000-2
§8.4.3.6 requires nonnegative entries.
Vendored hayro patch (5, 6): `usable_dash_pattern` bounds the pieces a
pattern can cut the path into, in user space where kurbo dashes:
(control-polygon length / period + restarts) x entries per period, where
restarts are kurbo's: every MoveTo and ClosePath. Over 250,000 the stroke
is drawn solid, as it is for a pattern with a negative or non-finite
entry or a period that isn't positive. Measured in release builds: a
250,000-piece stroke takes about 16 ms and 75 MB at its peak (a million
took 80 ms and 290 MB). A non-finite phase becomes 0.
Every case now renders in about 0.2 s and a few MB. Synthetic regression
tests in render. Both were red before the fix (the first killed at
2.7 GB, the second stalled past 20 s), and four mutations each turn them
red: the first estimate (one piece per subpath) with a 100k cap runs out
of memory, a 100k cap draws a fine legitimate dash solid, dropping the
negative-entry check draws `[-1 3]` dashed, and dropping it together
with the period check hangs. (The period check alone is defence in
depth: hayro-interpret turns 0 entries into 0.01, so a pattern without
negative entries always has a positive period.)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The app, crates, binaries, bundle id, packaging, icons, screenshots and docs are now
PdfCraft / pdfcraft (repo storytold/pdfcraft). On first launch the desktop app moves the
PrintCraft settings and crash-recovery folders to the new name, and both apps read settings
saved under the old storage key, so upgrades keep recent files, preferences and unsaved work.
The signing test fixtures keep their PrintCraft subject names (they are binary).
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The second full nightly fuzz run on main found two more:
- hayro-syntax computed a Flate/LZW predictor row as columns × bits in
wrapping arithmetic: /Columns 9223372036854775807 asked for 2^61 bytes and
aborted on any machine. Row widths are now saturating usize (colors × bits
also overflowed u8), a zero width is refused (it divided by zero), and a
row longer than the data is not allocated (empty output, as before).
- hayro stroked /LW 9223372036854775807 as given, and stroke expansion
allocated 10 GB. Widths are clamped to 4 × (canvas width + height) device
pixels, which covers the same pixels.
Both findings now run in 8 MB. Synthetic regression test in render.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fuzzed inputs drove one check-one process to 2-6 GB (the 16 GB CI runner
running two of them was killed every night). Three causes, each with a test:
- lopdf (bootstrap inspector) sized PNG predictor rows from /Columns before
reading data: /Columns 4294967295 allocated two 4 GiB rows. Vendored lopdf
0.45.0 refuses rows longer than the data (and uses saturating row widths);
its load-time stream decoding also gets a 256 MiB limit (it had none).
- hayro sized a tiling pattern's cell pixmap from /XStep x scale, bounded only
by u16: a step far beyond /BBox allocated 2 GB. The vendored patch keeps
the cell within 3000 px a side, lowering the resolution instead.
- cos decoded cross-reference streams up to the general 1 GiB limit; they now
stop at what 8,388,607 objects' rows can hold (~200 MB at most), and the
section is reconstructed (logged) instead.
On the nightly's seed (41) the worst captured input fell from 6.2 GB to 6 MB,
and a 300 s fuzz run had no process above 300 MB (before: many at 2-5 GB).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
