PDF export rasterises every slide as a backdrop and a content layer and
hands both to the writer. Profiling (gdb stack sampling) showed most of
the main thread outside rendering going to per-pixel work on those
rasters: compositing the backdrop onto white (not vectorised), the
writer hashing every backdrop to deduplicate it, and the straight-alpha
conversion scanning every content pixel.
- Composite the backdrop with saturating adds and skip 16-pixel blocks
that are already opaque; skip blocks with no partly transparent pixel
in the straight-alpha conversion. Same formulas, same bytes.
- Keep the last opaque backdrops (up to 32 MiB of pixels) and hand the
writer the same image again when a slide renders an identical
backdrop. The writer stored equal backdrops once already; now it also
skips hashing and compressing the copies. Output is byte-identical.
deckcraft-cli bench, export .pdf step, 15 interleaved runs each on a
16-thread Ryzen (A/A noise +0.9 %):
- sample x4 (32 slides): median 531 -> 322 ms (-39 %), min 529 -> 317 ms
- same deck, every slide a different background: 557 -> 441 ms (-21 %)
- sample (8 slides): 149 -> 103 ms (-31 %)
Exported PDFs are byte-identical.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`deckcraft-cli bench` capped the "export .pdf" step at three runs while
its help promises the median of N runs for every step. Use --runs for
it too, so a PDF export benchmark can take enough samples to beat noise.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Real msiexec runs (per-user and all-users, silent and wizard) found two defects
that the metadata fixtures cannot see:
- "Only for me" published ALLUSERS=2, so FindRelatedProducts treated the run as
per-machine and skipped the existing per-user copy. An interactive upgrade
left two registered products and uninstall left files behind. The page now
leaves ALLUSERS unset for a per-user choice.
- Windows Installer resolves Program Files once, for the context it started in,
so choosing "Everyone" in the wizard installed into the user profile with a
machine-wide registration. The wizard now sets the new-install folder itself
after the choice (Program Files, Program Files (x86) for 32-bit, or the
per-user Programs folder). Silent installs are unchanged.
check-install-policy.ps1 now fails if either regresses.
A harness runs frames until no repaint is asked for, so the preview had
already landed when its first assertion ran. One raw frame now shows the
preview pending and one render started, and the paging test pins a pending
job and checks a page turn starts no other render.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Turning the page with Presenter View on re-rendered the next-slide preview
synchronously, single-threaded, on the UI thread. The preview now renders
on a worker thread and arrives on a later frame; the web build (no
threads), or a worker that dies, renders it inline. At most one render runs:
paging faster than it asks for the newest preview once it is done, and a
finished one no longer wanted is dropped. Job and texture are keyed by the
presentation as well as the slide, so a preview rendered before an edit is
rendered again. While a preview is on its way the box stays empty rather
than saying "End of slide show".
Rebased onto the show module split: show/presenter.rs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An ellipse at the EMU limit, compounding group scales or custom-geometry
guides could reach vello's flattener with an outline billions of pixels
long; it sizes its buffer by that length and the allocation aborts the
process. geometry::drawable() rejects a shape with a non-finite path
point, then measures the outline's control box (which always holds the
curve, unlike the tight bounding box when its extremum overflows) unioned
with the frame; past 1e7 device pixels the shape is skipped with a warning.
Rebased onto the render module split: the guard now lives in geometry.rs
and renderer.rs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A full-screen show hides a still pointer after 2.5 s, which makes decks
with clickable, branching choices hard to use. Set Up Slide Show now has
Keep mouse cursor visible (`show.keepCursor {on?}`): the show then keeps
the normal pointer, and the pointing hand over links and actions, while
links still follow. It is a user preference kept in ui.json, since PPTX
has no place for it; off by default, Reading View unchanged.
Co-Authored-By: Claude <noreply@anthropic.com>
Found while testing crates/engine/src/cmd/file.rs:390-400 (parse_slide_range)
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Documents name non-RIBBI weights by their Windows family ("Montserrat ExtraBold" / Regular).
The bundled Montserrat is variable, so its instances had no Windows names; the installed static
files that do carry them were skipped as already present, and the lookup fell back to Inter
Regular without reporting the font missing. Each named instance other than Regular, Italic,
Bold and Bold Italic now also answers to "<family> <weight>" with a Regular or Italic style,
as Windows lists them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Scale is measured from the size the tab first showed, like the Size and Position
dialog, and runs shape.sizePosition (honouring Lock aspect ratio). From switches
Horizontal/Vertical between the top-left corner and the centre. The lock checkbox
used a fresh id every frame so it never held; it now keeps its state.
A buFont named like Noto Sans Symbols was treated as a legacy symbol font, so its
bullet characters mapped to a generic dot. Only Wingdings and Symbol map by code now;
the bullet's own face is chosen by the usual cascade instead of a fixed Inter fallback.
Every effect in one column is taller than a 1440p screen, and the popup
did not scroll, so the last entries could not be reached. Cap the menu at
the window height minus the title bar and ribbon, and scroll inside it.
Co-Authored-By: Grok 4.7 <noreply@x.ai>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PowerPoint writes autoRev on p:cBhvr/p:cTn, not the effect node, so Grow/Shrink with auto-reverse grew and held. Color Pulse writes it there as part of the effect and is left out. Tests patch autoRev onto the behaviour node of a saved deck and read it back. Not checked against real PowerPoint: no PowerPoint COM here.
aiongg
NetterSilverandClaude Opus 5.5
Jeremy
NaseqandClaude Opus 5.5
FaabyyandClaude
neoandClaude Opus 5.5
Dominic Anaeto
Harrison Booth
skylar taylor-barrick
aiongg