- Cryptomatte planes: a plane a render holds is not offered for its gone channels, and the entry
limit holds while renders hold every plane.
- Footage frames and whole stills held elsewhere are not offered to the memory accountant.
- Purge (clear frames) makes decodes in flight stale: their frames are not kept or joined.
- Plane store: a size refusal of a merged decode is not handed to a request that fits alone;
slimming a shared set and copies for the disk cache take memory fallibly, and planes the disk
cache can't take are not copied.
- Disk cache: the plane encoder's scratch is taken fallibly; no memory skips the entry.
- Region of interest: rows follow the projected transform of a 2D layer under a 3D parent, and
a frame of another size than its footage is read whole.
- The box filter's work bound is checked without a wall clock.
The session's figure set during the reading was replaced by the older reading's
answer; it is now published only while nothing set a limit meanwhile. A reading
that gave no answer now makes earlier refusals for size out of date, so the
second reading is asked for by the next request.
- The memory accountant is offered only layer buffers and preview frames nobody
else holds, so what it counts as freed is freed; the frame cache's recycled
buffers count and go first.
- A Cryptomatte plane over the matte cache's own limit is kept only where the
common budget trims it.
- Disk cache entries read back take their buffers fallibly: memory the system
can't give is a miss.
- The Layer panel, sampleImage and effect inputs read whole footage frames under
a viewer region of interest.
- A 16K Cryptomatte source at half resolution is box-filtered, not looked up
nearest-neighbour.
- The System Compatibility Report shows the cache budget in force.
- A Project panel thumbnail render never starts under a key held back from an
earlier state.
- Convert Expression to Keyframes makes no keys on a property that can't be
animated.
The limit of one read is decided here (the 8 GiB floor, or the memory
budget when that is more) and handed to craft-exr as the read's max_bytes,
so the planes it reserves and the largest block it decompresses follow it.
craft-exr reserves the planes fallibly, so the reader's own fallible plane
allocation is not needed: memory the system can't give is a read that
failed just now (read_error), a plane over the limit a refusal for size.
Conflicts: Cargo manifests (craft-libs pins from main, plus memory-stats, exr and
half for effectcraft-media), exr_channels.rs (the plane store's version; main's
craft-exr channel decode is replaced by the plane store until craft-exr reads
from a seekable source), pool.rs (frame_px takes the rows of #3271 and the
forget generation of #2042), session open/forget (main's forget_files plus the
reload counter), Project panel thumbnails (main's Thumb types, keyed by content).
* Import: VectorCraft documents import through the PDF they carry
File > Import reads a .vectorcraft file through the PDF that VectorCraft
saves in it with Create PDF-Compatible File, one page per artboard, as
footage or as a composition with one layer per PDF layer. The PDF is read
with craft-doc from craft-libs (plain or compressed files, the PDF inline
or after the JSON). A PDF after the JSON may be as large as a PDF file
EffectCraft imports (512 MiB). A document saved without that option gives
an import error that names it.
craft-doc, craft-vectorcraft-format (tests only) and craft-ecproj-format
are pinned to craft-libs fa1a6574; craft-ecproj-format is unchanged
between ca1b5e2e and that commit.
* Import: show file errors returned by the import dialog
* Keep dropped-file import errors alongside local drop errors
* Import: preserve every file error across background jobs and UI notices
* Import: queue pending errors with bounded notification details
* Import: a damaged VectorCraft preview no longer stops the carried PDF
The importer read every rendition with craft_doc::renditions and then
picked the PDF, so a preview whose base64 was invalid or whose data lay
past the end of the file made the whole import fail. It now reads only
the PDF with craft_doc::rendition, from storytold/craft-libs#546, and
never decodes the preview. A document without a PDF still reports
NoCarriedPdf.
* Import: dismissing a picked import's error status acknowledges its details
A picked or menu import queued its error details with the status
report_import_errors set ("first (and 1 more)"), and show_result_errors
then replaced that status with "first; second". The two never matched,
so clicking the status left the details queued and they showed again.
Import commands now move the queued details to the status that replaces
it, also when After Effects project errors lead it. Other commands'
statuses never take over an import's details.
Conflict in crates/render/src/pipeline.rs: kept both the footage-rows
reader of this branch and the frame-mix span reader from main.
Adapted three menus tests to DialogAnswer.paths now being a Result.
Merge reversed-layer-audio-3239. The mixer now plays negative stretch backwards, so
stretch_is_mixed is gone and the Timeline waveform no longer treats reversed layers as
silent: columns follow the layer time mapping, and Backwards on a reversed layer undoes
the reversal. Tests assert mix/waveform agreement for reversed stretch and double reversal.
- CancellableFootage calls off vector_frame too: a closed project's SVG / PDF thumbnail stops
before it rasterises, like a viewer frame nobody waits for.
- replace_project forgets every footage file once and clears the layer cache and the
Cryptomatte probe once, instead of once per footage item.
- The through-Close test covers comp thumbnails (layer cache and footage) as well as footage
ones, and fails with a timeout instead of hanging.
- Docs: forget is also how a closed project's files go.
* Render: memoized audio detection, bounded stretch, preroll, waveform, cache entries, 3D bounds, solids and selector time
comp_has_audio walked a precomp again for every layer using it; each comp
is answered once. A tiny stored stretch made one audio block fetch a vast
source range; playback speed is capped at 100x. Effect preroll is capped at
10 s. Waveforms stop where the audio ends and summarise at most a day.
Disk cache entries are decompressed only when their stated size matches
what their dimensions need (and at most 1 GiB). Advanced 3D triangle bounds
are clamped on both sides before narrowing. Solids larger than 16 Mpx are
made at a lower scale. A camera-facing layer's parent recursion is bounded.
A layer's Expression Selectors share one time allowance. A disk cache
folder under the temp folder is made private and may not be a link.
* address review: audio detection by breadth-first search, private cache folder resolved before the temp check, oversized layers not written, solid cap after rounding, selector static amount, viewport i32 clamp
* address review: private-folder check skips the temp folder itself; override value in expression fallback; assert alias cache opens
* address review: refuse the shared temp folder itself as a disk cache folder
* address review: shared counted selector budget, 30 s preroll, cache hardening (shared temp roots, matte discard, exact LZ4 length, render version 7), orient nesting guard
A precomp instance whose Essential Property names another LUT file renders
with that file, but a layer reading the instance through a layer parameter
was keyed on the nested comp's own file, so editing the overriding file left
that layer cached.
A layer graded with Apply Color LUT or Lumetri's Input LUT and Look was
cached by the parameter text alone, so the same frame kept showing the old
LUT from memory and from the disk cache after the file changed. A layer's
cache key (and the key of a layer reading it, through a precomp too) now
holds each LUT file's size and modification time, and a render lists the
LUT files it read, so a frame kept on disk made with an older file is not
used.
In an 8 bpc comp, a movie used by both a plain and a masked layer has an 8-bit and a float entry in the media pool, decoded separately. Rare and only a speed cost, so it is noted where the 8-bit path starts rather than changed.
AutoPick::ticket chooses a side and reports the frame's time once; dropped without one it ends the frame unmeasured. The viewer's frames, the renderer's top-level frames and frames rendered in passes use it, so an Auto CPU frame turned away by GPU admission (#3028), a cancelled export frame or any later exit no longer leaves its frame counted as running. The frame server keeps its choice across calls and still ends a replaced frame itself.
Layer jobs keep their float buffer until the writer compresses it, so the 64-job limit could hold ~8.5 GB of evicted 4K buffers. Layer and matte writes now also have a byte budget, like frames; a single bigger buffer is still taken when nothing else is queued.
footage_buf8 asked for a movie's 8-bit levels outside Renderer::read, so the media pool noted the movie in no recording: a layer buffer made of it listed no files and was still marked complete, and could be served stale after the movie changed on disk.
A GPU frame that failed or failed its health check, and a remote frame replaced before it finished, stayed counted as running on its side, so every later frame there was timed as one of several side by side and that side looked cheaper. AutoPick::abandoned releases them.
An 8 bpc render at Full quantises plain footage straight away, so the
media pool now keeps those movie frames as the 8-bit levels it is
quantised to (a quarter of the memory) and the renderer expands them
into the layer buffer: the same pixels to the bit. Other depths,
reduced renders, masks, colour profiles, fields, frame blending,
Cineon, Invert Alpha, linear light and 10/12-bit or float movies keep
the float path.
First playback of plain 1080p clips at Full caches ~43% faster on the
CPU, and the footage cache holds 4x the frames.
A preview renders up to 12 frames at once, and Auto only learns a frame's
time when it is done. A starting preview sent every lane to the GPU for
its warm-up, and on an integrated GPU 2–4× slower than the CPU each first
playback stalled. Now at most one GPU warm-up frame (and one re-probe frame
per side) runs at once, a frame's time counts per frame running on its side
at once, and the warm-up starts on the CPU, so the frames on screen come
first. Iris Xe, 1080p: Auto went from 13–34% slower than CPU only to 0–17%.
A footage layer's frame was copied from the footage cache, converted from
its colour profile, then quantised to the project depth in masked_source:
three passes over a 33 MB 1080p float frame. Footage without masks now
comes quantised from one fused pass (Pipe::copy_converted); other source()
callers still get unquantised pixels. Bit-identical (tested at 8/16/32 bpc).
A render thread handed each frame to the disk cache already copied pixel by
pixel and LZ4-compressed (~9 ms per 1080p frame) before it could start the
next one, and layer buffers were encoded even when the queue then dropped
them. The queue now holds raw RGBA (one memcpy via ColorImage::as_raw) and
Arc<Buf>s; the writer thread encodes them. Every frame of a 120-frame bench
run now reaches disk (113–119 before). cache_bench prints the flush time.