* perf: reuse large buffers so a pinned glibc trim threshold costs no speed (#10479)
With MALLOC_TRIM_THRESHOLD_ set (#8743), glibc maps every allocation over
128 KiB fresh, so tiles, compositor bands and filter scratch were page-faulted
in on every call: viewport compose 3.6x slower, Gaussian Blur up to 1.8x,
.pcraft save 2.4x.
- photocraft_raster::scratch: a process-wide, capped pool of large Vec
buffers (u8/u16/f32/[f32; 4]). A miss drops the oldest buffers worth its
size down to a 256 MiB floor, so peak memory stays near glibc's defaults.
The engine releases the pool after any command that closes a document,
Headless::close does too, and an idle thread releases it after 2 s unused.
- Tiles, compose Buffer/masks/bands, algo Image and the blur paths, engine
adjustment/filter scratch, GPU tile upload and io import buffers go through
the pool.
- .pcraft save reuses one zstd FrameCompressor per worker (byte-identical).
- interactive_bench reports minor page faults per run.
36 MP, trim threshold set vs glibc defaults: viewport 5.6 vs 6.1 ms (was
22.7), Blur r4 445 vs 437 ms (was 852), Curves CPU proxy 35.0 vs 35.7 ms
(was 50), flatten 146 vs 147 ms (was 237), save .pcraft 251 vs 230 ms (was
545). RSS after closing a 5400x7200 document: 21 MB with trim set.
* fix: corpus tests build again; return remaining pooled buffers (#10479 review)
- Corpus-gated io tests and the Photoshop-oracle test moved `.px` out of a
`Buffer`, which now implements `Drop` (E0509 with `--features corpus`):
use `into_px()`.
- io: masks, the merged image, extra channels, smart-filter masks,
smart-object pixels and patterns give their `interleave` buffers back to
the scratch pool after writing them.
- compose: recycle the mask values of a pass-through group, `effect_shape`
and both `split_parts` callers.
- Styled Free Transform preview copies each band out instead of moving it,
so the band returns to the pool.
- interactive_bench: the fault count covers each whole run, untimed setup
included; say so.
* fix: recycle the masks of adjustment-atop and shape layers (#10479 review)
* fix: recycle effect hiding masks (#10479 review)
* fix: recycle knockout, CMYK and GPU shape masks; release the pool when Split/Merge Channels close a document (#10479 review)
* fix: release the pool when a closed document's cancelled workers exit; idle-release switch for tests; don't pool PSD planes (#10479 review)
* fix: release the pool for workers cancelled before their document closes; take the pool lock for budget and idle-release changes (#10479 review)
* fix: idle release rechecks activity under the lock; try_take_exact releases the pool and retries before failing (#10479 review)
* test: cover a successful try_take_exact miss; name the failed-reservation test for what it checks (#10479 review)
* fix: pool the final GPU effect shape buffer and return it after use (#10479 review)
* docs: describe scratch pool lifecycle regression invariants
* test: verify polling releases cancelled document scratch buffers
* perf: recycle remaining transform selection and PSD scratch buffers
* fix: free PSD gather bands before allocating channel planes
* test: assert exact scratch take reuses pooled storage
Loading the Intel Vulkan driver on Windows (igvk64.dll) crashes the process
with STATUS_ACCESS_VIOLATION (#4). The app leaves Vulkan out on Intel, but
the tests use wgpu's defaults, so on such a machine `cargo test` dies in the
first GPU test (ui-egui's background_redraw, adjust_preview_gpu, ...).
WGPU_BACKEND=dx12 avoids it for the kittest harnesses, but seven tests and
four benches built their instance with wgpu::Instance::default(), which takes
every backend and ignores the environment. They now use
InstanceDescriptor::new_without_display_handle_from_env(): the same defaults
with WGPU_BACKEND and friends applied, so nothing changes where it is unset.
Document the workaround in docs/development.md and the testkit driver notes.
Blurs run as box passes now filter the whole area at once instead of tile by tile: the
horizontal passes run over the source rows only and the vertical passes over column strips,
with the canvas edge repeated virtually instead of materialising a 3-sigma margin around every
tile (at radius 500 that margin was ~25 MP of floats per 4 MP tile, copied and transposed
several times). The results are written to the layer's tile columns in parallel.
Output matches the previous running sums (8-bit: identical except 114 of 2.1 M samples one
level apart; 16-bit: within 2/65535) and a direct-convolution oracle at every depth.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* Adjustment dialogs: preview through a clipped adjustment layer on the GPU (#77)
While an Image › Adjustments dialog is open on a pixel layer, the canvas shows the
document with a temporary adjustment layer (the dialog's settings, the selection as
its mask) clipped to the target instead of re-running the command on a CPU proxy.
The canvas compositor draws it, so on the GPU canvas a change recomposites only the
target's area within the selection, and of that only what the view shows (the rest
catches up as the view moves); zoomed out on a large document it composites a copy
reduced to about screen resolution, built once per session. The base document and
region are built once per session; a change swaps the layer's adjustment in a
shallow copy. OK still runs the real command (one history step); Cancel drops it.
Kept on the CPU proxy where the layer composite can't match the command: non-pixel
or clipped targets, Blend If, visible untouched area, channel / Quick Mask targets,
CMYK/Lab/other modes, colour-making kinds in Grayscale, and documents the GPU
compositor doesn't cover. The navigator waits for settings to settle.
Tests: preview vs command within 1/255 (worst 0.37/255) for all 13 kinds at
8/16/32-bit, RGB and Gray, with/without selection; nothing changes outside the
region; dialog flow (cache, keys, damage, Cancel untouched, OK one step); GPU
parity of the preview and of the zoomed-out proxy.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* interactive_bench: per-change Image › Adjustments dialog preview rows (#77)
Times the old CPU proxy preview against the clipped-layer preview at full
resolution, at 100 % view and zoomed out, with and without a selection.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Compose: render in place and in bands; area-reduced and proxy thumbnails
render_tiled wrote every tile into a Vec and then copied them into the output:
two full-size float copies (6.3 GB for a 14000² document). Tiles are now
written straight into the output band by band.
New streaming APIs so large documents never need a full-size float
composite: render_bands (sink per band of ~8 MP), render_reduced (banded
area average, bit-identical to the old flatten-then-average thumbnail) and
flatten_to_surface (Flatten Image, Merge Visible, Copy Merged and the other
flattened-surface users).
thumbnail() reduces a downsampled proxy (moved from ui-egui to
compose::proxy, now also scaling spot channels) for documents over 16 MP
when that is faithful (no layer effects or vector masks), so navigator,
channel and histogram thumbnails of a 196 MP document cost milliseconds.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* GPU canvas: request the adapter's real limits; banded CPU fallback; navigator thumbnail
egui-wgpu asks for an 8192 px texture limit, so the wgpu compositor rejected
every document wider or taller than 8192 px and each refresh fell back to a
full CPU flatten. The desktop and web apps now request the adapter's own
max_texture_dimension_2d and max_buffer_size (gpu_canvas::use_adapter_limits);
the benches use the same limits.
The CPU fallback (documents beyond the GPU limit, or features it lacks) is
GpuCanvas::refresh: it renders and uploads the composite band by band
instead of a full float composite plus a full RGBA8 copy. The CPU canvas
texture is reduced in bands too, and the Navigator on the GPU canvas uses a
cached thumbnail instead of a full-resolution CPU composite per revision.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* IO: banded export and import; no full-size copies in encoders; larger decode limits
- Flat export renders the composite in bands and quantises each band; alpha
is dropped in place when every pixel is opaque. The single-layer native
path copies surface rows a band at a time (no f32 read_region and no
per-pixel Vec). Buffers come from try_reserve: Err, not an abort.
- PSD export encodes the merged image band by band into its planes (no
flatten, no second plane copy); near-opaque documents are re-encoded
without the white matte so the output is unchanged. Layer interleaved
bytes are freed before their planes are compressed.
- Flat import converts in ~32 MB bands (and not at all when already RGBA);
the PSD merged-image un-matte works per tile row.
- Encoders borrow the image when no conversion is needed (Image::converted)
instead of cloning it; PNG only copies 16-bit data for the byte swap.
- Decode limits: 2^30 pixels and 8 GiB on 64-bit targets (a 20000² JPEG
was refused at 2^28).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Filters: write tile results in memory-bounded groups
apply_tiled collected every tile's float result before writing any (3.1 GB
for a 14000² layer). Results are now produced and written in groups of about
256 MB (at least one tile per core).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* History: bound pixel memory by Preferences > Performance > Memory Usage
History::max_bytes caps the current document's tiles plus the tiles only
history holds (layers, masks, alpha channels); History::trim drops the
oldest states beyond it after every recorded step and when preferences
change, always keeping the last step undoable. The engine sets it from
memory_usage_mb, which was unused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Large-image benchmark (14k / 20k): open, refresh, thumbnails, filter, stroke, saves
One operation per run on a cached synthetic photo, with the GPU device created
the way the app creates it, so /usr/bin/time -l reports that operation's peak
memory. Documented in docs/development.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
CI's `cargo fmt --all -- --check` has failed on every run because the
repo had no rustfmt config and was never formatted. Use the same settings
as EffectCraft/FilmCraft (max_width 160, use_small_heuristics Max), which
also keeps the churn small. Formatting only; no code changes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Harris+BRIEF/RANSAC panorama with bundle adjustment, seams and multi-band
blending (Auto-Align/Auto-Blend share it); Debevec-Malik HDR merge with
ghost removal and tone mapping; Brown-Conrady lens correction (filter,
batch, smart filter); wide-angle constraint mesh warp; Crop and
Straighten Photos; Camera Raw Filter (Basic, curves, mixer, grading,
detail, effects) with dialog; Vanishing Point planes (engine only).
Fixes warp_triangles tile binning.
Checkpoint also includes compiling in-progress work of the remaining agents.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Gradients/Patterns/Styles/Shapes/Tool Presets/Clone Source panels with
headless preset commands and a Custom Shape tool; Color card tabs match
Essentials. Character and Paragraph Styles (doc model, .pcraft, PSD style
sheets), Glyphs panel + type.insertText, Check Spelling (SCOWL-50 list).
Checkpoint also includes compiling in-progress work of the remaining agents.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>