- lightcraft-raw: RawImage::develop_binned(k) — k×k superpixel binning of the CFA
(Bayer k even, X-Trans k multiple of 3), black/white normalised row by row, default
crop only, clipping preserved for highlight reconstruction. 2560 px previews of a
24 MP raw bin 2×, thumbnails 4–8×; exports / 1:1 still demosaic (AHD) at full size.
- highlight::reconstruct and its pyramid fill are rayon-parallel (bit-exact).
- X-Trans directional demosaic: precomputed per-phase neighbour offsets and no
reflection in the interior, no per-pixel allocation (bit-exact, tested against the
reference formulation): 2.9 s → 0.75 s CPU at 24 MP.
- files::load_bytes runs on a rayon worker so its short parallel loops don't each
wait for the pool to wake (dominant cost on a loaded machine); LIGHTCRAFT_PROFILE
prints per-stage source timings; files::bin_factor picks the block size.
- render_bench ONLY=source: per-file decode / normalise / demosaic / binning / load.
- Fix: u64 overflow on hostile IFD sizes in DNG/ARW/NEF raw-IFD selection (proptest).
load_bytes(2560) CPU time, ARW a7 III 24 MP: 2347 → 240 ms; X-T20: 3120 → 1220 ms.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`media::develop` renders on the GPU when available (views, render_now, exports;
thumbnails stay on the CPU) and falls back to the CPU otherwise. `app.gpu
{enabled?}` toggles it at runtime, `LIGHTCRAFT_GPU=0` disables it for the process,
`ui.inspect` reports the adapter under perf.gpu.
Performance: dropped device buffers are recycled once their thread has submitted
(fresh 100-300 MB allocations per pass cost as much as the passes), Gaussian
passes ping-pong between two buffers and read all channels of a pixel at once,
and long box radii slide over longer chunks. LIGHTCRAFT_PROFILE prints real GPU
stage timings. render_bench prints CPU and GPU columns and the CPU/GPU difference.
24 MP ARW, loaded machine, min of 7: cold loupe 415 -> 32 ms, exposure drag
22-42 -> 3.9 ms, NR drag (1920) 301+ -> 18 ms, export render 1037+ -> 291 ms;
max 1 LSB difference. docs/gpu-pipeline.md describes the implementation.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The guided base, clarity band, texture band and dark channel (and the paired blurs inside the
guided filter) are independent; each alone scales poorly (subsampled grids, short rows), so they now
run under raster::par_join. Same arithmetic, identical pixels. render_bench gains a per-feature
breakdown of the per-pixel stage (ONLY=finish).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Exports above 2560 px were rendered from the 2560 px preview proxy and upsampled. SourceLevel::Full
(native size; procedural demo scenes at their nominal size) is now selected for outputs larger than
a preview; the media cache keeps only the last full-size original. render_bench gains a batch-export
scenario (decode + render + JPEG per file).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- pipeline: StageCache + render_cached: resampled source, white-balanced/retouched/denoised image
and each spatial plane (log L, guided base, clarity, texture, dark channel + airlight) are keyed
by exactly their inputs; keeps a draft-size and a full-size entry warm.
- Exposure moved after the spatial filters (applied per pixel; planes shifted by EV): identical up
to rounding (test: <= 2 LSB vs. exposing the source), so exposure drags hit the cache too.
- engine: RenderJob::draft / with_stages; ui: loupe and before views own a StageCache, loupe
renders at Draft quality while interacting.
- engine example render_bench (min-of-N wall + process CPU time).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brandon Thomas