Held for the owner: #995 and #999 overlap, and #995's author offered on 10-12 to
close#995 once his commit was in #999. docs/parity-checklist.md: header keeps the
#1485 note and drops the #995 part.
GPU handles(): only output-device proofs stay on the CPU; Show All Masks now renders zoom windows (upstream); proof-boundary tolerance in the GPU equivalence test; device-profile draft test.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
On an Intel Mac with an AMD Radeon Pro 5300M the Metal compiler crashes
on candle's mlx_gemm kernel, candle-metal-kernels panics on the compile
error, and the worker reported "the AI model failed unexpectedly" and
loaded the model on Metal again for every next job, so AI masks never
worked there.
The worker now retries a job that fails on the GPU (a panic or a candle
error) once on the CPU, after freeing the model. When that works, the
model stays on the CPU for the session and the job's outcome carries a
notice that segment_poll shows once; when the CPU fails too, the GPU
was not the cause and the next job uses it again. Missing or damaged
model files, render and remote-worker errors are not retried. Only
builds with a GPU path (macOS) retry; LIGHTCRAFT_GPU=0 and
LIGHTCRAFT_GPU_BACKEND=off now keep SAM 3 on the CPU from the start
(lightcraft_gpu::env_off_reason), like AI Denoise and the face models.
lightcraft-segment: Sam3::load_cpu, GPU_BUILD; text_logits returns the
crate's Error so Describe failures are classified like clicks.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Conflict only in docs/parity-checklist.md: the generated summary is
refreshed with `cargo xtask parity --write`, and the Last updated line
keeps main's format with this PR's change note (today's date) instead of
the duplicated note the branch carried. ⌥-drag preview still wins over
AllMasks and the selected mask in view_overlay; overlay kinds 6–9 unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Holding ⌥ (Option) while dragging a slider shows a diagnostic view in the loupe, as in
Lightroom; releasing ⌥ or the slider brings the photo back and nothing is saved:
- Blacks: shadow clipping (white; black where every channel is 0, the colour of the
unclipped channels where only some are).
- Whites / Exposure: highlight clipping (black; white where every channel is 255, the
clipped channels' colour where only some are).
- Sharpening Masking: the weight sharpening gets at each pixel, white on black.
- Sharpening Amount / Radius / Detail, luminance NR: the image in grey.
New `Overlay` variants (ShadowClip, HighlightClip, SharpenMask, Grey) are drawn by the shared
`visualize::apply`, so CPU and GPU agree; the sharpening mask needs the detail planes and
renders on the CPU (like HdrRange). The mask uses the same weight function as the render
(`finish::sharpen_weight`). With a display profile the clipping views are tested in sRGB, like
the histogram, and diagnostic views keep the photo's histogram instead of the view's.
Tracker: LR-VIEW-CLIPPING ✅ → 🟡 (J only frames the photo, it doesn't paint clipped pixels)
and LR-EDIT-DETAIL-SHARPEN ✅ → 🟡 (Radius and Detail don't reach the render).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An accelerator hook in the face interpreter takes Conv nodes; lightcraft-gpu implements it with a WGSL
implicit-GEMM convolution (any stride, padding, grouping; split-K for small maps; a direct kernel for depthwise
layers), each model's weights resident on a device of its own. After loading, detection and recognition check
the GPU in the background (agreement within 1e-3 and faster than the CPU) and otherwise stay on the CPU, which
remains the reference; the first GPU error drops it. Sources: own design.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Visualize Spots, Point Color range and Visualize HDR are drawn on the
loupe's 1:1-8:1 window like mask overlays: the range and HDR views act
per pixel, the spots high pass reads within the window's margin. New
pipeline test: each overlay's window equals the whole-frame render's
pixels. The GPU finish kernel now soft-proofs output-space proofs
(destination and display gamut warnings) like the CPU; an output-device
profile's table still renders on the CPU. New GPU equivalence test.
Stacked on pr/zoom-detail (#1027).
Sources: own design.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The loupe sized its slider-drag drafts by its own copy of try_render's
"CPU only" list (soft proof, HDR range). A new CPU-only overlay added to
try_render (e.g. #995's sharpening mask) would have left the loupe
drafting full size on the CPU for its first drag. lightcraft_gpu::handles
is now the single rule: try_render returns NotTried by it and the loupe
asks it, so the two can't drift apart.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`export_falls_back_to_the_cpu_when_gpu_work_is_lost` asserted on
`lightcraft_gpu::last_fallback()`, one value for the whole process that
every other engine test's GPU render overwrites between the export and
the assertion (out of device memory under parallel load, once "a GPU
render panicked"). It also switched the GPU off for the whole process
through the injected fault and cleared everyone's state with
`reset_failures()`.
A render now reports its own fallback: `lightcraft_gpu::try_render`
returns `Result<Rendered, Fallback>` (`NotTried` when the GPU was not
used, `Failed(reason)` when it was tried and the CPU must render), and
`render` is a thin `.ok()` over it. `media::develop_reporting` passes
the reason on as `RenderResult::gpu_fallback`, an export carries it as
`Exported::gpu_fallback`, and each file in an `app.export` batch result
has `gpuFallback` when its render fell back. `last_fallback()` stays as
the process-wide diagnostic for `ui.inspect` → `perf.gpuFallback`.
A simulated failure (`inject_fault`) is thread-local, and so are its
consequences now: a fatal injected fault stops GPU use on the injecting
thread only (`available()` false, `unavailable_reason()` set there);
`reset_failures()` clears that and nothing else, so neither reaches the
other tests in the process. Real failures still stop the whole process.
The engine test asserts on its own export's reason and checks that
another thread still has the GPU; `crates/gpu/tests/fallback.rs` uses
`try_render` and checks the same; a new batch test checks `gpuFallback`
lands on the file whose render fell back. Docs: gpu-pipeline.md,
control-protocol.md (app.export result), parity.md.
A render still running on a worker thread while the process exits crashes in the
driver (SIGSEGV in the NVIDIA Vulkan driver on Linux), where neither the panic
hook nor catch_unwind helps: pool workers were detached, the devices are statics
that are never dropped, and nothing on the way out waited for GPU work.
- gpu: every entry into a device (render, device creation, pool trim, denoise
runner and tiles) is counted as work in flight (exit.rs). begin_shutdown()
lets nothing new in; wait_idle(timeout) waits for what is inside.
- preview: JobPool keeps its workers. shutdown(timeout) and Drop discard the
queued jobs and wait, bounded, for the running ones (2 s on drop); never for
the calling thread, and not twice.
- engine: a render job caught by the shutdown gives up instead of rendering on
the CPU, so quitting doesn't wait for it.
- app: on_exit ends with LightcraftApp::shutdown (2 s for everything), after
settings and library are saved; the macOS menu-bar Quit goes through on_exit.
Not changed: the export, merge, denoise and import threads are still detached;
their GPU work goes through the gate, their CPU work is cut off at exit as before.
Tests: pool drop waits / is bounded / survives a panicked job and its own job
dropping it; the gate; GPU renders refused and the device not created after
begin_shutdown; the app's shutdown leaves no render worker. The crash itself
needs Linux + NVIDIA and was not reproduced here (Windows, DX12): 9 of 9 quits
of the app mid-render exit with code 0 in under 0.8 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Light panel: HDR switch, headroom limit and an SDR Rendition flyout
(shown with HDR on); Visualize HDR beside the switch.
- `Overlay::HdrRange` / `view.visualizeHdr`: the SDR render dimmed to grey,
with pixels the HDR rendition puts above SDR white painted in four bands by
stops above it. Rendered on the CPU from the float HDR render.
- Export dialog: "HDR output" for JPEG (gain map) and 32-bit TIFF.
- Parity (Q. HDR 2 ✅ 3 🟡, LR-EDIT-HDR-MODE ✅, LR-EXP-HDR 🟡), README, ROADMAP.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- `DevelopSettings.hdr`: enabled, headroom limit (`max_ev`) and SDR-rendition
offsets (brightness, contrast, highlights, shadows, whites, clarity), with
control specs (`hdr.maxEv`, `hdr.sdr*`) and copied with the Light group.
- `ToneMap::hdr` / `display_hdr`: the SDR curve through the midtones, with
highlights lifted up to `2^max_ev`; identical to the SDR tables at peak 1.
- `OutputDepth::F32Hdr`: unclamped linear output up to the peak. The finish
stage carries the peak through highlight desaturation, gamut mapping,
encoding, curves, grain and LUT profiles; with HDR off every branch is the
SDR code path, bit for bit.
- `settings_for`: every non-HDR render (CPU and GPU) uses the SDR rendition.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Recover the unpublished denoise work against current upstream. Preserve later opt-in model setup, shared Rust HTTPS downloads, pending-photo resume and an Amount-independent switch. Keep original RAWs unchanged and share the current color finish between ordinary and denoised sources.
Bound public tiling inputs, refuse corrupt-cache exports, and retain synthetic/real-model coverage. Include two Windows path spelling fixes in existing tests. No model weights, face code, media or Tract runtime are included.
Validation: cargo xtask ci (all seven gates); installed RawNIND CPU/GPU agreement 1.38e-6; native CPU preview/export on a public 24 MP Sony RAW, unchanged original SHA-256.
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Generated with Codex.
Keeping a 288 MB upload with each view's stages made them thrash against the stage
budget; not keeping it made every new window (a pan crossing a snap, the first
window of the Before side, any control that invalidates the sampled pixels) upload
and convert the source again, about 64 ms against 6-12 ms. A source too big for a
view's stages is now kept once for all views of the photo (shared_source), outside
the views' stages, and given back when the app idles (trim_pool(0)).
source_uploads and shared_source_bytes expose it for tests and diagnostics.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The GPU renderer renders windows (no more CPU fallback): the vignette gets the
window's offset and the whole output's size (VIG_VIEW), the airlight and noise
reduction scale come from the whole frame, and a window grown for spots or
strokes is cut back after the readback. equivalence.rs renders every settings
case as a window plus a corner vignette and a spot reading from outside, within
the existing 1 LSB. Auto Mask strokes join spots in growing the window (they read
the pixel under each dab). An Auto Mask dab's reference pixel is chosen from its
position at f32 precision, like the GPU kernel: an f64 rounding error that
depended on the window's start could put a boundary dab on either pixel.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
A zoomed view needs the pixels it shows at the zoom scale, not a whole-frame
texture stretched over them. RenderRequest gains an optional PixelWindow of the
virtual full output that max_w x max_h describes; only the window is
materialized. It is Frame:🪟 the crop rectangle narrows to the window
(flips as drawn), so sampling, masks, spots, red eye and the lens warp need no
change. The post-crop vignette keeps the whole frame's geometry (Frame::view)
and dehaze uses the whole frame's airlight (estimated on a 384 px render), so a
window matches the same pixels of the full render. The GPU renderer declines
windowed requests (CPU for now).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The pipeline honoured a section's eye only where it applies Effects,
Optics, Geometry and Calibration, so turning the eye off on Light, Color
or Detail changed nothing. DevelopSettings::effective() resets the
switched-off sections to their defaults (Light with the tone curve; Color
with white balance, mixers, grading and point colour; Detail), and both the
CPU and GPU renderers render those. Borrowed, so free, when every section
is on.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01912WaK6Rmew3wxZuPyX7E8
wgpu loads the driver of every backend in an instance's set while it
enumerates adapters. The compute device (crates/gpu) hard-coded
Backends::PRIMARY (Vulkan + DX12 + Metal) and ignored WGPU_BACKEND, and
was created from LightcraftApp::new on a background thread at launch, so
a crashing Vulkan driver (Intel UHD 630, igvk64.dll) killed the process
before the window appeared; the window (egui-wgpu) also listed Vulkan.
- lightcraft_gpu::backend: DX12 only on Windows (Metal on macOS, as before
on Linux) for both the compute device and the window;
LIGHTCRAFT_GPU_BACKEND=dx12|vulkan|metal|gl|auto|off, else WGPU_BACKEND;
backends missing from the build fall back to the default.
- The device is created only when GPU rendering is enabled, after the
first frame; the preference is applied before the window opens.
- Crash sentinel: gpu-init.marker in the settings folder around device
creation; if left behind, the next launch starts with GPU rendering off
and a notice saying how to turn it back on.
- Tests: backend parsing/selection, sentinel, window defaults, headless
snapshot with LIGHTCRAFT_GPU=0 / LIGHTCRAFT_GPU_BACKEND=off.
- Docs: gpu-pipeline.md (backends, env vars, troubleshooting), whats-new,
parity LR-BEHAV-GPU.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A GPU render no longer hands back an image it cannot vouch for:
- set the wgpu device-lost callback (wgpu never reports a lost device as an
error, so a driver reset went unnoticed); readback waits with a timeout and
checks the poll result and map callback instead of blocking forever
- each render runs in OutOfMemory/Validation/Internal error scopes; errors
fail that render (OOM and limit failures for that render only, device
errors for the process) and it is redone on the CPU
- every buffer is checked against the device limit when created (no
validation error, no process-wide disable); LIGHTCRAFT_GPU_LIMITS=webgpu|<MiB>
creates the device with lower limits to reproduce small adapters
- the output buffer is cleared and the per-pixel kernel writes alpha 255:
any unwritten pixel (work silently dropped, the issue's all-black export)
fails the render; an all-black result is re-rendered on the CPU
- renders submit per stage and every 16 M invocations, the per-pixel stage
runs in ~4 MP bands: no long submission for a slow iGPU's watchdog to reset
- ui.inspect perf.gpuReason / perf.gpuFallback, app.gpu reason/lastFallback,
System Info and Settings say why the GPU is off (e.g. llvmpipe skipped) and
why the last render fell back
Tests: crates/gpu/tests/fallback.rs (limit, mid-render limit, validation
error, dropped work, black photo), engine export falls back to the CPU image,
banded 3000x2000 GPU/CPU equivalence.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
gpu (its render already runs under catch_unwind and falls back to the CPU):
- a failed readback (device lost, out of memory) logs, marks the GPU
broken and returns zeros, so crate::render discards the result and
renders on the CPU, instead of panicking;
- an unknown kernel name does the same instead of panic!;
- finish-parameter fields are looked up as Option (debug_assert on an
unknown name) instead of panic!;
- the warp sampler matches (mode, warp) like the CPU pipeline;
- Buf::raw keeps one documented expect (Some from construction to Drop).
merge:
- pyramid building (push_pull, Blender::finish), camera graph growth,
bundle adjustment, focal median and alignment use let-else / filter_map
instead of expect;
- synth::bracket_dngs / pano_views return Result (a missing demo scene
or a DNG write error is an error); lightcraft-cli synth-merge and the
engine merge tests handle it.
lightcraft-gpu, lightcraft-merge and lightcraft-cli now #![deny]
unwrap/expect/panic/unimplemented/todo/unreachable; clippy.toml as in the
steering PR.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Soft proofing (S in the loupe; S in grids stays Expand/Collapse Stack):
- pipeline Proof {space, destWarning, displayWarning}: gamut-map into the
proof space, show it in the output space; red/blue gamut warnings
(neutral clipping excluded). CPU only; the GPU path declines proofs.
- RenderJob::with_proof; paper-white surround and Proof Preview label;
Soft Proofing strip under the histogram (profile, warnings, Create
Proof Copy = a named virtual copy). view.softProof for control/MCP.
- photo.virtualCopy takes an optional name.
Mask components: MaskComponent.name; mask.component {component, action:
invert|duplicate|delete|rename|op}; right-click / … menu per component,
double-click rename, Intersect button.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
RenderRequest gains a sample format (OutputDepth: U8 for previews, U16, F32Linear); deep renders
run the per-pixel stage with the exact sRGB curve and return Rendered::deep alongside the 8-bit
image (GPU stays 8-bit: deep exports render on the CPU). Export encodes deep renders: 16-bit PNG and
TIFF, 32-bit float linear TIFF with a linear ICC profile we generate, 10-bit AVIF (YCbCr planes into
ravif). Output sharpening and watermarks work on deep images. ExportOptions.bitDepth with per-format
defaults (TIFF 16-bit, others 8-bit), export dialog bit depth choice (remembered with the last
export), MCP export schema. Tests: a smooth gradient has 1024 levels in a 16-bit TIFF vs 13 in
8-bit, float TIFF decodes to the same light as 16-bit, per-format depth rules, AVIF 10-bit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- memory::budget (default min(25% RAM, 1.5 GiB); LIGHTCRAFT_MEMORY_MB, app.memoryBudget):
half for engine caches (sources + rendered previews, LRU across them via a global
use tick; the photo on screen stays), a quarter for decodes in flight (WorkGate:
thumbnails, prefetch and import probes wait; loupe/export loads never wait), an
eighth for the GPU free pool.
- GPU: best-fit buffer reuse (<= 25% slack), buffers released outside a render are
pooled at once (no stranded retired lists), pool trimmed after 3 s idle.
- Decodes drop file bytes and raw samples early, colour/crop/orient in place.
- macOS: apps install malloc_zone_pressure_relief as the release hook (after each
decode, import and idle trim): freed large blocks no longer stay resident.
14 raws + loupe + 12 steps (M4 Pro): peak RSS 1.9-2.3 GB -> 0.83-0.99 GB, footprint
2.5-2.6 GB -> 0.91-0.96 GB, heap peak 930 -> 656 MB; stepping median ~13 ms.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Once the current photo's loupe render is done, the next and previous photos in filmstrip
order get a background loupe job (Slot::Prefetch, priority below on-screen thumbnails,
once per settings): the decoded 2560 px source is kept in the engine's bounded preview
LRU and the view render is cached, so stepping to a neighbour shows its developed image
in ~50 ms and then re-renders only on the GPU.
- lightcraft_gpu::warm_up (called when the app is created) builds the device and kernels on
a background thread; gpu::ready() answers without blocking and ui.inspect uses it, so the
UI thread never waits for device creation.
- docs/gpu-pipeline.md: opening-a-photo section.
Stepping through 24 MP raws in the loupe (real app): full render 220–580 ms → 45–62 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>
New L3 crate `lightcraft-gpu`: a wgpu device (Metal/Vulkan/DX12; none on wasm32),
WGSL kernels compiled once, buffer/dispatch/readback helpers, and a renderer that
runs the per-pixel stage (tone map, highlights/shadows, clarity/texture/sharpen,
dehaze, local mask terms, OkLCh colour ops, grading, vignette, gamut map, sRGB
encode, curves, grain) on the device. Earlier stages still run on the CPU and are
uploaded (per-stage hybrid). Stage results are cached per view next to the CPU
StageCache (new `StageCache::extension`).
The CPU pipeline stays the oracle and shares its resolved parameters with the
kernels: `finish::FinishParams` / `mask_terms`, `pipeline::plan` (stage-cache keys),
`local::plane_sigmas` / `nr_params` / `wb_matrix_for`, exact LUTs and OkLab
matrices exported for the shader prelude. Equivalence tests (skip without an
adapter): mean |d| < 0.5 LSB, max <= 3 LSB; measured max 1 LSB on Apple M4 Pro.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
