mirror of
https://github.com/storytold/effectcraft.git
synced 2026-11-22 02:11:07 +01:00
4e9fcee1c5bc7252e06cbff72f2bf746adf6a41c
4
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
|
389894e69b |
gpu: run Content-Aware Fill's PatchMatch and flow on the GPU (#1856)
* gpu: run Content-Aware Fill's PatchMatch on the device Content-Aware Fill's synthesis is the stage that makes the tool slow: at 1920×1080 with a 300×200 hole, PatchMatch costs ≈ 1.1 s per frame on the CPU, and it is what Object, Surface and every refining frame spend their time in. Add a `FillBackend` trait to `effectcraft-raster` whose stages (the membrane solve, the flow fields, PatchMatch) default to the CPU reference, and `Accelerator::inpaint()` on the compositor, which is how the engine asks for a device backend. `effectcraft_gpu::inpaint` implements PatchMatch: the multi-scale pyramid, the EM rounds, the propagation (a two-pass checkerboard instead of the CPU's scan order), the random search, the SSD, the histogram-based σ² and the weighted vote, as compute kernels in `shaders/inpaint.wgsl`. The fill's frames are read back to the CPU before it runs (analyses composite on the CPU by design), so the backend is about the fill's own work; the orchestration, the flow-guided propagation and the per-frame flow control stay on the CPU. Only the hole's bounds cross back to the CPU, which is what makes the offload pay. Measured on an Intel Arc (MTL) iGPU, 1920×1080, 300×200 hole: CPU 1087 ms GPU 397 ms (2.7×, 41.0 dB PSNR against the CPU's synthesis) The membrane (harmonic) solve deliberately stays on the CPU: the same measurement has it at ≈ 90 ms on the device against ≈ 55 ms on the CPU, because a fill's full-frame data movement (upload, the known pyramid, a copy) does not shrink with the hole, while the relaxation does. The device takes only the stage where its parallel search pays; `membrane_tex` still runs on the device inside PatchMatch, where it is the coarsest level's start and never crosses back. Closes #1839. * gpu: fix the fill's vote histogram and run its flow on the device Review follow-ups on the Content-Aware Fill GPU port, plus the fill's flow. Fixes: * imp_hist kept about one sample in 256: each of the 256 threads scanned a strided subset of the pixels and retained only those whose bin selected it, so the vote's sigma² came from a sample correlated with the pixel index (and a small hole usually fell back to 1.0, making the weights near-uniform). One workgroup now accumulates the whole population through shared atomic bins, with a test that pins the sample count. * imp_pm_prop and imp_pm_rand were dispatched over grow(near, 1) but with near's origin, so the ring's left column and top row were never written (NaN under test, stale pooled data in production) while a strip past it was. `over` now derives the shader's origin from the box it sizes the dispatch from, so a dispatch's region and its origin can no longer disagree. * Gpu::inpaint() requires its kernels to have built. Otherwise every stage paid an upload and a failed dispatch to fall back to the CPU, while the fill still reported the GPU path. * imp_rnd's XOR chain parenthesises its multiplicative terms: WGSL's grammar makes each XOR operand a unary expression, so `a * b ^ c * d` is rejected by spec-compliant parsers (naga accepts it). * FillBackend documents that its stages may be called concurrently. * The coarsest membrane's sweep count (64, against the CPU's 200) is now justified: matching 200 measured ~6% slower over the fill with no change in the result. And the flow: Gpu::flow runs block_flow's coarse-to-fine search, sub-pixel fit and vector median on the device. The luma is the CPU's (flow::luma_pair), so the field matches block_flow's to a float rounding and the propagation sees the same motion; the hole completion stays on the CPU with the membrane. A four-frame 1920x1080 fill goes from 1.27x to 1.42x against the CPU, and the device path is covered for holes from 16x16 to 300x200. * gpu: keep the fill vote's histogram range on HDR input imp_hist binned every distance at a fixed 256 bins per unit and clamped to 1023, so the vote's sigma2 topped out at 3.998. A 32-bpc frame's channels can exceed 1, and masked_frame forwards them unclamped: a patch distance of 1600 then reported sigma2 ~ 4, exp(-1600/(2*sigma2)) rounded to zero in f32, and imp_pm_vote kept the prior image instead of synthesising. The histogram now reports the largest finite distance it saw (atomicMax on the float bits, carried in bin 0's unused .g channel). SDR input keeps the fixed 0..4 range with identical bins; HDR input is re-binned over the observed range and percentile() maps the bin centre back through it. Nothing clamps the input. Also group over/sweeps/pm_level's arguments into small structs: the workspace denies too_many_arguments and the lint policy (#2496) does not allow new #[allow]s. No behaviour change on those paths. * engine: don't report Edge Blend fills as device fills Edge Blend calls only the membrane, which the GPU backend leaves to the CPU reference, so run_plan's "Filling (GPU)" status was wrong for it whenever a backend was present. Object and Surface still run the flow and PatchMatch on the device and keep the label. CodeRabbit's #1856 finding. * gpu: run the fill's device stages inside error scopes An allocation or validation failure inside patch_image/flow_image left an invalid texture whose later use raised an uncaptured error, retiring the shared compositor instead of falling back to the CPU for that frame. Wrap both stages in Gpu::within_memory, whose scopes are native thread-local, so the fill's rayon workers do not share them. --------- Co-authored-by: FredApps <153173854+FredApps@users.noreply.github.com> |
||
|
|
08362b9c04 |
GPU: free effect working textures after each call instead of at frame end (#1523)
* GPU: free effect working textures after each call instead of at frame end A GPU frame keeps its encoder open while the CPU renderer runs layers with GPU effects (Gpu::effects -> run_chain). Each call's Enc handed its textures back to the shared pool, where they waited for every open encoder, so all effect intermediates of a frame stayed allocated until the frame ended (17.7 GB by layer 17 of a heavy 4K comp, then OOM). Effect calls now borrow a pool owner (Enc::lent) that is given back once the call's work is submitted or discarded; its kept textures are trimmed to the pool budget, so the rest are freed after each call. * GPU: one pool budget for idle effect calls, trimmed at the end of a render Review of #1523: each idle lent owner kept up to POOL_BUDGET of textures of its own, outside the shared accounting, and there were as many idle owners as effect calls ever ran at once (16 frame threads: ~12 GB after a 4K render). Giving a lent owner back now moves its kept textures to the shared free list, so one POOL_BUDGET bounds what the free list and every finished effect call keep. They skip Pool::release's wait for open encoders: only the lent Enc recorded work on them, and that work is submitted or discarded. Each lend is a fresh owner id, so a lease that outlives its call (an image handed to another encoder) is released through Pool::release; this does not rely on the result being read back. Accelerator::render_end (called when an export ends) drops the kept textures. * GPU: evict the pool's oldest free textures, guard escaped leases Review of #1523 (follow-up to 0a884fb6): - Pool::make_free rejected a texture when the free list was full and never evicted, so once sizes that no longer occur (per-frame blur bounds, padded layer sizes) filled the budget, every later free was a drop and every take an allocation. It now evicts the least recently freed textures until the new one fits. - A lease read by another Enc (dispatch, copy, readback, display) or handed to a cache is marked escaped and always released through the guarded Pool::release, so a lent call can no longer reuse a texture that a still open encoder reads. - Docs: POOL_BUDGET caps the free list only; textures waiting for open encoders and those of live Encs are outside it. - export_async calls render_end from a drop guard, so the pool is also trimmed when an export panics or its future is dropped. - Tests use a private context with its own pool budget instead of a global override; new tests for stale sizes and an image read by an open encoder. * GPU: evict only stale pool sizes, count evicted textures until the GPU frees them Evicting the least recently freed texture for every newcomer thrashed when a frame frees more than the budget in a repeating order (blur/glow intermediates): each free evicted the size the next call takes first, so misses came back to the old level and evict + reallocate raised the peak. The free list now evicts only sizes nobody asked for in the last 2 output frames (Accelerator::frame_end, called by the export per frame); otherwise the incoming texture is dropped and the kept set keeps hitting. Evicted bytes count against the budget until the work submitted so far is done, so evictions run at most one budget ahead of the GPU. * GPU: lower the shared pool budget to 256 MB On 4K RTX 3090 runs 256 MB matches the old code's peak reserved on d_blurglow (9204 MB) and is 256-512 MB above it on c_effects (6592 vs 6080 MB), while staying faster. * GPU: end frames on the multi-frame render paths, route recording through Enc::record Only the export pipeline called frame_end, so off export (preview, remote frames) the pool's clock stood still, no size went stale and a full free list dropped every newcomer. passes::comp_frame (exports), remote frames and the preview workers now end each frame; test without pool_frame_end. Docs: the frame clock is coarse with many frames in flight; dying only throttles eviction, approximately; no cache path escapes images. Enc::record is the single way to get the encoder outside context.rs and marks the images it names (Enc::uses). * GPU: wait for the empty submission while evicted pool textures still count stale_sizes_do_not_starve_a_steady_size failed with 0 hits and 20 misses on hardware adapters: pool_frame_end submitted an empty batch and then only did a non-blocking Poll, and the queue's work-done callback did not run before the next frame, so Pool::dying never drained and make_free stopped evicting stale sizes. While dying is non-zero, pool_frame_end now waits on that empty submission for up to the readback timeout (the pool lock is already released). A wait timeout just moves on and does not retire the context or its readbacks; other poll errors still retire them. Without dying textures it keeps the cheap Poll. Enc::crop records through Enc::record(&[src, &dst]) so both pooled images are registered like on every other recording path (review). * GPU: do not repeat the pool frame-end wait until evicted textures drain A wait that frees none of the evicted bytes (timeout, or a WebGPU poll that does not block) marks the pool stalled; later frames only poll until dying drops. Test covers the no-repeat and partial-progress cases; dying decrement saturates. * Remote frames: end a pending frame that a new render replaces A render arriving while a multi-pass frame waits for readbacks dropped that frame without `frame_end`, so the GPU pool's clock skipped it. Errors return before `frame_begin`. * Render: end the frame even when a comp frame panics `passes::comp_frame` called `frame_end` only after a normal return, so a frame that panicked (or whose future was dropped between passes) never aged the GPU pool. * GPU pool: log the reuse rate at render end; test reuse across effect calls and frames Log cumulative hits/misses (and reuse %) once per render before trimming. Tests cover reuse within an open frame across effect calls, and across repeated frames. --------- Co-authored-by: Matej Kozlovský <MatejKozlovsky@users.noreply.github.com> Co-authored-by: bflatastic <bombayapirat@gmail.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: FredApps <153173854+FredApps@users.noreply.github.com> |
||
|
|
b7b75ed37a | style: cargo fmt on the moved files | ||
|
|
a7660a1e37 | refactor(engine): move jobs, offload, remote to background/ |