Repository navigation
GPU compositor: effect stacks round-trip through the CPU, intermediates allocate fresh textures, and Noise on an adjustment layer re-uploads its footprint every frame #687
Description
Activity
Partial: #701 fixes two of the four — the adjustment footprint no longer re-keys each frame under a time-dependent effect (127 MB/frame re-upload gone), and over-budget upload eviction is now counted, logged once and settable via
GpuContext::set_upload_budget.Deliberately not attempted: stacks staying GPU-resident (#1) and per-frame intermediate texture reuse (#2). Both are the measured wins but need a DX12/Vulkan/Metal device to know a change helped; every box here has only GL (declined by the compositor) or llvmpipe, so any number we would post would be fictional. The footprint fix does cut one full-frame upload per frame for finished comps; the residency rewrite and the intermediate pool want a run of
effectcraft-cli bench --gpubefore and after on hardware like yours — a maintainer with such a machine should drive that, not a guess-and-check patch.Thanks @vbasky, nice work on #701!
Heads-up so we don't duplicate effort: #688 already covers all four items from this issue, including the two in #701 (the footprint key ignoring time-dependent effects, and the upload-cache budget with a single warning when it's full). It also adds the two GPU parts you mentioned leaving to someone with real hardware: resident effect stacks and in-frame texture reuse.
I measured it on an RTX 3090 test PC (4K, 90 frames,
RAYON_NUM_THREADS=1so frames stay on the GPU): 81.1 s → 49.3 s (1.65×), with peak VRAM 17.2 GB → 9.4 GB. Output is pixel-identical to main's GPU render, and CPU output is byte-identical to main.@bflatastic, happy to go whichever way is easier. If you'd rather merge #701 first, I'll rebase #688 on top of it so it only carries the GPU parts.
@MatejKozlovsky thanks — we will not duplicate the GPU-resident stacks or in-frame texture reuse. Those belong in #688, and your 3090 numbers are the measurement we could not take here.
#701 stays the small cache-key / upload-budget slice (items 3–4 only). If maintainers land that first, rebasing #688 on top so it only carries the GPU parts is the easier split. If #688 merges first, we can close #701 as already covered.
@bflatastic either order is fine from this side.
Version: main (after #641/#648) · Area: GPU compositor (
crates/gpu), layer cache (crates/render/src/cache.rs)What happens
On the GPU compositor (
--gpu, Mercury GPU Acceleration), a 4K comp with effect stacks spends most of itstime moving pixels rather than compositing:
styled_layer), whichruns each run of GPU-capable effects as upload → kernels → readback, then uploads the result again to
place it.
that made it waits in the shared pool until that encoder submits, so a frame allocates a new texture for
almost every intermediate (a 4K RGBA f32 texture is 127 MB).
masks, without effects) is keyed with the effects included, so Noise or Grain on a finishing adjustment
layer changes the key every frame and the GPU re-uploads the full-frame footprint each time.
textures are evicted and re-uploaded every frame without any log line, and an embedder cannot size the
budget to the device.
Measured
On a test PC (Ryzen 7 5800X, RTX 3090 24 GB, Windows 10),
effectcraft-cli bench --gpuon a 4K compwith effect stacks (main, median of 5): GPU cold 1181 ms/frame, GPU warm 208 ms/frame. A 3 s / 90-frame
4K export with
--gpu(RAYON_NUM_THREADS=1, so frames stay on the GPU) takes 81.1 s with a 17.2 GBVRAM peak.
Expected
Effect stacks stay resident on the GPU, a frame reuses its own intermediates, an adjustment layer's
footprint keeps its key across frames, and a full upload cache is reported.