The shell's gpu-stage feature is on by default and Preferences > Stage >
Renderer defaults to Graphics card (GPU), with the CPU drawing whenever
there is no usable card, a tile is past the GPU's limits or anything
fails. The setting is stored under a new key: every earlier build wrote
the old one as Cpu without anyone choosing it. CI installs lavapipe so the
parity tests run on a software adapter.
Measured (release, RTX 5090): 4K filtered sample 3.9 ms on Vulkan and
4.2 ms on DX12 (CPU 31.5 ms); 200 bitmap fills 4.4 / 4.3 ms (CPU 37.2 ms).
Docs: ADR 0031, performance, roadmap, architecture, CHANGELOG.
The GPU stage shares eframe's wgpu device (egui_wgpu::RenderState) and
draws each stage tile into a storage texture the shell registers with
egui as it is. Fallbacks: a headless device with readback (window on
OpenGL), then the CPU oracle. Every call runs in thread-local error
scopes; nothing is installed on the window's device; a work bound per
dispatch guards the driver watchdog.
The engine stays GPU-free: Raster::Native carries an opaque image, cached
under its own 384 MB budget; only the shell downcasts it.
perf, gpu_bench on RTX 5090 (Vulkan), 32 threads:
4K, 1,000 shapes: CPU 16.2-19.6 ms, read-back 8.4, direct 2.1-2.4 ms
4K, 10k shapes: CPU 25.8-28.1 ms, read-back 23.2, stage tiles 10.2 ms
4K, masked 10k: CPU 74.8-80.0 ms, stage tiles 10.9 ms
Binning stays on the CPU (ADR 0029: not the bottleneck as stage tiles).