GpuContext::request() now applies InstanceDescriptor::with_env() (WGPU_BACKEND,
WGPU_DX12_COMPILER, WGPU_VALIDATION, WGPU_DEBUG, ...), the same variables the
desktop app honours via egui-wgpu, and requests PowerPreference::from_env(),
defaulting to HighPerformance (an unknown WGPU_POWER_PREF logs a warning).
WGPU_ADAPTER_NAME (CLI/headless only; the desktop ignores it) picks the first
adapter whose name contains the trimmed value, case-insensitively. Empty or
blank counts as unset. Only the name is matched (backend via WGPU_BACKEND). A non-matching name is an
error that says so and lists each distinct adapter once as 'name (backend)', instead of wgpu::util::initialize_adapter_from_env's panic;
no adapters at all gets its own message.
Runs every scenario N times (default 1, unchanged). Every run is still
appended to results.jsonl; with N > 1 one "kind": "summary" row per key
follows, whose metrics are the medians of the successful runs (user_s and
sys_s included, byte/instruction counts rounded) and whose summary holds
median, p90 (nearest rank, = max for N < 10), min and max. A failed run marks
the summary failed so the gate refuses it.
--compare uses the latest row per key; the table labels each side "median
of N" or "1 run" and warns when they differ, and --fail-over refuses such
keys. --repeat with --with is rejected. docs/testing.md documents the format.
Entries whose layer buffer was freed were pruned only once the cache went over its budget
(1.5 GiB), so an animated layer, which makes a new buffer every frame, left a dead 4K texture
per frame allocated until the cache filled up.
- The cap applied whenever an accelerator was attached, so Software Only
projects (which render on the CPU) rendered one frame at a time. It
now needs the project's GPU acceleration too, as #898 gates it.
- Only the render is capped: each window of max(cap, threads) frames
renders in chunks of the cap, then encodes in parallel, holding at
most one window of images.
- An invalid EFFECTCRAFT_GPU_INFLIGHT (0, negative, not a number, too
large) logs a warning before using the default.
- The default stays one frame: the GPU backend reports no free or total
device memory portably, so it is not sized from VRAM.
Three frames at once still filled an 8 GB card on the #879 project at
1080p (7.9 GB, the same as no cap); one frame peaked at 6.3 GB and two at
7.0 to 7.8 GB. One matches how the export scheduler in #898 renders image
sequences (one GPU worker, synchronous readback). EFFECTCRAFT_GPU_INFLIGHT
still raises it.
With a GPU accelerator every frame of a batch rendered in parallel, each
holding its own working textures, so a 30-frame 1080p batch could take
the whole card (20+ GB on an RTX 3090) and run shared-memory GPUs out of
memory. Cx::frames now renders GPU frames in consecutive chunks of 3
(EFFECTCRAFT_GPU_INFLIGHT overrides it), the same cap as the export
scheduler in #898; output order and the CPU path are unchanged.
Fixes#879
AuxChannels::crypto_rank_planes resolves a layer's (ID, coverage) planes once; crypto_ranks keeps its signature on top of it. The CPU effect and the GPU effect's aux image (effectcraft_effects::crypto_aux_image, rayon) index the planes per pixel. Output is bit-identical.
Matej Kozlovský