Imported video is a Group[Video, Audio] that rendered as a Vello-baked
Vector layer, re-uploading the full frame to Vello's image atlas every
frame (~17ms/frame at 1080p, hitting playback and export alike). Extract
video frames out of the Group/clip scene recursion into
VideoRenderInstances so they composite via the GPU Video path; mixed
video+vector containers fall back to Vello (correct, unaccelerated).
Also route video through hardware sRGB decode: upload raw sRGB bytes to an
Rgba8UnormSrgb texture and blit with a non-unpremultiplying shader variant
(blit_straight), removing the per-frame per-pixel CPU sRGB->linear pass.
Add an F3 GPU-timestamp timer and a per-frame video texture cache.
Drops the live composite of a 1080p video from ~17ms to ~2-3ms.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
`raster_layer_cache` (one ~w·h·16-byte Rgba16Float CanvasPair per keyframe) had no
size cap — scrubbing a long timeline grew VRAM without bound (~33 MB/frame at 1080p),
the largest unbounded consumer. Added a recency LRU (RASTER_LAYER_CACHE_MAX = 12):
`ensure_layer_texture` bumps the frame to most-recent and evicts the oldest past the
budget; the shown frame (and any rendered this pass) is always most-recent so it's
never the victim. Evicted textures re-upload cheaply from the resident/faulted-in
pixels on revisit. `remove_layer_texture` keeps the LRU in sync.
F3 debug overlay now reports the tracked VRAM (raster cache MB + frame count), pushed
from the GPU brush whenever the cache changes (wgpu exposes no allocator query).
Skyler Lehmkuhl