@graview/render
The spatial renderer: three planes, an affine transform per plane, and connectors drawn in SVG over the scene.
Two paths, and which one draws
A scene draws on the GPU where the browser can, and on the DOM path
everywhere else — silently, per page. Scene and the Shell default to
renderer="auto": gpuCapturePresent() asks, without a device, whether the
page has WebGPU, an element-capture call on its queue, a canvas paint
event, and hit-testing that follows the drawn geometry (below); where it
does, ./gpu is fetched and attachGpu() takes the canvas.
Its first-frame check falls back for good if no adapter is given, no paint
arrives within FIRST_PAINT_WAIT_MS, a frame is refused, the device is lost,
or the middle of the first FIRST_FRAMES frames' work is over
FIRST_FRAME_BUDGET_MS. A page without any of it — Firefox, Safari,
Chromium before element capture — never fetches ./gpu and says nothing.
?renderer=gpu|dom on the address (or graviewRenderer set on the page
before it loads) decides auto for an operator; GRAVIEW_RENDERER does the
same for every page of a harness run.
Element capture (CanvasDrawElement) is on by default in Chromium 158 and
behind --enable-blink-features=CanvasDrawElement before it. Canary 158
respelled the API (content="drawable" on the canvas, drawable on each
view, a plain { texture } destination); both spellings are written.
The two paths draw one picture. Each view is an ordinary element on
both: on the DOM path it carries a CSS matrix3d, opacity and filter; on
the GPU path it is a child of the canvas, captured, and the shader draws the
same treatment from the same numbers (treatmentOf, recededOpacity). The
accessibility tree is the same on both, and scripts/verify-renderer.mjs
holds that, the picture, and the silence.
Blur. Receded planes are blurred, the focused plane always sharp. On the
GPU the shader draws it at no measurable cost; the DOM path's CSS
filter: blur() is dropped for the rest of the page if its frames run over
a frame and a half while the picture moves. blur={false} on the Scene or
the Shell turns it off, and so do prefers-reduced-motion,
prefers-reduced-transparency and the reader's Motion setting
(planeBlurOn).
Capture costs about 0.22 ms a view on Chromium 158, so retaking every live
view every frame holds sixty frames a second only below about 64 views (6 fps
at 512). attachGpu() keeps a page under that: it retakes a view only when
the browser's paint event says it changed, skips what is off the canvas,
spends at most CAPTURE_BUDGET retakes a frame, and hands a view that keeps
changing faster than that to the DOM path, drawn at the same transform under
the canvas. The discography holds about 500 fps with 512 live views that way
(scripts/verify-scale.mjs).
Hit-testing. On Chromium 158 a view reported with updateElementGeometry
is hit-tested where it was drawn — elementsFromPoint, hover and presses
land on the drawn view — so the browser answers and the canvas says
data-graview-hit="platform", which lets the views' content take the
pointer. Earlier builds (the layoutSubtree spelling, Chrome 155 with
capture switched on among them) accept the geometry and ignore it: a view's
box stays where it was laid out, so a press, a screen reader and an
automation all land somewhere other than where the view was drawn.
gpuCapturePresent() is false there and the page draws on the DOM path.
PointerRouter (data-graview-hit="router") remains for a host driving the
Compositor itself on such a build. A pan moves the canvas by a transform,
as it moves the DOM path's stage, and redraws nothing until it lands.
What is stable here
PLANE_STYLES,styleFor,mixStyles,transformFor— the plane model.planFrame— what to capture, what to draw, where each view lands.PointerRouter,hitTest— hit-testing a click against drawn geometry.
The compositor, the WGSL and the html-in-canvas platform bindings move with
the browser feature they are built on.