How rendering works
The loop
Every binding runs the same loop:
if chart.needs_redraw() != Invalidation::None {
chart.build_frame(now_ms, &renderer.measurer()); // &mut chart, ends here
renderer.render(chart.frame(), &chart.data_view())?; // two shared borrows
}needs_redraw stays None while nothing changes, so an idle chart costs
nothing. Building a frame takes about 20 µs, independent of the data size.
Frames
A Frame has three layers — background (grid), main (series, axes) and overlay
(crosshair) — each a flat list of draw commands with a revision number.
Series are not expanded into rectangles: the frame says draw this series,
visible range i..j, with this transform, and the renderer reads the data
columns directly.
When only the crosshair moves, only the overlay's revision changes and the GPU renderer reuses a cached texture of the other two layers.
Data on the GPU
The wgpu renderer keeps series data in GPU buffers, in chunks of 4096 bars. It uploads a chunk the first time it is drawn and afterwards only the ranges that changed. Panning and zooming change a few numbers per chunk; nothing is re-uploaded.
Prices are stored relative to a reference price per chunk, so 32-bit floats on the GPU keep full precision even for large prices with small ticks.
Same pixels everywhere
The coordinate math lives in ferrochart_core::transform_helpers and is
mirrored operation by operation in the WGSL shaders, so the CPU and GPU
renderers round identically. Golden-image tests compare both against the same
reference images. (One subtlety: GPU compilers may fuse a * b + c into a
single instruction with different rounding, so the transforms are written as
(a + b) * c, which cannot be fused.)