GPU-driven rendering: merge GPU buffers to eliminate per-VItem CPU submission overhead

[AzurIce]

Background

The current rendering pipeline maintains separate GPU buffers and bind groups for each VItem. Every frame requires per-VItem calls to write_buffer, set_bind_group, and draw.

Benchmarks from benches/benches/gpu_render.rs on the main branch show that CPU submission is the dominant bottleneck, not the GPU:

VItem count	GPU total (submit+wait)	CPU submit (no GPU wait)	GPU pure (diff)
25	5.6 ms	1.6 ms	~4.0 ms
100	8.9 ms	4.1 ms	~4.8 ms
400	22.3 ms	25.2 ms	— (CPU already exceeds GPU)
1600	88.8 ms	92.7 ms	—
3600	256 ms	220 ms	~36 ms

CPU cost scales linearly at ~55μs/VItem. At 1600 VItems, CPU accounts for ~89% of total frame time. Root cause: 3600 VItems × 7 write_buffer calls = 25,200 API calls, each with fixed overhead (lock, validation, staging allocation).

Approach

Adopt a GPU-driven rendering strategy — merge all VItem data into a single set of contiguous GPU buffers:

1. Every-frame rebuild: each frame, pack all VItem points, fill_rgbas, stroke_rgbas, and stroke_widths into large contiguous buffers, with an ItemInfo index table recording each item's offset and count
2. Instanced drawing: use draw(0..4, 0..N) to render all N VItems in a single draw call; the vertex shader uses instance_index to look up per-item clip box and plane data
3. Binary search in compute shader: each compute thread binary-searches item_infos to determine which item it belongs to, then performs 3D→2D projection and atomic clip box updates

This reduces per-VItem O(N) draw calls / bind group switches / write_buffer calls to O(1).

Related

• PR: Experimental GPU-driven rendering #138