These figures are regenerated by generate_gallery.py,
which reads the archived CSVs directly. Each plot describes native padded
stereo (4352 × 2176), synthetic camera motion at 90 Hz, and offscreen parser,
upload, and approximate PLANAR rendering on Pico 4. The measurements exclude
PC encoding, network, compositor, display scanout, and physical pose input.
An interval is a scheduling interval; it is not latency.
Empirical completion-time distributions for the six single-pass arms, with the 11.111 ms 90 Hz deadline as a reference.
Per-frame completion traces for the paired single-pass and full-draw arms.
Median and p99 measured parse, upload, and GPU execution components. Fence wait is omitted because it overlaps GPU execution and adding both would double-count wall time.
Exact integer counts of frames above 11.111 ms for the six final comparison arms.
Per-frame skipped-tile fraction and oldest retained tile age for single-pass and multi-pass arms.
Historical v1 lifetime-maximum, v1 forced-2 ms, and v2 expiring-estimate traces. These controls document starvation behavior and are not production claims.
Artificial 2 ms and 3 ms admission-budget pressure controls. The 90 Hz deadline is not used as the budget threshold in this figure.
These are lossless offscreen GPU readbacks or capture frames, linked from the existing benchmark evidence:
- Single-pass all-prefix frame
- Centre-only frame 19
- Historical motion proof, yaw 0°
- Historical motion proof, frame 75, yaw ≈ +60°
- Historical motion proof, frame 225, yaw ≈ −60°
The capture links are visual evidence only; they do not establish live 240 Hz operation or end-to-end display latency.
Regenerate from the repository root with:
python3 docs/visual-results/generate_gallery.pyThe separate archives keep final paired runs, historical scheduler controls, and new artificial-pressure tests distinct. Run labels can repeat across archives, so they must never be merged by name alone.