Direct RGBA receive: packet-copy experiment

8 September 2026 · Apple M2 · 8 logical CPUs · 24 GiB memory · Darwin 23.6.0

Method

This report measures an experimental direct RGBA receiver. Native camera publication remains active. Noisedeck B receives packets directly and does not capture the camera.

The source uses a 1920×1080 checkerboard and an identity marker. The receiver uses ordinary media rendering through WebGL2. The transport permits one outstanding frame.

The controller requests five seconds of warmup and 30 minutes of measurement. Input and final markers identify each selected render.

The packet-copy kernel checks alpha bytes with word loads, then uses memcpy for the payload. It retains the original scalar path for overlapping buffers and failed alpha checks. The receiver retains one outstanding frame.

The browser is Chrome/152.0.7977.83. The powered baseline used the same browser version. The runs occurred at different times. They do not estimate repeatability or isolate all host conditions.

Measurement results

The first table reports final observations. Input and final callbacks can fall in different seconds. Their complete-second counts remain separate.

The run passed diagnostic and cleanup checks. The average 50-frame checkpoint is true. The complete-second 50-frame checkpoint is false. The report makes no sustained 60-frame acceptance claim.

Complete seconds and the partial tail remain separate below. The average uses the actual duration and includes the tail. Sequence gaps total 3359. They do not identify where a packet was lost.

Marker age

Marker age uses the decoded source wall-time stamp and the observation wall clock. Final age ends when the read starts. Read completion time includes later callback scheduling. These values do not measure physical display output.

Native elapsed times

These counts cover the native recorder lifetime, including setup, warmup and cleanup. The elapsed span is 1807.379 seconds. Native time has no measured alignment bound to B, so the native chart has a separate origin.

Ingest gap measures time between checked ingress callbacks. Tap copy and camera publication measure synchronous calls. Write submit measures the uv_write call. Write callback and ACK callback measure elapsed time from that lease’s submission. These are wall durations, not CPU or network durations.

Power samples

Each recorded sample reports AC power. Samples occur before measurement, during progress and after measurement. They do not prove continuous AC power or the state of browser Energy Saver between samples.

Pixel controls and limits

Three positive full-pixel gates and three poisoned-negative gates passed in this context. Each gate compares 8,294,400 bytes per surface. The timed run compares markers, not every pixel.

The fixture includes changes that the released product does not contain. These results do not establish released-product performance, external camera delivery or performance over several hours. This report does not estimate statistical uncertainty or prove a cause for earlier stalls.

Hardware metadata comes from the retained record for this host. This run does not measure process memory or CPU use. Browser telemetry retains 2506 slow events and omits 0. Per-cell totals remain available.

Source references

Allowlisted measurement data · Artifact manifest · SHA-256 checksums