Native video, decoded on the GPU.

Every measured process cleared source cadence

All six admitted streams decoded, converted, and made 288 frames application-ready. The slowest observed process still delivered 1.86× the 60 fps source cadence at 2160p and 3.73× at 1080p. Hardware counters confirmed the selected YCbCr path in every sample, while the forced-compute samples recorded zero hardware conversions and every row recorded zero CPU pixel readback.

These are unpaced saturation measurements: they establish decode-to-RGBA capacity, not compositor presentation deadlines, end-to-end playback pacing, or single-frame latency.

Decode and conversion capacity

Each cell is the complete observed range across both seven-process execution-order campaigns. The final column uses the slowest value from either conversion path.

Hardware YCbCr and manual compute are both viable

Each fresh process primed one path with 120 untimed frames, rewound the same decoder, then measured 288 frames on each path. A second seven-process campaign reversed the order. Ratios divide hardware-path throughput by manual-compute throughput, so values above 1.0 favor hardware.

No row is promoted as a stable speedup or regression. All twelve order-specific campaigns exceed the predeclared 15% full-spread gate. The paired medians place hardware YCbCr between 0.935× and 1.152× of manual compute, but those values remain characterization rather than a device-selection policy.

Codec and profile coverage

Codec names do not bypass capability admission. OA queries the live device and accepts only the profiles, chroma formats, bit depths, levels, and picture features the complete path can support honestly.

The sustained matrix is intentionally 8-bit. Native 10-bit paths are verified through P010 code-value differentials against FFmpeg and CPU color oracles; this checkpoint does not claim sustained HDR or 10-bit playback performance.

Measured-device limits remain explicit

Correctness and safety gates

Measurement provenance

Earlier timing from the session was rejected after concurrent VLC UHD 10-bit playback was disclosed. The published replacement used an uncontended clean detached worktree, primed both paths, retained both execution orders, and reports the remaining spread.

Evidence boundary

This checkpoint qualifies native decode and decode-to-RGBA conversion on one Intel Iris Xe configuration. It is not evidence for remote desktop, compositor scanout, encode, cross-vendor support, universal codec availability, single-frame latency, or universal performance. Runtime capability queries remain authoritative on every device.

Reproduce

Smoke the complete codec matrix before timing. Repeat both pair orders for1080p60-h264, 1080p60-h265, 1080p60-av1,1080p60-vp9, 2160p60-h265, and 2160p60-av1. Require a passing test, a positive hardware-conversion count only on the hardware path, and pixel_readback_bytes=0. A failed 15% spread gate stays characterization.

cmake --build build/release --target BenchVideoRealtime TestVideoDecoder -j
./bin/release/test/vision/codec/testVideoDecoder

OA_VIDEO_BENCH_VARIANT=2160p60-av1 \
OA_VIDEO_BENCH_PAIR_ORDER=hardware-first \
OA_VIDEO_BENCH_PRIME_FRAMES=120 \
./bin/release/test/vision/codec/benchVideoRealtime \
  --gtest_filter=BenchVideoRealtime.DecodePresentYcbcrPair

# Publication: wrap the command with tools/diagnostics/oaBench.py.
# Use one excluded warmup, seven measured fresh processes, 5 s cooldown,
# a 65 C package start gate, and the hardware_over_compute metric.
# Repeat both pair orders for every admitted variant.