Research-2112: SYCL motion-add-UV fixed-point oracle — 2026-09-25¶

Status: Complete

Authority inspected: integration head 84f8ab38133325d90042329e89f2b24719d252e7, Intel oneAPI DPC++ 2026.0, Level Zero runtime 1.17.39758, Intel Arc A380.

Scope: motion_sycl(motion_add_uv=true) test arithmetic and the ADR-1206 960x540 registration. No production kernel, public API, model, snapshot, Netflix golden assertion, benchmark, tuning, or training change.

Reproduction¶

The original test used one absolute 2e-4 budget between CPU float_motion and fixed-point SYCL motion_sycl. Rebuilding the same source at 960x540 and running it on the idle Arc A380 reproduced the tracked failure:

cpu(float_motion)=50.27938271
sycl=50.27915223
delta=2.30e-04
tol=2.00e-04
exit=1

The small 256x144 build passed. The failure was deterministic and exercised the exact source and device path named by T-SYCL-MOTION-ADD-UV-TOLERANCE-RESOLUTION-2026-09-06.

Arithmetic inventory¶

The two extractors implement different numerical contracts:

The semantic option is shared, but CPU float is therefore not an arithmetic-identical oracle. A fixture-calibrated cross-extractor tolerance cannot distinguish an expected float reduction residual from coefficient, border, rounding, UV inclusion, or normalization drift in the device kernel.

Resolution-independent bound¶

The scalar test oracle independently implements the fixed coefficients, reflect-101 index mapping, both rounding stages, exact per-plane SAD, and YUV420 ceiling geometry. It calculates both Y-only and Y+U+V scores.

For an 8-bit input, a fixed blurred sample is at most 255 * 256 = 65280. The public picture allocator limits each side to 32768, so a plane SAD is less than 65280 * 32768 * 32768 < 2^46; conversion to binary64 is exact. Division by 256 is also exact. The combined score has five rounded operations: one area division for each of Y, U, and V, followed by two additions. With u = DBL_EPSILON / 2 and gamma_5 = 5u / (1 - 5u), two independent host evaluations differ by at most:

2 * gamma_5 * max(1, |expected|)

This bound depends on normalized score magnitude, not pixel count. At 960x540 the oracle values are 24.028187542197145 for Y and 50.279152229214894 for Y+U+V; the corresponding bounds are 2.668e-14 and 5.582e-14. The unmodified device path matches both.

Hypotheses and falsification¶

1. Wrong oracle, correct kernel. Confirmed: fixed oracle and device agree at both fixture sizes while float and fixed differ at the large size.
2. Coefficient or rounding mismatch. Falsified on the unmodified source by the exact-oracle comparisons.
3. 960x540 tail or mirror defect. Falsified by the large exact-oracle run; its non-multiple-of-16/32 geometry agrees.
4. Atomic reduction instability. Falsified by ten serial hardware test executions (five per fixture size), all passing with the unchanged result.

Red cap¶

The regression was planted by changing only the first production coefficient from 3571 to 3572 while leaving the oracle untouched. The 960x540 test failed:

y delta=1.581e-03, bound=2.668e-14
uv delta=3.291e-03, bound=5.582e-14

The coefficient was restored with no production diff. This proves the test is sensitive to fixed arithmetic drift rather than merely confirming that UV adds a positive value.

Verification¶

• RED: old 960x540 float-vs-fixed test, exit 1 at 2.30e-4.
• GREEN: named 256x144 and registered 960x540 Meson tests on Arc A380.
• RED CAP: one-unit coefficient mutation rejected by the large test.
• The large variant remains in the fast, gpu, and sycl suites and is globally serialized by the existing GPU-test rule.
• The C test is clean under clang-format 22 and strict clang-tidy 22; the changed Meson file configures and builds both targets. Documentation, ADR index, changelog, rebase-note, source-citation and state-row gates pass.

Result¶

The tracked gap was a test-oracle defect, not a production-kernel defect. test_sycl_motion_add_uv_parity now validates the actual fixed-point contract under a derived resolution-independent host-roundoff bound, and the 960x540 variant is registered instead of excluded.