Skip to content

ADR-1001: SYCL parity round 5 — CAMBI CPU vs. SYCL parity gate

  • Status: Accepted
  • Date: 2026-06-04
  • Deciders: Lusoris
  • Tags: sycl, test, gpu, parity, kernel-coverage, cambi, fork-local

Context

ADR-0957 (round 4) claimed "16 of 18 registered SYCL extractors have an active CPU vs. SYCL parity test", treating the two dormant SpEED twins as the only gap. On inspection, integer_cambi_sycl.cpp was counted as covered by test_integer_cambi_sycl.c, but that file explicitly states: "This is NOT a numerical-correctness test." It verifies only registration + finite/non-negative output, not that the SYCL kernel produces scores within ADR-0214 tolerance of the CPU scalar path. The true active coverage was therefore 15 of 18, with cambi_sycl being the silent gap.

CAMBI feeds the vmaf_v0.6.1neg and vmaf_4k_v0.6.1 model lineages. A silent SYCL drift in the multi-scale banding detector would bias every CHUG re-extract that uses those models on Intel Arc without triggering any CI gate.

Decision

Add core/test/test_sycl_cambi_parity.c — a CPU vs. SYCL parity test for cambi_sycl using a 256x256 quantised-gradient fixture (above the 216x216 CAMBI_MIN_WIDTH_HEIGHT minimum) at places=4 (1e-4) tolerance per ADR-0214. Register it in core/test/meson.build inside the existing if get_option('enable_sycl') block, suite ['fast', 'gpu']. The test skips with a [skip: no SYCL device] message when no oneAPI runtime is present, matching the pattern from all prior rounds.

Alternatives considered

Option Pros Cons Why not chosen
Elevate test_integer_cambi_sycl.c to a parity test Single file, no extra meson entry The smoke test has a different scope documented in its header; changing it in place would confuse future readers about when the parity assertion was added and by which PR New file keeps the audit trail cleaner
Raise tolerance to 5e-3 (ssimulacra2 precedent) Simpler uniform rule CAMBI is integer arithmetic; empirically bit-tight. 1e-4 is the correct ADR-0214 default —

Consequences

  • Positive: cambi_sycl now has an active CI-gated parity gate alongside all other registered SYCL extractors. ADR-0957 counter corrects to 16/18 active + 2 dormant SpEED, i.e. cambi_sycl is no longer an uncovered gap.
  • Negative: minor CI cost increase — one additional GPU test per SYCL-enabled run. Fixture is 256x256 and single-frame; wall-clock impact is negligible.
  • Neutral: test_integer_cambi_sycl.c retains its smoke-test role unchanged; this ADR adds a complementary numerical gate, not a replacement.

References

  • ADR-0214 — GPU parity tolerance table (FEATURE_TOLERANCE).
  • ADR-0957 — SYCL kernel coverage round 4 (where the CAMBI gap was first mis-counted as covered).
  • ADR-0884 — SYCL kernel coverage round 2 (template for this series).
  • req: the user requested parity coverage for all missing SYCL kernels in the round-5 task brief; CAMBI was listed among the suspected gaps.