Skip to content

SYCL integer VIF warning cleanup — 2026-09-21

Finding and scope

core/src/feature/sycl/integer_vif_sycl.cpp hid strict-analysis debt behind file-wide and function-level NOLINT ranges. Removing those suppressions exposed three HISS-04 overlength phases, external-linkage findings for every private helper, designated-initializer ordering warnings, and five forced-loop unroll requests that oneAPI 2026 could not honour for every supported device target.

This is a fork-local implementation cleanup under ADR-0141 and ADR-1142. It changes no feature name, option, public header, metric definition, tolerance, golden assertion, build option, or FFmpeg patch surface. No ADR is needed: there is no policy choice or user-visible behaviour change, and the only acceptable implementation removes the diagnostics while preserving the integer arithmetic and callback contract.

Preservation contract

  • Keep every device-kernel operand fp64-free. The host stores the configured gain as double, then narrows it to float before kernel dispatch; score aggregation remains host-side double.
  • Preserve tap order, multiply/add widths, rounding constants, shift order, mirror indexing, subgroup reduction order, and sycl::fmin gain limiting.
  • Preserve ceiling downsample stride (width + 1U) / 2U and matching device allocation for odd intermediate widths.
  • Preserve separate and fused execution modes, SIMD-16/SIMD-32 selection, graph callback ordering, allocation cleanup, and feature-name derivation.
  • Keep the Netflix copyright notice and all Netflix golden assertions untouched.

The vertical convolution now has a typed accumulator phase. The horizontal kernel has a work-item phase with a trivially copyable launch-parameter record. Initialization is split into resource, device-selection, and graph-registration phases while retaining the original cleanup points. Short anonymous namespaces provide C++ internal linkage without a blanket suppression and without forming an overlength HISS scope. Failed #pragma unroll requests are removed; the compiler remains free to unroll profitable target-specific instances.

Validation

The owned source was checked without a suppression, baseline, threshold, or golden-data change:

  • praetorctl audit -base e0d31c5bcb43a7aab7fb845a592488635b061566 -touched core/src/feature/sycl/integer_vif_sycl.cpp reports the touched file clean and the repository ratchet unchanged at 1411.
  • scripts/ci/clang-tidy-sycl.sh -p core/build-sycl-vif core/src/feature/sycl/integer_vif_sycl.cpp --quiet reports zero diagnostics in the owned source under the full configured profile.
  • An independent oneAPI 2026.0 icpx compile using -fsycl-targets=spir64_gen,spir64, -fp-model=precise, and the current -Xsycl-target-backend=spir64_gen spelling exits zero with zero diagnostics.
  • On an Intel Arc A380 that reports no native fp64, test_sycl_vif_parity, test_sycl_vif_parity_large, test_vif_skip_scale0, and test_vif_lifecycle pass (4/4).
  • Relinking the same build with the untouched base-commit object gives exact full-precision equality for default and fused modes: 720/720 emitted metrics each on the 576x324 8-bit 48-frame fixture, and 45/45 each on a deterministic 258x146 10-bit three-frame fixture whose later scales have odd dimensions. Maximum absolute delta is zero in all four comparisons.

The existing parity and lifecycle tests already exercise the public seam, so no expected value or duplicate test harness is added. The local dev-llm-review command was also attempted as a read-only second opinion, but the installed repository virtualenv has a broken entry point (ModuleNotFoundError: vmaf_dev_llm.cli); no local-model claim is included.