Skip to content

Research-2047: Option-aware feature context registration

Reproduction and scope

Base: 2143174c7bf933f97aec610f423c559a885b8e9a. Both vmaf_use_feature() and vmaf_use_features_from_model() construct contexts with parsed options before calling feature_extractor_vector_append(). The old vector compares advertised base feature names and destroys the second context before considering its options. Real motion_force_zero=false and true contexts generate different SAD keys, but registration returns success and keeps only one. The existing 18-case feature-extractor suite passes because its twin-overlap test uses equivalent defaults.

This is an active registration defect, independent of GPU availability. The public regression registers both motion variants, extracts a blank frame, flushes, and reads both VMAF_integer_feature_motion2_score and integer_motion2_force_0. The existing name builder aliases the nondefault motion2 key; no naming or numerical formula changes are part of this fix.

Correction and invariants

For each shared advertised base, compare canonical keys generated from each extractor's parsed feature parameters. Preserve ADR-0385's first shared match and first-registration ownership, including equivalent CPU/GPU twins. The existing absent-list fallback still compares extractor-name/options keys. Defaults and aliases normalize through the same existing name builder that extractors use to populate their feature-name dictionaries.

Comparison strings have local RAII ownership with the C allocator's deleter. Failure to allocate either returns -ENOMEM, leaving the incoming context with the caller. Pointer-table doubling checks both its unsigned capacity and size_t byte limit before arithmetic or allocation. Failure leaves the existing pointer, count, capacity and contexts unchanged. The same helper is tested directly at its representable bounds; no huge allocation is required.

The published pointer array still uses malloc/realloc/free, preserving the existing C-visible layout, order and normal doubling. The prologue's obsolete std::vector/try/catch claims were removed. Its five no-malloc markers suppressed a check not enabled by the current profile and are removed. No public declarations, feature calculation, callback type or Netflix golden assertion changes.

Regression controls

  • Fourteen private Linux cases: distinct real motion options; explicit defaults; canonical/alias equality; identical custom options; equivalent and distinct mock backend twins; absent-list fallback; count/byte boundary arithmetic; real 8→16→32 growth; both comparison allocation failures and actual realloc failure with retry. Portable hosts run the eleven cases that do not depend on GNU link wrapping.
  • One public case lives in a separate executable linked only to libvmaf. A combined private/public executable can export internal vector symbols and interpose on an ELF shared library, invalidating an old-library control. The separate target avoids that test-harness defect.
  • GNU __wrap_* entry points exist only in the Linux test target. Their exact declaration markers preserve the linker ABI; their unusedFunction markers cover entry points invoked by linker substitution. No production test hook, diagnostic category exclusion or global suppression is introduced.
  • Existing test_feature_extractor remains unchanged with all 18 cases.

Validation receipts, exact commands, source hashes, failed harness iterations and limits are retained under .workingdir2/evidence/2026-09-08-fex-context-vector/. Disposable builds and logs are under .workingdir2/cache/fex-context-vector-20260908/.

Validation results

  • GCC 15.2 release CPU build with default LTO: 14 private cases, one public shared-library case, and the unchanged 18-case feature-extractor suite pass.
  • ASan/UBSan with leak detection: the same 33 cases pass in a separate debug build with LTO disabled for instrumentation.
  • Exact base-vector controls fail at the intended checks: private registration count and public option-specific score retrieval. The shared-only executable resolves the retained old library through LD_LIBRARY_PATH; its loader receipt is preserved alongside the full link commands.
  • clang-tidy 22.1.8: production warnings 3→0, uncited markers 5→0; both new test TUs and the private helper are clean. Cppcheck 2.21.1 with the official POSIX model, all checks and exhaustive configured analysis exits zero for all touched native sources and their real test caller. The scoped writer tightens only the measured production allowance and registers the clean tests.
  • A 32-bit compile check was attempted but this image lacks the 32-bit libstdc++ bits/c++config.h. No 32-bit execution is claimed; native boundary tests cover the actual host limits, and the source checks byte and count limits independently before doubling. No toolchain was downloaded.

Reproduce

In a configured CPU build with generated dependencies available:

ninja -C build test/test_feature_extractor test/test_fex_ctx_vector test/test_fex_ctx_vector_public
meson test -C build --no-rebuild --print-errorlogs \
  test_feature_extractor test_fex_ctx_vector test_fex_ctx_vector_public

Use the retained old-object/old-library adapter for negative controls: it compiles the exact base vector with current compatible build flags, then substitutes only that vector in a disposable executable/shared library. Normal insertion and numerical kernels remain from the configured build.

No new ADR is needed: this restores option identity while preserving the existing ADR-0385/ADR-0723 registration and ownership design. Alternatives such as comparing raw dictionary spelling or backend-name prefixes would misidentify defaults/aliases or require a separate backend list.

Limits

CPU-only tools container; no GPU device/runtime acceptance, native Windows acceptance or full-tree lint claim. SIMD and numerical kernels are unchanged. The parent integration owns complete native and golden gates. The separate model-registration dictionary failure path remains outside this commit.