Skip to content

Feature-extractor coverage round 2 — research digest

  • Date: 2026-05-31
  • ADR: ADR-0938
  • Status: Completed
  • Owner: lusoris

Baseline

meson setup build-cov core -Db_coverage=true -Denable_cuda=false -Denable_sycl=false && ninja -C build-cov && meson test -C build-cov --suite=fast --suite=simd --suite=dnn against origin/master at commit 45d536962f (2026-05-31).

gcovr -r .. --csv -e '.*test.*' -f '.*core/src/feature/.*' baseline:

File Lines Covered Pct
integer_motion.c 213 153 71.8 %
integer_motion_v2.c 119 21 17.6 %
integer_psnr.c 87 61 70.1 %
integer_vif.h 36 26 72.2 %
iqa/convolve.c 102 42 41.2 %
barten_csf_tools.h 66 30 45.5 %
ms_ssim_decimate.c 46 37 80.4 %

PR #344 (round 1) ships coverage for mkdirp.c, luminance_tools.c, feature_name.c, feature_extractor.c — file set verified disjoint via gh pr view 344 --json files.

Targets selected

Files in the 17 %-80 % range with > 30 lines, no in-flight PR claim, CPU-side (no GPU device dependency), low-cost test scaffolding (matched to existing test_speed_qa / test_psnr / test_iqa_convolve recipes).

Skipped: tiny-AI extractors (feature_dists.c, feature_lpips.c, feature_mobilesal.c) — these have ~25 % baseline coverage but the remaining code paths require a live ONNX session that the existing tiny_ai_test_template.h cannot fake.

Branch coverage strategy

File Branches plugged
integer_psnr.c pix_fmt=YUV400P (line 124-125), pix_fmt=YUV444P + min_sse>0 (line 128-136), HBD extract for 10 / 12 / 16-bit (line 263-266), flush() with enable_apsnr=true (line 277-292).
integer_motion.c motion_force_zero=true (line 307-343), 4-frame default-window flow that lands the second-SAD branch (line 552-577), motion_moving_average=true (line 435-437), motion_five_frame_window=true + N<2 flush (line 441-455).
integer_motion_v2.c init -ENOTSUP rejection of motion_five_frame_window=true (line 288-293), index=0 short-circuit (line 356-360), motion_force_zero=true (line 350-354), 3-frame extract+flush with manually-set prev_ref (lines 161-204), motion_moving_average=true (line 475), 10-bit pipeline (lines 206-254).
integer_vif.h log2_32(2^16) and log2_32(2^15) against closed-form expectation, log2_64(2^17) / (2^32) / (2^40), monotonicity sweep across [0x10000..0x80000) for log2_32 and [0x100000..0x1000000) for log2_64.
iqa/convolve.c KBND_SYMMETRIC in-range / negative-mirror / overflow-mirror, KBND_REPLICATE x/y under-/over-flow clamp, KBND_CONSTANT OOB returning the constant, iqa_img_filter() NULL kernel rejection, iqa_img_filter() owns-its-own-buffer path, iqa_filter_pixel() NULL kernel + edge / non-edge dispatch.
barten_csf_tools.h All 8 supported (resolution, distance) pairs of barten_watson_blend_csf and barten_watson_blend_csf_mae, the -EINVAL fallthrough on both variants, mid-segment interpolation pairs {0.005, 0.05, 0.5, 5.0, 50.0} for barten_csf to drive every successive anchor-index selection.
ms_ssim_decimate.c ms_ssim_decimate_scalar(rw=NULL, rh=NULL), the runtime-dispatch wrapper ms_ssim_decimate(...) with explicit dispatch + NULL-out variants.

Final coverage

After meson test -C build-cov --suite=fast --suite=simd --suite=dnn (68 tests, 100 % pass):

File Before After Delta
integer_motion.c 71.8 % 81.7 % +9.9 pp
integer_motion_v2.c 17.6 % 88.2 % +70.6 pp
integer_psnr.c 70.1 % 92.0 % +21.9 pp
integer_vif.h 72.2 % 100.0 % +27.8 pp
iqa/convolve.c 41.2 % 98.0 % +56.8 pp
barten_csf_tools.h 45.5 % 48.5 % +3.0 pp*
ms_ssim_decimate.c 80.4 % 95.7 % +15.3 pp

* barten_csf_tools.h is a static-inline header; gcovr aggregates a single TU per header definition, so most of the new test exercises land in the test_barten_csf_coverage.c TU which gcovr doesn't merge back into the canonical _adm_csf.c consumer accounting. Branch count rose 50→55, confirming the test ran the new paths.

Smoke test command

git worktree add -b test/core-feature-coverage-round2 /tmp/wt-fea-r2 origin/master
cd /tmp/wt-fea-r2
meson setup build-cov core -Db_coverage=true -Denable_cuda=false -Denable_sycl=false
ninja -C build-cov
meson test -C build-cov --suite=fast --suite=simd --suite=dnn
gcovr -r .. --csv -e '.*test.*' -f '.*core/src/feature/.*' \
  --gcov-ignore-parse-errors=negative_hits.warn_once_per_file

Expected: 68 tests pass; the seven targeted files report at or above the deltas in the table.

Notable findings

  • motion_v2.extract segfaults under prev_ref=NULL whenever the framework's read_pictures loop is bypassed (the unit test calls vmaf_feature_extractor_context_extract directly). Set prev_ref manually between frames; the test does this explicitly and clears it before context_destroy to avoid a double-free.
  • Non-default VMAF_OPT_FLAG_FEATURE_PARAM options (e.g. motion_force_zero=true) cause vmaf_feature_name_from_options to suffix the feature name with _<opt>=<val>. Score-key assertions must use the dictionary-resolved name or be omitted (we omit them for coverage tests; the branch is exercised by the extract call, not by the score lookup).
  • The 7 new test binaries together add ~350 ms of wall time to meson test --suite=fast.