feat(adm): make the x86 float ADM wavelet and CSF kernels exact and dispatch them (ADR-1473)

.standards-baseline.json

@@ -1,9 +1,9 @@
 {
   "version": 1,
-  "generated_at": "2026-10-02T17:17:10Z",
+  "generated_at": "2026-10-02T17:32:39Z",
   "repository": "VMAFx/vmafx",
-  "commit_sha": "baa4abfeba3ca277a16f601ad92977f748f8d9c6",
-  "total_infractions": 231,
+  "commit_sha": "58376c70f3bbaad0d4ccd710b7ffeed31a4e7d31",
+  "total_infractions": 229,
   "infractions": [
     {
       "rule_id": "HISS-04",
@@ -674,22 +674,6 @@
       "message": "process::abort ends the process from library code; return an error to fn main instead",
       "fingerprint": "bindings/rust/vmafx-sys/src/safe.rs:339:HISS-07"
     },
-    {
-      "rule_id": "HISS-04",
-      "file_path": "compat/python-vmaf/__init__.py",
-      "line_number": 204,
-      "symbol": "call_vmafexec_multi_features",
-      "message": "Function 'call_vmafexec_multi_features' (77 LOC) exceeds HISS-04 / NASA Rule 4 limit of 60 LOC",
-      "fingerprint": "compat/python-vmaf/__init__.py:204:HISS-04"
-    },
-    {
-      "rule_id": "HISS-04",
-      "file_path": "compat/python-vmaf/__init__.py",
-      "line_number": 303,
-      "symbol": "call_vmafexec",
-      "message": "Function 'call_vmafexec' (118 LOC) exceeds HISS-04 / NASA Rule 4 limit of 60 LOC",
-      "fingerprint": "compat/python-vmaf/__init__.py:303:HISS-04"
-    },
     {
       "rule_id": "HISS-04",
       "file_path": "compat/python-vmaf/core/cross_validation.py",

CHANGELOG.md

@@ -450,6 +450,22 @@ make `core/AGENTS.md` a generated index over `AGENTS.d/` topic pages ([ADR-1454]
   on 71,840 inputs.
 
 
+- **`float_adm` uses AVX2 and AVX-512 on x86, with unchanged scores
+  (ADR-1473).** The wavelet and the contrast-sensitivity stage run through
+  kernels that return the scalar code's bits: every four-tap sum starts at
+  `+0` and multiplies before it adds, and the filtered CSF value is a double
+  product narrowed to float. `--cpumask` selects scalar (63), AVX2 (48) or
+  AVX-512 (0); every `float_adm` output and the float model scores are
+  identical on all three and identical to the previous release. One thread on
+  a Ryzen 9 9950X3D: 1.45 to 1.11 ms per 576x324 frame, 17.9 to 15.1 ms at
+  1920x1080, 76.2 to 62.9 ms at 3840x2160. The kernels existed but nothing
+  called them, and they differed from the scalar code (a negative zero where
+  it returns a positive one, a float product where it multiplies in double).
+  Two unused reduction kernels that could not reproduce the scalar sums
+  (`float_adm_csf_den_scale_avx2` / `_avx512`, `float_adm_sum_cube_avx2` /
+  `_avx512`) are removed.
+
+
 - **The cross-backend parity gate covers every registered CUDA, SYCL and HIP
   twin.** `speed_temporal` was the one feature whose three GPU twins were
   registered and compared by no gate cell. It is now a gate feature, with a
@@ -1292,6 +1308,15 @@ make `core/AGENTS.md` a generated index over `AGENTS.d/` topic pages ([ADR-1454]
   [ADR-1250](docs/adr/1250-eupl-fork-relicense.md)).
 
 
+- **`compat/python-vmaf/__init__.py` meets the HISS standard and carries its
+  SPDX line.** `call_vmafexec()` (118 lines) and
+  `call_vmafexec_multi_features()` (77 lines) are assembled from pure helpers
+  that each emit one part of the command; the command text is unchanged and
+  now pinned by two tests. HISS baseline: two rows fewer
+  ([ADR-1142](docs/adr/1142-whole-codebase-standards.md),
+  [ADR-1250](docs/adr/1250-eupl-fork-relicense.md)).
+
+
 - Added missing SPDX-License-Identifier declarations across 387 clean source
   and header files in accordance with ADR-1250 and repository provenance,
   skipping 131 files with baselined debt, 5 vendored Pelorus mirror paths,
@@ -2483,6 +2508,16 @@ make `core/AGENTS.md` a generated index over `AGENTS.d/` topic pages ([ADR-1454]
   [`psnr_hvs`](docs/metrics/psnr-hvs.md#cpu-instruction-sets).
 
 
+- **The Python harness accepts `motion_force_zero` with more than one model.**
+  `ExternalProgramCaller.call_vmafexec()` raised `AssertionError` for the
+  second model when `motion_force_zero=True`: the loop over the models
+  replaced the argument with the string `"true"` and then failed its own type
+  check (the same statement is in Netflix upstream). Every model now gets the
+  overload. A run with one model, including every golden test, produces the
+  same command as before
+  (`T-PYTHON-CALL-VMAFEXEC-FORCE-ZERO-SECOND-MODEL-2026-10-02`).
+
+
 - **`vmaf` no longer hangs after an out-of-memory on the device.**
   `vmaf_read_pictures()` kept the pair of pictures it was given when it failed
   before it reached an extractor (a non-increasing index, pictures that

README.md

@@ -184,7 +184,7 @@ generated agent context against `AGENTS.md`.
 private scratch-link policy.
 
 **Debt Baseline**: `.standards-baseline.json` anchors the debt ratchet at
-231 recorded infractions; audit forbids growth.
+229 recorded infractions; audit forbids growth.
 
 [praetor-docs-badge]: https://github.com/vmafx/vmafx/actions/workflows/praetor-docs.yml/badge.svg
 [praetor-docs-runs]: https://github.com/vmafx/vmafx/actions/workflows/praetor-docs.yml

changelog.d/changed/float-adm-x86-simd-exact-dispatched.md

@@ -0,0 +1,14 @@
+- **`float_adm` uses AVX2 and AVX-512 on x86, with unchanged scores
+  (ADR-1473).** The wavelet and the contrast-sensitivity stage run through
+  kernels that return the scalar code's bits: every four-tap sum starts at
+  `+0` and multiplies before it adds, and the filtered CSF value is a double
+  product narrowed to float. `--cpumask` selects scalar (63), AVX2 (48) or
+  AVX-512 (0); every `float_adm` output and the float model scores are
+  identical on all three and identical to the previous release. One thread on
+  a Ryzen 9 9950X3D: 1.45 to 1.11 ms per 576x324 frame, 17.9 to 15.1 ms at
+  1920x1080, 76.2 to 62.9 ms at 3840x2160. The kernels existed but nothing
+  called them, and they differed from the scalar code (a negative zero where
+  it returns a positive one, a float product where it multiplies in double).
+  Two unused reduction kernels that could not reproduce the scalar sums
+  (`float_adm_csf_den_scale_avx2` / `_avx512`, `float_adm_sum_cube_avx2` /
+  `_avx512`) are removed.

changelog.d/changed/std-python-harness-init.md

@@ -0,0 +1,7 @@
+- **`compat/python-vmaf/__init__.py` meets the HISS standard and carries its
+  SPDX line.** `call_vmafexec()` (118 lines) and
+  `call_vmafexec_multi_features()` (77 lines) are assembled from pure helpers
+  that each emit one part of the command; the command text is unchanged and
+  now pinned by two tests. HISS baseline: two rows fewer
+  ([ADR-1142](docs/adr/1142-whole-codebase-standards.md),
+  [ADR-1250](docs/adr/1250-eupl-fork-relicense.md)).

changelog.d/fixed/python-call-vmafexec-force-zero-second-model.md

@@ -0,0 +1,8 @@
+- **The Python harness accepts `motion_force_zero` with more than one model.**
+  `ExternalProgramCaller.call_vmafexec()` raised `AssertionError` for the
+  second model when `motion_force_zero=True`: the loop over the models
+  replaced the argument with the string `"true"` and then failed its own type
+  check (the same statement is in Netflix upstream). Every model now gets the
+  overload. A run with one model, including every golden test, produces the
+  same command as before
+  (`T-PYTHON-CALL-VMAFEXEC-FORCE-ZERO-SECOND-MODEL-2026-10-02`).

compat/python-vmaf/AGENTS.md

@@ -106,6 +106,20 @@ python/vmaf/
   `noref_feature_extractor` preserve reduction accumulation order; `local_explainer` /
   `nn_train_test_model` keep RNG draw order. `result.scores_key_wildcard_match()`
   doctests redistributed across helpers; `doctest.testmod()` finds all 12.
+- **`__init__.py` command builders are assembled from pure helpers
+  (`T-PYTHON-CALL-VMAFEXEC-FORCE-ZERO-SECOND-MODEL-2026-10-02`).**
+  `ExternalProgramCaller.call_vmafexec()` and `call_vmafexec_multi_features()`
+  keep their signatures; the command text comes from
+  `_vmafexec_base_command`, `_vmafexec_feature_flags`, `_vmafexec_model_flags`
+  (which calls `_vmafexec_model_overloads` once per model),
+  `_vmafexec_run_flags`, `_multi_features_run_arguments` and
+  `_feature_argument`. Keep the helpers pure: upstream's loop overwrote
+  `motion_force_zero` with the string `"true"` and then failed its own
+  `isinstance(..., bool)` assertion on the second model. An upstream change to
+  a flag goes into the helper that emits it; the order of the parts is pinned
+  by `test_full_command_is_pinned` and
+  `test_full_multi_features_command_is_pinned` in
+  `python/test/python_harness_coverage_test.py`.
 - **Bounded frame loops replace `while True` (HISS-02).** PyPSNR loop in
   `core/feature_extractor.py` walks `min(ref.num_frms, dis.num_frms)` (`YuvReader` validated);
   no `try/except StopIteration`. `tools/scanf.py` bounds loop in header; `readiter()`