perf(adm): split adm_p_norm==3.0 fast-path (ADR-0463)

CHANGELOG.md

@@ -101,6 +101,15 @@
   [the HIP backend guide](docs/backends/hip/overview.md#measured-on-a-gfx1036-2026-10-01).
 
 
+- **Restore `adm_sum_cube_s_p3`, `adm_csf_den_scale_s_p3`, and `adm_cm_s_p3` fast-path
+  functions in `adm_tools.c` (ADR-0463 / BUG-048 B3).**
+  The specialized `adm_p_norm == 3.0` fast-paths eliminate all per-pixel `powf()`
+  calls and branch overhead on the hot path for default VMAF evaluation.
+  Scores remain 100% bit-identical to the baseline generic path across all 48
+  frames on the Netflix 576x324 reference pair at `--precision max`.
+  Dispatched once per scale in `core/src/feature/adm.c`.
+
+
 - **`vmaf` reads its two inputs ahead of scoring, on one thread each
   (ADR-1366).** The CLI used to read the reference frame, then the distorted
   frame, then score the pair, all on one thread; at 3840x2160 the two reads

changelog.d/changed/perf-adm-p3-fast-path.md

@@ -0,0 +1,7 @@
+- **Restore `adm_sum_cube_s_p3`, `adm_csf_den_scale_s_p3`, and `adm_cm_s_p3` fast-path
+  functions in `adm_tools.c` (ADR-0463 / BUG-048 B3).**
+  The specialized `adm_p_norm == 3.0` fast-paths eliminate all per-pixel `powf()`
+  calls and branch overhead on the hot path for default VMAF evaluation.
+  Scores remain 100% bit-identical to the baseline generic path across all 48
+  frames on the Netflix 576x324 reference pair at `--precision max`.
+  Dispatched once per scale in `core/src/feature/adm.c`.