fix(cuda): build float_ms_ssim's level 0 on the device instead of round-tripping every plane through the host

.config/lint-exceptions.d/clang-tidy-coverage.toml

@@ -63,66 +63,6 @@ path = "cmd/vmafx-node/bpf/rclone_bypass.bpf.c"
 reason = "eBPF program: includes vmlinux.h, generated from the running kernel's BTF and absent on every runner, so clang-tidy stops at the include; compiled by clang -target bpf through bpf2go."
 expires = 2027-03-31
 
-[[exception]]
-path = "compat/python-vmaf/matlab/STMAD_2011_MatlabCode/ical_stat.c"
-reason = "MATLAB MEX source: includes mex.h / matrix.h from the proprietary MATLAB SDK, present on no runner or image, so clang-tidy stops in the preprocessor; meson never builds it (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22 in docs/state.md)."
-expires = 2027-03-31
-
-[[exception]]
-path = "compat/python-vmaf/matlab/STMAD_2011_MatlabCode/ical_std.c"
-reason = "MATLAB MEX source: includes mex.h / matrix.h from the proprietary MATLAB SDK, present on no runner or image, so clang-tidy stops in the preprocessor; meson never builds it (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22 in docs/state.md)."
-expires = 2027-03-31
-
-[[exception]]
-path = "compat/python-vmaf/matlab/strred/matlabPyrTools/MEX/convolve.c"
-reason = "MATLAB MEX source: includes mex.h / matrix.h from the proprietary MATLAB SDK, present on no runner or image, so clang-tidy stops in the preprocessor; meson never builds it (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22 in docs/state.md)."
-expires = 2027-03-31
-
-[[exception]]
-path = "compat/python-vmaf/matlab/strred/matlabPyrTools/MEX/corrDn.c"
-reason = "MATLAB MEX source: includes mex.h / matrix.h from the proprietary MATLAB SDK, present on no runner or image, so clang-tidy stops in the preprocessor; meson never builds it (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22 in docs/state.md)."
-expires = 2027-03-31
-
-[[exception]]
-path = "compat/python-vmaf/matlab/strred/matlabPyrTools/MEX/edges-orig.c"
-reason = "MATLAB MEX source: includes mex.h / matrix.h from the proprietary MATLAB SDK, present on no runner or image, so clang-tidy stops in the preprocessor; meson never builds it (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22 in docs/state.md)."
-expires = 2027-03-31
-
-[[exception]]
-path = "compat/python-vmaf/matlab/strred/matlabPyrTools/MEX/edges.c"
-reason = "MATLAB MEX source: includes mex.h / matrix.h from the proprietary MATLAB SDK, present on no runner or image, so clang-tidy stops in the preprocessor; meson never builds it (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22 in docs/state.md)."
-expires = 2027-03-31
-
-[[exception]]
-path = "compat/python-vmaf/matlab/strred/matlabPyrTools/MEX/histo.c"
-reason = "MATLAB MEX source: includes mex.h / matrix.h from the proprietary MATLAB SDK, present on no runner or image, so clang-tidy stops in the preprocessor; meson never builds it (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22 in docs/state.md)."
-expires = 2027-03-31
-
-[[exception]]
-path = "compat/python-vmaf/matlab/strred/matlabPyrTools/MEX/innerProd.c"
-reason = "MATLAB MEX source: includes mex.h / matrix.h from the proprietary MATLAB SDK, present on no runner or image, so clang-tidy stops in the preprocessor; meson never builds it (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22 in docs/state.md)."
-expires = 2027-03-31
-
-[[exception]]
-path = "compat/python-vmaf/matlab/strred/matlabPyrTools/MEX/pointOp.c"
-reason = "MATLAB MEX source: includes mex.h / matrix.h from the proprietary MATLAB SDK, present on no runner or image, so clang-tidy stops in the preprocessor; meson never builds it (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22 in docs/state.md)."
-expires = 2027-03-31
-
-[[exception]]
-path = "compat/python-vmaf/matlab/strred/matlabPyrTools/MEX/range2.c"
-reason = "MATLAB MEX source: includes mex.h / matrix.h from the proprietary MATLAB SDK, present on no runner or image, so clang-tidy stops in the preprocessor; meson never builds it (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22 in docs/state.md)."
-expires = 2027-03-31
-
-[[exception]]
-path = "compat/python-vmaf/matlab/strred/matlabPyrTools/MEX/upConv.c"
-reason = "MATLAB MEX source: includes mex.h / matrix.h from the proprietary MATLAB SDK, present on no runner or image, so clang-tidy stops in the preprocessor; meson never builds it (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22 in docs/state.md)."
-expires = 2027-03-31
-
-[[exception]]
-path = "compat/python-vmaf/matlab/strred/matlabPyrTools/MEX/wrap.c"
-reason = "MATLAB MEX source: includes mex.h / matrix.h from the proprietary MATLAB SDK, present on no runner or image, so clang-tidy stops in the preprocessor; meson never builds it (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22 in docs/state.md)."
-expires = 2027-03-31
-
 [[exception]]
 path = "core/src/feature/metal/float_adm.metal"
 reason = "Metal Shading Language kernel. clang-tidy reads the languages the clang front end accepts; clang has no Metal language mode (`clang -x metal` answers \"language not recognized\" on clang 23.1.1, and the -x language list in https://clang.llvm.org/docs/ClangCommandLineReference.html names none; clang-tidy, https://clang.llvm.org/extra/clang-tidy/, runs on that front end). The kernels are compiled by Apple's `metal` compiler (-std=metal3.1) in the macOS Metal build leg and their arithmetic headers are held against the CPU by the device-free host-replay tests. Revisit when upstream clang gains a Metal mode."

.github/workflows/lint-and-format.yml

@@ -200,6 +200,7 @@ jobs:
           # exist on disk before clang-tidy parses the translation units.
           meson compile -C build
           python3 scripts/ci/write-compile-commands.py --build-dir build
+          python3 scripts/ci/gen-mex-compile-commands.py build
       - name: Run clang-tidy on changed files
         if: steps.cpp-changes.outputs.any == 'true'
         # Hardening per Semgrep `yaml.github-actions.security.run-shell-injection`
@@ -317,18 +318,12 @@ jobs:
           #     replay. `(^|/)testdata/` is already the tree-wide fixture
           #     exclusion in `.pre-commit-config.yaml`; this is the same rule
           #     applied to the changed-files tidy lane.
-          #   - compat/python-vmaf/matlab/ — upstream Netflix training-harness
-          #     MATLAB MEX sources. They include `mex.h` / `matrix.h` from the
-          #     proprietary MATLAB SDK, which does not exist on the runner, so
-          #     every one of them dies in the preprocessor
-          #     (`'mex.h' file not found`) and the diagnostics that follow are
-          #     artefacts of reading a `.c` file as C++ after the include
-          #     failed, not findings about the code. Same posture as the
-          #     CUDA/SYCL/HIP/eBPF families above, and the same exclusion these
-          #     files already carry in every hygiene hook in
-          #     `.pre-commit-config.yaml` (see the `check-copyright` entry).
-          #     The whole-tree ratchet does not measure them either — they have
-          #     no entry in `build/compile_commands.json`.
+          #   - compat/python-vmaf/matlab/ is NOT excluded: the MEX sources include the
+          #     MATLAB SDK's `mex.h` / `matrix.h`, which no runner has, so the step
+          #     `Generate compile_commands.json` appends entries that point at the
+          #     self-authored stubs in `scripts/ci/lint-stubs/matlab/`
+          #     (`scripts/ci/gen-mex-compile-commands.py`, the `cpu` lane does the
+          #     same; T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22).
           #
           # One filter shared by every trigger path below, so adding a family
           # is a single edit.
@@ -359,8 +354,7 @@ jobs:
               -e '^core/src/compat/win32/' \
               -e '^core/tools/test/test_vmaf_windows_utf8_argv\.cpp$' \
               -e '^cmd/vmafx-node/bpf/' \
-              -e '^\.config/hiss/testdata/' \
-              -e '^compat/python-vmaf/matlab/' |
+              -e '^\.config/hiss/testdata/' |
               python3 scripts/ci/pelorus_mirror.py filter
           }
           if [ "$GH_EVENT_NAME" = "pull_request" ]; then

CHANGELOG.md

@@ -550,6 +550,9 @@
   (the scaffold-audit P1 record) where they cited ADR-0613, the vmaf-tune dynamic optimizer.
 
 
+- ADR-0643, ADR-0665, ADR-0666 and ADR-0673 lose the unfilled allocator template block that preceded their real text (their first heading read `<fill in title>`); a status update records it. Four ADR numbers that other ADRs cite and that had no file (ADR-0228, 0636, 0867, 0979) get a short record each, written from the commits and ADRs that name them.
+
+
 - 34 ADR status headers read in the one form the drift gate parses (`- **Status**: Accepted`): 12 bullet variants and 22 table or heading headers (no value changed; empty `Supersedes` rows dropped). ADR-0003 and ADR-0019 link their successor ADR-1277. ADR-1129, ADR-1225 and ADR-0954 gain a dated status update: the pins they quote have moved (`build-config.env` is the authority) and the HIP dispatch strategy file was removed by #2030. `docs/state.md` cites ADR-0639 (scaffold-audit P1) where it said ADR-0613.
 
 
@@ -940,6 +943,13 @@ make `core/AGENTS.md` a generated index over `AGENTS.d/` topic pages ([ADR-1454]
   sweep).
 
 
+- **The CUDA MS-SSIM twin no longer copies frames through the host.**
+  `float_ms_ssim_cuda` converted every plane of every frame on the host
+  (a device-to-host copy, a wait, `picture_copy()` and an upload); the
+  conversion now runs on the device with the same arithmetic, so its scores
+  are unchanged and the frame stays on the GPU.
+
+
 - **The parity gate covers `speed_chroma` on CUDA, at `5e-6`.**
   `speed_chroma_cuda` reproduces the CPU extractor's arithmetic and rounds
   `log2` correctly; the CPU extractor calls the C library's `log2f`, and
@@ -2826,6 +2836,13 @@ make `core/AGENTS.md` a generated index over `AGENTS.d/` topic pages ([ADR-1454]
   scoped write under a clang-tidy other than the baseline's stops with exit 5.
 
 
+- The `cpu` clang-tidy lane and the changed-files job measure the ten MATLAB MEX
+  sources of `compat/python-vmaf/matlab/` against self-authored stub `mex.h` and
+  `matrix.h` (`scripts/ci/lint-stubs/matlab/`); their lint exceptions are removed.
+  The first measurement fixed the mechanical findings and a defect in `ical_std.c`
+  (`mxDestroyArray()` was called on a matrix's data pointer).
+
+
 - **The Metal host code is clean under clang-tidy.** The macOS `metal` lane measures 0 findings on its 27 translation units (it started at 1640 once `write-compile-commands.py` stopped dropping the `.mm` files). `core/src/metal/objc_handle.h` now holds the one `uintptr_t` slot to Metal object bridge and `vmaf_metal_library_load()` the one metallib loader, replacing 17 copies; file-scope helpers sit in anonymous namespaces; no score, public API or FFmpeg patch changes. See [tidy lanes](docs/development/tidy-lanes.md).
 
 

Makefile

@@ -346,8 +346,11 @@ TIDY_RATCHET_EXTRA_arm64 := --extra-arg=--target=$(AARCH64_TARGET) \
 # translation units out of compile_commands.json: write-compile-commands.py
 # exports only the native c/cpp_COMPILER rules, so without this second pass the
 # cuda and hip lanes measure the host files only and the sycl lane measures zero
-# SYCL feature TUs, recording an empty backend in its baseline.
-TIDY_RATCHET_COMPDB_cpu :=
+# SYCL feature TUs, recording an empty backend in its baseline. The `cpu` lane adds
+# the MATLAB MEX sources, which meson never builds and which need the stub headers
+# of scripts/ci/lint-stubs/matlab/ (T-TIDY-MATLAB-MEX-UNMEASURED-2026-09-22).
+TIDY_RATCHET_COMPDB_cpu := $(PYTHON_INTERPRETER) scripts/ci/gen-mex-compile-commands.py \
+	"$(TIDY_RATCHET_BUILD_DIR)"
 TIDY_RATCHET_COMPDB_clang :=
 TIDY_RATCHET_COMPDB_cuda := $(PYTHON_INTERPRETER) scripts/ci/gen-gpu-compile-commands.py \
 	"$(TIDY_RATCHET_BUILD_DIR)"

changelog.d/changed/adr-audit-backfill-2026-10-06.md

@@ -0,0 +1 @@
+- ADR-0643, ADR-0665, ADR-0666 and ADR-0673 lose the unfilled allocator template block that preceded their real text (their first heading read `<fill in title>`); a status update records it. Four ADR numbers that other ADRs cite and that had no file (ADR-0228, 0636, 0867, 0979) get a short record each, written from the commits and ADRs that name them.

changelog.d/changed/cuda-ms-ssim-device-level0.md

@@ -0,0 +1,5 @@
+- **The CUDA MS-SSIM twin no longer copies frames through the host.**
+  `float_ms_ssim_cuda` converted every plane of every frame on the host
+  (a device-to-host copy, a wait, `picture_copy()` and an upload); the
+  conversion now runs on the device with the same arithmetic, so its scores
+  are unchanged and the frame stays on the GPU.

changelog.d/changed/tidy-matlab-mex-lane.md

@@ -0,0 +1,5 @@
+- The `cpu` clang-tidy lane and the changed-files job measure the ten MATLAB MEX
+  sources of `compat/python-vmaf/matlab/` against self-authored stub `mex.h` and
+  `matrix.h` (`scripts/ci/lint-stubs/matlab/`); their lint exceptions are removed.
+  The first measurement fixed the mechanical findings and a defect in `ical_std.c`
+  (`mxDestroyArray()` was called on a matrix's data pointer).

compat/python-vmaf/matlab/STMAD_2011_MatlabCode/ical_stat.c

@@ -9,7 +9,8 @@
 static double block_mean(const double *xVal, int colLen, int i, int j)
 {
     double mean = 0;
-    int iB, jB;
+    int iB;
+    int jB;
 
     for (iB = i; iB < i + 16; iB++) {
         for (jB = j; jB < j + 16; jB++) {
@@ -26,8 +27,14 @@ static void block_moments(const double *xVal, int colLen, int i, int j, double *
                           double *outSkw, double *outKrt)
 {
     double mean = block_mean(xVal, colLen, i, j);
-    double stdev, skw, krt, stmp, tmp, tmp1;
-    int iB, jB;
+    double stdev;
+    double skw;
+    double krt;
+    double stmp;
+    double tmp;
+    double tmp1;
+    int iB;
+    int jB;
 
     stdev = 0;
     skw = 0;
@@ -61,7 +68,7 @@ static void block_moments(const double *xVal, int colLen, int i, int j, double *
     *outKrt = krt;
 }
 
-void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
+void mexFunction(int /* nlhs */, mxArray *plhs[], int /* nrhs */, const mxArray *prhs[])
 {
     /* We have input of one double type matrix*/
     /* this function calculates the local mean, std, skewness, and kurtosis*/
@@ -71,10 +78,19 @@ void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
 
     /*Declarations*/
     mxArray *xData;
-    double *xVal, *outStd, *outSkw, *outKrt;
-    double stdev, skw, krt;
-    int i, j, iB, jB;
-    int rowLen, colLen;
+    double *xVal;
+    double *outStd;
+    double *outSkw;
+    double *outKrt;
+    double stdev;
+    double skw;
+    double krt;
+    int i;
+    int j;
+    int iB;
+    int jB;
+    int rowLen;
+    int colLen;
 
     /*Copy input pointer x*/
     xData = prhs[0];
@@ -111,6 +127,4 @@ void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
             }
         }
     }
-
-    return;
 }

compat/python-vmaf/matlab/STMAD_2011_MatlabCode/ical_std.c

@@ -11,8 +11,14 @@
 static void block_mean_and_std(const double *xVal, const double *yVal, double *outStd,
                                double *outMean, int rowLen, int colLen)
 {
-    double mean, mean2, stdev, tmp1;
-    int i, j, iB, jB;
+    double mean;
+    double mean2;
+    double stdev;
+    double tmp1;
+    int i;
+    int j;
+    int iB;
+    int jB;
 
     /*Copy matrix while multiplying each point by 2*/
     for (i = 0; i < rowLen - 15; i += 4) {
@@ -54,8 +60,13 @@ static void block_mean_and_std(const double *xVal, const double *yVal, double *o
 static void block_modified_std(const double *yVal, double *outStdMod, double *TMP, int rowLen,
                                int colLen)
 {
-    double mean, stdev, tmp1;
-    int i, j, iB, jB;
+    double mean;
+    double stdev;
+    double tmp1;
+    int i;
+    int j;
+    int iB;
+    int jB;
 
     for (i = 0; i < rowLen - 15; i += 4) {
         for (j = 0; j < colLen - 15; j += 4) {
@@ -94,7 +105,10 @@ static void block_modified_std(const double *yVal, double *outStdMod, double *TM
 static void block_min_filter(double *outStdMod, const double *TMP, int rowLen, int colLen)
 {
     double mean;
-    int i, j, iB, jB;
+    int i;
+    int j;
+    int iB;
+    int jB;
 
     for (i = 0; i < rowLen - 15; i += 4) {
         for (j = 0; j < colLen - 15; j += 4) {
@@ -115,7 +129,7 @@ static void block_min_filter(double *outStdMod, const double *TMP, int rowLen, i
     }
 }
 
-void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
+void mexFunction(int /* nlhs */, mxArray *plhs[], int /* nrhs */, const mxArray *prhs[])
 {
     /* We have input of one double type matrix*/
     /* this function calculates the local mean, std, skewness, and kurtosis*/
@@ -124,19 +138,24 @@ void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
     /*---Inside mexFunction---*/
 
     /*Declarations*/
-    mxArray *xData, *yData;
-    double *xVal, *yVal, *outStd, *outStdMod, *outMean;
+    mxArray *xData;
+    mxArray *yData;
+    double *xVal;
+    double *yVal;
+    double *outStd;
+    double *outStdMod;
+    double *outMean;
+    mxArray *tmp_array;
     double *TMP;
-    int rowLen, colLen;
+    int rowLen;
+    int colLen;
 
     /*Copy input pointer x*/
     xData = prhs[0];
     yData = prhs[1];
 
     /*Get matrix x*/
     xVal = mxGetPr(xData);
-    rowLen = mxGetN(xData);
-    colLen = mxGetM(xData);
 
     /*Get matrix y*/
     yVal = mxGetPr(yData);
@@ -156,7 +175,8 @@ void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
     plhs[2] = mxCreateDoubleMatrix(colLen, rowLen, mxREAL); /*mxReal is our data-type*/
     outMean = mxGetPr(plhs[2]);
 
-    TMP = mxGetPr(mxCreateDoubleMatrix(colLen, rowLen, mxREAL));
+    tmp_array = mxCreateDoubleMatrix(colLen, rowLen, mxREAL);
+    TMP = mxGetPr(tmp_array);
 
     block_mean_and_std(xVal, yVal, outStd, outMean, rowLen, colLen);
 
@@ -167,7 +187,5 @@ void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
 
     block_min_filter(outStdMod, TMP, rowLen, colLen);
 
-    mxDestroyArray(TMP);
-
-    return;
+    mxDestroyArray(tmp_array);
 }