Research-0064: GPU-lane NOLINT audit — which suppressions are actually live¶

Date: 2026-09-22
Owner: Lusoris
Tags: lint, audit, gpu, sycl, cuda, hip, metal, touched-file-rule
Companion ADR: ADR-1290
Ledger: .workingdir/BUGS.md BUG-029, BUG-041

Question¶

ADR-0141 §2 allows a NOLINT only where refactoring would break a load-bearing invariant, and requires an inline ADR-NNNN citation naming it. ADR-0278 discharged that for the CPU lane. The GPU lane was never swept. How many uncited markers are there, and — the question the earlier attempts skipped — which of them actually suppress a diagnostic?

Citing a marker that suppresses nothing is worse than leaving it: it manufactures a justification for a suppression the code does not need, and the citation makes it look audited.

How many¶

count_uncited_nolints() in scripts/ci/tidy-ratchet.py is the function the ratchet gates on, so it is the authority. On fb06193b3, tree-wide:

Source	Count	Note
Filed in `BUGS.md` (headline)	50	Matches no counting method the tool implements
Filed in `BUGS.md` (legacy note)	6	Correct for what the lanes could see
Sum of the lane baselines	6	cuda 3 + hip 1 + sycl 2
`count_uncited_nolints()`, tree-wide	21	The real figure

The 15-marker gap is exactly the BUG-041 blind spot: 14 markers under core/src/feature/sycl/ + core/src/sycl/ and 1 under core/src/feature/metal/ live in files no lane measures — the sycl lane's compilation database contained zero SYCL feature TUs, and there is no Metal lane. A gate that cannot see a file reports it as clean.

A same-line grep (NOLINT without ADR- on that line) over the GPU trees returns 246 of 468 markers, which is presumably where a figure like "50" came from; it is not what the tool measures, because a citation also counts on the previous or next line and inside the enclosing block comment.

Method¶

For each file, every NOLINT* token was rewritten to an inert spelling and clang-tidy re-run over the real compilation database for that lane (--cuda-host-only -nocudalib for cuda, -x hip -D__HIP_PLATFORM_AMD__=1 for hip, scripts/ci/clang-tidy-sycl.sh for sycl), then the file restored. A marker whose check appears in the diagnostic set is live; one whose check does not is dead.

Two traps worth recording:

Parse the whole bracket. Diagnostics carry every enabled alias, e.g. [performance-no-int-to-ptr,-warnings-as-errors]. A regex anchored on [a-z-]+\] captures only the last token and silently reports the check as absent — which is how a first pass wrongly concluded the two CUDA Driver-API brackets were dead.
Match error: as well as warning:. The ten checks in .clang-tidy's WarningsAsErrors list — performance-no-int-to-ptr among them — are reported at error level.

Control: a probe TU confirms misc-const-correctness fires under this exact config, so its absence on the SYCL kernels is a real negative, not a disabled check.

Result¶

21 markers: 7 dead, 1 live but misdescribed, 13 live and load-bearing.

Dead — removed, not cited¶

Site	Check	Why it never fires
`cuda/ssimulacra2_cuda.c:301`	`readability-function-size`	Guards `ss2c_setup_gaussian`, which is under the thresholds. Its siblings at 406 / 905 / 1006 do exceed them and are cited — so the ledger's "copy theirs" advice would have cemented a false citation
`sycl/integer_adm_sycl.cpp:480`	`bugprone-branch-clone`	The two arms are not clones: one reads `uint8_t` at `e_in_stride`, the other `uint16_t` at `e_in_stride / 2`
`sycl/integer_adm_sycl.cpp` ×2	`misc-const-correctness`	On `csf_accum_ptr` / `cm_accum_ptr`. The check proposes a `const` pointee; the pointee is mutated through `atomic_ref`, so the check correctly stays silent
`sycl/integer_adm_sycl.cpp` ×2	`misc-const-correctness`	On `total_csf` / `total_cm`, which the following loop accumulates into
`sycl/integer_motion_sycl.cpp` ×1	`misc-const-correctness`	Same shape as the two above

The five misc-const-correctness comments all claimed clang-tidy "cannot see the writes through SYCL atomic_ref". It can. docs/state.md (T-SYCL-LINT-SWEEP-2026-09-16) had promoted that claim to "the measured justification" for these markers; it is corrected there.

Live but misdescribed — fixed properly¶

sycl/integer_adm_sycl.cpp launch_dwt_hori_pair carried NOLINTNEXTLINE(misc-unused-parameters) reading "kept for ABI symmetry", naming buf_stride. buf_stride is used (auto e_buf_stride = buf_stride). The diagnostic is at the parameter h_add, which nothing in the body reads: the horizontal pass derives its own rounding addend, rnd = 1 << (h_shift - 1), which equals every h_add the shift table supplies (h_shift 16 → 32768, 15 → 16384, all four rows). The CPU reference does the same — add_shift_HP is a local constant beside shift_HP in integer_adm.c, not a parameter — so this is not a parity surface either. The parameter, its DwtShifts field and its call argument are removed; the TU rebuilds under icpx 2026.0 and no arithmetic changes.

There is no "ABI symmetry" to preserve: the vertical sibling's v_add plays a different role (lo_val -= dwt_lo_sum * e_v_add), not a rounding addend.

Live and load-bearing — cited¶

Sites	Check	Citation
`cuda/integer_adm_cuda.c`, `cuda/integer_vif_cuda.c`	`performance-no-int-to-ptr` (28 + 18 diagnostics)	ADR-0141 §2 — `CUdeviceptr` is an integer handle; the cast is the CUDA Driver API contract
7 SYCL TUs	`misc-use-anonymous-namespace`, `misc-use-internal-linkage`	ADR-0141 §2 — entry-point addresses live in an `extern "C" VmafFeatureExtractor` struct, and a namespace cannot appear inside that linkage specification
`metal/ssimulacra2_metal.mm`	`readability-function-size`	ADR-0141 §2 — line-for-line scalar-diff parity port. Not verifiable here: Metal has no Linux toolchain, so this one is cited on the same grounds as its four already-cited siblings in the same file rather than on a measurement
`test_hip_float_vif_parity.c`	`readability-function-size`	ADR-1264 — the `-ENOSYS` scaffold skip contract is checked after each of four HIP entry points
`test_sycl_motion3_parity.c` ×2	`readability-function-size`	ADR-0141 §2 + ADR-0278 — same wording as the three already-cited siblings in the same file

Tree-wide uncited after the sweep: 0.

Loose end¶

count_uncited_nolints()'s block-comment scan runs forward only, from the marker to the closing */. A citation written above the marker in the same comment — the natural prose order, and what four of these sites had — does not count. That is why several markers with honest ADR references still measured as uncited. This sweep moved the tokens into the window rather than widening the scan: loosening the counter would reduce the count without improving a single comment, and the current rule ("the citation is adjacent to the marker") is defensible on its own terms. Recorded here so the next reader does not mistake it for a bug in the sweep.