Skip to content

Research-0734: CUDA __mul24 silent-corruption sweep (CUDA 13.3 fix surface)

Status: Complete Date: 2026-05-28 Author: cuda-reviewer agent

Summary

Zero __mul24 (or __umul24 / __mul24hi) calls exist anywhere in core/src/feature/cuda/ or core/src/cuda/. The CUDA 13.3 silent-corruption bug (present since CUDA 11.1) has no exposure in this codebase.

Scope

All .cu, .cuh, .c, .h files under:

  • /core/src/feature/cuda/ — 68 files (kernels for ADM, VIF, motion, SSIM, MS-SSIM, PSNR, PSNR-HVS, CAMBI, CIEDE, moment, SSIMULACRA2, SpEED, float-twin extractors)
  • /core/src/cuda/ — 10 files (runtime: common, picture, dispatch, kernel template)

Inventory

grep -rn '__mul24\|__umul24\|__mul24hi' \
    core/src/feature/cuda/ core/src/cuda/
# → no output; exit 1 (grep found nothing)

Total __mul24 sites: 0

Classification breakdown

Classification Count
(a) Both args runtime-variable — SAFE 0
(b) One arg compile-time constant — BUG IMPACTED 0
(c) Both args compile-time constant — constexpr-foldable 0

CPU-reference cross-check

The kernels that were most likely to have used __mul24 as a "performance optimisation for 24-bit-range data" are:

  • integer_vif/filter1d.cu — uses standard C integer multiply for accumulation; no __mul24.
  • integer_adm/adm_dwt2.cu, adm_decouple.cu, adm_csf.cu — uses standard * throughout.
  • integer_motion/motion_score.cu — uses standard * and >> for SAD/mean.
  • integer_ssim/ssim_score.cu, integer_ms_ssim/ms_ssim_score.cu — uses standard * for cross-product terms.

In each case the CPU reference path (core/src/feature/*.c) also uses plain C multiply operators. No divergence between CPU and CUDA multiply semantics exists.

Evidence

The bug surface condition is:

  1. __mul24(val, CONSTANT) where CONSTANT is a compile-time-constant integer.
  2. Either operand may silently produce wrong results on CUDA 11.1–13.2 JIT.
  3. Fixed in CUDA 13.3 per NVIDIA release notes (surfaced in PR #64 digest).

Because the codebase contains zero calls to any __mul24 variant, condition (1) is never met. No divergence evidence can be produced because there is nothing to test.

Score-correctness conclusion

No score impact. The CUDA 13.3 fix has no numerical effect on this fork because the affected intrinsic is absent from the entire CUDA kernel tree.

Recommendation

No in-tree code change is required. The audit finding is documented here and in docs/backends/cuda/overview.md (§ Known gaps / CUDA version notes) so future contributors know the search was performed.

The AGENTS.md invariant note for core/src/feature/cuda/ is updated to record the prohibition on introducing __mul24 — any future use must cite a CODEOWNERS-approved rationale noting CUDA version constraints.

Reproducer

# From the worktree root:
grep -rn '__mul24\|__umul24\|__mul24hi' \
    core/src/feature/cuda/ core/src/cuda/
# Expected: no output (exit 1)