ADR-1205: The ssimulacra2 FMA unification extends to the scalar fallback and every GPU host copy¶

Status: Proposed
Date: 2026-09-06
Deciders: Lusoris
Tags: cuda, sycl, hip, metal, simd, correctness, feature-extractor, reproducibility

Context¶

ADR-0891 unified the YCbCr -> linear-RGB conversion in ssimulacra2 on a single-rounded fused multiply-add so the vector kernels and their scalar tails agree bit-for-bit. The change landed in the four SIMD implementations (AVX2, AVX-512, NEON, SVE2) and in the private scalar reference inside core/test/test_ssimulacra2_simd.c, but not in any of the five shipped copies that are not SIMD:

Path	Conversion
`test_ssimulacra2_simd.c` reference	`fmaf()`
AVX2 / AVX-512 / NEON / SVE2 (and their scalar tails)	fused multiply-add
`core/src/feature/ssimulacra2.c` (shipped scalar fallback)	plain mul-then-add
`cuda/ssimulacra2_cuda.c`, `hip/ssimulacra2_hip.c`, `metal/ssimulacra2_metal.mm`, `sycl/ssimulacra2_sycl.cpp`	plain mul-then-add

Because the SIMD test compares the SIMD kernels against its own private reference rather than against the shipped scalar function, it asserted bit-exactness, passed, and could not see that no shipped non-SIMD path matched.

The consequences are larger than the 1-ULP seed suggests, because the ssimulacra2 pipeline is ill-conditioned downstream: the edge-diff term computes |img - blur(img)|, a catastrophic cancellation, and the pooling takes a 4-norm, which is dominated by the few largest survivors. Measured on an RTX 4090 at 256x144, a ~1.8e-07 difference in linear RGB on 3677 of 36864 pixels grew to 1.3e-06 in XYB, 6.6e-06 after the recursive Gaussian, and 2.62e-03 in the final score against a places=4 (1e-4) gate. Forcing the CPU to its scalar path made the CPU and CUDA scores bit-identical (delta exactly 0.0), which proves the GPU kernels and host helpers were correct all along and the divergent side was the CPU SIMD reference.

This is therefore not only a GPU-parity defect: on a host without AVX2 the shipped scalar fallback produced different ssimulacra2 scores than on a host with it.

Decision¶

We will use fmaf() in the same three positions and the same order as ADR-0891 in all five remaining copies — the shipped CPU scalar fallback and the CUDA, HIP, Metal and SYCL host conversions — so that every shipped path computes one value.

Alternatives considered¶

Option	Pros	Cons	Why not chosen
Add `fmaf()` to the five missed copies (chosen)	Every shipped path agrees; removes a host-dependent score; no change to already-correct SIMD output	Five files must stay in sync	—
Remove FMA from the SIMD kernels instead	Also unifies; arguably simpler arithmetic	Reverses ADR-0891, costs SIMD throughput, and changes scores on the overwhelmingly common AVX2/AVX-512 hosts	Rejected — would move the majority path, not the minority
Compile the affected TUs with `-ffp-contract=fast` and rely on the compiler	No source change	Empirically ineffective — rebuilding with contraction off left the delta at 2.62e-03, so contraction was never the mechanism; also leaves behaviour at the mercy of compiler and flags	Rejected — measured and falsified
Widen the ssimulacra2 parity tolerance	Zero code change	Hides a real cross-host reproducibility bug	Rejected

Consequences¶

Positive: CPU-vs-CUDA ssimulacra2 agreement improves from 2.62e-03 to ~2.8e-09, uniformly across 256x144 .. 1280x720. The scalar fallback now produces the same score as the SIMD paths, so ssimulacra2 no longer depends on the host ISA.
Negative: ssimulacra2 scores produced by a non-SIMD host before this change do not reproduce after it. That path was the incorrect one, and no published fork artifact is known to have come from a non-SIMD host; scores from AVX2/AVX-512/NEON/SVE2 hosts are unchanged.
Neutral / follow-ups: test_ssimulacra2_simd.c should eventually assert against the shipped scalar function rather than a private copy of it; until it does, this class of drift can recur. Tracked as a follow-up rather than fixed here to keep this change reviewable.

References¶

ADR-0891 — the original FMA unification this completes.
Reproducer: meson test -C build --suite=gpu test_cuda_ssimulacra2_parity on a CUDA host.
Source: req — user direction to fix the outstanding GPU parity failures.