The four lanes are computed together: exponents, 24x24-bit products with
round-to-odd, alignment by truncation and a signed horizontal sum. One
pairwise maximum finds both the alignment exponent and any inf or NaN,
which go to the reference. The final rounding is branch-free and in
integers, since the host rounding mode may be the game's.
About three times the throughput of the reference on Apple M-series
(5.2 vs 15.2 ns per call). VFPUDot checks it against the reference on
four million inputs picked to cover cancellation, ties, subnormals and
the overflow edges; a billion more matched offline.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
