[tannevaled]

The README said float64 Exp below 256 Ki was slower than NumPy (0.82×, measured with 16 cores). Measured today on Zen 3, that is the allocation, not Exp. This PR corrects the claim and adds the benchmark that shows it.

The evidence:

• On one core: float64 Exp is 1.2–1.3× faster than NumPy at every size (1 Ki to 4 Mi).
• With GOMAXPROCS = 16: the same serial code (below ParThreshold, nothing is split) gets slower than with 1, at 89 µs against 73 at 16 Ki.
• The attribution: Add goes from 14 to 23–26 µs at 16 Ps, while AddInto, which allocates nothing, stays at 4.3 µs. The per-result allocation costs more with more Ps.
• In a Workspace (new BenchmarkExpWS), the difference disappears.

float64 Exp, Zen 3, load ≤ 5	1 Ki	16 Ki	256 Ki	4 Mi
NumPy 2.5.4 (one thread)	6.0 µs	85.5 µs	1.32 ms	25.0 ms
ours, 1 P, allocating	4.9 µs	73 µs	1.11 ms	19.3 ms
ours, 16 P, allocating	5.8–6.6 µs	83–95 µs	0.63–0.78 ms	4.8–6.7 ms
ours, 16 P, Workspace	4.1 µs	64 µs	0.18 ms	1.47 ms

Changes:

• README: float64 Exp moves under "faster". The "slower" list keeps the allocating forms, noting that their cost grows with GOMAXPROCS.
• docs/perf.md: the table above. The note about the other architectures now says the kernels live in go-simd/floats since v0.37.0, NEON included (proven, not benchmarked).
• benchmarks: BenchmarkExpWS.

No library code changes.

🤖 Generated with Claude Code