[antonwolfy]

The ASV suite has no benchmark for add or subtract and no micro-benchmark for a complex dtype. The contiguous complex fast path in #278 makes complex add/subtract up to 2.4x faster, and the suite doesn't see it: its only complex coverage is the mandelbrot kernels, which moved by 0–2% with that change. This PR adds benchmarks that do.

Changes

• micro/bench_micro.py: BenchAdd and BenchSubtract, written like BenchArctan2 and BenchPower (a new result array on each call), for float32, float64, complex64 and complex128 at 10k, 100k and 1M. The 10k and 100k sizes run mkl_umath's own loops and 1M calls oneMKL VM, since the threshold is n > 100000. A _uniform helper draws the real and imaginary parts independently.
• micro/bench_inplace.py (new): BenchAddInplace and BenchSubtractInplace for complex64 and complex128 at 10k and 100k, with modes out=a, out=b and z += z. Each mode takes its own loop in perf: add contiguous fast path to complex add/subtract fallback loop #278. z += z starts from zeros, because repeated doubling would overflow to inf.
• README.md: coverage table.

That's 48 new benchmarks, 24 per file. On the host below they add about 30 s per pass.

Testing

Core Ultra 5 236V (Lunar Lake), conda-forge icx 2026.1.1, NumPy 2.5.4, MKL 2026.1.0, MKL_NUM_THREADS=1, pinned to one P-core with turbo off. asv run --python=same against two builds of main 8ca254c (byte-identical, as an A/A control) and #278 at 1346ce2, with passes ordered main, main, #278, #278, main, main and samples appended.

• main vs main: nothing flagged, ratios 0.93–1.04.
• main vs perf: add contiguous fast path to complex add/subtract fallback loop #278: 27 of 48 flagged as faster, 26 of them at 0.80 or below, so a change that undoes the fast path would trip the 20% regression threshold in asv.conf.json.

Median time relative to main, so below 1 is faster:

Benchmark	#278 / main
BenchAdd / BenchSubtract, complex at 10k and complex64 at 100k	0.70–0.79
BenchAdd / BenchSubtract, complex128 at 100k	0.98
BenchAdd / BenchSubtract, complex at 1M (oneMKL)	0.99–1.00
BenchAdd / BenchSubtract, float32 / float64	0.94–1.07 (A/A spread 0.93–1.04)
In place, complex64	0.42–0.66
In place, complex128 at 10k, and z += z at 100k	0.54–0.79
In place, complex128 out=a / out=b at 100k	1.03–1.10, not significant per ASV

The complex128 in-place cases at 100k use two 1.6 MB arrays, more than the 2.5 MB L2, and ASV's spread there is ±5–10 µs. The same pattern outside ASV (fresh setup, one warm-up call, one timed call, 1,800 samples per build) gives 0.90–0.92 with an A/A of 1.00, so I read the ASV result as noise at that size.

Median times

Medians of 20 single-call samples, turbo off.

dtype	size	add main	add #278	subtract main	subtract #278
float32	10k	3.26 µs	3.30 µs	3.47 µs	3.25 µs
float32	100k	24.6 µs	24.5 µs	24.6 µs	24.5 µs
float32	1M	666 µs	658 µs	669 µs	664 µs
float64	10k	6.60 µs	7.07 µs	6.93 µs	6.56 µs
float64	100k	62.0 µs	65.9 µs	62.7 µs	63.2 µs
float64	1M	1.316 ms	1.320 ms	1.299 ms	1.298 ms
complex64	10k	9.20 µs	6.42 µs	9.18 µs	6.55 µs
complex64	100k	84.9 µs	66.8 µs	85.6 µs	66.9 µs
complex64	1M	1.260 ms	1.242 ms	1.285 ms	1.275 ms
complex128	10k	13.3 µs	10.1 µs	13.5 µs	10.0 µs
complex128	100k	249 µs	245 µs	251 µs	245 µs
complex128	1M	2.651 ms	2.655 ms	2.628 ms	2.605 ms

dtype	size	mode	add main	add #278	subtract main	subtract #278
complex64	10k	out=a	7.93 µs	5.08 µs	7.75 µs	5.07 µs
complex64	10k	out=b	7.82 µs	4.96 µs	7.75 µs	5.05 µs
complex64	10k	z += z	6.74 µs	3.46 µs	6.65 µs	3.50 µs
complex64	100k	out=a	75.1 µs	48.4 µs	75.3 µs	48.8 µs
complex64	100k	out=b	74.2 µs	48.7 µs	74.1 µs	48.5 µs
complex64	100k	z += z	62.2 µs	25.9 µs	60.3 µs	26.1 µs
complex128	10k	out=a	12.9 µs	9.98 µs	12.8 µs	9.94 µs
complex128	10k	out=b	12.7 µs	10.1 µs	12.8 µs	9.97 µs
complex128	10k	z += z	11.3 µs	6.13 µs	11.0 µs	6.19 µs
complex128	100k	out=a	143 µs	152 µs	139 µs	152 µs
complex128	100k	out=b	144 µs	148 µs	144 µs	148 µs
complex128	100k	z += z	114 µs	61.4 µs	104 µs	61.7 µs

pre-commit passes on the changed files.

Notes

• Unrelated, found while checking coverage: BenchArctan2 and BenchPower time stock NumPy. arctan2 and power are commented out in generate_umath.py, so patch_numpy_umath() leaves them alone. power gets mkl_umath loops in Add oneMKL power loops for float32 and float64 #285; both are left as they are here.