1-D Array Benchmarks

imcombiners.kernels exposes _1d functions for generic value vectors as described in 1-D tutorial. These are meant to be fast for typical 1-D workloads in astronomy, and may replace NumPy or Bottleneck for operation-specific hot paths.

The _1d rejection APIs call direct Rust slice entry points. Those entry points and the image-stack kernels share the same per-output-vector decision logic, so 1-D calls avoid (N, 1, 1) packaging overhead without maintaining a separate rejection algorithm.

Run the benchmark locally with:

uv run --extra bench python benchmarks/benchmark_1d.py

The benchmark reports one row per (length, dtype, operation). np, bn, and imc are median elapsed microseconds for NumPy, Bottleneck, and imcombiners. np/imc and bn/imc are speedup factors for imc: values above 1 mean imc is faster than that comparison path, and values below 1 mean imc is slower. float64 is the more typical general NumPy workload; float32 remains important for image stacks and CCD pipelines.

Every generated benchmark vector contains NaNs. The compared functions are the NaN-aware variants: np.nan*, bn.nan*, and the _1d imcombiners kernels. Timing columns use {:.2f} microseconds. Speedup columns use {:.1f}x.

length	dtype	op	np (us)	bn (us)	imc (us)	np/imc	bn/imc
10^2	float64	mean	4.04	0.15	0.14	29.1x	1.1x
10^2	float64	median	6.27	0.20	0.31	20.1x	0.7x
10^2	float64	sum	1.69	0.15	0.14	12.0x	1.0x
10^2	float64	min	1.51	0.09	0.13	11.5x	0.7x
10^2	float64	max	1.49	0.09	0.13	11.7x	0.7x
10^2	float64	var	8.98	0.29	0.17	52.0x	1.7x
10^4	float64	mean	11.24	11.14	2.88	3.9x	3.9x
10^4	float64	median	30.60	14.51	15.94	1.9x	0.9x
10^4	float64	sum	6.81	12.53	2.99	2.3x	4.2x
10^4	float64	min	2.11	5.05	1.33	1.6x	3.8x
10^4	float64	max	2.21	5.17	1.39	1.6x	3.7x
10^4	float64	var	26.82	24.73	3.52	7.6x	7.0x
10^7	float64	mean	6316.96	11306.96	2848.58	2.2x	4.0x
10^7	float64	median	56071.50	52099.04	17486.00	3.2x	3.0x
10^7	float64	sum	4798.46	11374.21	2852.79	1.7x	4.0x
10^7	float64	min	840.62	4499.88	368.04	2.3x	12.2x
10^7	float64	max	856.88	4594.58	357.67	2.4x	12.8x
10^7	float64	var	16004.79	24882.92	3500.87	4.6x	7.1x
10^2	float32	mean	5.08	0.15	0.12	43.5x	1.2x
10^2	float32	median	7.20	0.21	0.29	24.6x	0.7x
10^2	float32	sum	1.70	0.15	0.12	14.3x	1.2x
10^2	float32	min	1.51	0.08	0.10	14.8x	0.8x
10^2	float32	max	1.47	0.08	0.11	14.0x	0.7x
10^2	float32	var	9.93	0.28	0.14	68.9x	2.0x
10^4	float32	mean	11.76	11.25	2.91	4.0x	3.9x
10^4	float32	median	31.05	14.63	15.84	2.0x	0.9x
10^4	float32	sum	5.52	11.36	2.86	1.9x	4.0x
10^4	float32	min	1.73	4.57	1.19	1.5x	3.8x
10^4	float32	max	1.68	4.47	1.19	1.4x	3.8x
10^4	float32	var	24.93	24.82	2.89	8.6x	8.6x
10^7	float32	mean	5466.42	12166.38	2999.17	1.8x	4.1x
10^7	float32	median	68780.25	67508.33	14898.29	4.6x	4.5x
10^7	float32	sum	3471.17	11381.50	2854.54	1.2x	4.0x
10^7	float32	min	410.62	4545.87	220.29	1.9x	20.6x
10^7	float32	max	449.17	4585.04	194.96	2.3x	23.5x
10^7	float32	var	13546.38	24994.04	2873.79	4.7x	8.7x

imc is generally very fast, and often faster than both NumPy and Bottleneck. Thus, if the user has already loaded imc, less reason to import Bottleneck for faster NaN-aware reductions.

Fast variance calculation implies why imc is optimal for sigma-clip style rejections.