1-D Array Benchmarks

Reduce 1-D array into a single value.

Run the benchmark locally with (the script prints plain finite-data results first, then NaN-aware results):

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

The benchmark prints one row per (length, dtype, function). np, bn, and rd are trimmed median elapsed µs for NumPy, Bottleneck, and reducers: the fastest and slowest samples are dropped before taking the median. The np/rd and bn/rd columns are speedups relative to reducers; a value above 1 means reducers is faster. A speedup is bolded when reducers ratio > 0.95.

Each run uses 3 warmups and 15 repeats by default.

average rows use one weight per value, generated from a uniform distribution on [0.5, 1.5). The percentile/quantile rows (and their nan* forms) request three positions, [16, 50, 84] (the median and +/-1 sigma).

Plain, float64

These rows compare the plain path on finite data. std uses ddof=1. For average, Bottleneck has no weighted-average reducer.

function	np (µs)	rd (µs)	np/rd
`mean`	1.15	0.13	9.11x
`average`	2.81	0.29	9.62x
`median`	4.72	0.25	18.70x
`sum`	0.80	0.14	5.89x
`min`	0.79	0.14	5.80x
`max`	0.80	0.14	5.70x
`minmax`	1.62	0.15	10.57x
`var`	3.70	0.19	19.54x
`std`	3.94	0.25	15.80x
`percentile`	14.38	3.58	4.02x
`quantile`	14.01	3.82	3.67x

function	np (µs)	rd (µs)	np/rd
`mean`	1.18	0.14	8.42x
`average`	2.79	0.31	8.86x
`median`	4.87	0.37	13.06x
`sum`	0.78	0.13	6.09x
`min`	0.81	0.15	5.57x
`max`	0.79	0.14	5.64x
`minmax`	1.62	0.17	9.66x
`var`	3.60	0.20	17.64x
`std`	3.86	0.27	14.21x
`percentile`	14.48	3.92	3.69x
`quantile`	13.97	4.23	3.30x

function	np (µs)	rd (µs)	np/rd
`mean`	2.46	1.53	1.61x
`average`	7.89	2.34	3.37x
`median`	28.92	15.96	1.81x
`sum`	2.00	1.53	1.31x
`min`	1.37	1.40	0.98x
`max`	1.35	1.40	0.96x
`minmax`	2.75	1.71	1.60x
`var`	9.28	3.61	2.57x
`std`	9.73	3.86	2.52x
`percentile`	51.84	39.41	1.32x
`quantile`	50.49	39.32	1.28x

function	np (µs)	rd (µs)	np/rd
`mean`	1242.83	1405.92	0.88x
`average`	4483.08	1895.38	2.37x
`median`	70200.50	15501.21	4.53x
`sum`	1249.71	1412.08	0.89x
`min`	951.12	397.88	2.39x
`max`	966.00	420.12	2.30x
`minmax`	1912.75	425.13	4.50x
`var`	5283.50	3547.08	1.49x
`std`	5274.63	3551.08	1.49x
`percentile`	114862.92	37314.75	3.08x
`quantile`	116236.92	37130.62	3.13x

Plain, float32

These rows compare the plain path on finite float32 data. std uses ddof=1. For average, Bottleneck has no weighted-average reducer.

function	np (µs)	rd (µs)	np/rd
`mean`	1.95	0.15	13.29x
`average`	2.66	0.33	8.11x
`median`	5.47	0.23	23.40x
`sum`	0.77	0.15	5.27x
`min`	0.77	0.15	5.04x
`max`	0.77	0.15	5.10x
`minmax`	1.56	0.17	9.22x
`var`	4.53	0.20	22.91x
`std`	4.91	0.26	18.59x
`percentile`	14.51	3.52	4.13x
`quantile`	14.26	3.80	3.76x

function	np (µs)	rd (µs)	np/rd
`mean`	1.94	0.17	11.71x
`average`	2.74	0.36	7.64x
`median`	5.74	0.38	15.28x
`sum`	0.77	0.15	5.01x
`min`	0.80	0.16	4.87x
`max`	0.78	0.16	4.93x
`minmax`	1.62	0.18	8.79x
`var`	4.54	0.23	19.77x
`std`	4.89	0.29	16.89x
`percentile`	14.75	3.86	3.82x
`quantile`	14.11	4.21	3.35x

function	np (µs)	rd (µs)	np/rd
`mean`	3.21	1.67	1.92x
`average`	6.58	3.44	1.91x
`median`	29.79	14.69	2.03x
`sum`	1.90	1.71	1.11x
`min`	1.03	1.40	0.74x
`max`	1.01	1.39	0.72x
`minmax`	2.08	1.71	1.22x
`var`	8.46	3.97	2.13x
`std`	8.92	3.97	2.24x
`percentile`	50.17	37.05	1.35x
`quantile`	50.16	37.03	1.35x

function	np (µs)	rd (µs)	np/rd
`mean`	1149.54	1644.46	0.70x
`average`	3323.54	3105.29	1.07x
`median`	77016.87	14371.58	5.36x
`sum`	1143.33	1593.33	0.72x
`min`	439.37	318.08	1.38x
`max`	453.38	308.62	1.47x
`minmax`	883.67	357.13	2.47x
`var`	3592.25	3757.00	0.96x
`std`	3614.29	3708.42	0.97x
`percentile`	114627.62	35568.00	3.22x
`quantile`	114870.21	35316.00	3.25x

NaN-aware, float64

These rows compare rd.nan* against np.nan* and bn.nan*. nanstd uses ddof=1. For nanaverage, the NumPy baseline masks a and weights before calling np.average; Bottleneck has no weighted-average reducer.

function	np (µs)	bn (µs)	rd (µs)	np/rd	bn/rd
`nanmean`	4.25	0.07	0.15	28.82x	0.49x
`nanaverage`	3.06	-	0.30	10.08x	-
`nanmedian`	6.47	0.10	0.24	26.92x	0.43x
`nansum`	1.74	0.07	0.15	11.84x	0.49x
`nanmin`	1.55	0.06	0.15	10.60x	0.39x
`nanmax`	1.57	0.06	0.15	10.71x	0.39x
`nanminmax`	3.14	0.11	0.17	19.00x	0.67x
`nanvar`	9.57	0.12	0.20	47.22x	0.59x
`nanstd`	10.35	0.16	0.29	36.11x	0.57x
`nanpercentile`	16.65	-	3.50	4.76x	-
`nanquantile`	16.08	-	3.79	4.24x	-

function	np (µs)	bn (µs)	rd (µs)	np/rd	bn/rd
`nanmean`	4.33	0.16	0.17	25.06x	0.92x
`nanaverage`	3.18	-	0.32	9.88x	-
`nanmedian`	6.79	0.21	0.36	18.69x	0.59x
`nansum`	1.76	0.15	0.16	11.07x	0.97x
`nanmin`	1.59	0.09	0.16	10.28x	0.56x
`nanmax`	1.58	0.09	0.16	9.86x	0.55x
`nanminmax`	3.26	0.17	0.18	18.13x	0.93x
`nanvar`	9.62	0.29	0.24	39.57x	1.20x
`nanstd`	10.22	0.35	0.31	32.49x	1.10x
`nanpercentile`	26.78	-	4.04	6.63x	-
`nanquantile`	16.75	-	5.62	2.98x	-

function	np (µs)	bn (µs)	rd (µs)	np/rd	bn/rd
`nanmean`	17.68	19.39	5.89	3.00x	3.29x
`nanaverage`	20.52	-	3.40	6.04x	-
`nanmedian`	37.00	23.96	20.74	1.78x	1.16x
`nansum`	8.51	14.79	3.99	2.13x	3.70x
`nanmin`	2.28	5.23	1.45	1.58x	3.62x
`nanmax`	2.29	5.31	1.55	1.48x	3.43x
`nanminmax`	5.70	10.27	1.83	3.11x	5.61x
`nanvar`	35.03	31.62	7.16	4.89x	4.41x
`nanstd`	34.39	34.96	8.14	4.23x	4.30x
`nanpercentile`	67.52	-	57.25	1.18x	-
`nanquantile`	80.25	-	42.12	1.91x	-

function	np (µs)	bn (µs)	rd (µs)	np/rd	bn/rd
`nanmean`	6860.37	12986.58	5381.96	1.27x	2.41x
`nanaverage`	12966.96	-	2600.71	4.99x	-
`nanmedian`	55754.29	62285.00	17180.88	3.25x	3.63x
`nansum`	5483.00	12445.46	3228.96	1.70x	3.85x
`nanmin`	973.17	5182.71	423.42	2.30x	12.24x
`nanmax`	955.54	5059.75	437.67	2.18x	11.56x
`nanminmax`	1976.50	9903.83	407.29	4.85x	24.32x
`nanvar`	17407.50	25080.29	6146.75	2.83x	4.08x
`nanstd`	17286.87	24883.71	6145.12	2.81x	4.05x
`nanpercentile`	118428.96	-	38616.58	3.07x	-
`nanquantile`	118718.75	-	38961.83	3.05x	-

NaN-aware, float32

These rows compare float32 rd.nan* against np.nan* and bn.nan*. nanstd uses ddof=1. For nanaverage, the NumPy baseline masks a and weights before calling np.average; Bottleneck has no weighted-average reducer.

function	np (µs)	bn (µs)	rd (µs)	np/rd	bn/rd
`nanmean`	5.32	0.07	0.18	30.27x	0.42x
`nanaverage`	3.17	-	0.36	8.81x	-
`nanmedian`	9.78	0.11	0.26	38.22x	0.41x
`nansum`	1.83	0.07	0.17	10.81x	0.43x
`nanmin`	1.61	0.06	0.17	9.56x	0.37x
`nanmax`	1.59	0.06	0.17	9.31x	0.33x
`nanminmax`	3.18	0.11	0.18	17.33x	0.60x
`nanvar`	10.80	0.12	0.24	44.89x	0.50x
`nanstd`	11.36	0.16	0.29	39.84x	0.56x
`nanpercentile`	17.05	-	3.49	4.88x	-
`nanquantile`	17.03	-	3.85	4.42x	-

function	np (µs)	bn (µs)	rd (µs)	np/rd	bn/rd
`nanmean`	5.30	0.16	0.19	28.19x	0.83x
`nanaverage`	3.25	-	0.38	8.46x	-
`nanmedian`	7.92	0.21	0.38	20.87x	0.56x
`nansum`	1.82	0.15	0.18	9.92x	0.84x
`nanmin`	1.60	0.09	0.18	9.08x	0.51x
`nanmax`	1.56	0.09	0.18	8.69x	0.49x
`nanminmax`	3.14	0.17	0.20	15.54x	0.82x
`nanvar`	11.39	0.30	0.27	42.08x	1.12x
`nanstd`	11.64	0.34	0.33	35.53x	1.03x
`nanpercentile`	17.44	-	3.97	4.40x	-
`nanquantile`	16.96	-	4.23	4.01x	-

function	np (µs)	bn (µs)	rd (µs)	np/rd	bn/rd
`nanmean`	12.26	11.44	3.18	3.85x	3.60x
`nanaverage`	14.80	-	3.86	3.83x	-
`nanmedian`	32.12	20.91	17.91	1.79x	1.17x
`nansum`	6.11	12.31	2.98	2.05x	4.13x
`nanmin`	1.85	4.77	1.37	1.35x	3.48x
`nanmax`	1.87	4.50	1.24	1.50x	3.62x
`nanminmax`	3.83	9.52	1.62	2.36x	5.86x
`nanvar`	27.69	25.00	6.33	4.37x	3.95x
`nanstd`	27.56	25.06	6.18	4.46x	4.05x
`nanpercentile`	56.69	-	39.34	1.44x	-
`nanquantile`	57.54	-	39.80	1.45x	-

function	np (µs)	bn (µs)	rd (µs)	np/rd	bn/rd
`nanmean`	5383.63	11343.13	3067.08	1.76x	3.70x
`nanaverage`	10844.63	-	3854.04	2.81x	-
`nanmedian`	74681.58	58510.08	14937.17	5.00x	3.92x
`nansum`	3568.08	11215.92	3017.08	1.18x	3.72x
`nanmin`	426.00	4730.37	308.08	1.38x	15.35x
`nanmax`	449.50	4656.54	310.79	1.45x	14.98x
`nanminmax`	907.71	9253.00	318.00	2.85x	29.10x
`nanvar`	14054.17	24633.04	5999.33	2.34x	4.11x
`nanstd`	14135.12	25003.75	5960.54	2.37x	4.19x
`nanpercentile`	111802.79	-	35897.71	3.11x	-
`nanquantile`	110976.12	-	35670.29	3.11x	-