reducers 0.3.0

Rust-backed reductions for NumPy arrays (plain + NaN-aware)

reducers

Reduction functions + Rust(rs), shortname rd.

Rust-backed reduction functions for NumPy arrays - plain (numpy-like) and NaN-aware.

• Full documentation: https://ysbach.github.io/reducers/
• Rust API reference: https://docs.rs/reducers

The Goal of this toy project was:

1. much faster than numpy in many use cases,
2. much faster than bottleneck in many use cases, and
3. especially maximum performance for median and variance calculations, which are often bottlenecks in data processing pipelines.

reducers might be slower than numpy or bottleneck for small arrays. However, the most time-consuming reductions like large arrays or deep stacks, median, percentile or quantile, var and std are frequently several times (>100 times for nanpercentiles) faster than numpy and bottleneck.

Install

pip install reducers

For Rust crate use:

[dependencies]
reducers = "<version>"

After Installation

Run the autotuner once on your machine where reducers will run:

python -m reducers.autotuner

It saves parallel-grain settings for that CPU and workload profile. Future import reducers calls apply those settings automatically. The built-in defaults are still valid; use python -m reducers.autotuner --reset to remove the saved tuning file and return to them.

import numpy as np
import reducers as rd

a = np.array([1.0, 2.0, np.nan, np.inf, 5.0])

rd.mean(a)                      # nan: plain reducers propagate NaN/inf
rd.nanmean(a)                   # inf: skip NaN, keep inf
rd.nanmean(a, ignore_inf=True)  # finite-only
rd.nanminmax(a)                 # one fused 1-D scan for nanmin + nanmax
rd.nanpercentile(a, [16, 50, 84])

Axis reductions cover the layouts this package optimizes:

rng = np.random.default_rng(20250311)
stack = rng.normal(size=(31, 256, 256)).astype("f4")
rows = rng.normal(size=(256, 256, 31)).astype("f4")

rd.nanmedian(stack, axis=0)      # stack reduction -> shape (256, 256)
rd.nanmean(rows, axis=-1)        # contiguous trailing-axis reduction

• For [nan]var and [nan]std, return_mean=True returns the already-computed mean with the variance or standard deviation to avoid duplicate work when both are needed.
• [nan]sum(a, weights=w) can do similar: return_sum_weights=True and return_unweighted_sum=True expose quantities already available during the fused weighted scan, avoiding separate sum(a * w), sum(w), or sum(a) passes when a caller needs them together.

std, mean = rd.nanstd(a, ddof=1, return_mean=True)
weighted_sum, sum_of_weights = rd.nansum(a, weights=w, return_sum_weights=True)
weighted_sum, unweighted_sum = rd.nansum(a, weights=w, return_unweighted_sum=True)
weighted_sum, unweighted_sum, sum_of_weights = rd.nansum(
    a, weights=w, return_unweighted_sum=True, return_sum_weights=True
)

Dual use: the kernel modules are pure Rust (no PyO3/NumPy) and usable as a crate.

Maximum-performance Python calls

For fixed production hot loops, import the low-level Python API as rdl:

import reducers.lowlevel as rdl

The rdl functions call the same Rust kernels but skip the high-level Python normalization layer. Weighted hot loops can choose the narrow fused primitive for the output terms they actually need:

weighted_sum = rdl.weighted_sum_only_skip_nonfinite(a, w)
weighted_sum, sum_weights = rdl.weighted_sum_and_weights_skip_nonfinite(a, w)
weighted_sum, sum_weights, unweighted_sum = rdl.weighted_sum_skip_nonfinite(a, w)
average = rdl.weighted_average_skip_nonfinite(a, w)

See the documentation for details on achieving maximum performance.

Current limits

• axis may be None (default, whole-array), 0 or -1 (identical to a.ndim - 1); other axes raise NotImplementedError. This keeps hidden transpose/copy costs out of the API and lets the Rust kernels specialize for the supported layouts.

• NumPy-like subset: There are many unsupported parameters like out, keepdims, where, dtype, or percentile method (linear only). Adding them will not likely be considered unless there is a strong use case, as they add complexity and maintenance burden. The main focus is on the core reduction logic and, more importantly, performance.

See the documentation for detailed API semantics, performance notes, axis behavior, and release wheels: https://ysbach.github.io/reducers/.