ferray-ufunc

NumPy-equivalent element-wise universal functions (ufuncs) and their reductions for the ferray N-dimensional array library.

Part of the ferray workspace — powered by pulp runtime SIMD dispatch (SSE2/AVX2/AVX-512/NEON).

Overview

ufuncs are generic free functions that preserve the input array's element type and rank: sin on an Array2<f64> returns an Array2<f64>. The crate ships 100+ such functions, organized into families under src/ops/:

• Arithmetic — add, subtract, multiply, divide, true_divide, floor_divide, power, remainder, mod_, fmod, divmod, negative, absolute, sign, reciprocal, square, sqrt, cbrt, gcd, lcm, cross, plus cumulative ops (cumsum, cumprod, diff, ediff1d, gradient, trapezoid) and add_reduce / nan_*_reduce reductions.
• Trigonometric (ops::trig) — sin, cos, tan, arcsin, arccos, arctan, arctan2, hypot, sinh, cosh, tanh, arcsinh, arccosh, arctanh, degrees/radians, deg2rad/rad2deg, unwrap, plus fast paths sin_fast / cos_fast.
• Exp / log (ops::explog) — exp, exp2, expm1, exp_fast, log, log2, log10, log1p, logaddexp, logaddexp2.
• Rounding (ops::rounding) — floor, ceil, trunc, rint, round, around, fix (with integer/bool-preserving *_int variants per REQ-24).
• Float intrinsics (ops::floatintrinsic) — signbit, copysign, nextafter, spacing, frexp, ldexp, modf, isnan, isinf, isfinite, isposinf/isneginf, clip, nan_to_num, maximum/minimum, fmax/fmin, float_power.
• Complex (ops::complex) — real, imag, conj/conjugate, angle, abs, predicates (isreal, iscomplex, isscalar), plus the *_complex transcendentals (sin_complex, exp_complex, ln_complex, sqrt_complex, power_complex, …).
• Bitwise (ops::bitwise) — bitwise_and/or/xor, invert/ bitwise_not, left_shift, right_shift, bitwise_count.
• Comparison (ops::comparison) — equal, not_equal, less, less_equal, greater, greater_equal, isclose, allclose, array_equal, array_equiv (with *_broadcast variants).
• Logical (ops::logical) — logical_and/or/xor/not, all, any.
• Datetime / timedelta (ops::datetime) — add_datetime_timedelta, sub_datetime, sub_datetime_timedelta, add_timedelta, sub_timedelta, isnat_datetime, isnat_timedelta.
• Special / misc — sinc, i0, convolve (+ interp).

Dtype promotion (promoted::*) follows NumPy: integer/bool inputs to float ufuncs promote to the smallest safe-cast float (add_promoted, sin_promote, hypot_promote, …). First-class ufunc objects (Ufunc, add_ufunc) expose .reduce / .accumulate / .outer / .at. Operator overloads (+, -, *, /, %, &, |, ^, !, <<, >>) are provided via operator_overloads.

NumPy correspondence

SIMD

Contiguous inner loops dispatch to SIMD kernels at runtime via pulp, which selects the best available instruction set (SSE2 / AVX2 / AVX-512 / NEON) for the host CPU. Setting FERRAY_FORCE_SCALAR=1 forces the scalar fallback path, used to cross-verify SIMD results against scalar reference output in tests. Large compute-bound arrays are additionally parallelized with rayon.

Feature flags

Example

use ferray_ufunc::{add, sin, exp, copysign};
use ferray_core::prelude::*;

let a = Array1::<f64>::linspace(0.0, 6.28, 1000)?;

// Element-wise transcendentals (preserve shape and dtype)
let s = sin(&a)?;
let e = exp(&s)?;

// Binary ufuncs with broadcasting
let sum = add(&s, &e)?;

// Float intrinsic: copy the sign of `s` onto `e`
let signed = copysign(&e, &s)?;
# Ok::<(), ferray_core::FerrayError>(())

MSRV & edition

Edition 2024, MSRV 1.88. Licensed under MIT OR Apache-2.0.