/* * Single-precision polynomial evaluation function for scalar * atan(x) and atan2(y,x). * * Copyright (c) 2021-2023, Arm Limited. * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception */ #ifndef PL_MATH_ATANF_COMMON_H #define PL_MATH_ATANF_COMMON_H #include "math_config.h" #include "poly_scalar_f32.h" /* Polynomial used in fast atanf(x) and atan2f(y,x) implementations The order 7 polynomial P approximates (atan(sqrt(x))-sqrt(x))/x^(3/2). */ static inline float eval_poly (float z, float az, float shift) { /* Use 2-level Estrin scheme for P(z^2) with deg(P)=7. However, a standard implementation using z8 creates spurious underflow in the very last fma (when z^8 is small enough). Therefore, we split the last fma into a mul and and an fma. Horner and single-level Estrin have higher errors that exceed threshold. */ float z2 = z * z; float z4 = z2 * z2; /* Then assemble polynomial. */ float y = fmaf ( z4, z4 * pairwise_poly_3_f32 (z2, z4, __atanf_poly_data.poly + 4), pairwise_poly_3_f32 (z2, z4, __atanf_poly_data.poly)); /* Finalize: y = shift + z * P(z^2). */ return fmaf (y, z2 * az, az) + shift; } #endif // PL_MATH_ATANF_COMMON_H