xref: /freebsd/contrib/arm-optimized-routines/pl/math/atanf_common.h (revision 5a02ffc32e777041dd2dad4e651ed2a0865a0a5d)
1 /*
2  * Single-precision polynomial evaluation function for scalar
3  * atan(x) and atan2(y,x).
4  *
5  * Copyright (c) 2021-2023, Arm Limited.
6  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
7  */
8 
9 #ifndef PL_MATH_ATANF_COMMON_H
10 #define PL_MATH_ATANF_COMMON_H
11 
12 #include "math_config.h"
13 #include "poly_scalar_f32.h"
14 
15 /* Polynomial used in fast atanf(x) and atan2f(y,x) implementations
16    The order 7 polynomial P approximates (atan(sqrt(x))-sqrt(x))/x^(3/2).  */
17 static inline float
eval_poly(float z,float az,float shift)18 eval_poly (float z, float az, float shift)
19 {
20   /* Use 2-level Estrin scheme for P(z^2) with deg(P)=7. However,
21      a standard implementation using z8 creates spurious underflow
22      in the very last fma (when z^8 is small enough).
23      Therefore, we split the last fma into a mul and and an fma.
24      Horner and single-level Estrin have higher errors that exceed
25      threshold.  */
26   float z2 = z * z;
27   float z4 = z2 * z2;
28 
29   /* Then assemble polynomial.  */
30   float y = fmaf (
31       z4, z4 * pairwise_poly_3_f32 (z2, z4, __atanf_poly_data.poly + 4),
32       pairwise_poly_3_f32 (z2, z4, __atanf_poly_data.poly));
33   /* Finalize:
34      y = shift + z * P(z^2).  */
35   return fmaf (y, z2 * az, az) + shift;
36 }
37 
38 #endif // PL_MATH_ATANF_COMMON_H
39