1 /*
2 * Single-precision polynomial evaluation function for SVE atan(x) and
3 * 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_SV_ATANF_COMMON_H
10 #define PL_MATH_SV_ATANF_COMMON_H
11
12 #include "math_config.h"
13 #include "sv_math.h"
14
15 #define P(i) sv_f32 (__atanf_poly_data.poly[i])
16
17 /* Polynomial used in fast SVE atanf(x) and atan2f(y,x) implementations
18 The order 7 polynomial P approximates (f(sqrt(x))-sqrt(x))/x^(3/2). */
19 static inline sv_f32_t
__sv_atanf_common(svbool_t pg,svbool_t red,sv_f32_t z,sv_f32_t az,sv_f32_t shift)20 __sv_atanf_common (svbool_t pg, svbool_t red, sv_f32_t z, sv_f32_t az,
21 sv_f32_t shift)
22 {
23 /* Use full Estrin scheme for P(z^2) with deg(P)=7. */
24
25 /* First compute square powers of z. */
26 sv_f32_t z2 = svmul_f32_x (pg, z, z);
27 sv_f32_t z4 = svmul_f32_x (pg, z2, z2);
28 sv_f32_t z8 = svmul_f32_x (pg, z4, z4);
29
30 /* Then assemble polynomial. */
31 sv_f32_t p_4_7 = sv_fma_f32_x (pg, z4, (sv_fma_f32_x (pg, z2, P (7), P (6))),
32 (sv_fma_f32_x (pg, z2, P (5), P (4))));
33 sv_f32_t p_0_3 = sv_fma_f32_x (pg, z4, (sv_fma_f32_x (pg, z2, P (3), P (2))),
34 (sv_fma_f32_x (pg, z2, P (1), P (0))));
35 sv_f32_t y = sv_fma_f32_x (pg, z8, p_4_7, p_0_3);
36
37 /* Finalize. y = shift + z + z^3 * P(z^2). */
38 sv_f32_t z3 = svmul_f32_x (pg, z2, az);
39 y = sv_fma_f32_x (pg, y, z3, az);
40
41 /* Apply shift as indicated by 'red' predicate. */
42 y = svadd_f32_m (red, y, shift);
43
44 return y;
45 }
46
47 #endif // PL_MATH_SV_ATANF_COMMON_H
48