xref: /freebsd/contrib/arm-optimized-routines/pl/math/sv_log1pf_inline.h (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
1 /*
2  * Helper for SVE routines which calculate log(1 + x) and do not
3  * need special-case handling
4  *
5  * Copyright (c) 2023, Arm Limited.
6  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
7  */
8 
9 #ifndef PL_MATH_SV_LOG1PF_INLINE_H
10 #define PL_MATH_SV_LOG1PF_INLINE_H
11 
12 #include "v_math.h"
13 #include "math_config.h"
14 #include "poly_sve_f32.h"
15 
16 static const struct sv_log1pf_data
17 {
18   float32_t poly[9];
19   float32_t ln2;
20   float32_t scale_back;
21 } sv_log1pf_data = {
22   /* Polynomial generated using FPMinimax in [-0.25, 0.5].  */
23   .poly = { -0x1p-1f, 0x1.5555aap-2f, -0x1.000038p-2f, 0x1.99675cp-3f,
24 	    -0x1.54ef78p-3f, 0x1.28a1f4p-3f, -0x1.0da91p-3f, 0x1.abcb6p-4f,
25 	    -0x1.6f0d5ep-5f },
26   .scale_back = 0x1.0p-23f,
27   .ln2 = 0x1.62e43p-1f,
28 };
29 
30 static inline svfloat32_t
31 eval_poly (svfloat32_t m, const float32_t *c, svbool_t pg)
32 {
33   svfloat32_t p_12 = svmla_x (pg, sv_f32 (c[0]), m, sv_f32 (c[1]));
34   svfloat32_t m2 = svmul_x (pg, m, m);
35   svfloat32_t q = svmla_x (pg, m, m2, p_12);
36   svfloat32_t p = sv_pw_horner_6_f32_x (pg, m, m2, c + 2);
37   p = svmul_x (pg, m2, p);
38 
39   return svmla_x (pg, q, m2, p);
40 }
41 
42 static inline svfloat32_t
43 sv_log1pf_inline (svfloat32_t x, svbool_t pg)
44 {
45   const struct sv_log1pf_data *d = ptr_barrier (&sv_log1pf_data);
46 
47   svfloat32_t m = svadd_x (pg, x, 1.0f);
48 
49   svint32_t ks = svsub_x (pg, svreinterpret_s32 (m),
50 			  svreinterpret_s32 (svdup_f32 (0.75f)));
51   ks = svand_x (pg, ks, 0xff800000);
52   svuint32_t k = svreinterpret_u32 (ks);
53   svfloat32_t s = svreinterpret_f32 (
54       svsub_x (pg, svreinterpret_u32 (svdup_f32 (4.0f)), k));
55 
56   svfloat32_t m_scale
57       = svreinterpret_f32 (svsub_x (pg, svreinterpret_u32 (x), k));
58   m_scale
59       = svadd_x (pg, m_scale, svmla_x (pg, sv_f32 (-1.0f), sv_f32 (0.25f), s));
60   svfloat32_t p = eval_poly (m_scale, d->poly, pg);
61   svfloat32_t scale_back = svmul_x (pg, svcvt_f32_x (pg, k), d->scale_back);
62   return svmla_x (pg, p, scale_back, d->ln2);
63 }
64 
65 #endif //  PL_MATH_SV_LOG1PF_INLINE_H