xref: /freebsd/contrib/arm-optimized-routines/pl/math/v_log1p_inline.h (revision 5a02ffc32e777041dd2dad4e651ed2a0865a0a5d)
1 /*
2  * Helper for vector double-precision routines which calculate log(1 + x) and do
3  * not need special-case handling
4  *
5  * Copyright (c) 2022-2023, Arm Limited.
6  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
7  */
8 #ifndef PL_MATH_V_LOG1P_INLINE_H
9 #define PL_MATH_V_LOG1P_INLINE_H
10 
11 #include "v_math.h"
12 #include "poly_advsimd_f64.h"
13 
14 struct v_log1p_data
15 {
16   float64x2_t poly[19], ln2[2];
17   uint64x2_t hf_rt2_top, one_m_hf_rt2_top, umask;
18   int64x2_t one_top;
19 };
20 
21 /* Coefficients generated using Remez, deg=20, in [sqrt(2)/2-1, sqrt(2)-1].  */
22 #define V_LOG1P_CONSTANTS_TABLE                                               \
23   {                                                                           \
24     .poly = { V2 (-0x1.ffffffffffffbp-2), V2 (0x1.55555555551a9p-2),          \
25 	      V2 (-0x1.00000000008e3p-2), V2 (0x1.9999999a32797p-3),          \
26 	      V2 (-0x1.555555552fecfp-3), V2 (0x1.249248e071e5ap-3),          \
27 	      V2 (-0x1.ffffff8bf8482p-4), V2 (0x1.c71c8f07da57ap-4),          \
28 	      V2 (-0x1.9999ca4ccb617p-4), V2 (0x1.7459ad2e1dfa3p-4),          \
29 	      V2 (-0x1.554d2680a3ff2p-4), V2 (0x1.3b4c54d487455p-4),          \
30 	      V2 (-0x1.2548a9ffe80e6p-4), V2 (0x1.0f389a24b2e07p-4),          \
31 	      V2 (-0x1.eee4db15db335p-5), V2 (0x1.e95b494d4a5ddp-5),          \
32 	      V2 (-0x1.15fdf07cb7c73p-4), V2 (0x1.0310b70800fcfp-4),          \
33 	      V2 (-0x1.cfa7385bdb37ep-6) },                                   \
34     .ln2 = { V2 (0x1.62e42fefa3800p-1), V2 (0x1.ef35793c76730p-45) },         \
35     .hf_rt2_top = V2 (0x3fe6a09e00000000),                                    \
36     .one_m_hf_rt2_top = V2 (0x00095f6200000000),                              \
37     .umask = V2 (0x000fffff00000000), .one_top = V2 (0x3ff)                   \
38   }
39 
40 #define BottomMask v_u64 (0xffffffff)
41 
42 static inline float64x2_t
log1p_inline(float64x2_t x,const struct v_log1p_data * d)43 log1p_inline (float64x2_t x, const struct v_log1p_data *d)
44 {
45   /* Helper for calculating log(x + 1). Copied from v_log1p_2u5.c, with several
46      modifications:
47      - No special-case handling - this should be dealt with by the caller.
48      - Pairwise Horner polynomial evaluation for improved accuracy.
49      - Optionally simulate the shortcut for k=0, used in the scalar routine,
50        using v_sel, for improved accuracy when the argument to log1p is close to
51        0. This feature is enabled by defining WANT_V_LOG1P_K0_SHORTCUT as 1 in
52        the source of the caller before including this file.
53      See v_log1pf_2u1.c for details of the algorithm.  */
54   float64x2_t m = vaddq_f64 (x, v_f64 (1));
55   uint64x2_t mi = vreinterpretq_u64_f64 (m);
56   uint64x2_t u = vaddq_u64 (mi, d->one_m_hf_rt2_top);
57 
58   int64x2_t ki
59       = vsubq_s64 (vreinterpretq_s64_u64 (vshrq_n_u64 (u, 52)), d->one_top);
60   float64x2_t k = vcvtq_f64_s64 (ki);
61 
62   /* Reduce x to f in [sqrt(2)/2, sqrt(2)].  */
63   uint64x2_t utop = vaddq_u64 (vandq_u64 (u, d->umask), d->hf_rt2_top);
64   uint64x2_t u_red = vorrq_u64 (utop, vandq_u64 (mi, BottomMask));
65   float64x2_t f = vsubq_f64 (vreinterpretq_f64_u64 (u_red), v_f64 (1));
66 
67   /* Correction term c/m.  */
68   float64x2_t cm = vdivq_f64 (vsubq_f64 (x, vsubq_f64 (m, v_f64 (1))), m);
69 
70 #ifndef WANT_V_LOG1P_K0_SHORTCUT
71 #error                                                                         \
72   "Cannot use v_log1p_inline.h without specifying whether you need the k0 shortcut for greater accuracy close to 0"
73 #elif WANT_V_LOG1P_K0_SHORTCUT
74   /* Shortcut if k is 0 - set correction term to 0 and f to x. The result is
75      that the approximation is solely the polynomial.  */
76   uint64x2_t k0 = vceqzq_f64 (k);
77   cm = v_zerofy_f64 (cm, k0);
78   f = vbslq_f64 (k0, x, f);
79 #endif
80 
81   /* Approximate log1p(f) on the reduced input using a polynomial.  */
82   float64x2_t f2 = vmulq_f64 (f, f);
83   float64x2_t p = v_pw_horner_18_f64 (f, f2, d->poly);
84 
85   /* Assemble log1p(x) = k * log2 + log1p(f) + c/m.  */
86   float64x2_t ylo = vfmaq_f64 (cm, k, d->ln2[1]);
87   float64x2_t yhi = vfmaq_f64 (f, k, d->ln2[0]);
88   return vfmaq_f64 (vaddq_f64 (ylo, yhi), f2, p);
89 }
90 
91 #endif // PL_MATH_V_LOG1P_INLINE_H
92