1 /*
2  * Double-precision vector asinh(x) function.
3  *
4  * Copyright (c) 2022-2023, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "v_math.h"
9 #include "estrin.h"
10 #include "pl_sig.h"
11 #include "pl_test.h"
12 
13 #if V_SUPPORTED
14 
15 #define OneTop 0x3ff	/* top12(asuint64(1.0f)).  */
16 #define HugeBound 0x5fe /* top12(asuint64(0x1p511)).  */
17 #define TinyBound 0x3e5 /* top12(asuint64(0x1p-26)).  */
18 #define AbsMask v_u64 (0x7fffffffffffffff)
19 #define C(i) v_f64 (__asinh_data.poly[i])
20 
21 /* Constants & data for log.  */
22 #define OFF 0x3fe6000000000000
23 #define Ln2 v_f64 (0x1.62e42fefa39efp-1)
24 #define A(i) v_f64 (__sv_log_data.poly[i])
25 #define T(i) __log_data.tab[i]
26 #define N (1 << LOG_TABLE_BITS)
27 
28 static NOINLINE v_f64_t
29 special_case (v_f64_t x, v_f64_t y, v_u64_t special)
30 {
31   return v_call_f64 (asinh, x, y, special);
32 }
33 
34 struct entry
35 {
36   v_f64_t invc;
37   v_f64_t logc;
38 };
39 
40 static inline struct entry
41 lookup (v_u64_t i)
42 {
43   struct entry e;
44 #ifdef SCALAR
45   e.invc = T (i).invc;
46   e.logc = T (i).logc;
47 #else
48   e.invc[0] = T (i[0]).invc;
49   e.logc[0] = T (i[0]).logc;
50   e.invc[1] = T (i[1]).invc;
51   e.logc[1] = T (i[1]).logc;
52 #endif
53   return e;
54 }
55 
56 static inline v_f64_t
57 log_inline (v_f64_t x)
58 {
59   /* Double-precision vector log, copied from math/v_log.c with some cosmetic
60      modification and special-cases removed. See that file for details of the
61      algorithm used.  */
62   v_u64_t ix = v_as_u64_f64 (x);
63   v_u64_t tmp = ix - OFF;
64   v_u64_t i = (tmp >> (52 - LOG_TABLE_BITS)) % N;
65   v_s64_t k = v_as_s64_u64 (tmp) >> 52;
66   v_u64_t iz = ix - (tmp & 0xfffULL << 52);
67   v_f64_t z = v_as_f64_u64 (iz);
68   struct entry e = lookup (i);
69   v_f64_t r = v_fma_f64 (z, e.invc, v_f64 (-1.0));
70   v_f64_t kd = v_to_f64_s64 (k);
71   v_f64_t hi = v_fma_f64 (kd, Ln2, e.logc + r);
72   v_f64_t r2 = r * r;
73   v_f64_t y = v_fma_f64 (A (3), r, A (2));
74   v_f64_t p = v_fma_f64 (A (1), r, A (0));
75   y = v_fma_f64 (A (4), r2, y);
76   y = v_fma_f64 (y, r2, p);
77   y = v_fma_f64 (y, r2, hi);
78   return y;
79 }
80 
81 /* Double-precision implementation of vector asinh(x).
82    asinh is very sensitive around 1, so it is impractical to devise a single
83    low-cost algorithm which is sufficiently accurate on a wide range of input.
84    Instead we use two different algorithms:
85    asinh(x) = sign(x) * log(|x| + sqrt(x^2 + 1)      if |x| >= 1
86 	    = sign(x) * (|x| + |x|^3 * P(x^2))       otherwise
87    where log(x) is an optimized log approximation, and P(x) is a polynomial
88    shared with the scalar routine. The greatest observed error 3.29 ULP, in
89    |x| >= 1:
90    __v_asinh(0x1.2cd9d717e2c9bp+0) got 0x1.ffffcfd0e234fp-1
91 				  want 0x1.ffffcfd0e2352p-1.  */
92 VPCS_ATTR v_f64_t V_NAME (asinh) (v_f64_t x)
93 {
94   v_u64_t ix = v_as_u64_f64 (x);
95   v_u64_t iax = ix & AbsMask;
96   v_f64_t ax = v_as_f64_u64 (iax);
97   v_u64_t top12 = iax >> 52;
98 
99   v_u64_t gt1 = v_cond_u64 (top12 >= OneTop);
100   v_u64_t special = v_cond_u64 (top12 >= HugeBound);
101 
102 #if WANT_SIMD_EXCEPT
103   v_u64_t tiny = v_cond_u64 (top12 < TinyBound);
104   special |= tiny;
105 #endif
106 
107   /* Option 1: |x| >= 1.
108      Compute asinh(x) according by asinh(x) = log(x + sqrt(x^2 + 1)).
109      If WANT_SIMD_EXCEPT is enabled, sidestep special values, which will
110      overflow, by setting special lanes to 1. These will be fixed later.  */
111   v_f64_t option_1 = v_f64 (0);
112   if (likely (v_any_u64 (gt1)))
113     {
114 #if WANT_SIMD_EXCEPT
115       v_f64_t xm = v_sel_f64 (special, v_f64 (1), ax);
116 #else
117       v_f64_t xm = ax;
118 #endif
119       option_1 = log_inline (xm + v_sqrt_f64 (xm * xm + 1));
120     }
121 
122   /* Option 2: |x| < 1.
123      Compute asinh(x) using a polynomial.
124      If WANT_SIMD_EXCEPT is enabled, sidestep special lanes, which will
125      overflow, and tiny lanes, which will underflow, by setting them to 0. They
126      will be fixed later, either by selecting x or falling back to the scalar
127      special-case. The largest observed error in this region is 1.47 ULPs:
128      __v_asinh(0x1.fdfcd00cc1e6ap-1) got 0x1.c1d6bf874019bp-1
129 				    want 0x1.c1d6bf874019cp-1.  */
130   v_f64_t option_2 = v_f64 (0);
131   if (likely (v_any_u64 (~gt1)))
132     {
133 #if WANT_SIMD_EXCEPT
134       ax = v_sel_f64 (tiny | gt1, v_f64 (0), ax);
135 #endif
136       v_f64_t x2 = ax * ax;
137       v_f64_t z2 = x2 * x2;
138       v_f64_t z4 = z2 * z2;
139       v_f64_t z8 = z4 * z4;
140       v_f64_t p = ESTRIN_17 (x2, z2, z4, z8, z8 * z8, C);
141       option_2 = v_fma_f64 (p, x2 * ax, ax);
142 #if WANT_SIMD_EXCEPT
143       option_2 = v_sel_f64 (tiny, x, option_2);
144 #endif
145     }
146 
147   /* Choose the right option for each lane.  */
148   v_f64_t y = v_sel_f64 (gt1, option_1, option_2);
149   /* Copy sign.  */
150   y = v_as_f64_u64 (v_bsl_u64 (AbsMask, v_as_u64_f64 (y), ix));
151 
152   if (unlikely (v_any_u64 (special)))
153     return special_case (x, y, special);
154   return y;
155 }
156 VPCS_ALIAS
157 
158 PL_SIG (V, D, 1, asinh, -10.0, 10.0)
159 PL_TEST_ULP (V_NAME (asinh), 2.80)
160 PL_TEST_EXPECT_FENV (V_NAME (asinh), WANT_SIMD_EXCEPT)
161 /* Test vector asinh 3 times, with control lane < 1, > 1 and special.
162    Ensures the v_sel is choosing the right option in all cases.  */
163 #define V_ASINH_INTERVAL(lo, hi, n)                                            \
164   PL_TEST_INTERVAL_C (V_NAME (asinh), lo, hi, n, 0.5)                          \
165   PL_TEST_INTERVAL_C (V_NAME (asinh), lo, hi, n, 2)                            \
166   PL_TEST_INTERVAL_C (V_NAME (asinh), lo, hi, n, 0x1p600)
167 V_ASINH_INTERVAL (0, 0x1p-26, 50000)
168 V_ASINH_INTERVAL (0x1p-26, 1, 50000)
169 V_ASINH_INTERVAL (1, 0x1p511, 50000)
170 V_ASINH_INTERVAL (0x1p511, inf, 40000)
171 V_ASINH_INTERVAL (-0, -0x1p-26, 50000)
172 V_ASINH_INTERVAL (-0x1p-26, -1, 50000)
173 V_ASINH_INTERVAL (-1, -0x1p511, 50000)
174 V_ASINH_INTERVAL (-0x1p511, -inf, 40000)
175 #endif
176