xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/advsimd/asinh.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1 /*
2  * Double-precision vector asinh(x) function.
3  *
4  * Copyright (c) 2022-2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "test_defs.h"
9 #include "test_sig.h"
10 #include "v_math.h"
11 
12 const static struct data
13 {
14   uint64x2_t huge_bound, abs_mask, off, mask;
15 #if WANT_SIMD_EXCEPT
16   float64x2_t tiny_bound;
17 #endif
18   float64x2_t lc0, lc2;
19   double lc1, lc3, ln2, lc4;
20 
21   float64x2_t c0, c2, c4, c6, c8, c10, c12, c14, c16, c17;
22   double c1, c3, c5, c7, c9, c11, c13, c15;
23 
24 } data = {
25 
26 #if WANT_SIMD_EXCEPT
27   .tiny_bound = V2 (0x1p-26),
28 #endif
29   /* Even terms of polynomial s.t. asinh(x) is approximated by
30      asinh(x) ~= x + x^3 * (C0 + C1 * x + C2 * x^2 + C3 * x^3 + ...).
31      Generated using Remez, f = (asinh(sqrt(x)) - sqrt(x))/x^(3/2).  */
32 
33   .c0 = V2 (-0x1.55555555554a7p-3),
34   .c1 = 0x1.3333333326c7p-4,
35   .c2 = V2 (-0x1.6db6db68332e6p-5),
36   .c3 = 0x1.f1c71b26fb40dp-6,
37   .c4 = V2 (-0x1.6e8b8b654a621p-6),
38   .c5 = 0x1.1c4daa9e67871p-6,
39   .c6 = V2 (-0x1.c9871d10885afp-7),
40   .c7 = 0x1.7a16e8d9d2ecfp-7,
41   .c8 = V2 (-0x1.3ddca533e9f54p-7),
42   .c9 = 0x1.0becef748dafcp-7,
43   .c10 = V2 (-0x1.b90c7099dd397p-8),
44   .c11 = 0x1.541f2bb1ffe51p-8,
45   .c12 = V2 (-0x1.d217026a669ecp-9),
46   .c13 = 0x1.0b5c7977aaf7p-9,
47   .c14 = V2 (-0x1.e0f37daef9127p-11),
48   .c15 = 0x1.388b5fe542a6p-12,
49   .c16 = V2 (-0x1.021a48685e287p-14),
50   .c17 = V2 (0x1.93d4ba83d34dap-18),
51 
52   .lc0 = V2 (-0x1.ffffffffffff7p-2),
53   .lc1 = 0x1.55555555170d4p-2,
54   .lc2 = V2 (-0x1.0000000399c27p-2),
55   .lc3 = 0x1.999b2e90e94cap-3,
56   .lc4 = -0x1.554e550bd501ep-3,
57   .ln2 = 0x1.62e42fefa39efp-1,
58 
59   .off = V2 (0x3fe6900900000000),
60   .huge_bound = V2 (0x5fe0000000000000),
61   .abs_mask = V2 (0x7fffffffffffffff),
62   .mask = V2 (0xfffULL << 52),
63 };
64 
65 static float64x2_t NOINLINE VPCS_ATTR
special_case(float64x2_t x,float64x2_t y,uint64x2_t abs_mask,uint64x2_t special)66 special_case (float64x2_t x, float64x2_t y, uint64x2_t abs_mask,
67 	      uint64x2_t special)
68 {
69   /* Copy sign.  */
70   y = vbslq_f64 (abs_mask, y, x);
71   return v_call_f64 (asinh, x, y, special);
72 }
73 
74 #define N (1 << V_LOG_TABLE_BITS)
75 #define IndexMask (N - 1)
76 
77 struct entry
78 {
79   float64x2_t invc;
80   float64x2_t logc;
81 };
82 
83 static inline struct entry
lookup(uint64x2_t i)84 lookup (uint64x2_t i)
85 {
86   /* Since N is a power of 2, n % N = n & (N - 1).  */
87   struct entry e;
88   uint64_t i0 = (vgetq_lane_u64 (i, 0) >> (52 - V_LOG_TABLE_BITS)) & IndexMask;
89   uint64_t i1 = (vgetq_lane_u64 (i, 1) >> (52 - V_LOG_TABLE_BITS)) & IndexMask;
90   float64x2_t e0 = vld1q_f64 (&__v_log_data.table[i0].invc);
91   float64x2_t e1 = vld1q_f64 (&__v_log_data.table[i1].invc);
92   e.invc = vuzp1q_f64 (e0, e1);
93   e.logc = vuzp2q_f64 (e0, e1);
94   return e;
95 }
96 
97 static inline float64x2_t
log_inline(float64x2_t xm,const struct data * d)98 log_inline (float64x2_t xm, const struct data *d)
99 {
100 
101   uint64x2_t u = vreinterpretq_u64_f64 (xm);
102   uint64x2_t u_off = vsubq_u64 (u, d->off);
103 
104   int64x2_t k = vshrq_n_s64 (vreinterpretq_s64_u64 (u_off), 52);
105   uint64x2_t iz = vsubq_u64 (u, vandq_u64 (u_off, d->mask));
106   float64x2_t z = vreinterpretq_f64_u64 (iz);
107 
108   struct entry e = lookup (u_off);
109 
110   /* log(x) = log1p(z/c-1) + log(c) + k*Ln2.  */
111   float64x2_t r = vfmaq_f64 (v_f64 (-1.0), z, e.invc);
112   float64x2_t kd = vcvtq_f64_s64 (k);
113 
114   /* hi = r + log(c) + k*Ln2.  */
115   float64x2_t ln2_and_lc4 = vld1q_f64 (&d->ln2);
116   float64x2_t hi = vfmaq_laneq_f64 (vaddq_f64 (e.logc, r), kd, ln2_and_lc4, 0);
117 
118   /* y = r2*(A0 + r*A1 + r2*(A2 + r*A3 + r2*A4)) + hi.  */
119   float64x2_t odd_coeffs = vld1q_f64 (&d->lc1);
120   float64x2_t r2 = vmulq_f64 (r, r);
121   float64x2_t y = vfmaq_laneq_f64 (d->lc2, r, odd_coeffs, 1);
122   float64x2_t p = vfmaq_laneq_f64 (d->lc0, r, odd_coeffs, 0);
123   y = vfmaq_laneq_f64 (y, r2, ln2_and_lc4, 1);
124   y = vfmaq_f64 (p, r2, y);
125   return vfmaq_f64 (hi, y, r2);
126 }
127 
128 /* Double-precision implementation of vector asinh(x).
129    asinh is very sensitive around 1, so it is impractical to devise a single
130    low-cost algorithm which is sufficiently accurate on a wide range of input.
131    Instead we use two different algorithms:
132    asinh(x) = sign(x) * log(|x| + sqrt(x^2 + 1)      if |x| >= 1
133 	    = sign(x) * (|x| + |x|^3 * P(x^2))       otherwise
134    where log(x) is an optimized log approximation, and P(x) is a polynomial
135    shared with the scalar routine. The greatest observed error 2.79 ULP, in
136    |x| >= 1:
137    _ZGVnN2v_asinh(0x1.2cd9d73ea76a6p+0) got 0x1.ffffd003219dap-1
138 				       want  0x1.ffffd003219ddp-1.  */
V_NAME_D1(asinh)139 VPCS_ATTR float64x2_t V_NAME_D1 (asinh) (float64x2_t x)
140 {
141   const struct data *d = ptr_barrier (&data);
142   float64x2_t ax = vabsq_f64 (x);
143 
144   uint64x2_t gt1 = vcgeq_f64 (ax, v_f64 (1));
145 
146 #if WANT_SIMD_EXCEPT
147   uint64x2_t iax = vreinterpretq_u64_f64 (ax);
148   uint64x2_t special = vcgeq_u64 (iax, (d->huge_bound));
149   uint64x2_t tiny = vcltq_f64 (ax, d->tiny_bound);
150   special = vorrq_u64 (special, tiny);
151 #else
152   uint64x2_t special = vcgeq_f64 (ax, vreinterpretq_f64_u64 (d->huge_bound));
153 #endif
154 
155   /* Option 1: |x| >= 1.
156      Compute asinh(x) according by asinh(x) = log(x + sqrt(x^2 + 1)).
157      If WANT_SIMD_EXCEPT is enabled, sidestep special values, which will
158      overflow, by setting special lanes to 1. These will be fixed later.  */
159   float64x2_t option_1 = v_f64 (0);
160   if (likely (v_any_u64 (gt1)))
161     {
162 #if WANT_SIMD_EXCEPT
163       float64x2_t xm = v_zerofy_f64 (ax, special);
164 #else
165       float64x2_t xm = ax;
166 #endif
167       option_1 = log_inline (
168 	  vaddq_f64 (xm, vsqrtq_f64 (vfmaq_f64 (v_f64 (1), xm, xm))), d);
169     }
170 
171   /* Option 2: |x| < 1.
172      Compute asinh(x) using a polynomial.
173      If WANT_SIMD_EXCEPT is enabled, sidestep special lanes, which will
174      overflow, and tiny lanes, which will underflow, by setting them to 0. They
175      will be fixed later, either by selecting x or falling back to the scalar
176      special-case. The largest observed error in this region is 1.47 ULPs:
177      _ZGVnN2v_asinh(0x1.fdfcd00cc1e6ap-1) got 0x1.c1d6bf874019bp-1
178 					 want 0x1.c1d6bf874019cp-1.  */
179   float64x2_t option_2 = v_f64 (0);
180 
181   if (likely (v_any_u64 (vceqzq_u64 (gt1))))
182     {
183 
184 #if WANT_SIMD_EXCEPT
185       ax = v_zerofy_f64 (ax, vorrq_u64 (tiny, gt1));
186 #endif
187       float64x2_t x2 = vmulq_f64 (ax, ax), z2 = vmulq_f64 (x2, x2);
188       /* Order-17 Pairwise Horner scheme.  */
189       float64x2_t c13 = vld1q_f64 (&d->c1);
190       float64x2_t c57 = vld1q_f64 (&d->c5);
191       float64x2_t c911 = vld1q_f64 (&d->c9);
192       float64x2_t c1315 = vld1q_f64 (&d->c13);
193 
194       float64x2_t p01 = vfmaq_laneq_f64 (d->c0, x2, c13, 0);
195       float64x2_t p23 = vfmaq_laneq_f64 (d->c2, x2, c13, 1);
196       float64x2_t p45 = vfmaq_laneq_f64 (d->c4, x2, c57, 0);
197       float64x2_t p67 = vfmaq_laneq_f64 (d->c6, x2, c57, 1);
198       float64x2_t p89 = vfmaq_laneq_f64 (d->c8, x2, c911, 0);
199       float64x2_t p1011 = vfmaq_laneq_f64 (d->c10, x2, c911, 1);
200       float64x2_t p1213 = vfmaq_laneq_f64 (d->c12, x2, c1315, 0);
201       float64x2_t p1415 = vfmaq_laneq_f64 (d->c14, x2, c1315, 1);
202       float64x2_t p1617 = vfmaq_f64 (d->c16, x2, d->c17);
203 
204       float64x2_t p = vfmaq_f64 (p1415, z2, p1617);
205       p = vfmaq_f64 (p1213, z2, p);
206       p = vfmaq_f64 (p1011, z2, p);
207       p = vfmaq_f64 (p89, z2, p);
208 
209       p = vfmaq_f64 (p67, z2, p);
210       p = vfmaq_f64 (p45, z2, p);
211 
212       p = vfmaq_f64 (p23, z2, p);
213 
214       p = vfmaq_f64 (p01, z2, p);
215       option_2 = vfmaq_f64 (ax, p, vmulq_f64 (ax, x2));
216 #if WANT_SIMD_EXCEPT
217       option_2 = vbslq_f64 (tiny, x, option_2);
218 #endif
219     }
220 
221   /* Choose the right option for each lane.  */
222   float64x2_t y = vbslq_f64 (gt1, option_1, option_2);
223   if (unlikely (v_any_u64 (special)))
224     {
225       return special_case (x, y, d->abs_mask, special);
226     }
227   /* Copy sign.  */
228   return vbslq_f64 (d->abs_mask, y, x);
229 }
230 
231 TEST_SIG (V, D, 1, asinh, -10.0, 10.0)
232 TEST_ULP (V_NAME_D1 (asinh), 2.29)
233 TEST_DISABLE_FENV_IF_NOT (V_NAME_D1 (asinh), WANT_SIMD_EXCEPT)
234 TEST_SYM_INTERVAL (V_NAME_D1 (asinh), 0, 0x1p-26, 50000)
235 TEST_SYM_INTERVAL (V_NAME_D1 (asinh), 0x1p-26, 1, 50000)
236 TEST_SYM_INTERVAL (V_NAME_D1 (asinh), 1, 0x1p511, 50000)
237 TEST_SYM_INTERVAL (V_NAME_D1 (asinh), 0x1p511, inf, 40000)
238 /* Test vector asinh 3 times, with control lane < 1, > 1 and special.
239    Ensures the v_sel is choosing the right option in all cases.  */
240 TEST_CONTROL_VALUE (V_NAME_D1 (asinh), 0.5)
241 TEST_CONTROL_VALUE (V_NAME_D1 (asinh), 2)
242 TEST_CONTROL_VALUE (V_NAME_D1 (asinh), 0x1p600)
243