xref: /freebsd/contrib/arm-optimized-routines/pl/math/v_cosh_2u.c (revision 7ef62cebc2f965b0f640263e179276928885e33d)
1 /*
2  * Double-precision vector cosh(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 "pl_sig.h"
10 #include "pl_test.h"
11 #include "v_exp_tail.h"
12 
13 #define C1 v_f64 (C1_scal)
14 #define C2 v_f64 (C2_scal)
15 #define C3 v_f64 (C3_scal)
16 #define InvLn2 v_f64 (InvLn2_scal)
17 #define Ln2hi v_f64 (Ln2hi_scal)
18 #define Ln2lo v_f64 (Ln2lo_scal)
19 #define IndexMask v_u64 (IndexMask_scal)
20 #define Shift v_f64 (Shift_scal)
21 #define Thres v_f64 (Thres_scal)
22 
23 #define AbsMask 0x7fffffffffffffff
24 #define Half v_f64 (0.5)
25 #define SpecialBound                                                           \
26   0x4086000000000000 /* 0x1.6p9, above which exp overflows.  */
27 
28 #if V_SUPPORTED
29 
30 static inline v_f64_t
31 exp_inline (v_f64_t x)
32 {
33   /* Helper for approximating exp(x). Copied from v_exp_tail, with no
34      special-case handling or tail.  */
35 
36   /* n = round(x/(ln2/N)).  */
37   v_f64_t z = v_fma_f64 (x, InvLn2, Shift);
38   v_u64_t u = v_as_u64_f64 (z);
39   v_f64_t n = z - Shift;
40 
41   /* r = x - n*ln2/N.  */
42   v_f64_t r = x;
43   r = v_fma_f64 (-Ln2hi, n, r);
44   r = v_fma_f64 (-Ln2lo, n, r);
45 
46   v_u64_t e = u << (52 - V_EXP_TAIL_TABLE_BITS);
47   v_u64_t i = u & IndexMask;
48 
49   /* y = tail + exp(r) - 1 ~= r + C1 r^2 + C2 r^3 + C3 r^4.  */
50   v_f64_t y = v_fma_f64 (C3, r, C2);
51   y = v_fma_f64 (y, r, C1);
52   y = v_fma_f64 (y, r, v_f64 (1)) * r;
53 
54   /* s = 2^(n/N).  */
55   u = v_lookup_u64 (Tab, i);
56   v_f64_t s = v_as_f64_u64 (u + e);
57 
58   return v_fma_f64 (y, s, s);
59 }
60 
61 /* Approximation for vector double-precision cosh(x) using exp_inline.
62    cosh(x) = (exp(x) + exp(-x)) / 2.
63    The greatest observed error is in the scalar fall-back region, so is the same
64    as the scalar routine, 1.93 ULP:
65    __v_cosh(0x1.628af341989dap+9) got 0x1.fdf28623ef921p+1021
66 				 want 0x1.fdf28623ef923p+1021.
67 
68    The greatest observed error in the non-special region is 1.54 ULP:
69    __v_cosh(0x1.8e205b6ecacf7p+2) got 0x1.f711dcb0c77afp+7
70 				 want 0x1.f711dcb0c77b1p+7.  */
71 VPCS_ATTR v_f64_t V_NAME (cosh) (v_f64_t x)
72 {
73   v_u64_t ix = v_as_u64_f64 (x);
74   v_u64_t iax = ix & AbsMask;
75   v_u64_t special = v_cond_u64 (iax > SpecialBound);
76 
77   /* If any inputs are special, fall back to scalar for all lanes.  */
78   if (unlikely (v_any_u64 (special)))
79     return v_call_f64 (cosh, x, x, v_u64 (-1));
80 
81   v_f64_t ax = v_as_f64_u64 (iax);
82   /* Up to the point that exp overflows, we can use it to calculate cosh by
83      exp(|x|) / 2 + 1 / (2 * exp(|x|)).  */
84   v_f64_t t = exp_inline (ax);
85   return t * Half + Half / t;
86 }
87 VPCS_ALIAS
88 
89 PL_SIG (V, D, 1, cosh, -10.0, 10.0)
90 PL_TEST_ULP (V_NAME (cosh), 1.43)
91 PL_TEST_EXPECT_FENV_ALWAYS (V_NAME (cosh))
92 PL_TEST_INTERVAL (V_NAME (cosh), 0, 0x1.6p9, 100000)
93 PL_TEST_INTERVAL (V_NAME (cosh), -0, -0x1.6p9, 100000)
94 PL_TEST_INTERVAL (V_NAME (cosh), 0x1.6p9, inf, 1000)
95 PL_TEST_INTERVAL (V_NAME (cosh), -0x1.6p9, -inf, 1000)
96 #endif
97