xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/advsimd/exp10.c (revision dd21556857e8d40f66bf5ad54754d9d52669ebf7)
1 /*
2  * Double-precision vector 10^x function.
3  *
4  * Copyright (c) 2023-2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #define _GNU_SOURCE
9 #include "mathlib.h"
10 #include "v_math.h"
11 #include "test_sig.h"
12 #include "test_defs.h"
13 
14 /* Value of |x| above which scale overflows without special treatment.  */
15 #define SpecialBound 306.0 /* floor (log10 (2^1023)) - 1.  */
16 /* Value of n above which scale overflows even with special treatment.  */
17 #define ScaleBound 163840.0 /* 1280.0 * N.  */
18 
19 const static struct data
20 {
21   float64x2_t poly[4];
22   float64x2_t log10_2, log2_10_hi, log2_10_lo, shift;
23 #if !WANT_SIMD_EXCEPT
24   float64x2_t special_bound, scale_thresh;
25 #endif
26 } data = {
27   /* Coefficients generated using Remez algorithm.
28      rel error: 0x1.5ddf8f28p-54
29      abs error: 0x1.5ed266c8p-54 in [ -log10(2)/256, log10(2)/256 ]
30      maxerr: 1.14432 +0.5 ulp.  */
31   .poly = { V2 (0x1.26bb1bbb5524p1), V2 (0x1.53524c73cecdap1),
32 	    V2 (0x1.047060efb781cp1), V2 (0x1.2bd76040f0d16p0) },
33   .log10_2 = V2 (0x1.a934f0979a371p8),	   /* N/log2(10).  */
34   .log2_10_hi = V2 (0x1.34413509f79ffp-9), /* log2(10)/N.  */
35   .log2_10_lo = V2 (-0x1.9dc1da994fd21p-66),
36   .shift = V2 (0x1.8p+52),
37 #if !WANT_SIMD_EXCEPT
38   .scale_thresh = V2 (ScaleBound),
39   .special_bound = V2 (SpecialBound),
40 #endif
41 };
42 
43 #define N (1 << V_EXP_TABLE_BITS)
44 #define IndexMask v_u64 (N - 1)
45 
46 #if WANT_SIMD_EXCEPT
47 
48 # define TinyBound v_u64 (0x2000000000000000) /* asuint64 (0x1p-511).  */
49 # define BigBound v_u64 (0x4070000000000000)  /* asuint64 (0x1p8).  */
50 # define Thres v_u64 (0x2070000000000000)     /* BigBound - TinyBound.  */
51 
52 static float64x2_t VPCS_ATTR NOINLINE
53 special_case (float64x2_t x, float64x2_t y, uint64x2_t cmp)
54 {
55   /* If fenv exceptions are to be triggered correctly, fall back to the scalar
56      routine for special lanes.  */
57   return v_call_f64 (exp10, x, y, cmp);
58 }
59 
60 #else
61 
62 # define SpecialOffset v_u64 (0x6000000000000000) /* 0x1p513.  */
63 /* SpecialBias1 + SpecialBias1 = asuint(1.0).  */
64 # define SpecialBias1 v_u64 (0x7000000000000000)  /* 0x1p769.  */
65 # define SpecialBias2 v_u64 (0x3010000000000000)  /* 0x1p-254.  */
66 
67 static inline float64x2_t VPCS_ATTR
68 special_case (float64x2_t s, float64x2_t y, float64x2_t n,
69 	      const struct data *d)
70 {
71   /* 2^(n/N) may overflow, break it up into s1*s2.  */
72   uint64x2_t b = vandq_u64 (vcltzq_f64 (n), SpecialOffset);
73   float64x2_t s1 = vreinterpretq_f64_u64 (vsubq_u64 (SpecialBias1, b));
74   float64x2_t s2 = vreinterpretq_f64_u64 (
75       vaddq_u64 (vsubq_u64 (vreinterpretq_u64_f64 (s), SpecialBias2), b));
76   uint64x2_t cmp = vcagtq_f64 (n, d->scale_thresh);
77   float64x2_t r1 = vmulq_f64 (s1, s1);
78   float64x2_t r0 = vmulq_f64 (vfmaq_f64 (s2, y, s2), s1);
79   return vbslq_f64 (cmp, r1, r0);
80 }
81 
82 #endif
83 
84 /* Fast vector implementation of exp10.
85    Maximum measured error is 1.64 ulp.
86    _ZGVnN2v_exp10(0x1.ccd1c9d82cc8cp+0) got 0x1.f8dab6d7fed0cp+5
87 				       want 0x1.f8dab6d7fed0ap+5.  */
88 float64x2_t VPCS_ATTR V_NAME_D1 (exp10) (float64x2_t x)
89 {
90   const struct data *d = ptr_barrier (&data);
91   uint64x2_t cmp;
92 #if WANT_SIMD_EXCEPT
93   /* If any lanes are special, mask them with 1 and retain a copy of x to allow
94      special_case to fix special lanes later. This is only necessary if fenv
95      exceptions are to be triggered correctly.  */
96   float64x2_t xm = x;
97   uint64x2_t iax = vreinterpretq_u64_f64 (vabsq_f64 (x));
98   cmp = vcgeq_u64 (vsubq_u64 (iax, TinyBound), Thres);
99   if (unlikely (v_any_u64 (cmp)))
100     x = vbslq_f64 (cmp, v_f64 (1), x);
101 #else
102   cmp = vcageq_f64 (x, d->special_bound);
103 #endif
104 
105   /* n = round(x/(log10(2)/N)).  */
106   float64x2_t z = vfmaq_f64 (d->shift, x, d->log10_2);
107   uint64x2_t u = vreinterpretq_u64_f64 (z);
108   float64x2_t n = vsubq_f64 (z, d->shift);
109 
110   /* r = x - n*log10(2)/N.  */
111   float64x2_t r = x;
112   r = vfmsq_f64 (r, d->log2_10_hi, n);
113   r = vfmsq_f64 (r, d->log2_10_lo, n);
114 
115   uint64x2_t e = vshlq_n_u64 (u, 52 - V_EXP_TABLE_BITS);
116   uint64x2_t i = vandq_u64 (u, IndexMask);
117 
118   /* y = exp10(r) - 1 ~= C0 r + C1 r^2 + C2 r^3 + C3 r^4.  */
119   float64x2_t r2 = vmulq_f64 (r, r);
120   float64x2_t p = vfmaq_f64 (d->poly[0], r, d->poly[1]);
121   float64x2_t y = vfmaq_f64 (d->poly[2], r, d->poly[3]);
122   p = vfmaq_f64 (p, y, r2);
123   y = vmulq_f64 (r, p);
124 
125   /* s = 2^(n/N).  */
126   u = v_lookup_u64 (__v_exp_data, i);
127   float64x2_t s = vreinterpretq_f64_u64 (vaddq_u64 (u, e));
128 
129   if (unlikely (v_any_u64 (cmp)))
130 #if WANT_SIMD_EXCEPT
131     return special_case (xm, vfmaq_f64 (s, y, s), cmp);
132 #else
133     return special_case (s, y, n, d);
134 #endif
135 
136   return vfmaq_f64 (s, y, s);
137 }
138 
139 #if WANT_EXP10_TESTS
140 TEST_SIG (S, D, 1, exp10, -9.9, 9.9)
141 TEST_SIG (V, D, 1, exp10, -9.9, 9.9)
142 TEST_ULP (V_NAME_D1 (exp10), 1.15)
143 TEST_DISABLE_FENV_IF_NOT (V_NAME_D1 (exp10), WANT_SIMD_EXCEPT)
144 TEST_SYM_INTERVAL (V_NAME_D1 (exp10), 0, SpecialBound, 5000)
145 TEST_SYM_INTERVAL (V_NAME_D1 (exp10), SpecialBound, ScaleBound, 5000)
146 TEST_SYM_INTERVAL (V_NAME_D1 (exp10), ScaleBound, inf, 10000)
147 #endif
148