xref: /freebsd/contrib/arm-optimized-routines/pl/math/v_exp10_2u.c (revision 397e83df75e0fcd0d3fcb95ae4d794cb7600fc89)
1 /*
2  * Double-precision vector 10^x function.
3  *
4  * Copyright (c) 2023, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "mathlib.h"
9 #include "v_math.h"
10 #include "pl_sig.h"
11 #include "pl_test.h"
12 
13 /* Value of |x| above which scale overflows without special treatment.  */
14 #define SpecialBound 306.0 /* floor (log10 (2^1023)) - 1.  */
15 /* Value of n above which scale overflows even with special treatment.  */
16 #define ScaleBound 163840.0 /* 1280.0 * N.  */
17 
18 const static struct data
19 {
20   float64x2_t poly[4];
21   float64x2_t log10_2, log2_10_hi, log2_10_lo, shift;
22 #if !WANT_SIMD_EXCEPT
23   float64x2_t special_bound, scale_thresh;
24 #endif
25 } data = {
26   /* Coefficients generated using Remez algorithm.
27      rel error: 0x1.5ddf8f28p-54
28      abs error: 0x1.5ed266c8p-54 in [ -log10(2)/256, log10(2)/256 ]
29      maxerr: 1.14432 +0.5 ulp.  */
30   .poly = { V2 (0x1.26bb1bbb5524p1), V2 (0x1.53524c73cecdap1),
31 	    V2 (0x1.047060efb781cp1), V2 (0x1.2bd76040f0d16p0) },
32   .log10_2 = V2 (0x1.a934f0979a371p8),	   /* N/log2(10).  */
33   .log2_10_hi = V2 (0x1.34413509f79ffp-9), /* log2(10)/N.  */
34   .log2_10_lo = V2 (-0x1.9dc1da994fd21p-66),
35   .shift = V2 (0x1.8p+52),
36 #if !WANT_SIMD_EXCEPT
37   .scale_thresh = V2 (ScaleBound),
38   .special_bound = V2 (SpecialBound),
39 #endif
40 };
41 
42 #define N (1 << V_EXP_TABLE_BITS)
43 #define IndexMask v_u64 (N - 1)
44 
45 #if WANT_SIMD_EXCEPT
46 
47 # define TinyBound v_u64 (0x2000000000000000) /* asuint64 (0x1p-511).  */
48 # define BigBound v_u64 (0x4070000000000000)  /* asuint64 (0x1p8).  */
49 # define Thres v_u64 (0x2070000000000000)     /* BigBound - TinyBound.  */
50 
51 static float64x2_t VPCS_ATTR NOINLINE
52 special_case (float64x2_t x, float64x2_t y, uint64x2_t cmp)
53 {
54   /* If fenv exceptions are to be triggered correctly, fall back to the scalar
55      routine for special lanes.  */
56   return v_call_f64 (exp10, x, y, cmp);
57 }
58 
59 #else
60 
61 # define SpecialOffset v_u64 (0x6000000000000000) /* 0x1p513.  */
62 /* SpecialBias1 + SpecialBias1 = asuint(1.0).  */
63 # define SpecialBias1 v_u64 (0x7000000000000000)  /* 0x1p769.  */
64 # define SpecialBias2 v_u64 (0x3010000000000000)  /* 0x1p-254.  */
65 
66 static inline float64x2_t VPCS_ATTR
67 special_case (float64x2_t s, float64x2_t y, float64x2_t n,
68 	      const struct data *d)
69 {
70   /* 2^(n/N) may overflow, break it up into s1*s2.  */
71   uint64x2_t b = vandq_u64 (vcltzq_f64 (n), SpecialOffset);
72   float64x2_t s1 = vreinterpretq_f64_u64 (vsubq_u64 (SpecialBias1, b));
73   float64x2_t s2 = vreinterpretq_f64_u64 (
74       vaddq_u64 (vsubq_u64 (vreinterpretq_u64_f64 (s), SpecialBias2), b));
75   uint64x2_t cmp = vcagtq_f64 (n, d->scale_thresh);
76   float64x2_t r1 = vmulq_f64 (s1, s1);
77   float64x2_t r0 = vmulq_f64 (vfmaq_f64 (s2, y, s2), s1);
78   return vbslq_f64 (cmp, r1, r0);
79 }
80 
81 #endif
82 
83 /* Fast vector implementation of exp10.
84    Maximum measured error is 1.64 ulp.
85    _ZGVnN2v_exp10(0x1.ccd1c9d82cc8cp+0) got 0x1.f8dab6d7fed0cp+5
86 				       want 0x1.f8dab6d7fed0ap+5.  */
87 float64x2_t VPCS_ATTR V_NAME_D1 (exp10) (float64x2_t x)
88 {
89   const struct data *d = ptr_barrier (&data);
90   uint64x2_t cmp;
91 #if WANT_SIMD_EXCEPT
92   /* If any lanes are special, mask them with 1 and retain a copy of x to allow
93      special_case to fix special lanes later. This is only necessary if fenv
94      exceptions are to be triggered correctly.  */
95   float64x2_t xm = x;
96   uint64x2_t iax = vreinterpretq_u64_f64 (vabsq_f64 (x));
97   cmp = vcgeq_u64 (vsubq_u64 (iax, TinyBound), Thres);
98   if (unlikely (v_any_u64 (cmp)))
99     x = vbslq_f64 (cmp, v_f64 (1), x);
100 #else
101   cmp = vcageq_f64 (x, d->special_bound);
102 #endif
103 
104   /* n = round(x/(log10(2)/N)).  */
105   float64x2_t z = vfmaq_f64 (d->shift, x, d->log10_2);
106   uint64x2_t u = vreinterpretq_u64_f64 (z);
107   float64x2_t n = vsubq_f64 (z, d->shift);
108 
109   /* r = x - n*log10(2)/N.  */
110   float64x2_t r = x;
111   r = vfmsq_f64 (r, d->log2_10_hi, n);
112   r = vfmsq_f64 (r, d->log2_10_lo, n);
113 
114   uint64x2_t e = vshlq_n_u64 (u, 52 - V_EXP_TABLE_BITS);
115   uint64x2_t i = vandq_u64 (u, IndexMask);
116 
117   /* y = exp10(r) - 1 ~= C0 r + C1 r^2 + C2 r^3 + C3 r^4.  */
118   float64x2_t r2 = vmulq_f64 (r, r);
119   float64x2_t p = vfmaq_f64 (d->poly[0], r, d->poly[1]);
120   float64x2_t y = vfmaq_f64 (d->poly[2], r, d->poly[3]);
121   p = vfmaq_f64 (p, y, r2);
122   y = vmulq_f64 (r, p);
123 
124   /* s = 2^(n/N).  */
125   u = v_lookup_u64 (__v_exp_data, i);
126   float64x2_t s = vreinterpretq_f64_u64 (vaddq_u64 (u, e));
127 
128   if (unlikely (v_any_u64 (cmp)))
129 #if WANT_SIMD_EXCEPT
130     return special_case (xm, vfmaq_f64 (s, y, s), cmp);
131 #else
132     return special_case (s, y, n, d);
133 #endif
134 
135   return vfmaq_f64 (s, y, s);
136 }
137 
138 PL_SIG (S, D, 1, exp10, -9.9, 9.9)
139 PL_SIG (V, D, 1, exp10, -9.9, 9.9)
140 PL_TEST_ULP (V_NAME_D1 (exp10), 1.15)
141 PL_TEST_EXPECT_FENV (V_NAME_D1 (exp10), WANT_SIMD_EXCEPT)
142 PL_TEST_SYM_INTERVAL (V_NAME_D1 (exp10), 0, SpecialBound, 5000)
143 PL_TEST_SYM_INTERVAL (V_NAME_D1 (exp10), SpecialBound, ScaleBound, 5000)
144 PL_TEST_SYM_INTERVAL (V_NAME_D1 (exp10), ScaleBound, inf, 10000)
145