xref: /freebsd/contrib/arm-optimized-routines/pl/math/v_exp10f_2u4.c (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
1 /*
2  * Single-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 #include "poly_advsimd_f32.h"
13 
14 #define ScaleBound 192.0f
15 
16 static const struct data
17 {
18   float32x4_t poly[5];
19   float32x4_t log10_2_and_inv, shift;
20 
21 #if !WANT_SIMD_EXCEPT
22   float32x4_t scale_thresh;
23 #endif
24 } data = {
25   /* Coefficients generated using Remez algorithm with minimisation of relative
26      error.
27      rel error: 0x1.89dafa3p-24
28      abs error: 0x1.167d55p-23 in [-log10(2)/2, log10(2)/2]
29      maxerr: 1.85943 +0.5 ulp.  */
30   .poly = { V4 (0x1.26bb16p+1f), V4 (0x1.5350d2p+1f), V4 (0x1.04744ap+1f),
31 	    V4 (0x1.2d8176p+0f), V4 (0x1.12b41ap-1f) },
32   .shift = V4 (0x1.8p23f),
33 
34   /* Stores constants 1/log10(2), log10(2)_high, log10(2)_low, 0.  */
35   .log10_2_and_inv = { 0x1.a934fp+1, 0x1.344136p-2, -0x1.ec10cp-27, 0 },
36 #if !WANT_SIMD_EXCEPT
37   .scale_thresh = V4 (ScaleBound)
38 #endif
39 };
40 
41 #define ExponentBias v_u32 (0x3f800000)
42 
43 #if WANT_SIMD_EXCEPT
44 
45 # define SpecialBound 38.0f	       /* rint(log10(2^127)).  */
46 # define TinyBound v_u32 (0x20000000) /* asuint (0x1p-63).  */
47 # define BigBound v_u32 (0x42180000)  /* asuint (SpecialBound).  */
48 # define Thres v_u32 (0x22180000)     /* BigBound - TinyBound.  */
49 
50 static float32x4_t VPCS_ATTR NOINLINE
51 special_case (float32x4_t x, float32x4_t y, uint32x4_t cmp)
52 {
53   /* If fenv exceptions are to be triggered correctly, fall back to the scalar
54      routine to special lanes.  */
55   return v_call_f32 (exp10f, x, y, cmp);
56 }
57 
58 #else
59 
60 # define SpecialBound 126.0f /* rint (log2 (2^127 / (1 + sqrt (2)))).  */
61 # define SpecialOffset v_u32 (0x82000000)
62 # define SpecialBias v_u32 (0x7f000000)
63 
64 static float32x4_t VPCS_ATTR NOINLINE
65 special_case (float32x4_t poly, float32x4_t n, uint32x4_t e, uint32x4_t cmp1,
66 	      float32x4_t scale, const struct data *d)
67 {
68   /* 2^n may overflow, break it up into s1*s2.  */
69   uint32x4_t b = vandq_u32 (vclezq_f32 (n), SpecialOffset);
70   float32x4_t s1 = vreinterpretq_f32_u32 (vaddq_u32 (b, SpecialBias));
71   float32x4_t s2 = vreinterpretq_f32_u32 (vsubq_u32 (e, b));
72   uint32x4_t cmp2 = vcagtq_f32 (n, d->scale_thresh);
73   float32x4_t r2 = vmulq_f32 (s1, s1);
74   float32x4_t r1 = vmulq_f32 (vfmaq_f32 (s2, poly, s2), s1);
75   /* Similar to r1 but avoids double rounding in the subnormal range.  */
76   float32x4_t r0 = vfmaq_f32 (scale, poly, scale);
77   float32x4_t r = vbslq_f32 (cmp1, r1, r0);
78   return vbslq_f32 (cmp2, r2, r);
79 }
80 
81 #endif
82 
83 /* Fast vector implementation of single-precision exp10.
84    Algorithm is accurate to 2.36 ULP.
85    _ZGVnN4v_exp10f(0x1.be2b36p+1) got 0x1.7e79c4p+11
86 				 want 0x1.7e79cp+11.  */
87 float32x4_t VPCS_ATTR V_NAME_F1 (exp10) (float32x4_t x)
88 {
89   const struct data *d = ptr_barrier (&data);
90 #if WANT_SIMD_EXCEPT
91   /* asuint(x) - TinyBound >= BigBound - TinyBound.  */
92   uint32x4_t cmp = vcgeq_u32 (
93       vsubq_u32 (vreinterpretq_u32_f32 (vabsq_f32 (x)), TinyBound), Thres);
94   float32x4_t xm = x;
95   /* If any lanes are special, mask them with 1 and retain a copy of x to allow
96      special case handler to fix special lanes later. This is only necessary if
97      fenv exceptions are to be triggered correctly.  */
98   if (unlikely (v_any_u32 (cmp)))
99     x = v_zerofy_f32 (x, cmp);
100 #endif
101 
102   /* exp10(x) = 2^n * 10^r = 2^n * (1 + poly (r)),
103      with poly(r) in [1/sqrt(2), sqrt(2)] and
104      x = r + n * log10 (2), with r in [-log10(2)/2, log10(2)/2].  */
105   float32x4_t z = vfmaq_laneq_f32 (d->shift, x, d->log10_2_and_inv, 0);
106   float32x4_t n = vsubq_f32 (z, d->shift);
107   float32x4_t r = vfmsq_laneq_f32 (x, n, d->log10_2_and_inv, 1);
108   r = vfmsq_laneq_f32 (r, n, d->log10_2_and_inv, 2);
109   uint32x4_t e = vshlq_n_u32 (vreinterpretq_u32_f32 (z), 23);
110 
111   float32x4_t scale = vreinterpretq_f32_u32 (vaddq_u32 (e, ExponentBias));
112 
113 #if !WANT_SIMD_EXCEPT
114   uint32x4_t cmp = vcagtq_f32 (n, v_f32 (SpecialBound));
115 #endif
116 
117   float32x4_t r2 = vmulq_f32 (r, r);
118   float32x4_t poly
119       = vfmaq_f32 (vmulq_f32 (r, d->poly[0]),
120 		   v_pairwise_poly_3_f32 (r, r2, d->poly + 1), r2);
121 
122   if (unlikely (v_any_u32 (cmp)))
123 #if WANT_SIMD_EXCEPT
124     return special_case (xm, vfmaq_f32 (scale, poly, scale), cmp);
125 #else
126     return special_case (poly, n, e, cmp, scale, d);
127 #endif
128 
129   return vfmaq_f32 (scale, poly, scale);
130 }
131 
132 PL_SIG (S, F, 1, exp10, -9.9, 9.9)
133 PL_SIG (V, F, 1, exp10, -9.9, 9.9)
134 PL_TEST_ULP (V_NAME_F1 (exp10), 1.86)
135 PL_TEST_EXPECT_FENV (V_NAME_F1 (exp10), WANT_SIMD_EXCEPT)
136 PL_TEST_SYM_INTERVAL (V_NAME_F1 (exp10), 0, SpecialBound, 5000)
137 PL_TEST_SYM_INTERVAL (V_NAME_F1 (exp10), SpecialBound, ScaleBound, 5000)
138 PL_TEST_SYM_INTERVAL (V_NAME_F1 (exp10), ScaleBound, inf, 10000)
139