xref: /freebsd/contrib/arm-optimized-routines/pl/math/sv_erfc_1u8.c (revision 59c8e88e72633afbc47a4ace0d2170d00d51f7dc)
1 /*
2  * Double-precision vector erfc(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 "sv_math.h"
9 #include "pl_sig.h"
10 #include "pl_test.h"
11 
12 static const struct data
13 {
14   uint64_t off_idx, off_arr;
15   double max, shift;
16   double p20, p40, p41, p42;
17   double p51, p52;
18   double q5, r5;
19   double q6, r6;
20   double q7, r7;
21   double q8, r8;
22   double q9, r9;
23   uint64_t table_scale;
24 } data = {
25   /* Set an offset so the range of the index used for lookup is 3487, and it
26      can be clamped using a saturated add on an offset index.
27      Index offset is 0xffffffffffffffff - asuint64(shift) - 3487.  */
28   .off_idx = 0xbd3ffffffffff260,
29   .off_arr = 0xfffffffffffff260, /* 0xffffffffffffffff - 3487.  */
30   .max = 0x1.b3ep+4,		 /* 3487/128.  */
31   .shift = 0x1p45,
32   .table_scale = 0x37f0000000000000, /* asuint64(0x1p-128).  */
33   .p20 = 0x1.5555555555555p-2,	     /* 1/3, used to compute 2/3 and 1/6.  */
34   .p40 = -0x1.999999999999ap-4,	     /* 1/10.  */
35   .p41 = -0x1.999999999999ap-2,	     /* 2/5.  */
36   .p42 = 0x1.1111111111111p-3,	     /* 2/15.  */
37   .p51 = -0x1.c71c71c71c71cp-3,	     /* 2/9.  */
38   .p52 = 0x1.6c16c16c16c17p-5,	     /* 2/45.  */
39   /* Qi = (i+1) / i, for i = 5, ..., 9.  */
40   .q5 = 0x1.3333333333333p0,
41   .q6 = 0x1.2aaaaaaaaaaabp0,
42   .q7 = 0x1.2492492492492p0,
43   .q8 = 0x1.2p0,
44   .q9 = 0x1.1c71c71c71c72p0,
45   /* Ri = -2 * i / ((i+1)*(i+2)), for i = 5, ..., 9.  */
46   .r5 = -0x1.e79e79e79e79ep-3,
47   .r6 = -0x1.b6db6db6db6dbp-3,
48   .r7 = -0x1.8e38e38e38e39p-3,
49   .r8 = -0x1.6c16c16c16c17p-3,
50   .r9 = -0x1.4f2094f2094f2p-3,
51 };
52 
53 /* Optimized double-precision vector erfc(x).
54    Approximation based on series expansion near x rounded to
55    nearest multiple of 1/128.
56    Let d = x - r, and scale = 2 / sqrt(pi) * exp(-r^2). For x near r,
57 
58    erfc(x) ~ erfc(r) - scale * d * poly(r, d), with
59 
60    poly(r, d) = 1 - r d + (2/3 r^2 - 1/3) d^2 - r (1/3 r^2 - 1/2) d^3
61 		+ (2/15 r^4 - 2/5 r^2 + 1/10) d^4
62 		- r * (2/45 r^4 - 2/9 r^2 + 1/6) d^5
63 		+ p6(r) d^6 + ... + p10(r) d^10
64 
65    Polynomials p6(r) to p10(r) are computed using recurrence relation
66 
67    2(i+1)p_i + 2r(i+2)p_{i+1} + (i+2)(i+3)p_{i+2} = 0,
68    with p0 = 1, and p1(r) = -r.
69 
70    Values of erfc(r) and scale are read from lookup tables. Stored values
71    are scaled to avoid hitting the subnormal range.
72 
73    Note that for x < 0, erfc(x) = 2.0 - erfc(-x).
74 
75    Maximum measured error: 1.71 ULP
76    _ZGVsMxv_erfc(0x1.46cfe976733p+4) got 0x1.e15fcbea3e7afp-608
77 				    want 0x1.e15fcbea3e7adp-608.  */
78 svfloat64_t SV_NAME_D1 (erfc) (svfloat64_t x, const svbool_t pg)
79 {
80   const struct data *dat = ptr_barrier (&data);
81 
82   svfloat64_t a = svabs_x (pg, x);
83 
84   /* Clamp input at |x| <= 3487/128.  */
85   a = svmin_x (pg, a, dat->max);
86 
87   /* Reduce x to the nearest multiple of 1/128.  */
88   svfloat64_t shift = sv_f64 (dat->shift);
89   svfloat64_t z = svadd_x (pg, a, shift);
90 
91   /* Saturate index for the NaN case.  */
92   svuint64_t i = svqadd (svreinterpret_u64 (z), dat->off_idx);
93 
94   /* Lookup erfc(r) and 2/sqrt(pi)*exp(-r^2) in tables.  */
95   i = svadd_x (pg, i, i);
96   const float64_t *p = &__erfc_data.tab[0].erfc - 2 * dat->off_arr;
97   svfloat64_t erfcr = svld1_gather_index (pg, p, i);
98   svfloat64_t scale = svld1_gather_index (pg, p + 1, i);
99 
100   /* erfc(x) ~ erfc(r) - scale * d * poly(r, d).  */
101   svfloat64_t r = svsub_x (pg, z, shift);
102   svfloat64_t d = svsub_x (pg, a, r);
103   svfloat64_t d2 = svmul_x (pg, d, d);
104   svfloat64_t r2 = svmul_x (pg, r, r);
105 
106   /* poly (d, r) = 1 + p1(r) * d + p2(r) * d^2 + ... + p9(r) * d^9.  */
107   svfloat64_t p1 = r;
108   svfloat64_t third = sv_f64 (dat->p20);
109   svfloat64_t twothird = svmul_x (pg, third, 2.0);
110   svfloat64_t sixth = svmul_x (pg, third, 0.5);
111   svfloat64_t p2 = svmls_x (pg, third, r2, twothird);
112   svfloat64_t p3 = svmad_x (pg, r2, third, -0.5);
113   p3 = svmul_x (pg, r, p3);
114   svfloat64_t p4 = svmla_x (pg, sv_f64 (dat->p41), r2, dat->p42);
115   p4 = svmls_x (pg, sv_f64 (dat->p40), r2, p4);
116   svfloat64_t p5 = svmla_x (pg, sv_f64 (dat->p51), r2, dat->p52);
117   p5 = svmla_x (pg, sixth, r2, p5);
118   p5 = svmul_x (pg, r, p5);
119   /* Compute p_i using recurrence relation:
120      p_{i+2} = (p_i + r * Q_{i+1} * p_{i+1}) * R_{i+1}.  */
121   svfloat64_t qr5 = svld1rq (svptrue_b64 (), &dat->q5);
122   svfloat64_t qr6 = svld1rq (svptrue_b64 (), &dat->q6);
123   svfloat64_t qr7 = svld1rq (svptrue_b64 (), &dat->q7);
124   svfloat64_t qr8 = svld1rq (svptrue_b64 (), &dat->q8);
125   svfloat64_t qr9 = svld1rq (svptrue_b64 (), &dat->q9);
126   svfloat64_t p6 = svmla_x (pg, p4, p5, svmul_lane (r, qr5, 0));
127   p6 = svmul_lane (p6, qr5, 1);
128   svfloat64_t p7 = svmla_x (pg, p5, p6, svmul_lane (r, qr6, 0));
129   p7 = svmul_lane (p7, qr6, 1);
130   svfloat64_t p8 = svmla_x (pg, p6, p7, svmul_lane (r, qr7, 0));
131   p8 = svmul_lane (p8, qr7, 1);
132   svfloat64_t p9 = svmla_x (pg, p7, p8, svmul_lane (r, qr8, 0));
133   p9 = svmul_lane (p9, qr8, 1);
134   svfloat64_t p10 = svmla_x (pg, p8, p9, svmul_lane (r, qr9, 0));
135   p10 = svmul_lane (p10, qr9, 1);
136   /* Compute polynomial in d using pairwise Horner scheme.  */
137   svfloat64_t p90 = svmla_x (pg, p9, d, p10);
138   svfloat64_t p78 = svmla_x (pg, p7, d, p8);
139   svfloat64_t p56 = svmla_x (pg, p5, d, p6);
140   svfloat64_t p34 = svmla_x (pg, p3, d, p4);
141   svfloat64_t p12 = svmla_x (pg, p1, d, p2);
142   svfloat64_t y = svmla_x (pg, p78, d2, p90);
143   y = svmla_x (pg, p56, d2, y);
144   y = svmla_x (pg, p34, d2, y);
145   y = svmla_x (pg, p12, d2, y);
146 
147   y = svmls_x (pg, erfcr, scale, svmls_x (pg, d, d2, y));
148 
149   /* Offset equals 2.0 if sign, else 0.0.  */
150   svuint64_t sign = svand_x (pg, svreinterpret_u64 (x), 0x8000000000000000);
151   svfloat64_t off = svreinterpret_f64 (svlsr_x (pg, sign, 1));
152   /* Handle sign and scale back in a single fma.  */
153   svfloat64_t fac = svreinterpret_f64 (svorr_x (pg, sign, dat->table_scale));
154 
155   return svmla_x (pg, off, fac, y);
156 }
157 
158 PL_SIG (SV, D, 1, erfc, -6.0, 28.0)
159 PL_TEST_ULP (SV_NAME_D1 (erfc), 1.21)
160 PL_TEST_SYM_INTERVAL (SV_NAME_D1 (erfc), 0.0, 0x1p-26, 40000)
161 PL_TEST_INTERVAL (SV_NAME_D1 (erfc), 0x1p-26, 28.0, 40000)
162 PL_TEST_INTERVAL (SV_NAME_D1 (erfc), -0x1p-26, -6.0, 40000)
163 PL_TEST_INTERVAL (SV_NAME_D1 (erfc), 28.0, inf, 40000)
164 PL_TEST_INTERVAL (SV_NAME_D1 (erfc), 6.0, -inf, 40000)
165