xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/advsimd/erfcf.c (revision dd21556857e8d40f66bf5ad54754d9d52669ebf7)
1 /*
2  * Single-precision vector erfc(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 #include "v_math.h"
9 #include "test_sig.h"
10 #include "test_defs.h"
11 
12 static const struct data
13 {
14   uint32x4_t offset, table_scale;
15   float32x4_t max, shift;
16   float coeffs[4];
17   float32x4_t third, two_over_five, tenth;
18 #if WANT_SIMD_EXCEPT
19   float32x4_t uflow_bound;
20 #endif
21 
22 } data = {
23   /* Set an offset so the range of the index used for lookup is 644, and it can
24      be clamped using a saturated add.  */
25   .offset = V4 (0xb7fffd7b),	       /* 0xffffffff - asuint(shift) - 644.  */
26   .table_scale = V4 (0x28000000 << 1), /* asuint (2^-47) << 1.  */
27   .max = V4 (10.0625f),		       /* 10 + 1/16 = 644/64.  */
28   .shift = V4 (0x1p17f),
29   /* Store 1/3, 2/3 and 2/15 in a single register for use with indexed muls and
30      fmas.  */
31   .coeffs = { 0x1.555556p-2f, 0x1.555556p-1f, 0x1.111112p-3f, 0 },
32   .third = V4 (0x1.555556p-2f),
33   .two_over_five = V4 (-0x1.99999ap-2f),
34   .tenth = V4 (-0x1.99999ap-4f),
35 #if WANT_SIMD_EXCEPT
36   .uflow_bound = V4 (0x1.2639cp+3f),
37 #endif
38 };
39 
40 #define TinyBound 0x41000000 /* 0x1p-62f << 1.  */
41 #define Thres 0xbe000000     /* asuint(infinity) << 1 - TinyBound.  */
42 #define Off 0xfffffd7b	     /* 0xffffffff - 644.  */
43 
44 struct entry
45 {
46   float32x4_t erfc;
47   float32x4_t scale;
48 };
49 
50 static inline struct entry
51 lookup (uint32x4_t i)
52 {
53   struct entry e;
54   float32x2_t t0
55       = vld1_f32 (&__v_erfcf_data.tab[vgetq_lane_u32 (i, 0) - Off].erfc);
56   float32x2_t t1
57       = vld1_f32 (&__v_erfcf_data.tab[vgetq_lane_u32 (i, 1) - Off].erfc);
58   float32x2_t t2
59       = vld1_f32 (&__v_erfcf_data.tab[vgetq_lane_u32 (i, 2) - Off].erfc);
60   float32x2_t t3
61       = vld1_f32 (&__v_erfcf_data.tab[vgetq_lane_u32 (i, 3) - Off].erfc);
62   float32x4_t e1 = vcombine_f32 (t0, t1);
63   float32x4_t e2 = vcombine_f32 (t2, t3);
64   e.erfc = vuzp1q_f32 (e1, e2);
65   e.scale = vuzp2q_f32 (e1, e2);
66   return e;
67 }
68 
69 #if WANT_SIMD_EXCEPT
70 static float32x4_t VPCS_ATTR NOINLINE
71 special_case (float32x4_t x, float32x4_t y, uint32x4_t cmp)
72 {
73   return v_call_f32 (erfcf, x, y, cmp);
74 }
75 #endif
76 
77 /* Optimized single-precision vector erfcf(x).
78    Approximation based on series expansion near x rounded to
79    nearest multiple of 1/64.
80    Let d = x - r, and scale = 2 / sqrt(pi) * exp(-r^2). For x near r,
81 
82    erfc(x) ~ erfc(r) - scale * d * poly(r, d), with
83 
84    poly(r, d) = 1 - r d + (2/3 r^2 - 1/3) d^2 - r (1/3 r^2 - 1/2) d^3
85 		+ (2/15 r^4 - 2/5 r^2 + 1/10) d^4
86 
87    Values of erfc(r) and scale are read from lookup tables. Stored values
88    are scaled to avoid hitting the subnormal range.
89 
90    Note that for x < 0, erfc(x) = 2.0 - erfc(-x).
91    Maximum error: 1.63 ULP (~1.0 ULP for x < 0.0).
92    _ZGVnN4v_erfcf(0x1.1dbf7ap+3) got 0x1.f51212p-120
93 				want 0x1.f51216p-120.  */
94 NOINLINE VPCS_ATTR float32x4_t V_NAME_F1 (erfc) (float32x4_t x)
95 {
96   const struct data *dat = ptr_barrier (&data);
97 
98 #if WANT_SIMD_EXCEPT
99   /* |x| < 2^-62. Avoid fabs by left-shifting by 1.  */
100   uint32x4_t ix = vreinterpretq_u32_f32 (x);
101   uint32x4_t cmp = vcltq_u32 (vaddq_u32 (ix, ix), v_u32 (TinyBound));
102   /* x >= ~9.19 (into subnormal case and uflow case). Comparison is done in
103      integer domain to avoid raising exceptions in presence of nans.  */
104   uint32x4_t uflow = vcgeq_s32 (vreinterpretq_s32_f32 (x),
105 				vreinterpretq_s32_f32 (dat->uflow_bound));
106   cmp = vorrq_u32 (cmp, uflow);
107   float32x4_t xm = x;
108   /* If any lanes are special, mask them with 0 and retain a copy of x to allow
109      special case handler to fix special lanes later. This is only necessary if
110      fenv exceptions are to be triggered correctly.  */
111   if (unlikely (v_any_u32 (cmp)))
112     x = v_zerofy_f32 (x, cmp);
113 #endif
114 
115   float32x4_t a = vabsq_f32 (x);
116   a = vminq_f32 (a, dat->max);
117 
118   /* Lookup erfc(r) and scale(r) in tables, e.g. set erfc(r) to 0 and scale to
119      2/sqrt(pi), when x reduced to r = 0.  */
120   float32x4_t shift = dat->shift;
121   float32x4_t z = vaddq_f32 (a, shift);
122 
123   /* Clamp index to a range of 644. A naive approach would use a subtract and
124      min. Instead we offset the table address and the index, then use a
125      saturating add.  */
126   uint32x4_t i = vqaddq_u32 (vreinterpretq_u32_f32 (z), dat->offset);
127 
128   struct entry e = lookup (i);
129 
130   /* erfc(x) ~ erfc(r) - scale * d * poly(r, d).  */
131   float32x4_t r = vsubq_f32 (z, shift);
132   float32x4_t d = vsubq_f32 (a, r);
133   float32x4_t d2 = vmulq_f32 (d, d);
134   float32x4_t r2 = vmulq_f32 (r, r);
135 
136   float32x4_t p1 = r;
137   float32x4_t coeffs = vld1q_f32 (dat->coeffs);
138   float32x4_t p2 = vfmsq_laneq_f32 (dat->third, r2, coeffs, 1);
139   float32x4_t p3
140       = vmulq_f32 (r, vfmaq_laneq_f32 (v_f32 (-0.5), r2, coeffs, 0));
141   float32x4_t p4 = vfmaq_laneq_f32 (dat->two_over_five, r2, coeffs, 2);
142   p4 = vfmsq_f32 (dat->tenth, r2, p4);
143 
144   float32x4_t y = vfmaq_f32 (p3, d, p4);
145   y = vfmaq_f32 (p2, d, y);
146   y = vfmaq_f32 (p1, d, y);
147   y = vfmsq_f32 (e.erfc, e.scale, vfmsq_f32 (d, d2, y));
148 
149   /* Offset equals 2.0f if sign, else 0.0f.  */
150   uint32x4_t sign = vshrq_n_u32 (vreinterpretq_u32_f32 (x), 31);
151   float32x4_t off = vreinterpretq_f32_u32 (vshlq_n_u32 (sign, 30));
152   /* Copy sign and scale back in a single fma. Since the bit patterns do not
153      overlap, then logical or and addition are equivalent here.  */
154   float32x4_t fac = vreinterpretq_f32_u32 (
155       vsraq_n_u32 (vshlq_n_u32 (sign, 31), dat->table_scale, 1));
156 
157 #if WANT_SIMD_EXCEPT
158   if (unlikely (v_any_u32 (cmp)))
159     return special_case (xm, vfmaq_f32 (off, fac, y), cmp);
160 #endif
161 
162   return vfmaq_f32 (off, fac, y);
163 }
164 
165 HALF_WIDTH_ALIAS_F1 (erfc)
166 
167 TEST_SIG (V, F, 1, erfc, -4.0, 10.0)
168 TEST_DISABLE_FENV_IF_NOT (V_NAME_F1 (erfc), WANT_SIMD_EXCEPT)
169 TEST_ULP (V_NAME_F1 (erfc), 1.14)
170 TEST_SYM_INTERVAL (V_NAME_F1 (erfc), 0, 0x1p-26, 40000)
171 TEST_INTERVAL (V_NAME_F1 (erfc), 0x1p-26, 10.0625, 40000)
172 TEST_INTERVAL (V_NAME_F1 (erfc), -0x1p-26, -4.0, 40000)
173 TEST_INTERVAL (V_NAME_F1 (erfc), 10.0625, inf, 40000)
174 TEST_INTERVAL (V_NAME_F1 (erfc), -4.0, -inf, 40000)
175