xref: /freebsd/contrib/arm-optimized-routines/pl/math/erfcf_1u7.c (revision 725a9f47324d42037db93c27ceb40d4956872f3e)
1 /*
2  * Single-precision 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 "math_config.h"
9 #include "pl_sig.h"
10 #include "pl_test.h"
11 
12 #define Shift 0x1p17f
13 #define OneThird 0x1.555556p-2f
14 #define TwoThird 0x1.555556p-1f
15 
16 #define TwoOverFifteen 0x1.111112p-3f
17 #define TwoOverFive 0x1.99999ap-2f
18 #define Tenth 0x1.99999ap-4f
19 
20 #define SignMask 0x7fffffff
21 
22 /* Fast erfcf approximation based on series expansion near x rounded to
23    nearest multiple of 1/64.
24    Let d = x - r, and scale = 2 / sqrt(pi) * exp(-r^2). For x near r,
25 
26    erfc(x) ~ erfc(r) - scale * d * poly(r, d), with
27 
28    poly(r, d) = 1 - r d + (2/3 r^2 - 1/3) d^2 - r (1/3 r^2 - 1/2) d^3
29 		+ (2/15 r^4 - 2/5 r^2 + 1/10) d^4
30 
31    Values of erfc(r) and scale are read from lookup tables. Stored values
32    are scaled to avoid hitting the subnormal range.
33 
34    Note that for x < 0, erfc(x) = 2.0 - erfc(-x).
35 
36    Maximum error: 1.63 ULP (~1.0 ULP for x < 0.0).
37    erfcf(0x1.1dbf7ap+3) got 0x1.f51212p-120
38 		       want 0x1.f51216p-120.  */
39 float
40 erfcf (float x)
41 {
42   /* Get top words and sign.  */
43   uint32_t ix = asuint (x);
44   uint32_t ia = ix & SignMask;
45   uint32_t sign = ix & ~SignMask;
46 
47   /* |x| < 0x1.0p-26 => accurate to 0.5 ULP (top12(0x1p-26) = 0x328).  */
48   if (unlikely (ia < 0x32800000))
49     return 1.0f - x; /* Small case.  */
50 
51   /* For |x| < 10.0625, the following approximation holds.  */
52   if (likely (ia < 0x41210000))
53     {
54       /* Lookup erfc(r) and scale(r) in tables, e.g. set erfc(r) to 1 and scale
55 	 to 2/sqrt(pi), when x reduced to r = 0.  */
56       float a = asfloat (ia);
57       float z = a + Shift;
58       uint32_t i = asuint (z) - asuint (Shift);
59       float r = z - Shift;
60 
61       /* These values are scaled by 2^-47.  */
62       float erfcr = __erfcf_data.tab[i].erfc;
63       float scale = __erfcf_data.tab[i].scale;
64 
65       /* erfc(x) ~ erfc(r) - scale * d * poly (r, d).  */
66       float d = a - r;
67       float d2 = d * d;
68       float r2 = r * r;
69       float p1 = -r;
70       float p2 = fmaf (TwoThird, r2, -OneThird);
71       float p3 = -r * fmaf (OneThird, r2, -0.5f);
72       float p4 = fmaf (fmaf (TwoOverFifteen, r2, -TwoOverFive), r2, Tenth);
73       float y = fmaf (p4, d, p3);
74       y = fmaf (y, d, p2);
75       y = fmaf (y, d, p1);
76       y = fmaf (-fmaf (y, d2, d), scale, erfcr);
77       /* Handle sign and scale back in a single fma.  */
78       float off = asfloat (sign >> 1);
79       float fac = asfloat (asuint (0x1p-47f) | sign);
80       y = fmaf (y, fac, off);
81       /* The underflow exception needs to be signaled explicitly when
82 	 result gets into subormnal range.  */
83       if (x >= 0x1.2639cp+3f)
84 	force_eval_float (opt_barrier_float (0x1p-123f) * 0x1p-123f);
85       return y;
86     }
87 
88   /* erfcf(nan)=nan, erfcf(+inf)=0 and erfcf(-inf)=2.  */
89   if (unlikely (ia >= 0x7f800000))
90     return asfloat (sign >> 1) + 1.0f / x; /* Special cases.  */
91 
92   /* Above this threshold erfcf is constant and needs to raise underflow
93      exception for positive x.  */
94   return sign ? 2.0f : __math_uflowf (0);
95 }
96 
97 PL_SIG (S, F, 1, erfc, -4.0, 10.0)
98 PL_TEST_ULP (erfcf, 1.14)
99 PL_TEST_SYM_INTERVAL (erfcf, 0, 0x1p-26, 40000)
100 PL_TEST_INTERVAL (erfcf, 0x1p-26, 10.0625, 40000)
101 PL_TEST_INTERVAL (erfcf, -0x1p-26, -4.0, 40000)
102 PL_TEST_INTERVAL (erfcf, 10.0625, inf, 40000)
103 PL_TEST_INTERVAL (erfcf, -4.0, -inf, 40000)
104