blob: 0fe1f7d20efd58d44e5acfec2db0a466bd8ce42b [file] [log] [blame]
George Landerda55ef92015-11-19 12:05:06 +00001/*
Szabolcs Nagy1b945972018-05-14 14:46:40 +01002 * Single-precision e^x function.
George Landerda55ef92015-11-19 12:05:06 +00003 *
Szabolcs Nagy11253b02018-11-12 11:10:57 +00004 * Copyright (c) 2017-2018, Arm Limited.
5 * SPDX-License-Identifier: MIT
George Landerda55ef92015-11-19 12:05:06 +00006 */
7
George Landerda55ef92015-11-19 12:05:06 +00008#include <math.h>
Szabolcs Nagy86067a72017-08-11 15:35:12 +01009#include <stdint.h>
10#include "math_config.h"
11
12/*
13EXP2F_TABLE_BITS = 5
14EXP2F_POLY_ORDER = 3
15
16ULP error: 0.502 (nearest rounding.)
17Relative error: 1.69 * 2^-34 in [-ln2/64, ln2/64] (before rounding.)
18Wrong count: 170635 (all nearest rounding wrong results with fma.)
19Non-nearest ULP error: 1 (rounded ULP error)
20*/
21
22#define N (1 << EXP2F_TABLE_BITS)
23#define InvLn2N __exp2f_data.invln2_scaled
24#define T __exp2f_data.tab
25#define C __exp2f_data.poly_scaled
26
27static inline uint32_t
28top12 (float x)
29{
30 return asuint (x) >> 20;
31}
George Landerda55ef92015-11-19 12:05:06 +000032
33float
Szabolcs Nagy0d51c042018-04-25 13:26:22 +010034expf (float x)
George Landerda55ef92015-11-19 12:05:06 +000035{
Szabolcs Nagy86067a72017-08-11 15:35:12 +010036 uint32_t abstop;
37 uint64_t ki, t;
38 /* double_t for better performance on targets with FLT_EVAL_METHOD==2. */
39 double_t kd, xd, z, r, r2, y, s;
George Landerda55ef92015-11-19 12:05:06 +000040
Szabolcs Nagy86067a72017-08-11 15:35:12 +010041 xd = (double_t) x;
42 abstop = top12 (x) & 0x7ff;
Szabolcs Nagyb21378a2018-05-11 13:29:44 +010043 if (unlikely (abstop >= top12 (88.0f)))
Szabolcs Nagy86067a72017-08-11 15:35:12 +010044 {
45 /* |x| >= 88 or x is nan. */
46 if (asuint (x) == asuint (-INFINITY))
47 return 0.0f;
48 if (abstop >= top12 (INFINITY))
49 return x + x;
50 if (x > 0x1.62e42ep6f) /* x > log(0x1p128) ~= 88.72 */
51 return __math_oflowf (0);
52 if (x < -0x1.9fe368p6f) /* x < log(0x1p-150) ~= -103.97 */
53 return __math_uflowf (0);
54#if WANT_ERRNO_UFLOW
55 if (x < -0x1.9d1d9ep6f) /* x < log(0x1p-149) ~= -103.28 */
56 return __math_may_uflowf (0);
George Landerda55ef92015-11-19 12:05:06 +000057#endif
Szabolcs Nagy86067a72017-08-11 15:35:12 +010058 }
George Landerda55ef92015-11-19 12:05:06 +000059
Szabolcs Nagy86067a72017-08-11 15:35:12 +010060 /* x*N/Ln2 = k + r with r in [-1/2, 1/2] and int k. */
61 z = InvLn2N * xd;
George Landerda55ef92015-11-19 12:05:06 +000062
Szabolcs Nagy86067a72017-08-11 15:35:12 +010063 /* Round and convert z to int, the result is in [-150*N, 128*N] and
Szabolcs Nagy39b01912018-05-10 15:35:06 +010064 ideally nearest int is used, otherwise the magnitude of r can be
65 bigger which gives larger approximation error. */
Szabolcs Nagy86067a72017-08-11 15:35:12 +010066#if TOINT_INTRINSICS
67 kd = roundtoint (z);
68 ki = converttoint (z);
Szabolcs Nagy39b01912018-05-10 15:35:06 +010069#else
Szabolcs Nagy86067a72017-08-11 15:35:12 +010070# define SHIFT __exp2f_data.shift
Szabolcs Nagy5eb9d192018-05-11 13:20:04 +010071 kd = eval_as_double (z + SHIFT);
Szabolcs Nagy86067a72017-08-11 15:35:12 +010072 ki = asuint64 (kd);
73 kd -= SHIFT;
74#endif
75 r = z - kd;
George Landerda55ef92015-11-19 12:05:06 +000076
Szabolcs Nagy86067a72017-08-11 15:35:12 +010077 /* exp(x) = 2^(k/N) * 2^(r/N) ~= s * (C0*r^3 + C1*r^2 + C2*r + 1) */
78 t = T[ki % N];
79 t += ki << (52 - EXP2F_TABLE_BITS);
80 s = asdouble (t);
81 z = C[0] * r + C[1];
82 r2 = r * r;
83 y = C[2] * r + 1;
84 y = z * r2 + y;
85 y = y * s;
Szabolcs Nagy04884bd2018-12-07 14:58:51 +000086 return eval_as_float (y);
George Landerda55ef92015-11-19 12:05:06 +000087}
Szabolcs Nagyb7d568d2018-06-06 12:26:56 +010088#if USE_GLIBC_ABI
89strong_alias (expf, __expf_finite)
90hidden_alias (expf, __ieee754_expf)
91#endif