1*f3087befSAndrew Turner /*
2*f3087befSAndrew Turner * Single-precision vector exp(x) - 1 function.
3*f3087befSAndrew Turner *
4*f3087befSAndrew Turner * Copyright (c) 2022-2024, Arm Limited.
5*f3087befSAndrew Turner * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6*f3087befSAndrew Turner */
7*f3087befSAndrew Turner
8*f3087befSAndrew Turner #include "v_math.h"
9*f3087befSAndrew Turner #include "test_sig.h"
10*f3087befSAndrew Turner #include "test_defs.h"
11*f3087befSAndrew Turner #include "v_expm1f_inline.h"
12*f3087befSAndrew Turner
13*f3087befSAndrew Turner static const struct data
14*f3087befSAndrew Turner {
15*f3087befSAndrew Turner struct v_expm1f_data d;
16*f3087befSAndrew Turner #if WANT_SIMD_EXCEPT
17*f3087befSAndrew Turner uint32x4_t thresh;
18*f3087befSAndrew Turner #else
19*f3087befSAndrew Turner float32x4_t oflow_bound;
20*f3087befSAndrew Turner #endif
21*f3087befSAndrew Turner } data = {
22*f3087befSAndrew Turner .d = V_EXPM1F_DATA,
23*f3087befSAndrew Turner #if !WANT_SIMD_EXCEPT
24*f3087befSAndrew Turner /* Value above which expm1f(x) should overflow. Absolute value of the
25*f3087befSAndrew Turner underflow bound is greater than this, so it catches both cases - there is
26*f3087befSAndrew Turner a small window where fallbacks are triggered unnecessarily. */
27*f3087befSAndrew Turner .oflow_bound = V4 (0x1.5ebc4p+6),
28*f3087befSAndrew Turner #else
29*f3087befSAndrew Turner /* asuint(oflow_bound) - asuint(0x1p-23), shifted left by 1 for absolute
30*f3087befSAndrew Turner compare. */
31*f3087befSAndrew Turner .thresh = V4 (0x1d5ebc40),
32*f3087befSAndrew Turner #endif
33*f3087befSAndrew Turner };
34*f3087befSAndrew Turner
35*f3087befSAndrew Turner /* asuint(0x1p-23), shifted by 1 for abs compare. */
36*f3087befSAndrew Turner #define TinyBound v_u32 (0x34000000 << 1)
37*f3087befSAndrew Turner
38*f3087befSAndrew Turner static float32x4_t VPCS_ATTR NOINLINE
special_case(float32x4_t x,uint32x4_t special,const struct data * d)39*f3087befSAndrew Turner special_case (float32x4_t x, uint32x4_t special, const struct data *d)
40*f3087befSAndrew Turner {
41*f3087befSAndrew Turner return v_call_f32 (
42*f3087befSAndrew Turner expm1f, x, expm1f_inline (v_zerofy_f32 (x, special), &d->d), special);
43*f3087befSAndrew Turner }
44*f3087befSAndrew Turner
45*f3087befSAndrew Turner /* Single-precision vector exp(x) - 1 function.
46*f3087befSAndrew Turner The maximum error is 1.62 ULP:
47*f3087befSAndrew Turner _ZGVnN4v_expm1f(0x1.85f83p-2) got 0x1.da9f4p-2
48*f3087befSAndrew Turner want 0x1.da9f44p-2. */
V_NAME_F1(expm1)49*f3087befSAndrew Turner float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (expm1) (float32x4_t x)
50*f3087befSAndrew Turner {
51*f3087befSAndrew Turner const struct data *d = ptr_barrier (&data);
52*f3087befSAndrew Turner
53*f3087befSAndrew Turner #if WANT_SIMD_EXCEPT
54*f3087befSAndrew Turner uint32x4_t ix = vreinterpretq_u32_f32 (x);
55*f3087befSAndrew Turner /* If fp exceptions are to be triggered correctly, fall back to scalar for
56*f3087befSAndrew Turner |x| < 2^-23, |x| > oflow_bound, Inf & NaN. Add ix to itself for
57*f3087befSAndrew Turner shift-left by 1, and compare with thresh which was left-shifted offline -
58*f3087befSAndrew Turner this is effectively an absolute compare. */
59*f3087befSAndrew Turner uint32x4_t special
60*f3087befSAndrew Turner = vcgeq_u32 (vsubq_u32 (vaddq_u32 (ix, ix), TinyBound), d->thresh);
61*f3087befSAndrew Turner #else
62*f3087befSAndrew Turner /* Handles very large values (+ve and -ve), +/-NaN, +/-Inf. */
63*f3087befSAndrew Turner uint32x4_t special = vcagtq_f32 (x, d->oflow_bound);
64*f3087befSAndrew Turner #endif
65*f3087befSAndrew Turner
66*f3087befSAndrew Turner if (unlikely (v_any_u32 (special)))
67*f3087befSAndrew Turner return special_case (x, special, d);
68*f3087befSAndrew Turner
69*f3087befSAndrew Turner /* expm1(x) ~= p * t + (t - 1). */
70*f3087befSAndrew Turner return expm1f_inline (x, &d->d);
71*f3087befSAndrew Turner }
72*f3087befSAndrew Turner
73*f3087befSAndrew Turner HALF_WIDTH_ALIAS_F1 (expm1)
74*f3087befSAndrew Turner
75*f3087befSAndrew Turner TEST_SIG (V, F, 1, expm1, -9.9, 9.9)
76*f3087befSAndrew Turner TEST_ULP (V_NAME_F1 (expm1), 1.13)
77*f3087befSAndrew Turner TEST_DISABLE_FENV_IF_NOT (V_NAME_F1 (expm1), WANT_SIMD_EXCEPT)
78*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (expm1), 0, 0x1p-23, 1000)
79*f3087befSAndrew Turner TEST_INTERVAL (V_NAME_F1 (expm1), -0x1p-23, 0x1.5ebc4p+6, 1000000)
80*f3087befSAndrew Turner TEST_INTERVAL (V_NAME_F1 (expm1), -0x1p-23, -0x1.9bbabcp+6, 1000000)
81*f3087befSAndrew Turner TEST_INTERVAL (V_NAME_F1 (expm1), 0x1.5ebc4p+6, inf, 1000)
82*f3087befSAndrew Turner TEST_INTERVAL (V_NAME_F1 (expm1), -0x1.9bbabcp+6, -inf, 1000)
83