xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/sve/exp10f.c (revision dd21556857e8d40f66bf5ad54754d9d52669ebf7)
1 /*
2  * Single-precision SVE 10^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 #define _GNU_SOURCE
9 #include "sv_math.h"
10 #include "test_sig.h"
11 #include "test_defs.h"
12 #include "sv_poly_f32.h"
13 
14 /* For x < -Thres, the result is subnormal and not handled correctly by
15    FEXPA.  */
16 #define Thres 37.9
17 
18 static const struct data
19 {
20   float log2_10_lo, c0, c2, c4;
21   float c1, c3, log10_2;
22   float shift, log2_10_hi, thres;
23 } data = {
24   /* Coefficients generated using Remez algorithm with minimisation of relative
25      error.
26      rel error: 0x1.89dafa3p-24
27      abs error: 0x1.167d55p-23 in [-log10(2)/2, log10(2)/2]
28      maxerr: 0.52 +0.5 ulp.  */
29   .c0 = 0x1.26bb16p+1f,
30   .c1 = 0x1.5350d2p+1f,
31   .c2 = 0x1.04744ap+1f,
32   .c3 = 0x1.2d8176p+0f,
33   .c4 = 0x1.12b41ap-1f,
34   /* 1.5*2^17 + 127, a shift value suitable for FEXPA.  */
35   .shift = 0x1.803f8p17f,
36   .log10_2 = 0x1.a934fp+1,
37   .log2_10_hi = 0x1.344136p-2,
38   .log2_10_lo = -0x1.ec10cp-27,
39   .thres = Thres,
40 };
41 
42 static inline svfloat32_t
43 sv_exp10f_inline (svfloat32_t x, const svbool_t pg, const struct data *d)
44 {
45   /* exp10(x) = 2^(n/N) * 10^r = 2^n * (1 + poly (r)),
46      with poly(r) in [1/sqrt(2), sqrt(2)] and
47      x = r + n * log10(2) / N, with r in [-log10(2)/2N, log10(2)/2N].  */
48 
49   svfloat32_t lane_consts = svld1rq (svptrue_b32 (), &d->log2_10_lo);
50 
51   /* n = round(x/(log10(2)/N)).  */
52   svfloat32_t shift = sv_f32 (d->shift);
53   svfloat32_t z = svmad_x (pg, sv_f32 (d->log10_2), x, shift);
54   svfloat32_t n = svsub_x (svptrue_b32 (), z, shift);
55 
56   /* r = x - n*log10(2)/N.  */
57   svfloat32_t r = svmsb_x (pg, sv_f32 (d->log2_10_hi), n, x);
58   r = svmls_lane (r, n, lane_consts, 0);
59 
60   svfloat32_t scale = svexpa (svreinterpret_u32 (z));
61 
62   /* Polynomial evaluation: poly(r) ~ exp10(r)-1.  */
63   svfloat32_t p12 = svmla_lane (sv_f32 (d->c1), r, lane_consts, 2);
64   svfloat32_t p34 = svmla_lane (sv_f32 (d->c3), r, lane_consts, 3);
65   svfloat32_t r2 = svmul_x (svptrue_b32 (), r, r);
66   svfloat32_t p14 = svmla_x (pg, p12, p34, r2);
67   svfloat32_t p0 = svmul_lane (r, lane_consts, 1);
68   svfloat32_t poly = svmla_x (pg, p0, r2, p14);
69 
70   return svmla_x (pg, scale, scale, poly);
71 }
72 
73 static svfloat32_t NOINLINE
74 special_case (svfloat32_t x, svbool_t special, const struct data *d)
75 {
76   return sv_call_f32 (exp10f, x, sv_exp10f_inline (x, svptrue_b32 (), d),
77 		      special);
78 }
79 
80 /* Single-precision SVE exp10f routine. Implements the same algorithm
81    as AdvSIMD exp10f.
82    Worst case error is 1.02 ULPs.
83    _ZGVsMxv_exp10f(-0x1.040488p-4) got 0x1.ba5f9ep-1
84 				  want 0x1.ba5f9cp-1.  */
85 svfloat32_t SV_NAME_F1 (exp10) (svfloat32_t x, const svbool_t pg)
86 {
87   const struct data *d = ptr_barrier (&data);
88   svbool_t special = svacgt (pg, x, d->thres);
89   if (unlikely (svptest_any (special, special)))
90     return special_case (x, special, d);
91   return sv_exp10f_inline (x, pg, d);
92 }
93 
94 #if WANT_EXP10_TESTS
95 TEST_SIG (SV, F, 1, exp10, -9.9, 9.9)
96 TEST_ULP (SV_NAME_F1 (exp10), 0.52)
97 TEST_DISABLE_FENV (SV_NAME_F1 (exp10))
98 TEST_SYM_INTERVAL (SV_NAME_F1 (exp10), 0, Thres, 50000)
99 TEST_SYM_INTERVAL (SV_NAME_F1 (exp10), Thres, inf, 50000)
100 #endif
101 CLOSE_SVE_ATTR
102