xref: /freebsd/contrib/arm-optimized-routines/pl/math/sv_exp10f_1u5.c (revision 5a02ffc32e777041dd2dad4e651ed2a0865a0a5d)
1 /*
2  * Single-precision SVE 2^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 "sv_math.h"
9 #include "include/mathlib.h"
10 #include "pl_sig.h"
11 #include "pl_test.h"
12 #include "poly_sve_f32.h"
13 
14 /* For x < -SpecialBound, the result is subnormal and not handled correctly by
15    FEXPA.  */
16 #define SpecialBound 37.9
17 
18 static const struct data
19 {
20   float poly[5];
21   float shift, log10_2, log2_10_hi, log2_10_lo, special_bound;
22 } data = {
23   /* Coefficients generated using Remez algorithm with minimisation of relative
24      error.
25      rel error: 0x1.89dafa3p-24
26      abs error: 0x1.167d55p-23 in [-log10(2)/2, log10(2)/2]
27      maxerr: 0.52 +0.5 ulp.  */
28   .poly = { 0x1.26bb16p+1f, 0x1.5350d2p+1f, 0x1.04744ap+1f, 0x1.2d8176p+0f,
29 	    0x1.12b41ap-1f },
30   /* 1.5*2^17 + 127, a shift value suitable for FEXPA.  */
31   .shift = 0x1.903f8p17f,
32   .log10_2 = 0x1.a934fp+1,
33   .log2_10_hi = 0x1.344136p-2,
34   .log2_10_lo = -0x1.ec10cp-27,
35   .special_bound = SpecialBound,
36 };
37 
38 static svfloat32_t NOINLINE
special_case(svfloat32_t x,svfloat32_t y,svbool_t special)39 special_case (svfloat32_t x, svfloat32_t y, svbool_t special)
40 {
41   return sv_call_f32 (exp10f, x, y, special);
42 }
43 
44 /* Single-precision SVE exp10f routine. Implements the same algorithm
45    as AdvSIMD exp10f.
46    Worst case error is 1.02 ULPs.
47    _ZGVsMxv_exp10f(-0x1.040488p-4) got 0x1.ba5f9ep-1
48 				  want 0x1.ba5f9cp-1.  */
SV_NAME_F1(exp10)49 svfloat32_t SV_NAME_F1 (exp10) (svfloat32_t x, const svbool_t pg)
50 {
51   const struct data *d = ptr_barrier (&data);
52   /* exp10(x) = 2^(n/N) * 10^r = 2^n * (1 + poly (r)),
53      with poly(r) in [1/sqrt(2), sqrt(2)] and
54      x = r + n * log10(2) / N, with r in [-log10(2)/2N, log10(2)/2N].  */
55 
56   /* Load some constants in quad-word chunks to minimise memory access (last
57      lane is wasted).  */
58   svfloat32_t log10_2_and_inv = svld1rq (svptrue_b32 (), &d->log10_2);
59 
60   /* n = round(x/(log10(2)/N)).  */
61   svfloat32_t shift = sv_f32 (d->shift);
62   svfloat32_t z = svmla_lane (shift, x, log10_2_and_inv, 0);
63   svfloat32_t n = svsub_x (pg, z, shift);
64 
65   /* r = x - n*log10(2)/N.  */
66   svfloat32_t r = svmls_lane (x, n, log10_2_and_inv, 1);
67   r = svmls_lane (r, n, log10_2_and_inv, 2);
68 
69   svbool_t special = svacgt (pg, x, d->special_bound);
70   svfloat32_t scale = svexpa (svreinterpret_u32 (z));
71 
72   /* Polynomial evaluation: poly(r) ~ exp10(r)-1.  */
73   svfloat32_t r2 = svmul_x (pg, r, r);
74   svfloat32_t poly
75       = svmla_x (pg, svmul_x (pg, r, d->poly[0]),
76 		 sv_pairwise_poly_3_f32_x (pg, r, r2, d->poly + 1), r2);
77 
78   if (unlikely (svptest_any (pg, special)))
79     return special_case (x, svmla_x (pg, scale, scale, poly), special);
80 
81   return svmla_x (pg, scale, scale, poly);
82 }
83 
84 PL_SIG (SV, F, 1, exp10, -9.9, 9.9)
85 PL_TEST_ULP (SV_NAME_F1 (exp10), 0.52)
86 PL_TEST_SYM_INTERVAL (SV_NAME_F1 (exp10), 0, SpecialBound, 50000)
87 PL_TEST_SYM_INTERVAL (SV_NAME_F1 (exp10), SpecialBound, inf, 50000)
88