1*f3087befSAndrew Turner /*
2*f3087befSAndrew Turner * SVE helper for single-precision routines which calculate exp(x) and do
3*f3087befSAndrew Turner * not need special-case handling
4*f3087befSAndrew Turner *
5*f3087befSAndrew Turner * Copyright (c) 2023-2025, Arm Limited.
6*f3087befSAndrew Turner * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
7*f3087befSAndrew Turner */
8*f3087befSAndrew Turner
9*f3087befSAndrew Turner #ifndef MATH_SV_EXPF_INLINE_H
10*f3087befSAndrew Turner #define MATH_SV_EXPF_INLINE_H
11*f3087befSAndrew Turner
12*f3087befSAndrew Turner #include "sv_math.h"
13*f3087befSAndrew Turner #include "test_sig.h"
14*f3087befSAndrew Turner #include "test_defs.h"
15*f3087befSAndrew Turner
16*f3087befSAndrew Turner struct sv_expf_data
17*f3087befSAndrew Turner {
18*f3087befSAndrew Turner float c1, c3, inv_ln2;
19*f3087befSAndrew Turner float ln2_lo, c0, c2, c4;
20*f3087befSAndrew Turner float ln2_hi, shift;
21*f3087befSAndrew Turner };
22*f3087befSAndrew Turner
23*f3087befSAndrew Turner /* Coefficients copied from the polynomial in AdvSIMD variant, reversed for
24*f3087befSAndrew Turner compatibility with polynomial helpers. Shift is 1.5*2^17 + 127. */
25*f3087befSAndrew Turner #define SV_EXPF_DATA \
26*f3087befSAndrew Turner { \
27*f3087befSAndrew Turner /* Coefficients copied from the polynomial in AdvSIMD variant. */ \
28*f3087befSAndrew Turner .c0 = 0x1.ffffecp-1f, .c1 = 0x1.fffdb6p-2f, .c2 = 0x1.555e66p-3f, \
29*f3087befSAndrew Turner .c3 = 0x1.573e2ep-5f, .c4 = 0x1.0e4020p-7f, .inv_ln2 = 0x1.715476p+0f, \
30*f3087befSAndrew Turner .ln2_hi = 0x1.62e4p-1f, .ln2_lo = 0x1.7f7d1cp-20f, \
31*f3087befSAndrew Turner .shift = 0x1.803f8p17f, \
32*f3087befSAndrew Turner }
33*f3087befSAndrew Turner
34*f3087befSAndrew Turner #define C(i) sv_f32 (d->poly[i])
35*f3087befSAndrew Turner
36*f3087befSAndrew Turner static inline svfloat32_t
expf_inline(svfloat32_t x,const svbool_t pg,const struct sv_expf_data * d)37*f3087befSAndrew Turner expf_inline (svfloat32_t x, const svbool_t pg, const struct sv_expf_data *d)
38*f3087befSAndrew Turner {
39*f3087befSAndrew Turner /* exp(x) = 2^n (1 + poly(r)), with 1 + poly(r) in [1/sqrt(2),sqrt(2)]
40*f3087befSAndrew Turner x = ln2*n + r, with r in [-ln2/2, ln2/2]. */
41*f3087befSAndrew Turner
42*f3087befSAndrew Turner svfloat32_t lane_consts = svld1rq (svptrue_b32 (), &d->ln2_lo);
43*f3087befSAndrew Turner
44*f3087befSAndrew Turner /* n = round(x/(ln2/N)). */
45*f3087befSAndrew Turner svfloat32_t z = svmad_x (pg, sv_f32 (d->inv_ln2), x, d->shift);
46*f3087befSAndrew Turner svfloat32_t n = svsub_x (pg, z, d->shift);
47*f3087befSAndrew Turner
48*f3087befSAndrew Turner /* r = x - n*ln2/N. */
49*f3087befSAndrew Turner svfloat32_t r = svmsb_x (pg, sv_f32 (d->ln2_hi), n, x);
50*f3087befSAndrew Turner r = svmls_lane (r, n, lane_consts, 0);
51*f3087befSAndrew Turner
52*f3087befSAndrew Turner /* scale = 2^(n/N). */
53*f3087befSAndrew Turner svfloat32_t scale = svexpa (svreinterpret_u32 (z));
54*f3087befSAndrew Turner
55*f3087befSAndrew Turner /* poly(r) = exp(r) - 1 ~= C0 r + C1 r^2 + C2 r^3 + C3 r^4 + C4 r^5. */
56*f3087befSAndrew Turner svfloat32_t p12 = svmla_lane (sv_f32 (d->c1), r, lane_consts, 2);
57*f3087befSAndrew Turner svfloat32_t p34 = svmla_lane (sv_f32 (d->c3), r, lane_consts, 3);
58*f3087befSAndrew Turner svfloat32_t r2 = svmul_x (svptrue_b32 (), r, r);
59*f3087befSAndrew Turner svfloat32_t p14 = svmla_x (pg, p12, p34, r2);
60*f3087befSAndrew Turner svfloat32_t p0 = svmul_lane (r, lane_consts, 1);
61*f3087befSAndrew Turner svfloat32_t poly = svmla_x (pg, p0, r2, p14);
62*f3087befSAndrew Turner
63*f3087befSAndrew Turner return svmla_x (pg, scale, scale, poly);
64*f3087befSAndrew Turner }
65*f3087befSAndrew Turner
66*f3087befSAndrew Turner #endif // MATH_SV_EXPF_INLINE_H
67