1*412f47f9SXin Li /*
2*412f47f9SXin Li * Double-precision SVE log10(x) function.
3*412f47f9SXin Li *
4*412f47f9SXin Li * Copyright (c) 2022-2023, Arm Limited.
5*412f47f9SXin Li * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6*412f47f9SXin Li */
7*412f47f9SXin Li
8*412f47f9SXin Li #include "sv_math.h"
9*412f47f9SXin Li #include "pl_sig.h"
10*412f47f9SXin Li #include "pl_test.h"
11*412f47f9SXin Li #include "poly_sve_f64.h"
12*412f47f9SXin Li
13*412f47f9SXin Li #define Min 0x0010000000000000
14*412f47f9SXin Li #define Max 0x7ff0000000000000
15*412f47f9SXin Li #define Thres 0x7fe0000000000000 /* Max - Min. */
16*412f47f9SXin Li #define Off 0x3fe6900900000000
17*412f47f9SXin Li #define N (1 << V_LOG10_TABLE_BITS)
18*412f47f9SXin Li
19*412f47f9SXin Li static svfloat64_t NOINLINE
special_case(svfloat64_t x,svfloat64_t y,svbool_t special)20*412f47f9SXin Li special_case (svfloat64_t x, svfloat64_t y, svbool_t special)
21*412f47f9SXin Li {
22*412f47f9SXin Li return sv_call_f64 (log10, x, y, special);
23*412f47f9SXin Li }
24*412f47f9SXin Li
25*412f47f9SXin Li /* SVE log10 algorithm.
26*412f47f9SXin Li Maximum measured error is 2.46 ulps.
27*412f47f9SXin Li SV_NAME_D1 (log10)(0x1.131956cd4b627p+0) got 0x1.fffbdf6eaa669p-6
28*412f47f9SXin Li want 0x1.fffbdf6eaa667p-6. */
SV_NAME_D1(log10)29*412f47f9SXin Li svfloat64_t SV_NAME_D1 (log10) (svfloat64_t x, const svbool_t pg)
30*412f47f9SXin Li {
31*412f47f9SXin Li svuint64_t ix = svreinterpret_u64 (x);
32*412f47f9SXin Li svbool_t special = svcmpge (pg, svsub_x (pg, ix, Min), Thres);
33*412f47f9SXin Li
34*412f47f9SXin Li /* x = 2^k z; where z is in range [Off,2*Off) and exact.
35*412f47f9SXin Li The range is split into N subintervals.
36*412f47f9SXin Li The ith subinterval contains z and c is near its center. */
37*412f47f9SXin Li svuint64_t tmp = svsub_x (pg, ix, Off);
38*412f47f9SXin Li svuint64_t i = svlsr_x (pg, tmp, 51 - V_LOG10_TABLE_BITS);
39*412f47f9SXin Li i = svand_x (pg, i, (N - 1) << 1);
40*412f47f9SXin Li svfloat64_t k = svcvt_f64_x (pg, svasr_x (pg, svreinterpret_s64 (tmp), 52));
41*412f47f9SXin Li svfloat64_t z = svreinterpret_f64 (
42*412f47f9SXin Li svsub_x (pg, ix, svand_x (pg, tmp, 0xfffULL << 52)));
43*412f47f9SXin Li
44*412f47f9SXin Li /* log(x) = k*log(2) + log(c) + log(z/c). */
45*412f47f9SXin Li svfloat64_t invc = svld1_gather_index (pg, &__v_log10_data.table[0].invc, i);
46*412f47f9SXin Li svfloat64_t logc
47*412f47f9SXin Li = svld1_gather_index (pg, &__v_log10_data.table[0].log10c, i);
48*412f47f9SXin Li
49*412f47f9SXin Li /* We approximate log(z/c) with a polynomial P(x) ~= log(x + 1):
50*412f47f9SXin Li r = z/c - 1 (we look up precomputed 1/c)
51*412f47f9SXin Li log(z/c) ~= P(r). */
52*412f47f9SXin Li svfloat64_t r = svmad_x (pg, invc, z, -1.0);
53*412f47f9SXin Li
54*412f47f9SXin Li /* hi = log(c) + k*log(2). */
55*412f47f9SXin Li svfloat64_t w = svmla_x (pg, logc, r, __v_log10_data.invln10);
56*412f47f9SXin Li svfloat64_t hi = svmla_x (pg, w, k, __v_log10_data.log10_2);
57*412f47f9SXin Li
58*412f47f9SXin Li /* y = r2*(A0 + r*A1 + r2*(A2 + r*A3 + r2*A4)) + hi. */
59*412f47f9SXin Li svfloat64_t r2 = svmul_x (pg, r, r);
60*412f47f9SXin Li svfloat64_t y = sv_pw_horner_4_f64_x (pg, r, r2, __v_log10_data.poly);
61*412f47f9SXin Li
62*412f47f9SXin Li if (unlikely (svptest_any (pg, special)))
63*412f47f9SXin Li return special_case (x, svmla_x (svnot_z (pg, special), hi, r2, y),
64*412f47f9SXin Li special);
65*412f47f9SXin Li return svmla_x (pg, hi, r2, y);
66*412f47f9SXin Li }
67*412f47f9SXin Li
68*412f47f9SXin Li PL_SIG (SV, D, 1, log10, 0.01, 11.1)
69*412f47f9SXin Li PL_TEST_ULP (SV_NAME_D1 (log10), 1.97)
70*412f47f9SXin Li PL_TEST_INTERVAL (SV_NAME_D1 (log10), -0.0, -0x1p126, 100)
71*412f47f9SXin Li PL_TEST_INTERVAL (SV_NAME_D1 (log10), 0x1p-149, 0x1p-126, 4000)
72*412f47f9SXin Li PL_TEST_INTERVAL (SV_NAME_D1 (log10), 0x1p-126, 0x1p-23, 50000)
73*412f47f9SXin Li PL_TEST_INTERVAL (SV_NAME_D1 (log10), 0x1p-23, 1.0, 50000)
74*412f47f9SXin Li PL_TEST_INTERVAL (SV_NAME_D1 (log10), 1.0, 100, 50000)
75*412f47f9SXin Li PL_TEST_INTERVAL (SV_NAME_D1 (log10), 100, inf, 50000)
76