1*412f47f9SXin Li /*
2*412f47f9SXin Li * Double-precision vector log(x) function - inline version
3*412f47f9SXin Li *
4*412f47f9SXin Li * Copyright (c) 2024, 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 "math_config.h"
10*412f47f9SXin Li
11*412f47f9SXin Li #ifndef SV_LOG_INLINE_POLY_ORDER
12*412f47f9SXin Li # error Cannot use inline log helper without specifying poly order (options are 4 or 5)
13*412f47f9SXin Li #endif
14*412f47f9SXin Li
15*412f47f9SXin Li #if SV_LOG_INLINE_POLY_ORDER == 4
16*412f47f9SXin Li # define POLY \
17*412f47f9SXin Li { \
18*412f47f9SXin Li -0x1.ffffffffcbad3p-2, 0x1.555555578ed68p-2, -0x1.0000d3a1e7055p-2, \
19*412f47f9SXin Li 0x1.999392d02a63ep-3 \
20*412f47f9SXin Li }
21*412f47f9SXin Li #elif SV_LOG_INLINE_POLY_ORDER == 5
22*412f47f9SXin Li # define POLY \
23*412f47f9SXin Li { \
24*412f47f9SXin Li -0x1.ffffffffffff7p-2, 0x1.55555555170d4p-2, -0x1.0000000399c27p-2, \
25*412f47f9SXin Li 0x1.999b2e90e94cap-3, -0x1.554e550bd501ep-3 \
26*412f47f9SXin Li }
27*412f47f9SXin Li #else
28*412f47f9SXin Li # error Can only choose order 4 or 5 for log poly
29*412f47f9SXin Li #endif
30*412f47f9SXin Li
31*412f47f9SXin Li struct sv_log_inline_data
32*412f47f9SXin Li {
33*412f47f9SXin Li double poly[SV_LOG_INLINE_POLY_ORDER];
34*412f47f9SXin Li double ln2;
35*412f47f9SXin Li uint64_t off, sign_exp_mask;
36*412f47f9SXin Li };
37*412f47f9SXin Li
38*412f47f9SXin Li #define SV_LOG_CONSTANTS \
39*412f47f9SXin Li { \
40*412f47f9SXin Li .poly = POLY, .ln2 = 0x1.62e42fefa39efp-1, \
41*412f47f9SXin Li .sign_exp_mask = 0xfff0000000000000, .off = 0x3fe6900900000000 \
42*412f47f9SXin Li }
43*412f47f9SXin Li
44*412f47f9SXin Li #define P(i) sv_f64 (d->poly[i])
45*412f47f9SXin Li #define N (1 << V_LOG_TABLE_BITS)
46*412f47f9SXin Li
47*412f47f9SXin Li static inline svfloat64_t
sv_log_inline(svbool_t pg,svfloat64_t x,const struct sv_log_inline_data * d)48*412f47f9SXin Li sv_log_inline (svbool_t pg, svfloat64_t x, const struct sv_log_inline_data *d)
49*412f47f9SXin Li {
50*412f47f9SXin Li svuint64_t ix = svreinterpret_u64 (x);
51*412f47f9SXin Li
52*412f47f9SXin Li /* x = 2^k z; where z is in range [Off,2*Off) and exact.
53*412f47f9SXin Li The range is split into N subintervals.
54*412f47f9SXin Li The ith subinterval contains z and c is near its center. */
55*412f47f9SXin Li svuint64_t tmp = svsub_x (pg, ix, d->off);
56*412f47f9SXin Li /* Calculate table index = (tmp >> (52 - V_LOG_TABLE_BITS)) % N.
57*412f47f9SXin Li The actual value of i is double this due to table layout. */
58*412f47f9SXin Li svuint64_t i
59*412f47f9SXin Li = svand_x (pg, svlsr_x (pg, tmp, (51 - V_LOG_TABLE_BITS)), (N - 1) << 1);
60*412f47f9SXin Li svint64_t k
61*412f47f9SXin Li = svasr_x (pg, svreinterpret_s64 (tmp), 52); /* Arithmetic shift. */
62*412f47f9SXin Li svuint64_t iz = svsub_x (pg, ix, svand_x (pg, tmp, 0xfffULL << 52));
63*412f47f9SXin Li svfloat64_t z = svreinterpret_f64 (iz);
64*412f47f9SXin Li
65*412f47f9SXin Li /* Lookup in 2 global lists (length N). */
66*412f47f9SXin Li svfloat64_t invc = svld1_gather_index (pg, &__v_log_data.table[0].invc, i);
67*412f47f9SXin Li svfloat64_t logc = svld1_gather_index (pg, &__v_log_data.table[0].logc, i);
68*412f47f9SXin Li
69*412f47f9SXin Li /* log(x) = log1p(z/c-1) + log(c) + k*Ln2. */
70*412f47f9SXin Li svfloat64_t r = svmad_x (pg, invc, z, -1);
71*412f47f9SXin Li svfloat64_t kd = svcvt_f64_x (pg, k);
72*412f47f9SXin Li /* hi = r + log(c) + k*Ln2. */
73*412f47f9SXin Li svfloat64_t hi = svmla_x (pg, svadd_x (pg, logc, r), kd, __v_log_data.ln2);
74*412f47f9SXin Li /* y = r2*(A0 + r*A1 + r2*(A2 + r*A3 + r2*A4)) + hi. */
75*412f47f9SXin Li svfloat64_t r2 = svmul_x (pg, r, r);
76*412f47f9SXin Li svfloat64_t y = svmla_x (pg, P (2), r, P (3));
77*412f47f9SXin Li svfloat64_t p = svmla_x (pg, P (0), r, P (1));
78*412f47f9SXin Li #if SV_LOG_INLINE_POLY_ORDER == 5
79*412f47f9SXin Li y = svmla_x (pg, P (4), r2);
80*412f47f9SXin Li #endif
81*412f47f9SXin Li y = svmla_x (pg, p, r2, y);
82*412f47f9SXin Li return svmla_x (pg, hi, r2, y);
83*412f47f9SXin Li }
84