1/*
2 * Double-precision vector log2 function.
3 *
4 * Copyright (c) 2022-2023, Arm Limited.
5 * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6 */
7
8#include "v_math.h"
9#include "pl_sig.h"
10#include "pl_test.h"
11#include "poly_advsimd_f64.h"
12
13#define N (1 << V_LOG2_TABLE_BITS)
14
15static const struct data
16{
17  uint64x2_t min_norm;
18  uint32x4_t special_bound;
19  float64x2_t poly[5];
20  float64x2_t invln2;
21  uint64x2_t sign_exp_mask;
22} data = {
23  /* Each coefficient was generated to approximate log(r) for |r| < 0x1.fp-9
24     and N = 128, then scaled by log2(e) in extended precision and rounded back
25     to double precision.  */
26  .poly = { V2 (-0x1.71547652b83p-1), V2 (0x1.ec709dc340953p-2),
27	    V2 (-0x1.71547651c8f35p-2), V2 (0x1.2777ebe12dda5p-2),
28	    V2 (-0x1.ec738d616fe26p-3) },
29  .invln2 = V2 (0x1.71547652b82fep0),
30  .min_norm = V2 (0x0010000000000000), /* asuint64(0x1p-1022).  */
31  .special_bound = V4 (0x7fe00000),    /* asuint64(inf) - min_norm.  */
32  .sign_exp_mask = V2 (0xfff0000000000000),
33};
34
35#define Off v_u64 (0x3fe6900900000000)
36#define IndexMask (N - 1)
37
38struct entry
39{
40  float64x2_t invc;
41  float64x2_t log2c;
42};
43
44static inline struct entry
45lookup (uint64x2_t i)
46{
47  struct entry e;
48  uint64_t i0 = (i[0] >> (52 - V_LOG2_TABLE_BITS)) & IndexMask;
49  uint64_t i1 = (i[1] >> (52 - V_LOG2_TABLE_BITS)) & IndexMask;
50  float64x2_t e0 = vld1q_f64 (&__v_log2_data.table[i0].invc);
51  float64x2_t e1 = vld1q_f64 (&__v_log2_data.table[i1].invc);
52  e.invc = vuzp1q_f64 (e0, e1);
53  e.log2c = vuzp2q_f64 (e0, e1);
54  return e;
55}
56
57static float64x2_t VPCS_ATTR NOINLINE
58special_case (float64x2_t x, float64x2_t y, float64x2_t w, float64x2_t r2,
59	      uint32x2_t special)
60{
61  return v_call_f64 (log2, x, vfmaq_f64 (w, r2, y), vmovl_u32 (special));
62}
63
64/* Double-precision vector log2 routine. Implements the same algorithm as
65   vector log10, with coefficients and table entries scaled in extended
66   precision. The maximum observed error is 2.58 ULP:
67   _ZGVnN2v_log2(0x1.0b556b093869bp+0) got 0x1.fffb34198d9dap-5
68				      want 0x1.fffb34198d9ddp-5.  */
69float64x2_t VPCS_ATTR V_NAME_D1 (log2) (float64x2_t x)
70{
71  const struct data *d = ptr_barrier (&data);
72  uint64x2_t ix = vreinterpretq_u64_f64 (x);
73  uint32x2_t special = vcge_u32 (vsubhn_u64 (ix, d->min_norm),
74				 vget_low_u32 (d->special_bound));
75
76  /* x = 2^k z; where z is in range [Off,2*Off) and exact.
77     The range is split into N subintervals.
78     The ith subinterval contains z and c is near its center.  */
79  uint64x2_t tmp = vsubq_u64 (ix, Off);
80  int64x2_t k = vshrq_n_s64 (vreinterpretq_s64_u64 (tmp), 52);
81  uint64x2_t iz = vsubq_u64 (ix, vandq_u64 (tmp, d->sign_exp_mask));
82  float64x2_t z = vreinterpretq_f64_u64 (iz);
83
84  struct entry e = lookup (tmp);
85
86  /* log2(x) = log1p(z/c-1)/log(2) + log2(c) + k.  */
87
88  float64x2_t r = vfmaq_f64 (v_f64 (-1.0), z, e.invc);
89  float64x2_t kd = vcvtq_f64_s64 (k);
90  float64x2_t w = vfmaq_f64 (e.log2c, r, d->invln2);
91
92  float64x2_t r2 = vmulq_f64 (r, r);
93  float64x2_t y = v_pw_horner_4_f64 (r, r2, d->poly);
94  w = vaddq_f64 (kd, w);
95
96  if (unlikely (v_any_u32h (special)))
97    return special_case (x, y, w, r2, special);
98  return vfmaq_f64 (w, r2, y);
99}
100
101PL_SIG (V, D, 1, log2, 0.01, 11.1)
102PL_TEST_ULP (V_NAME_D1 (log2), 2.09)
103PL_TEST_EXPECT_FENV_ALWAYS (V_NAME_D1 (log2))
104PL_TEST_INTERVAL (V_NAME_D1 (log2), -0.0, -0x1p126, 100)
105PL_TEST_INTERVAL (V_NAME_D1 (log2), 0x1p-149, 0x1p-126, 4000)
106PL_TEST_INTERVAL (V_NAME_D1 (log2), 0x1p-126, 0x1p-23, 50000)
107PL_TEST_INTERVAL (V_NAME_D1 (log2), 0x1p-23, 1.0, 50000)
108PL_TEST_INTERVAL (V_NAME_D1 (log2), 1.0, 100, 50000)
109PL_TEST_INTERVAL (V_NAME_D1 (log2), 100, inf, 50000)
110