xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/advsimd/log2f.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1*f3087befSAndrew Turner /*
2*f3087befSAndrew Turner  * Single-precision vector log2 function.
3*f3087befSAndrew Turner  *
4*f3087befSAndrew Turner  * Copyright (c) 2022-2024, Arm Limited.
5*f3087befSAndrew Turner  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6*f3087befSAndrew Turner  */
7*f3087befSAndrew Turner 
8*f3087befSAndrew Turner #include "v_math.h"
9*f3087befSAndrew Turner #include "test_sig.h"
10*f3087befSAndrew Turner #include "test_defs.h"
11*f3087befSAndrew Turner 
12*f3087befSAndrew Turner static const struct data
13*f3087befSAndrew Turner {
14*f3087befSAndrew Turner   float32x4_t c0, c2, c4, c6, c8;
15*f3087befSAndrew Turner   uint32x4_t off, offset_lower_bound;
16*f3087befSAndrew Turner   uint16x8_t special_bound;
17*f3087befSAndrew Turner   uint32x4_t mantissa_mask;
18*f3087befSAndrew Turner   float c1, c3, c5, c7;
19*f3087befSAndrew Turner } data = {
20*f3087befSAndrew Turner   /* Coefficients generated using Remez algorithm approximate
21*f3087befSAndrew Turner      log2(1+r)/r for r in [ -1/3, 1/3 ].
22*f3087befSAndrew Turner      rel error: 0x1.c4c4b0cp-26.  */
23*f3087befSAndrew Turner   .c0 = V4 (0x1.715476p0f), /* (float)(1 / ln(2)).  */
24*f3087befSAndrew Turner   .c1 = -0x1.715458p-1f,
25*f3087befSAndrew Turner   .c2 = V4 (0x1.ec701cp-2f),
26*f3087befSAndrew Turner   .c3 = -0x1.7171a4p-2f,
27*f3087befSAndrew Turner   .c4 = V4 (0x1.27a0b8p-2f),
28*f3087befSAndrew Turner   .c5 = -0x1.e5143ep-3f,
29*f3087befSAndrew Turner   .c6 = V4 (0x1.9d8ecap-3f),
30*f3087befSAndrew Turner   .c7 = -0x1.c675bp-3f,
31*f3087befSAndrew Turner   .c8 = V4 (0x1.9e495p-3f),
32*f3087befSAndrew Turner   /* Lower bound is the smallest positive normal float 0x00800000. For
33*f3087befSAndrew Turner      optimised register use subnormals are detected after offset has been
34*f3087befSAndrew Turner      subtracted, so lower bound is 0x0080000 - offset (which wraps around).  */
35*f3087befSAndrew Turner   .offset_lower_bound = V4 (0x00800000 - 0x3f2aaaab),
36*f3087befSAndrew Turner   .special_bound = V8 (0x7f00), /* top16(asuint32(inf) - 0x00800000).  */
37*f3087befSAndrew Turner   .off = V4 (0x3f2aaaab),	/* 0.666667.  */
38*f3087befSAndrew Turner   .mantissa_mask = V4 (0x007fffff),
39*f3087befSAndrew Turner };
40*f3087befSAndrew Turner 
41*f3087befSAndrew Turner static float32x4_t VPCS_ATTR NOINLINE
special_case(float32x4_t n,uint32x4_t u_off,float32x4_t p,float32x4_t r,uint16x4_t cmp,const struct data * d)42*f3087befSAndrew Turner special_case (float32x4_t n, uint32x4_t u_off, float32x4_t p, float32x4_t r,
43*f3087befSAndrew Turner 	      uint16x4_t cmp, const struct data *d)
44*f3087befSAndrew Turner {
45*f3087befSAndrew Turner   /* Fall back to scalar code.  */
46*f3087befSAndrew Turner   return v_call_f32 (log2f, vreinterpretq_f32_u32 (vaddq_u32 (u_off, d->off)),
47*f3087befSAndrew Turner 		     vfmaq_f32 (n, p, r), vmovl_u16 (cmp));
48*f3087befSAndrew Turner }
49*f3087befSAndrew Turner 
50*f3087befSAndrew Turner /* Fast implementation for single precision AdvSIMD log2,
51*f3087befSAndrew Turner    relies on same argument reduction as AdvSIMD logf.
52*f3087befSAndrew Turner    Maximum error: 2.48 ULPs
53*f3087befSAndrew Turner    _ZGVnN4v_log2f(0x1.558174p+0) got 0x1.a9be84p-2
54*f3087befSAndrew Turner 				want 0x1.a9be8p-2.  */
V_NAME_F1(log2)55*f3087befSAndrew Turner float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (log2) (float32x4_t x)
56*f3087befSAndrew Turner {
57*f3087befSAndrew Turner   const struct data *d = ptr_barrier (&data);
58*f3087befSAndrew Turner 
59*f3087befSAndrew Turner   /* To avoid having to mov x out of the way, keep u after offset has been
60*f3087befSAndrew Turner      applied, and recover x by adding the offset back in the special-case
61*f3087befSAndrew Turner      handler.  */
62*f3087befSAndrew Turner   uint32x4_t u_off = vreinterpretq_u32_f32 (x);
63*f3087befSAndrew Turner 
64*f3087befSAndrew Turner   /* x = 2^n * (1+r), where 2/3 < 1+r < 4/3.  */
65*f3087befSAndrew Turner   u_off = vsubq_u32 (u_off, d->off);
66*f3087befSAndrew Turner   float32x4_t n = vcvtq_f32_s32 (
67*f3087befSAndrew Turner       vshrq_n_s32 (vreinterpretq_s32_u32 (u_off), 23)); /* signextend.  */
68*f3087befSAndrew Turner 
69*f3087befSAndrew Turner   uint16x4_t special = vcge_u16 (vsubhn_u32 (u_off, d->offset_lower_bound),
70*f3087befSAndrew Turner 				 vget_low_u16 (d->special_bound));
71*f3087befSAndrew Turner 
72*f3087befSAndrew Turner   uint32x4_t u = vaddq_u32 (vandq_u32 (u_off, d->mantissa_mask), d->off);
73*f3087befSAndrew Turner   float32x4_t r = vsubq_f32 (vreinterpretq_f32_u32 (u), v_f32 (1.0f));
74*f3087befSAndrew Turner 
75*f3087befSAndrew Turner   /* y = log2(1+r) + n.  */
76*f3087befSAndrew Turner   float32x4_t r2 = vmulq_f32 (r, r);
77*f3087befSAndrew Turner 
78*f3087befSAndrew Turner   float32x4_t c1357 = vld1q_f32 (&d->c1);
79*f3087befSAndrew Turner   float32x4_t c01 = vfmaq_laneq_f32 (d->c0, r, c1357, 0);
80*f3087befSAndrew Turner   float32x4_t c23 = vfmaq_laneq_f32 (d->c2, r, c1357, 1);
81*f3087befSAndrew Turner   float32x4_t c45 = vfmaq_laneq_f32 (d->c4, r, c1357, 2);
82*f3087befSAndrew Turner   float32x4_t c67 = vfmaq_laneq_f32 (d->c6, r, c1357, 3);
83*f3087befSAndrew Turner   float32x4_t p68 = vfmaq_f32 (c67, r2, d->c8);
84*f3087befSAndrew Turner   float32x4_t p48 = vfmaq_f32 (c45, r2, p68);
85*f3087befSAndrew Turner   float32x4_t p28 = vfmaq_f32 (c23, r2, p48);
86*f3087befSAndrew Turner   float32x4_t p = vfmaq_f32 (c01, r2, p28);
87*f3087befSAndrew Turner 
88*f3087befSAndrew Turner   if (unlikely (v_any_u16h (special)))
89*f3087befSAndrew Turner     return special_case (n, u_off, p, r, special, d);
90*f3087befSAndrew Turner   return vfmaq_f32 (n, p, r);
91*f3087befSAndrew Turner }
92*f3087befSAndrew Turner 
93*f3087befSAndrew Turner HALF_WIDTH_ALIAS_F1 (log2)
94*f3087befSAndrew Turner 
95*f3087befSAndrew Turner TEST_SIG (V, F, 1, log2, 0.01, 11.1)
96*f3087befSAndrew Turner TEST_ULP (V_NAME_F1 (log2), 1.99)
97*f3087befSAndrew Turner TEST_INTERVAL (V_NAME_F1 (log2), -0.0, -0x1p126, 100)
98*f3087befSAndrew Turner TEST_INTERVAL (V_NAME_F1 (log2), 0x1p-149, 0x1p-126, 4000)
99*f3087befSAndrew Turner TEST_INTERVAL (V_NAME_F1 (log2), 0x1p-126, 0x1p-23, 50000)
100*f3087befSAndrew Turner TEST_INTERVAL (V_NAME_F1 (log2), 0x1p-23, 1.0, 50000)
101*f3087befSAndrew Turner TEST_INTERVAL (V_NAME_F1 (log2), 1.0, 100, 50000)
102*f3087befSAndrew Turner TEST_INTERVAL (V_NAME_F1 (log2), 100, inf, 50000)
103