xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/advsimd/logf.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1 /*
2  * Single-precision vector log function.
3  *
4  * Copyright (c) 2019-2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 #include "v_math.h"
8 #include "test_defs.h"
9 #include "test_sig.h"
10 
11 static const struct data
12 {
13   float32x4_t c2, c4, c6, ln2;
14   uint32x4_t off, offset_lower_bound, mantissa_mask;
15   uint16x8_t special_bound;
16   float c1, c3, c5, c0;
17 } data = {
18   /* 3.34 ulp error.  */
19   .c0 = -0x1.3e737cp-3f,
20   .c1 = 0x1.5a9aa2p-3f,
21   .c2 = V4 (-0x1.4f9934p-3f),
22   .c3 = 0x1.961348p-3f,
23   .c4 = V4 (-0x1.00187cp-2f),
24   .c5 = 0x1.555d7cp-2f,
25   .c6 = V4 (-0x1.ffffc8p-2f),
26   .ln2 = V4 (0x1.62e43p-1f),
27   /* Lower bound is the smallest positive normal float 0x00800000. For
28      optimised register use subnormals are detected after offset has been
29      subtracted, so lower bound is 0x0080000 - offset (which wraps around).  */
30   .offset_lower_bound = V4 (0x00800000 - 0x3f2aaaab),
31   .special_bound = V8 (0x7f00), /* top16(asuint32(inf) - 0x00800000).  */
32   .off = V4 (0x3f2aaaab),	/* 0.666667.  */
33   .mantissa_mask = V4 (0x007fffff)
34 };
35 
36 static float32x4_t VPCS_ATTR NOINLINE
special_case(float32x4_t p,uint32x4_t u_off,float32x4_t y,float32x4_t r2,uint16x4_t cmp,const struct data * d)37 special_case (float32x4_t p, uint32x4_t u_off, float32x4_t y, float32x4_t r2,
38 	      uint16x4_t cmp, const struct data *d)
39 {
40   /* Fall back to scalar code.  */
41   return v_call_f32 (logf, vreinterpretq_f32_u32 (vaddq_u32 (u_off, d->off)),
42 		     vfmaq_f32 (p, y, r2), vmovl_u16 (cmp));
43 }
44 
V_NAME_F1(log)45 float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (log) (float32x4_t x)
46 {
47   const struct data *d = ptr_barrier (&data);
48   float32x4_t c1350 = vld1q_f32 (&d->c1);
49 
50   /* To avoid having to mov x out of the way, keep u after offset has been
51      applied, and recover x by adding the offset back in the special-case
52      handler.  */
53   uint32x4_t u_off = vsubq_u32 (vreinterpretq_u32_f32 (x), d->off);
54 
55   /* x = 2^n * (1+r), where 2/3 < 1+r < 4/3.  */
56   float32x4_t n = vcvtq_f32_s32 (
57       vshrq_n_s32 (vreinterpretq_s32_u32 (u_off), 23)); /* signextend.  */
58   uint16x4_t cmp = vcge_u16 (vsubhn_u32 (u_off, d->offset_lower_bound),
59 			     vget_low_u16 (d->special_bound));
60 
61   uint32x4_t u = vaddq_u32 (vandq_u32 (u_off, d->mantissa_mask), d->off);
62   float32x4_t r = vsubq_f32 (vreinterpretq_f32_u32 (u), v_f32 (1.0f));
63 
64   /* y = log(1+r) + n*ln2.  */
65   float32x4_t r2 = vmulq_f32 (r, r);
66   /* n*ln2 + r + r2*(P1 + r*P2 + r2*(P3 + r*P4 + r2*(P5 + r*P6 + r2*P7))).  */
67   float32x4_t p = vfmaq_laneq_f32 (d->c2, r, c1350, 0);
68   float32x4_t q = vfmaq_laneq_f32 (d->c4, r, c1350, 1);
69   float32x4_t y = vfmaq_laneq_f32 (d->c6, r, c1350, 2);
70   p = vfmaq_laneq_f32 (p, r2, c1350, 3);
71 
72   q = vfmaq_f32 (q, p, r2);
73   y = vfmaq_f32 (y, q, r2);
74   p = vfmaq_f32 (r, d->ln2, n);
75 
76   if (unlikely (v_any_u16h (cmp)))
77     return special_case (p, u_off, y, r2, cmp, d);
78   return vfmaq_f32 (p, y, r2);
79 }
80 
81 HALF_WIDTH_ALIAS_F1 (log)
82 
83 TEST_SIG (V, F, 1, log, 0.01, 11.1)
84 TEST_ULP (V_NAME_F1 (log), 2.9)
85 TEST_DISABLE_FENV_IF_NOT (V_NAME_F1 (log), WANT_SIMD_EXCEPT)
86 TEST_INTERVAL (V_NAME_F1 (log), 0, 0xffff0000, 10000)
87 TEST_INTERVAL (V_NAME_F1 (log), 0x1p-4, 0x1p4, 500000)
88 TEST_INTERVAL (V_NAME_F1 (log), 0, inf, 50000)
89