2  * Single-precision vector log10 function.
4  * Copyright (c) 2020-2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
21       [-1/3, 1/3] (offset=2/3). Max. relative error: 0x1.068ee468p-25.  */
22   .c0 = V4 (-0x1.bcb79cp-3f),
23   .c1 = 0x1.2879c8p-3f,
24   .c2 = V4 (-0x1.bcd472p-4f),
25   .c3 = 0x1.6408f8p-4f,
26   .c4 = V4 (-0x1.246f8p-4f),
27   .c5 = 0x1.f0e514p-5f,
28   .c6 = V4 (-0x1.0fc92cp-4f),
29   .c7 = 0x1.f5f76ap-5f,
30   .ln2 = V4 (0x1.62e43p-1f),
31   .inv_ln10 = V4 (0x1.bcb7b2p-2f),
33      optimised register use subnormals are detected after offset has been
34      subtracted, so lower bound is 0x0080000 - offset (which wraps around).  */
35   .offset_lower_bound = V4 (0x00800000 - 0x3f2aaaab),
36   .special_bound = V8 (0x7f00), /* top16(asuint32(inf) - 0x00800000).  */
42 special_case (float32x4_t y, uint32x4_t u_off, float32x4_t p, float32x4_t r2,  in special_case()  argument
46   return v_call_f32 (log10f, vreinterpretq_f32_u32 (vaddq_u32 (u_off, d->off)),  in special_case()
47 		     vfmaq_f32 (y, p, r2), vmovl_u16 (cmp));  in special_case()
51    uses a similar approach as AdvSIMD logf with the same offset (i.e., 2/3) and
54    _ZGVnN4v_log10f(0x1.555c16p+0) got 0x1.ffe2fap-4
55 				 want 0x1.ffe2f4p-4.  */
59   float32x4_t c1357 = vld1q_f32 (&d->c1);  in V_NAME_F1()
60   /* To avoid having to mov x out of the way, keep u after offset has been  in V_NAME_F1()
61      applied, and recover x by adding the offset back in the special-case  in V_NAME_F1()
66   u_off = vsubq_u32 (u_off, d->off);  in V_NAME_F1()
70   uint16x4_t special = vcge_u16 (vsubhn_u32 (u_off, d->offset_lower_bound),  in V_NAME_F1()
71 				 vget_low_u16 (d->special_bound));  in V_NAME_F1()
73   uint32x4_t u = vaddq_u32 (vandq_u32 (u_off, d->mantissa_mask), d->off);  in V_NAME_F1()
76   /* y = log10(1+r) + n * log10(2).  */  in V_NAME_F1()
79   float32x4_t c01 = vfmaq_laneq_f32 (d->c0, r, c1357, 0);  in V_NAME_F1()
80   float32x4_t c23 = vfmaq_laneq_f32 (d->c2, r, c1357, 1);  in V_NAME_F1()
81   float32x4_t c45 = vfmaq_laneq_f32 (d->c4, r, c1357, 2);  in V_NAME_F1()
82   float32x4_t c67 = vfmaq_laneq_f32 (d->c6, r, c1357, 3);  in V_NAME_F1()
88   /* y = Log10(2) * n + poly * InvLn(10).  */  in V_NAME_F1()
89   float32x4_t y = vfmaq_f32 (r, d->ln2, n);  in V_NAME_F1()  local
90   y = vmulq_f32 (y, d->inv_ln10);  in V_NAME_F1()
93     return special_case (y, u_off, poly, r2, special, d);  in V_NAME_F1()
94   return vfmaq_f32 (y, poly, r2);  in V_NAME_F1()
101 TEST_INTERVAL (V_NAME_F1 (log10), -0.0, -inf, 100)
102 TEST_INTERVAL (V_NAME_F1 (log10), 0, 0x1p-126, 100)
103 TEST_INTERVAL (V_NAME_F1 (log10), 0x1p-126, 0x1p-23, 50000)
104 TEST_INTERVAL (V_NAME_F1 (log10), 0x1p-23, 1.0, 50000)