xref: /aosp_15_r20/external/arm-optimized-routines/pl/math/v_log2f_2u5.c (revision 412f47f9e737e10ed5cc46ec6a8d7fa2264f8a14)
1*412f47f9SXin Li /*
2*412f47f9SXin Li  * Single-precision vector log2 function.
3*412f47f9SXin Li  *
4*412f47f9SXin Li  * Copyright (c) 2022-2023, 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 "v_math.h"
9*412f47f9SXin Li #include "poly_advsimd_f32.h"
10*412f47f9SXin Li #include "pl_sig.h"
11*412f47f9SXin Li #include "pl_test.h"
12*412f47f9SXin Li 
13*412f47f9SXin Li static const struct data
14*412f47f9SXin Li {
15*412f47f9SXin Li   uint32x4_t min_norm;
16*412f47f9SXin Li   uint16x8_t special_bound;
17*412f47f9SXin Li   uint32x4_t off, mantissa_mask;
18*412f47f9SXin Li   float32x4_t poly[9];
19*412f47f9SXin Li } data = {
20*412f47f9SXin Li   /* Coefficients generated using Remez algorithm approximate
21*412f47f9SXin Li      log2(1+r)/r for r in [ -1/3, 1/3 ].
22*412f47f9SXin Li      rel error: 0x1.c4c4b0cp-26.  */
23*412f47f9SXin Li   .poly = { V4 (0x1.715476p0f), /* (float)(1 / ln(2)).  */
24*412f47f9SXin Li 	    V4 (-0x1.715458p-1f), V4 (0x1.ec701cp-2f), V4 (-0x1.7171a4p-2f),
25*412f47f9SXin Li 	    V4 (0x1.27a0b8p-2f), V4 (-0x1.e5143ep-3f), V4 (0x1.9d8ecap-3f),
26*412f47f9SXin Li 	    V4 (-0x1.c675bp-3f), V4 (0x1.9e495p-3f) },
27*412f47f9SXin Li   .min_norm = V4 (0x00800000),
28*412f47f9SXin Li   .special_bound = V8 (0x7f00), /* asuint32(inf) - min_norm.  */
29*412f47f9SXin Li   .off = V4 (0x3f2aaaab),	/* 0.666667.  */
30*412f47f9SXin Li   .mantissa_mask = V4 (0x007fffff),
31*412f47f9SXin Li };
32*412f47f9SXin Li 
33*412f47f9SXin Li static float32x4_t VPCS_ATTR NOINLINE
special_case(float32x4_t x,float32x4_t n,float32x4_t p,float32x4_t r,uint16x4_t cmp)34*412f47f9SXin Li special_case (float32x4_t x, float32x4_t n, float32x4_t p, float32x4_t r,
35*412f47f9SXin Li 	      uint16x4_t cmp)
36*412f47f9SXin Li {
37*412f47f9SXin Li   /* Fall back to scalar code.  */
38*412f47f9SXin Li   return v_call_f32 (log2f, x, vfmaq_f32 (n, p, r), vmovl_u16 (cmp));
39*412f47f9SXin Li }
40*412f47f9SXin Li 
41*412f47f9SXin Li /* Fast implementation for single precision AdvSIMD log2,
42*412f47f9SXin Li    relies on same argument reduction as AdvSIMD logf.
43*412f47f9SXin Li    Maximum error: 2.48 ULPs
44*412f47f9SXin Li    _ZGVnN4v_log2f(0x1.558174p+0) got 0x1.a9be84p-2
45*412f47f9SXin Li 				want 0x1.a9be8p-2.  */
V_NAME_F1(log2)46*412f47f9SXin Li float32x4_t VPCS_ATTR V_NAME_F1 (log2) (float32x4_t x)
47*412f47f9SXin Li {
48*412f47f9SXin Li   const struct data *d = ptr_barrier (&data);
49*412f47f9SXin Li   uint32x4_t u = vreinterpretq_u32_f32 (x);
50*412f47f9SXin Li   uint16x4_t special = vcge_u16 (vsubhn_u32 (u, d->min_norm),
51*412f47f9SXin Li 				 vget_low_u16 (d->special_bound));
52*412f47f9SXin Li 
53*412f47f9SXin Li   /* x = 2^n * (1+r), where 2/3 < 1+r < 4/3.  */
54*412f47f9SXin Li   u = vsubq_u32 (u, d->off);
55*412f47f9SXin Li   float32x4_t n = vcvtq_f32_s32 (
56*412f47f9SXin Li       vshrq_n_s32 (vreinterpretq_s32_u32 (u), 23)); /* signextend.  */
57*412f47f9SXin Li   u = vaddq_u32 (vandq_u32 (u, d->mantissa_mask), d->off);
58*412f47f9SXin Li   float32x4_t r = vsubq_f32 (vreinterpretq_f32_u32 (u), v_f32 (1.0f));
59*412f47f9SXin Li 
60*412f47f9SXin Li   /* y = log2(1+r) + n.  */
61*412f47f9SXin Li   float32x4_t r2 = vmulq_f32 (r, r);
62*412f47f9SXin Li   float32x4_t p = v_pw_horner_8_f32 (r, r2, d->poly);
63*412f47f9SXin Li 
64*412f47f9SXin Li   if (unlikely (v_any_u16h (special)))
65*412f47f9SXin Li     return special_case (x, n, p, r, special);
66*412f47f9SXin Li   return vfmaq_f32 (n, p, r);
67*412f47f9SXin Li }
68*412f47f9SXin Li 
69*412f47f9SXin Li PL_SIG (V, F, 1, log2, 0.01, 11.1)
70*412f47f9SXin Li PL_TEST_ULP (V_NAME_F1 (log2), 1.99)
71*412f47f9SXin Li PL_TEST_EXPECT_FENV_ALWAYS (V_NAME_F1 (log2))
72*412f47f9SXin Li PL_TEST_INTERVAL (V_NAME_F1 (log2), -0.0, -0x1p126, 100)
73*412f47f9SXin Li PL_TEST_INTERVAL (V_NAME_F1 (log2), 0x1p-149, 0x1p-126, 4000)
74*412f47f9SXin Li PL_TEST_INTERVAL (V_NAME_F1 (log2), 0x1p-126, 0x1p-23, 50000)
75*412f47f9SXin Li PL_TEST_INTERVAL (V_NAME_F1 (log2), 0x1p-23, 1.0, 50000)
76*412f47f9SXin Li PL_TEST_INTERVAL (V_NAME_F1 (log2), 1.0, 100, 50000)
77*412f47f9SXin Li PL_TEST_INTERVAL (V_NAME_F1 (log2), 100, inf, 50000)
78