1*412f47f9SXin Li /*
2*412f47f9SXin Li * Single-precision vector log function - inline version
3*412f47f9SXin Li *
4*412f47f9SXin Li * Copyright (c) 2019-2023, 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
10*412f47f9SXin Li struct v_logf_data
11*412f47f9SXin Li {
12*412f47f9SXin Li float32x4_t poly[7];
13*412f47f9SXin Li float32x4_t ln2;
14*412f47f9SXin Li uint32x4_t off, mantissa_mask;
15*412f47f9SXin Li };
16*412f47f9SXin Li
17*412f47f9SXin Li #define V_LOGF_CONSTANTS \
18*412f47f9SXin Li { \
19*412f47f9SXin Li .poly \
20*412f47f9SXin Li = { V4 (-0x1.3e737cp-3f), V4 (0x1.5a9aa2p-3f), V4 (-0x1.4f9934p-3f), \
21*412f47f9SXin Li V4 (0x1.961348p-3f), V4 (-0x1.00187cp-2f), V4 (0x1.555d7cp-2f), \
22*412f47f9SXin Li V4 (-0x1.ffffc8p-2f) }, \
23*412f47f9SXin Li .ln2 = V4 (0x1.62e43p-1f), .off = V4 (0x3f2aaaab), \
24*412f47f9SXin Li .mantissa_mask = V4 (0x007fffff) \
25*412f47f9SXin Li }
26*412f47f9SXin Li
27*412f47f9SXin Li #define P(i) d->poly[7 - i]
28*412f47f9SXin Li
29*412f47f9SXin Li static inline float32x4_t
v_logf_inline(float32x4_t x,const struct v_logf_data * d)30*412f47f9SXin Li v_logf_inline (float32x4_t x, const struct v_logf_data *d)
31*412f47f9SXin Li {
32*412f47f9SXin Li float32x4_t n, p, q, r, r2, y;
33*412f47f9SXin Li uint32x4_t u;
34*412f47f9SXin Li
35*412f47f9SXin Li u = vreinterpretq_u32_f32 (x);
36*412f47f9SXin Li
37*412f47f9SXin Li /* x = 2^n * (1+r), where 2/3 < 1+r < 4/3. */
38*412f47f9SXin Li u = vsubq_u32 (u, d->off);
39*412f47f9SXin Li n = vcvtq_f32_s32 (
40*412f47f9SXin Li vshrq_n_s32 (vreinterpretq_s32_u32 (u), 23)); /* signextend. */
41*412f47f9SXin Li u = vandq_u32 (u, d->mantissa_mask);
42*412f47f9SXin Li u = vaddq_u32 (u, d->off);
43*412f47f9SXin Li r = vsubq_f32 (vreinterpretq_f32_u32 (u), v_f32 (1.0f));
44*412f47f9SXin Li
45*412f47f9SXin Li /* y = log(1+r) + n*ln2. */
46*412f47f9SXin Li r2 = vmulq_f32 (r, r);
47*412f47f9SXin Li /* n*ln2 + r + r2*(P1 + r*P2 + r2*(P3 + r*P4 + r2*(P5 + r*P6 + r2*P7))). */
48*412f47f9SXin Li p = vfmaq_f32 (P (5), P (6), r);
49*412f47f9SXin Li q = vfmaq_f32 (P (3), P (4), r);
50*412f47f9SXin Li y = vfmaq_f32 (P (1), P (2), r);
51*412f47f9SXin Li p = vfmaq_f32 (p, P (7), r2);
52*412f47f9SXin Li q = vfmaq_f32 (q, p, r2);
53*412f47f9SXin Li y = vfmaq_f32 (y, q, r2);
54*412f47f9SXin Li p = vfmaq_f32 (r, d->ln2, n);
55*412f47f9SXin Li
56*412f47f9SXin Li return vfmaq_f32 (p, y, r2);
57*412f47f9SXin Li }
58*412f47f9SXin Li
59*412f47f9SXin Li #undef P
60