xref: /aosp_15_r20/external/arm-optimized-routines/pl/math/v_expm1f_inline.h (revision 412f47f9e737e10ed5cc46ec6a8d7fa2264f8a14)
1 /*
2  * Helper for single-precision routines which calculate exp(x) - 1 and do not
3  * need special-case handling
4  *
5  * Copyright (c) 2022-2024, Arm Limited.
6  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
7  */
8 
9 #ifndef PL_MATH_V_EXPM1F_INLINE_H
10 #define PL_MATH_V_EXPM1F_INLINE_H
11 
12 #include "v_math.h"
13 #include "math_config.h"
14 #include "poly_advsimd_f32.h"
15 
16 struct v_expm1f_data
17 {
18   float32x4_t poly[5];
19   float invln2_and_ln2[4];
20   float32x4_t shift;
21   int32x4_t exponent_bias;
22 };
23 
24 /* Coefficients generated using fpminimax with degree=5 in [-log(2)/2,
25    log(2)/2]. Exponent bias is asuint(1.0f).
26    invln2_and_ln2 Stores constants: invln2, ln2_lo, ln2_hi, 0.  */
27 #define V_EXPM1F_DATA                                                         \
28   {                                                                           \
29     .poly = { V4 (0x1.fffffep-2), V4 (0x1.5554aep-3), V4 (0x1.555736p-5),     \
30 	      V4 (0x1.12287cp-7), V4 (0x1.6b55a2p-10) },                      \
31     .shift = V4 (0x1.8p23f), .exponent_bias = V4 (0x3f800000),                \
32     .invln2_and_ln2 = { 0x1.715476p+0f, 0x1.62e4p-1f, 0x1.7f7d1cp-20f, 0 },   \
33   }
34 
35 static inline float32x4_t
expm1f_inline(float32x4_t x,const struct v_expm1f_data * d)36 expm1f_inline (float32x4_t x, const struct v_expm1f_data *d)
37 {
38   /* Helper routine for calculating exp(x) - 1.
39      Copied from v_expm1f_1u6.c, with all special-case handling removed - the
40      calling routine should handle special values if required.  */
41 
42   /* Reduce argument: f in [-ln2/2, ln2/2], i is exact.  */
43   float32x4_t invln2_and_ln2 = vld1q_f32 (d->invln2_and_ln2);
44   float32x4_t j
45       = vsubq_f32 (vfmaq_laneq_f32 (d->shift, x, invln2_and_ln2, 0), d->shift);
46   int32x4_t i = vcvtq_s32_f32 (j);
47   float32x4_t f = vfmsq_laneq_f32 (x, j, invln2_and_ln2, 1);
48   f = vfmsq_laneq_f32 (f, j, invln2_and_ln2, 2);
49 
50   /* Approximate expm1(f) with polynomial P, expm1(f) ~= f + f^2 * P(f).
51      Uses Estrin scheme, where the main _ZGVnN4v_expm1f routine uses
52      Horner.  */
53   float32x4_t f2 = vmulq_f32 (f, f);
54   float32x4_t f4 = vmulq_f32 (f2, f2);
55   float32x4_t p = v_estrin_4_f32 (f, f2, f4, d->poly);
56   p = vfmaq_f32 (f, f2, p);
57 
58   /* t = 2^i.  */
59   int32x4_t u = vaddq_s32 (vshlq_n_s32 (i, 23), d->exponent_bias);
60   float32x4_t t = vreinterpretq_f32_s32 (u);
61   /* expm1(x) ~= p * t + (t - 1).  */
62   return vfmaq_f32 (vsubq_f32 (t, v_f32 (1.0f)), p, t);
63 }
64 
65 #endif // PL_MATH_V_EXPM1F_INLINE_H
66