xref: /freebsd/contrib/arm-optimized-routines/pl/math/include/mathlib.h (revision 66fd12cf4896eb08ad8e7a2627537f84ead84dd3)
1 // clang-format off
2 /*
3  * Public API.
4  *
5  * Copyright (c) 2015-2023, Arm Limited.
6  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
7  */
8 
9 #ifndef _MATHLIB_H
10 #define _MATHLIB_H
11 
12 float acoshf (float);
13 float asinhf (float);
14 float atan2f (float, float);
15 float atanf (float);
16 float atanhf (float);
17 float cbrtf (float);
18 float coshf (float);
19 float erfcf (float);
20 float erff (float);
21 float expm1f (float);
22 float log10f (float);
23 float log1pf (float);
24 float sinhf (float);
25 float tanf (float);
26 float tanhf (float);
27 
28 double acosh (double);
29 double asinh (double);
30 double atan (double);
31 double atan2 (double, double);
32 double atanh (double);
33 double cbrt (double);
34 double cosh (double);
35 double erfc (double);
36 double expm1 (double);
37 double log10 (double);
38 double log1p (double);
39 double sinh (double);
40 double tanh (double);
41 
42 float __s_acoshf (float);
43 float __s_asinhf (float);
44 float __s_atanf (float);
45 float __s_atan2f (float, float);
46 float __s_atanhf (float);
47 float __s_cbrtf (float);
48 float __s_coshf (float);
49 float __s_erfcf (float);
50 float __s_erff (float);
51 float __s_expm1f (float);
52 float __s_log10f (float);
53 float __s_log1pf (float);
54 float __s_log2f (float);
55 float __s_sinhf (float);
56 float __s_tanf (float);
57 float __s_tanhf (float);
58 
59 double __s_acosh (double);
60 double __s_asinh (double);
61 double __s_atan (double);
62 double __s_atan2 (double, double);
63 double __s_atanh (double);
64 double __s_cbrt (double);
65 double __s_cosh (double);
66 double __s_erf (double);
67 double __s_erfc (double);
68 double __s_expm1 (double);
69 double __s_log10 (double);
70 double __s_log1p (double);
71 double __s_log2 (double);
72 double __s_sinh (double);
73 double __s_tan (double);
74 double __s_tanh (double);
75 
76 #if __aarch64__
77 #if __GNUC__ >= 5
78 typedef __Float32x4_t __f32x4_t;
79 typedef __Float64x2_t __f64x2_t;
80 #elif __clang_major__*100+__clang_minor__ >= 305
81 typedef __attribute__((__neon_vector_type__(4))) float __f32x4_t;
82 typedef __attribute__((__neon_vector_type__(2))) double __f64x2_t;
83 #else
84 #error Unsupported compiler
85 #endif
86 
87 /* Vector functions following the base PCS.  */
88 __f32x4_t __v_acoshf (__f32x4_t);
89 __f64x2_t __v_acosh (__f64x2_t);
90 __f32x4_t __v_asinhf (__f32x4_t);
91 __f64x2_t __v_asinh (__f64x2_t);
92 __f32x4_t __v_atanf (__f32x4_t);
93 __f64x2_t __v_atan (__f64x2_t);
94 __f32x4_t __v_atan2f (__f32x4_t, __f32x4_t);
95 __f64x2_t __v_atan2 (__f64x2_t, __f64x2_t);
96 __f32x4_t __v_atanhf (__f32x4_t);
97 __f64x2_t __v_atanh (__f64x2_t);
98 __f32x4_t __v_cbrtf (__f32x4_t);
99 __f64x2_t __v_cbrt (__f64x2_t);
100 __f32x4_t __v_coshf (__f32x4_t);
101 __f64x2_t __v_cosh (__f64x2_t);
102 __f32x4_t __v_erff (__f32x4_t);
103 __f64x2_t __v_erf (__f64x2_t);
104 __f32x4_t __v_erfcf (__f32x4_t);
105 __f64x2_t __v_erfc (__f64x2_t);
106 __f32x4_t __v_expm1f (__f32x4_t);
107 __f64x2_t __v_expm1 (__f64x2_t);
108 __f32x4_t __v_log10f (__f32x4_t);
109 __f64x2_t __v_log10 (__f64x2_t);
110 __f32x4_t __v_log1pf (__f32x4_t);
111 __f64x2_t __v_log1p (__f64x2_t);
112 __f32x4_t __v_log2f (__f32x4_t);
113 __f64x2_t __v_log2 (__f64x2_t);
114 __f32x4_t __v_sinhf (__f32x4_t);
115 __f64x2_t __v_sinh (__f64x2_t);
116 __f32x4_t __v_tanf (__f32x4_t);
117 __f64x2_t __v_tan (__f64x2_t);
118 __f32x4_t __v_tanhf (__f32x4_t);
119 __f64x2_t __v_tanh (__f64x2_t);
120 
121 #if __GNUC__ >= 9 || __clang_major__ >= 8
122 #define __vpcs __attribute__((__aarch64_vector_pcs__))
123 
124 /* Vector functions following the vector PCS.  */
125 __vpcs __f32x4_t __vn_acoshf (__f32x4_t);
126 __vpcs __f64x2_t __vn_acosh (__f64x2_t);
127 __vpcs __f32x4_t __vn_asinhf (__f32x4_t);
128 __vpcs __f64x2_t __vn_asinh (__f64x2_t);
129 __vpcs __f32x4_t __vn_atanf (__f32x4_t);
130 __vpcs __f64x2_t __vn_atan (__f64x2_t);
131 __vpcs __f32x4_t __vn_atan2f (__f32x4_t, __f32x4_t);
132 __vpcs __f64x2_t __vn_atan2 (__f64x2_t, __f64x2_t);
133 __vpcs __f32x4_t __vn_atanhf (__f32x4_t);
134 __vpcs __f64x2_t __vn_atanh (__f64x2_t);
135 __vpcs __f32x4_t __vn_cbrtf (__f32x4_t);
136 __vpcs __f64x2_t __vn_cbrt (__f64x2_t);
137 __vpcs __f32x4_t __vn_coshf (__f32x4_t);
138 __vpcs __f64x2_t __vn_cosh (__f64x2_t);
139 __vpcs __f32x4_t __vn_erff (__f32x4_t);
140 __vpcs __f64x2_t __vn_erf (__f64x2_t);
141 __vpcs __f32x4_t __vn_erfcf (__f32x4_t);
142 __vpcs __f64x2_t __vn_erfc (__f64x2_t);
143 __vpcs __f32x4_t __vn_expm1f (__f32x4_t);
144 __vpcs __f64x2_t __vn_expm1 (__f64x2_t);
145 __vpcs __f32x4_t __vn_log10f (__f32x4_t);
146 __vpcs __f64x2_t __vn_log10 (__f64x2_t);
147 __vpcs __f32x4_t __vn_log1pf (__f32x4_t);
148 __vpcs __f64x2_t __vn_log1p (__f64x2_t);
149 __vpcs __f32x4_t __vn_log2f (__f32x4_t);
150 __vpcs __f64x2_t __vn_log2 (__f64x2_t);
151 __vpcs __f32x4_t __vn_sinhf (__f32x4_t);
152 __vpcs __f64x2_t __vn_sinh (__f64x2_t);
153 __vpcs __f32x4_t __vn_tanf (__f32x4_t);
154 __vpcs __f64x2_t __vn_tan (__f64x2_t);
155 __vpcs __f32x4_t __vn_tanhf (__f32x4_t);
156 __vpcs __f64x2_t __vn_tanh (__f64x2_t);
157 
158 /* Vector functions following the vector PCS using ABI names.  */
159 __vpcs __f32x4_t _ZGVnN4v_acoshf (__f32x4_t);
160 __vpcs __f64x2_t _ZGVnN2v_acosh (__f64x2_t);
161 __vpcs __f32x4_t _ZGVnN4v_asinhf (__f32x4_t);
162 __vpcs __f64x2_t _ZGVnN2v_asinh (__f64x2_t);
163 __vpcs __f32x4_t _ZGVnN4v_atanf (__f32x4_t);
164 __vpcs __f64x2_t _ZGVnN2v_atan (__f64x2_t);
165 __vpcs __f32x4_t _ZGVnN4vv_atan2f (__f32x4_t, __f32x4_t);
166 __vpcs __f64x2_t _ZGVnN2vv_atan2 (__f64x2_t, __f64x2_t);
167 __vpcs __f32x4_t _ZGVnN4v_atanhf (__f32x4_t);
168 __vpcs __f64x2_t _ZGVnN2v_atanh (__f64x2_t);
169 __vpcs __f32x4_t _ZGVnN4v_cbrtf (__f32x4_t);
170 __vpcs __f64x2_t _ZGVnN2v_cbrt (__f64x2_t);
171 __vpcs __f32x4_t _ZGVnN4v_coshf (__f32x4_t);
172 __vpcs __f64x2_t _ZGVnN2v_cosh (__f64x2_t);
173 __vpcs __f32x4_t _ZGVnN4v_erff (__f32x4_t);
174 __vpcs __f64x2_t _ZGVnN2v_erf (__f64x2_t);
175 __vpcs __f32x4_t _ZGVnN4v_erfcf (__f32x4_t);
176 __vpcs __f64x2_t _ZGVnN2v_erfc (__f64x2_t);
177 __vpcs __f32x4_t _ZGVnN4v_expm1f (__f32x4_t);
178 __vpcs __f64x2_t _ZGVnN2v_expm1 (__f64x2_t);
179 __vpcs __f32x4_t _ZGVnN4v_log10f (__f32x4_t);
180 __vpcs __f64x2_t _ZGVnN2v_log10 (__f64x2_t);
181 __vpcs __f32x4_t _ZGVnN4v_log1pf (__f32x4_t);
182 __vpcs __f64x2_t _ZGVnN2v_log1p (__f64x2_t);
183 __vpcs __f32x4_t _ZGVnN4v_log2f (__f32x4_t);
184 __vpcs __f64x2_t _ZGVnN2v_log2 (__f64x2_t);
185 __vpcs __f32x4_t _ZGVnN4v_sinhf (__f32x4_t);
186 __vpcs __f64x2_t _ZGVnN2v_sinh (__f64x2_t);
187 __vpcs __f32x4_t _ZGVnN4v_tanf (__f32x4_t);
188 __vpcs __f64x2_t _ZGVnN2v_tan (__f64x2_t);
189 __vpcs __f32x4_t _ZGVnN4v_tanhf (__f32x4_t);
190 __vpcs __f64x2_t _ZGVnN2v_tanh (__f64x2_t);
191 
192 #endif
193 
194 #if WANT_SVE_MATH
195 #include <arm_sve.h>
196 svfloat32_t __sv_atan2f_x (svfloat32_t, svfloat32_t, svbool_t);
197 svfloat32_t __sv_atanf_x (svfloat32_t, svbool_t);
198 svfloat64_t __sv_atan_x (svfloat64_t, svbool_t);
199 svfloat64_t __sv_atan2_x (svfloat64_t, svfloat64_t, svbool_t);
200 svfloat32_t __sv_cosf_x (svfloat32_t, svbool_t);
201 svfloat64_t __sv_cos_x (svfloat64_t, svbool_t);
202 svfloat32_t __sv_erff_x (svfloat32_t, svbool_t);
203 svfloat64_t __sv_erf_x (svfloat64_t, svbool_t);
204 svfloat64_t __sv_erfc_x (svfloat64_t, svbool_t);
205 svfloat32_t __sv_expf_x (svfloat32_t, svbool_t);
206 svfloat32_t __sv_logf_x (svfloat32_t, svbool_t);
207 svfloat64_t __sv_log_x (svfloat64_t, svbool_t);
208 svfloat32_t __sv_log10f_x (svfloat32_t, svbool_t);
209 svfloat64_t __sv_log10_x (svfloat64_t, svbool_t);
210 svfloat32_t __sv_log2f_x (svfloat32_t, svbool_t);
211 svfloat64_t __sv_log2_x (svfloat64_t, svbool_t);
212 svfloat32_t __sv_powif_x (svfloat32_t, svint32_t, svbool_t);
213 svfloat64_t __sv_powi_x (svfloat64_t, svint64_t, svbool_t);
214 svfloat32_t __sv_sinf_x (svfloat32_t, svbool_t);
215 svfloat64_t __sv_sin_x (svfloat64_t, svbool_t);
216 svfloat32_t __sv_tanf_x (svfloat32_t, svbool_t);
217 /* SVE ABI names.  */
218 svfloat32_t _ZGVsMxvv_atan2f (svfloat32_t, svfloat32_t, svbool_t);
219 svfloat32_t _ZGVsMxv_atanf (svfloat32_t, svbool_t);
220 svfloat64_t _ZGVsMxv_atan (svfloat64_t, svbool_t);
221 svfloat64_t _ZGVsMxvv_atan2 (svfloat64_t, svfloat64_t, svbool_t);
222 svfloat32_t _ZGVsMxv_cosf (svfloat32_t, svbool_t);
223 svfloat64_t _ZGVsMxv_cos (svfloat64_t, svbool_t);
224 svfloat32_t _ZGVsMxv_erff (svfloat32_t, svbool_t);
225 svfloat64_t _ZGVsMxv_erf (svfloat64_t, svbool_t);
226 svfloat64_t _ZGVsMxv_erfc (svfloat64_t, svbool_t);
227 svfloat32_t _ZGVsMxv_expf (svfloat32_t, svbool_t);
228 svfloat32_t _ZGVsMxv_logf (svfloat32_t, svbool_t);
229 svfloat64_t _ZGVsMxv_log (svfloat64_t, svbool_t);
230 svfloat32_t _ZGVsMxv_log10f (svfloat32_t, svbool_t);
231 svfloat64_t _ZGVsMxv_log10 (svfloat64_t, svbool_t);
232 svfloat32_t _ZGVsMxv_log2f (svfloat32_t, svbool_t);
233 svfloat64_t _ZGVsMxv_log2 (svfloat64_t, svbool_t);
234 svfloat32_t _ZGVsMxvv_powi(svfloat32_t, svint32_t, svbool_t);
235 svfloat64_t _ZGVsMxvv_powk(svfloat64_t, svint64_t, svbool_t);
236 svfloat32_t _ZGVsMxv_sinf (svfloat32_t, svbool_t);
237 svfloat64_t _ZGVsMxv_sin (svfloat64_t, svbool_t);
238 svfloat32_t _ZGVsMxv_tanf (svfloat32_t, svbool_t);
239 #endif
240 
241 #endif
242 
243 #endif
244 // clang-format on
245