xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/NEMath.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2016-2022 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #ifndef ARM_COMPUTE_NEMATH_H
25 #define ARM_COMPUTE_NEMATH_H
26 
27 #include <arm_neon.h>
28 #include <array>
29 
30 namespace arm_compute
31 {
32 /** Calculate floor of a vector.
33  *
34  * @param[in] val Input vector value in F32 format.
35  *
36  * @return The calculated floor vector.
37  */
38 float32x4_t vfloorq_f32(float32x4_t val);
39 
40 /** Calculate round value of a vector to nearest with ties to even.
41  *
42  * @param[in] val Input vector value in F32 format.
43  *
44  * @return The calculated round vector.
45  */
46 float32x4_t vroundq_rte_f32(float32x4_t val);
47 
48 /** Calculate inverse square root.
49  *
50  * @param[in] x Input value.
51  *
52  * @return The calculated inverse square root.
53  */
54 float32x2_t vinvsqrt_f32(float32x2_t x);
55 
56 /** Calculate inverse square root.
57  *
58  * @param[in] x Input value.
59  *
60  * @return The calculated inverse square root.
61  */
62 float32x4_t vinvsqrtq_f32(float32x4_t x);
63 
64 /** Calculate reciprocal.
65  *
66  * @param[in] x Input value.
67  *
68  * @return The calculated reciprocal.
69  */
70 float32x2_t vinv_f32(float32x2_t x);
71 
72 /** Calculate reciprocal.
73  *
74  * @param[in] x Input value.
75  *
76  * @return The calculated reciprocal.
77  */
78 float32x4_t vinvq_f32(float32x4_t x);
79 
80 /** Perform a 7th degree polynomial approximation using Estrin's method.
81  *
82  * @param[in] x      Input vector value in F32 format.
83  * @param[in] coeffs Polynomial coefficients table.
84  *
85  * @return The calculated approximation.
86  */
87 float32x4_t vtaylor_polyq_f32(float32x4_t x, const std::array<float32x4_t, 8> &coeffs);
88 
89 /** Calculate exponential
90  *
91  * @param[in] x Input vector value in F32 format.
92  *
93  * @return The calculated exponent.
94  */
95 float32x4_t vexpq_f32(float32x4_t x);
96 
97 /** Calculate error function
98  *
99  * @param[in] x Input vector in F32 format.
100  *
101  * @return The calculated erf.
102  */
103 float32x4_t verfq_f32(float32x4_t x);
104 
105 /** Calculate logarithm
106  *
107  * @param[in] x Input vector value in F32 format.
108  *
109  * @return The calculated logarithm.
110  */
111 float32x4_t vlogq_f32(float32x4_t x);
112 
113 /** Calculate hyperbolic tangent.
114  *
115  * tanh(x) = (e^2x - 1)/(e^2x + 1)
116  *
117  * @note We clamp x to [-5,5] to avoid overflowing issues.
118  *
119  * @param[in] val Input vector value in F32 format.
120  *
121  * @return The calculated Hyperbolic Tangent.
122  */
123 float32x4_t vtanhq_f32(float32x4_t val);
124 
125 /** Calculate n power of a number.
126  *
127  * pow(x,n) = e^(n*log(x))
128  *
129  * @param[in] val Input vector value in F32 format.
130  * @param[in] n   Powers to raise the input to.
131  *
132  * @return The calculated power.
133  */
134 float32x4_t vpowq_f32(float32x4_t val, float32x4_t n);
135 
136 /** Round to the nearest division by a power-of-two using exponent
137  *
138  * @note This function calculates the following expression: (x + 2^n -1 ) / 2^n where n = exponent
139  *
140  * @param[in] x        Vector of 4 elements
141  * @param[in] exponent Vector of 4 elements with integer value used to round to nearest division by a power-of-two
142  *
143  * @return the nearest division by a power-of-two using exponent
144  */
145 int32x4_t rounding_divide_by_pow2(int32x4_t x, int32x4_t exponent);
146 
147 /** Round to the nearest division by a power-of-two using exponent
148  *
149  * @note This function calculates the following expression: (x + 2^n -1 ) / 2^n where n = exponent
150  *
151  * @param[in] x        Vector of 4 elements
152  * @param[in] exponent Integer value used to round to nearest division by a power-of-two
153  *
154  * @return the nearest division by a power-of-two using exponent
155  */
156 int32x4_t rounding_divide_by_pow2(int32x4_t x, int exponent);
157 
158 /** Round to the nearest division by a power-of-two using exponent
159  *
160  * @note This function calculates the following expression: (x + 2^n -1 ) / 2^n where n = exponent
161  *
162  * @param[in] x        Element to divide.
163  * @param[in] exponent Integer value used to round to nearest division by a power-of-two
164  *
165  * @return the nearest division by a power-of-two using exponent
166  */
167 int32_t rounding_divide_by_pow2(int32_t x, int exponent);
168 
169 /** Converts from uint8x16 to float32x4x4_t
170  *
171  * @param[in] in Vector of uint8 to be converted
172  *
173  * @return Converted vector of float
174  */
175 float32x4x4_t convert_uint8x16_to_float32x4x4(const uint8x16_t &in);
176 
177 /** Converts from int8x16 to float32x4x4_t
178  *
179  * @param[in] in Vector of int8 to be converted
180  *
181  * @return Converted vector of float
182  */
183 float32x4x4_t convert_int8x16_to_float32x4x4(const int8x16_t &in);
184 
185 /** Converts to float32x4x4_t from the specified templated 16 elements vectors
186  *
187  * @param[in] in Vector of float to be converted
188  *
189  * @return Converted vector of float
190  */
191 template <typename T>
192 float32x4x4_t convert_to_float32x4x4(const T &in);
193 
194 /** Converts from two float32x4x3_t to just one uint8x8x3_t
195  *
196  * @param[in]  in1 First input vector of float to be converted
197  * @param[in]  in2 Second input vector of float to be converted
198  * @param[out] out Converted output vector uint8 to store the result
199  */
200 void convert_float32x4x3_to_uint8x8x3(const float32x4x3_t &in1, const float32x4x3_t &in2, uint8x8x3_t &out);
201 
202 /** Converts from two float32x4x4_t to just one uint8x16_t
203  *
204  * @param[in]  in  Vector of float to be converted
205  * @param[out] out Converted vector of uint8 to store the result
206  */
207 void convert_float32x4x4_to_uint8x16(const float32x4x4_t &in, uint8x16_t &out);
208 
209 /** Converts from float32x4x4_t to just one int8x16_t
210  *
211  * @param[in]  in  Vector of float to be converted
212  * @param[out] out Converted vector of uint8 to store the result
213  */
214 void convert_float32x4x4_to_int8x16(const float32x4x4_t &in, int8x16_t &out);
215 
216 /** Converts from float vector to integer vector
217  *
218  * @param[in] in Float vector to converted
219  *
220  * @return The converted integer vector
221  */
222 template <typename float_vec_type, typename int_vec_type>
223 int_vec_type convert_float_to_int(const float_vec_type &in);
224 
225 /** Converts from integer vector to float vector
226  *
227  * @param[in] in Integer vector to converted
228  *
229  * @return The converted float vector
230  */
231 template <typename float_vec_type, typename int_vec_type>
232 float_vec_type convert_int_to_float(const int_vec_type &in);
233 
234 /** Calculate sine.
235  *
236  * @param[in] val Input vector value in radians, F32 format.
237  *
238  * @return The calculated sine.
239  */
240 float32x4_t vsinq_f32(float32x4_t val);
241 
242 /** Calculate sine.
243  *
244  * @param[in] val Input vector value in radians, F32 format.
245  *
246  * @return The calculated sine.
247  */
248 float32x2_t vsin_f32(float32x2_t val);
249 
250 /** Reduce a vector to be a scalar by accumulating all lanes in the vector
251  *
252  * @param[in] v Vector to be reduced.
253  *
254  * @return the wrapped-around number.
255  */
256 float vreduce(const float32x4_t &v);
257 
258 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
259 /** Calculate hyperbolic tangent.
260  *
261  * tanh(x) = (e^2x - 1)/(e^2x + 1)
262  *
263  * @note We clamp x to [-5,5] to avoid overflowing issues.
264  *
265  * @param[in] val Input vector value in F16 format.
266  *
267  * @return The calculated Hyperbolic Tangent.
268  */
269 float16x8_t vtanhq_f16(float16x8_t val);
270 
271 /** Calculate round value of a vector to nearest with ties to even.
272  *
273  * @param[in] val Input vector value in F16 format.
274  *
275  * @return The calculated round vector.
276  */
277 float16x8_t vroundq_rte_f16(float16x8_t val);
278 
279 /** Calculate reciprocal.
280  *
281  * @param[in] x Input value.
282  *
283  * @return The calculated reciprocal.
284  */
285 float16x4_t vinv_f16(float16x4_t x);
286 
287 /** Calculate reciprocal.
288  *
289  * @param[in] x Input value.
290  *
291  * @return The calculated reciprocal.
292  */
293 float16x8_t vinvq_f16(float16x8_t x);
294 
295 /** Calculate inverse square root.
296  *
297  * @param[in] x Input value.
298  *
299  * @return The calculated inverse square root.
300  */
301 float16x4_t vinvsqrt_f16(float16x4_t x);
302 
303 /** Calculate inverse square root.
304  *
305  * @param[in] x Input value.
306  *
307  * @return The calculated inverse square root.
308  */
309 float16x8_t vinvsqrtq_f16(float16x8_t x);
310 
311 /** Calculate exponential
312  *
313  * @param[in] x Input vector value in F16 format.
314  *
315  * @return The calculated exponent.
316  */
317 float16x8_t vexpq_f16(float16x8_t x);
318 
319 /** Calculate error function
320  *
321  * @param[in] x Input vector in F16 format.
322  *
323  * @return The calculated erf.
324  */
325 float16x8_t verfq_f16(float16x8_t x);
326 
327 /** Calculate n power of a number.
328  *
329  * pow(x,n) = e^(n*log(x))
330  *
331  * @param[in] val Input vector value in F16 format.
332  * @param[in] n   Powers to raise the input to.
333  *
334  * @return The calculated power.
335  */
336 float16x8_t vpowq_f16(float16x8_t val, float16x8_t n);
337 
338 /** Calculate sine.
339  *
340  * @param[in] val Input vector value in radians, F16 format.
341  *
342  * @return The calculated sine.
343  */
344 float16x8_t vsinq_f16(float16x8_t val);
345 
346 /** Reduce a vector to be a scalar by accumulating all lanes in the vector
347  *
348  * @param[in] v Vector to be reduced.
349  *
350  * @return the wrapped-around number.
351  */
352 float16_t vreduce(const float16x8_t &v);
353 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
354 } // namespace arm_compute
355 #include "src/core/NEON/NEMath.inl"
356 #endif /* ARM_COMPUTE_NEMATH_H */
357