1 /* 2 * Copyright (c) 2016-2020 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 logarithm 98 * 99 * @param[in] x Input vector value in F32 format. 100 * 101 * @return The calculated logarithm. 102 */ 103 float32x4_t vlogq_f32(float32x4_t x); 104 105 /** Calculate hyperbolic tangent. 106 * 107 * tanh(x) = (e^2x - 1)/(e^2x + 1) 108 * 109 * @note We clamp x to [-5,5] to avoid overflowing issues. 110 * 111 * @param[in] val Input vector value in F32 format. 112 * 113 * @return The calculated Hyperbolic Tangent. 114 */ 115 float32x4_t vtanhq_f32(float32x4_t val); 116 117 /** Calculate n power of a number. 118 * 119 * pow(x,n) = e^(n*log(x)) 120 * 121 * @param[in] val Input vector value in F32 format. 122 * @param[in] n Powers to raise the input to. 123 * 124 * @return The calculated power. 125 */ 126 float32x4_t vpowq_f32(float32x4_t val, float32x4_t n); 127 128 /** Round to the nearest division by a power-of-two using exponent 129 * 130 * @note This function calculates the following expression: (x + 2^n -1 ) / 2^n where n = exponent 131 * 132 * @param[in] x Vector of 4 elements 133 * @param[in] exponent Vector of 4 elements with integer value used to round to nearest division by a power-of-two 134 * 135 * @return the nearest division by a power-of-two using exponent 136 */ 137 int32x4_t rounding_divide_by_pow2(int32x4_t x, int32x4_t exponent); 138 139 /** Round to the nearest division by a power-of-two using exponent 140 * 141 * @note This function calculates the following expression: (x + 2^n -1 ) / 2^n where n = exponent 142 * 143 * @param[in] x Vector of 4 elements 144 * @param[in] exponent Integer value used to round to nearest division by a power-of-two 145 * 146 * @return the nearest division by a power-of-two using exponent 147 */ 148 int32x4_t rounding_divide_by_pow2(int32x4_t x, int exponent); 149 150 /** Round to the nearest division by a power-of-two using exponent 151 * 152 * @note This function calculates the following expression: (x + 2^n -1 ) / 2^n where n = exponent 153 * 154 * @param[in] x Element to divide. 155 * @param[in] exponent Integer value used to round to nearest division by a power-of-two 156 * 157 * @return the nearest division by a power-of-two using exponent 158 */ 159 int32_t rounding_divide_by_pow2(int32_t x, int exponent); 160 161 /** Converts from uint8x16 to float32x4x4_t 162 * 163 * @param[in] in Vector of uint8 to be converted 164 * 165 * @return Converted vector of float 166 */ 167 float32x4x4_t convert_uint8x16_to_float32x4x4(const uint8x16_t &in); 168 169 /** Converts from int8x16 to float32x4x4_t 170 * 171 * @param[in] in Vector of int8 to be converted 172 * 173 * @return Converted vector of float 174 */ 175 float32x4x4_t convert_int8x16_to_float32x4x4(const int8x16_t &in); 176 177 /** Converts to float32x4x4_t from the specified templated 16 elements vectors 178 * 179 * @param[in] in Vector of float to be converted 180 * 181 * @return Converted vector of float 182 */ 183 template <typename T> 184 float32x4x4_t convert_to_float32x4x4(const T &in); 185 186 /** Converts from two float32x4x3_t to just one uint8x8x3_t 187 * 188 * @param[in] in1 First input vector of float to be converted 189 * @param[in] in2 Second input vector of float to be converted 190 * @param[out] out Converted output vector uint8 to store the result 191 */ 192 void convert_float32x4x3_to_uint8x8x3(const float32x4x3_t &in1, const float32x4x3_t &in2, uint8x8x3_t &out); 193 194 /** Converts from two float32x4x4_t to just one uint8x16_t 195 * 196 * @param[in] in Vector of float to be converted 197 * @param[out] out Converted vector of uint8 to store the result 198 */ 199 void convert_float32x4x4_to_uint8x16(const float32x4x4_t &in, uint8x16_t &out); 200 201 /** Converts from float32x4x4_t to just one int8x16_t 202 * 203 * @param[in] in Vector of float to be converted 204 * @param[out] out Converted vector of uint8 to store the result 205 */ 206 void convert_float32x4x4_to_int8x16(const float32x4x4_t &in, int8x16_t &out); 207 208 /** Calculate sine. 209 * 210 * @param[in] val Input vector value in radians, F32 format. 211 * 212 * @return The calculated sine. 213 */ 214 float32x4_t vsinq_f32(float32x4_t val); 215 216 /** Calculate sine. 217 * 218 * @param[in] val Input vector value in radians, F32 format. 219 * 220 * @return The calculated sine. 221 */ 222 float32x2_t vsin_f32(float32x2_t val); 223 224 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC 225 /** Calculate hyperbolic tangent. 226 * 227 * tanh(x) = (e^2x - 1)/(e^2x + 1) 228 * 229 * @note We clamp x to [-5,5] to avoid overflowing issues. 230 * 231 * @param[in] val Input vector value in F16 format. 232 * 233 * @return The calculated Hyperbolic Tangent. 234 */ 235 float16x8_t vtanhq_f16(float16x8_t val); 236 237 /** Calculate round value of a vector to nearest with ties to even. 238 * 239 * @param[in] val Input vector value in F16 format. 240 * 241 * @return The calculated round vector. 242 */ 243 float16x8_t vroundq_rte_f16(float16x8_t val); 244 245 /** Calculate reciprocal. 246 * 247 * @param[in] x Input value. 248 * 249 * @return The calculated reciprocal. 250 */ 251 float16x4_t vinv_f16(float16x4_t x); 252 253 /** Calculate reciprocal. 254 * 255 * @param[in] x Input value. 256 * 257 * @return The calculated reciprocal. 258 */ 259 float16x8_t vinvq_f16(float16x8_t x); 260 261 /** Calculate inverse square root. 262 * 263 * @param[in] x Input value. 264 * 265 * @return The calculated inverse square root. 266 */ 267 float16x4_t vinvsqrt_f16(float16x4_t x); 268 269 /** Calculate inverse square root. 270 * 271 * @param[in] x Input value. 272 * 273 * @return The calculated inverse square root. 274 */ 275 float16x8_t vinvsqrtq_f16(float16x8_t x); 276 277 /** Calculate exponential 278 * 279 * @param[in] x Input vector value in F16 format. 280 * 281 * @return The calculated exponent. 282 */ 283 float16x8_t vexpq_f16(float16x8_t x); 284 285 /** Calculate n power of a number. 286 * 287 * pow(x,n) = e^(n*log(x)) 288 * 289 * @param[in] val Input vector value in F16 format. 290 * @param[in] n Powers to raise the input to. 291 * 292 * @return The calculated power. 293 */ 294 float16x8_t vpowq_f16(float16x8_t val, float16x8_t n); 295 296 /** Calculate sine. 297 * 298 * @param[in] val Input vector value in radians, F16 format. 299 * 300 * @return The calculated sine. 301 */ 302 float16x8_t vsinq_f16(float16x8_t val); 303 304 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ 305 } // namespace arm_compute 306 #include "src/core/NEON/NEMath.inl" 307 #endif /* ARM_COMPUTE_NEMATH_H */ 308