Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 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 | |
| 25 | #ifndef __ARM_COMPUTE_NEDIRECTCONVOLUTIONDETAIL_H__ |
| 26 | #define __ARM_COMPUTE_NEDIRECTCONVOLUTIONDETAIL_H__ |
| 27 | |
| 28 | #include "arm_compute/core/AccessWindowStatic.h" |
| 29 | #include "arm_compute/core/NEON/NEFixedPoint.h" |
| 30 | |
| 31 | #include <arm_neon.h> |
| 32 | |
| 33 | namespace arm_compute |
| 34 | { |
| 35 | namespace detail |
| 36 | { |
| 37 | /** Loads a 3x3 matrix as a row (float). |
| 38 | * |
| 39 | * @param[in] ptr Pointer to a float 3x3 matrix. |
| 40 | * |
| 41 | * @return The loaded matrix. |
| 42 | */ |
| 43 | inline float32x4x3_t load_matrix_row(const float *ptr) |
| 44 | { |
| 45 | const float32x4x3_t r = |
| 46 | { |
| 47 | { |
| 48 | vld1q_dup_f32(ptr), |
| 49 | vld1q_dup_f32(1 + ptr), |
| 50 | vld1q_dup_f32(2 + ptr) |
| 51 | } |
| 52 | }; |
| 53 | return r; |
| 54 | } |
| 55 | |
| 56 | /** Loads a 3x3 matrix as a row (qint8_t). |
| 57 | * |
| 58 | * @param[in] ptr Pointer to a qint8 3x3 matrix. |
| 59 | * |
| 60 | * @return The loaded matrix. |
| 61 | */ |
| 62 | inline qint8x8x3_t load_matrix_row(const qint8_t *ptr) |
| 63 | { |
| 64 | /* ptr is a pointer to a row in a 3x3 matrix, the function returns 3 vectors holding exactly the same value in all lanes: |
| 65 | r.val[0] contains the first element, r.val[1] the second element and r.val[2] the third element (in all lanes) */ |
| 66 | const qint8x8x3_t r = |
| 67 | { |
| 68 | { |
| 69 | vld1_dup_qs8(ptr), |
| 70 | vld1_dup_qs8(1 + ptr), |
| 71 | vld1_dup_qs8(2 + ptr) |
| 72 | } |
| 73 | }; |
| 74 | return r; |
| 75 | } |
| 76 | |
| 77 | /** Perform a convolve3x3 on float32. |
| 78 | * |
| 79 | * @param[in] in_top Pointer to the first row of the input. |
| 80 | * @param[in] in_mid Pointer to the second row of the input. |
| 81 | * @param[in] in_low Pointer to the third row of the input. |
| 82 | * @param[in] m0 First row of the filter. |
| 83 | * @param[in] m1 Second row of the filter. |
| 84 | * @param[in] m2 Third row of the filter. |
| 85 | * @param[in] fixed_point_position (Optional) Fixed point position. |
| 86 | * |
| 87 | */ |
| 88 | template <unsigned int stridex> |
| 89 | float32x4x2_t convolve_3x3(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position); |
| 90 | |
| 91 | template <> |
| 92 | inline float32x4x2_t convolve_3x3<1>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position) |
| 93 | { |
| 94 | ARM_COMPUTE_UNUSED(fixed_point_position); |
| 95 | |
| 96 | const float32x4x3_t vtop = |
| 97 | { |
| 98 | { |
| 99 | vld1q_f32(in_top), |
| 100 | vld1q_f32(in_top + 4), |
| 101 | vld1q_f32(in_top + 8) |
| 102 | } |
| 103 | }; |
| 104 | const float32x4x3_t vmid = |
| 105 | { |
| 106 | { |
| 107 | vld1q_f32(in_mid), |
| 108 | vld1q_f32(in_mid + 4), |
| 109 | vld1q_f32(in_mid + 8) |
| 110 | } |
| 111 | }; |
| 112 | const float32x4x3_t vlow = |
| 113 | { |
| 114 | { |
| 115 | vld1q_f32(in_low), |
| 116 | vld1q_f32(in_low + 4), |
| 117 | vld1q_f32(in_low + 8) |
| 118 | } |
| 119 | }; |
| 120 | float32x4x2_t out = |
| 121 | { |
| 122 | { |
| 123 | vmulq_f32(vtop.val[0], m0.val[0]), |
| 124 | vmulq_f32(vtop.val[1], m0.val[0]) |
| 125 | } |
| 126 | }; |
| 127 | out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vtop.val[0], vtop.val[1], 1), m0.val[1]); |
| 128 | out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vtop.val[0], vtop.val[1], 2), m0.val[2]); |
| 129 | |
| 130 | out.val[0] = vmlaq_f32(out.val[0], vmid.val[0], m1.val[0]); |
| 131 | out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vmid.val[0], vmid.val[1], 1), m1.val[1]); |
| 132 | out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vmid.val[0], vmid.val[1], 2), m1.val[2]); |
| 133 | |
| 134 | out.val[0] = vmlaq_f32(out.val[0], vlow.val[0], m2.val[0]); |
| 135 | out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vlow.val[0], vlow.val[1], 1), m2.val[1]); |
| 136 | out.val[0] = vmlaq_f32(out.val[0], vextq_f32(vlow.val[0], vlow.val[1], 2), m2.val[2]); |
| 137 | |
| 138 | out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vtop.val[1], vtop.val[2], 1), m0.val[1]); |
| 139 | out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vtop.val[1], vtop.val[2], 2), m0.val[2]); |
| 140 | |
| 141 | out.val[1] = vmlaq_f32(out.val[1], vmid.val[1], m1.val[0]); |
| 142 | out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vmid.val[1], vmid.val[2], 1), m1.val[1]); |
| 143 | out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vmid.val[1], vmid.val[2], 2), m1.val[2]); |
| 144 | |
| 145 | out.val[1] = vmlaq_f32(out.val[1], vlow.val[1], m2.val[0]); |
| 146 | out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vlow.val[1], vlow.val[2], 1), m2.val[1]); |
| 147 | out.val[1] = vmlaq_f32(out.val[1], vextq_f32(vlow.val[1], vlow.val[2], 2), m2.val[2]); |
| 148 | return out; |
| 149 | } |
| 150 | |
| 151 | template <> |
| 152 | inline float32x4x2_t convolve_3x3<2>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position) |
| 153 | { |
| 154 | float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position); |
| 155 | out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 2), out.val[0], 1); |
| 156 | out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[1], 0), out.val[0], 2); |
| 157 | out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[1], 2), out.val[0], 3); |
| 158 | return out; |
| 159 | } |
| 160 | |
| 161 | template <> |
| 162 | inline float32x4x2_t convolve_3x3<3>(const float *in_top, const float *in_mid, const float *in_low, const float32x4x3_t &m0, const float32x4x3_t &m1, const float32x4x3_t &m2, int fixed_point_position) |
| 163 | { |
| 164 | float32x4x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position); |
| 165 | out.val[0] = vsetq_lane_f32(vgetq_lane_f32(out.val[0], 3), out.val[0], 1); |
| 166 | return out; |
| 167 | } |
| 168 | |
| 169 | /** Perform a convolve3x3 on qint16. |
| 170 | * |
| 171 | * @param[in] in_top Pointer to the first row of the input. |
| 172 | * @param[in] in_mid Pointer to the second row of the input. |
| 173 | * @param[in] in_low Pointer to the third row of the input. |
| 174 | * @param[in] m0 First row of the filter. |
| 175 | * @param[in] m1 Second row of the filter. |
| 176 | * @param[in] m2 Third row of the filter. |
| 177 | * @param[in] fixed_point_position (Optional) Fixed point position. |
| 178 | * |
| 179 | */ |
| 180 | template <unsigned int stridex> |
| 181 | qint16x8x2_t convolve_3x3(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position); |
| 182 | |
| 183 | template <> |
| 184 | inline qint16x8x2_t convolve_3x3<1>(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position) |
| 185 | { |
| 186 | ARM_COMPUTE_UNUSED(fixed_point_position); |
| 187 | |
| 188 | const qint8x8x3_t vtop = |
| 189 | { |
| 190 | { |
| 191 | vld1_qs8(in_top), |
| 192 | vld1_qs8(in_top + 8), |
| 193 | vld1_qs8(in_top + 16) |
| 194 | } |
| 195 | }; |
| 196 | const qint8x8x3_t vmid = |
| 197 | { |
| 198 | { |
| 199 | vld1_qs8(in_mid), |
| 200 | vld1_qs8(in_mid + 8), |
| 201 | vld1_qs8(in_mid + 16) |
| 202 | } |
| 203 | }; |
| 204 | const qint8x8x3_t vlow = |
| 205 | { |
| 206 | { |
| 207 | vld1_qs8(in_low), |
| 208 | vld1_qs8(in_low + 8), |
| 209 | vld1_qs8(in_low + 16) |
| 210 | } |
| 211 | }; |
| 212 | qint16x8x2_t out = |
| 213 | { |
| 214 | { |
| 215 | vmull_qs8(vtop.val[0], m0.val[0], fixed_point_position), |
| 216 | vmull_qs8(vtop.val[1], m0.val[0], fixed_point_position) |
| 217 | } |
| 218 | }; |
| 219 | out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vtop.val[0], vtop.val[1], 1), m0.val[1], fixed_point_position); |
| 220 | out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vtop.val[0], vtop.val[1], 2), m0.val[2], fixed_point_position); |
| 221 | out.val[0] = vqmlal_qs8(out.val[0], vmid.val[0], m1.val[0], fixed_point_position); |
| 222 | out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vmid.val[0], vmid.val[1], 1), m1.val[1], fixed_point_position); |
| 223 | out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vmid.val[0], vmid.val[1], 2), m1.val[2], fixed_point_position); |
| 224 | out.val[0] = vqmlal_qs8(out.val[0], vlow.val[0], m2.val[0], fixed_point_position); |
| 225 | out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vlow.val[0], vlow.val[1], 1), m2.val[1], fixed_point_position); |
| 226 | out.val[0] = vqmlal_qs8(out.val[0], vext_s8(vlow.val[0], vlow.val[1], 2), m2.val[2], fixed_point_position); |
| 227 | out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vtop.val[1], vtop.val[2], 1), m0.val[1], fixed_point_position); |
| 228 | out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vtop.val[1], vtop.val[2], 2), m0.val[2], fixed_point_position); |
| 229 | out.val[1] = vqmlal_qs8(out.val[1], vmid.val[1], m1.val[0], fixed_point_position); |
| 230 | out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vmid.val[1], vmid.val[2], 1), m1.val[1], fixed_point_position); |
| 231 | out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vmid.val[1], vmid.val[2], 2), m1.val[2], fixed_point_position); |
| 232 | out.val[1] = vqmlal_qs8(out.val[1], vlow.val[1], m2.val[0], fixed_point_position); |
| 233 | out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vlow.val[1], vlow.val[2], 1), m2.val[1], fixed_point_position); |
| 234 | out.val[1] = vqmlal_qs8(out.val[1], vext_s8(vlow.val[1], vlow.val[2], 2), m2.val[2], fixed_point_position); |
| 235 | return out; |
| 236 | } |
| 237 | |
| 238 | template <> |
| 239 | inline qint16x8x2_t convolve_3x3<2>(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position) |
| 240 | { |
| 241 | qint16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position); |
| 242 | out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 2), out.val[0], 1); |
| 243 | out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 4), out.val[0], 2); |
| 244 | out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 6), out.val[0], 3); |
| 245 | out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 0), out.val[0], 4); |
| 246 | out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 2), out.val[0], 5); |
| 247 | out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 4), out.val[0], 6); |
| 248 | out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 6), out.val[0], 7); |
| 249 | return out; |
| 250 | } |
| 251 | |
| 252 | template <> |
| 253 | inline qint16x8x2_t convolve_3x3<3>(const qint8_t *in_top, const qint8_t *in_mid, const qint8_t *in_low, const qint8x8x3_t &m0, const qint8x8x3_t &m1, const qint8x8x3_t &m2, int fixed_point_position) |
| 254 | { |
| 255 | qint16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position); |
| 256 | out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 3), out.val[0], 1); |
| 257 | out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[0], 6), out.val[0], 2); |
| 258 | out.val[0] = vsetq_lane_s16(vgetq_lane_s16(out.val[1], 1), out.val[0], 3); |
| 259 | return out; |
| 260 | } |
| 261 | |
| 262 | /** Stores a float32x4x2_t array into a memory location. |
| 263 | * |
| 264 | * @param[in] buffer Pointer to the memory location where the values will be stored. |
| 265 | * @param[in] values Values that will be stored. |
| 266 | * |
| 267 | */ |
| 268 | template <unsigned int stridex> |
| 269 | void store_results(float *buffer, const float32x4x2_t &values); |
| 270 | |
| 271 | template <> |
| 272 | inline void store_results<1>(float *buffer, const float32x4x2_t &values) |
| 273 | { |
| 274 | vst1q_f32(buffer, values.val[0]); |
| 275 | vst1q_f32(buffer + 4, values.val[1]); |
| 276 | } |
| 277 | |
| 278 | template <> |
| 279 | inline void store_results<2>(float *buffer, const float32x4x2_t &values) |
| 280 | { |
| 281 | vst1q_f32(buffer, values.val[0]); |
| 282 | } |
| 283 | |
| 284 | template <> |
| 285 | inline void store_results<3>(float *buffer, const float32x4x2_t &values) |
| 286 | { |
| 287 | vst1_f32(buffer, vget_low_f32(values.val[0])); |
| 288 | } |
| 289 | |
| 290 | /** Stores a qint16_t array into a memory location. |
| 291 | * |
| 292 | * @param[in] buffer Pointer to the memory location where the values will be stored. |
| 293 | * @param[in] values Values that will be stored. |
| 294 | * |
| 295 | */ |
| 296 | template <unsigned int stridex> |
| 297 | void store_results(qint16_t *buffer, const qint16x8x2_t &values); |
| 298 | |
| 299 | template <> |
| 300 | inline void store_results<1>(qint16_t *buffer, const qint16x8x2_t &values) |
| 301 | { |
| 302 | vst1q_qs16(buffer, values.val[0]); |
| 303 | vst1q_qs16(buffer + 8, values.val[1]); |
| 304 | } |
| 305 | |
| 306 | template <> |
| 307 | inline void store_results<2>(qint16_t *buffer, const qint16x8x2_t &values) |
| 308 | { |
| 309 | vst1q_qs16(buffer, values.val[0]); |
| 310 | } |
| 311 | |
| 312 | template <> |
| 313 | inline void store_results<3>(qint16_t *buffer, const qint16x8x2_t &values) |
| 314 | { |
| 315 | vst1_qs16(buffer, vget_low_s16(values.val[0])); |
| 316 | } |
| 317 | |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 318 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 319 | /** Loads a 3x3 matrix as a row (float16_t). |
| 320 | * |
| 321 | * @param[in] ptr Pointer to a float 3x3 matrix. |
| 322 | * |
| 323 | * @return The loaded matrix. |
| 324 | */ |
| 325 | inline float16x8x3_t load_matrix_row(const float16_t *ptr) |
| 326 | { |
| 327 | /* ptr is a pointer to a row in a 3x3 matrix, the function returns 3 vectors holding exactly the same value in all lanes: |
| 328 | r.val[0] contains the first element, r.val[1] the second element and r.val[2] the third element (in all lanes) */ |
| 329 | const float16x8x3_t r = |
| 330 | { |
| 331 | { |
| 332 | vld1q_dup_f16(ptr), |
| 333 | vld1q_dup_f16(1 + ptr), |
| 334 | vld1q_dup_f16(2 + ptr) |
| 335 | } |
| 336 | }; |
| 337 | return r; |
| 338 | } |
| 339 | |
| 340 | /** Perform a convolve3x3 on float16. |
| 341 | * |
| 342 | * @param[in] in_top Pointer to the first row of the input. |
| 343 | * @param[in] in_mid Pointer to the second row of the input. |
| 344 | * @param[in] in_low Pointer to the third row of the input. |
| 345 | * @param[in] m0 First row of the filter. |
| 346 | * @param[in] m1 Second row of the filter. |
| 347 | * @param[in] m2 Third row of the filter. |
| 348 | * @param[in] fixed_point_position (Optional) Fixed point position. |
| 349 | * |
| 350 | */ |
| 351 | template <unsigned int stridex> |
| 352 | float16x8x2_t convolve_3x3(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2, |
| 353 | int fixed_point_position); |
| 354 | |
| 355 | template <> |
| 356 | inline float16x8x2_t convolve_3x3<1>(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2, |
| 357 | int fixed_point_position) |
| 358 | { |
| 359 | ARM_COMPUTE_UNUSED(fixed_point_position); |
| 360 | |
| 361 | const float16x8x3_t vtop = |
| 362 | { |
| 363 | { |
| 364 | vld1q_f16(in_top), |
| 365 | vld1q_f16(in_top + 8), |
| 366 | vld1q_f16(in_top + 16) |
| 367 | } |
| 368 | }; |
| 369 | const float16x8x3_t vmid = |
| 370 | { |
| 371 | { |
| 372 | vld1q_f16(in_mid), |
| 373 | vld1q_f16(in_mid + 8), |
| 374 | vld1q_f16(in_mid + 16) |
| 375 | } |
| 376 | }; |
| 377 | const float16x8x3_t vlow = |
| 378 | { |
| 379 | { |
| 380 | vld1q_f16(in_low), |
| 381 | vld1q_f16(in_low + 8), |
| 382 | vld1q_f16(in_low + 16) |
| 383 | } |
| 384 | }; |
| 385 | float16x8x2_t out = |
| 386 | { |
| 387 | { |
| 388 | vmulq_f16(vtop.val[0], m0.val[0]), |
| 389 | vmulq_f16(vtop.val[1], m0.val[0]) |
| 390 | } |
| 391 | }; |
| 392 | out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vtop.val[0], vtop.val[1], 1), m0.val[1])); |
| 393 | out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vtop.val[0], vtop.val[1], 2), m0.val[2])); |
| 394 | out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vmid.val[0], m1.val[0])); |
| 395 | out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vmid.val[0], vmid.val[1], 1), m1.val[1])); |
| 396 | out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vmid.val[0], vmid.val[1], 2), m1.val[2])); |
| 397 | out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vlow.val[0], m2.val[0])); |
| 398 | out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vlow.val[0], vlow.val[1], 1), m2.val[1])); |
| 399 | out.val[0] = vaddq_f16(out.val[0], vmulq_f16(vextq_f16(vlow.val[0], vlow.val[1], 2), m2.val[2])); |
| 400 | out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vtop.val[1], vtop.val[2], 1), m0.val[1])); |
| 401 | out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vtop.val[1], vtop.val[2], 2), m0.val[2])); |
| 402 | out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vmid.val[1], m1.val[0])); |
| 403 | out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vmid.val[1], vmid.val[2], 1), m1.val[1])); |
| 404 | out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vmid.val[1], vmid.val[2], 2), m1.val[2])); |
| 405 | out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vlow.val[1], m2.val[0])); |
| 406 | out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vlow.val[1], vlow.val[2], 1), m2.val[1])); |
| 407 | out.val[1] = vaddq_f16(out.val[1], vmulq_f16(vextq_f16(vlow.val[1], vlow.val[2], 2), m2.val[2])); |
| 408 | return out; |
| 409 | } |
| 410 | |
| 411 | template <> |
| 412 | inline float16x8x2_t convolve_3x3<2>(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2, |
| 413 | int fixed_point_position) |
| 414 | { |
| 415 | float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position); |
| 416 | out.val[0] = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 2), out.val[0], 1); |
| 417 | out.val[0] = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 0), out.val[0], 2); |
| 418 | out.val[0] = vsetq_lane_f16(vgetq_lane_f16(out.val[1], 2), out.val[0], 3); |
| 419 | return out; |
| 420 | } |
| 421 | |
| 422 | template <> |
| 423 | inline float16x8x2_t convolve_3x3<3>(const float16_t *in_top, const float16_t *in_mid, const float16_t *in_low, const float16x8x3_t &m0, const float16x8x3_t &m1, const float16x8x3_t &m2, |
| 424 | int fixed_point_position) |
| 425 | { |
| 426 | float16x8x2_t out = convolve_3x3<1>(in_top, in_mid, in_low, m0, m1, m2, fixed_point_position); |
| 427 | out.val[0] = vsetq_lane_f16(vgetq_lane_f16(out.val[0], 3), out.val[0], 1); |
| 428 | return out; |
| 429 | } |
| 430 | |
| 431 | /** Stores a float16x8x2_t array into a memory location. |
| 432 | * |
| 433 | * @param[in] buffer Pointer to the memory location where the values will be stored. |
| 434 | * @param[in] values Values that will be stored. |
| 435 | * |
| 436 | */ |
| 437 | template <unsigned int stridex> |
| 438 | void store_results(float16_t *buffer, const float16x8x2_t &values); |
| 439 | |
| 440 | template <> |
| 441 | inline void store_results<1>(float16_t *buffer, const float16x8x2_t &values) |
| 442 | { |
| 443 | vst1q_f16(buffer, values.val[0]); |
| 444 | vst1q_f16(buffer + 8, values.val[1]); |
| 445 | } |
| 446 | |
| 447 | template <> |
| 448 | inline void store_results<2>(float16_t *buffer, const float16x8x2_t &values) |
| 449 | { |
| 450 | vst1q_f16(buffer, values.val[0]); |
| 451 | } |
| 452 | |
| 453 | template <> |
| 454 | inline void store_results<3>(float16_t *buffer, const float16x8x2_t &values) |
| 455 | { |
| 456 | vst1_f16(buffer, vget_low_f16(values.val[0])); |
| 457 | } |
Ioan-Cristian Szabo | 5edbd1c | 2017-11-13 13:34:08 +0000 | [diff] [blame] | 458 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 459 | |
| 460 | /** Get the number of elements processed on 3x3 convolution. |
| 461 | * |
| 462 | * @param[in] num_elems_written_per_iteration Number of elements written per iteration on 3x3 convolution. |
| 463 | * |
| 464 | * @return The number of elements processed. |
| 465 | */ |
| 466 | template <unsigned int stridex> |
| 467 | int get_input_num_elems_processed(unsigned int num_elems_written_per_iteration); |
| 468 | |
| 469 | template <> |
| 470 | inline int get_input_num_elems_processed<1>(unsigned int num_elems_written_per_iteration) |
| 471 | { |
| 472 | return num_elems_written_per_iteration; |
| 473 | } |
| 474 | |
| 475 | template <> |
| 476 | inline int get_input_num_elems_processed<2>(unsigned int num_elems_written_per_iteration) |
| 477 | { |
| 478 | return num_elems_written_per_iteration << 1; |
| 479 | } |
| 480 | |
| 481 | template <> |
| 482 | inline int get_input_num_elems_processed<3>(unsigned int num_elems_written_per_iteration) |
| 483 | { |
| 484 | return num_elems_written_per_iteration * 3; |
| 485 | } |
Anthony Barbier | 1568621 | 2017-12-12 17:17:50 +0000 | [diff] [blame] | 486 | inline int get_input_num_elems_processed(unsigned int num_elems_written_per_iteration, unsigned int stridex) |
| 487 | { |
| 488 | switch(stridex) |
| 489 | { |
| 490 | case 1: |
| 491 | return get_input_num_elems_processed<1>(num_elems_written_per_iteration); |
| 492 | case 2: |
| 493 | return get_input_num_elems_processed<2>(num_elems_written_per_iteration); |
| 494 | case 3: |
| 495 | return get_input_num_elems_processed<3>(num_elems_written_per_iteration); |
| 496 | default: |
| 497 | ARM_COMPUTE_ERROR("stridex not supported"); |
| 498 | return 0; |
| 499 | } |
| 500 | } |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 501 | } |
| 502 | } // namespace arm_compute |
Anthony Barbier | 1568621 | 2017-12-12 17:17:50 +0000 | [diff] [blame] | 503 | #endif /* __ARM_COMPUTE_NEDIRECTCONVOLUTIONDETAIL_H__ */ |