Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +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 | #ifndef __ARM_COMPUTE_TEST_TENSOR_OPERATIONS_H__ |
| 25 | #define __ARM_COMPUTE_TEST_TENSOR_OPERATIONS_H__ |
| 26 | |
| 27 | #include "FixedPoint.h" |
| 28 | #include "Tensor.h" |
| 29 | #include "Types.h" |
| 30 | #include "Utils.h" |
| 31 | |
| 32 | #include "FixedPoint.h" |
| 33 | #include "Types.h" |
| 34 | #include "arm_compute/core/FixedPoint.h" |
| 35 | #include "arm_compute/core/Types.h" |
| 36 | #include "tests/validation/FixedPoint.h" |
| 37 | |
| 38 | #include <algorithm> |
| 39 | #include <array> |
| 40 | #include <cmath> |
| 41 | |
| 42 | namespace arm_compute |
| 43 | { |
| 44 | namespace test |
| 45 | { |
| 46 | namespace validation |
| 47 | { |
| 48 | namespace tensor_operations |
| 49 | { |
| 50 | namespace |
| 51 | { |
| 52 | bool is_valid_pixel(int i, int min, int max) |
| 53 | { |
| 54 | return (i >= min && i < max); |
| 55 | } |
| 56 | |
| 57 | // 3D convolution for floating point type |
| 58 | template <typename T, typename std::enable_if<std::is_floating_point<T>::value, int>::type * = nullptr> |
| 59 | void convolution3d(const T *in, const T *weights, const T *bias, T *out, int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights, int8_t fixed_point_position) |
| 60 | { |
| 61 | const int half_width_weights = width_weights / 2; |
| 62 | const int half_height_weights = height_weights / 2; |
| 63 | |
| 64 | // Reset accumulator |
| 65 | T acc = static_cast<T>(0); |
| 66 | |
| 67 | // Compute a 2D convolution for each IFM and accumulate the result |
| 68 | for(int ifm = 0; ifm < depth_in; ++ifm) |
| 69 | { |
| 70 | // Compute the offset for the input slice |
| 71 | const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in; |
| 72 | |
| 73 | // Compute 2D convolution |
| 74 | for(int yk = -half_height_weights; yk <= half_height_weights; ++yk) |
| 75 | { |
| 76 | for(int xk = -half_width_weights; xk <= half_width_weights; ++xk) |
| 77 | { |
| 78 | // Check if the pixel is out-of-bound |
| 79 | if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in)) |
| 80 | { |
| 81 | const int idx = xk + half_width_weights; |
| 82 | const int idy = yk + half_height_weights; |
| 83 | |
| 84 | const T i_value = in[offset_slice_in + xk + yk * width_in]; |
| 85 | const T w_value = weights[idx + idy * width_weights + ifm * width_weights * height_weights]; |
| 86 | |
| 87 | acc += i_value * w_value; |
| 88 | } |
| 89 | } |
| 90 | } |
| 91 | } |
| 92 | |
| 93 | // Accumulate the bias and store the result |
| 94 | *out = acc + (*bias); |
| 95 | } |
| 96 | |
| 97 | // 3D convolution for fixed point type |
| 98 | template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type * = nullptr> |
| 99 | void convolution3d(const T *in, const T *weights, const T *bias, T *out, int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights, |
| 100 | int8_t fixed_point_position) |
| 101 | { |
| 102 | const int half_width_weights = width_weights / 2; |
| 103 | const int half_height_weights = height_weights / 2; |
| 104 | |
| 105 | using namespace fixed_point_arithmetic; |
| 106 | using promoted_type = typename fixed_point_arithmetic::traits::promote<T>::type; |
| 107 | |
| 108 | // Reset accumulator |
| 109 | fixed_point<promoted_type> acc(0, fixed_point_position); |
| 110 | |
| 111 | // Compute a 2D convolution for each IFM and accumulate the result |
| 112 | for(int ifm = 0; ifm < depth_in; ++ifm) |
| 113 | { |
| 114 | // Compute the offset for the input slice |
| 115 | const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in; |
| 116 | |
| 117 | // Compute 2D convolution |
| 118 | for(int yk = -half_height_weights; yk <= half_height_weights; ++yk) |
| 119 | { |
| 120 | for(int xk = -half_width_weights; xk <= half_width_weights; ++xk) |
| 121 | { |
| 122 | // Check if the pixel is out-of-bound |
| 123 | if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in)) |
| 124 | { |
| 125 | const int idx = xk + half_width_weights; |
| 126 | const int idy = yk + half_height_weights; |
| 127 | |
| 128 | const fixed_point<promoted_type> i_value(in[offset_slice_in + xk + yk * width_in], fixed_point_position, true); |
| 129 | const fixed_point<promoted_type> w_value(weights[idx + idy * width_weights + ifm * width_weights * height_weights], fixed_point_position, true); |
| 130 | const fixed_point<promoted_type> iw = i_value * w_value; |
| 131 | acc = iw + acc; |
| 132 | } |
| 133 | } |
| 134 | } |
| 135 | } |
| 136 | |
| 137 | // Get the bias |
| 138 | const fixed_point<promoted_type> b(*bias, fixed_point_position, true); |
| 139 | |
| 140 | // Accumulate the bias and covert back |
| 141 | acc = acc + b; |
| 142 | fixed_point<T> res(acc); |
| 143 | *out = res.raw(); |
| 144 | } |
| 145 | |
| 146 | template <typename T> |
| 147 | void vector_matrix_multiply(const T *in, const T *weights, const T *bias, T *out, int cols_weights, int rows_weights, uint8_t fixed_point_position) |
| 148 | { |
| 149 | for(int x = 0; x < cols_weights; ++x) |
| 150 | { |
| 151 | T acc = 0.0f; |
| 152 | for(int y = 0; y < rows_weights; ++y) |
| 153 | { |
| 154 | acc += in[y] * weights[x + y * cols_weights]; |
| 155 | } |
| 156 | out[x] = acc + bias[x]; |
| 157 | } |
| 158 | } |
| 159 | |
| 160 | template <> |
| 161 | void vector_matrix_multiply(const int8_t *in, const int8_t *weights, const int8_t *bias, int8_t *out, int cols_weights, int rows_weights, uint8_t fixed_point_position) |
| 162 | { |
| 163 | using namespace fixed_point_arithmetic; |
| 164 | using promoted_type = typename fixed_point_arithmetic::traits::promote<int8_t>::type; |
| 165 | |
| 166 | for(int x = 0; x < cols_weights; ++x) |
| 167 | { |
| 168 | // Reset accumulator |
| 169 | fixed_point<promoted_type> acc(0, fixed_point_position); |
| 170 | |
| 171 | for(int y = 0; y < rows_weights; ++y) |
| 172 | { |
| 173 | const fixed_point<promoted_type> i_value(in[y], fixed_point_position, true); |
| 174 | const fixed_point<promoted_type> w_value(weights[x + y * cols_weights], fixed_point_position, true); |
| 175 | const fixed_point<promoted_type> iw = i_value * w_value; |
| 176 | acc = iw + acc; |
| 177 | } |
| 178 | |
| 179 | // Get the bias |
| 180 | const fixed_point<int8_t> b(bias[x], fixed_point_position, true); |
| 181 | |
| 182 | // Convert back and accumulate the bias |
| 183 | fixed_point<int8_t> res(acc); |
| 184 | res = res + b; |
| 185 | |
| 186 | // Store the result |
| 187 | out[x] = res.raw(); |
| 188 | } |
| 189 | } |
| 190 | |
| 191 | /** Apply 2D spatial filter on a single element of @p in at coordinates @p coord |
| 192 | * |
| 193 | * - filter sizes have to be odd number |
| 194 | * - Valid region assumed |
| 195 | * - Row major order of filter assumed |
| 196 | * - TO_ZERO rounding policy assumed |
| 197 | * - SATURATE convert policy assumed |
| 198 | * |
| 199 | */ |
| 200 | template <typename T1, typename T2, typename T3> |
| 201 | void apply_2d_spatial_filter(Coordinates coord, const Tensor<T1> &in, Tensor<T3> &out, const TensorShape &filter_shape, const T2 *filter_itr, float scale) |
| 202 | { |
| 203 | using intermediate_type = typename common_promoted_signed_type<T1, T2, T3>::intermediate_type; |
| 204 | intermediate_type val = 0; |
| 205 | int x = coord.x(); |
| 206 | int y = coord.y(); |
| 207 | for(size_t j = y - filter_shape[1] / 2; j <= y + filter_shape[1] / 2; ++j) |
| 208 | { |
| 209 | for(size_t i = x - filter_shape[0] / 2; i <= x + filter_shape[0] / 2; ++i) |
| 210 | { |
| 211 | coord.set(0, i); |
| 212 | coord.set(1, j); |
| 213 | val += static_cast<intermediate_type>(*filter_itr) * static_cast<intermediate_type>(in[coord2index(in.shape(), coord)]); |
| 214 | ++filter_itr; |
| 215 | } |
| 216 | } |
| 217 | coord.set(0, x); |
| 218 | coord.set(1, y); |
| 219 | double rounded_val = cpp11::trunc(val * static_cast<double>(scale)); |
| 220 | out[coord2index(in.shape(), coord)] = saturate_cast<T3>(rounded_val); |
| 221 | } |
| 222 | } // namespace |
| 223 | |
| 224 | // Integral Image |
| 225 | void integral_image(const Tensor<uint8_t> &in, Tensor<uint32_t> &out) |
| 226 | { |
| 227 | // Length of dimensions |
| 228 | const size_t width = in.shape().x(); |
| 229 | const size_t height = in.shape().y(); |
| 230 | const size_t depth = in.shape().z() * in.shape()[3] * in.shape()[4] * in.shape()[5]; |
| 231 | |
| 232 | const size_t image_size = width * height; |
| 233 | |
| 234 | for(size_t z = 0; z < depth; ++z) |
| 235 | { |
| 236 | size_t current_image = z * image_size; |
| 237 | |
| 238 | //First element of each image |
| 239 | out[current_image] = in[current_image]; |
| 240 | |
| 241 | // First row of each image (add only pixel on the left) |
| 242 | for(size_t x = 1; x < width; ++x) |
| 243 | { |
| 244 | out[current_image + x] = static_cast<uint32_t>(in[current_image + x]) + out[current_image + x - 1]; |
| 245 | } |
| 246 | |
| 247 | // Subsequent rows |
| 248 | for(size_t y = 1; y < height; ++y) |
| 249 | { |
| 250 | size_t current_row = current_image + (width * y); |
| 251 | |
| 252 | // First element of each row (add only pixel up) |
| 253 | out[current_row] = static_cast<uint32_t>(in[current_row]) + out[current_row - width]; |
| 254 | |
| 255 | // Following row elements |
| 256 | for(size_t x = 1; x < width; ++x) |
| 257 | { |
| 258 | size_t current_pixel = current_row + x; |
| 259 | |
| 260 | // out = in + up(out) + left(out) - up_left(out) |
| 261 | out[current_pixel] = static_cast<uint32_t>(in[current_pixel]) + out[current_pixel - 1] |
| 262 | + out[current_pixel - width] - out[current_pixel - width - 1]; |
| 263 | } |
| 264 | } |
| 265 | } |
| 266 | } |
| 267 | |
| 268 | // Absolute difference |
| 269 | template <typename T1, typename T2, typename T3> |
| 270 | void absolute_difference(const Tensor<T1> &in1, const Tensor<T2> &in2, Tensor<T3> &out) |
| 271 | { |
| 272 | using intermediate_type = typename common_promoted_signed_type<T1, T2, T3>::intermediate_type; |
| 273 | |
| 274 | for(int i = 0; i < in1.num_elements(); ++i) |
| 275 | { |
| 276 | intermediate_type val = std::abs(static_cast<intermediate_type>(in1[i]) - static_cast<intermediate_type>(in2[i])); |
| 277 | out[i] = saturate_cast<T3>(val); |
| 278 | } |
| 279 | } |
| 280 | |
| 281 | // Accumulate |
| 282 | template <typename T1, typename T2> |
| 283 | void accumulate(const Tensor<T1> &in, Tensor<T2> &out) |
| 284 | { |
| 285 | using intermediate_type = typename common_promoted_signed_type<T1, T2>::intermediate_type; |
| 286 | |
| 287 | for(int i = 0; i < in.num_elements(); ++i) |
| 288 | { |
| 289 | intermediate_type val = static_cast<intermediate_type>(out[i]) + static_cast<intermediate_type>(in[i]); |
| 290 | out[i] = saturate_cast<T2>(val); |
| 291 | } |
| 292 | } |
| 293 | |
| 294 | // Accumulate squared |
| 295 | template <typename T1, typename T2> |
| 296 | void accumulate_squared(const Tensor<T1> &in, Tensor<T2> &out, uint32_t shift) |
| 297 | { |
| 298 | if(shift > 15) |
| 299 | { |
| 300 | ARM_COMPUTE_ERROR("Shift in accumulate_squared must be within the range [0, 15]"); |
| 301 | } |
| 302 | using intermediate_type = typename common_promoted_signed_type<T1, T2>::intermediate_type; |
| 303 | intermediate_type denom = 1 << shift; |
| 304 | |
| 305 | for(int i = 0; i < in.num_elements(); ++i) |
| 306 | { |
| 307 | intermediate_type val = static_cast<intermediate_type>(out[i]) + (static_cast<intermediate_type>(in[i]) * static_cast<intermediate_type>(in[i]) / denom); |
| 308 | out[i] = saturate_cast<T2>(val); |
| 309 | } |
| 310 | } |
| 311 | |
| 312 | // Accumulate weighted |
| 313 | template <typename T> |
| 314 | void accumulate_weighted(const Tensor<T> &in, Tensor<T> &out, float alpha) |
| 315 | { |
| 316 | if(alpha < 0.f || alpha > 1.f) |
| 317 | { |
| 318 | ARM_COMPUTE_ERROR("Weight (alpha) specified in accumulate_weighted must be within the range [0, 1]"); |
| 319 | } |
| 320 | using intermediate_type = typename common_promoted_signed_type<T>::intermediate_type; |
| 321 | |
| 322 | for(int i = 0; i < in.num_elements(); ++i) |
| 323 | { |
| 324 | double val = (1. - static_cast<double>(alpha)) * static_cast<intermediate_type>(out[i]) + static_cast<double>(alpha) * static_cast<intermediate_type>(in[i]); |
| 325 | out[i] = static_cast<T>(val); |
| 326 | } |
| 327 | } |
| 328 | |
| 329 | // Arithmetic addition |
| 330 | template <typename T1, typename T2, typename T3> |
| 331 | void arithmetic_addition(const Tensor<T1> &in1, const Tensor<T2> &in2, Tensor<T3> &out, ConvertPolicy convert_policy) |
| 332 | { |
| 333 | using intermediate_type = typename common_promoted_signed_type<T1, T2, T3>::intermediate_type; |
| 334 | |
| 335 | for(int i = 0; i < in1.num_elements(); ++i) |
| 336 | { |
| 337 | intermediate_type val = static_cast<intermediate_type>(in1[i]) + static_cast<intermediate_type>(in2[i]); |
| 338 | out[i] = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T3>(val) : static_cast<T3>(val); |
| 339 | } |
| 340 | } |
| 341 | |
| 342 | // Arithmetic Subtraction |
| 343 | template <typename T1, typename T2, typename T3> |
| 344 | void arithmetic_subtraction(const Tensor<T1> &in1, const Tensor<T2> &in2, Tensor<T3> &out, ConvertPolicy convert_policy) |
| 345 | { |
| 346 | using intermediate_type = typename common_promoted_signed_type<T1, T2, T3>::intermediate_type; |
| 347 | |
| 348 | for(int i = 0; i < in1.num_elements(); ++i) |
| 349 | { |
| 350 | intermediate_type val = static_cast<intermediate_type>(in1[i]) - static_cast<intermediate_type>(in2[i]); |
| 351 | out[i] = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T3>(val) : static_cast<T3>(val); |
| 352 | } |
| 353 | } |
| 354 | |
| 355 | // Bitwise and |
| 356 | template <typename T, typename = typename std::enable_if<std::is_integral<T>::value>::type> |
| 357 | void bitwise_and(const Tensor<T> &in1, const Tensor<T> &in2, Tensor<T> &out) |
| 358 | { |
| 359 | for(int i = 0; i < in1.num_elements(); ++i) |
| 360 | { |
| 361 | out[i] = in1[i] & in2[i]; |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | // Bitwise or |
| 366 | template <typename T, typename = typename std::enable_if<std::is_integral<T>::value>::type> |
| 367 | void bitwise_or(const Tensor<T> &in1, const Tensor<T> &in2, Tensor<T> &out) |
| 368 | { |
| 369 | for(int i = 0; i < in1.num_elements(); ++i) |
| 370 | { |
| 371 | out[i] = in1[i] | in2[i]; |
| 372 | } |
| 373 | } |
| 374 | |
| 375 | // Bitwise xor |
| 376 | template <typename T, typename = typename std::enable_if<std::is_integral<T>::value>::type> |
| 377 | void bitwise_xor(const Tensor<T> &in1, const Tensor<T> &in2, Tensor<T> &out) |
| 378 | { |
| 379 | for(int i = 0; i < in1.num_elements(); ++i) |
| 380 | { |
| 381 | out[i] = in1[i] ^ in2[i]; |
| 382 | } |
| 383 | } |
| 384 | |
| 385 | // Bitwise not |
| 386 | template <typename T, typename = typename std::enable_if<std::is_integral<T>::value>::type> |
| 387 | void bitwise_not(const Tensor<T> &in, Tensor<T> &out) |
| 388 | { |
| 389 | for(int i = 0; i < in.num_elements(); ++i) |
| 390 | { |
| 391 | out[i] = ~in[i]; |
| 392 | } |
| 393 | } |
| 394 | |
| 395 | // 3-by-3 box filter |
| 396 | template <typename T, typename = typename std::enable_if<std::is_integral<T>::value>::type> |
| 397 | void box3x3(const Tensor<T> &in, Tensor<T> &out) |
| 398 | { |
| 399 | const std::array<T, 9> filter{ { 1, 1, 1, 1, 1, 1, 1, 1, 1 } }; |
| 400 | float scale = 1.f / static_cast<float>(filter.size()); |
| 401 | const ValidRegion valid_region = shape_to_valid_region_undefined_border(in.shape(), BorderSize(1)); |
| 402 | for(int element_idx = 0; element_idx < in.num_elements(); ++element_idx) |
| 403 | { |
| 404 | const Coordinates id = index2coord(in.shape(), element_idx); |
| 405 | if(is_in_valid_region(valid_region, id)) |
| 406 | { |
| 407 | apply_2d_spatial_filter(id, in, out, TensorShape(3U, 3U), filter.data(), scale); |
| 408 | } |
| 409 | } |
| 410 | } |
| 411 | |
| 412 | // Depth conversion |
| 413 | template <typename T1, typename T2> |
| 414 | void depth_convert(const Tensor<T1> &in, Tensor<T2> &out, ConvertPolicy policy, uint32_t shift) |
| 415 | { |
| 416 | ARM_COMPUTE_ERROR("The conversion is not supported"); |
| 417 | } |
| 418 | |
| 419 | template <> |
| 420 | void depth_convert<int8_t, float>(const Tensor<int8_t> &in, Tensor<float> &out, ConvertPolicy policy, uint32_t shift) |
| 421 | { |
| 422 | const int8_t fixed_point_position = static_cast<int8_t>(in.fixed_point_position()); |
| 423 | for(int i = 0; i < in.num_elements(); ++i) |
| 424 | { |
| 425 | out[i] = static_cast<float>(in[i]) * (1.0f / (1 << fixed_point_position)); |
| 426 | } |
| 427 | } |
| 428 | |
| 429 | template <> |
| 430 | void depth_convert<float, int8_t>(const Tensor<float> &in, Tensor<int8_t> &out, ConvertPolicy policy, uint32_t shift) |
| 431 | { |
| 432 | const int8_t fixed_point_position = static_cast<int8_t>(in.fixed_point_position()); |
| 433 | for(int i = 0; i < in.num_elements(); ++i) |
| 434 | { |
| 435 | float val = in[i] * (1 << fixed_point_position) + 0.5f; |
| 436 | out[i] = ((policy == ConvertPolicy::SATURATE) ? saturate_cast<int8_t>(val) : static_cast<int8_t>(val)); |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | template <> |
| 441 | void depth_convert<uint8_t, uint16_t>(const Tensor<uint8_t> &in, Tensor<uint16_t> &out, ConvertPolicy policy, uint32_t shift) |
| 442 | { |
| 443 | for(int i = 0; i < in.num_elements(); ++i) |
| 444 | { |
| 445 | out[i] = static_cast<uint16_t>(in[i]) << shift; |
| 446 | } |
| 447 | } |
| 448 | |
| 449 | template <> |
| 450 | void depth_convert<uint8_t, int16_t>(const Tensor<uint8_t> &in, Tensor<int16_t> &out, ConvertPolicy policy, uint32_t shift) |
| 451 | { |
| 452 | for(int i = 0; i < in.num_elements(); ++i) |
| 453 | { |
| 454 | out[i] = static_cast<int16_t>(in[i]) << shift; |
| 455 | } |
| 456 | } |
| 457 | |
| 458 | template <> |
| 459 | void depth_convert<uint8_t, int32_t>(const Tensor<uint8_t> &in, Tensor<int32_t> &out, ConvertPolicy policy, uint32_t shift) |
| 460 | { |
| 461 | for(int i = 0; i < in.num_elements(); ++i) |
| 462 | { |
| 463 | out[i] = static_cast<int32_t>(in[i]) << shift; |
| 464 | } |
| 465 | } |
| 466 | |
| 467 | template <> |
| 468 | void depth_convert<uint16_t, uint8_t>(const Tensor<uint16_t> &in, Tensor<uint8_t> &out, ConvertPolicy policy, uint32_t shift) |
| 469 | { |
| 470 | for(int i = 0; i < in.num_elements(); ++i) |
| 471 | { |
| 472 | uint16_t val = in[i] >> shift; |
| 473 | out[i] = ((policy == ConvertPolicy::SATURATE) ? saturate_cast<uint8_t>(val) : static_cast<uint8_t>(val)); |
| 474 | } |
| 475 | } |
| 476 | |
| 477 | template <> |
| 478 | void depth_convert<uint16_t, uint32_t>(const Tensor<uint16_t> &in, Tensor<uint32_t> &out, ConvertPolicy policy, uint32_t shift) |
| 479 | { |
| 480 | for(int i = 0; i < in.num_elements(); ++i) |
| 481 | { |
| 482 | out[i] = static_cast<uint32_t>(in[i]) << shift; |
| 483 | } |
| 484 | } |
| 485 | |
| 486 | template <> |
| 487 | void depth_convert<int16_t, uint8_t>(const Tensor<int16_t> &in, Tensor<uint8_t> &out, ConvertPolicy policy, uint32_t shift) |
| 488 | { |
| 489 | for(int i = 0; i < in.num_elements(); ++i) |
| 490 | { |
| 491 | int16_t val = in[i] >> shift; |
| 492 | out[i] = ((policy == ConvertPolicy::SATURATE) ? saturate_cast<uint8_t>(val) : static_cast<uint8_t>(val)); |
| 493 | } |
| 494 | } |
| 495 | template <> |
| 496 | void depth_convert<int16_t, int32_t>(const Tensor<int16_t> &in, Tensor<int32_t> &out, ConvertPolicy policy, uint32_t shift) |
| 497 | { |
| 498 | for(int i = 0; i < in.num_elements(); ++i) |
| 499 | { |
| 500 | out[i] = static_cast<int32_t>(in[i]) << shift; |
| 501 | } |
| 502 | } |
| 503 | |
| 504 | // Matrix multiplication for floating point type |
| 505 | template <typename T, typename std::enable_if<std::is_floating_point<T>::value, int>::type * = nullptr> |
| 506 | void gemm(const Tensor<T> &in1, const Tensor<T> &in2, const Tensor<T> &in3, Tensor<T> &out, float alpha, float beta) |
| 507 | { |
| 508 | const int M = out.shape().y(); |
| 509 | const int N = out.shape().x(); |
| 510 | const int K = in1.shape().x(); |
| 511 | |
| 512 | for(int r = 0; r < M; ++r) |
| 513 | { |
| 514 | for(int c = 0; c < N; ++c) |
| 515 | { |
| 516 | T acc = 0.0f; |
| 517 | |
| 518 | for(int k = 0; k < K; ++k) |
| 519 | { |
| 520 | const T a0 = in1[r * K + k]; |
| 521 | const T b0 = in2[k * N + c]; |
| 522 | |
| 523 | acc += a0 * b0; |
| 524 | } |
| 525 | |
| 526 | // Finalize the result: A * B * alpha + C * beta |
| 527 | const T c0 = in3[c + r * N]; |
| 528 | out[c + r * N] = alpha * acc + beta * c0; |
| 529 | } |
| 530 | } |
| 531 | } |
| 532 | |
| 533 | // Matrix multiplication for fixed point type |
| 534 | template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type * = nullptr> |
| 535 | void gemm(const Tensor<T> &in1, const Tensor<T> &in2, const Tensor<T> &in3, Tensor<T> &out, float alpha, float beta) |
| 536 | { |
| 537 | using namespace fixed_point_arithmetic; |
| 538 | |
| 539 | using promoted_type = typename fixed_point_arithmetic::traits::promote<T>::type; |
| 540 | |
| 541 | const int M = out.shape().y(); |
| 542 | const int N = out.shape().x(); |
| 543 | const int K = in1.shape().x(); |
| 544 | const int8_t fixed_point_position = static_cast<int8_t>(in1.fixed_point_position()); |
| 545 | |
| 546 | const fixed_point<T> alpha_q(alpha, fixed_point_position); |
| 547 | const fixed_point<T> beta_q(beta, fixed_point_position); |
| 548 | |
| 549 | for(int r = 0; r < M; ++r) |
| 550 | { |
| 551 | for(int c = 0; c < N; ++c) |
| 552 | { |
| 553 | fixed_point<promoted_type> acc_q(0, fixed_point_position); |
| 554 | |
| 555 | for(int k = 0; k < K; ++k) |
| 556 | { |
| 557 | const fixed_point<promoted_type> a0_q(in1[r * K + k], fixed_point_position, true); |
| 558 | const fixed_point<promoted_type> b0_q(in2[k * N + c], fixed_point_position, true); |
| 559 | const fixed_point<promoted_type> axb_q = a0_q * b0_q; |
| 560 | |
| 561 | acc_q = axb_q + acc_q; |
| 562 | } |
| 563 | |
| 564 | // Finalize the result: A * B * alpha + C * beta |
| 565 | const fixed_point<T> c0_q(in3[c + r * N], fixed_point_position, true); |
| 566 | |
| 567 | fixed_point<T> res_q(acc_q); |
| 568 | res_q = alpha_q * res_q; |
| 569 | res_q = (c0_q * beta_q) + res_q; |
| 570 | |
| 571 | // Store the result |
| 572 | out[c + r * N] = res_q.raw(); |
| 573 | } |
| 574 | } |
| 575 | } |
| 576 | |
| 577 | // Pixel-wise multiplication |
| 578 | template <typename T1, typename T2, typename T3> |
| 579 | void pixel_wise_multiplication(const Tensor<T1> &in1, const Tensor<T2> &in2, Tensor<T3> &out, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy) |
| 580 | { |
| 581 | if(scale < 0) |
| 582 | { |
| 583 | ARM_COMPUTE_ERROR("Scale of pixel-wise multiplication must be non-negative"); |
| 584 | } |
| 585 | using intermediate_type = typename common_promoted_signed_type<T1, T2, T3>::intermediate_type; |
| 586 | for(int i = 0; i < in1.num_elements(); ++i) |
| 587 | { |
| 588 | double val = static_cast<intermediate_type>(in1[i]) * static_cast<intermediate_type>(in2[i]) * static_cast<double>(scale); |
| 589 | if(std::is_floating_point<T3>::value) |
| 590 | { |
| 591 | out[i] = val; |
| 592 | } |
| 593 | else |
| 594 | { |
| 595 | double rounded_val = 0; |
| 596 | switch(rounding_policy) |
| 597 | { |
| 598 | case(RoundingPolicy::TO_ZERO): |
| 599 | rounded_val = cpp11::trunc(val); |
| 600 | break; |
| 601 | case(RoundingPolicy::TO_NEAREST_UP): |
| 602 | rounded_val = cpp11::round_half_up(val); |
| 603 | break; |
| 604 | case(RoundingPolicy::TO_NEAREST_EVEN): |
| 605 | rounded_val = cpp11::round_half_even(val); |
| 606 | break; |
| 607 | default: |
| 608 | ARM_COMPUTE_ERROR("Unsupported rounding policy"); |
| 609 | } |
| 610 | out[i] = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T3>(rounded_val) : static_cast<T3>(rounded_val); |
| 611 | } |
| 612 | } |
| 613 | } |
| 614 | |
| 615 | // Fixed-point Pixel-wise Multiplication |
| 616 | template <typename T, typename = typename std::enable_if<std::is_integral<T>::value>::type> |
| 617 | void fixed_point_pixel_wise_multiplication(const Tensor<T> &in1, const Tensor<T> &in2, Tensor<T> &out, int scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy) |
| 618 | { |
| 619 | using namespace fixed_point_arithmetic; |
| 620 | |
| 621 | const int fixed_point_position = in1.fixed_point_position(); |
| 622 | |
| 623 | ARM_COMPUTE_ERROR_ON_MSG(in1.data_type() != in2.data_type() || in1.data_type() != out.data_type(), |
| 624 | "Tensors must all have the same DataType"); |
| 625 | ARM_COMPUTE_ERROR_ON_MSG(fixed_point_position != in2.fixed_point_position() || fixed_point_position != out.fixed_point_position(), |
| 626 | "Fixed-point position must be the same for both inputs and outputs"); |
| 627 | |
| 628 | // Validate fixed_point_position |
| 629 | ARM_COMPUTE_ERROR_ON((in1.data_type() == DataType::QS8) && (fixed_point_position == 0 || fixed_point_position > 7)); |
| 630 | ARM_COMPUTE_ERROR_ON((in1.data_type() == DataType::QS16) && (fixed_point_position == 0 || fixed_point_position > 15)); |
| 631 | |
| 632 | fixed_point<T> fp_scale(scale, fixed_point_position); |
| 633 | const bool is_sat = convert_policy == ConvertPolicy::SATURATE; |
| 634 | const bool do_scaling = scale != 1; |
| 635 | |
| 636 | for(int i = 0; i < in1.num_elements(); ++i) |
| 637 | { |
| 638 | fixed_point<T> val1(in1[i], fixed_point_position, true); |
| 639 | fixed_point<T> val2(in2[i], fixed_point_position, true); |
| 640 | fixed_point<T> res = (is_sat) ? val1 * val2 : mul<OverflowPolicy::WRAP>(val1, val2); |
| 641 | if(do_scaling) |
| 642 | { |
| 643 | res = (is_sat) ? res * fp_scale : mul<OverflowPolicy::WRAP>(res, fp_scale); |
| 644 | } |
| 645 | out[i] = res.raw(); |
| 646 | } |
| 647 | } |
| 648 | |
| 649 | // Threshold |
| 650 | template <typename T> |
| 651 | void threshold(const Tensor<T> &in, Tensor<T> &out, uint8_t threshold, uint8_t false_value, uint8_t true_value, ThresholdType type, uint8_t upper) |
| 652 | { |
| 653 | switch(type) |
| 654 | { |
| 655 | case ThresholdType::BINARY: |
| 656 | for(int i = 0; i < in.num_elements(); ++i) |
| 657 | { |
| 658 | out[i] = ((in[i] > threshold) ? true_value : false_value); |
| 659 | } |
| 660 | break; |
| 661 | case ThresholdType::RANGE: |
| 662 | for(int i = 0; i < in.num_elements(); ++i) |
| 663 | { |
| 664 | if(in[i] > upper) |
| 665 | { |
| 666 | out[i] = false_value; |
| 667 | } |
| 668 | else if(in[i] < threshold) |
| 669 | { |
| 670 | out[i] = false_value; |
| 671 | } |
| 672 | else |
| 673 | { |
| 674 | out[i] = true_value; |
| 675 | } |
| 676 | } |
| 677 | break; |
| 678 | default: |
| 679 | ARM_COMPUTE_ERROR("Thresholding type not recognised"); |
| 680 | break; |
| 681 | } |
| 682 | } |
| 683 | |
| 684 | // Activation Layer for floating point type |
| 685 | template <typename T, typename std::enable_if<std::is_floating_point<T>::value, int>::type * = nullptr> |
| 686 | void activation_layer(const Tensor<T> &in, Tensor<T> &out, ActivationLayerInfo act_info) |
| 687 | { |
| 688 | const T a = static_cast<T>(act_info.a()); |
| 689 | const T b = static_cast<T>(act_info.b()); |
| 690 | |
| 691 | for(int i = 0; i < in.num_elements(); ++i) |
| 692 | { |
| 693 | T x = in[i]; |
| 694 | switch(act_info.activation()) |
| 695 | { |
| 696 | case ActivationLayerInfo::ActivationFunction::ABS: |
| 697 | out[i] = std::abs(x); |
| 698 | break; |
| 699 | case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: |
| 700 | out[i] = std::min<T>(a, std::max<T>(0, x)); |
| 701 | break; |
| 702 | case ActivationLayerInfo::ActivationFunction::LINEAR: |
| 703 | out[i] = a * x + b; |
| 704 | break; |
| 705 | case ActivationLayerInfo::ActivationFunction::LOGISTIC: |
| 706 | out[i] = static_cast<T>(1) / (static_cast<T>(1) + std::exp(-x)); |
| 707 | break; |
| 708 | case ActivationLayerInfo::ActivationFunction::RELU: |
| 709 | out[i] = std::max<T>(0, x); |
| 710 | break; |
| 711 | case ActivationLayerInfo::ActivationFunction::SOFT_RELU: |
| 712 | out[i] = std::log(static_cast<T>(1) + std::exp(x)); |
| 713 | break; |
| 714 | case ActivationLayerInfo::ActivationFunction::SQRT: |
| 715 | out[i] = std::sqrt(x); |
| 716 | break; |
| 717 | case ActivationLayerInfo::ActivationFunction::SQUARE: |
| 718 | out[i] = x * x; |
| 719 | break; |
| 720 | case ActivationLayerInfo::ActivationFunction::TANH: |
| 721 | out[i] = a * std::tanh(b * x); |
| 722 | break; |
| 723 | default: |
| 724 | ARM_COMPUTE_ERROR("Activation function not recognised"); |
| 725 | break; |
| 726 | } |
| 727 | } |
| 728 | } |
| 729 | |
| 730 | // Activation Layer for fixed point type |
| 731 | template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type * = nullptr> |
| 732 | void activation_layer(const Tensor<T> &in, Tensor<T> &out, ActivationLayerInfo act_info) |
| 733 | { |
| 734 | using namespace fixed_point_arithmetic; |
| 735 | int fixed_point_position = in.fixed_point_position(); |
| 736 | ActivationLayerInfo::ActivationFunction act_func = act_info.activation(); |
| 737 | const fixed_point<T> a(act_info.a(), fixed_point_position); |
| 738 | const fixed_point<T> b(act_info.b(), fixed_point_position); |
| 739 | const fixed_point<T> const_0(0, fixed_point_position); |
| 740 | const fixed_point<T> const_1(1, fixed_point_position); |
| 741 | |
| 742 | for(int i = 0; i < in.num_elements(); ++i) |
| 743 | { |
| 744 | fixed_point<T> x(in[i], fixed_point_position, true); |
| 745 | switch(act_func) |
| 746 | { |
| 747 | case ActivationLayerInfo::ActivationFunction::ABS: |
| 748 | out[i] = abs(x).raw(); |
| 749 | break; |
| 750 | case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: |
| 751 | out[i] = min(a, max(const_0, x)).raw(); |
| 752 | break; |
| 753 | case ActivationLayerInfo::ActivationFunction::LINEAR: |
| 754 | out[i] = add(b, mul(a, x)).raw(); |
| 755 | break; |
| 756 | case ActivationLayerInfo::ActivationFunction::LOGISTIC: |
| 757 | out[i] = (const_1 / (const_1 + exp(-x))).raw(); |
| 758 | break; |
| 759 | case ActivationLayerInfo::ActivationFunction::RELU: |
| 760 | out[i] = max(const_0, x).raw(); |
| 761 | break; |
| 762 | case ActivationLayerInfo::ActivationFunction::SOFT_RELU: |
| 763 | out[i] = log(const_1 + exp(x)).raw(); |
| 764 | break; |
| 765 | case ActivationLayerInfo::ActivationFunction::SQRT: |
| 766 | out[i] = (const_1 / inv_sqrt(x)).raw(); |
| 767 | break; |
| 768 | case ActivationLayerInfo::ActivationFunction::SQUARE: |
| 769 | out[i] = mul(x, x).raw(); |
| 770 | break; |
| 771 | case ActivationLayerInfo::ActivationFunction::TANH: |
| 772 | out[i] = tanh(x).raw(); |
| 773 | break; |
| 774 | default: |
| 775 | ARM_COMPUTE_ERROR("Activation function not recognised"); |
| 776 | break; |
| 777 | } |
| 778 | } |
| 779 | } |
| 780 | |
| 781 | // Batch Normalization Layer for fixed point type |
| 782 | template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type * = nullptr> |
| 783 | void batch_normalization_layer(const Tensor<T> &in, Tensor<T> &out, const Tensor<T> &mean, const Tensor<T> &var, const Tensor<T> &beta, const Tensor<T> &gamma, float epsilon, int fixed_point_position) |
| 784 | { |
| 785 | const int cols = static_cast<int>(in.shape()[0]); |
| 786 | const int rows = static_cast<int>(in.shape()[1]); |
| 787 | const int depth = static_cast<int>(in.shape()[2]); |
| 788 | int upper_dims = in.shape().total_size() / (cols * rows * depth); |
| 789 | |
| 790 | for(int r = 0; r < upper_dims; ++r) |
| 791 | { |
| 792 | for(int i = 0; i < depth; ++i) |
| 793 | { |
| 794 | for(int k = 0; k < rows; ++k) |
| 795 | { |
| 796 | for(int l = 0; l < cols; ++l) |
| 797 | { |
| 798 | const int pos = l + k * cols + i * rows * cols + r * cols * rows * depth; |
| 799 | fixed_point_arithmetic::fixed_point<T> in_qs8(in[pos], fixed_point_position, true); |
| 800 | fixed_point_arithmetic::fixed_point<T> var_qs8(var[i], fixed_point_position, true); |
| 801 | fixed_point_arithmetic::fixed_point<T> mean_qs8(mean[i], fixed_point_position, true); |
| 802 | fixed_point_arithmetic::fixed_point<T> beta_qs8(beta[i], fixed_point_position, true); |
| 803 | fixed_point_arithmetic::fixed_point<T> gamma_qs8(gamma[i], fixed_point_position, true); |
| 804 | fixed_point_arithmetic::fixed_point<T> epsilon_qs8(epsilon, fixed_point_position); |
| 805 | |
| 806 | auto denominator = fixed_point_arithmetic::inv_sqrt(var_qs8 + epsilon_qs8); |
| 807 | auto numerator = in_qs8 - mean_qs8; |
| 808 | auto x_bar = numerator * denominator; |
| 809 | x_bar = beta_qs8 + x_bar * gamma_qs8; |
| 810 | out[pos] = x_bar.raw(); |
| 811 | } |
| 812 | } |
| 813 | } |
| 814 | } |
| 815 | } |
| 816 | |
| 817 | // Batch Normalization Layer for floating point type |
| 818 | template <typename T, typename std::enable_if<std::is_floating_point<T>::value, int>::type * = nullptr> |
| 819 | void batch_normalization_layer(const Tensor<T> &in, Tensor<T> &out, const Tensor<T> &mean, const Tensor<T> &var, const Tensor<T> &beta, const Tensor<T> &gamma, float epsilon, int fixed_point_position) |
| 820 | { |
| 821 | const int cols = static_cast<int>(in.shape()[0]); |
| 822 | const int rows = static_cast<int>(in.shape()[1]); |
| 823 | const int depth = static_cast<int>(in.shape()[2]); |
| 824 | int upper_dims = in.shape().total_size() / (cols * rows * depth); |
| 825 | |
| 826 | for(int r = 0; r < upper_dims; ++r) |
| 827 | { |
| 828 | for(int i = 0; i < depth; ++i) |
| 829 | { |
| 830 | for(int k = 0; k < rows; ++k) |
| 831 | { |
| 832 | for(int l = 0; l < cols; ++l) |
| 833 | { |
| 834 | const int pos = l + k * cols + i * rows * cols + r * cols * rows * depth; |
| 835 | const float denominator = sqrt(var[i] + epsilon); |
| 836 | const float numerator = in[pos] - mean[i]; |
| 837 | const float x_bar = numerator / denominator; |
| 838 | out[pos] = beta[i] + x_bar * gamma[i]; |
| 839 | } |
| 840 | } |
| 841 | } |
| 842 | } |
| 843 | } |
| 844 | |
| 845 | // Convolution layer |
| 846 | template <typename T> |
| 847 | void convolution_layer(const Tensor<T> &in, const Tensor<T> &weights, const Tensor<T> &bias, Tensor<T> &out, const PadStrideInfo &conv_info) |
| 848 | { |
| 849 | const int width_in = in.shape().x(); |
| 850 | const int height_in = in.shape().y(); |
| 851 | const int depth_in = in.shape().z(); |
| 852 | const int width_out = out.shape().x(); |
| 853 | const int height_out = out.shape().y(); |
| 854 | const int depth_out = out.shape().z(); |
| 855 | const int width_weights = weights.shape().x(); |
| 856 | const int height_weights = weights.shape().y(); |
| 857 | const int depth_weights = weights.shape().z(); |
| 858 | const int pad_xi = std::min(static_cast<int>(conv_info.pad().first), width_weights / 2); |
| 859 | const int pad_yi = std::min(static_cast<int>(conv_info.pad().second), height_weights / 2); |
| 860 | const int start_xi = width_weights / 2 - pad_xi; |
| 861 | const int start_yi = height_weights / 2 - pad_yi; |
| 862 | const int end_xi = width_in - start_xi; |
| 863 | const int end_yi = height_in - start_yi; |
| 864 | const int stride_xi = conv_info.stride().first; |
| 865 | const int stride_yi = conv_info.stride().second; |
| 866 | const int num_batches = in.shape().total_size() / (width_in * height_in * depth_in); |
| 867 | |
| 868 | for(int r = 0; r < num_batches; ++r) |
| 869 | { |
| 870 | for(int yi = start_yi; yi < end_yi; yi += stride_yi) |
| 871 | { |
| 872 | for(int xi = start_xi; xi < end_xi; xi += stride_xi) |
| 873 | { |
| 874 | for(int ofm = 0; ofm < depth_out; ++ofm) |
| 875 | { |
| 876 | // Compute input and output offsets |
| 877 | const int offset_in = r * width_in * height_in * depth_in; |
| 878 | const int xo = (xi - start_xi) / stride_xi; |
| 879 | const int yo = (yi - start_yi) / stride_yi; |
| 880 | const int offset_out = xo + yo * width_out + ofm * width_out * height_out + r * width_out * height_out * depth_out; |
| 881 | |
| 882 | // Compute 3D convolution |
| 883 | convolution3d(in.data() + offset_in, |
| 884 | weights.data() + ofm * width_weights * height_weights * depth_weights, |
| 885 | bias.data() + ofm, |
| 886 | out.data() + offset_out, |
| 887 | xi, yi, |
| 888 | width_in, height_in, depth_in, |
| 889 | width_weights, height_weights, |
| 890 | static_cast<int8_t>(in.fixed_point_position())); |
| 891 | } |
| 892 | } |
| 893 | } |
| 894 | } |
| 895 | } |
| 896 | |
| 897 | // Fully connected layer |
| 898 | template <typename T> |
| 899 | void fully_connected_layer(const Tensor<T> &in, const Tensor<T> &weights, const Tensor<T> &bias, Tensor<T> &out) |
| 900 | { |
| 901 | ARM_COMPUTE_ERROR_ON(weights.shape().x() != out.shape().x()); |
| 902 | ARM_COMPUTE_ERROR_ON(weights.shape().y() != in.shape().x() * in.shape().y() * in.shape().z()); |
| 903 | const int cols_weights = weights.shape().x(); |
| 904 | const int rows_weights = weights.shape().y(); |
| 905 | const int num_batches = in.shape().total_size() / rows_weights; |
| 906 | |
| 907 | for(int k = 0; k < num_batches; ++k) |
| 908 | { |
| 909 | vector_matrix_multiply<T>(in.data() + k * rows_weights, |
| 910 | weights.data(), |
| 911 | bias.data(), |
| 912 | out.data() + k * cols_weights, |
| 913 | cols_weights, |
| 914 | rows_weights, |
| 915 | in.fixed_point_position()); |
| 916 | } |
| 917 | } |
| 918 | |
| 919 | // Normalization Layer for floating point type |
| 920 | template <typename T, typename std::enable_if<std::is_floating_point<T>::value, int>::type * = nullptr> |
| 921 | void normalization_layer(const Tensor<T> &in, Tensor<T> &out, NormalizationLayerInfo norm_info) |
| 922 | { |
| 923 | const uint32_t norm_size = norm_info.norm_size(); |
| 924 | NormType type = norm_info.type(); |
| 925 | float beta = norm_info.beta(); |
| 926 | uint32_t kappa = norm_info.kappa(); |
| 927 | |
| 928 | const int cols = static_cast<int>(in.shape()[0]); |
| 929 | const int rows = static_cast<int>(in.shape()[1]); |
| 930 | const int depth = static_cast<int>(in.shape()[2]); |
| 931 | int upper_dims = in.shape().total_size() / (cols * rows); |
| 932 | |
| 933 | float coeff = norm_info.scale_coeff(); |
| 934 | int radius_cols = norm_size / 2; |
| 935 | // IN_MAP_1D and CROSS_MAP normalize over a single axis only |
| 936 | int radius_rows = (NormType::IN_MAP_2D == type) ? norm_size / 2 : 0; |
| 937 | |
| 938 | if(type == NormType::CROSS_MAP) |
| 939 | { |
| 940 | // Remove also depth from upper dimensions since it is the axes we want |
| 941 | // to use for normalization |
| 942 | upper_dims /= depth; |
| 943 | for(int r = 0; r < upper_dims; ++r) |
| 944 | { |
| 945 | for(int i = 0; i < rows; ++i) |
| 946 | { |
| 947 | for(int k = 0; k < cols; ++k) |
| 948 | { |
| 949 | for(int l = 0; l < depth; ++l) |
| 950 | { |
| 951 | float accumulated_scale = 0.f; |
| 952 | for(int j = -radius_cols; j <= radius_cols; ++j) |
| 953 | { |
| 954 | const int z = l + j; |
| 955 | if(z >= 0 && z < depth) |
| 956 | { |
| 957 | const T value = in[k + i * cols + z * rows * cols + r * cols * rows * depth]; |
| 958 | accumulated_scale += value * value; |
| 959 | } |
| 960 | } |
| 961 | out[k + i * cols + l * rows * cols + r * cols * rows * depth] = kappa + accumulated_scale * coeff; |
| 962 | } |
| 963 | } |
| 964 | } |
| 965 | } |
| 966 | } |
| 967 | else |
| 968 | { |
| 969 | for(int r = 0; r < upper_dims; ++r) |
| 970 | { |
| 971 | for(int i = 0; i < rows; ++i) |
| 972 | { |
| 973 | for(int k = 0; k < cols; ++k) |
| 974 | { |
| 975 | float accumulated_scale = 0.f; |
| 976 | for(int j = -radius_rows; j <= radius_rows; ++j) |
| 977 | { |
| 978 | const int y = i + j; |
| 979 | for(int l = -radius_cols; l <= radius_cols; ++l) |
| 980 | { |
| 981 | const int x = k + l; |
| 982 | if((x >= 0 && y >= 0) && (x < cols && y < rows)) |
| 983 | { |
| 984 | const T value = in[x + y * cols + r * cols * rows]; |
| 985 | accumulated_scale += value * value; |
| 986 | } |
| 987 | } |
| 988 | } |
| 989 | out[k + i * cols + r * cols * rows] = kappa + accumulated_scale * coeff; |
| 990 | } |
| 991 | } |
| 992 | } |
| 993 | } |
| 994 | |
| 995 | if(beta == 1.f) |
| 996 | { |
| 997 | for(int i = 0; i < out.num_elements(); ++i) |
| 998 | { |
| 999 | out[i] = in[i] / out[i]; |
| 1000 | } |
| 1001 | } |
| 1002 | else if(beta == 0.5f) |
| 1003 | { |
| 1004 | for(int i = 0; i < out.num_elements(); ++i) |
| 1005 | { |
| 1006 | out[i] = in[i] / std::sqrt(out[i]); |
| 1007 | } |
| 1008 | } |
| 1009 | else |
| 1010 | { |
| 1011 | for(int i = 0; i < out.num_elements(); ++i) |
| 1012 | { |
| 1013 | out[i] = in[i] * std::exp(std::log(out[i]) * -beta); |
| 1014 | } |
| 1015 | } |
| 1016 | } |
| 1017 | // Normalization Layer for fixed-point types |
| 1018 | template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type * = nullptr> |
| 1019 | void normalization_layer(const Tensor<T> &in, Tensor<T> &out, NormalizationLayerInfo norm_info) |
| 1020 | { |
| 1021 | using namespace fixed_point_arithmetic; |
| 1022 | |
| 1023 | const int fixed_point_position = in.fixed_point_position(); |
| 1024 | |
| 1025 | const uint32_t norm_size = norm_info.norm_size(); |
| 1026 | NormType type = norm_info.type(); |
| 1027 | fixed_point<T> beta(norm_info.beta(), fixed_point_position); |
| 1028 | fixed_point<T> kappa(norm_info.kappa(), fixed_point_position); |
| 1029 | |
| 1030 | const int cols = static_cast<int>(in.shape()[0]); |
| 1031 | const int rows = static_cast<int>(in.shape()[1]); |
| 1032 | const int depth = static_cast<int>(in.shape()[2]); |
| 1033 | int upper_dims = in.shape().total_size() / (cols * rows); |
| 1034 | |
| 1035 | fixed_point<T> coeff(norm_info.scale_coeff(), fixed_point_position); |
| 1036 | int radius_cols = norm_size / 2; |
| 1037 | // IN_MAP_1D and CROSS_MAP normalize over a single axis only |
| 1038 | int radius_rows = (NormType::IN_MAP_2D == type) ? norm_size / 2 : 0; |
| 1039 | |
| 1040 | if(type == NormType::CROSS_MAP) |
| 1041 | { |
| 1042 | // Remove also depth from upper dimensions since it is the axes we want |
| 1043 | // to use for normalization |
| 1044 | upper_dims /= depth; |
| 1045 | for(int r = 0; r < upper_dims; ++r) |
| 1046 | { |
| 1047 | for(int i = 0; i < rows; ++i) |
| 1048 | { |
| 1049 | for(int k = 0; k < cols; ++k) |
| 1050 | { |
| 1051 | for(int l = 0; l < depth; ++l) |
| 1052 | { |
| 1053 | fixed_point<T> accumulated_scale(0.f, fixed_point_position); |
| 1054 | for(int j = -radius_cols; j <= radius_cols; ++j) |
| 1055 | { |
| 1056 | const int z = l + j; |
| 1057 | if(z >= 0 && z < depth) |
| 1058 | { |
| 1059 | const T value = in[k + i * cols + z * rows * cols + r * cols * rows * depth]; |
| 1060 | const fixed_point<T> fp_value(value, fixed_point_position, true); |
| 1061 | accumulated_scale = add(accumulated_scale, mul(fp_value, fp_value)); |
| 1062 | } |
| 1063 | } |
| 1064 | accumulated_scale = add(kappa, mul(accumulated_scale, coeff)); |
| 1065 | out[k + i * cols + l * rows * cols + r * cols * rows * depth] = accumulated_scale.raw(); |
| 1066 | } |
| 1067 | } |
| 1068 | } |
| 1069 | } |
| 1070 | } |
| 1071 | else |
| 1072 | { |
| 1073 | for(int r = 0; r < upper_dims; ++r) |
| 1074 | { |
| 1075 | for(int i = 0; i < rows; ++i) |
| 1076 | { |
| 1077 | for(int k = 0; k < cols; ++k) |
| 1078 | { |
| 1079 | fixed_point<T> accumulated_scale(0.f, fixed_point_position); |
| 1080 | for(int j = -radius_rows; j <= radius_rows; ++j) |
| 1081 | { |
| 1082 | const int y = i + j; |
| 1083 | for(int l = -radius_cols; l <= radius_cols; ++l) |
| 1084 | { |
| 1085 | const int x = k + l; |
| 1086 | if((x >= 0 && y >= 0) && (x < cols && y < rows)) |
| 1087 | { |
| 1088 | const T value = in[x + y * cols + r * cols * rows]; |
| 1089 | const fixed_point<T> fp_value(value, fixed_point_position, true); |
| 1090 | accumulated_scale = add(accumulated_scale, mul(fp_value, fp_value)); |
| 1091 | } |
| 1092 | } |
| 1093 | } |
| 1094 | accumulated_scale = add(kappa, mul(accumulated_scale, coeff)); |
| 1095 | out[k + i * cols + r * cols * rows] = accumulated_scale.raw(); |
| 1096 | } |
| 1097 | } |
| 1098 | } |
| 1099 | } |
| 1100 | |
| 1101 | if(norm_info.beta() == 1.f) |
| 1102 | { |
| 1103 | for(int i = 0; i < out.num_elements(); ++i) |
| 1104 | { |
| 1105 | fixed_point<T> res = div(fixed_point<T>(in[i], fixed_point_position, true), fixed_point<T>(out[i], fixed_point_position, true)); |
| 1106 | out[i] = res.raw(); |
| 1107 | } |
| 1108 | } |
| 1109 | else |
| 1110 | { |
| 1111 | const fixed_point<T> beta(norm_info.beta(), fixed_point_position); |
| 1112 | for(int i = 0; i < out.num_elements(); ++i) |
| 1113 | { |
| 1114 | fixed_point<T> res = pow(fixed_point<T>(out[i], fixed_point_position, true), beta); |
| 1115 | res = div(fixed_point<T>(in[i], fixed_point_position, true), res); |
| 1116 | out[i] = res.raw(); |
| 1117 | } |
| 1118 | } |
| 1119 | } |
| 1120 | |
| 1121 | // Pooling layer |
| 1122 | template <typename T> |
| 1123 | void pooling_layer(const Tensor<T> &in, Tensor<T> &out, PoolingLayerInfo pool_info, int fixed_point_position) |
| 1124 | { |
| 1125 | const int pool_size = pool_info.pool_size(); |
| 1126 | PoolingType type = pool_info.pool_type(); |
| 1127 | int pool_stride_x = 0; |
| 1128 | int pool_stride_y = 0; |
| 1129 | int pad_x = 0; |
| 1130 | int pad_y = 0; |
| 1131 | std::tie(pool_stride_x, pool_stride_y) = pool_info.pad_stride_info().stride(); |
| 1132 | std::tie(pad_x, pad_y) = pool_info.pad_stride_info().pad(); |
| 1133 | |
| 1134 | const int cols_in = static_cast<int>(in.shape()[0]); |
| 1135 | const int rows_in = static_cast<int>(in.shape()[1]); |
| 1136 | |
| 1137 | const int cols_out = static_cast<int>(out.shape()[0]); |
| 1138 | const int rows_out = static_cast<int>(out.shape()[1]); |
| 1139 | |
| 1140 | int upper_dims = in.shape().total_size() / (cols_in * rows_in); |
| 1141 | |
| 1142 | int pooled_height = static_cast<int>(ceil(static_cast<float>(rows_in + 2 * pad_x - pool_size) / pool_stride_x)) + 1; |
| 1143 | int pooled_width = static_cast<int>(ceil(static_cast<float>(cols_in + 2 * pad_y - pool_size) / pool_stride_y)) + 1; |
| 1144 | |
| 1145 | if((pooled_height - 1) * pool_stride_x >= rows_in + pad_x) |
| 1146 | { |
| 1147 | --pooled_height; |
| 1148 | } |
| 1149 | if((pooled_width - 1) * pool_stride_y >= cols_in + pad_y) |
| 1150 | { |
| 1151 | --pooled_width; |
| 1152 | } |
| 1153 | |
| 1154 | if(type == PoolingType::MAX) |
| 1155 | { |
| 1156 | for(int r = 0; r < upper_dims; ++r) |
| 1157 | { |
| 1158 | for(int i = 0; i < pooled_height; ++i) |
| 1159 | { |
| 1160 | for(int k = 0; k < pooled_width; ++k) |
| 1161 | { |
| 1162 | int hstart = i * pool_stride_x - pad_x; |
| 1163 | int wstart = k * pool_stride_y - pad_y; |
| 1164 | int hend = std::min(hstart + pool_size, rows_in); |
| 1165 | int wend = std::min(wstart + pool_size, cols_in); |
| 1166 | hstart = std::max(hstart, 0); |
| 1167 | wstart = std::max(wstart, 0); |
| 1168 | |
| 1169 | T max_val = std::numeric_limits<T>::lowest(); |
| 1170 | for(int y = hstart; y < hend; ++y) |
| 1171 | { |
| 1172 | for(int x = wstart; x < wend; ++x) |
| 1173 | { |
| 1174 | T val = in[r * cols_in * rows_in + y * cols_in + x]; |
| 1175 | if(val > max_val) |
| 1176 | { |
| 1177 | max_val = val; |
| 1178 | } |
| 1179 | } |
| 1180 | } |
| 1181 | |
| 1182 | out[r * rows_out * cols_out + i * pooled_width + k] = max_val; |
| 1183 | } |
| 1184 | } |
| 1185 | } |
| 1186 | } |
| 1187 | else // Average pooling |
| 1188 | { |
| 1189 | for(int r = 0; r < upper_dims; ++r) |
| 1190 | { |
| 1191 | for(int i = 0; i < pooled_height; ++i) |
| 1192 | { |
| 1193 | for(int k = 0; k < pooled_width; ++k) |
| 1194 | { |
| 1195 | T avg_val = 0; |
| 1196 | |
| 1197 | int hstart = i * pool_stride_x - pad_x; |
| 1198 | int wstart = k * pool_stride_y - pad_y; |
| 1199 | int hend = std::min(hstart + pool_size, cols_in + pad_x); |
| 1200 | int wend = std::min(wstart + pool_size, rows_in + pad_y); |
| 1201 | int pool = (hend - hstart) * (wend - wstart); |
| 1202 | hstart = std::max(hstart, 0); |
| 1203 | wstart = std::max(wstart, 0); |
| 1204 | hend = std::min(hend, rows_in); |
| 1205 | wend = std::min(wend, cols_in); |
| 1206 | |
| 1207 | if(std::is_floating_point<T>::value) |
| 1208 | { |
| 1209 | for(int y = hstart; y < hend; ++y) |
| 1210 | { |
| 1211 | for(int x = wstart; x < wend; ++x) |
| 1212 | { |
| 1213 | avg_val += in[r * cols_in * rows_in + y * cols_in + x]; |
| 1214 | } |
| 1215 | } |
| 1216 | out[r * rows_out * cols_out + i * pooled_width + k] = avg_val / pool; |
| 1217 | } |
| 1218 | else |
| 1219 | { |
| 1220 | static std::array<qint8_t, 10> scale_values_q8 = |
| 1221 | { { 0x0, 0x0, 0x40, 0x2A, 0x20, 0x19, 0x15, 0x12, 0x10, 0xE } }; |
| 1222 | |
| 1223 | for(int y = hstart; y < hend; ++y) |
| 1224 | { |
| 1225 | for(int x = wstart; x < wend; ++x) |
| 1226 | { |
| 1227 | avg_val = sqadd_qs8(avg_val, in[r * cols_in * rows_in + y * cols_in + x]); |
| 1228 | } |
| 1229 | } |
| 1230 | out[r * rows_out * cols_out + i * pooled_width + k] = sqmul_qs8(avg_val, (scale_values_q8[pool] >> (7 - fixed_point_position)), fixed_point_position); |
| 1231 | } |
| 1232 | } |
| 1233 | } |
| 1234 | } |
| 1235 | } |
| 1236 | } |
| 1237 | |
| 1238 | // Softmax Layer |
| 1239 | template <typename T, typename std::enable_if<std::is_floating_point<T>::value, int>::type * = nullptr> |
| 1240 | void softmax_layer(const Tensor<T> &in, Tensor<T> &out) |
| 1241 | { |
| 1242 | const int cols = static_cast<int>(in.shape()[0]); |
| 1243 | const int upper_dims = in.shape().total_size() / cols; |
| 1244 | for(int r = 0; r < upper_dims; ++r) |
| 1245 | { |
| 1246 | // Find max |
| 1247 | T max = std::numeric_limits<T>::lowest(); |
| 1248 | for(int c = 0; c < cols; ++c) |
| 1249 | { |
| 1250 | const T x = in[r * cols + c]; |
| 1251 | if(x > max) |
| 1252 | { |
| 1253 | max = x; |
| 1254 | } |
| 1255 | } |
| 1256 | |
| 1257 | // Regularize |
| 1258 | T sum = 0; |
| 1259 | for(int c = 0; c < cols; ++c) |
| 1260 | { |
| 1261 | const T res = exp(in[r * cols + c] - max); |
| 1262 | out[r * cols + c] = res; |
| 1263 | sum += res; |
| 1264 | } |
| 1265 | |
| 1266 | // Normalize |
| 1267 | const T norm_val = 1 / sum; |
| 1268 | for(int c = 0; c < cols; ++c) |
| 1269 | { |
| 1270 | out[r * cols + c] *= norm_val; |
| 1271 | } |
| 1272 | } |
| 1273 | } |
| 1274 | template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type * = nullptr> |
| 1275 | void softmax_layer(const Tensor<T> &in, Tensor<T> &out) |
| 1276 | { |
| 1277 | using namespace fixed_point_arithmetic; |
| 1278 | using promoted_T = typename test::traits::promote<T>::type; |
| 1279 | |
| 1280 | const int fixed_point_position = in.fixed_point_position(); |
| 1281 | const int cols = static_cast<int>(in.shape()[0]); |
| 1282 | const int upper_dims = in.shape().total_size() / cols; |
| 1283 | |
| 1284 | for(int r = 0; r < upper_dims; ++r) |
| 1285 | { |
| 1286 | // Find max |
| 1287 | fixed_point<T> max(std::numeric_limits<T>::lowest(), fixed_point_position, true); |
| 1288 | for(int c = 0; c < cols; ++c) |
| 1289 | { |
| 1290 | const fixed_point<T> x(in[r * cols + c], fixed_point_position, true); |
| 1291 | if(x > max) |
| 1292 | { |
| 1293 | max = x; |
| 1294 | } |
| 1295 | } |
| 1296 | |
| 1297 | // Regularize |
| 1298 | fixed_point<promoted_T> sum(0, fixed_point_position); |
| 1299 | for(int c = 0; c < cols; ++c) |
| 1300 | { |
| 1301 | const fixed_point<T> x(in[r * cols + c], fixed_point_position, true); |
| 1302 | fixed_point<T> res = exp(x - max); |
| 1303 | out[r * cols + c] = res.raw(); |
| 1304 | sum = add(sum, static_cast<fixed_point<promoted_T>>(res)); |
| 1305 | } |
| 1306 | |
| 1307 | // Normalize |
| 1308 | fixed_point<T> sat_sum(sum); |
| 1309 | for(int c = 0; c < cols; ++c) |
| 1310 | { |
| 1311 | const fixed_point<T> x(out[r * cols + c], fixed_point_position, true); |
| 1312 | out[r * cols + c] = div(x, sat_sum).raw(); |
| 1313 | } |
| 1314 | } |
| 1315 | } |
| 1316 | |
| 1317 | // Fixed point operations |
| 1318 | template <typename T> |
| 1319 | void fixed_point_operation(const Tensor<T> &in, Tensor<T> &out, FixedPointOp op) |
| 1320 | { |
| 1321 | int p = in.fixed_point_position(); |
| 1322 | switch(op) |
| 1323 | { |
| 1324 | case FixedPointOp::EXP: |
| 1325 | for(int i = 0; i < in.num_elements(); ++i) |
| 1326 | { |
| 1327 | out[i] = fixed_point_arithmetic::exp(fixed_point_arithmetic::fixed_point<T>(in[i], p, true)).raw(); |
| 1328 | } |
| 1329 | break; |
| 1330 | case FixedPointOp::LOG: |
| 1331 | for(int i = 0; i < in.num_elements(); ++i) |
| 1332 | { |
| 1333 | out[i] = fixed_point_arithmetic::log(fixed_point_arithmetic::fixed_point<T>(in[i], p, true)).raw(); |
| 1334 | } |
| 1335 | break; |
| 1336 | case FixedPointOp::INV_SQRT: |
| 1337 | for(int i = 0; i < in.num_elements(); ++i) |
| 1338 | { |
| 1339 | out[i] = fixed_point_arithmetic::inv_sqrt(fixed_point_arithmetic::fixed_point<T>(in[i], p, true)).raw(); |
| 1340 | } |
| 1341 | break; |
| 1342 | case FixedPointOp::RECIPROCAL: |
| 1343 | for(int i = 0; i < in.num_elements(); ++i) |
| 1344 | { |
| 1345 | out[i] = fixed_point_arithmetic::div(fixed_point_arithmetic::fixed_point<T>(1, p), fixed_point_arithmetic::fixed_point<T>(in[i], p, true)).raw(); |
| 1346 | } |
| 1347 | break; |
| 1348 | default: |
| 1349 | ARM_COMPUTE_ERROR("Fixed point operation not supported"); |
| 1350 | break; |
| 1351 | } |
| 1352 | } |
| 1353 | |
| 1354 | // Tensor print |
| 1355 | template <typename T> |
| 1356 | void print(const Tensor<T> &in, std::ostream &out) |
| 1357 | { |
| 1358 | out << "\n"; |
| 1359 | for(int i = 0; i < in.num_elements(); ++i) |
| 1360 | { |
| 1361 | out << in[i] << " "; |
| 1362 | } |
| 1363 | out << "\n"; |
| 1364 | } |
| 1365 | } // namespace tensor_operations |
| 1366 | } // namespace validation |
| 1367 | } // namespace test |
| 1368 | } // namespace arm_compute |
| 1369 | |
| 1370 | #endif /* __ARM_COMPUTE_TEST_TENSOR_OPERATIONS_H__ */ |