Dana Zlotnik | ebbae94 | 2022-02-03 12:52:15 +0200 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2019-2022 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 | #include "src/cpu/kernels/depthwiseconv2d/generic/neon/impl.h" |
| 25 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
| 26 | #include "src/core/NEON/wrapper/wrapper.h" |
| 27 | |
| 28 | namespace arm_compute |
| 29 | { |
| 30 | namespace cpu |
| 31 | { |
| 32 | namespace |
| 33 | { |
| 34 | constexpr auto data_layout = DataLayout::NHWC; |
| 35 | const size_t width_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); |
| 36 | const size_t height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); |
| 37 | const size_t channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); |
| 38 | |
| 39 | constexpr auto dim_manual_loop = Window::Dimension(0, 0, 0); |
| 40 | constexpr auto dim_single_unit_step = Window::Dimension(0, 1, 1); |
| 41 | constexpr size_t vector_size = 8; |
| 42 | |
| 43 | struct DepthwiseConvolutionRunInfo |
| 44 | { |
| 45 | const size_t num_read_elements_per_iteration; |
| 46 | const uint32_t x_start; |
| 47 | const uint32_t x_end; |
| 48 | const uint32_t x_step; |
| 49 | const uint32_t x_leftover_start; |
| 50 | const size_t input_stride_y; |
| 51 | const size_t input_stride_z; |
| 52 | const size_t input_max_offset; |
| 53 | const size_t weights_width; |
| 54 | const size_t weights_height; |
| 55 | const size_t weights_stride_y; |
| 56 | const size_t weights_stride_z; |
| 57 | const size_t conv_stride_x; |
| 58 | const size_t conv_stride_y; |
| 59 | const size_t conv_pad_left; |
| 60 | const size_t conv_pad_top; |
| 61 | const size_t input_height; |
| 62 | const size_t input_width; |
| 63 | const size_t input_depth; |
| 64 | |
| 65 | DepthwiseConvolutionRunInfo(const ITensorInfo &input, const ITensorInfo &weights, const PadStrideInfo &conv_info, const Window &w, uint32_t depth_multiplier = 1) // NOLINT |
| 66 | : num_read_elements_per_iteration((depth_multiplier == 1 ? (vector_size / element_size_from_data_type(input.data_type())) : 1)), |
| 67 | x_start(w.x().start()), |
| 68 | x_end(w.x().end()), |
| 69 | x_step(static_cast<uint32_t>(num_read_elements_per_iteration * depth_multiplier)), |
| 70 | x_leftover_start(std::max(static_cast<int32_t>(w.x().end() + 1) - static_cast<int32_t>(x_step), int32_t(0))), |
| 71 | input_stride_y(input.strides_in_bytes().y()), |
| 72 | input_stride_z(input.strides_in_bytes().z()), |
| 73 | input_max_offset(input.strides_in_bytes().z() * input.dimension(height_idx) - (input.padding().bottom + input.padding().top) * input.strides_in_bytes().y()), |
| 74 | weights_width(weights.dimension(width_idx)), |
| 75 | weights_height(weights.dimension(height_idx)), |
| 76 | weights_stride_y(weights.strides_in_bytes().y()), |
| 77 | weights_stride_z(weights.strides_in_bytes().z()), |
| 78 | conv_stride_x(conv_info.stride().first), |
| 79 | conv_stride_y(conv_info.stride().second), |
| 80 | conv_pad_left(conv_info.pad_left()), |
| 81 | conv_pad_top(conv_info.pad_top()), |
| 82 | input_height(input.dimension(height_idx)), |
| 83 | input_width(input.dimension(width_idx)), |
| 84 | input_depth(input.dimension(channel_idx)) |
| 85 | { |
| 86 | } |
| 87 | }; |
| 88 | |
| 89 | inline int32x4_t saturating_doubling_high_mul(const int32x4_t &a, const int32_t &b) |
| 90 | { |
| 91 | return vqrdmulhq_n_s32(a, b); |
| 92 | } |
| 93 | |
| 94 | inline int32_t saturating_doubling_high_mul(const int32_t &a, const int32_t &b) |
| 95 | { |
| 96 | return vget_lane_s32(vqrdmulh_n_s32(vdup_n_s32(a), b), 0); |
| 97 | } |
| 98 | |
| 99 | inline int32x4_t rounding_divide_by_exp2(const int32x4_t &x, const int exponent) |
| 100 | { |
| 101 | const int32x4_t shift = vdupq_n_s32(-exponent); |
| 102 | const int32x4_t fixup = vshrq_n_s32(vandq_s32(x, shift), 31); |
| 103 | const int32x4_t fixed = vqaddq_s32(x, fixup); |
| 104 | return vrshlq_s32(fixed, shift); |
| 105 | } |
| 106 | |
| 107 | inline int32x2_t rounding_divide_by_exp2(const int32x2_t &x, const int exponent) |
| 108 | { |
| 109 | const int32x2_t shift = vdup_n_s32(-exponent); |
| 110 | const int32x2_t fixup = vshr_n_s32(vand_s32(x, shift), 31); |
| 111 | const int32x2_t fixed = vqadd_s32(x, fixup); |
| 112 | return vrshl_s32(fixed, shift); |
| 113 | } |
| 114 | |
| 115 | inline int32_t rounding_divide_by_exp2(const int32_t &x, const int exponent) |
| 116 | { |
| 117 | const int32x2_t xs = vdup_n_s32(x); |
| 118 | return vget_lane_s32(rounding_divide_by_exp2(xs, exponent), 0); |
| 119 | } |
| 120 | |
| 121 | inline bool is_valid_input_region(int32_t base_w, uint32_t base_h, uint32_t w, uint32_t h, const DepthwiseConvolutionRunInfo &run_info, const Size2D &dilation) |
| 122 | { |
| 123 | const int32_t current_h = base_h + h * dilation.y(); |
| 124 | const bool is_valid_h = current_h >= 0 && current_h < static_cast<int32_t>(run_info.input_height); |
| 125 | |
| 126 | const int32_t current_w = base_w + w * dilation.x(); |
| 127 | const bool is_valid_w = current_w >= 0 && current_w < static_cast<int32_t>(run_info.input_width); |
| 128 | |
| 129 | return is_valid_h && is_valid_w; |
| 130 | } |
| 131 | |
| 132 | template <typename T> |
| 133 | void depthwise_loop_multiplier1_fp(const ITensor *src, const ITensor *weights, const ITensor *biases, ITensor *dst, const PadStrideInfo &conv_info, |
| 134 | const Size2D &dilation, const Window &window, bool has_biases) |
| 135 | { |
| 136 | constexpr auto element_per_vector = vector_size / sizeof(T); |
| 137 | using VectorType = typename wrapper::traits::neon_vector<T, element_per_vector>::type; |
| 138 | using TagType = typename wrapper::traits::neon_vector<T, element_per_vector>::tag_type; |
| 139 | |
| 140 | const auto run_info = DepthwiseConvolutionRunInfo(*src->info(), *weights->info(), conv_info, window); |
| 141 | |
| 142 | const VectorType zero_vector = wrapper::vdup_n(static_cast<T>(0), TagType{}); |
| 143 | |
| 144 | Window execution_window = window; |
| 145 | execution_window.set(Window::DimX, dim_single_unit_step); |
| 146 | |
| 147 | Window win_input = window; |
| 148 | win_input.set(Window::DimX, dim_manual_loop); |
| 149 | win_input.set(Window::DimY, dim_manual_loop); |
| 150 | win_input.set(Window::DimZ, dim_manual_loop); |
| 151 | |
| 152 | Window win_weights = win_input; |
| 153 | win_weights.set(Window::DimW, dim_manual_loop); |
| 154 | |
| 155 | Window win_output = window; |
| 156 | win_output.set(Window::DimX, dim_manual_loop); |
| 157 | |
| 158 | Iterator input_it(src, win_input); |
| 159 | Iterator weights_it(weights, win_weights); |
| 160 | Iterator output_it(dst, win_output); |
| 161 | Iterator biases_it{}; |
| 162 | |
| 163 | if(has_biases) |
| 164 | { |
| 165 | biases_it = Iterator(biases, win_weights); |
| 166 | } |
| 167 | |
| 168 | execute_window_loop(execution_window, [&](const Coordinates & id) |
| 169 | { |
| 170 | const int32_t input_y = id.y() * run_info.conv_stride_x - run_info.conv_pad_left; |
| 171 | const int32_t input_z = id.z() * run_info.conv_stride_y - run_info.conv_pad_top; |
| 172 | const int64_t base_input_offset = input_y * run_info.input_stride_y + input_z * run_info.input_stride_z; |
| 173 | |
| 174 | auto const base_weights_ptr = weights_it.ptr(); |
| 175 | uint32_t x = run_info.x_start; |
| 176 | |
| 177 | for(; x < run_info.x_leftover_start; x += run_info.x_step) |
| 178 | { |
| 179 | VectorType acc = zero_vector; |
| 180 | auto weights_ptr = base_weights_ptr; |
| 181 | int64_t input_offset = base_input_offset; |
| 182 | |
| 183 | for(uint32_t h = 0; h < run_info.weights_height; ++h) |
| 184 | { |
| 185 | int64_t offs = input_offset + x * sizeof(T); |
| 186 | for(uint32_t w = 0; w < run_info.weights_width; ++w) |
| 187 | { |
| 188 | const bool is_valid_region = is_valid_input_region(input_y, input_z, w, h, run_info, dilation); |
| 189 | const auto input_vals = is_valid_region ? |
| 190 | wrapper::vload(reinterpret_cast<T *>(input_it.ptr() + std::min(static_cast<size_t>(offs), run_info.input_max_offset))) : |
| 191 | zero_vector; |
| 192 | const auto weights_vals = wrapper::vload(reinterpret_cast<T *>(weights_ptr + w * run_info.weights_stride_y) + x); |
| 193 | acc = wrapper::vmla(acc, weights_vals, input_vals); |
| 194 | |
| 195 | offs += dilation.x() * run_info.input_stride_y; |
| 196 | } |
| 197 | |
| 198 | weights_ptr += run_info.weights_stride_z; |
| 199 | input_offset += dilation.y() * run_info.input_stride_z; |
| 200 | } |
| 201 | |
| 202 | if(has_biases) |
| 203 | { |
| 204 | const auto biases_vals = wrapper::vload(reinterpret_cast<T *>(biases_it.ptr()) + x); |
| 205 | acc = wrapper::vadd(acc, biases_vals); |
| 206 | } |
| 207 | |
| 208 | wrapper::vstore(reinterpret_cast<T *>(output_it.ptr()) + x, acc); |
| 209 | } |
| 210 | |
| 211 | for(; x < run_info.x_end; ++x) |
| 212 | { |
| 213 | auto acc_scalar = T{ 0 }; |
| 214 | auto weights_ptr = base_weights_ptr; |
| 215 | int64_t input_offset = base_input_offset; |
| 216 | |
| 217 | for(size_t h = 0; h < run_info.weights_height; ++h) |
| 218 | { |
| 219 | int64_t offs = input_offset + x * sizeof(T); |
| 220 | for(size_t w = 0; w < run_info.weights_width; ++w) |
| 221 | { |
| 222 | const bool is_valid_region = is_valid_input_region(input_y, input_z, w, h, run_info, dilation); |
| 223 | const auto input_vals = is_valid_region ? *reinterpret_cast<T *>(input_it.ptr() + std::min(static_cast<size_t>(offs), run_info.input_max_offset)) : 0; |
| 224 | const auto weights_vals = *(reinterpret_cast<T *>(weights_ptr + w * run_info.weights_stride_y) + x); |
| 225 | |
| 226 | acc_scalar += (input_vals * weights_vals); |
| 227 | |
| 228 | offs += dilation.x() * run_info.input_stride_y; |
| 229 | } |
| 230 | |
| 231 | weights_ptr += run_info.weights_stride_z; |
| 232 | input_offset += dilation.y() * run_info.input_stride_z; |
| 233 | } |
| 234 | |
| 235 | if(has_biases) |
| 236 | { |
| 237 | const auto biases_vals = *(reinterpret_cast<T *>(biases_it.ptr()) + x); |
| 238 | acc_scalar += biases_vals; |
| 239 | } |
| 240 | *(reinterpret_cast<T *>(output_it.ptr()) + x) = acc_scalar; |
| 241 | } |
| 242 | }, |
| 243 | input_it, weights_it, biases_it, output_it); |
| 244 | } |
| 245 | |
| 246 | template <typename T> |
| 247 | void depthwise_loop_generic_fp(const ITensor *src, const ITensor *weights, const ITensor *biases, ITensor *dst, const PadStrideInfo &conv_info, |
| 248 | const Size2D &dilation, unsigned int depth_multiplier, const Window &window, bool has_biases) |
| 249 | { |
| 250 | const auto run_info = DepthwiseConvolutionRunInfo(*src->info(), *weights->info(), conv_info, window, depth_multiplier); |
| 251 | |
| 252 | Window execution_window = window; |
| 253 | execution_window.set(Window::DimX, Window::Dimension(0, run_info.input_depth, 1)); |
| 254 | |
| 255 | Window win_input = execution_window; |
| 256 | win_input.set(Window::DimX, Window::Dimension(0, run_info.input_depth, 1)); |
| 257 | win_input.set(Window::DimY, dim_manual_loop); |
| 258 | win_input.set(Window::DimZ, dim_manual_loop); |
| 259 | |
| 260 | Window win_weights = window; |
| 261 | win_weights.set_dimension_step(Window::DimX, run_info.x_step); |
| 262 | win_weights.set(Window::DimY, dim_manual_loop); |
| 263 | win_weights.set(Window::DimZ, dim_manual_loop); |
| 264 | win_weights.set(Window::DimW, dim_manual_loop); |
| 265 | |
| 266 | Window win_output = window; |
| 267 | win_output.set_dimension_step(Window::DimX, run_info.x_step); |
| 268 | |
| 269 | Iterator input_it(src, win_input); |
| 270 | Iterator weights_it(weights, win_weights); |
| 271 | Iterator output_it(dst, win_output); |
| 272 | Iterator biases_it{}; |
| 273 | |
| 274 | if(has_biases) |
| 275 | { |
| 276 | biases_it = Iterator(biases, win_weights); |
| 277 | } |
| 278 | |
| 279 | execute_window_loop(execution_window, [&](const Coordinates & id) |
| 280 | { |
| 281 | std::vector<T> acc(depth_multiplier, static_cast<T>(0)); |
| 282 | |
| 283 | const int input_y = id.y() * run_info.conv_stride_x - run_info.conv_pad_left; |
| 284 | const int input_z = id.z() * run_info.conv_stride_y - run_info.conv_pad_top; |
| 285 | int input_offset = input_y * run_info.input_stride_y + input_z * run_info.input_stride_z; |
| 286 | |
| 287 | auto weights_ptr = weights_it.ptr(); |
| 288 | for(size_t h = 0; h < run_info.weights_height; ++h) |
| 289 | { |
| 290 | int offs = input_offset; |
| 291 | for(size_t w = 0; w < run_info.weights_width; ++w) |
| 292 | { |
| 293 | const bool is_valid_region = is_valid_input_region(input_y, input_z, w, h, run_info, dilation); |
| 294 | const auto input_val = is_valid_region ? *(reinterpret_cast<T *>(input_it.ptr() + std::min(static_cast<size_t>(offs), run_info.input_max_offset))) : T(0); |
| 295 | |
| 296 | for(size_t m = 0; m < depth_multiplier; ++m) |
| 297 | { |
| 298 | const auto weights_val = *(reinterpret_cast<T *>(weights_ptr + m * sizeof(T) + w * run_info.weights_stride_y)); |
| 299 | acc.at(m) = support::cpp11::fma(weights_val, input_val, acc.at(m)); |
| 300 | } |
| 301 | |
| 302 | offs += dilation.x() * run_info.input_stride_y; |
| 303 | } |
| 304 | |
| 305 | weights_ptr += run_info.weights_stride_z; |
| 306 | input_offset += dilation.y() * run_info.input_stride_z; |
| 307 | } |
| 308 | |
| 309 | if(has_biases) |
| 310 | { |
| 311 | for(size_t m = 0; m < depth_multiplier; ++m) |
| 312 | { |
| 313 | const auto biases_val = *(reinterpret_cast<T *>(biases_it.ptr() + m * sizeof(T))); |
| 314 | *(reinterpret_cast<T *>(output_it.ptr() + m * sizeof(T))) = acc.at(m) + biases_val; |
| 315 | } |
| 316 | } |
| 317 | else |
| 318 | { |
| 319 | for(size_t m = 0; m < depth_multiplier; ++m) |
| 320 | { |
| 321 | *(reinterpret_cast<T *>(output_it.ptr() + m * sizeof(T))) = acc.at(m); |
| 322 | } |
| 323 | } |
| 324 | }, |
| 325 | input_it, weights_it, biases_it, output_it); |
| 326 | } |
| 327 | |
| 328 | template <typename T, typename TW> |
| 329 | void depthwise_loop_multiplier1_quantized(const ITensor *src, const ITensor *weights, const ITensor *biases, ITensor *dst, const PadStrideInfo &conv_info, |
| 330 | const Size2D &dilation, std::vector<int> output_multiplier, std::vector<int> output_shift, const Window &window, bool has_biases) // NOLINT |
| 331 | { |
| 332 | ARM_COMPUTE_UNUSED(output_multiplier, output_shift); |
| 333 | constexpr auto element_per_vector = vector_size / sizeof(T); |
| 334 | using VectorType = typename wrapper::traits::neon_vector<T, element_per_vector>::type; |
| 335 | using TagType = typename wrapper::traits::neon_vector<T, element_per_vector>::tag_type; |
| 336 | using AccType = int32_t; |
| 337 | using AccArrayType = std::array<AccType, element_per_vector>; |
| 338 | |
| 339 | const auto out_of_bound_value = PixelValue(static_cast<uint64_t>(0), src->info()->data_type(), src->info()->quantization_info()).get<T>(); |
| 340 | const auto out_of_bound_vector = wrapper::vdup_n(static_cast<T>(out_of_bound_value), TagType{}); |
| 341 | |
| 342 | const auto run_info = DepthwiseConvolutionRunInfo(*src->info(), *weights->info(), conv_info, window); |
| 343 | |
| 344 | const int32_t input_qoffset = src->info()->quantization_info().uniform().offset; |
| 345 | const int32_t weights_qoffset = weights->info()->quantization_info().uniform().offset; |
| 346 | const int32_t output_qoffset = dst->info()->quantization_info().uniform().offset; |
| 347 | const int32_t k_offset = run_info.weights_width * run_info.weights_height * input_qoffset * weights_qoffset; |
| 348 | |
| 349 | Window execution_window = window; |
| 350 | execution_window.set(Window::DimX, dim_single_unit_step); |
| 351 | |
| 352 | Window win_input = window; |
| 353 | win_input.set(Window::DimX, dim_manual_loop); |
| 354 | win_input.set(Window::DimY, dim_manual_loop); |
| 355 | win_input.set(Window::DimZ, dim_manual_loop); |
| 356 | |
| 357 | Window win_weights = win_input; |
| 358 | win_weights.set(Window::DimW, dim_manual_loop); |
| 359 | |
| 360 | Window win_output = window; |
| 361 | win_output.set(Window::DimX, dim_manual_loop); |
| 362 | |
| 363 | Iterator input_it(src, win_input); |
| 364 | Iterator weights_it(weights, win_weights); |
| 365 | Iterator output_it(dst, win_output); |
| 366 | Iterator biases_it{}; |
| 367 | |
| 368 | if(has_biases) |
| 369 | { |
| 370 | biases_it = Iterator(biases, win_weights); |
| 371 | } |
| 372 | |
| 373 | execute_window_loop(execution_window, [&](const Coordinates & id) |
| 374 | { |
| 375 | const int32_t input_y = id.y() * run_info.conv_stride_x - run_info.conv_pad_left; |
| 376 | const int32_t input_z = id.z() * run_info.conv_stride_y - run_info.conv_pad_top; |
| 377 | const int64_t base_input_offset = input_y * run_info.input_stride_y + input_z * run_info.input_stride_z; |
| 378 | auto const base_weights_ptr = weights_it.ptr(); |
| 379 | size_t x = run_info.x_start; |
| 380 | |
| 381 | for(; x < run_info.x_leftover_start; x += run_info.x_step) |
| 382 | { |
| 383 | AccArrayType acc{}; |
| 384 | AccArrayType in_sum{}; |
| 385 | AccArrayType we_sum{}; |
| 386 | |
| 387 | auto weights_ptr = base_weights_ptr; |
| 388 | auto input_offset = base_input_offset; |
| 389 | |
| 390 | for(size_t h = 0; h < run_info.weights_height; ++h) |
| 391 | { |
| 392 | int64_t offs = input_offset + x * sizeof(T); |
| 393 | for(size_t w = 0; w < run_info.weights_width; ++w) |
| 394 | { |
| 395 | const bool is_valid_region = is_valid_input_region(input_y, input_z, w, h, run_info, dilation); |
| 396 | const auto input_vals = is_valid_region ? |
| 397 | wrapper::vload(reinterpret_cast<T *>(input_it.ptr() + std::min(static_cast<size_t>(offs), run_info.input_max_offset))) : |
| 398 | out_of_bound_vector; |
| 399 | const auto weights_vals = wrapper::vload(reinterpret_cast<TW *>(weights_ptr + w * run_info.weights_stride_y) + x); |
| 400 | |
| 401 | for(size_t i = 0; i < element_per_vector; ++i) |
| 402 | { |
| 403 | acc.at(i) += input_vals[i] * weights_vals[i]; |
| 404 | in_sum.at(i) += input_vals[i]; |
| 405 | we_sum.at(i) += weights_vals[i]; |
| 406 | } |
| 407 | |
| 408 | offs += dilation.x() * run_info.input_stride_y; |
| 409 | } |
| 410 | |
| 411 | weights_ptr += run_info.weights_stride_z; |
| 412 | input_offset += dilation.y() * run_info.input_stride_z; |
| 413 | } |
| 414 | |
| 415 | VectorType out_vals = wrapper::vdup_n(static_cast<T>(0), TagType{}); |
| 416 | for(size_t i = 0; i < element_per_vector; ++i) |
| 417 | { |
| 418 | acc.at(i) -= in_sum.at(i) * weights_qoffset; |
| 419 | acc.at(i) -= we_sum.at(i) * input_qoffset; |
| 420 | acc.at(i) += k_offset; |
| 421 | |
| 422 | if(has_biases) |
| 423 | { |
| 424 | acc.at(i) += *(reinterpret_cast<int32_t *>(biases_it.ptr() + i * sizeof(int32_t)) + x); |
| 425 | } |
| 426 | |
| 427 | const int32_t out_mul = output_multiplier.at(x + i); |
| 428 | const int32_t out_shift = output_shift.at(x + i); |
| 429 | if(out_shift < 0) |
| 430 | { |
| 431 | acc.at(i) = saturating_doubling_high_mul(acc.at(i) * (1 << (-out_shift)), out_mul) + output_qoffset; |
| 432 | } |
| 433 | else |
| 434 | { |
| 435 | acc.at(i) = rounding_divide_by_exp2(saturating_doubling_high_mul(acc.at(i), out_mul), out_shift) + output_qoffset; |
| 436 | } |
| 437 | out_vals[i] = static_cast<T>(utility::clamp<AccType, T>(acc.at(i))); |
| 438 | } |
| 439 | |
| 440 | wrapper::vstore(reinterpret_cast<T *>(output_it.ptr()) + x, out_vals); |
| 441 | } |
| 442 | |
| 443 | // left-over |
| 444 | for(; x < run_info.x_end; ++x) |
| 445 | { |
| 446 | AccType acc = 0; |
| 447 | AccType in_sum = 0; |
| 448 | AccType we_sum = 0; |
| 449 | |
| 450 | auto weights_ptr = base_weights_ptr; |
| 451 | auto input_offset = base_input_offset; |
| 452 | |
| 453 | for(size_t h = 0; h < run_info.weights_height; ++h) |
| 454 | { |
| 455 | int64_t offs = input_offset + x * sizeof(T); |
| 456 | for(size_t w = 0; w < run_info.weights_width; ++w) |
| 457 | { |
| 458 | const bool is_valid_region = is_valid_input_region(input_y, input_z, w, h, run_info, dilation); |
| 459 | const auto input_val = is_valid_region ? |
| 460 | *reinterpret_cast<T *>(input_it.ptr() + std::min(static_cast<size_t>(offs), run_info.input_max_offset)) : |
| 461 | out_of_bound_value; |
| 462 | const auto weights_val = *(reinterpret_cast<TW *>(weights_ptr + w * run_info.weights_stride_y) + x); |
| 463 | |
| 464 | acc += input_val * weights_val; |
| 465 | in_sum += input_val; |
| 466 | we_sum += weights_val; |
| 467 | |
| 468 | offs += dilation.x() * run_info.input_stride_y; |
| 469 | } |
| 470 | |
| 471 | weights_ptr += run_info.weights_stride_z; |
| 472 | input_offset += dilation.y() * run_info.input_stride_z; |
| 473 | } |
| 474 | |
| 475 | T out_vals{ 0 }; |
| 476 | |
| 477 | acc -= in_sum * weights_qoffset; |
| 478 | acc -= we_sum * input_qoffset; |
| 479 | acc += k_offset; |
| 480 | |
| 481 | if(has_biases) |
| 482 | { |
| 483 | acc += *(reinterpret_cast<int32_t *>(biases_it.ptr()) + x); |
| 484 | } |
| 485 | |
| 486 | const int32_t out_mul = output_multiplier.at(x); |
| 487 | const int32_t out_shift = output_shift.at(x); |
| 488 | |
| 489 | if(out_shift < 0) |
| 490 | { |
| 491 | acc = saturating_doubling_high_mul(acc * (1 << (-out_shift)), out_mul) + output_qoffset; |
| 492 | } |
| 493 | else |
| 494 | { |
| 495 | acc = rounding_divide_by_exp2(saturating_doubling_high_mul(acc, out_mul), out_shift) + output_qoffset; |
| 496 | } |
| 497 | |
| 498 | out_vals = static_cast<T>(utility::clamp<AccType, T>(acc)); |
| 499 | *(reinterpret_cast<T *>(output_it.ptr()) + x) = out_vals; |
| 500 | } |
| 501 | }, |
| 502 | input_it, weights_it, biases_it, output_it); |
| 503 | } |
| 504 | |
| 505 | template <typename T, typename TW> |
| 506 | void depthwise_loop_generic_quantized(const ITensor *src, const ITensor *weights, const ITensor *biases, ITensor *dst, const PadStrideInfo &conv_info, |
| 507 | const Size2D &dilation, unsigned int depth_multiplier, std::vector<int> output_multiplier, std::vector<int> output_shift, const Window &window, bool has_biases) // NOLINT |
| 508 | { |
| 509 | using AccType = int32_t; |
| 510 | |
| 511 | const auto run_info = DepthwiseConvolutionRunInfo(*src->info(), *weights->info(), conv_info, window, depth_multiplier); |
| 512 | |
| 513 | const auto out_of_bound_value = PixelValue(static_cast<uint64_t>(0), src->info()->data_type(), src->info()->quantization_info()).get<T>(); |
| 514 | |
| 515 | const int32_t input_qoffset = src->info()->quantization_info().uniform().offset; |
| 516 | const int32_t weights_qoffset = weights->info()->quantization_info().uniform().offset; |
| 517 | const int32_t output_qoffset = dst->info()->quantization_info().uniform().offset; |
| 518 | const int32_t k_offset = run_info.weights_width * run_info.weights_height * input_qoffset * weights_qoffset; |
| 519 | |
| 520 | Window execution_window = window; |
| 521 | execution_window.set(Window::DimX, Window::Dimension(0, run_info.input_depth, 1)); |
| 522 | |
| 523 | Window win_input = execution_window; |
| 524 | win_input.set(Window::DimY, dim_manual_loop); |
| 525 | win_input.set(Window::DimZ, dim_manual_loop); |
| 526 | |
| 527 | Window win_weights = window; |
| 528 | win_weights.set_dimension_step(Window::DimX, run_info.x_step); |
| 529 | win_weights.set(Window::DimY, dim_manual_loop); |
| 530 | win_weights.set(Window::DimZ, dim_manual_loop); |
| 531 | win_weights.set(Window::DimW, dim_manual_loop); |
| 532 | |
| 533 | Window win_output = window; |
| 534 | win_output.set_dimension_step(Window::DimX, run_info.x_step); |
| 535 | |
| 536 | Iterator input_it(src, win_input); |
| 537 | Iterator weights_it(weights, win_weights); |
| 538 | Iterator output_it(dst, win_output); |
| 539 | Iterator biases_it{}; |
| 540 | |
| 541 | if(has_biases) |
| 542 | { |
| 543 | biases_it = Iterator(biases, win_weights); |
| 544 | } |
| 545 | |
| 546 | execute_window_loop(execution_window, [&](const Coordinates & id) |
| 547 | { |
| 548 | std::vector<AccType> acc(depth_multiplier, 0); |
| 549 | std::vector<AccType> we_sum(depth_multiplier, 0); |
| 550 | AccType in_sum = 0; |
| 551 | |
| 552 | const int32_t input_y = id.y() * run_info.conv_stride_x - run_info.conv_pad_left; |
| 553 | const int32_t input_z = id.z() * run_info.conv_stride_y - run_info.conv_pad_top; |
| 554 | int64_t input_offset = input_y * run_info.input_stride_y + input_z * run_info.input_stride_z; |
| 555 | |
| 556 | auto weights_ptr = weights_it.ptr(); |
| 557 | for(size_t h = 0; h < run_info.weights_height; ++h) |
| 558 | { |
| 559 | int offs = input_offset; |
| 560 | for(size_t w = 0; w < run_info.weights_width; ++w) |
| 561 | { |
| 562 | const bool is_valid_region = is_valid_input_region(input_y, input_z, w, h, run_info, dilation); |
| 563 | const auto input_val = is_valid_region ? *(reinterpret_cast<T *>(input_it.ptr() + std::min(static_cast<size_t>(offs), run_info.input_max_offset))) : out_of_bound_value; |
| 564 | |
| 565 | for(size_t m = 0; m < depth_multiplier; ++m) |
| 566 | { |
| 567 | const auto weights_val = *(reinterpret_cast<TW *>(weights_ptr + m * sizeof(T) + w * run_info.weights_stride_y)); |
| 568 | acc.at(m) += input_val * weights_val; |
| 569 | |
| 570 | we_sum.at(m) += weights_val; |
| 571 | } |
| 572 | |
| 573 | offs += dilation.x() * run_info.input_stride_y; |
| 574 | in_sum += input_val; |
| 575 | } |
| 576 | |
| 577 | weights_ptr += run_info.weights_stride_z; |
| 578 | input_offset += dilation.y() * run_info.input_stride_z; |
| 579 | } |
| 580 | |
| 581 | for(size_t m = 0; m < depth_multiplier; ++m) |
| 582 | { |
| 583 | acc.at(m) -= in_sum * weights_qoffset; |
| 584 | acc.at(m) -= we_sum.at(m) * input_qoffset; |
| 585 | acc.at(m) += k_offset; |
| 586 | |
| 587 | if(has_biases) |
| 588 | { |
| 589 | acc.at(m) += *(reinterpret_cast<int32_t *>(biases_it.ptr() + m * sizeof(int32_t))); |
| 590 | } |
| 591 | |
| 592 | const int32_t out_mul = output_multiplier.at(id.x() * depth_multiplier + m); |
| 593 | const int32_t out_shift = output_shift.at(id.x() * depth_multiplier + m); |
| 594 | if(out_shift < 0) |
| 595 | { |
| 596 | acc.at(m) = saturating_doubling_high_mul(acc.at(m) * (1 << (-out_shift)), out_mul) + output_qoffset; |
| 597 | } |
| 598 | else |
| 599 | { |
| 600 | acc.at(m) = rounding_divide_by_exp2(saturating_doubling_high_mul(acc.at(m), out_mul), out_shift) + output_qoffset; |
| 601 | } |
| 602 | *(reinterpret_cast<T *>(output_it.ptr() + m * sizeof(T))) = static_cast<T>(utility::clamp<AccType, T>(acc.at(m))); |
| 603 | } |
| 604 | }, |
| 605 | input_it, weights_it, biases_it, output_it); |
| 606 | } |
| 607 | |
| 608 | template <typename T, typename TW> |
| 609 | void depthwise_loop_pow2_quantized_per_tensor(const ITensor *src, const ITensor *weights, const ITensor *biases, ITensor *dst, const PadStrideInfo &conv_info, |
| 610 | const Size2D &dilation, unsigned int depth_multiplier, std::vector<int> output_multiplier, std::vector<int> output_shift, const Window &window, bool has_biases) // NOLINT |
| 611 | { |
| 612 | constexpr int half_vec = vector_size / 2; |
| 613 | |
| 614 | using AccType = int32_t; |
| 615 | using AccVectorType = typename wrapper::traits::neon_vector<AccType, half_vec>::type; |
| 616 | using AccVectorTagType = typename wrapper::traits::neon_vector<AccType, half_vec>::tag_type; |
| 617 | using TagType = typename wrapper::traits::neon_vector<T, vector_size>::tag_type; |
| 618 | |
| 619 | const auto run_info = DepthwiseConvolutionRunInfo(*src->info(), *weights->info(), conv_info, window, depth_multiplier); |
| 620 | |
| 621 | const auto input_qoffset_vec = wrapper::vreinterpret(wrapper::vmovl(wrapper::vdup_n(static_cast<T>(src->info()->quantization_info().uniform().offset), TagType{}))); |
| 622 | const auto weights_qoffset_vec = wrapper::vreinterpret(wrapper::vmovl(wrapper::vdup_n(static_cast<TW>(weights->info()->quantization_info().uniform().offset), TagType{}))); |
| 623 | const auto output_qoffset_vec = wrapper::vdup_n(dst->info()->quantization_info().uniform().offset, arm_compute::wrapper::traits::vector_128_tag{}); |
| 624 | |
| 625 | const auto lower = wrapper::vdup_n(static_cast<AccType>(std::numeric_limits<T>::lowest()), AccVectorTagType{}); |
| 626 | const auto upper = wrapper::vdup_n(static_cast<AccType>(std::numeric_limits<T>::max()), AccVectorTagType{}); |
| 627 | const auto zero = wrapper::vdup_n(static_cast<AccType>(0), AccVectorTagType{}); |
| 628 | |
| 629 | const auto out_mul = output_multiplier.at(0); |
| 630 | const auto out_shift = output_shift.at(0); |
| 631 | |
| 632 | Window execution_window = window; |
| 633 | execution_window.set(Window::DimX, Window::Dimension(0, run_info.input_depth, 1)); |
| 634 | |
| 635 | Window win_input = execution_window; |
| 636 | win_input.set(Window::DimY, dim_manual_loop); |
| 637 | win_input.set(Window::DimZ, dim_manual_loop); |
| 638 | |
| 639 | Window win_weights = window; |
| 640 | win_weights.set_dimension_step(Window::DimX, run_info.x_step); |
| 641 | win_weights.set(Window::DimY, dim_manual_loop); |
| 642 | win_weights.set(Window::DimZ, dim_manual_loop); |
| 643 | win_weights.set(Window::DimW, dim_manual_loop); |
| 644 | |
| 645 | Window win_output = window; |
| 646 | win_output.set_dimension_step(Window::DimX, run_info.x_step); |
| 647 | |
| 648 | Iterator input_it(src, win_input); |
| 649 | Iterator weights_it(weights, win_weights); |
| 650 | Iterator output_it(dst, win_output); |
| 651 | Iterator biases_it{}; |
| 652 | |
| 653 | if(has_biases) |
| 654 | { |
| 655 | biases_it = Iterator(biases, win_weights); |
| 656 | } |
| 657 | |
| 658 | std::vector<AccVectorType> acc0(depth_multiplier / vector_size); |
| 659 | std::vector<AccVectorType> acc1(depth_multiplier / vector_size); |
| 660 | |
| 661 | execute_window_loop(execution_window, [&](const Coordinates & id) |
| 662 | { |
| 663 | std::fill(begin(acc0), end(acc0), zero); |
| 664 | std::fill(begin(acc1), end(acc1), zero); |
| 665 | |
| 666 | const int32_t input_y = id.y() * run_info.conv_stride_x - run_info.conv_pad_left; |
| 667 | const int32_t input_z = id.z() * run_info.conv_stride_y - run_info.conv_pad_top; |
| 668 | int64_t input_offset = input_y * run_info.input_stride_y + input_z * run_info.input_stride_z; |
| 669 | |
| 670 | auto weights_ptr = weights_it.ptr(); |
| 671 | for(size_t h = 0; h < run_info.weights_height; ++h) |
| 672 | { |
| 673 | const int32_t current_h = input_z + h * dilation.y(); |
| 674 | if(current_h >= 0 && current_h < static_cast<int32_t>(run_info.input_height)) |
| 675 | { |
| 676 | int offs = input_offset; |
| 677 | for(size_t w = 0; w < run_info.weights_width; ++w) |
| 678 | { |
| 679 | const int32_t current_w = input_y + w * dilation.x(); |
| 680 | if(current_w >= 0 && current_w < static_cast<int32_t>(run_info.input_width)) |
| 681 | { |
| 682 | const auto input_8x8 = wrapper::vdup_n(*(reinterpret_cast<T *>(input_it.ptr() + std::min(static_cast<size_t>(offs), run_info.input_max_offset))), TagType{}); |
| 683 | const auto input_s16x8 = wrapper::vreinterpret(wrapper::vmovl(input_8x8)); |
| 684 | const auto input_no_offs = wrapper::vsub(input_s16x8, input_qoffset_vec); |
| 685 | |
| 686 | for(size_t m = 0, i = 0; m < depth_multiplier; m += vector_size, ++i) |
| 687 | { |
| 688 | const auto weights_8x8 = wrapper::vload(reinterpret_cast<TW *>(weights_ptr + m * sizeof(T) + w * run_info.weights_stride_y)); |
| 689 | const auto weights_s16x8 = wrapper::vreinterpret(wrapper::vmovl(weights_8x8)); |
| 690 | const auto weights_no_offs = wrapper::vsub(weights_s16x8, weights_qoffset_vec); |
| 691 | |
| 692 | acc0.at(i) = wrapper::vmlal(acc0.at(i), wrapper::vgetlow(input_no_offs), wrapper::vgetlow(weights_no_offs)); |
| 693 | acc1.at(i) = wrapper::vmlal(acc1.at(i), wrapper::vgethigh(input_no_offs), wrapper::vgethigh(weights_no_offs)); |
| 694 | } |
| 695 | } |
| 696 | |
| 697 | offs += dilation.x() * run_info.input_stride_y; |
| 698 | } |
| 699 | } |
| 700 | |
| 701 | weights_ptr += run_info.weights_stride_z; |
| 702 | input_offset += dilation.y() * run_info.input_stride_z; |
| 703 | } |
| 704 | |
| 705 | for(size_t m = 0, i = 0; m < depth_multiplier; m += vector_size, ++i) |
| 706 | { |
| 707 | if(has_biases) |
| 708 | { |
| 709 | const auto bias_val0 = wrapper::vloadq(reinterpret_cast<int32_t *>(biases_it.ptr() + m * sizeof(int32_t))); |
| 710 | const auto bias_val1 = wrapper::vloadq(reinterpret_cast<int32_t *>(biases_it.ptr() + (m + half_vec) * sizeof(int32_t))); |
| 711 | |
| 712 | acc0.at(i) = wrapper::vadd(acc0.at(i), bias_val0); |
| 713 | acc1.at(i) = wrapper::vadd(acc1.at(i), bias_val1); |
| 714 | } |
| 715 | |
| 716 | if(out_shift < 0) |
| 717 | { |
| 718 | acc0.at(i) = wrapper::vadd(saturating_doubling_high_mul(acc0.at(i) * (1 << (-out_shift)), out_mul), output_qoffset_vec); |
| 719 | acc1.at(i) = wrapper::vadd(saturating_doubling_high_mul(acc1.at(i) * (1 << (-out_shift)), out_mul), output_qoffset_vec); |
| 720 | } |
| 721 | else |
| 722 | { |
| 723 | acc0.at(i) = wrapper::vadd(rounding_divide_by_exp2(saturating_doubling_high_mul(acc0.at(i), out_mul), out_shift), output_qoffset_vec); |
| 724 | acc1.at(i) = wrapper::vadd(rounding_divide_by_exp2(saturating_doubling_high_mul(acc1.at(i), out_mul), out_shift), output_qoffset_vec); |
| 725 | } |
| 726 | |
| 727 | acc0.at(i) = wrapper::vmin(wrapper::vmax(acc0.at(i), lower), upper); |
| 728 | acc1.at(i) = wrapper::vmin(wrapper::vmax(acc1.at(i), lower), upper); |
| 729 | |
| 730 | const auto out_val = wrapper::vcombine(wrapper::vmovn(acc0.at(i)), |
| 731 | wrapper::vmovn(acc1.at(i))); |
| 732 | |
| 733 | if(std::is_same<T, uint8_t>::value) |
| 734 | { |
| 735 | wrapper::vstore(reinterpret_cast<uint8_t *>(output_it.ptr() + m * sizeof(uint8_t)), wrapper::vqmovn(vreinterpretq_u16_s16(out_val))); |
| 736 | } |
| 737 | else |
| 738 | { |
| 739 | wrapper::vstore(reinterpret_cast<int8_t *>(output_it.ptr() + m * sizeof(int8_t)), wrapper::vqmovn(out_val)); |
| 740 | } |
| 741 | } |
| 742 | }, |
| 743 | input_it, weights_it, biases_it, output_it); |
| 744 | } |
| 745 | } // namespace |
| 746 | template <typename T, typename TW> |
| 747 | void run_depthwise_float(const ITensor *src, const ITensor *weights, const ITensor *biases, |
| 748 | ITensor *dst, const Window &window, bool has_biases, const ConvolutionInfo &info) |
| 749 | { |
| 750 | PadStrideInfo conv_info = info.pad_stride_info; |
| 751 | unsigned int depth_multiplier = info.depth_multiplier; |
| 752 | Size2D dilation = info.dilation; |
| 753 | |
| 754 | if(depth_multiplier == 1) |
| 755 | { |
| 756 | depthwise_loop_multiplier1_fp<T>(src, weights, biases, dst, conv_info, dilation, window, has_biases); |
| 757 | } |
| 758 | else |
| 759 | { |
| 760 | depthwise_loop_generic_fp<T>(src, weights, biases, dst, conv_info, dilation, depth_multiplier, window, has_biases); |
| 761 | } |
| 762 | } |
| 763 | template void run_depthwise_float<float, float>(const ITensor *src, const ITensor *weights, const ITensor *biases, |
| 764 | ITensor *dst, const Window &window, bool has_biases, const ConvolutionInfo &info); |
| 765 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 766 | template void run_depthwise_float<float16_t, float16_t>(const ITensor *src, const ITensor *weights, const ITensor *biases, |
| 767 | ITensor *dst, const Window &window, bool has_biases, const ConvolutionInfo &info); |
| 768 | #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 769 | |
| 770 | template <typename T, typename TW> |
| 771 | void run_depthwise_quanitized8bit(const ITensor *src, const ITensor *weights, const ITensor *biases, |
| 772 | ITensor *dst, const Window &window, bool has_biases, const ConvolutionInfo &info) |
| 773 | { |
| 774 | PadStrideInfo conv_info = info.pad_stride_info; |
| 775 | unsigned int depth_multiplier = info.depth_multiplier; |
| 776 | Size2D dilation = info.dilation; |
| 777 | std::vector<int> output_multiplier; |
| 778 | std::vector<int> output_shift; |
| 779 | |
| 780 | const auto input_scale = src->info()->quantization_info().uniform().scale; |
| 781 | const auto output_scale = dst->info()->quantization_info().uniform().scale; |
| 782 | auto weights_scale = weights->info()->quantization_info().scale(); |
| 783 | |
| 784 | if(!is_data_type_quantized_per_channel(weights->info()->data_type())) |
| 785 | { |
| 786 | for(size_t i = 1; i < weights->info()->dimension(channel_idx); ++i) |
| 787 | { |
| 788 | weights_scale.push_back(weights_scale.front()); |
| 789 | } |
| 790 | } |
| 791 | |
| 792 | for(const auto &s : weights_scale) |
| 793 | { |
| 794 | int32_t out_mult = 0; |
| 795 | int32_t out_shift = 0; |
| 796 | const float multiplier = input_scale * s / output_scale; |
| 797 | arm_compute::quantization::calculate_quantized_multiplier(multiplier, &out_mult, &out_shift); |
| 798 | |
| 799 | output_multiplier.push_back(out_mult); |
| 800 | output_shift.push_back(out_shift); |
| 801 | } |
| 802 | |
| 803 | if(depth_multiplier == 1) |
| 804 | { |
| 805 | depthwise_loop_multiplier1_quantized<T, TW>(src, weights, biases, dst, conv_info, dilation, output_multiplier, output_shift, window, has_biases); |
| 806 | } |
| 807 | else |
| 808 | { |
| 809 | const bool is_pow2 = ((depth_multiplier & (depth_multiplier - 1)) == 0); |
| 810 | const bool is_quantized_per_tensor = !(is_data_type_quantized_per_channel(weights->info()->data_type())); |
| 811 | |
| 812 | if(is_pow2 && is_quantized_per_tensor && depth_multiplier >= 8) |
| 813 | { |
| 814 | depthwise_loop_pow2_quantized_per_tensor<T, TW>(src, weights, biases, dst, conv_info, dilation, depth_multiplier, output_multiplier, output_shift, window, has_biases); |
| 815 | } |
| 816 | else |
| 817 | { |
| 818 | depthwise_loop_generic_quantized<T, TW>(src, weights, biases, dst, conv_info, dilation, depth_multiplier, output_multiplier, output_shift, window, has_biases); |
| 819 | } |
| 820 | } |
| 821 | } |
| 822 | template void run_depthwise_quanitized8bit<uint8_t, uint8_t>(const ITensor *src, const ITensor *weights, const ITensor *biases, |
| 823 | ITensor *dst, const Window &window, bool has_biases, const ConvolutionInfo &info); |
| 824 | template void run_depthwise_quanitized8bit<int8_t, int8_t>(const ITensor *src, const ITensor *weights, const ITensor *biases, |
| 825 | ITensor *dst, const Window &window, bool has_biases, const ConvolutionInfo &info); |
| 826 | template void run_depthwise_quanitized8bit<uint8_t, int8_t>(const ITensor *src, const ITensor *weights, const ITensor *biases, |
| 827 | ITensor *dst, const Window &window, bool has_biases, const ConvolutionInfo &info); |
| 828 | } // namespace cpu |
| 829 | } // namespace arm_compute |