Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017-2018 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 "arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h" |
| 25 | |
| 26 | #include "arm_compute/core/NEON/kernels/arm32/NEGEMMAArch32Kernel.h" |
| 27 | #include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64Kernel.h" |
| 28 | #include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64NativeKernel.h" |
| 29 | #include "arm_compute/core/PixelValue.h" |
| 30 | #include "arm_compute/core/Size2D.h" |
| 31 | #include "arm_compute/core/Utils.h" |
| 32 | #include "arm_compute/core/Validate.h" |
| 33 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
| 34 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 35 | #include "support/ToolchainSupport.h" |
| 36 | |
| 37 | namespace arm_compute |
| 38 | { |
| 39 | #include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp" |
| 40 | #include "arm_compute/core/NEON/kernels/assembly/kernels/a32_sgemm_8x6.hpp" |
| 41 | #include "arm_compute/core/NEON/kernels/assembly/kernels/a64_sgemm_12x8.hpp" |
| 42 | } // namespace arm_compute |
| 43 | |
| 44 | #include <cmath> |
| 45 | #include <tuple> |
| 46 | |
| 47 | namespace |
| 48 | { |
| 49 | arm_compute::TensorShape get_reshaped_weights_shape(const arm_compute::ITensorInfo *weights, bool append_bias) |
| 50 | { |
| 51 | const unsigned int mat_weights_cols = weights->dimension(3); |
| 52 | const unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0); |
| 53 | return arm_compute::TensorShape(mat_weights_cols, mat_weights_rows); |
| 54 | } |
| 55 | } // namespace |
| 56 | |
| 57 | namespace arm_compute |
| 58 | { |
| 59 | NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager) |
| 60 | : _memory_group(std::move(memory_manager)), _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false) |
| 61 | { |
| 62 | } |
| 63 | |
| 64 | void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW) |
| 65 | { |
| 66 | // Perform validation step |
| 67 | ARM_COMPUTE_ERROR_ON_NULLPTR(weights, output); |
| 68 | ARM_COMPUTE_ERROR_THROW_ON(NEConvolutionLayerReshapeWeights::validate(weights->info(), |
| 69 | (biases != nullptr) ? biases->info() : nullptr, |
| 70 | output->info(), |
| 71 | transpose1xW)); |
| 72 | |
| 73 | // Check if bias are present, if yes they will be embedded to the weights matrix |
| 74 | const bool append_biases = (biases != nullptr) && !is_data_type_quantized_asymmetric(weights->info()->data_type()); |
| 75 | //const unsigned bias_element = (append_biases) ? 1 : 0; |
| 76 | const ITensor *biases_to_use = (append_biases) ? biases : nullptr; |
| 77 | |
| 78 | _transpose1xW = transpose1xW; |
| 79 | |
| 80 | if(transpose1xW) |
| 81 | { |
| 82 | // Create tensor to store the reshaped weights |
| 83 | TensorInfo info_wr = weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(get_reshaped_weights_shape(weights->info(), append_biases)); |
| 84 | |
| 85 | _weights_reshaped.allocator()->init(info_wr); |
| 86 | _memory_group.manage(&_weights_reshaped); |
| 87 | |
| 88 | _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped); |
| 89 | _weights_transposed_kernel.configure(&_weights_reshaped, output); |
| 90 | |
| 91 | _weights_reshaped.allocator()->allocate(); |
| 92 | } |
| 93 | else |
| 94 | { |
| 95 | _weights_reshape_kernel.configure(weights, biases_to_use, output); |
| 96 | } |
| 97 | |
| 98 | output->info()->set_quantization_info(weights->info()->quantization_info()); |
| 99 | } |
| 100 | |
| 101 | Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW) |
| 102 | { |
| 103 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); |
| 104 | ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); |
| 105 | if(!is_data_type_quantized_asymmetric(weights->data_type())) |
| 106 | { |
| 107 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output); |
| 108 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output); |
| 109 | } |
| 110 | // Check if bias are present, if yes they will be embedded to the weights matrix |
| 111 | const bool append_bias = (biases != nullptr); |
| 112 | |
| 113 | if(append_bias) |
| 114 | { |
| 115 | ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(weights->data_type())); |
| 116 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); |
| 117 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases); |
| 118 | ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3)); |
| 119 | ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); |
| 120 | } |
| 121 | |
| 122 | // Checks performed when biases are present |
| 123 | if(append_bias) |
| 124 | { |
| 125 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); |
| 126 | ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3)); |
| 127 | ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); |
| 128 | } |
| 129 | |
| 130 | if(transpose1xW) |
| 131 | { |
| 132 | TensorInfo weights_reshaped = weights->clone()->set_tensor_shape(get_reshaped_weights_shape(weights, append_bias)); |
| 133 | ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, &weights_reshaped)); |
| 134 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(&weights_reshaped, output)); |
| 135 | } |
| 136 | else |
| 137 | { |
| 138 | ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, output)); |
| 139 | } |
| 140 | |
| 141 | return Status{}; |
| 142 | } |
| 143 | |
| 144 | void NEConvolutionLayerReshapeWeights::run() |
| 145 | { |
| 146 | _memory_group.acquire(); |
| 147 | |
| 148 | NEScheduler::get().schedule(&_weights_reshape_kernel, 3); |
| 149 | |
| 150 | if(_transpose1xW) |
| 151 | { |
| 152 | NEScheduler::get().schedule(&_weights_transposed_kernel, Window::DimY); |
| 153 | } |
| 154 | |
| 155 | _memory_group.release(); |
| 156 | } |
| 157 | |
| 158 | namespace |
| 159 | { |
| 160 | TensorShape get_reshaped_weights_shape_conv(const ITensorInfo *weights, bool append_bias, bool is_fully_connected_convolution) |
| 161 | { |
| 162 | unsigned int mat_weights_cols = weights->dimension(3); |
| 163 | unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0); |
| 164 | |
| 165 | if(is_fully_connected_convolution) |
| 166 | { |
| 167 | // Create tensor to store the reshaped weights |
| 168 | return TensorShape(mat_weights_cols, mat_weights_rows); |
| 169 | } |
| 170 | else |
| 171 | { |
| 172 | // Create tensor to store transposed weights |
| 173 | const float transpose_width = 16.0f / weights->element_size(); |
| 174 | return TensorShape(mat_weights_rows * static_cast<unsigned int>(transpose_width), static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width))); |
| 175 | } |
| 176 | } |
| 177 | |
| 178 | Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, DataType &dt, |
| 179 | bool &append_bias, |
| 180 | bool &are_weights_reshaped, unsigned int &kernel_width, unsigned int &kernel_height, |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 181 | bool &is_fully_connected_convolution, bool &is_interleaved, bool &is_quantized, |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 182 | unsigned int &mat_weights_cols, unsigned int &mat_weights_rows, |
| 183 | unsigned int &conv_w, unsigned int &conv_h) |
| 184 | { |
| 185 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); |
| 186 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); |
| 187 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights); |
| 188 | ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && weights->dimension(2) != input->dimension(2)); |
| 189 | ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); |
| 190 | ARM_COMPUTE_RETURN_ERROR_ON(weights_info.are_reshaped() && is_data_type_quantized_asymmetric(input->data_type())); |
| 191 | |
| 192 | dt = input->data_type(); |
| 193 | is_quantized = is_data_type_quantized_asymmetric(dt); |
| 194 | |
| 195 | if(biases != nullptr) |
| 196 | { |
| 197 | if(is_quantized) |
| 198 | { |
| 199 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); |
| 200 | } |
| 201 | else |
| 202 | { |
| 203 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); |
| 204 | } |
| 205 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); |
| 206 | ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && biases->dimension(0) != weights->dimension(3)); |
| 207 | ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); |
| 208 | } |
| 209 | |
| 210 | append_bias = (biases != nullptr) && (!is_quantized); |
| 211 | are_weights_reshaped = weights_info.are_reshaped(); |
| 212 | kernel_width = (are_weights_reshaped) ? weights_info.kernel_size().first : weights->dimension(0); |
| 213 | kernel_height = (are_weights_reshaped) ? weights_info.kernel_size().second : weights->dimension(1); |
| 214 | mat_weights_cols = weights->dimension(3); |
| 215 | mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0); |
| 216 | |
| 217 | std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height, |
| 218 | conv_info); |
| 219 | |
| 220 | // Check if its a "fully connected" convolution |
| 221 | is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1)); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 222 | is_interleaved = (!is_fully_connected_convolution && !is_quantized); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 223 | |
| 224 | return Status{}; |
| 225 | } |
| 226 | } // namespace |
| 227 | |
| 228 | NEGEMMConvolutionLayer::NEGEMMConvolutionLayer(const std::shared_ptr<IMemoryManager> &memory_manager) |
| 229 | : _memory_group(memory_manager), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_optimised_kernel(nullptr), _mm_gemmlowp(memory_manager), |
| 230 | _gemmlowp_output_stage(), _output_col2im_kernel(), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _tmp_output(), _workspace(), _append_bias(false), |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 231 | _is_fully_connected_convolution(false), _are_weights_reshaped(false), _is_quantized(false), _is_interleaved(false) |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 232 | { |
| 233 | } |
| 234 | |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 235 | void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weights, ITensor *output, bool is_interleaved, const GEMMReshapeInfo &reshape_info) |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 236 | { |
| 237 | if(_is_quantized) |
| 238 | { |
| 239 | // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() |
| 240 | // Extract and negate input and weights offset |
| 241 | const QuantizationInfo input_quantization_info = input->info()->quantization_info(); |
| 242 | const QuantizationInfo weights_quantization_info = weights->info()->quantization_info(); |
| 243 | |
| 244 | input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset)); |
| 245 | weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset)); |
| 246 | |
| 247 | _mm_gemmlowp.configure(input, weights, output, GEMMInfo(false, false, true /* Reshape weights only for the first run*/)); |
| 248 | |
| 249 | // Revert back QuantizatioInfo as input and weights could be used in other convolution layers |
| 250 | input->info()->set_quantization_info(input_quantization_info); |
| 251 | weights->info()->set_quantization_info(weights_quantization_info); |
| 252 | } |
| 253 | else |
| 254 | { |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 255 | _mm_kernel.configure(input, weights, output, 1.f, is_interleaved, reshape_info); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 256 | } |
| 257 | } |
| 258 | |
| 259 | void NEGEMMConvolutionLayer::configure_asm_mm(const struct CPUInfo &ci, int M, int N, int K) |
| 260 | { |
| 261 | ARM_COMPUTE_UNUSED(ci); |
| 262 | ARM_COMPUTE_UNUSED(M); |
| 263 | ARM_COMPUTE_UNUSED(N); |
| 264 | ARM_COMPUTE_UNUSED(K); |
| 265 | #if defined(__arm__) || defined(__aarch64__) |
| 266 | #if defined(__arm__) |
| 267 | GemmInterleaved<sgemm_8x6, float, float> gemm(&ci, M, N, K, false, false); |
| 268 | #elif defined(__aarch64__) |
| 269 | GemmInterleaved<sgemm_12x8, float, float> gemm(&ci, M, N, K, false, false); |
| 270 | #endif /* defined(__arm__) || defined(__aarch64__) */ |
| 271 | |
| 272 | constexpr size_t alignment = 4096; |
| 273 | _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8)); |
| 274 | _memory_group.manage(&_workspace); |
| 275 | #endif /* defined(__arm__) || defined(__aarch64__) */ |
| 276 | } |
| 277 | |
| 278 | void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) |
| 279 | { |
| 280 | // Perform validate step |
| 281 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); |
| 282 | |
| 283 | DataType dt{}; |
| 284 | unsigned int kernel_width = 0; |
| 285 | unsigned int kernel_height = 0; |
| 286 | unsigned int mat_weights_cols = 0; |
| 287 | unsigned int mat_weights_rows = 0; |
| 288 | unsigned int conv_w = 0; |
| 289 | unsigned int conv_h = 0; |
| 290 | |
| 291 | Status status = validate_and_initialize_values(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), conv_info, weights_info, dt, _append_bias, _are_weights_reshaped, |
| 292 | kernel_width, kernel_height, |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 293 | _is_fully_connected_convolution, _is_interleaved, _is_quantized, |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 294 | mat_weights_cols, mat_weights_rows, conv_w, conv_h); |
| 295 | |
| 296 | ARM_COMPUTE_ERROR_THROW_ON(status); |
| 297 | |
| 298 | const unsigned int fixed_point_position = input->info()->fixed_point_position(); |
| 299 | const ITensor *biases_to_use = (_append_bias) ? biases : nullptr; |
| 300 | |
| 301 | #if defined(__arm__) |
| 302 | if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32) |
| 303 | { |
| 304 | _mm_optimised_kernel = support::cpp14::make_unique<NEGEMMAArch32Kernel>(); |
| 305 | } |
| 306 | #elif defined(__aarch64__) |
| 307 | if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32) |
| 308 | { |
| 309 | _mm_optimised_kernel = support::cpp14::make_unique<NEGEMMAArch64Kernel>(); |
| 310 | } |
| 311 | #endif /* defined(__arm__) || defined(__aarch64__) */ |
| 312 | |
| 313 | // Reshape weights if needed |
| 314 | if(_mm_optimised_kernel != nullptr) |
| 315 | { |
| 316 | if(_are_weights_reshaped) |
| 317 | { |
| 318 | mat_weights_cols = weights_info.num_kernels(); |
| 319 | mat_weights_rows = weights->info()->dimension(1); |
| 320 | } |
| 321 | else |
| 322 | { |
| 323 | TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows }; |
| 324 | |
| 325 | // Create tensor to store the reshaped weights |
| 326 | _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position)); |
| 327 | _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */); |
| 328 | weights = &_weights_reshaped; |
| 329 | } |
| 330 | } |
| 331 | else |
| 332 | { |
| 333 | if(_are_weights_reshaped) |
| 334 | { |
| 335 | if(_is_fully_connected_convolution || _is_quantized) |
| 336 | { |
| 337 | mat_weights_cols = weights_info.num_kernels(); |
| 338 | mat_weights_rows = weights->info()->dimension(1); |
| 339 | } |
| 340 | else |
| 341 | { |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 342 | mat_weights_cols = weights_info.num_kernels(); |
| 343 | mat_weights_rows = weights_info.kernel_size().first * weights_info.kernel_size().second * input->info()->dimension(2) + (_append_bias ? 1 : 0); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 344 | } |
| 345 | } |
| 346 | else |
| 347 | { |
| 348 | TensorShape reshaped_weights_shape; |
| 349 | |
| 350 | if(_is_fully_connected_convolution || _is_quantized) |
| 351 | { |
| 352 | reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows }; |
| 353 | } |
| 354 | else |
| 355 | { |
| 356 | // Create tensor to store transposed weights |
| 357 | const float transpose_width = 16.0f / input->info()->element_size(); |
| 358 | reshaped_weights_shape = TensorShape{ mat_weights_rows *static_cast<unsigned int>(transpose_width), |
| 359 | static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)) }; |
| 360 | } |
| 361 | |
| 362 | // Create tensor to store the reshaped weights |
| 363 | _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position)); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 364 | _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, _is_interleaved /* 1xW transpose */); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 365 | weights = &_weights_reshaped; |
| 366 | } |
| 367 | } |
| 368 | |
| 369 | // Create tensor to store im2col reshaped inputs |
| 370 | const unsigned int mat_input_cols = mat_weights_rows; |
| 371 | const unsigned int mat_input_rows = conv_w * conv_h; |
| 372 | |
| 373 | TensorShape shape_im2col(input->info()->tensor_shape()); |
| 374 | shape_im2col.set(0, mat_input_cols); |
| 375 | shape_im2col.set(1, mat_input_rows); |
| 376 | shape_im2col.set(2, 1); |
| 377 | _input_im2col_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col)); |
| 378 | _memory_group.manage(&_input_im2col_reshaped); |
| 379 | |
| 380 | // Create tensor (interleave) to prepare input tensor for GEMM |
| 381 | if(!_is_fully_connected_convolution && _mm_optimised_kernel == nullptr) |
| 382 | { |
| 383 | TensorShape shape_interleaved(shape_im2col); |
| 384 | shape_interleaved.set(0, shape_interleaved.x() * 4); |
| 385 | shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); |
| 386 | _input_interleaved_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_interleaved)); |
| 387 | _memory_group.manage(&_input_interleaved_reshaped); |
| 388 | } |
| 389 | |
| 390 | // Create GEMM output tensor |
| 391 | TensorShape shape_gemm(_input_im2col_reshaped.info()->tensor_shape()); |
| 392 | shape_gemm.set(0, mat_weights_cols); |
| 393 | shape_gemm.set(1, mat_input_rows); |
| 394 | const DataType gemm_data_type = _is_quantized ? DataType::S32 : dt; |
| 395 | // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input. |
| 396 | TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position()); |
| 397 | info_gemm.set_quantization_info(output->info()->quantization_info()); |
| 398 | _gemm_output.allocator()->init(info_gemm); |
| 399 | _memory_group.manage(&_gemm_output); |
| 400 | |
| 401 | // Configure kernels |
| 402 | // Configure im2col |
| 403 | _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias); |
| 404 | |
| 405 | // Configure matrix multiply |
| 406 | if(_mm_optimised_kernel != nullptr) |
| 407 | { |
| 408 | struct CPUInfo ci = NEScheduler::get().cpu_info(); |
| 409 | |
| 410 | const int M = _gemm_output.info()->tensor_shape().y(); |
| 411 | const int N = _gemm_output.info()->tensor_shape().x(); |
| 412 | const int K = _input_im2col_reshaped.info()->tensor_shape().x(); |
| 413 | |
| 414 | #if defined(__aarch64__) |
| 415 | if((N <= 128) && (K <= 128)) |
| 416 | { |
| 417 | _mm_optimised_kernel = support::cpp14::make_unique<NEGEMMAArch64NativeKernel>(); |
| 418 | } |
| 419 | else |
| 420 | #endif /* defined(__aarch64__) */ |
| 421 | { |
| 422 | configure_asm_mm(ci, M, N, K); |
| 423 | } |
| 424 | |
| 425 | // Configure matrix multiplication kernel |
| 426 | _mm_optimised_kernel->configure(&_input_im2col_reshaped, weights, &_gemm_output, &_workspace); |
| 427 | |
| 428 | _workspace.allocator()->allocate(); |
| 429 | } |
| 430 | else |
| 431 | { |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 432 | if(_is_interleaved) |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 433 | { |
| 434 | // Configure GEMMInterleave4x4. _input_interleaved_reshaped will be auto configured in the kernel |
| 435 | _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped); |
| 436 | |
| 437 | // Configure GEMM |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 438 | configure_mm(&_input_interleaved_reshaped, weights, &_gemm_output, _is_interleaved, GEMMReshapeInfo(_input_im2col_reshaped.info()->dimension(1), 0 /* no transpose */, |
| 439 | _input_im2col_reshaped.info()->dimension(0))); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 440 | _input_interleaved_reshaped.allocator()->allocate(); |
| 441 | } |
| 442 | else |
| 443 | { |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 444 | configure_mm(&_input_im2col_reshaped, weights, &_gemm_output, _is_interleaved); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 445 | } |
| 446 | } |
| 447 | |
| 448 | _input_im2col_reshaped.allocator()->allocate(); |
| 449 | |
| 450 | // Configure output stage for quantized case |
| 451 | if(_is_quantized) |
| 452 | { |
| 453 | const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info(); |
| 454 | |
| 455 | float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale; |
| 456 | int output_multiplier, output_shift; |
| 457 | quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); |
| 458 | _memory_group.manage(&_tmp_output); |
| 459 | _gemmlowp_output_stage.configure(&_gemm_output, biases, &_tmp_output, output_multiplier, output_shift, output_quant_info.offset); |
| 460 | } |
| 461 | |
| 462 | // Configure Col2Im |
| 463 | _output_col2im_kernel.configure(_is_quantized ? &_tmp_output : &_gemm_output, output, Size2D(conv_w, conv_h)); |
| 464 | if(_is_quantized) |
| 465 | { |
| 466 | _tmp_output.allocator()->allocate(); |
| 467 | } |
| 468 | _gemm_output.allocator()->allocate(); |
| 469 | |
| 470 | ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one"); |
| 471 | |
| 472 | // Allocate intermediate tensor |
| 473 | if(!_are_weights_reshaped) |
| 474 | { |
| 475 | _weights_reshaped.allocator()->allocate(); |
| 476 | } |
| 477 | } |
| 478 | |
| 479 | Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, |
| 480 | const WeightsInfo &weights_info) |
| 481 | { |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 482 | ARM_COMPUTE_UNUSED(output); |
| 483 | |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 484 | DataType dt{}; |
| 485 | bool append_bias{}; |
| 486 | bool are_weights_reshaped{}; |
| 487 | bool is_fully_connected_convolution{}; |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 488 | bool is_interleaved{}; |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 489 | bool is_quantized{}; |
| 490 | unsigned int kernel_width = 0; |
| 491 | unsigned int kernel_height = 0; |
| 492 | unsigned int mat_weights_cols = 0; |
| 493 | unsigned int mat_weights_rows = 0; |
| 494 | unsigned int conv_w = 0; |
| 495 | unsigned int conv_h = 0; |
| 496 | |
| 497 | Status status = validate_and_initialize_values(input, weights, biases, conv_info, weights_info, dt, append_bias, are_weights_reshaped, kernel_width, kernel_height, |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 498 | is_fully_connected_convolution, is_interleaved, is_quantized, mat_weights_cols, mat_weights_rows, |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 499 | conv_w, conv_h); |
| 500 | |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 501 | const Size2D kernel_weights = Size2D(kernel_width, kernel_height); |
| 502 | |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 503 | ARM_COMPUTE_RETURN_ON_ERROR(status); |
| 504 | |
| 505 | std::unique_ptr<ITensorInfo> reshaped_weights = weights->clone(); |
| 506 | bool optimised_kernel = false; |
| 507 | |
| 508 | #if defined(__arm__) |
| 509 | if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32) |
| 510 | { |
| 511 | optimised_kernel = true; |
| 512 | } |
| 513 | #elif defined(__aarch64__) |
| 514 | if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32) |
| 515 | { |
| 516 | optimised_kernel = true; |
| 517 | } |
| 518 | #endif /* defined(__arm__) || defined(__aarch64__) */ |
| 519 | |
| 520 | // Reshape weights if needed |
| 521 | if(optimised_kernel) |
| 522 | { |
| 523 | if(are_weights_reshaped) |
| 524 | { |
| 525 | mat_weights_cols = weights_info.num_kernels(); |
| 526 | mat_weights_rows = weights->dimension(1); |
| 527 | } |
| 528 | else |
| 529 | { |
| 530 | TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows }; |
| 531 | |
| 532 | // Create tensor to store the reshaped weights |
| 533 | reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution)); |
| 534 | ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */)); |
| 535 | weights = reshaped_weights.get(); |
| 536 | } |
| 537 | } |
| 538 | else |
| 539 | { |
| 540 | if(are_weights_reshaped) |
| 541 | { |
| 542 | const unsigned int transpose_width = 16 / input->element_size(); |
| 543 | mat_weights_cols = weights_info.num_kernels(); |
| 544 | mat_weights_rows = weights->dimension(0) / transpose_width + (append_bias ? 1 : 0); |
| 545 | } |
| 546 | else |
| 547 | { |
| 548 | TensorShape reshaped_weights_shape; |
| 549 | |
| 550 | if(is_fully_connected_convolution || is_quantized) |
| 551 | { |
| 552 | reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows }; |
| 553 | } |
| 554 | else |
| 555 | { |
| 556 | // Create tensor to store transposed weights |
| 557 | const float transpose_width = 16.0f / input->element_size(); |
| 558 | reshaped_weights_shape = TensorShape{ mat_weights_rows *static_cast<unsigned int>(transpose_width), |
| 559 | static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)) }; |
| 560 | } |
| 561 | |
| 562 | // Create tensor to store the reshaped weights |
| 563 | reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution)); |
| 564 | ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */)); |
| 565 | weights = reshaped_weights.get(); |
| 566 | } |
| 567 | } |
| 568 | |
| 569 | // Validate im2col |
| 570 | const unsigned int mat_input_cols = mat_weights_rows; |
| 571 | const unsigned int mat_input_rows = conv_w * conv_h; |
| 572 | TensorShape shape_im2col = input->tensor_shape(); |
| 573 | shape_im2col.set(0, mat_input_cols); |
| 574 | shape_im2col.set(1, mat_input_rows); |
| 575 | shape_im2col.set(2, 1); |
| 576 | TensorInfo im2_col_info = input->clone()->set_tensor_shape(shape_im2col); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 577 | ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2_col_info, kernel_weights, conv_info, append_bias, false)); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 578 | |
| 579 | // Create GEMM output tensor |
| 580 | TensorShape shape_gemm(im2_col_info.tensor_shape()); |
| 581 | shape_gemm.set(0, mat_weights_cols); |
| 582 | shape_gemm.set(1, mat_input_rows); |
| 583 | TensorInfo gemm_output_info = input->clone()->set_tensor_shape(shape_gemm); |
| 584 | |
| 585 | // Validate GEMM interleave and multiply |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 586 | if(is_interleaved) |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 587 | { |
| 588 | TensorShape shape_interleaved = shape_im2col; |
| 589 | shape_interleaved.set(0, shape_interleaved.x() * 4); |
| 590 | shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); |
| 591 | TensorInfo input_interleaved_info = input->clone()->set_tensor_shape(shape_interleaved); |
| 592 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(&im2_col_info, &input_interleaved_info)); |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 593 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&input_interleaved_info, weights, &gemm_output_info, 1.f, is_interleaved, GEMMReshapeInfo())); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 594 | } |
| 595 | else |
| 596 | { |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 597 | ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&im2_col_info, weights, &gemm_output_info, 1.f, is_interleaved, GEMMReshapeInfo())); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 598 | } |
| 599 | |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 600 | return Status{}; |
| 601 | } |
| 602 | |
| 603 | void NEGEMMConvolutionLayer::run() |
| 604 | { |
| 605 | // Run weights reshaping (Runs once for every configure) |
| 606 | if(!_are_weights_reshaped) |
| 607 | { |
| 608 | _are_weights_reshaped = true; |
| 609 | _reshape_weights.run(); |
| 610 | } |
| 611 | |
| 612 | _memory_group.acquire(); |
| 613 | |
| 614 | // Run input reshaping |
| 615 | NEScheduler::get().schedule(&_input_im2col_kernel, Window::DimY); |
| 616 | |
| 617 | // Runs matrix multiply on reshaped matrices |
| 618 | if(_mm_optimised_kernel != nullptr) |
| 619 | { |
| 620 | NEScheduler::get().schedule(_mm_optimised_kernel.get(), Window::DimY); |
| 621 | } |
| 622 | else |
| 623 | { |
Ioan-Cristian Szabo | b4e3e1c | 2017-11-30 17:17:17 +0000 | [diff] [blame^] | 624 | if(_is_interleaved) |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 625 | { |
| 626 | // Run interleave |
| 627 | NEScheduler::get().schedule(&_input_interleave_kernel, Window::DimY); |
| 628 | } |
| 629 | |
| 630 | // Runs matrix multiply on reshaped matrices |
| 631 | if(_is_quantized) |
| 632 | { |
| 633 | _mm_gemmlowp.run(); |
| 634 | } |
| 635 | else |
| 636 | { |
| 637 | NEScheduler::get().schedule(&_mm_kernel, Window::DimY); |
| 638 | } |
| 639 | } |
| 640 | |
| 641 | // Run output stage for quantized case |
| 642 | if(_is_quantized) |
| 643 | { |
| 644 | _gemmlowp_output_stage.run(); |
| 645 | } |
| 646 | |
| 647 | // Reshape output matrix |
| 648 | NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY); |
| 649 | |
| 650 | _memory_group.release(); |
| 651 | } |
| 652 | } // namespace arm_compute |