Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2019-2020 Arm Limited. |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | |
| 25 | #include "arm_compute/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.h" |
| 26 | |
| 27 | #include "arm_compute/core/CPP/Validate.h" |
| 28 | #include "arm_compute/core/ITensor.h" |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 29 | #include "arm_compute/core/NEON/kernels/assembly/NEDepthwiseConvolutionAssemblyKernelWrapper.h" |
| 30 | #include "arm_compute/core/NEON/kernels/convolution/depthwise/depthwise_dilated.hpp" |
| 31 | #include "arm_compute/core/NEON/kernels/convolution/depthwise/depthwise_quantized_dilated.hpp" |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 32 | #include "arm_compute/core/Utils.h" |
| 33 | #include "arm_compute/core/utils/misc/InfoHelpers.h" |
| 34 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 35 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
| 36 | |
| 37 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 38 | |
Georgios Pinitas | 4c75851 | 2019-07-10 19:49:11 +0100 | [diff] [blame] | 39 | #include <set> |
| 40 | |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 41 | namespace arm_compute |
| 42 | { |
| 43 | namespace |
| 44 | { |
Georgios Pinitas | 4c75851 | 2019-07-10 19:49:11 +0100 | [diff] [blame] | 45 | std::unique_ptr<depthwise::IDepthwiseConvolution> get_qasymm8_convolver(int kernel_size, int stride_x, |
| 46 | int n_batches, int in_rows, int in_cols, int n_channels, |
| 47 | int dilation_factor, neon_convolution_kernels::ActivationFunction activation, |
| 48 | const qasymm8::QAsymm8Params &wqinfo, const qasymm8::QAsymm8Params &iqinfo, const qasymm8::QAsymm8Params &oqinfo, |
| 49 | const qasymm8::QAsymm8RescaleParams &rescale_params, |
| 50 | int padding_top, int padding_left, int padding_bottom, int padding_right) |
| 51 | { |
| 52 | switch(kernel_size) |
| 53 | { |
| 54 | case 3: |
| 55 | { |
| 56 | switch(stride_x) |
| 57 | { |
| 58 | case 1: |
| 59 | return arm_compute::support::cpp14::make_unique<depthwise::QAsymm8DilatedDepthwiseConvolution<2, 2, 3, 3, 1, 1>>( |
| 60 | n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, wqinfo, iqinfo, oqinfo, rescale_params, padding_top, padding_left, padding_bottom, padding_right); |
| 61 | case 2: |
| 62 | return arm_compute::support::cpp14::make_unique<depthwise::QAsymm8DilatedDepthwiseConvolution<2, 2, 3, 3, 2, 2>>( |
| 63 | n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, wqinfo, iqinfo, oqinfo, rescale_params, padding_top, padding_left, padding_bottom, padding_right); |
| 64 | default: |
| 65 | return nullptr; |
| 66 | } |
| 67 | } |
| 68 | case 5: |
| 69 | { |
| 70 | switch(stride_x) |
| 71 | { |
| 72 | case 1: |
| 73 | return arm_compute::support::cpp14::make_unique<depthwise::QAsymm8DilatedDepthwiseConvolution<2, 2, 5, 5, 1, 1>>( |
| 74 | n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, wqinfo, iqinfo, oqinfo, rescale_params, padding_top, padding_left, padding_bottom, padding_right); |
| 75 | case 2: |
| 76 | return arm_compute::support::cpp14::make_unique<depthwise::QAsymm8DilatedDepthwiseConvolution<2, 2, 5, 5, 2, 2>>( |
| 77 | n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, wqinfo, iqinfo, oqinfo, rescale_params, padding_top, padding_left, padding_bottom, padding_right); |
| 78 | default: |
| 79 | return nullptr; |
| 80 | } |
| 81 | } |
| 82 | default: |
| 83 | return nullptr; |
| 84 | } |
| 85 | } |
| 86 | |
Giuseppe Rossini | f01201a | 2019-11-06 14:57:49 +0000 | [diff] [blame] | 87 | std::unique_ptr<depthwise::IDepthwiseConvolution> get_qsymm8_perchannel_convolver(int kernel_size, int stride_x, |
| 88 | int n_batches, int in_rows, int in_cols, int n_channels, |
| 89 | neon_convolution_kernels::ActivationFunction activation, |
| 90 | const qsymm8::QSymm8PerChannelParams &wqinfo, const qasymm8::QAsymm8Params &iqinfo, const qasymm8::QAsymm8Params &oqinfo, |
| 91 | const qsymm8::QSymm8PerChannelRescaleParams &rescale_params, |
| 92 | int padding_top, int padding_left, int padding_bottom, int padding_right) |
| 93 | { |
| 94 | switch(kernel_size) |
| 95 | { |
| 96 | case 3: |
| 97 | { |
| 98 | switch(stride_x) |
| 99 | { |
| 100 | case 1: |
| 101 | return arm_compute::support::cpp14::make_unique<depthwise::QSymm8HybridPerChannelDepthwiseConvolution<2, 2, 3, 3, 1, 1>>( |
| 102 | n_batches, in_rows, in_cols, n_channels, activation, wqinfo, iqinfo, oqinfo, rescale_params, padding_top, padding_left, padding_bottom, padding_right); |
| 103 | case 2: |
| 104 | return arm_compute::support::cpp14::make_unique<depthwise::QSymm8HybridPerChannelDepthwiseConvolution<2, 2, 3, 3, 2, 2>>( |
| 105 | n_batches, in_rows, in_cols, n_channels, activation, wqinfo, iqinfo, oqinfo, rescale_params, padding_top, padding_left, padding_bottom, padding_right); |
| 106 | default: |
| 107 | return nullptr; |
| 108 | } |
| 109 | } |
| 110 | case 5: |
| 111 | { |
| 112 | switch(stride_x) |
| 113 | { |
| 114 | case 1: |
| 115 | return arm_compute::support::cpp14::make_unique<depthwise::QSymm8HybridPerChannelDepthwiseConvolution<2, 2, 5, 5, 1, 1>>( |
| 116 | n_batches, in_rows, in_cols, n_channels, activation, wqinfo, iqinfo, oqinfo, rescale_params, padding_top, padding_left, padding_bottom, padding_right); |
| 117 | case 2: |
| 118 | return arm_compute::support::cpp14::make_unique<depthwise::QSymm8HybridPerChannelDepthwiseConvolution<2, 2, 5, 5, 2, 2>>( |
| 119 | n_batches, in_rows, in_cols, n_channels, activation, wqinfo, iqinfo, oqinfo, rescale_params, padding_top, padding_left, padding_bottom, padding_right); |
| 120 | default: |
| 121 | return nullptr; |
| 122 | } |
| 123 | } |
| 124 | default: |
| 125 | return nullptr; |
| 126 | } |
| 127 | } |
| 128 | |
Georgios Pinitas | 4c75851 | 2019-07-10 19:49:11 +0100 | [diff] [blame] | 129 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 130 | std::unique_ptr<depthwise::IDepthwiseConvolution> get_fp16_convolver(int kernel_size, int stride_x, |
| 131 | int n_batches, int in_rows, int in_cols, int n_channels, |
| 132 | int dilation_factor, neon_convolution_kernels::ActivationFunction activation, |
| 133 | int padding_top, int padding_left, int padding_bottom, int padding_right) |
| 134 | { |
| 135 | switch(kernel_size) |
| 136 | { |
| 137 | case 3: |
| 138 | { |
| 139 | switch(stride_x) |
| 140 | { |
| 141 | case 1: |
| 142 | return arm_compute::support::cpp14::make_unique<depthwise::DilatedDepthwiseConvolution<3, 3, 3, 3, 1, 1, float16_t, float16_t, float16_t>>( |
| 143 | n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, padding_top, padding_left, padding_bottom, padding_right); |
| 144 | case 2: |
| 145 | return arm_compute::support::cpp14::make_unique<depthwise::DilatedDepthwiseConvolution<3, 3, 3, 3, 2, 2, float16_t, float16_t, float16_t>>( |
| 146 | n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, padding_top, padding_left, padding_bottom, padding_right); |
| 147 | default: |
| 148 | return nullptr; |
| 149 | } |
| 150 | } |
| 151 | case 5: |
| 152 | { |
| 153 | switch(stride_x) |
| 154 | { |
| 155 | case 1: |
| 156 | return arm_compute::support::cpp14::make_unique<depthwise::DilatedDepthwiseConvolution<3, 3, 5, 5, 1, 1, float16_t, float16_t, float16_t>>( |
| 157 | n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, padding_top, padding_left, padding_bottom, padding_right); |
| 158 | case 2: |
| 159 | return arm_compute::support::cpp14::make_unique<depthwise::DilatedDepthwiseConvolution<3, 3, 5, 5, 2, 2, float16_t, float16_t, float16_t>>( |
| 160 | n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, padding_top, padding_left, padding_bottom, padding_right); |
| 161 | default: |
| 162 | return nullptr; |
| 163 | } |
| 164 | } |
| 165 | default: |
| 166 | return nullptr; |
| 167 | } |
| 168 | } |
| 169 | #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 170 | |
| 171 | std::unique_ptr<depthwise::IDepthwiseConvolution> get_fp32_convolver(int kernel_size, int stride_x, |
| 172 | int n_batches, int in_rows, int in_cols, int n_channels, |
| 173 | int dilation_factor, neon_convolution_kernels::ActivationFunction activation, |
| 174 | int padding_top, int padding_left, int padding_bottom, int padding_right) |
| 175 | { |
| 176 | switch(kernel_size) |
| 177 | { |
| 178 | case 3: |
| 179 | { |
| 180 | switch(stride_x) |
| 181 | { |
| 182 | case 1: |
| 183 | return arm_compute::support::cpp14::make_unique<depthwise::DilatedDepthwiseConvolution<4, 4, 3, 3, 1, 1, float, float, float>>( |
| 184 | n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, padding_top, padding_left, padding_bottom, padding_right); |
| 185 | case 2: |
| 186 | return arm_compute::support::cpp14::make_unique<depthwise::DilatedDepthwiseConvolution<3, 3, 3, 3, 2, 2, float, float, float>>( |
| 187 | n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, padding_top, padding_left, padding_bottom, padding_right); |
| 188 | default: |
| 189 | return nullptr; |
| 190 | } |
| 191 | } |
| 192 | case 5: |
| 193 | { |
| 194 | switch(stride_x) |
| 195 | { |
| 196 | case 1: |
| 197 | return arm_compute::support::cpp14::make_unique<depthwise::DilatedDepthwiseConvolution<4, 4, 5, 5, 1, 1, float, float, float>>( |
| 198 | n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, padding_top, padding_left, padding_bottom, padding_right); |
| 199 | case 2: |
| 200 | return arm_compute::support::cpp14::make_unique<depthwise::DilatedDepthwiseConvolution<3, 3, 5, 5, 2, 2, float, float, float>>( |
| 201 | n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, padding_top, padding_left, padding_bottom, padding_right); |
| 202 | default: |
| 203 | return nullptr; |
| 204 | } |
| 205 | } |
| 206 | default: |
| 207 | return nullptr; |
| 208 | } |
| 209 | } |
| 210 | |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 211 | std::unique_ptr<depthwise::IDepthwiseConvolution> create_convolver(const ITensor *input, |
| 212 | const ITensor *weights, |
| 213 | ITensor *output, |
| 214 | PadStrideInfo conv_info, |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 215 | ActivationLayerInfo act_info, |
| 216 | const Size2D &dilation) |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 217 | { |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 218 | ARM_COMPUTE_UNUSED(dilation); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 219 | const DataType data_type = input->info()->data_type(); |
| 220 | const TensorShape shape = input->info()->tensor_shape(); |
| 221 | |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 222 | const int n_batches = shape[3]; |
| 223 | const int in_rows = shape.z(); |
| 224 | const int in_cols = shape.y(); |
| 225 | const int n_channels = shape.x(); |
| 226 | const int dilation_factor = dilation.x(); |
| 227 | const int padding_top = conv_info.pad_top(); |
| 228 | const int padding_left = conv_info.pad_left(); |
| 229 | const int padding_bottom = conv_info.pad_bottom(); |
| 230 | const int padding_right = conv_info.pad_right(); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 231 | |
Giuseppe Rossini | f01201a | 2019-11-06 14:57:49 +0000 | [diff] [blame] | 232 | const bool is_uniform_quantized = (data_type == DataType::QASYMM8) && (weights->info()->data_type() == DataType::QASYMM8); |
| 233 | const bool is_perchannel_quantized = (data_type == DataType::QASYMM8) && (weights->info()->data_type() == DataType::QSYMM8_PER_CHANNEL); |
| 234 | |
Georgios Pinitas | 4c75851 | 2019-07-10 19:49:11 +0100 | [diff] [blame] | 235 | const unsigned int stride_x = conv_info.stride().first; |
| 236 | const unsigned int kernel_size = weights->info()->tensor_shape().y(); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 237 | |
| 238 | // Map activation function |
| 239 | neon_convolution_kernels::ActivationFunction activation = neon_convolution_kernels::ActivationFunction::None; |
| 240 | if(arm_compute::utils::info_helpers::is_relu(act_info)) |
| 241 | { |
| 242 | activation = neon_convolution_kernels::ActivationFunction::ReLU; |
| 243 | } |
| 244 | else if(arm_compute::utils::info_helpers::is_relu6(act_info)) |
| 245 | { |
| 246 | activation = neon_convolution_kernels::ActivationFunction::ReLU6; |
| 247 | } |
| 248 | |
| 249 | // Create quantized convolver |
Giuseppe Rossini | f01201a | 2019-11-06 14:57:49 +0000 | [diff] [blame] | 250 | if(is_uniform_quantized) |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 251 | { |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 252 | const UniformQuantizationInfo input_qinfo = input->info()->quantization_info().uniform(); |
| 253 | const UniformQuantizationInfo weights_qinfo = weights->info()->quantization_info().uniform(); |
| 254 | const UniformQuantizationInfo output_qinfo = output->info()->quantization_info().uniform(); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 255 | |
| 256 | // Check that quantization info are in the range [0, 255] |
| 257 | ARM_COMPUTE_ERROR_ON(input_qinfo.offset < 0 || input_qinfo.offset > 255); |
| 258 | ARM_COMPUTE_ERROR_ON(weights_qinfo.offset < 0 || weights_qinfo.offset > 255); |
| 259 | ARM_COMPUTE_ERROR_ON(output_qinfo.offset < 0 || output_qinfo.offset > 255); |
| 260 | const qasymm8::QAsymm8Params iqinfo{ static_cast<uint8_t>(input_qinfo.offset), input_qinfo.scale }; |
| 261 | const qasymm8::QAsymm8Params wqinfo{ static_cast<uint8_t>(weights_qinfo.offset), weights_qinfo.scale }; |
| 262 | const qasymm8::QAsymm8Params oqinfo{ static_cast<uint8_t>(output_qinfo.offset), output_qinfo.scale }; |
| 263 | |
| 264 | // Calculate rescale parameters |
| 265 | const float fmultipler = iqinfo.scale * wqinfo.scale / oqinfo.scale; |
Michalis Spyrou | e7be8a0 | 2019-12-12 16:16:09 +0000 | [diff] [blame] | 266 | int32_t qmultiplier = 0; |
| 267 | int32_t qshift = 0; |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 268 | quantization::calculate_quantized_multiplier_less_than_one(fmultipler, &qmultiplier, &qshift); |
| 269 | qasymm8::QAsymm8RescaleParams rescale_params(qshift, qmultiplier, fmultipler); |
| 270 | |
Georgios Pinitas | 4c75851 | 2019-07-10 19:49:11 +0100 | [diff] [blame] | 271 | return get_qasymm8_convolver(kernel_size, stride_x, n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, |
| 272 | wqinfo, iqinfo, oqinfo, rescale_params, padding_top, padding_left, padding_bottom, padding_right); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 273 | } |
Giuseppe Rossini | f01201a | 2019-11-06 14:57:49 +0000 | [diff] [blame] | 274 | else if(is_perchannel_quantized) |
| 275 | { |
| 276 | const UniformQuantizationInfo input_qinfo = input->info()->quantization_info().uniform(); |
| 277 | const QuantizationInfo weights_qinfo = weights->info()->quantization_info(); |
| 278 | const UniformQuantizationInfo output_qinfo = output->info()->quantization_info().uniform(); |
| 279 | |
| 280 | // Check that quantization info are in the range [0, 255] |
| 281 | ARM_COMPUTE_ERROR_ON(input_qinfo.offset < 0 || input_qinfo.offset > 255); |
| 282 | ARM_COMPUTE_ERROR_ON(output_qinfo.offset < 0 || output_qinfo.offset > 255); |
| 283 | const qasymm8::QAsymm8Params iqinfo{ static_cast<uint8_t>(input_qinfo.offset), input_qinfo.scale }; |
| 284 | const qsymm8::QSymm8PerChannelParams wqinfo{ weights_qinfo.scale() }; |
| 285 | const qasymm8::QAsymm8Params oqinfo{ static_cast<uint8_t>(output_qinfo.offset), output_qinfo.scale }; |
| 286 | |
| 287 | // Calculate rescale parameters |
Michalis Spyrou | e7be8a0 | 2019-12-12 16:16:09 +0000 | [diff] [blame] | 288 | std::vector<float> fmultipliers; |
| 289 | std::vector<int32_t> qmultipliers; |
| 290 | std::vector<int32_t> qshifts; |
Giuseppe Rossini | f01201a | 2019-11-06 14:57:49 +0000 | [diff] [blame] | 291 | |
| 292 | for(auto const s : wqinfo.scales) |
| 293 | { |
| 294 | const float fmultipler = iqinfo.scale * s / oqinfo.scale; |
Michalis Spyrou | e7be8a0 | 2019-12-12 16:16:09 +0000 | [diff] [blame] | 295 | int32_t qmultiplier = 0; |
| 296 | int32_t qshift = 0; |
Giuseppe Rossini | f01201a | 2019-11-06 14:57:49 +0000 | [diff] [blame] | 297 | quantization::calculate_quantized_multiplier_less_than_one(fmultipler, &qmultiplier, &qshift); |
| 298 | fmultipliers.push_back(fmultipler); |
| 299 | qmultipliers.push_back(qmultiplier); |
| 300 | qshifts.push_back(qshift); |
| 301 | } |
| 302 | |
| 303 | qsymm8::QSymm8PerChannelRescaleParams rescale_params(qshifts, qmultipliers, fmultipliers); |
| 304 | |
| 305 | return get_qsymm8_perchannel_convolver(kernel_size, stride_x, n_batches, in_rows, in_cols, n_channels, activation, |
| 306 | wqinfo, iqinfo, oqinfo, rescale_params, padding_top, padding_left, padding_bottom, padding_right); |
| 307 | } |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 308 | else |
| 309 | { |
| 310 | // Create float convolver |
| 311 | switch(data_type) |
| 312 | { |
| 313 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 314 | case DataType::F16: |
| 315 | { |
Georgios Pinitas | 4c75851 | 2019-07-10 19:49:11 +0100 | [diff] [blame] | 316 | return get_fp16_convolver(kernel_size, stride_x, n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, padding_top, padding_left, padding_bottom, padding_right); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 317 | } |
| 318 | #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 319 | case DataType::F32: |
| 320 | { |
Georgios Pinitas | 4c75851 | 2019-07-10 19:49:11 +0100 | [diff] [blame] | 321 | return get_fp32_convolver(kernel_size, stride_x, n_batches, in_rows, in_cols, n_channels, dilation_factor, activation, padding_top, padding_left, padding_bottom, padding_right); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 322 | } |
| 323 | default: |
| 324 | return nullptr; |
| 325 | } |
| 326 | } |
| 327 | } |
| 328 | } // namespace |
| 329 | |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 330 | struct NEDepthwiseConvolutionAssemblyDispatch::LocalImpl |
| 331 | { |
| 332 | std::unique_ptr<depthwise::IDepthwiseConvolution> _dwc_assembly_kernel{ nullptr }; |
| 333 | NEDepthwiseConvolutionAssemblyKernelWrapper _dwc_acl_kernel{}; |
| 334 | }; |
| 335 | |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 336 | #ifndef DOXYGEN_SKIP_THIS |
| 337 | NEDepthwiseConvolutionAssemblyDispatch::NEDepthwiseConvolutionAssemblyDispatch(std::shared_ptr<arm_compute::IMemoryManager> memory_manager) |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 338 | : _memory_group(std::move(memory_manager)), _input(nullptr), _weights(nullptr), _bias(nullptr), _output(nullptr), _packed_weights(), _workspace(), _is_prepared(false), |
| 339 | _pImpl(support::cpp14::make_unique<LocalImpl>()) |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 340 | { |
| 341 | } |
| 342 | #endif /* DOXYGEN_SKIP_THIS */ |
| 343 | |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 344 | NEDepthwiseConvolutionAssemblyDispatch::~NEDepthwiseConvolutionAssemblyDispatch() = default; |
| 345 | |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 346 | void NEDepthwiseConvolutionAssemblyDispatch::configure(const ITensor *input, |
| 347 | const ITensor *weights, |
| 348 | const ITensor *bias, |
| 349 | ITensor *output, |
| 350 | const PadStrideInfo &conv_info, |
| 351 | unsigned int depth_multiplier, |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 352 | const ActivationLayerInfo &act_info, |
| 353 | const Size2D &dilation) |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 354 | { |
| 355 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); |
| 356 | ARM_COMPUTE_UNUSED(depth_multiplier); |
| 357 | ARM_COMPUTE_ERROR_THROW_ON(NEDepthwiseConvolutionAssemblyDispatch::validate(input->info(), |
| 358 | weights->info(), |
| 359 | bias != nullptr ? bias->info() : nullptr, |
| 360 | output->info(), |
| 361 | conv_info, |
| 362 | depth_multiplier, |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 363 | act_info, |
| 364 | dilation)); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 365 | |
| 366 | // Output auto inizialitation if not yet initialized |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 367 | const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info, depth_multiplier, dilation); |
Pablo Tello | a28aebc | 2019-06-03 14:59:48 +0100 | [diff] [blame] | 368 | auto_init_if_empty(*output->info(), input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape).set_quantization_info(output->info()->quantization_info())); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 369 | |
| 370 | _input = input; |
| 371 | _weights = weights; |
| 372 | _bias = bias; |
| 373 | _output = output; |
| 374 | _is_prepared = false; |
| 375 | |
| 376 | // Create convolver |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 377 | _pImpl->_dwc_assembly_kernel = create_convolver(input, weights, output, conv_info, act_info, dilation); |
| 378 | ARM_COMPUTE_ERROR_ON(_pImpl->_dwc_assembly_kernel == nullptr); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 379 | |
| 380 | // Create assembly kernel wrapper |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 381 | _pImpl->_dwc_acl_kernel.configure(_pImpl->_dwc_assembly_kernel.get()); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 382 | |
| 383 | constexpr size_t alignment = 128; |
| 384 | |
| 385 | // Create workspace |
| 386 | const unsigned int num_threads = NEScheduler::get().num_threads(); |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 387 | const size_t workspace_size = _pImpl->_dwc_assembly_kernel->get_working_space_size(num_threads); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 388 | ARM_COMPUTE_ERROR_ON_MSG(workspace_size == 0, "Workspace size cannot be 0 !"); |
| 389 | _workspace.allocator()->init(TensorInfo(TensorShape{ workspace_size }, 1, DataType::S8), alignment); |
| 390 | _memory_group.manage(&_workspace); |
| 391 | _workspace.allocator()->allocate(); |
| 392 | |
| 393 | // Create packing tensor |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 394 | const size_t pack_tensor_size = _pImpl->_dwc_assembly_kernel->get_packed_params_size(); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 395 | ARM_COMPUTE_ERROR_ON_MSG(pack_tensor_size == 0, "Pack tensor size cannot be 0 !"); |
| 396 | _packed_weights.allocator()->init(TensorInfo(TensorShape{ pack_tensor_size }, 1, DataType::S8), alignment); |
| 397 | } |
| 398 | |
| 399 | Status NEDepthwiseConvolutionAssemblyDispatch::validate(const ITensorInfo *input, |
| 400 | const ITensorInfo *weights, |
| 401 | const ITensorInfo *bias, |
| 402 | const ITensorInfo *output, |
| 403 | const PadStrideInfo &conv_info, |
| 404 | unsigned int depth_multiplier, |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 405 | const ActivationLayerInfo &act_info, |
| 406 | const Size2D &dilation) |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 407 | { |
| 408 | ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input); |
| 409 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); |
Giuseppe Rossini | f01201a | 2019-11-06 14:57:49 +0000 | [diff] [blame] | 410 | if(weights->data_type() != DataType::QSYMM8_PER_CHANNEL) |
| 411 | { |
| 412 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); |
| 413 | } |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 414 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights); |
| 415 | |
Georgios Pinitas | 4c75851 | 2019-07-10 19:49:11 +0100 | [diff] [blame] | 416 | // Validate convolver |
| 417 | ARM_COMPUTE_RETURN_ERROR_ON(!is_optimized_supported(input, weights, conv_info, depth_multiplier, dilation)); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 418 | |
Georgios Pinitas | 4c75851 | 2019-07-10 19:49:11 +0100 | [diff] [blame] | 419 | // Validate activation |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 420 | const bool is_relu = arm_compute::utils::info_helpers::is_relu(act_info); |
| 421 | const bool is_relu6 = arm_compute::utils::info_helpers::is_relu6(act_info); |
| 422 | ARM_COMPUTE_RETURN_ERROR_ON(act_info.enabled() && !(is_relu || is_relu6)); |
| 423 | |
| 424 | // Check bias |
| 425 | if(bias != nullptr) |
| 426 | { |
Georgios Pinitas | 4c75851 | 2019-07-10 19:49:11 +0100 | [diff] [blame] | 427 | unsigned int channel_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 428 | ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); |
| 429 | ARM_COMPUTE_RETURN_ERROR_ON(bias->dimension(0) != weights->dimension(channel_idx)); |
| 430 | } |
| 431 | |
| 432 | // Check output |
| 433 | if(output->total_size() != 0) |
| 434 | { |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 435 | const TensorShape output_shape = misc::shape_calculator::compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier, dilation); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 436 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape); |
| 437 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 438 | } |
| 439 | |
Michele Di Giorgio | f29d1b7 | 2019-10-29 10:58:13 +0000 | [diff] [blame] | 440 | // The uniform quantization case will only have 1 scale value in the weights quantization info |
| 441 | const UniformQuantizationInfo input_qinfo = input->quantization_info().uniform(); |
| 442 | const QuantizationInfo weights_qinfo = weights->quantization_info(); |
| 443 | const UniformQuantizationInfo output_qinfo = output->quantization_info().uniform(); |
| 444 | for(auto const s : weights_qinfo.scale()) |
| 445 | { |
| 446 | const float fmultipler = input_qinfo.scale * s / output_qinfo.scale; |
| 447 | ARM_COMPUTE_RETURN_ERROR_ON(fmultipler > 1.f); |
| 448 | } |
| 449 | |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 450 | return Status{}; |
| 451 | } |
| 452 | |
| 453 | bool NEDepthwiseConvolutionAssemblyDispatch::is_optimized_supported(const ITensorInfo *input, |
| 454 | const ITensorInfo *weights, |
| 455 | PadStrideInfo conv_info, |
Usama Arif | 881f2de | 2019-04-12 10:29:17 +0100 | [diff] [blame] | 456 | unsigned int depth_multiplier, |
| 457 | const Size2D &dilation) |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 458 | { |
| 459 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights); |
| 460 | |
| 461 | // Reshape input shape if in NHWC format |
| 462 | const DataLayout data_layout = input->data_layout(); |
| 463 | TensorShape in_shape{ input->tensor_shape() }; |
| 464 | if(data_layout == DataLayout::NHWC) |
| 465 | { |
| 466 | in_shape.set(Window::DimX, input->tensor_shape().y()); |
| 467 | in_shape.set(Window::DimY, input->tensor_shape().z()); |
| 468 | in_shape.set(Window::DimZ, input->tensor_shape().x()); |
| 469 | } |
| 470 | |
| 471 | // Check data type |
Michele Di Giorgio | 13ec5f0 | 2020-01-02 12:11:13 +0000 | [diff] [blame] | 472 | // TODO (COMPMID-3004): Add assembly optimized routine for QASYMM8_SIGNED NEDepthwiseConvolutionLayer |
| 473 | const DataType input_type = input->data_type(); |
| 474 | const bool is_input_type_valid = is_data_type_float(input_type) || input_type == DataType::QASYMM8; |
| 475 | const DataType weights_type = weights->data_type(); |
| 476 | const bool is_weights_type_valid = is_data_type_float(weights_type) || weights_type == DataType::QASYMM8 || weights_type == DataType::QASYMM8_SIGNED |
| 477 | || weights_type == DataType::QSYMM8_PER_CHANNEL; |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 478 | |
| 479 | // Check weighs size |
Georgios Pinitas | 4c75851 | 2019-07-10 19:49:11 +0100 | [diff] [blame] | 480 | std::set<unsigned int> supported_kernel_sizes = { 3, 5 }; |
| 481 | const unsigned int width_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); |
| 482 | const unsigned int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); |
| 483 | const unsigned int kernel_w = weights->dimension(width_idx); |
| 484 | const unsigned int kernel_h = weights->dimension(height_idx); |
| 485 | bool weights_supported = (kernel_w == kernel_h) && (supported_kernel_sizes.count(kernel_w) != 0); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 486 | |
| 487 | // Check for supported strides |
| 488 | const auto &strides = conv_info.stride(); |
| 489 | bool supported_strides = (strides.first == strides.second) && ((strides.first == 1) || (strides.first == 2)); |
| 490 | |
| 491 | // Check for supported padding |
Georgios Pinitas | 4c75851 | 2019-07-10 19:49:11 +0100 | [diff] [blame] | 492 | const auto pad_top = conv_info.pad_top(); |
| 493 | const auto pad_right = conv_info.pad_right(); |
| 494 | const auto pad_bottom = conv_info.pad_bottom(); |
| 495 | const auto pad_left = conv_info.pad_left(); |
| 496 | PadStrideInfo same_pad = calculate_same_pad(in_shape, TensorShape(kernel_w, kernel_h), conv_info, DataLayout::NCHW, dilation); |
| 497 | bool is_same_padding = (pad_top == same_pad.pad_top()) && (pad_right == same_pad.pad_right()) && (pad_bottom == same_pad.pad_bottom()) && (pad_left == same_pad.pad_left()); |
| 498 | bool is_valid_padding = (pad_top == 0) && (pad_right == 0) && (pad_bottom == 0) && (pad_left == 0); |
| 499 | bool supported_padding = is_same_padding || is_valid_padding; |
| 500 | // TODO(COMPMID-2464): Enable once dilated conv with stride 2 is supported |
Giuseppe Rossini | f01201a | 2019-11-06 14:57:49 +0000 | [diff] [blame] | 501 | bool is_dilation_supported = ((dilation == Size2D(1U, 1U)) || ((dilation.x() == dilation.y()) && strides.first == 1)); |
| 502 | |
Michele Di Giorgio | 13ec5f0 | 2020-01-02 12:11:13 +0000 | [diff] [blame] | 503 | if(weights_type == DataType::QSYMM8_PER_CHANNEL) |
Giuseppe Rossini | f01201a | 2019-11-06 14:57:49 +0000 | [diff] [blame] | 504 | { |
| 505 | is_dilation_supported = is_dilation_supported && (dilation == Size2D(1U, 1U)); |
| 506 | } |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 507 | |
Michele Di Giorgio | 13ec5f0 | 2020-01-02 12:11:13 +0000 | [diff] [blame] | 508 | return is_input_type_valid && is_weights_type_valid && weights_supported && supported_strides && supported_padding && (depth_multiplier == 1) && is_dilation_supported; |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 509 | } |
| 510 | |
| 511 | void NEDepthwiseConvolutionAssemblyDispatch::run() |
| 512 | { |
| 513 | // Prepare assembly kernel |
| 514 | prepare(); |
| 515 | |
Georgios Pinitas | da953f2 | 2019-04-02 17:27:03 +0100 | [diff] [blame] | 516 | MemoryGroupResourceScope scope_mg(_memory_group); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 517 | |
| 518 | // Setup inputs/outputs |
| 519 | ARM_COMPUTE_ERROR_ON(_workspace.buffer() == nullptr); |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 520 | _pImpl->_dwc_assembly_kernel->set_working_space(static_cast<void *>(_workspace.buffer())); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 521 | |
| 522 | ARM_COMPUTE_ERROR_ON(_input->buffer() == nullptr); |
| 523 | const int input_element_size = _input->info()->element_size(); |
| 524 | const int input_batch_stride = _input->info()->strides_in_bytes()[3] / input_element_size; |
| 525 | const int input_row_stride = _input->info()->strides_in_bytes().z() / input_element_size; |
| 526 | const int input_col_stride = _input->info()->strides_in_bytes().y() / input_element_size; |
| 527 | const void *input_ptr = _input->buffer() + _input->info()->offset_first_element_in_bytes(); |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 528 | _pImpl->_dwc_assembly_kernel->set_input(input_ptr, input_batch_stride, input_row_stride, input_col_stride); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 529 | |
| 530 | ARM_COMPUTE_ERROR_ON(_output->buffer() == nullptr); |
| 531 | const int output_element_size = _output->info()->element_size(); |
| 532 | const int output_batch_stride = _output->info()->strides_in_bytes()[3] / output_element_size; |
| 533 | const int output_row_stride = _output->info()->strides_in_bytes().z() / output_element_size; |
| 534 | const int output_col_stride = _output->info()->strides_in_bytes().y() / output_element_size; |
| 535 | void *output_ptr = _output->buffer() + _output->info()->offset_first_element_in_bytes(); |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 536 | _pImpl->_dwc_assembly_kernel->set_output(output_ptr, output_batch_stride, output_row_stride, output_col_stride); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 537 | |
| 538 | // Schedule assembly kernel |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 539 | NEScheduler::get().schedule(&_pImpl->_dwc_acl_kernel, Window::DimX); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 540 | } |
| 541 | |
| 542 | void NEDepthwiseConvolutionAssemblyDispatch::prepare() |
| 543 | { |
| 544 | if(!_is_prepared) |
| 545 | { |
| 546 | _packed_weights.allocator()->allocate(); |
| 547 | ARM_COMPUTE_ERROR_ON(_packed_weights.buffer() == nullptr); |
| 548 | |
| 549 | // Pack weights and bias |
| 550 | const int weights_element_size = _weights->info()->element_size(); |
| 551 | const int weights_row_stride = _weights->info()->strides_in_bytes().z() / weights_element_size; |
| 552 | const int weights_col_stride = _weights->info()->strides_in_bytes().y() / weights_element_size; |
Georgios Pinitas | 30271c7 | 2019-06-24 14:56:34 +0100 | [diff] [blame] | 553 | _pImpl->_dwc_assembly_kernel->pack_params(_packed_weights.buffer(), |
| 554 | _weights->buffer() + _weights->info()->offset_first_element_in_bytes(), |
| 555 | weights_row_stride, |
| 556 | weights_col_stride, |
| 557 | (_bias != nullptr) ? _bias->buffer() : nullptr); |
| 558 | _pImpl->_dwc_assembly_kernel->set_packed_params_buffer(_packed_weights.buffer()); |
Georgios Pinitas | 47d39dc | 2019-03-11 14:03:23 +0000 | [diff] [blame] | 559 | |
| 560 | _weights->mark_as_unused(); |
| 561 | if(_bias != nullptr) |
| 562 | { |
| 563 | _bias->mark_as_unused(); |
| 564 | } |
| 565 | _is_prepared = true; |
| 566 | } |
| 567 | } |
| 568 | } // namespace arm_compute |