Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2021 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 | */ |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 24 | #include "src/cpu/operators/CpuFullyConnected.h" |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 25 | |
| 26 | #include "arm_compute/core/Helpers.h" |
| 27 | #include "arm_compute/core/ITensorPack.h" |
| 28 | #include "arm_compute/core/Validate.h" |
| 29 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 30 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
| 31 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
ramelg01 | 3ae3d88 | 2021-09-12 23:07:47 +0100 | [diff] [blame] | 32 | #include "src/common/utils/Log.h" |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 33 | #include "src/core/helpers/AutoConfiguration.h" |
| 34 | #include "src/core/helpers/MemoryHelpers.h" |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 35 | #include "src/cpu/kernels/CpuTransposeKernel.h" |
| 36 | #include "src/cpu/operators/CpuConvertFullyConnectedWeights.h" |
| 37 | #include "src/cpu/operators/CpuFlatten.h" |
| 38 | #include "src/cpu/operators/CpuGemm.h" |
| 39 | #include "src/cpu/operators/CpuGemmLowpMatrixMultiplyCore.h" |
| 40 | #include "src/cpu/utils/CpuAuxTensorHandler.h" |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 41 | |
| 42 | namespace arm_compute |
| 43 | { |
| 44 | namespace cpu |
| 45 | { |
| 46 | using namespace arm_compute::experimental; |
| 47 | using namespace arm_compute::misc::shape_calculator; |
| 48 | |
| 49 | namespace |
| 50 | { |
| 51 | // Get min, max bound of a quantized asymmetric dst tensor, with the effect of fused activation |
| 52 | std::pair<PixelValue, PixelValue> get_quantized_asymmetric_output_min_max(const QuantizationInfo &q_info, const ActivationLayerInfo &act_info, DataType data_type) |
| 53 | { |
| 54 | PixelValue type_min{}; |
| 55 | PixelValue type_max{}; |
| 56 | std::tie(type_min, type_max) = get_min_max(data_type); |
| 57 | const UniformQuantizationInfo q_unif = q_info.uniform(); |
| 58 | |
| 59 | if(act_info.enabled()) |
| 60 | { |
| 61 | switch(act_info.activation()) |
| 62 | { |
| 63 | case ActivationLayerInfo::ActivationFunction::RELU: |
| 64 | type_min = PixelValue(q_unif.offset); |
| 65 | break; |
| 66 | case ActivationLayerInfo::ActivationFunction::BOUNDED_RELU: |
| 67 | type_min = PixelValue(q_unif.offset); |
| 68 | type_max = PixelValue(act_info.a(), data_type, q_info); |
| 69 | break; |
| 70 | case ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU: |
| 71 | type_min = PixelValue(act_info.b(), data_type, q_info); |
| 72 | type_max = PixelValue(act_info.a(), data_type, q_info); |
| 73 | break; |
| 74 | default: |
| 75 | ARM_COMPUTE_ERROR("Activation function not supported."); |
| 76 | break; |
| 77 | } |
| 78 | } |
| 79 | |
| 80 | return std::make_pair(type_min, type_max); |
| 81 | } |
| 82 | |
| 83 | Status get_gemmlowp_output_stage_info(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, const ActivationLayerInfo &act, |
| 84 | GEMMLowpOutputStageInfo &gemmlowp_output_stage_info) |
| 85 | { |
| 86 | const auto data_type = src->data_type(); |
| 87 | const QuantizationInfo oq_info = dst->quantization_info(); |
| 88 | const UniformQuantizationInfo iq_unif = src->quantization_info().uniform(); |
| 89 | const UniformQuantizationInfo wq_unif = weights->quantization_info().uniform(); |
| 90 | const UniformQuantizationInfo oq_unif = oq_info.uniform(); |
| 91 | |
| 92 | float multiplier = (iq_unif.scale * wq_unif.scale) / oq_unif.scale; |
| 93 | int32_t output_multiplier; |
| 94 | int32_t output_shift; |
| 95 | |
| 96 | ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift)); |
| 97 | |
| 98 | PixelValue type_min{}; |
| 99 | PixelValue type_max{}; |
| 100 | std::tie(type_min, type_max) = get_quantized_asymmetric_output_min_max(oq_info, act, data_type); |
| 101 | |
| 102 | gemmlowp_output_stage_info.gemmlowp_multiplier = output_multiplier; |
| 103 | gemmlowp_output_stage_info.gemmlowp_shift = output_shift; |
| 104 | gemmlowp_output_stage_info.gemmlowp_offset = oq_unif.offset; |
| 105 | gemmlowp_output_stage_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; |
| 106 | gemmlowp_output_stage_info.gemmlowp_min_bound = type_min.get<int32_t>(); |
| 107 | gemmlowp_output_stage_info.gemmlowp_max_bound = type_max.get<int32_t>(); |
| 108 | |
| 109 | return Status{}; |
| 110 | } |
| 111 | |
| 112 | Status validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const ActivationLayerInfo &act) |
| 113 | { |
| 114 | if(is_data_type_quantized_asymmetric(src->data_type())) |
| 115 | { |
| 116 | // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() |
| 117 | // Extract and negate src and weights offset |
| 118 | const QuantizationInfo src_quantization_info(src->quantization_info().uniform().scale, -src->quantization_info().uniform().offset); |
| 119 | const QuantizationInfo weights_quantization_info(weights->quantization_info().uniform().scale, -weights->quantization_info().uniform().offset); |
| 120 | |
| 121 | GEMMLowpOutputStageInfo gemmlowp_output_stage_info; |
| 122 | ARM_COMPUTE_RETURN_ON_ERROR(get_gemmlowp_output_stage_info(src, weights, dst, act, gemmlowp_output_stage_info)); |
| 123 | |
| 124 | GEMMInfo gemm_info; |
| 125 | gemm_info.set_gemmlowp_output_stage(gemmlowp_output_stage_info); |
| 126 | |
| 127 | // Validate gemmlowp function |
| 128 | TensorInfo src_info = src->clone()->set_quantization_info(src_quantization_info); |
| 129 | TensorInfo weights_info = weights->clone()->set_quantization_info(weights_quantization_info); |
| 130 | ARM_COMPUTE_RETURN_ON_ERROR(CpuGemmLowpMatrixMultiplyCore::validate(&src_info, |
| 131 | &weights_info, |
| 132 | biases, |
| 133 | dst, |
| 134 | gemm_info)); |
| 135 | } |
| 136 | else |
| 137 | { |
| 138 | ARM_COMPUTE_RETURN_ON_ERROR(CpuGemm::validate(src, weights, biases, dst, 1.f, 1.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run */))); |
| 139 | } |
| 140 | |
| 141 | return Status{}; |
| 142 | } |
| 143 | } // namespace |
| 144 | |
| 145 | CpuFullyConnected::CpuFullyConnected() |
| 146 | : _flatten(nullptr), |
| 147 | _convert_weights(nullptr), |
| 148 | _transpose_weights(nullptr), |
| 149 | _mm_gemm(nullptr), |
| 150 | _mm_gemmlowp(nullptr), |
| 151 | _flattened_src(), |
| 152 | _converted_weights(), |
| 153 | _reshaped_weights(), |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 154 | _trans_weights(), |
| 155 | _trans_weights_idx(AuxTensorIdx::Count), |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 156 | _aux_mem(Count), |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 157 | _needs_weights_conversion(false), |
| 158 | _needs_weights_reshape(false), |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 159 | _is_fc_after_conv(false), |
| 160 | _is_quantized_asymmetric(false), |
| 161 | _is_prepared(false) |
| 162 | |
| 163 | { |
| 164 | } |
| 165 | |
| 166 | CpuFullyConnected::~CpuFullyConnected() = default; |
| 167 | |
| 168 | void CpuFullyConnected::configure_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act) |
| 169 | { |
| 170 | if(_is_quantized_asymmetric) |
| 171 | { |
| 172 | // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() |
| 173 | // Extract and negate src and weights offset |
| 174 | const QuantizationInfo src_quantization_info(src->quantization_info().uniform().scale, -src->quantization_info().uniform().offset); |
| 175 | const QuantizationInfo weights_quantization_info(weights->quantization_info().uniform().scale, -weights->quantization_info().uniform().offset); |
| 176 | |
| 177 | TensorInfo src_info = src->clone()->set_quantization_info(src_quantization_info); |
| 178 | TensorInfo weights_info = weights->clone()->set_quantization_info(weights_quantization_info); |
| 179 | |
| 180 | // Configure gemmlowp function and output stage for asymmetric quantized types |
| 181 | GEMMLowpOutputStageInfo gemmlowp_output_stage_info; |
| 182 | const Status status = get_gemmlowp_output_stage_info(&src_info, &weights_info, dst, act, gemmlowp_output_stage_info); |
| 183 | ARM_COMPUTE_ERROR_ON(status.error_code() != ErrorCode::OK); |
| 184 | |
| 185 | GEMMInfo gemm_info; |
| 186 | gemm_info.set_gemmlowp_output_stage(gemmlowp_output_stage_info); |
| 187 | gemm_info.set_activation_info(act); |
| 188 | _mm_gemmlowp = std::make_unique<CpuGemmLowpMatrixMultiplyCore>(); |
| 189 | _mm_gemmlowp->configure(&src_info, &weights_info, biases, dst, gemm_info); |
| 190 | } |
| 191 | else |
| 192 | { |
| 193 | // Configure matrix multiply kernel |
| 194 | GEMMInfo gemm_info(false, false, true /* Reshape weights only for the first run */); |
| 195 | gemm_info.set_activation_info(act); |
| 196 | _mm_gemm = std::make_unique<CpuGemm>(); |
| 197 | _mm_gemm->configure(src, weights, biases, dst, 1.f, 1.0f, gemm_info); |
| 198 | } |
| 199 | } |
| 200 | |
| 201 | void CpuFullyConnected::configure_conv_fc(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act) |
| 202 | { |
| 203 | ARM_COMPUTE_ERROR_ON((weights->dimension(1) != (src->dimension(0) * src->dimension(1) * src->dimension(2)))); |
| 204 | |
| 205 | // If the fully connected layer is called after a convolution layer, the src tensor must be linearized |
| 206 | |
| 207 | // Initialize output tensor for flatten |
| 208 | auto_init_if_empty(_flattened_src, src->clone()->set_tensor_shape(compute_flatten_shape(src))); |
| 209 | |
| 210 | _flatten = std::make_unique<CpuFlatten>(); |
| 211 | _flatten->configure(src, &_flattened_src); |
| 212 | |
| 213 | // Configure matrix multiply kernel |
| 214 | configure_mm(&_flattened_src, weights, biases, dst, act); |
| 215 | } |
| 216 | |
| 217 | void CpuFullyConnected::configure_fc_fc(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act) |
| 218 | { |
| 219 | ARM_COMPUTE_ERROR_ON(src->dimension(0) != weights->dimension(1)); |
| 220 | |
| 221 | // Configure matrix multiply kernel |
| 222 | configure_mm(src, weights, biases, dst, act); |
| 223 | } |
| 224 | |
| 225 | void CpuFullyConnected::configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, |
| 226 | FullyConnectedLayerInfo fc_info) |
| 227 | { |
| 228 | // Perform validate step |
| 229 | ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst); |
| 230 | ARM_COMPUTE_ERROR_THROW_ON(CpuFullyConnected::validate(src, |
| 231 | weights, |
| 232 | biases != nullptr ? biases : nullptr, |
| 233 | dst, |
| 234 | fc_info)); |
ramelg01 | 3ae3d88 | 2021-09-12 23:07:47 +0100 | [diff] [blame] | 235 | ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, fc_info); |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 236 | |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 237 | _needs_weights_conversion = false; |
| 238 | _needs_weights_reshape = fc_info.transpose_weights ? !fc_info.are_weights_reshaped : false; |
| 239 | _needs_weights_reshape = _needs_weights_reshape && !fc_info.retain_internal_weights; |
| 240 | _is_fc_after_conv = true; |
| 241 | _is_quantized_asymmetric = is_data_type_quantized_asymmetric(src->data_type()); |
| 242 | _is_prepared = false; |
| 243 | _trans_weights_idx = AuxTensorIdx::Count; |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 244 | |
| 245 | // With the Fully Connected layer we can have 4 different cases: |
| 246 | // 1) Convolution layer -> Fully Connected layer without batches |
| 247 | // 2) Fully Connected layer -> Fully Connected layer without batches |
| 248 | // 3) Convolution layer -> Fully Connected layer with batches |
| 249 | // 4) Fully Connected layer -> Fully Connected layer with batches |
| 250 | |
| 251 | const ITensorInfo *weights_to_use = weights; |
| 252 | |
| 253 | // Check if we have a fully connected layer with batches |
| 254 | const bool is_batched_fc_layer = dst->dimension(1) > 1; |
| 255 | if(is_batched_fc_layer) |
| 256 | { |
| 257 | _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(src->tensor_shape().cbegin() + 3, |
| 258 | src->tensor_shape().cend(), |
| 259 | dst->tensor_shape().cbegin() + 1)); |
| 260 | } |
| 261 | else |
| 262 | { |
| 263 | _is_fc_after_conv = src->num_dimensions() > 1; |
| 264 | } |
| 265 | |
| 266 | // Reshape weights if needed |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 267 | if(_needs_weights_reshape) |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 268 | { |
| 269 | // Reshape the weights |
| 270 | _transpose_weights = std::make_unique<kernels::CpuTransposeKernel>(); |
| 271 | _transpose_weights->configure(weights, &_reshaped_weights); |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 272 | weights_to_use = &_reshaped_weights; |
| 273 | _trans_weights_idx = AuxTensorIdx::TransposedWeights; |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 274 | } |
| 275 | |
| 276 | // Convert weights if needed |
| 277 | if(_is_fc_after_conv && (src->data_layout() != fc_info.weights_trained_layout)) |
| 278 | { |
| 279 | // Convert weights |
| 280 | _convert_weights = std::make_unique<CpuConvertFullyConnectedWeights>(); |
| 281 | _convert_weights->configure(weights_to_use, |
| 282 | &_converted_weights, |
| 283 | src->tensor_shape(), |
| 284 | fc_info.weights_trained_layout); |
| 285 | |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 286 | weights_to_use = &_converted_weights; |
| 287 | _needs_weights_conversion = true; |
| 288 | _trans_weights_idx = AuxTensorIdx::ConvertedWeights; |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 289 | } |
| 290 | |
| 291 | if(_is_fc_after_conv) |
| 292 | { |
| 293 | // Fully Connected layer after a Convolution Layer without batches |
| 294 | configure_conv_fc(src, weights_to_use, biases, dst, fc_info.activation_info); |
| 295 | } |
| 296 | else |
| 297 | { |
| 298 | // Fully Connected layer after a Fully Connected Layer without batches |
| 299 | configure_fc_fc(src, weights_to_use, biases, dst, fc_info.activation_info); |
| 300 | } |
| 301 | |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 302 | // Retain the tensorinfo with the weights to use |
| 303 | if(_needs_weights_reshape || _needs_weights_conversion) |
| 304 | { |
| 305 | _trans_weights = *weights_to_use; |
| 306 | } |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 307 | |
| 308 | // Set auxiliary memory requirements |
| 309 | auto gemm_mem_req = (_is_quantized_asymmetric) ? _mm_gemmlowp->workspace() : _mm_gemm->workspace(); |
| 310 | for(unsigned int i = 0; i < gemm_mem_req.size(); ++i) |
| 311 | { |
| 312 | _aux_mem[i] = gemm_mem_req[i]; |
| 313 | } |
| 314 | |
| 315 | if(_aux_mem[Pretranspose].size > 0) |
| 316 | { |
Giorgio Arena | 63e0beb | 2021-09-24 14:04:27 +0100 | [diff] [blame] | 317 | // Release permuted weights at the end of prepare as they are further transposed by the assembly dispatch |
| 318 | // Do not release them if biases are dynamic and data type is quantized, since the weights tensor will be used for biases offset calculation |
| 319 | _aux_mem[TransposedWeights] = MemoryInfo(offset_int_vec(TransposedWeights), (_is_quantized_asymmetric |
| 320 | && biases && !(biases->are_values_constant())) ? |
| 321 | MemoryLifetime::Persistent : |
| 322 | MemoryLifetime::Prepare, |
| 323 | _reshaped_weights.total_size()); |
| 324 | _aux_mem[ConvertedWeights] = MemoryInfo(offset_int_vec(ConvertedWeights), MemoryLifetime::Prepare, _converted_weights.total_size()); |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 325 | } |
| 326 | else |
| 327 | { |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 328 | _aux_mem[TransposedWeights] = MemoryInfo(offset_int_vec(TransposedWeights), _needs_weights_conversion ? MemoryLifetime::Prepare : MemoryLifetime::Persistent, _reshaped_weights.total_size()); |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 329 | _aux_mem[ConvertedWeights] = MemoryInfo(offset_int_vec(ConvertedWeights), MemoryLifetime::Persistent, _converted_weights.total_size()); |
| 330 | } |
| 331 | _aux_mem[FlattenedSrc] = MemoryInfo(offset_int_vec(FlattenedSrc), MemoryLifetime::Temporary, _flattened_src.total_size()); |
| 332 | } |
| 333 | |
| 334 | Status CpuFullyConnected::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, |
| 335 | FullyConnectedLayerInfo fc_info) |
| 336 | { |
| 337 | ARM_COMPUTE_UNUSED(fc_info.retain_internal_weights); |
| 338 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst); |
| 339 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32); |
| 340 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights, dst); |
| 341 | ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2); |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 342 | ARM_COMPUTE_RETURN_ERROR_ON(fc_info.activation_info.enabled() && is_data_type_quantized(src->data_type()) && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::RELU |
| 343 | && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU && fc_info.activation_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU); |
Giorgio Arena | 63e0beb | 2021-09-24 14:04:27 +0100 | [diff] [blame] | 344 | ARM_COMPUTE_RETURN_ERROR_ON(!weights->are_values_constant() && (!fc_info.are_weights_reshaped || fc_info.transpose_weights)); |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 345 | |
| 346 | bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true; |
| 347 | bool is_fc_after_conv = true; |
| 348 | |
| 349 | const ITensorInfo &flatten_src = TensorInfo(src->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_flatten_shape(src))); |
| 350 | const ITensorInfo &reshaped_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights))); |
| 351 | const ITensorInfo &converted_weights = weights_reshaped ? TensorInfo(weights->clone()->set_is_resizable(true).reset_padding()) : TensorInfo(*reshaped_weights.clone()); |
| 352 | |
| 353 | // With the Fully Connected layer we can have 4 different cases: |
| 354 | // 1) Convolution layer -> Fully Connected layer without batches |
| 355 | // 2) Fully Connected layer -> Fully Connected layer without batches |
| 356 | // 3) Convolution layer -> Fully Connected layer with batches |
| 357 | // 4) Fully Connected layer -> Fully Connected layer with batches |
| 358 | |
| 359 | const ITensorInfo *src_to_use = src; |
| 360 | const ITensorInfo *weights_to_use = weights; |
| 361 | |
| 362 | // Check if we have a fully connected layer with batches |
| 363 | const bool is_batched_fc_layer = dst->dimension(1) > 1; |
| 364 | |
Giorgio Arena | 63e0beb | 2021-09-24 14:04:27 +0100 | [diff] [blame] | 365 | if(biases != nullptr) |
| 366 | { |
| 367 | ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); |
| 368 | if(is_data_type_quantized(src->data_type())) |
| 369 | { |
| 370 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); |
| 371 | } |
| 372 | else |
| 373 | { |
| 374 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, biases); |
| 375 | } |
| 376 | } |
| 377 | |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 378 | if(is_batched_fc_layer) |
| 379 | { |
| 380 | is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(src->tensor_shape().cbegin() + 3, |
| 381 | src->tensor_shape().cend(), |
| 382 | dst->tensor_shape().cbegin() + 1)); |
| 383 | } |
| 384 | else |
| 385 | { |
| 386 | is_fc_after_conv = src->num_dimensions() > 1; |
| 387 | } |
| 388 | |
| 389 | if(!weights_reshaped) |
| 390 | { |
| 391 | // Validate reshape weights kernel |
| 392 | ARM_COMPUTE_RETURN_ON_ERROR(kernels::CpuTransposeKernel::validate(weights, &reshaped_weights)); |
| 393 | weights_to_use = &reshaped_weights; |
| 394 | } |
| 395 | |
| 396 | if(is_fc_after_conv && (src->data_layout() != fc_info.weights_trained_layout)) |
| 397 | { |
| 398 | // Validate convert weights kernel |
| 399 | ARM_COMPUTE_RETURN_ON_ERROR(CpuConvertFullyConnectedWeights::validate(weights_to_use, |
| 400 | &converted_weights, |
| 401 | src->tensor_shape(), |
| 402 | fc_info.weights_trained_layout)); |
| 403 | weights_to_use = &converted_weights; |
| 404 | } |
| 405 | |
| 406 | if(is_fc_after_conv) |
| 407 | { |
| 408 | // Fully Connected layer after a Convolution Layer without batches |
| 409 | ARM_COMPUTE_RETURN_ERROR_ON((weights_to_use->dimension(1) != (src->dimension(0) * src->dimension(1) * src->dimension(2)))); |
| 410 | |
| 411 | // Validate flatten kernel |
| 412 | ARM_COMPUTE_RETURN_ON_ERROR(CpuFlatten::validate(src, &flatten_src)); |
| 413 | src_to_use = &flatten_src; |
| 414 | } |
| 415 | else |
| 416 | { |
| 417 | // Fully Connected layer after a Fully Connected Layer without batches |
| 418 | ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != weights_to_use->dimension(1)); |
| 419 | } |
| 420 | // Validate matrix multiply kernel |
| 421 | ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(src_to_use, weights_to_use, biases, dst, fc_info.activation_info)); |
| 422 | |
| 423 | return Status{}; |
| 424 | } |
| 425 | |
| 426 | void CpuFullyConnected::run(ITensorPack &tensors) |
| 427 | { |
| 428 | prepare(tensors); |
| 429 | |
| 430 | auto src = tensors.get_const_tensor(ACL_SRC_0); |
| 431 | |
| 432 | CpuAuxTensorHandler flattened_src(offset_int_vec(FlattenedSrc), _flattened_src, tensors, false); |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 433 | CpuAuxTensorHandler transformed_wei(offset_int_vec(_trans_weights_idx), _trans_weights, tensors, false); |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 434 | |
| 435 | // Linearize src if it comes from a convolutional layer |
| 436 | if(_is_fc_after_conv) |
| 437 | { |
| 438 | ITensorPack flatten_pack{ { ACL_SRC, src }, { ACL_DST, flattened_src.get() } }; |
| 439 | _flatten->run(flatten_pack); |
| 440 | } |
| 441 | |
| 442 | ITensorPack gemm_pack = tensors; |
| 443 | gemm_pack.add_const_tensor(ACL_SRC_0, (_is_fc_after_conv) ? flattened_src.get() : src); |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 444 | if(_needs_weights_reshape || _needs_weights_conversion) |
| 445 | { |
| 446 | gemm_pack.add_const_tensor(ACL_SRC_1, transformed_wei.get()); |
| 447 | } |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 448 | |
| 449 | // Run matrix multiply |
| 450 | if(_is_quantized_asymmetric) |
| 451 | { |
| 452 | _mm_gemmlowp->run(gemm_pack); |
| 453 | } |
| 454 | else |
| 455 | { |
| 456 | _mm_gemm->run(gemm_pack); |
| 457 | } |
| 458 | } |
| 459 | |
| 460 | void CpuFullyConnected::prepare(ITensorPack &tensors) |
| 461 | { |
| 462 | if(!_is_prepared) |
| 463 | { |
| 464 | auto weights = tensors.get_const_tensor(ACL_SRC_1); |
| 465 | |
| 466 | CpuAuxTensorHandler reshaped_weights(offset_int_vec(TransposedWeights), _reshaped_weights, tensors, false); |
| 467 | CpuAuxTensorHandler converted_weights(offset_int_vec(ConvertedWeights), _converted_weights, tensors, false); |
| 468 | |
| 469 | // Pointer to current weights |
| 470 | const ITensor *cur_weights = weights; |
| 471 | |
| 472 | // Reshape of the weights (happens only once) |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 473 | if(_needs_weights_reshape) |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 474 | { |
| 475 | // Run reshape weights kernel and mark weights as unused |
| 476 | ITensorPack transpose_pack{ { ACL_SRC, weights }, { ACL_DST, reshaped_weights.get() } }; |
| 477 | NEScheduler::get().schedule_op(_transpose_weights.get(), Window::DimY, _transpose_weights->window(), transpose_pack); |
| 478 | |
| 479 | cur_weights->mark_as_unused(); |
| 480 | cur_weights = reshaped_weights.get(); |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 481 | } |
| 482 | |
| 483 | // Convert weights if needed (happens only once) |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 484 | if(_needs_weights_conversion) |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 485 | { |
| 486 | ITensorPack convert_pack{ { ACL_SRC, cur_weights }, { ACL_DST, converted_weights.get() } }; |
| 487 | _convert_weights->run(convert_pack); |
| 488 | |
| 489 | cur_weights->mark_as_unused(); |
| 490 | cur_weights = converted_weights.get(); |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 491 | } |
| 492 | |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 493 | ITensorPack gemm_pack = tensors; |
| 494 | gemm_pack.add_const_tensor(ACL_SRC_1, cur_weights); |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 495 | |
| 496 | // Prepare GEMM prepare and release unused weights |
| 497 | if(!_is_quantized_asymmetric) |
| 498 | { |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 499 | _mm_gemm->prepare(gemm_pack); |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 500 | } |
| 501 | else |
| 502 | { |
Georgios Pinitas | fa1db17 | 2021-08-12 06:28:09 +0100 | [diff] [blame] | 503 | _mm_gemmlowp->prepare(gemm_pack); |
Michele Di Giorgio | d9cdf14 | 2021-07-02 15:17:08 +0100 | [diff] [blame] | 504 | } |
| 505 | |
| 506 | _is_prepared = true; |
| 507 | } |
| 508 | } |
| 509 | |
| 510 | experimental::MemoryRequirements CpuFullyConnected::workspace() const |
| 511 | { |
| 512 | return _aux_mem; |
| 513 | } |
| 514 | } // namespace cpu |
| 515 | } // namespace arm_compute |