compute/ARMComputeEx/src/runtime/CL/functions/CLFullyConnectedLayerEx.cpp

   1 /*
   2  * Copyright (c) 2020 Samsung Electronics Co., Ltd. All Rights Reserved
   3  *
   4  * Licensed under the Apache License, Version 2.0 (the "License");
   5  * you may not use this file except in compliance with the License.
   6  * You may obtain a copy of the License at
   7  *
   8  *      http://www.apache.org/licenses/LICENSE-2.0
   9  *
  10  * Unless required by applicable law or agreed to in writing, software
  11  * distributed under the License is distributed on an "AS IS" BASIS,
  12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13  * See the License for the specific language governing permissions and
  14  * limitations under the License.
  15  */
  16
  17 /*
  18  * Copyright (c) 2017-2019 ARM Limited.
  19  *
  20  * SPDX-License-Identifier: MIT
  21  *
  22  * Permission is hereby granted, free of charge, to any person obtaining a copy
  23  * of this software and associated documentation files (the "Software"), to
  24  * deal in the Software without restriction, including without limitation the
  25  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  26  * sell copies of the Software, and to permit persons to whom the Software is
  27  * furnished to do so, subject to the following conditions:
  28  *
  29  * The above copyright notice and this permission notice shall be included in all
  30  * copies or substantial portions of the Software.
  31  *
  32  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  33  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  34  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  35  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  36  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  37  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  38  * SOFTWARE.
  39  */
  40
  41 #include "arm_compute/runtime/CL/functions/CLFullyConnectedLayerEx.h"
  42
  43 #include "arm_compute/core/Size2D.h"
  44 #include "arm_compute/core/Validate.h"
  45 #include "arm_compute/core/utils/misc/Cast.h"
  46 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
  47 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
  48 #include "arm_compute/runtime/CL/CLScheduler.h"
  49 #include "support/MemorySupport.h"
  50
  51 #include <algorithm>
  52
  53 namespace arm_compute
  54 {
  55 using namespace arm_compute::misc::shape_calculator;
  56 using namespace arm_compute::utils::cast;
  57
  58 namespace
  59 {
  60 Status construct_gemmlowp_output_stage(const ITensorInfo &input, const ITensorInfo &weights,
  61                                        const ITensorInfo &output,
  62                                        GEMMLowpOutputStageInfo &gemmlowp_output_stage)
  63 {
  64   gemmlowp_output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
  65   gemmlowp_output_stage.gemmlowp_offset = 0;
  66   gemmlowp_output_stage.gemmlowp_multiplier = 0;
  67   gemmlowp_output_stage.gemmlowp_shift = 0;
  68
  69   // Configure output stage for quantized case
  70   if (is_data_type_quantized_asymmetric(input.data_type()))
  71   {
  72     const UniformQuantizationInfo iq_info = input.quantization_info().uniform();
  73     const UniformQuantizationInfo wq_info = weights.quantization_info().uniform();
  74     const UniformQuantizationInfo oq_info = output.quantization_info().uniform();
  75
  76     const auto output_quant_info = (output.total_size() == 0) ? iq_info : oq_info;
  77
  78     const float multiplier = (iq_info.scale * wq_info.scale) / output_quant_info.scale;
  79     int output_multiplier = 0;
  80     int output_shift = 0;
  81     ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier_less_than_one(
  82         multiplier, &output_multiplier, &output_shift));
  83
  84     // Set the GEMMLowp output stage info
  85     gemmlowp_output_stage.gemmlowp_offset = output_quant_info.offset;
  86     gemmlowp_output_stage.gemmlowp_multiplier = output_multiplier;
  87     gemmlowp_output_stage.gemmlowp_shift = output_shift;
  88     gemmlowp_output_stage.gemmlowp_min_bound = 0;
  89     gemmlowp_output_stage.gemmlowp_max_bound = 255;
  90     gemmlowp_output_stage.gemmlowp_multipliers.push_back(output_multiplier);
  91     gemmlowp_output_stage.gemmlowp_shifts.push_back(output_shift);
  92   }
  93
  94   return Status{};
  95 }
  96
  97 Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const ITensorInfo *bias,
  98                    const ITensorInfo &output, const FullyConnectedLayerInfo &fc_info)
  99 {
 100   GEMMLowpOutputStageInfo gemmlowp_output_stage;
 101   ARM_COMPUTE_RETURN_ON_ERROR(
 102       construct_gemmlowp_output_stage(input, weights, output, gemmlowp_output_stage));
 103
 104   const GEMMInfo &gemm_info = GEMMInfo(false, // is_a_reshaped
 105                                        false, // is_b_reshaped
 106                                        true,  // reshape_b_only_on_first_run
 107                                        0,     // depth_output_gemm3d
 108                                        false, // reinterpret_input_as_3d
 109                                        fc_info.retain_internal_weights, // retain_internal_weights
 110                                        gemmlowp_output_stage,           // gemmlowp_output_stage
 111                                        fc_info.fp_mixed_precision,      // fp_mixed_precision
 112                                        true,                            // broadcast_bias
 113                                        ActivationLayerInfo());          // activation_info
 114
 115   if (is_data_type_quantized_asymmetric(input.data_type()))
 116   {
 117     const UniformQuantizationInfo iq_info = input.quantization_info().uniform();
 118     const UniformQuantizationInfo wq_info = weights.quantization_info().uniform();
 119
 120     // Since we need negative offsets for computing convolution, we need to change
 121     // QuantizationInfo()
 122     // Extract and negate input and weights offset
 123     const QuantizationInfo input_quantization_info(iq_info.scale, -iq_info.offset);
 124     const QuantizationInfo weights_quantization_info(wq_info.scale, -wq_info.offset);
 125
 126     // Validate gemmlowp function
 127     ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyCore::validate(
 128         &input.clone()->set_quantization_info(input_quantization_info),
 129         &weights.clone()->set_quantization_info(weights_quantization_info), bias, &output,
 130         gemm_info));
 131   }
 132   else
 133   {
 134     ARM_COMPUTE_RETURN_ON_ERROR(
 135         CLGEMM::validate(&input, &weights, bias, &output, 1.f, 1.f, gemm_info));
 136   }
 137
 138   return Status{};
 139 }
 140 } // namespace
 141
 142 void CLFullyConnectedLayerReshapeWeightsEx::configure(const ICLTensor *input, ICLTensor *output)
 143 {
 144   auto k = support::cpp14::make_unique<CLTransposeKernel>();
 145   k->configure(input, output);
 146   _kernel = std::move(k);
 147 }
 148
 149 Status CLFullyConnectedLayerReshapeWeightsEx::validate(const ITensorInfo *input,
 150                                                        const ITensorInfo *output)
 151 {
 152   return CLTransposeKernel::validate(input, output);
 153 }
 154
 155 CLFullyConnectedLayerEx::CLFullyConnectedLayerEx(std::shared_ptr<IMemoryManager> memory_manager,
 156                                                  IWeightsManager *weights_manager)
 157     : _memory_group(memory_manager), _weights_manager(weights_manager), _convert_weights(),
 158       _convert_weights_managed(), _reshape_weights_managed_function(), _flatten_layer(),
 159       _reshape_weights_function(), _mm_gemm(memory_manager, weights_manager),
 160       _mm_gemmlowp(memory_manager), _flatten_output(), _converted_weights_output(),
 161       _reshape_weights_output(), _are_weights_converted(true), _are_weights_reshaped(true),
 162       _is_fc_after_conv(true), _is_quantized(false), _is_prepared(false), _original_weights(nullptr)
 163 {
 164 }
 165 void CLFullyConnectedLayerEx::configure_mm(const ICLTensor *input, const ICLTensor *weights,
 166                                            const ICLTensor *bias, ICLTensor *output,
 167                                            const FullyConnectedLayerInfo &fc_info)
 168 {
 169   GEMMLowpOutputStageInfo gemmlowp_output_stage;
 170   construct_gemmlowp_output_stage(*input->info(), *weights->info(), *output->info(),
 171                                   gemmlowp_output_stage);
 172
 173   const GEMMInfo &gemm_info = GEMMInfo(false, // is_a_reshaped
 174                                        false, // is_b_reshaped
 175                                        true,  // reshape_b_only_on_first_run
 176                                        0,     // depth_output_gemm3d
 177                                        false, // reinterpret_input_as_3d
 178                                        fc_info.retain_internal_weights, // retain_internal_weights
 179                                        gemmlowp_output_stage,           // gemmlowp_output_stage
 180                                        fc_info.fp_mixed_precision,      // fp_mixed_precision
 181                                        true,                            // broadcast_bias
 182                                        ActivationLayerInfo());          // activation_info
 183
 184   if (_is_quantized)
 185   {
 186     // Since we need negative offsets for computing convolution, we need to change
 187     // QuantizationInfo()
 188     // Extract and negate input and weights offset
 189     const QuantizationInfo input_quantization_info = input->info()->quantization_info();
 190     const QuantizationInfo weights_quantization_info = weights->info()->quantization_info();
 191
 192     input->info()->set_quantization_info(QuantizationInfo(
 193         input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset));
 194     weights->info()->set_quantization_info(QuantizationInfo(
 195         weights_quantization_info.uniform().scale, -weights_quantization_info.uniform().offset));
 196
 197     // Configure gemmlowp function
 198     _mm_gemmlowp.configure(input, weights, bias, output, gemm_info);
 199
 200     // Revert back QuantizatioInfo as input and weights could be used in other fully connected
 201     // layers
 202     input->info()->set_quantization_info(input_quantization_info);
 203     weights->info()->set_quantization_info(weights_quantization_info);
 204   }
 205   else
 206   {
 207     // Configure matrix multiply kernel
 208     _mm_gemm.configure(input, weights, bias, output, 1.f, 1.f, gemm_info);
 209   }
 210 }
 211
 212 void CLFullyConnectedLayerEx::configure_conv_fc(const ICLTensor *input, const ICLTensor *weights,
 213                                                 const ICLTensor *bias, ICLTensor *output,
 214                                                 const FullyConnectedLayerInfo &fc_info)
 215 {
 216   ARM_COMPUTE_ERROR_ON(
 217       (weights->info()->dimension(1) !=
 218        (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))));
 219
 220   // If the fully connected layer is called after a convolution layer, the input tensor must be
 221   // linearized
 222
 223   // Initialize output tensor for flatten
 224   TensorShape shape_flatten = compute_flatten_shape(input->info());
 225   _flatten_output.allocator()->init(input->info()
 226                                         ->clone()
 227                                         ->set_is_resizable(true)
 228                                         .reset_padding()
 229                                         .set_tensor_shape(shape_flatten)
 230                                         .set_data_layout(DataLayout::NCHW));
 231
 232   // Configure flatten kernel
 233   _memory_group.manage(&_flatten_output);
 234   _flatten_layer.configure(input, &_flatten_output);
 235
 236   // Configure matrix multiply kernel
 237   configure_mm(&_flatten_output, weights, bias, output, fc_info);
 238
 239   // Allocate the output tensor for flatten once all the configure methods have been called
 240   _flatten_output.allocator()->allocate();
 241 }
 242
 243 void CLFullyConnectedLayerEx::configure_fc_fc(const ICLTensor *input, const ICLTensor *weights,
 244                                               const ICLTensor *bias, ICLTensor *output,
 245                                               const FullyConnectedLayerInfo &fc_info)
 246 {
 247   ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1));
 248
 249   // Configure matrix multiply kernel
 250   configure_mm(input, weights, bias, output, fc_info);
 251 }
 252
 253 void CLFullyConnectedLayerEx::configure(const ICLTensor *input, const ICLTensor *weights,
 254                                         const ICLTensor *biases, ICLTensor *output,
 255                                         FullyConnectedLayerInfo fc_info)
 256 {
 257   ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
 258
 259   // Perform validate step
 260   ARM_COMPUTE_ERROR_THROW_ON(CLFullyConnectedLayerEx::validate(
 261       input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(),
 262       fc_info));
 263
 264   _are_weights_converted = true;
 265   _are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
 266   _is_fc_after_conv = true;
 267   _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
 268   _is_prepared = fc_info.retain_internal_weights;
 269   _original_weights = weights;
 270
 271   if (_weights_manager)
 272   {
 273     _weights_manager->manage(weights);
 274   }
 275
 276   const ICLTensor *weights_to_use = weights;
 277
 278   // With the Fully Connected layer we can have 4 different cases:
 279   //  1) Convolution layer -> Fully Connected layer without batches
 280   //  2) Fully Connected layer -> Fully Connected layer without batches
 281   //  3) Convolution layer -> Fully Connected layer with batches
 282   //  4) Fully Connected layer -> Fully Connected layer with batches
 283
 284   // Check if we have a fully connected layer with batches
 285   const bool is_batched_fc_layer = output->info()->dimension(1) > 1;
 286   if (is_batched_fc_layer)
 287   {
 288     _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) &&
 289                         (std::equal(input->info()->tensor_shape().cbegin() + 3,
 290                                     input->info()->tensor_shape().cend(),
 291                                     output->info()->tensor_shape().cbegin() + 1));
 292   }
 293   else
 294   {
 295     _is_fc_after_conv = input->info()->num_dimensions() > 1;
 296   }
 297
 298   // Reshape weights if needed
 299   if (!_are_weights_reshaped)
 300   {
 301     if (_weights_manager && _weights_manager->are_weights_managed(weights))
 302     {
 303       _reshape_weights_managed_function.configure(weights);
 304       weights_to_use = utils::cast::polymorphic_downcast<ICLTensor *>(
 305           _weights_manager->acquire(weights, &_reshape_weights_managed_function));
 306     }
 307     else
 308     {
 309       // Reshape the weights
 310       _reshape_weights_function.configure(weights, &_reshape_weights_output);
 311       weights_to_use = &_reshape_weights_output;
 312     }
 313   }
 314
 315   // Convert weights if needed
 316   if (_is_fc_after_conv && (input->info()->data_layout() != fc_info.weights_trained_layout))
 317   {
 318     if (_weights_manager && _weights_manager->are_weights_managed(weights_to_use))
 319     {
 320       _convert_weights_managed.configure(weights_to_use, input->info()->tensor_shape(),
 321                                          fc_info.weights_trained_layout);
 322       weights_to_use = utils::cast::polymorphic_downcast<ICLTensor *>(
 323           _weights_manager->acquire(weights, &_convert_weights_managed));
 324     }
 325     else
 326     {
 327       // Convert weights
 328       _convert_weights.configure(weights_to_use, &_converted_weights_output,
 329                                  input->info()->tensor_shape(), fc_info.weights_trained_layout);
 330
 331       weights_to_use = &_converted_weights_output;
 332     }
 333     _are_weights_converted = false;
 334   }
 335
 336   if (_is_fc_after_conv)
 337   {
 338     // Fully Connected layer after a Convolution Layer without batches
 339     configure_conv_fc(input, weights_to_use, biases, output, fc_info);
 340   }
 341   else
 342   {
 343     // Fully Connected layer after a Fully Connected Layer without batches
 344     configure_fc_fc(input, weights_to_use, biases, output, fc_info);
 345   }
 346 }
 347
 348 Status CLFullyConnectedLayerEx::validate(const ITensorInfo *input, const ITensorInfo *weights,
 349                                          const ITensorInfo *biases, const ITensorInfo *output,
 350                                          FullyConnectedLayerInfo fc_info)
 351 {
 352   ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
 353   ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16,
 354                                                        DataType::F32);
 355   ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
 356   ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2);
 357
 358   bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
 359   bool is_fc_after_conv = true;
 360
 361   const ITensorInfo &flatten_input = TensorInfo(input->clone()
 362                                                     ->set_is_resizable(true)
 363                                                     .reset_padding()
 364                                                     .set_tensor_shape(compute_flatten_shape(input))
 365                                                     .set_data_layout(DataLayout::NCHW));
 366   const ITensorInfo &reshaped_weights =
 367       TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(
 368           compute_transposed_shape(*weights)));
 369   const ITensorInfo &converted_weights =
 370       weights_reshaped ? TensorInfo(weights->clone()->set_is_resizable(true).reset_padding())
 371                        : TensorInfo(*reshaped_weights.clone());
 372
 373   // With the Fully Connected layer we can have 4 different cases:
 374   //  1) Convolution layer -> Fully Connected layer without batches
 375   //  2) Fully Connected layer -> Fully Connected layer without batches
 376   //  3) Convolution layer -> Fully Connected layer with batches
 377   //  4) Fully Connected layer -> Fully Connected layer with batches
 378
 379   const ITensorInfo *input_to_use = input;
 380   const ITensorInfo *weights_to_use = weights;
 381
 382   // Check if we have a fully connected layer with batches
 383   const bool is_batched_fc_layer = output->dimension(1) > 1;
 384   if (is_batched_fc_layer)
 385   {
 386     is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) &&
 387                        (std::equal(input->tensor_shape().cbegin() + 3, input->tensor_shape().cend(),
 388                                    output->tensor_shape().cbegin() + 1));
 389   }
 390   else
 391   {
 392     is_fc_after_conv = input->num_dimensions() > 1;
 393   }
 394
 395   if (!weights_reshaped)
 396   {
 397     // Validate reshape weights kernel
 398     ARM_COMPUTE_RETURN_ON_ERROR(
 399         CLFullyConnectedLayerReshapeWeightsEx::validate(weights, &reshaped_weights));
 400     weights_to_use = &reshaped_weights;
 401   }
 402
 403   if (is_fc_after_conv && (input->data_layout() != fc_info.weights_trained_layout))
 404   {
 405     // Validate convert weights kernel
 406     ARM_COMPUTE_RETURN_ON_ERROR(CLConvertFullyConnectedWeights::validate(
 407         weights_to_use, &converted_weights, input->tensor_shape(), fc_info.weights_trained_layout));
 408     weights_to_use = &converted_weights;
 409   }
 410
 411   if (is_fc_after_conv)
 412   {
 413     // Fully Connected layer after a Convolution Layer without batches
 414     ARM_COMPUTE_RETURN_ERROR_ON(
 415         (weights_to_use->dimension(1) !=
 416          (input->dimension(0) * input->dimension(1) * input->dimension(2))));
 417
 418     // Validate flatten kernel
 419     ARM_COMPUTE_RETURN_ON_ERROR(CLFlattenLayer::validate(input, &flatten_input));
 420     input_to_use = &flatten_input;
 421   }
 422   else
 423   {
 424     // Fully Connected layer after a Fully Connected Layer without batches
 425     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != weights_to_use->dimension(1));
 426   }
 427
 428   // Validate matrix multiply kernel
 429   ARM_COMPUTE_RETURN_ON_ERROR(
 430       validate_mm(*input_to_use, *weights_to_use, biases, *output, fc_info));
 431
 432   return Status{};
 433 }
 434
 435 void CLFullyConnectedLayerEx::run()
 436 {
 437   if (!_is_prepared)
 438   {
 439     if (!_are_weights_reshaped)
 440       _reshape_weights_output.allocator()->allocate();
 441     if (!_are_weights_converted)
 442       _converted_weights_output.allocator()->allocate();
 443     _is_prepared = true;
 444   }
 445
 446   {
 447     if (!_weights_manager)
 448     {
 449       ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
 450     }
 451
 452     // Pointer to current weights
 453     const ICLTensor *cur_weights = _original_weights;
 454     // Reshape of the weights
 455     if (!_are_weights_reshaped)
 456     {
 457       if (_weights_manager && _weights_manager->are_weights_managed(cur_weights))
 458       {
 459         _original_weights = utils::cast::polymorphic_downcast<ICLTensor *>(
 460             _weights_manager->run(cur_weights, &_reshape_weights_managed_function));
 461       }
 462       else
 463       {
 464         _reshape_weights_function.run();
 465         cur_weights = &_reshape_weights_output;
 466       }
 467     }
 468
 469     // Convert weights if needed
 470     if (!_are_weights_converted)
 471     {
 472       if (_weights_manager && _weights_manager->are_weights_managed(cur_weights))
 473       {
 474         _weights_manager->run(cur_weights, &_convert_weights_managed);
 475       }
 476       else
 477       {
 478         _convert_weights.run();
 479       }
 480     }
 481
 482     // Prepare GEMM prepare
 483     if (!_is_quantized)
 484     {
 485       _mm_gemm.prepare();
 486     }
 487   }
 488
 489   MemoryGroupResourceScope scope_mg(_memory_group);
 490
 491   // Linearize input if it comes from a convolutional layer
 492   if (_is_fc_after_conv)
 493   {
 494     _flatten_layer.run();
 495   }
 496
 497   // Run matrix multiply
 498   if (_is_quantized)
 499   {
 500     _mm_gemmlowp.run();
 501   }
 502   else
 503   {
 504     _mm_gemm.run();
 505   }
 506 }
 507
 508 void CLFullyConnectedLayerEx::prepare()
 509 {
 510 #if 0 // TODO Remove this block
 511     if(!_is_prepared)
 512     {
 513         if(!_weights_manager)
 514         {
 515             ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
 516         }
 517
 518         auto release_unused = [](CLTensor * w)
 519         {
 520             if(!w->is_used())
 521             {
 522                 CLScheduler::get().queue().finish();
 523                 w->allocator()->free();
 524             }
 525         };
 526
 527         // Pointer to current weights
 528         const ICLTensor *cur_weights = _original_weights;
 529
 530         // Reshape of the weights if needed (happens only once)
 531         if(!_are_weights_reshaped)
 532         {
 533             if(_weights_manager && _weights_manager->are_weights_managed(_original_weights))
 534             {
 535                 cur_weights = utils::cast::polymorphic_downcast<ICLTensor *>(_weights_manager->run(cur_weights, &_reshape_weights_managed_function));
 536             }
 537             else
 538             {
 539                 // Run reshape weights kernel and mark weights as unused
 540                 _reshape_weights_output.allocator()->allocate();
 541                 _reshape_weights_function.run();
 542
 543                 cur_weights->mark_as_unused();
 544                 cur_weights = &_reshape_weights_output;
 545             }
 546             _are_weights_reshaped = true;
 547         }
 548
 549         // Convert weights if needed (happens only once)
 550         if(!_are_weights_converted)
 551         {
 552             if(_weights_manager && _weights_manager->are_weights_managed(cur_weights))
 553             {
 554                 _weights_manager->run(cur_weights, &_convert_weights_managed);
 555             }
 556             else
 557             {
 558                 _converted_weights_output.allocator()->allocate();
 559                 _convert_weights.run();
 560                 cur_weights->mark_as_unused();
 561             }
 562
 563             _are_weights_converted = true;
 564         }
 565
 566         // Release reshaped weights if unused
 567         release_unused(&_reshape_weights_output);
 568
 569         // Prepare GEMM prepare and release unused weights
 570         if(!_is_quantized)
 571         {
 572             _mm_gemm.prepare();
 573         }
 574
 575         // Release converted weights if unused
 576         release_unused(&_reshape_weights_output);
 577         release_unused(&_converted_weights_output);
 578
 579         _is_prepared = true;
 580     }
 581 #endif
 582 }
 583 } // namespace arm_compute