compute/ARMComputeEx/src/runtime/NEON/functions/NEFullyConnectedLayerEx.cpp

   1 /*
   2  * Copyright (c) 2019 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/NEON/functions/NEFullyConnectedLayerEx.h"
  42
  43 #include "arm_compute/core/Helpers.h"
  44 #include "arm_compute/core/Size2D.h"
  45 #include "arm_compute/core/Validate.h"
  46 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
  47 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
  48 #include "arm_compute/runtime/NEON/NEScheduler.h"
  49
  50 #include <algorithm>
  51 #include <cmath>
  52
  53 using namespace arm_compute;
  54 using namespace arm_compute::misc::shape_calculator;
  55
  56 namespace
  57 {
  58 Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const ITensorInfo &output)
  59 {
  60   if (is_data_type_quantized_asymmetric(input.data_type()))
  61   {
  62     // Since we need negative offsets for computing convolution, we need to change
  63     // QuantizationInfo()
  64     // Extract and negate input and weights offset
  65     const QuantizationInfo input_quantization_info(input.quantization_info().uniform().scale,
  66                                                    -input.quantization_info().uniform().offset);
  67     const QuantizationInfo weights_quantization_info(weights.quantization_info().uniform().scale,
  68                                                      -weights.quantization_info().uniform().offset);
  69
  70     // Validate gemmlowp function
  71     ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpMatrixMultiplyCore::validate(
  72         &input.clone()->set_quantization_info(input_quantization_info),
  73         &weights.clone()->set_quantization_info(weights_quantization_info), nullptr, &output));
  74   }
  75   else
  76   {
  77     ARM_COMPUTE_RETURN_ON_ERROR(NEGEMM::validate(
  78         &input, &weights, nullptr, &output, 1.f, 0.0f,
  79         GEMMInfo(false, false, false /* Reshape weights only for the first run */)));
  80   }
  81
  82   return Status{};
  83 }
  84 } // namespace
  85
  86 NEFullyConnectedLayerEx::NEFullyConnectedLayerEx(std::shared_ptr<IMemoryManager> memory_manager)
  87     : _memory_group(std::move(memory_manager)), _flatten_kernel(), _convert_weights(),
  88       _reshape_weights_function(), _mm_gemm(), _mm_gemmlowp(), _gemmlowp_output_stage(),
  89       _accumulate_biases_kernel(), _flatten_output(), _gemmlowp_output(),
  90       _converted_weights_output(), _reshape_weights_output(), _original_weights(nullptr),
  91       _are_weights_converted(true), _are_weights_reshaped(false), _is_fc_after_conv(false),
  92       _accumulate_biases(false), _is_quantized(false), _is_prepared(false)
  93 {
  94 }
  95
  96 void NEFullyConnectedLayerEx::configure_mm(const ITensor *input, const ITensor *weights,
  97                                            ITensor *output)
  98 {
  99   if (_is_quantized)
 100   {
 101     // Since we need negative offsets for computing convolution, we need to change
 102     // QuantizationInfo()
 103     // Extract and negate input and weights offset
 104     const QuantizationInfo input_quantization_info = input->info()->quantization_info();
 105     const QuantizationInfo weights_quantization_info = weights->info()->quantization_info();
 106
 107     input->info()->set_quantization_info(QuantizationInfo(
 108         input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset));
 109     weights->info()->set_quantization_info(QuantizationInfo(
 110         weights_quantization_info.uniform().scale, -weights_quantization_info.uniform().offset));
 111
 112     // Configure gemmlowp function
 113     _mm_gemmlowp.configure(input, weights, nullptr, output);
 114
 115     // Revert back QuantizatioInfo as input and weights could be used in other fully connected
 116     // layers
 117     input->info()->set_quantization_info(input_quantization_info);
 118     weights->info()->set_quantization_info(weights_quantization_info);
 119   }
 120   else
 121   {
 122     // Configure matrix multiply kernel
 123     _mm_gemm.configure(input, weights, nullptr, output, 1.f, 0.0f,
 124                        GEMMInfo(false, false, false /* Reshape weights only for the first run */));
 125   }
 126 }
 127
 128 void NEFullyConnectedLayerEx::configure_conv_fc(const ITensor *input, const ITensor *weights,
 129                                                 ITensor *output)
 130 {
 131   ARM_COMPUTE_ERROR_ON(
 132       (weights->info()->dimension(1) !=
 133        (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))));
 134
 135   // If the fully connected layer is called after a convolution layer, the input tensor must be
 136   // linearized
 137
 138   // Initialize output tensor for flatten
 139   TensorShape shape_flatten = compute_flatten_shape(input->info());
 140   _flatten_output.allocator()->init(
 141       input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(
 142           shape_flatten));
 143
 144   // Configure flatten kernel
 145   _memory_group.manage(&_flatten_output);
 146   _flatten_kernel.configure(input, &_flatten_output);
 147
 148   // Configure matrix multiply kernel
 149   configure_mm(&_flatten_output, weights, output);
 150
 151   // Allocate the output tensor for flatten once all the configure methods have been called
 152   _flatten_output.allocator()->allocate();
 153 }
 154
 155 void NEFullyConnectedLayerEx::configure_fc_fc(const ITensor *input, const ITensor *weights,
 156                                               ITensor *output)
 157 {
 158   ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1));
 159
 160   // Configure matrix multiply kernel
 161   configure_mm(input, weights, output);
 162 }
 163
 164 void NEFullyConnectedLayerEx::configure(const ITensor *input, const ITensor *weights,
 165                                         const ITensor *biases, ITensor *output,
 166                                         FullyConnectedLayerInfo fc_info)
 167 {
 168   ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
 169
 170   // Perform validate step
 171   ARM_COMPUTE_ERROR_THROW_ON(NEFullyConnectedLayerEx::validate(
 172       input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(),
 173       fc_info));
 174
 175   _are_weights_converted = true;
 176   _are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
 177   _is_fc_after_conv = true;
 178   _accumulate_biases = false;
 179   _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
 180   _original_weights = weights;
 181
 182   // Configure gemmlowp output
 183   if (_is_quantized)
 184   {
 185     _gemmlowp_output.allocator()->init(
 186         output->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(
 187             DataType::S32));
 188   }
 189
 190   // Configure accumulate biases kernel for non quantized asymmetric types
 191   if (biases != nullptr && !_is_quantized)
 192   {
 193     _accumulate_biases = true;
 194
 195     // Configure accumulate biases kernel
 196     _accumulate_biases_kernel.configure(output, biases);
 197   }
 198
 199   // With the Fully Connected layer we can have 4 different cases:
 200   //  1) Convolution layer -> Fully Connected layer without batches
 201   //  2) Fully Connected layer -> Fully Connected layer without batches
 202   //  3) Convolution layer -> Fully Connected layer with batches
 203   //  4) Fully Connected layer -> Fully Connected layer with batches
 204
 205   const ITensor *weights_to_use = weights;
 206
 207   // Check if we have a fully connected layer with batches
 208   const bool is_batched_fc_layer = output->info()->dimension(1) > 1;
 209   if (is_batched_fc_layer)
 210   {
 211     _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) &&
 212                         (std::equal(input->info()->tensor_shape().cbegin() + 3,
 213                                     input->info()->tensor_shape().cend(),
 214                                     output->info()->tensor_shape().cbegin() + 1));
 215   }
 216   else
 217   {
 218     _is_fc_after_conv = input->info()->num_dimensions() > 1;
 219   }
 220
 221   // Reshape weights if needed
 222   if (!_are_weights_reshaped)
 223   {
 224     // Reshape the weights
 225     _reshape_weights_function.configure(weights, &_reshape_weights_output);
 226     weights_to_use = &_reshape_weights_output;
 227   }
 228
 229   // Convert weights if needed
 230   if (_is_fc_after_conv && (input->info()->data_layout() != fc_info.weights_trained_layout))
 231   {
 232     // Convert weights
 233     _convert_weights.configure(weights_to_use, &_converted_weights_output,
 234                                input->info()->tensor_shape(), fc_info.weights_trained_layout);
 235
 236     weights_to_use = &_converted_weights_output;
 237     _are_weights_converted = false;
 238   }
 239
 240   ITensor *tmp_output = (_is_quantized) ? &_gemmlowp_output : output;
 241   if (_is_fc_after_conv)
 242   {
 243     // Fully Connected layer after a Convolution Layer without batches
 244     configure_conv_fc(input, weights_to_use, tmp_output);
 245   }
 246   else
 247   {
 248     // Fully Connected layer after a Fully Connected Layer without batches
 249     configure_fc_fc(input, weights_to_use, tmp_output);
 250   }
 251
 252   // Configure output stage for asymmetric quantized types
 253   if (_is_quantized)
 254   {
 255     float multiplier = input->info()->quantization_info().uniform().scale *
 256                        weights->info()->quantization_info().uniform().scale /
 257                        output->info()->quantization_info().uniform().scale;
 258     int output_multiplier;
 259     int output_shift;
 260     quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier,
 261                                                                &output_shift);
 262     _gemmlowp_output_stage.configure(&_gemmlowp_output, biases, output, output_multiplier,
 263                                      output_shift,
 264                                      output->info()->quantization_info().uniform().offset);
 265     _gemmlowp_output.allocator()->allocate();
 266   }
 267
 268   _are_weights_reshaped = _are_weights_reshaped || fc_info.retain_internal_weights;
 269 }
 270
 271 Status NEFullyConnectedLayerEx::validate(const ITensorInfo *input, const ITensorInfo *weights,
 272                                          const ITensorInfo *biases, const ITensorInfo *output,
 273                                          FullyConnectedLayerInfo fc_info)
 274 {
 275   ARM_COMPUTE_UNUSED(fc_info.retain_internal_weights);
 276   ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
 277   ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16,
 278                                                        DataType::F32);
 279   ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
 280   ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2);
 281
 282   bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
 283   bool is_fc_after_conv = true;
 284   bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
 285
 286   const ITensorInfo &flatten_input =
 287       TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(
 288           compute_flatten_shape(input)));
 289   const ITensorInfo &reshaped_weights =
 290       TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(
 291           compute_transposed_shape(*weights)));
 292   const ITensorInfo &converted_weights =
 293       weights_reshaped ? TensorInfo(weights->clone()->set_is_resizable(true).reset_padding())
 294                        : TensorInfo(*reshaped_weights.clone());
 295   const ITensorInfo &gemmlowp_output = TensorInfo(
 296       output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
 297
 298   // Configure accumulate biases kernel for non quantized asymmetric types
 299   if (biases != nullptr && !is_quantized)
 300   {
 301     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
 302     ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixAccumulateBiasesKernel::validate(output, biases));
 303   }
 304
 305   // With the Fully Connected layer we can have 4 different cases:
 306   //  1) Convolution layer -> Fully Connected layer without batches
 307   //  2) Fully Connected layer -> Fully Connected layer without batches
 308   //  3) Convolution layer -> Fully Connected layer with batches
 309   //  4) Fully Connected layer -> Fully Connected layer with batches
 310
 311   const ITensorInfo *input_to_use = input;
 312   const ITensorInfo *weights_to_use = weights;
 313   const ITensorInfo *tmp_output = (is_quantized) ? &gemmlowp_output : output;
 314
 315   // Check if we have a fully connected layer with batches
 316   const bool is_batched_fc_layer = output->dimension(1) > 1;
 317
 318   if (is_batched_fc_layer)
 319   {
 320     is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) &&
 321                        (std::equal(input->tensor_shape().cbegin() + 3, input->tensor_shape().cend(),
 322                                    output->tensor_shape().cbegin() + 1));
 323   }
 324   else
 325   {
 326     is_fc_after_conv = input->num_dimensions() > 1;
 327   }
 328
 329   if (!weights_reshaped)
 330   {
 331     // Validate reshape weights kernel
 332     ARM_COMPUTE_RETURN_ON_ERROR(
 333         NEFullyConnectedLayerReshapeWeights::validate(weights, &reshaped_weights));
 334     weights_to_use = &reshaped_weights;
 335   }
 336
 337   if (is_fc_after_conv && (input->data_layout() != fc_info.weights_trained_layout))
 338   {
 339     // Validate convert weights kernel
 340     ARM_COMPUTE_RETURN_ON_ERROR(NEConvertFullyConnectedWeights::validate(
 341         weights_to_use, &converted_weights, input->tensor_shape(), fc_info.weights_trained_layout));
 342     weights_to_use = &converted_weights;
 343   }
 344
 345   if (is_fc_after_conv)
 346   {
 347     // Fully Connected layer after a Convolution Layer without batches
 348     ARM_COMPUTE_RETURN_ERROR_ON(
 349         (weights_to_use->dimension(1) !=
 350          (input->dimension(0) * input->dimension(1) * input->dimension(2))));
 351
 352     // Validate flatten kernel
 353     ARM_COMPUTE_RETURN_ON_ERROR(NEFlattenLayerKernel::validate(input, &flatten_input));
 354     input_to_use = &flatten_input;
 355   }
 356   else
 357   {
 358     // Fully Connected layer after a Fully Connected Layer without batches
 359     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != weights_to_use->dimension(1));
 360   }
 361   // Validate matrix multiply kernel
 362   ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(*input_to_use, *weights_to_use, *tmp_output));
 363
 364   // Validate output stage for asymmetric quantized types
 365   if (is_quantized)
 366   {
 367     ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(
 368         &gemmlowp_output, biases, output));
 369   }
 370
 371   return Status{};
 372 }
 373
 374 void NEFullyConnectedLayerEx::run()
 375 {
 376   if (!_is_prepared)
 377   {
 378     if (!_are_weights_reshaped)
 379       _reshape_weights_output.allocator()->allocate();
 380     if (!_are_weights_converted)
 381       _converted_weights_output.allocator()->allocate();
 382     _is_prepared = true;
 383   }
 384
 385   {
 386     ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
 387
 388     // Reshape of the weights
 389     if (!_are_weights_reshaped)
 390     {
 391       _reshape_weights_function.run();
 392     }
 393
 394     // Convert weights if needed
 395     if (!_are_weights_converted)
 396     {
 397       _convert_weights.run();
 398     }
 399
 400     // Prepare GEMM prepare
 401     if (!_is_quantized)
 402     {
 403       _mm_gemm.prepare();
 404     }
 405   }
 406
 407   MemoryGroupResourceScope scope_mg(_memory_group);
 408
 409   // Linearize input if it comes from a convolutional layer
 410   if (_is_fc_after_conv)
 411   {
 412     NEScheduler::get().schedule(&_flatten_kernel, Window::DimY);
 413   }
 414
 415   // Run matrix multiply
 416   if (_is_quantized)
 417   {
 418     _mm_gemmlowp.run();
 419   }
 420   else
 421   {
 422     _mm_gemm.run();
 423   }
 424
 425   // Accumulate biases if provided
 426   if (_is_quantized)
 427   {
 428     _gemmlowp_output_stage.run();
 429   }
 430   else
 431   {
 432     if (_accumulate_biases)
 433     {
 434       NEScheduler::get().schedule(&_accumulate_biases_kernel, Window::DimY);
 435     }
 436   }
 437 }
 438
 439 void NEFullyConnectedLayerEx::prepare()
 440 {
 441 #if 0 // TODO Remove this block
 442   if (!_is_prepared)
 443   {
 444     ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
 445
 446     auto release_unused = [](Tensor *w) {
 447       if (!w->is_used())
 448       {
 449         w->allocator()->free();
 450       }
 451     };
 452
 453     // Pointer to current weights
 454     const ITensor *cur_weights = _original_weights;
 455
 456     // Reshape of the weights (happens only once)
 457     if (!_are_weights_reshaped)
 458     {
 459       // Run reshape weights kernel and mark weights as unused
 460       _reshape_weights_output.allocator()->allocate();
 461       _reshape_weights_function.run();
 462
 463       cur_weights->mark_as_unused();
 464       cur_weights = &_reshape_weights_output;
 465       _are_weights_reshaped = true;
 466     }
 467
 468     // Convert weights if needed (happens only once)
 469     if (!_are_weights_converted)
 470     {
 471       _converted_weights_output.allocator()->allocate();
 472       _convert_weights.run();
 473
 474       cur_weights->mark_as_unused();
 475       _are_weights_converted = true;
 476     }
 477
 478     // Release reshaped weights if unused
 479     release_unused(&_reshape_weights_output);
 480
 481     // Prepare GEMM prepare and release unused weights
 482     if (!_is_quantized)
 483     {
 484       _mm_gemm.prepare();
 485     }
 486
 487     // Release converted weights if unused
 488     release_unused(&_reshape_weights_output);
 489     release_unused(&_converted_weights_output);
 490
 491     _is_prepared = true;
 492   }
 493 #endif
 494 }