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(
  78       NEGEMM::validate(&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(), _converted_weights_output(),
  90     _reshape_weights_output(), _original_weights(nullptr), _are_weights_converted(true),
  91     _are_weights_reshaped(false), _is_fc_after_conv(false), _accumulate_biases(false),
  92     _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(shape_flatten));
 142
 143   // Configure flatten kernel
 144   _memory_group.manage(&_flatten_output);
 145   _flatten_kernel.configure(input, &_flatten_output);
 146
 147   // Configure matrix multiply kernel
 148   configure_mm(&_flatten_output, weights, output);
 149
 150   // Allocate the output tensor for flatten once all the configure methods have been called
 151   _flatten_output.allocator()->allocate();
 152 }
 153
 154 void NEFullyConnectedLayerEx::configure_fc_fc(const ITensor *input, const ITensor *weights,
 155                                               ITensor *output)
 156 {
 157   ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1));
 158
 159   // Configure matrix multiply kernel
 160   configure_mm(input, weights, output);
 161 }
 162
 163 void NEFullyConnectedLayerEx::configure(const ITensor *input, const ITensor *weights,
 164                                         const ITensor *biases, ITensor *output,
 165                                         FullyConnectedLayerInfo fc_info)
 166 {
 167   ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
 168
 169   // Perform validate step
 170   ARM_COMPUTE_ERROR_THROW_ON(NEFullyConnectedLayerEx::validate(
 171     input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(),
 172     fc_info));
 173
 174   _are_weights_converted = true;
 175   _are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
 176   _is_fc_after_conv = true;
 177   _accumulate_biases = false;
 178   _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
 179   _original_weights = weights;
 180
 181   // Configure gemmlowp output
 182   if (_is_quantized)
 183   {
 184     _gemmlowp_output.allocator()->init(
 185       output->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
 186   }
 187
 188   // Configure accumulate biases kernel for non quantized asymmetric types
 189   if (biases != nullptr && !_is_quantized)
 190   {
 191     _accumulate_biases = true;
 192
 193     // Configure accumulate biases kernel
 194     _accumulate_biases_kernel.configure(output, biases);
 195   }
 196
 197   // With the Fully Connected layer we can have 4 different cases:
 198   //  1) Convolution layer -> Fully Connected layer without batches
 199   //  2) Fully Connected layer -> Fully Connected layer without batches
 200   //  3) Convolution layer -> Fully Connected layer with batches
 201   //  4) Fully Connected layer -> Fully Connected layer with batches
 202
 203   const ITensor *weights_to_use = weights;
 204
 205   // Check if we have a fully connected layer with batches
 206   const bool is_batched_fc_layer = output->info()->dimension(1) > 1;
 207   if (is_batched_fc_layer)
 208   {
 209     _is_fc_after_conv =
 210       (TensorShape::num_max_dimensions >= 4) &&
 211       (std::equal(input->info()->tensor_shape().cbegin() + 3, input->info()->tensor_shape().cend(),
 212                   output->info()->tensor_shape().cbegin() + 1));
 213   }
 214   else
 215   {
 216     _is_fc_after_conv = input->info()->num_dimensions() > 1;
 217   }
 218
 219   // Reshape weights if needed
 220   if (!_are_weights_reshaped)
 221   {
 222     // Reshape the weights
 223     _reshape_weights_function.configure(weights, &_reshape_weights_output);
 224     weights_to_use = &_reshape_weights_output;
 225   }
 226
 227   // Convert weights if needed
 228   if (_is_fc_after_conv && (input->info()->data_layout() != fc_info.weights_trained_layout))
 229   {
 230     // Convert weights
 231     _convert_weights.configure(weights_to_use, &_converted_weights_output,
 232                                input->info()->tensor_shape(), fc_info.weights_trained_layout);
 233
 234     weights_to_use = &_converted_weights_output;
 235     _are_weights_converted = false;
 236   }
 237
 238   ITensor *tmp_output = (_is_quantized) ? &_gemmlowp_output : output;
 239   if (_is_fc_after_conv)
 240   {
 241     // Fully Connected layer after a Convolution Layer without batches
 242     configure_conv_fc(input, weights_to_use, tmp_output);
 243   }
 244   else
 245   {
 246     // Fully Connected layer after a Fully Connected Layer without batches
 247     configure_fc_fc(input, weights_to_use, tmp_output);
 248   }
 249
 250   // Configure output stage for asymmetric quantized types
 251   if (_is_quantized)
 252   {
 253     float multiplier = input->info()->quantization_info().uniform().scale *
 254                        weights->info()->quantization_info().uniform().scale /
 255                        output->info()->quantization_info().uniform().scale;
 256     int output_multiplier;
 257     int output_shift;
 258     quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier,
 259                                                                &output_shift);
 260     _gemmlowp_output_stage.configure(&_gemmlowp_output, biases, output, output_multiplier,
 261                                      output_shift,
 262                                      output->info()->quantization_info().uniform().offset);
 263     _gemmlowp_output.allocator()->allocate();
 264   }
 265
 266   _are_weights_reshaped = _are_weights_reshaped || fc_info.retain_internal_weights;
 267 }
 268
 269 Status NEFullyConnectedLayerEx::validate(const ITensorInfo *input, const ITensorInfo *weights,
 270                                          const ITensorInfo *biases, const ITensorInfo *output,
 271                                          FullyConnectedLayerInfo fc_info)
 272 {
 273   ARM_COMPUTE_UNUSED(fc_info.retain_internal_weights);
 274   ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
 275   ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16,
 276                                                        DataType::F32);
 277   ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
 278   ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2);
 279
 280   bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
 281   bool is_fc_after_conv = true;
 282   bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
 283
 284   const ITensorInfo &flatten_input =
 285     TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(
 286       compute_flatten_shape(input)));
 287   const ITensorInfo &reshaped_weights =
 288     TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(
 289       compute_transposed_shape(*weights)));
 290   const ITensorInfo &converted_weights =
 291     weights_reshaped ? TensorInfo(weights->clone()->set_is_resizable(true).reset_padding())
 292                      : TensorInfo(*reshaped_weights.clone());
 293   const ITensorInfo &gemmlowp_output = TensorInfo(
 294     output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
 295
 296   // Configure accumulate biases kernel for non quantized asymmetric types
 297   if (biases != nullptr && !is_quantized)
 298   {
 299     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
 300     ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixAccumulateBiasesKernel::validate(output, biases));
 301   }
 302
 303   // With the Fully Connected layer we can have 4 different cases:
 304   //  1) Convolution layer -> Fully Connected layer without batches
 305   //  2) Fully Connected layer -> Fully Connected layer without batches
 306   //  3) Convolution layer -> Fully Connected layer with batches
 307   //  4) Fully Connected layer -> Fully Connected layer with batches
 308
 309   const ITensorInfo *input_to_use = input;
 310   const ITensorInfo *weights_to_use = weights;
 311   const ITensorInfo *tmp_output = (is_quantized) ? &gemmlowp_output : output;
 312
 313   // Check if we have a fully connected layer with batches
 314   const bool is_batched_fc_layer = output->dimension(1) > 1;
 315
 316   if (is_batched_fc_layer)
 317   {
 318     is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) &&
 319                        (std::equal(input->tensor_shape().cbegin() + 3, input->tensor_shape().cend(),
 320                                    output->tensor_shape().cbegin() + 1));
 321   }
 322   else
 323   {
 324     is_fc_after_conv = input->num_dimensions() > 1;
 325   }
 326
 327   if (!weights_reshaped)
 328   {
 329     // Validate reshape weights kernel
 330     ARM_COMPUTE_RETURN_ON_ERROR(
 331       NEFullyConnectedLayerReshapeWeights::validate(weights, &reshaped_weights));
 332     weights_to_use = &reshaped_weights;
 333   }
 334
 335   if (is_fc_after_conv && (input->data_layout() != fc_info.weights_trained_layout))
 336   {
 337     // Validate convert weights kernel
 338     ARM_COMPUTE_RETURN_ON_ERROR(NEConvertFullyConnectedWeights::validate(
 339       weights_to_use, &converted_weights, input->tensor_shape(), fc_info.weights_trained_layout));
 340     weights_to_use = &converted_weights;
 341   }
 342
 343   if (is_fc_after_conv)
 344   {
 345     // Fully Connected layer after a Convolution Layer without batches
 346     ARM_COMPUTE_RETURN_ERROR_ON(
 347       (weights_to_use->dimension(1) !=
 348        (input->dimension(0) * input->dimension(1) * input->dimension(2))));
 349
 350     // Validate flatten kernel
 351     ARM_COMPUTE_RETURN_ON_ERROR(NEFlattenLayerKernel::validate(input, &flatten_input));
 352     input_to_use = &flatten_input;
 353   }
 354   else
 355   {
 356     // Fully Connected layer after a Fully Connected Layer without batches
 357     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != weights_to_use->dimension(1));
 358   }
 359   // Validate matrix multiply kernel
 360   ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(*input_to_use, *weights_to_use, *tmp_output));
 361
 362   // Validate output stage for asymmetric quantized types
 363   if (is_quantized)
 364   {
 365     ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(
 366       &gemmlowp_output, biases, output));
 367   }
 368
 369   return Status{};
 370 }
 371
 372 void NEFullyConnectedLayerEx::run()
 373 {
 374   if (!_is_prepared)
 375   {
 376     if (!_are_weights_reshaped)
 377       _reshape_weights_output.allocator()->allocate();
 378     if (!_are_weights_converted)
 379       _converted_weights_output.allocator()->allocate();
 380     _is_prepared = true;
 381   }
 382
 383   {
 384     ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
 385
 386     // Reshape of the weights
 387     if (!_are_weights_reshaped)
 388     {
 389       _reshape_weights_function.run();
 390     }
 391
 392     // Convert weights if needed
 393     if (!_are_weights_converted)
 394     {
 395       _convert_weights.run();
 396     }
 397
 398     // Prepare GEMM prepare
 399     if (!_is_quantized)
 400     {
 401       _mm_gemm.prepare();
 402     }
 403   }
 404
 405   MemoryGroupResourceScope scope_mg(_memory_group);
 406
 407   // Linearize input if it comes from a convolutional layer
 408   if (_is_fc_after_conv)
 409   {
 410     NEScheduler::get().schedule(&_flatten_kernel, Window::DimY);
 411   }
 412
 413   // Run matrix multiply
 414   if (_is_quantized)
 415   {
 416     _mm_gemmlowp.run();
 417   }
 418   else
 419   {
 420     _mm_gemm.run();
 421   }
 422
 423   // Accumulate biases if provided
 424   if (_is_quantized)
 425   {
 426     _gemmlowp_output_stage.run();
 427   }
 428   else
 429   {
 430     if (_accumulate_biases)
 431     {
 432       NEScheduler::get().schedule(&_accumulate_biases_kernel, Window::DimY);
 433     }
 434   }
 435 }
 436
 437 void NEFullyConnectedLayerEx::prepare()
 438 {
 439 #if 0 // TODO Remove this block
 440   if (!_is_prepared)
 441   {
 442     ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
 443
 444     auto release_unused = [](Tensor *w) {
 445       if (!w->is_used())
 446       {
 447         w->allocator()->free();
 448       }
 449     };
 450
 451     // Pointer to current weights
 452     const ITensor *cur_weights = _original_weights;
 453
 454     // Reshape of the weights (happens only once)
 455     if (!_are_weights_reshaped)
 456     {
 457       // Run reshape weights kernel and mark weights as unused
 458       _reshape_weights_output.allocator()->allocate();
 459       _reshape_weights_function.run();
 460
 461       cur_weights->mark_as_unused();
 462       cur_weights = &_reshape_weights_output;
 463       _are_weights_reshaped = true;
 464     }
 465
 466     // Convert weights if needed (happens only once)
 467     if (!_are_weights_converted)
 468     {
 469       _converted_weights_output.allocator()->allocate();
 470       _convert_weights.run();
 471
 472       cur_weights->mark_as_unused();
 473       _are_weights_converted = true;
 474     }
 475
 476     // Release reshaped weights if unused
 477     release_unused(&_reshape_weights_output);
 478
 479     // Prepare GEMM prepare and release unused weights
 480     if (!_is_quantized)
 481     {
 482       _mm_gemm.prepare();
 483     }
 484
 485     // Release converted weights if unused
 486     release_unused(&_reshape_weights_output);
 487     release_unused(&_converted_weights_output);
 488
 489     _is_prepared = true;
 490   }
 491 #endif
 492 }