compute/ARMComputeEx/src/runtime/NEON/functions/NEFullyConnectedHybridLayer.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/NEFullyConnectedHybridLayer.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   ARM_COMPUTE_RETURN_ON_ERROR(
  61       NEGEMMLowpMatrixMultiplyCore::validate(&input, &weights, nullptr, &output));
  62
  63   return Status{};
  64 }
  65 } // namespace
  66
  67 void NEFullyConnectedHybridLayerReshapeWeights::configure(const ITensor *input, ITensor *output)
  68 {
  69   auto k = support::cpp14::make_unique<NETransposeKernel>();
  70   k->configure(input, output);
  71   _kernel = std::move(k);
  72 }
  73
  74 Status NEFullyConnectedHybridLayerReshapeWeights::validate(const ITensorInfo *input,
  75                                                            const ITensorInfo *output)
  76 {
  77   return NETransposeKernel::validate(input, output);
  78 }
  79
  80 NEFullyConnectedHybridLayer::NEFullyConnectedHybridLayer(
  81     std::shared_ptr<IMemoryManager> memory_manager)
  82     : _memory_group(std::move(memory_manager)), _reshape_weights_function(), _quant_input_kernel(),
  83       _mm_gemmlowp(), _accumulate_biases_kernel(), _reshape_weights_output(), _quantized_input(),
  84       _scale_factor(), _original_weights(nullptr), _are_weights_reshaped(false),
  85       _accumulate_biases(false), _is_prepared(false)
  86 {
  87 }
  88
  89 void NEFullyConnectedHybridLayer::configure_mm(const ITensor *input, const ITensor *weights,
  90                                                ITensor *output)
  91 {
  92   ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1));
  93
  94   // Configure gemmlowp function
  95   _mm_gemmlowp.configure(input, weights, nullptr, output);
  96 }
  97
  98 void NEFullyConnectedHybridLayer::configure(const ITensor *input, const ITensor *weights,
  99                                             const ITensor *biases, ITensor *output,
 100                                             FullyConnectedLayerInfo fc_info)
 101 {
 102   ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
 103
 104   // Perform validate step
 105   ARM_COMPUTE_ERROR_THROW_ON(NEFullyConnectedHybridLayer::validate(
 106       input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(),
 107       fc_info));
 108
 109   _are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
 110   _accumulate_biases = false;
 111   _original_weights = weights;
 112
 113   // Configure accumulate biases kernel for non quantized asymmetric types
 114   if (biases != nullptr)
 115   {
 116     _accumulate_biases = true;
 117
 118     // Configure accumulate biases kernel
 119     _accumulate_biases_kernel.configure(output, biases);
 120   }
 121
 122   // With the Fully Connected layer we can have 4 different cases:
 123   //  1) Convolution layer -> Fully Connected layer without batches
 124   //  2) Fully Connected layer -> Fully Connected layer without batches
 125   //  3) Convolution layer -> Fully Connected layer with batches
 126   //  4) Fully Connected layer -> Fully Connected layer with batches
 127
 128   const ITensor *weights_to_use = weights;
 129
 130   // Check if we have a fully connected layer with batches
 131   const bool is_batched_fc_layer = output->info()->dimension(1) > 1;
 132   bool _is_fc_after_conv;
 133   if (is_batched_fc_layer)
 134   {
 135     _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) &&
 136                         (std::equal(input->info()->tensor_shape().cbegin() + 3,
 137                                     input->info()->tensor_shape().cend(),
 138                                     output->info()->tensor_shape().cbegin() + 1));
 139   }
 140   else
 141   {
 142     _is_fc_after_conv = input->info()->num_dimensions() > 1 && input->info()->dimension(1) > 1;
 143   }
 144   ARM_COMPUTE_ERROR_ON_MSG(_is_fc_after_conv,
 145                            "NEFullyConnectedHybridLayer does not support after conv");
 146   (void)_is_fc_after_conv;
 147
 148   // Reshape weights if needed
 149   if (!_are_weights_reshaped)
 150   {
 151     // Reshape the weights
 152     _reshape_weights_output.allocator()->init(
 153         weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(
 154             compute_transposed_shape(*weights->info())));
 155     _reshape_weights_function.configure(weights_to_use, &_reshape_weights_output);
 156     weights_to_use = &_reshape_weights_output;
 157   }
 158
 159   // Quantize input
 160   _quantized_input.allocator()->init(
 161       input->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(
 162           DataType::QASYMM8_SIGNED));
 163   _scale_factor.allocator()->init(
 164       TensorInfo(TensorShape{output->info()->dimension(1)}, 1, DataType::F32));
 165   _quant_input_kernel.configure(input, &_quantized_input, &_scale_factor);
 166
 167   // GEMM
 168   _gemmlowp_output.allocator()->init(
 169       output->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
 170   configure_mm(&_quantized_input, weights_to_use, &_gemmlowp_output);
 171
 172   // Multiply scale
 173   _multiply_scale_kernel.configure(&_gemmlowp_output, &_scale_factor, output,
 174                                    weights->info()->quantization_info().uniform().scale);
 175
 176   _are_weights_reshaped = _are_weights_reshaped || fc_info.retain_internal_weights;
 177
 178   _quantized_input.allocator()->allocate();
 179   _scale_factor.allocator()->allocate();
 180   _gemmlowp_output.allocator()->allocate();
 181 }
 182
 183 Status NEFullyConnectedHybridLayer::validate(const ITensorInfo *input, const ITensorInfo *weights,
 184                                              const ITensorInfo *biases, const ITensorInfo *output,
 185                                              FullyConnectedLayerInfo fc_info)
 186 {
 187   ARM_COMPUTE_UNUSED(fc_info.retain_internal_weights);
 188   ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
 189   ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
 190   ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QASYMM8_SIGNED);
 191   ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
 192   ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2);
 193   ARM_COMPUTE_RETURN_ERROR_ON(output->num_dimensions() > 2);
 194
 195   bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
 196
 197   const ITensorInfo &reshaped_weights =
 198       TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(
 199           compute_transposed_shape(*weights)));
 200
 201   // Configure accumulate biases kernel for non quantized asymmetric types
 202   if (biases != nullptr)
 203   {
 204     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
 205     ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixAccumulateBiasesKernel::validate(output, biases));
 206   }
 207
 208   // With the Fully Connected layer we can have 4 different cases:
 209   //  1) Convolution layer -> Fully Connected layer without batches
 210   //  2) Fully Connected layer -> Fully Connected layer without batches
 211   //  3) Convolution layer -> Fully Connected layer with batches
 212   //  4) Fully Connected layer -> Fully Connected layer with batches
 213
 214   const ITensorInfo *weights_to_use = weights;
 215
 216   if (!weights_reshaped)
 217   {
 218     // Validate reshape weights kernel
 219     ARM_COMPUTE_RETURN_ON_ERROR(
 220         NEFullyConnectedHybridLayerReshapeWeights::validate(weights_to_use, &reshaped_weights));
 221     weights_to_use = &reshaped_weights;
 222   }
 223
 224   // Fully Connected layer after a Fully Connected Layer without batches
 225   ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != weights_to_use->dimension(1));
 226
 227   // Validate quantization kernel
 228   const ITensorInfo &quantized_input =
 229       TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_data_type(
 230           DataType::QASYMM8_SIGNED));
 231   const ITensorInfo &scale_factor = TensorInfo(TensorShape{output->dimension(1)}, 1, DataType::F32);
 232   ARM_COMPUTE_RETURN_ON_ERROR(
 233       NEQuantizationSymmetricKernel::validate(input, &quantized_input, &scale_factor));
 234
 235   const ITensorInfo &gemmlowp_output = TensorInfo(
 236       output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
 237   // Validate matrix multiply kernel
 238   ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(quantized_input, *weights_to_use, gemmlowp_output));
 239
 240   ARM_COMPUTE_RETURN_ON_ERROR(NEMultiplyScaleFactorKernel::validate(
 241       &gemmlowp_output, &scale_factor, output, weights->quantization_info().uniform().scale));
 242
 243   return Status{};
 244 }
 245
 246 void NEFullyConnectedHybridLayer::run()
 247 {
 248   prepare();
 249
 250   MemoryGroupResourceScope scope_mg(_memory_group);
 251
 252   // Quantize input
 253   NEScheduler::get().schedule(&_quant_input_kernel, Window::DimY);
 254
 255   // Run matrix multiply
 256   _mm_gemmlowp.run();
 257
 258   // Multiply scale factor
 259   NEScheduler::get().schedule(&_multiply_scale_kernel, Window::DimY);
 260
 261   // Accumulate biases if provided
 262   if (_accumulate_biases)
 263   {
 264     NEScheduler::get().schedule(&_accumulate_biases_kernel, Window::DimY);
 265   }
 266 }
 267
 268 void NEFullyConnectedHybridLayer::prepare()
 269 {
 270   if (!_is_prepared)
 271   {
 272     ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
 273
 274     auto release_unused = [](Tensor *w) {
 275       if (!w->is_used())
 276       {
 277         w->allocator()->free();
 278       }
 279     };
 280
 281     // Reshape of the weights (happens only once)
 282     if (!_are_weights_reshaped)
 283     {
 284       // Run reshape weights kernel and mark weights as unused
 285       _reshape_weights_output.allocator()->allocate();
 286       _reshape_weights_function.run();
 287
 288       _are_weights_reshaped = true;
 289       // We can not release _original_weights because it can be used in other nodes
 290     }
 291
 292     // Prepare GEMM prepare and release unused weights
 293     _mm_gemmlowp.prepare();
 294
 295     // Release reshaped weights if unused
 296     release_unused(&_reshape_weights_output);
 297
 298     _is_prepared = true;
 299   }
 300 }