compute/ARMComputeEx/src/runtime/CL/functions/CLFullyConnectedHybridLayer.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/CL/functions/CLFullyConnectedHybridLayer.h"
  42
  43 #include "arm_compute/core/Size2D.h"
  44 #include "arm_compute/core/Validate.h"
  45 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
  46 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
  47 #include "arm_compute/runtime/CL/CLScheduler.h"
  48 #include "support/MemorySupport.h"
  49
  50 #include <algorithm>
  51
  52 using namespace arm_compute;
  53 using namespace arm_compute::misc::shape_calculator;
  54
  55 namespace
  56 {
  57 Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const ITensorInfo &output)
  58 {
  59   ARM_COMPUTE_UNUSED(input);
  60   ARM_COMPUTE_UNUSED(weights);
  61   ARM_COMPUTE_UNUSED(output);
  62   ARM_COMPUTE_RETURN_ON_ERROR(
  63       CLGEMMLowpMatrixMultiplyCore::validate(&input, &weights, nullptr, &output));
  64
  65   return Status{};
  66 }
  67 } // namespace
  68
  69 void CLFullyConnectedHybridLayerReshapeWeights::configure(const ICLTensor *input, ICLTensor *output)
  70 {
  71   auto k = support::cpp14::make_unique<CLTransposeKernel>();
  72   k->configure(input, output);
  73   _kernel = std::move(k);
  74 }
  75
  76 Status CLFullyConnectedHybridLayerReshapeWeights::validate(const ITensorInfo *input,
  77                                                            const ITensorInfo *output)
  78 {
  79   return CLTransposeKernel::validate(input, output);
  80 }
  81
  82 CLFullyConnectedHybridLayer::CLFullyConnectedHybridLayer(
  83     std::shared_ptr<IMemoryManager> memory_manager)
  84     : _memory_group(memory_manager), _reshape_weights_kernel(), _quant_input_kernel(),
  85       _mm_gemmlowp(memory_manager), _multiply_scale_kernel(), _accumulate_biases_kernel(),
  86       _reshape_weights_output(), _quantized_input(), _scale_factor(), _gemmlowp_output(),
  87       _are_weights_reshaped(true), _accumulate_biases(false), _is_prepared(false),
  88       _original_weights(nullptr)
  89 {
  90 }
  91 void CLFullyConnectedHybridLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights,
  92                                                ICLTensor *output, bool retain_internal_weights)
  93 {
  94   ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1));
  95
  96   ARM_COMPUTE_UNUSED(output);
  97   ARM_COMPUTE_UNUSED(retain_internal_weights);
  98   // Configure gemmlowp function
  99   _mm_gemmlowp.configure(input, weights, nullptr, output);
 100 }
 101
 102 void CLFullyConnectedHybridLayer::configure(const ICLTensor *input, const ICLTensor *weights,
 103                                             const ICLTensor *biases, ICLTensor *output,
 104                                             FullyConnectedLayerInfo fc_info)
 105 {
 106   ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
 107
 108   // Perform validate step
 109   ARM_COMPUTE_ERROR_THROW_ON(CLFullyConnectedHybridLayer::validate(
 110       input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(),
 111       fc_info));
 112
 113   _are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
 114   _accumulate_biases = false;
 115   _is_prepared = fc_info.retain_internal_weights;
 116   _original_weights = weights;
 117
 118   // Configure accumulate biases kernel for non quantized asymmetric types
 119   if (biases != nullptr)
 120   {
 121     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
 122
 123     _accumulate_biases = true;
 124
 125     // Configure accumulate biases kernel
 126     _accumulate_biases_kernel.set_target(CLScheduler::get().target());
 127     _accumulate_biases_kernel.configure(output, biases);
 128   }
 129
 130   const ICLTensor *weights_to_use = weights;
 131
 132   // With the Fully Connected layer we can have 4 different cases:
 133   //  1) Convolution layer -> Fully Connected layer without batches
 134   //  2) Fully Connected layer -> Fully Connected layer without batches
 135   //  3) Convolution layer -> Fully Connected layer with batches
 136   //  4) Fully Connected layer -> Fully Connected layer with batches
 137
 138   // Check if we have a fully connected layer with batches
 139   const bool is_batched_fc_layer = output->info()->dimension(1) > 1;
 140   bool is_fc_after_conv = false;
 141   if (is_batched_fc_layer)
 142   {
 143     is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) &&
 144                        (std::equal(input->info()->tensor_shape().cbegin() + 3,
 145                                    input->info()->tensor_shape().cend(),
 146                                    output->info()->tensor_shape().cbegin() + 1));
 147   }
 148   else
 149   {
 150     is_fc_after_conv = input->info()->num_dimensions() > 1 && input->info()->dimension(1) > 1;
 151   }
 152   ARM_COMPUTE_ERROR_ON_MSG(is_fc_after_conv,
 153                            "CLFullyConnectedHybridLayer does not support after conv");
 154   ARM_COMPUTE_UNUSED(is_fc_after_conv);
 155
 156   // Reshape weights if needed
 157   if (!_are_weights_reshaped)
 158   {
 159     // Reshape the weights
 160     _reshape_weights_output.allocator()->init(
 161         weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(
 162             compute_transposed_shape(*weights->info())));
 163     _reshape_weights_kernel.configure(weights_to_use, &_reshape_weights_output);
 164     weights_to_use = &_reshape_weights_output;
 165   }
 166
 167   // Extract scale factor
 168   _scale_factor.allocator()->init(
 169       TensorInfo(TensorShape{output->info()->dimension(1)}, 1, input->info()->data_type()));
 170   _memory_group.manage(&_scale_factor);
 171   _scale_factor_kernel.configure(input, &_scale_factor);
 172
 173   // Quantize input
 174   _quantized_input.allocator()->init(
 175       input->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(
 176           DataType::QASYMM8_SIGNED));
 177   _memory_group.manage(&_quantized_input);
 178   _quant_input_kernel.configure(input, &_scale_factor, &_quantized_input);
 179
 180   // GEMMLowp
 181   _gemmlowp_output.allocator()->init(
 182       output->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
 183   _memory_group.manage(&_gemmlowp_output);
 184   configure_mm(&_quantized_input, weights_to_use, &_gemmlowp_output,
 185                fc_info.retain_internal_weights);
 186   _quantized_input.allocator()->allocate();
 187
 188   // Multiply scale
 189   _multiply_scale_kernel.configure(&_gemmlowp_output, &_scale_factor, output,
 190                                    weights->info()->quantization_info().uniform().scale);
 191   _gemmlowp_output.allocator()->allocate();
 192   _scale_factor.allocator()->allocate();
 193
 194   _are_weights_reshaped = _are_weights_reshaped || fc_info.retain_internal_weights;
 195 }
 196
 197 Status CLFullyConnectedHybridLayer::validate(const ITensorInfo *input, const ITensorInfo *weights,
 198                                              const ITensorInfo *biases, const ITensorInfo *output,
 199                                              FullyConnectedLayerInfo fc_info)
 200 {
 201   ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
 202   ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
 203   ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QASYMM8_SIGNED);
 204   ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
 205   ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2);
 206
 207   bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
 208   bool is_fc_after_conv = true;
 209   const GPUTarget gpu_target = CLScheduler::get().target();
 210
 211   const ITensorInfo &reshaped_weights =
 212       TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(
 213           compute_transposed_shape(*weights)));
 214
 215   // Configure accumulate biases kernel for non quantized asymmetric types
 216   if (biases != nullptr)
 217   {
 218     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
 219     ARM_COMPUTE_RETURN_ON_ERROR(
 220         CLGEMMMatrixAccumulateBiasesKernel::validate(output, biases, gpu_target));
 221   }
 222
 223   // With the Fully Connected layer we can have 4 different cases:
 224   //  1) Convolution layer -> Fully Connected layer without batches
 225   //  2) Fully Connected layer -> Fully Connected layer without batches
 226   //  3) Convolution layer -> Fully Connected layer with batches
 227   //  4) Fully Connected layer -> Fully Connected layer with batches
 228
 229   const ITensorInfo *weights_to_use = weights;
 230
 231   // Check if we have a fully connected layer with batches
 232   const bool is_batched_fc_layer = output->dimension(1) > 1;
 233   if (is_batched_fc_layer)
 234   {
 235     is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) &&
 236                        (std::equal(input->tensor_shape().cbegin() + 3, input->tensor_shape().cend(),
 237                                    output->tensor_shape().cbegin() + 1));
 238   }
 239   else
 240   {
 241     is_fc_after_conv = input->num_dimensions() > 1 && input->dimension(1) > 1;
 242   }
 243   ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_fc_after_conv,
 244                                   "CLFullyConnectedHybridLayer does not support after conv");
 245
 246   if (!weights_reshaped)
 247   {
 248     // Validate reshape weights kernel
 249     ARM_COMPUTE_RETURN_ON_ERROR(
 250         CLFullyConnectedHybridLayerReshapeWeights::validate(weights_to_use, &reshaped_weights));
 251     weights_to_use = &reshaped_weights;
 252   }
 253
 254   // Validate Scale factor kernel
 255   const ITensorInfo &scale_factor =
 256       TensorInfo(TensorShape{output->dimension(1)}, 1, input->data_type());
 257   ARM_COMPUTE_RETURN_ON_ERROR(CLScaleFactorSymm8Kernel::validate(input, &scale_factor));
 258
 259   // Validate quantization symm8 kernel
 260   const ITensorInfo &quantized_input =
 261       TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_data_type(
 262           DataType::QASYMM8_SIGNED));
 263   ARM_COMPUTE_RETURN_ON_ERROR(
 264       CLQuantizationSymmetricKernel::validate(input, &scale_factor, &quantized_input));
 265
 266   // Fully Connected layer after a Fully Connected Layer without batches
 267   ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != weights_to_use->dimension(1));
 268
 269   // Validate matrix multiply kernel
 270   const ITensorInfo &gemmlowp_output = TensorInfo(
 271       output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32));
 272   ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(quantized_input, *weights_to_use, gemmlowp_output));
 273
 274   // Multiply scale
 275   ARM_COMPUTE_RETURN_ON_ERROR(
 276       CLMultiplyScaleFactorKernel::validate(&gemmlowp_output, &scale_factor, output));
 277
 278   return Status{};
 279 }
 280
 281 void CLFullyConnectedHybridLayer::run()
 282 {
 283   prepare();
 284
 285   MemoryGroupResourceScope scope_mg(_memory_group);
 286
 287   // Extract scale_factor
 288   CLScheduler::get().enqueue(_scale_factor_kernel);
 289
 290   // Quantize input
 291   CLScheduler::get().enqueue(_quant_input_kernel);
 292
 293   // Run matrix multiply
 294   _mm_gemmlowp.run();
 295
 296   // Multiply scale factor
 297   CLScheduler::get().enqueue(_multiply_scale_kernel);
 298
 299   // Accumulate biases if provided
 300   if (_accumulate_biases)
 301   {
 302     CLScheduler::get().enqueue(_accumulate_biases_kernel);
 303   }
 304 }
 305
 306 void CLFullyConnectedHybridLayer::prepare()
 307 {
 308   if (!_is_prepared)
 309   {
 310     ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
 311
 312     auto release_unused = [](CLTensor *w) {
 313       if (!w->is_used())
 314       {
 315         CLScheduler::get().queue().finish();
 316         w->allocator()->free();
 317       }
 318     };
 319
 320     // Reshape of the weights if needed (happens only once)
 321     if (!_are_weights_reshaped)
 322     {
 323       // Run reshape weights kernel and mark weights as unused
 324       _reshape_weights_output.allocator()->allocate();
 325       _reshape_weights_kernel.run();
 326
 327       _are_weights_reshaped = true;
 328       // We can not release _original_weights because it can be used in other nodes
 329     }
 330
 331     // Prepare GEMM prepare and release unused weights
 332     _mm_gemmlowp.prepare();
 333
 334     // Release reshaped weights if unused
 335     release_unused(&_reshape_weights_output);
 336
 337     _is_prepared = true;
 338   }
 339 }