src/runtime/CL/functions/CLFullyConnectedLayer.cpp

   1 /*
   2  * Copyright (c) 2017 ARM Limited.
   3  *
   4  * SPDX-License-Identifier: MIT
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a copy
   7  * of this software and associated documentation files (the "Software"), to
   8  * deal in the Software without restriction, including without limitation the
   9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
  10  * sell copies of the Software, and to permit persons to whom the Software is
  11  * furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in all
  14  * copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  */
  24 #include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h"
  25
  26 #include "arm_compute/core/Validate.h"
  27 #include "arm_compute/runtime/CL/CLScheduler.h"
  28
  29 #include <algorithm>
  30 #include <cmath>
  31
  32 using namespace arm_compute;
  33
  34 CLFullyConnectedLayer::CLFullyConnectedLayer()
  35     : _im2col_kernel(), _transpose_kernel(), _transpose1xW_kernel(), _interleave4x4_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(), _interleave4x4_output(), _transpose_output(),
  36       _transpose1xW_output(), _is_first_run(true), _transpose_weights(true), _fc_after_conv(true), _batched_fc_layer(false), _accumulate_biases(false)
  37 {
  38 }
  39
  40 void CLFullyConnectedLayer::configure_conv_fc_wb(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output)
  41 {
  42     ARM_COMPUTE_ERROR_ON(weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2)));
  43
  44     // If the fully connected layer is called after a convolution layer, the input tensor must be linearized
  45
  46     // Initialize output tensor for im2col
  47     TensorShape shape_im2col;
  48     shape_im2col.set(0, weights->info()->dimension(1));
  49     shape_im2col.set(1, input->info()->dimension(3));
  50     shape_im2col.set(2, input->info()->dimension(4));
  51     shape_im2col.set(3, input->info()->dimension(5));
  52     _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, input->info()->data_type()));
  53
  54     // Initialize output tensor for interleave 4x4
  55     TensorShape shape_interleaved = _im2col_output.info()->tensor_shape();
  56     shape_interleaved.set(0, shape_interleaved.x() * 4);
  57     shape_interleaved.set(1, std::ceil(static_cast<float>(shape_interleaved.y()) / 4));
  58     _interleave4x4_output.allocator()->init(TensorInfo(shape_interleaved, 1, input->info()->data_type()));
  59
  60     // Initialize output tensor for transpose 1xW
  61     TensorShape shape_transposed1xW(weights->info()->dimension(1) * 4, static_cast<size_t>(std::ceil(weights->info()->dimension(0) / 4.f)));
  62     _transpose1xW_output.allocator()->init(TensorInfo(shape_transposed1xW, 1, weights->info()->data_type()));
  63
  64     // Configure im2col kernel
  65     _im2col_kernel.configure(input, &_im2col_output, std::make_pair(1, 1), PadStrideInfo(1, 1, 0, 0), false);
  66
  67     // Configure interleave4x4 kernel
  68     _interleave4x4_kernel.configure(&_im2col_output, &_interleave4x4_output);
  69
  70     // Configure transpose 1xW kernel
  71     _transpose1xW_kernel.configure(weights, &_transpose1xW_output);
  72
  73     // Configure matrix multiply kernel
  74     _mm_kernel.configure(&_interleave4x4_output, &_transpose1xW_output, output, 1.0f);
  75
  76     // Allocate the tensors once all the configure methods have been called
  77     _im2col_output.allocator()->allocate();
  78     _interleave4x4_output.allocator()->allocate();
  79     _transpose1xW_output.allocator()->allocate();
  80 }
  81
  82 void CLFullyConnectedLayer::configure_fc_fc_wb(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output)
  83 {
  84     // Initialize output tensor for interleave 4x4
  85     TensorShape shape_interleaved = input->info()->tensor_shape();
  86     shape_interleaved.set(0, shape_interleaved.x() * 4);
  87     shape_interleaved.set(1, std::ceil(static_cast<float>(shape_interleaved.y()) / 4));
  88     _interleave4x4_output.allocator()->init(TensorInfo(shape_interleaved, 1, input->info()->data_type()));
  89
  90     // Initialize output tensor for transpose 1xW
  91     TensorShape shape_transposed1xW(weights->info()->dimension(1) * 4, static_cast<size_t>(std::ceil(weights->info()->dimension(0) / 4.f)));
  92     _transpose1xW_output.allocator()->init(TensorInfo(shape_transposed1xW, 1, weights->info()->data_type()));
  93
  94     // Configure interleave4x4 kernel
  95     _interleave4x4_kernel.configure(input, &_interleave4x4_output);
  96
  97     // Configure transpose 1xW kernel
  98     _transpose1xW_kernel.configure(weights, &_transpose1xW_output);
  99
 100     // Configure matrix multiply kernel
 101     _mm_kernel.configure(&_interleave4x4_output, &_transpose1xW_output, output, 1.0f);
 102
 103     // Allocate the tensors once all the configure methods have been called
 104     _interleave4x4_output.allocator()->allocate();
 105     _transpose1xW_output.allocator()->allocate();
 106 }
 107
 108 void CLFullyConnectedLayer::configure_conv_fc_nb(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output)
 109 {
 110     ARM_COMPUTE_ERROR_ON((weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))));
 111
 112     // If the fully connected layer is called after a convolution layer, the input tensor must be linearized
 113
 114     // Initialize output tensor for im2col
 115     TensorShape shape_im2col;
 116     shape_im2col.set(0, weights->info()->dimension(1));
 117     shape_im2col.set(1, 1);
 118     _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, input->info()->data_type()));
 119
 120     // Configure im2col kernel
 121     _im2col_kernel.configure(input, &_im2col_output, std::make_pair(1, 1), PadStrideInfo(1, 1, 0, 0), false);
 122
 123     // Configure matrix multiply kernel
 124     _mm_kernel.configure(&_im2col_output, weights, output, 1.0f);
 125
 126     // Allocate the output tensor for im2col once all the configure methods have been called
 127     _im2col_output.allocator()->allocate();
 128 }
 129
 130 void CLFullyConnectedLayer::configure_fc_fc_nb(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output)
 131 {
 132     ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1));
 133
 134     // Configure matrix multiply kernel
 135     _mm_kernel.configure(input, weights, output, 1.0f);
 136 }
 137
 138 void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights)
 139 {
 140     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
 141     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F32);
 142     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
 143     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() != 2);
 144
 145     const ICLTensor *weights_to_use = weights;
 146
 147     _is_first_run      = true;
 148     _transpose_weights = transpose_weights;
 149     _fc_after_conv     = true;
 150     _batched_fc_layer  = false;
 151     _accumulate_biases = false;
 152
 153     if(biases != nullptr)
 154     {
 155         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
 156
 157         _accumulate_biases = true;
 158
 159         // Configure accumulate biases kernel
 160         _accumulate_biases_kernel.configure(output, biases);
 161     }
 162
 163     // Check if we need to transpose the weights
 164     if(_transpose_weights)
 165     {
 166         // Initialize the output tensor for transpose
 167         TensorShape shape_transposed(weights->info()->dimension(1), weights->info()->dimension(0));
 168         _transpose_output.allocator()->init(TensorInfo(shape_transposed, 1, weights->info()->data_type()));
 169         _transpose_kernel.configure(weights, &_transpose_output);
 170
 171         weights_to_use = &_transpose_output;
 172     }
 173
 174     // With the Fully Connected layer we can have 4 different cases:
 175     //  1) Convolution layer -> Fully Connected layer without batches
 176     //  2) Fully Connected layer -> Fully Connected layer without batches
 177     //  3) Convolution layer -> Fully Connected layer with batches
 178     //  4) Fully Connected layer -> Fully Connected layer with batches
 179
 180     // Check if we have a fully connected layer with batches
 181     _batched_fc_layer = (output->info()->dimension(1) > 1);
 182
 183     if(_batched_fc_layer)
 184     {
 185         _fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->info()->tensor_shape().cbegin() + 3,
 186                                                                                input->info()->tensor_shape().cend(),
 187                                                                                output->info()->tensor_shape().cbegin() + 1));
 188
 189         if(_fc_after_conv)
 190         {
 191             // Fully Connected layer after a Convolution Layer with batches
 192             configure_conv_fc_wb(input, weights_to_use, output);
 193         }
 194         else
 195         {
 196             // Fully Connected layer after a Fully Connected Layer with batches
 197             configure_fc_fc_wb(input, weights_to_use, output);
 198         }
 199     }
 200     else
 201     {
 202         _fc_after_conv = (weights_to_use->info()->dimension(1) == (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2)));
 203
 204         if(_fc_after_conv)
 205         {
 206             // Fully Connected layer after a Convolution Layer without batches
 207             configure_conv_fc_nb(input, weights_to_use, output);
 208         }
 209         else
 210         {
 211             // Fully Connected layer after a Fully Connected Layer without batches
 212             configure_fc_fc_nb(input, weights_to_use, output);
 213         }
 214     }
 215
 216     // Allocate the transpose tensor if the transpose_weights flag is true and once all the configure methods have been called
 217     if(_transpose_weights)
 218     {
 219         _transpose_output.allocator()->allocate();
 220     }
 221 }
 222
 223 void CLFullyConnectedLayer::run()
 224 {
 225     // The reshape of the weights happens only once
 226     if(_is_first_run)
 227     {
 228         _is_first_run = false;
 229
 230         if(_transpose_weights)
 231         {
 232             CLScheduler::get().enqueue(_transpose_kernel);
 233         }
 234
 235         if(_batched_fc_layer)
 236         {
 237             CLScheduler::get().enqueue(_transpose1xW_kernel);
 238         }
 239     }
 240
 241     // Linearize input if it comes from a convolutional layer
 242     if(_fc_after_conv)
 243     {
 244         CLScheduler::get().enqueue(_im2col_kernel, false);
 245     }
 246
 247     // Interleave input
 248     if(_batched_fc_layer)
 249     {
 250         CLScheduler::get().enqueue(_interleave4x4_kernel, false);
 251     }
 252
 253     // Run matrix multiply
 254     CLScheduler::get().enqueue(_mm_kernel, !_accumulate_biases);
 255
 256     // Accumulate biases if provided
 257     if(_accumulate_biases)
 258     {
 259         CLScheduler::get().enqueue(_accumulate_biases_kernel);
 260     }
 261 }