1 /*
2 * Copyright (c) 2017-2020 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/GLES_COMPUTE/functions/GCFullyConnectedLayer.h"
25
26 #include "arm_compute/core/Validate.h"
27 #include "arm_compute/runtime/GLES_COMPUTE/GCScheduler.h"
28 #include "support/MemorySupport.h"
29
30 #include <algorithm>
31
32 using namespace arm_compute;
33
configure(const IGCTensor * input,IGCTensor * output)34 void GCFullyConnectedLayerReshapeWeights::configure(const IGCTensor *input, IGCTensor *output)
35 {
36 auto k = arm_compute::support::cpp14::make_unique<GCTransposeKernel>();
37 k->configure(input, output);
38 _kernel = std::move(k);
39 }
40
GCFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager,IWeightsManager * weights_manager)41 GCFullyConnectedLayer::GCFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager)
42 : _memory_group(std::move(memory_manager)), _weights_manager(std::move(weights_manager)), _im2col_kernel(), _reshape_weights_kernel(), _mm_kernel(), _accumulate_biases_kernel(), _im2col_output(),
43 _reshape_weights_output(), _original_weights(nullptr), _are_weights_reshaped(true), _is_fc_after_conv(true), _accumulate_biases(false)
44 {
45 }
46
configure_conv_fc(const IGCTensor * input,const IGCTensor * weights,IGCTensor * output)47 void GCFullyConnectedLayer::configure_conv_fc(const IGCTensor *input, const IGCTensor *weights, IGCTensor *output)
48 {
49 ARM_COMPUTE_ERROR_ON((weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))));
50
51 const DataType dt = input->info()->data_type();
52
53 // If the fully connected layer is called after a convolution layer, the input tensor must be linearized
54
55 // Initialize output tensor for im2col
56 TensorShape shape_im2col;
57 shape_im2col.set(0, input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2));
58 shape_im2col.set(1, input->info()->dimension(3));
59 shape_im2col.set(2, input->info()->dimension(4));
60 shape_im2col.set(3, input->info()->dimension(5));
61 _im2col_output.allocator()->init(TensorInfo(shape_im2col, 1, dt));
62
63 // Configure im2col kernel
64 _memory_group.manage(&_im2col_output);
65 _im2col_kernel.configure(input, &_im2col_output, Size2D(1, 1), PadStrideInfo(1, 1, 0, 0), false);
66
67 // Configure matrix multiply kernel
68 _mm_kernel.configure(&_im2col_output, weights, output, 1.0f, false);
69
70 // Allocate the output tensor for im2col once all the configure methods have been called
71 _im2col_output.allocator()->allocate();
72 }
73
configure_fc_fc(const IGCTensor * input,const IGCTensor * weights,IGCTensor * output)74 void GCFullyConnectedLayer::configure_fc_fc(const IGCTensor *input, const IGCTensor *weights, IGCTensor *output)
75 {
76 ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1));
77
78 // Configure matrix multiply kernel
79 _mm_kernel.configure(input, weights, output, 1.0f, false);
80 }
81
configure(const IGCTensor * input,const IGCTensor * weights,const IGCTensor * biases,IGCTensor * output,FullyConnectedLayerInfo fc_info)82 void GCFullyConnectedLayer::configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output,
83 FullyConnectedLayerInfo fc_info)
84 {
85 ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16);
86 ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
87 ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 2);
88
89 _original_weights = weights;
90 _are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
91 _is_fc_after_conv = true;
92 _accumulate_biases = false;
93
94 if(biases != nullptr)
95 {
96 ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
97
98 _accumulate_biases = true;
99
100 // Configure accumulate biases kernel
101 _accumulate_biases_kernel.configure(output, biases);
102 }
103
104 // With the Fully Connected layer we can have 4 different cases:
105 // 1) Convolution layer -> Fully Connected layer without batches
106 // 2) Fully Connected layer -> Fully Connected layer without batches
107 // 3) Convolution layer -> Fully Connected layer with batches
108 // 4) Fully Connected layer -> Fully Connected layer with batches
109
110 const IGCTensor *weights_to_use = weights;
111
112 if(!_are_weights_reshaped)
113 {
114 weights_to_use = &_reshape_weights_output;
115
116 // Reshape the weights
117 _reshape_weights_kernel.configure(weights, &_reshape_weights_output);
118 }
119
120 // Check if we have a fully connected layer with batches
121 const bool is_batched_fc_layer = output->info()->dimension(1) > 1;
122
123 if(is_batched_fc_layer)
124 {
125 _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->info()->tensor_shape().cbegin() + 3,
126 input->info()->tensor_shape().cend(),
127 output->info()->tensor_shape().cbegin() + 1));
128 }
129 else
130 {
131 _is_fc_after_conv = input->info()->num_dimensions() > 1;
132 }
133
134 if(_is_fc_after_conv)
135 {
136 // Fully Connected layer after a Convolution Layer without batches
137 configure_conv_fc(input, weights_to_use, output);
138 }
139 else
140 {
141 // Fully Connected layer after a Fully Connected Layer without batches
142 configure_fc_fc(input, weights_to_use, output);
143 }
144
145 ARM_COMPUTE_ERROR_ON(fc_info.retain_internal_weights && _reshape_weights_output.gc_buffer() == 0);
146 _are_weights_reshaped = _are_weights_reshaped || fc_info.retain_internal_weights;
147 }
148
run()149 void GCFullyConnectedLayer::run()
150 {
151 prepare();
152
153 MemoryGroupResourceScope scope_mg(_memory_group);
154
155 // Linearize input if it comes from a convolutional layer
156 if(_is_fc_after_conv)
157 {
158 GCScheduler::get().dispatch(_im2col_kernel, false);
159 }
160
161 if(!_are_weights_reshaped || _is_fc_after_conv)
162 {
163 GCScheduler::get().memory_barrier();
164 }
165
166 // Run matrix multiply
167 GCScheduler::get().dispatch(_mm_kernel, !_accumulate_biases);
168
169 // Accumulate biases if provided
170 if(_accumulate_biases)
171 {
172 GCScheduler::get().memory_barrier();
173
174 GCScheduler::get().dispatch(_accumulate_biases_kernel);
175 }
176 }
177
prepare()178 void GCFullyConnectedLayer::prepare()
179 {
180 // Reshape of the weights (happens only once)
181 if(!_are_weights_reshaped)
182 {
183 ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
184
185 // Run reshape weights kernel and mark weights as unused
186 _reshape_weights_output.allocator()->allocate();
187 _reshape_weights_kernel.run();
188
189 // Mark original weights tensor as unused
190 _original_weights->mark_as_unused();
191
192 _are_weights_reshaped = true;
193 }
194 }
195