• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/NEON/functions/NELocallyConnectedLayer.h"
25 
26 #include "arm_compute/core/PixelValue.h"
27 #include "arm_compute/core/Utils.h"
28 #include "arm_compute/core/Validate.h"
29 #include "arm_compute/runtime/NEON/NEScheduler.h"
30 #include "src/core/NEON/kernels/NEIm2ColKernel.h"
31 #include "src/core/NEON/kernels/NELocallyConnectedMatrixMultiplyKernel.h"
32 #include "src/core/NEON/kernels/NEWeightsReshapeKernel.h"
33 #include "support/MemorySupport.h"
34 
35 #include <cmath>
36 #include <tuple>
37 
38 namespace arm_compute
39 {
40 namespace
41 {
calculate_shapes(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * output,const PadStrideInfo & conv_info,TensorShape & shape_wr,TensorShape & shape_im2col,TensorShape & shape_gemm)42 void calculate_shapes(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
43                       TensorShape &shape_wr, TensorShape &shape_im2col, TensorShape &shape_gemm)
44 {
45     ARM_COMPUTE_UNUSED(output);
46 
47     const unsigned int kernel_width  = weights->dimension(0);
48     const unsigned int kernel_height = weights->dimension(1);
49 
50     bool has_bias = (biases != nullptr);
51 
52     // Get convolved dimensions
53     unsigned int conv_w = 0;
54     unsigned int conv_h = 0;
55     std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height,
56                                                  conv_info);
57 
58     const size_t mat_weights_cols = weights->dimension(3);
59     const size_t mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + ((has_bias) ? 1 : 0);
60     const size_t mat_weights_num  = weights->dimension(4);
61 
62     shape_wr = TensorShape(mat_weights_cols, mat_weights_rows, mat_weights_num);
63 
64     const size_t mat_input_cols = mat_weights_rows;
65     const size_t mat_input_rows = conv_w * conv_h;
66 
67     shape_im2col = input->tensor_shape();
68     shape_im2col.set(0, mat_input_cols);
69     shape_im2col.set(1, mat_input_rows);
70     shape_im2col.set(2, 1);
71 
72     shape_gemm = shape_im2col;
73     shape_gemm.set(0, mat_weights_cols);
74     shape_gemm.set(1, mat_input_rows);
75 }
76 } // namespace
77 NELocallyConnectedLayer::~NELocallyConnectedLayer() = default;
78 
NELocallyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)79 NELocallyConnectedLayer::NELocallyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
80     : _memory_group(std::move(memory_manager)), _input_im2col(), _weights_reshape_kernel(), _mm_kernel(), _output_col2im(), _input_im2col_reshaped(), _weights_reshaped(), _gemm_output(),
81       _is_prepared(false), _original_weights(nullptr)
82 {
83 }
84 
validate(const ITensorInfo * input,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * output,const PadStrideInfo & conv_info)85 Status NELocallyConnectedLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info)
86 {
87     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
88     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(2) != input->dimension(2));
89     ARM_COMPUTE_RETURN_ERROR_ON(!conv_info.padding_is_symmetric());
90 
91     bool has_bias = (biases != nullptr);
92 
93     if(has_bias)
94     {
95         ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3));
96         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 2);
97     }
98 
99     const unsigned int kernel_width  = weights->dimension(0);
100     const unsigned int kernel_height = weights->dimension(1);
101 
102     // Get convolved dimensions
103     unsigned int conv_w = 0;
104     unsigned int conv_h = 0;
105     std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height,
106                                                  conv_info);
107 
108     ARM_COMPUTE_RETURN_ERROR_ON_MSG((output->dimension(0) != conv_w) || (output->dimension(1) != conv_h), "Output shape does not match the expected one");
109     ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(4) != (conv_w * conv_h), "Weights shape does not match the expected one");
110 
111     // Calculate intermediate buffer shapes
112     TensorShape shape_wr;
113     TensorShape shape_im2col;
114     TensorShape shape_gemm;
115     calculate_shapes(input, weights, biases, output, conv_info, shape_wr, shape_im2col, shape_gemm);
116 
117     TensorInfo weights_reshaped_info(shape_wr, 1, weights->data_type());
118     TensorInfo input_im2col_reshaped_info(shape_im2col, 1, input->data_type());
119     TensorInfo gemm_output_info(shape_gemm, 1, input->data_type());
120 
121     ARM_COMPUTE_RETURN_ON_ERROR(NEIm2Col::validate(input, &input_im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, has_bias));
122     ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, &weights_reshaped_info));
123     ARM_COMPUTE_RETURN_ON_ERROR(NELocallyConnectedMatrixMultiplyKernel::validate(&input_im2col_reshaped_info, &weights_reshaped_info, &gemm_output_info));
124     ARM_COMPUTE_RETURN_ON_ERROR(NECol2Im::validate(&gemm_output_info, output, Size2D(conv_w, conv_h)));
125 
126     return Status{};
127 }
128 
configure(const ITensor * input,const ITensor * weights,const ITensor * biases,ITensor * output,const PadStrideInfo & conv_info)129 void NELocallyConnectedLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info)
130 {
131     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
132     ARM_COMPUTE_ERROR_THROW_ON(NELocallyConnectedLayer::validate(input->info(), weights->info(), biases == nullptr ? nullptr : biases->info(), output->info(), conv_info));
133 
134     bool _has_bias    = (biases != nullptr);
135     _is_prepared      = false;
136     _original_weights = weights;
137 
138     const unsigned int kernel_width  = weights->info()->dimension(0);
139     const unsigned int kernel_height = weights->info()->dimension(1);
140 
141     // Get convolved dimensions
142     unsigned int conv_w = 0;
143     unsigned int conv_h = 0;
144     std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height,
145                                                  conv_info);
146 
147     // Calculate intermediate buffer shapes
148     TensorShape shape_wr;
149     TensorShape shape_im2col;
150     TensorShape shape_gemm;
151     calculate_shapes(input->info(), weights->info(), biases == nullptr ? nullptr : biases->info(), output->info(), conv_info, shape_wr, shape_im2col, shape_gemm);
152 
153     _weights_reshaped.allocator()->init(TensorInfo(shape_wr, 1, weights->info()->data_type()));
154     _input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, input->info()->data_type()));
155     _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, input->info()->data_type()));
156 
157     // Manage intermediate buffers
158     _memory_group.manage(&_input_im2col_reshaped);
159     _memory_group.manage(&_gemm_output);
160 
161     // Configure kernels
162     _input_im2col.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias);
163     _weights_reshape_kernel = arm_compute::support::cpp14::make_unique<NEWeightsReshapeKernel>();
164     _weights_reshape_kernel->configure(weights, biases, &_weights_reshaped);
165     _mm_kernel = arm_compute::support::cpp14::make_unique<NELocallyConnectedMatrixMultiplyKernel>();
166     _mm_kernel->configure(&_input_im2col_reshaped, &_weights_reshaped, &_gemm_output);
167     _output_col2im.configure(&_gemm_output, output, Size2D(conv_w, conv_h));
168 
169     // Allocate intermediate tensors
170     _input_im2col_reshaped.allocator()->allocate();
171     _gemm_output.allocator()->allocate();
172 }
173 
run()174 void NELocallyConnectedLayer::run()
175 {
176     prepare();
177 
178     MemoryGroupResourceScope scope_mg(_memory_group);
179 
180     // Run input reshaping
181     _input_im2col.run();
182 
183     // Runs GEMM on reshaped matrices
184     NEScheduler::get().schedule(_mm_kernel.get(), Window::DimX);
185 
186     // Reshape output matrix
187     _output_col2im.run();
188 }
189 
prepare()190 void NELocallyConnectedLayer::prepare()
191 {
192     if(!_is_prepared)
193     {
194         ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
195 
196         // Run weights reshaping and mark original weights tensor as unused
197         _weights_reshaped.allocator()->allocate();
198         NEScheduler::get().schedule(_weights_reshape_kernel.get(), 3);
199         _original_weights->mark_as_unused();
200 
201         _is_prepared = true;
202     }
203 }
204 } // namespace arm_compute
205