Compute Library
 20.08
NELocallyConnectedLayer.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2017-2019 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  */
25 
27 #include "arm_compute/core/Utils.h"
30 
31 #include <cmath>
32 #include <tuple>
33 
34 using namespace arm_compute;
35 
36 namespace
37 {
38 void calculate_shapes(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
39  TensorShape &shape_wr, TensorShape &shape_im2col, TensorShape &shape_gemm)
40 {
41  ARM_COMPUTE_UNUSED(output);
42 
43  const unsigned int kernel_width = weights->dimension(0);
44  const unsigned int kernel_height = weights->dimension(1);
45 
46  bool has_bias = (biases != nullptr);
47 
48  // Get convolved dimensions
49  unsigned int conv_w = 0;
50  unsigned int conv_h = 0;
51  std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height,
52  conv_info);
53 
54  const size_t mat_weights_cols = weights->dimension(3);
55  const size_t mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + ((has_bias) ? 1 : 0);
56  const size_t mat_weights_num = weights->dimension(4);
57 
58  shape_wr = TensorShape(mat_weights_cols, mat_weights_rows, mat_weights_num);
59 
60  const size_t mat_input_cols = mat_weights_rows;
61  const size_t mat_input_rows = conv_w * conv_h;
62 
63  shape_im2col = input->tensor_shape();
64  shape_im2col.set(0, mat_input_cols);
65  shape_im2col.set(1, mat_input_rows);
66  shape_im2col.set(2, 1);
67 
68  shape_gemm = shape_im2col;
69  shape_gemm.set(0, mat_weights_cols);
70  shape_gemm.set(1, mat_input_rows);
71 }
72 } // namespace
73 
74 NELocallyConnectedLayer::NELocallyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager)
75  : _memory_group(std::move(memory_manager)), _input_im2col_kernel(), _weights_reshape_kernel(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(), _weights_reshaped(), _gemm_output(),
76  _is_prepared(false), _original_weights(nullptr)
77 {
78 }
79 
81 {
83  ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(2) != input->dimension(2));
84  ARM_COMPUTE_RETURN_ERROR_ON(!conv_info.padding_is_symmetric());
85 
86  bool has_bias = (biases != nullptr);
87 
88  if(has_bias)
89  {
90  ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3));
92  }
93 
94  const unsigned int kernel_width = weights->dimension(0);
95  const unsigned int kernel_height = weights->dimension(1);
96 
97  // Get convolved dimensions
98  unsigned int conv_w = 0;
99  unsigned int conv_h = 0;
100  std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height,
101  conv_info);
102 
103  ARM_COMPUTE_RETURN_ERROR_ON_MSG((output->dimension(0) != conv_w) || (output->dimension(1) != conv_h), "Output shape does not match the expected one");
104  ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(4) != (conv_w * conv_h), "Weights shape does not match the expected one");
105 
106  // Calculate intermediate buffer shapes
107  TensorShape shape_wr;
108  TensorShape shape_im2col;
109  TensorShape shape_gemm;
110  calculate_shapes(input, weights, biases, output, conv_info, shape_wr, shape_im2col, shape_gemm);
111 
112  TensorInfo weights_reshaped_info(shape_wr, 1, weights->data_type());
113  TensorInfo input_im2col_reshaped_info(shape_im2col, 1, input->data_type());
114  TensorInfo gemm_output_info(shape_gemm, 1, input->data_type());
115 
116  ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &input_im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, has_bias));
117  ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, &weights_reshaped_info));
118  ARM_COMPUTE_RETURN_ON_ERROR(NELocallyConnectedMatrixMultiplyKernel::validate(&input_im2col_reshaped_info, &weights_reshaped_info, &gemm_output_info));
119  ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(&gemm_output_info, output, Size2D(conv_w, conv_h)));
120 
121  return Status{};
122 }
123 
125 {
127  ARM_COMPUTE_ERROR_THROW_ON(NELocallyConnectedLayer::validate(input->info(), weights->info(), biases == nullptr ? nullptr : biases->info(), output->info(), conv_info));
128 
129  bool _has_bias = (biases != nullptr);
130  _is_prepared = false;
131  _original_weights = weights;
132 
133  const unsigned int kernel_width = weights->info()->dimension(0);
134  const unsigned int kernel_height = weights->info()->dimension(1);
135 
136  // Get convolved dimensions
137  unsigned int conv_w = 0;
138  unsigned int conv_h = 0;
139  std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height,
140  conv_info);
141 
142  // Calculate intermediate buffer shapes
143  TensorShape shape_wr;
144  TensorShape shape_im2col;
145  TensorShape shape_gemm;
146  calculate_shapes(input->info(), weights->info(), biases == nullptr ? nullptr : biases->info(), output->info(), conv_info, shape_wr, shape_im2col, shape_gemm);
147 
148  _weights_reshaped.allocator()->init(TensorInfo(shape_wr, 1, weights->info()->data_type()));
149  _input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, input->info()->data_type()));
150  _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, input->info()->data_type()));
151 
152  // Manage intermediate buffers
153  _memory_group.manage(&_input_im2col_reshaped);
154  _memory_group.manage(&_gemm_output);
155 
156  // Configure kernels
157  _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias);
158  _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped);
159  _mm_kernel.configure(&_input_im2col_reshaped, &_weights_reshaped, &_gemm_output);
160  _output_col2im_kernel.configure(&_gemm_output, output, Size2D(conv_w, conv_h));
161 
162  // Allocate intermediate tensors
163  _input_im2col_reshaped.allocator()->allocate();
164  _gemm_output.allocator()->allocate();
165 }
166 
168 {
169  prepare();
170 
171  MemoryGroupResourceScope scope_mg(_memory_group);
172 
173  // Run input reshaping
174  NEScheduler::get().schedule(&_input_im2col_kernel, Window::DimY);
175 
176  // Runs GEMM on reshaped matrices
177  NEScheduler::get().schedule(&_mm_kernel, Window::DimX);
178 
179  // Reshape output matrix
180  NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY);
181 }
182 
184 {
185  if(!_is_prepared)
186  {
187  ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
188 
189  // Run weights reshaping and mark original weights tensor as unused
190  _weights_reshaped.allocator()->allocate();
191  NEScheduler::get().schedule(&_weights_reshape_kernel, 3);
192  _original_weights->mark_as_unused();
193 
194  _is_prepared = true;
195  }
196 }
void configure(const ITensor *input, const ITensor *bias, ITensor *output)
Set the input and output of the kernel.
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
Shape of a tensor.
Definition: TensorShape.h:39
void init(const TensorAllocator &allocator, const Coordinates &coords, TensorInfo &sub_info)
Shares the same backing memory with another tensor allocator, while the tensor info might be differen...
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output)
Static function to check if given info will lead to a valid configuration of NELocallyConnectedMatrix...
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
virtual DataType data_type() const =0
Data type used for each element of the tensor.
bool is_used() const
Flags if the tensor is used or not.
Definition: ITensor.cpp:163
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
NELocallyConnectedLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Default constructor.
Store the tensor's metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Status class.
Definition: Error.h:52
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
Interface for NEON tensor.
Definition: ITensor.h:36
Copyright (c) 2017-2020 Arm Limited.
void configure(const ITensor *input, ITensor *output, const Size2D &convolved_dims)
Set the input and output of the kernel.
static Status validate(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output)
Static function to check if given info will lead to a valid configuration of NEWeightsReshapeKernel.
std::pair< unsigned int, unsigned int > scaled_dimensions(int width, int height, int kernel_width, int kernel_height, const PadStrideInfo &pad_stride_info, const Size2D &dilation=Size2D(1U, 1U))
Returns expected width and height of output scaled tensor depending on dimensions rounding mode.
Definition: Utils.cpp:395
TensorAllocator * allocator()
Return a pointer to the tensor's allocator.
Definition: Tensor.cpp:48
ITensorInfo * info() const override
Interface to be implemented by the child class to return the tensor's metadata.
Definition: Tensor.cpp:33
void mark_as_unused() const
Marks a tensor as unused.
Definition: ITensor.cpp:168
void manage(IMemoryManageable *obj) override
Sets a object to be managed by the given memory group.
Definition: MemoryGroup.h:79
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info)
Static function to check if given info will lead to a valid configuration of NELocallyConnectedLayer.
void allocate() override
Allocate size specified by TensorInfo of CPU memory.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
Padding and stride information class.
Definition: Types.h:689
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &convolved_dims)
Static function to check if given info will lead to a valid configuration of NECol2ImKernel.
void prepare() override
Prepare the function for executing.
void configure(const ITensor *input0, const ITensor *input1, ITensor *output)
Initialise the kernel's input and output.
void run() override
Run the kernels contained in the function.
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:163
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
Memory group resources scope handling class.
Definition: IMemoryGroup.h:82
virtual void schedule(ICPPKernel *kernel, const Hints &hints)=0
Runs the kernel in the same thread as the caller synchronously.
Class for specifying the size of an image or rectangle.
Definition: Size2D.h:34
void configure(const ITensor *input, ITensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation=Size2D(1U, 1U), unsigned int num_groups=1)
Set the input and output of the kernel.
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
Definition: Error.h:244
TensorShape & set(size_t dimension, size_t value, bool apply_dim_correction=true)
Accessor to set the value of one of the dimensions.
Definition: TensorShape.h:78
Store the tensor's metadata.
Definition: TensorInfo.h:45
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation=Size2D(1U, 1U), unsigned int num_groups=1)
Static function to check if given info will lead to a valid configuration of NEIm2ColKernel.
void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info)
Set the input and output tensors.
static IScheduler & get()
Access the scheduler singleton.
Definition: Scheduler.cpp:95