Compute Library
 21.02
NEConvolutionLayer.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2018-2021 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 #ifndef ARM_COMPUTE_NECONVOLUTIONLAYER_H
25 #define ARM_COMPUTE_NECONVOLUTIONLAYER_H
26 
28 
30 #include "arm_compute/core/Types.h"
32 
33 #include <memory>
34 
35 namespace arm_compute
36 {
37 // Forward declarations
38 class ITensor;
39 
40 /** Basic function to simulate a convolution layer. This function calls one of the following Neon functions:
41  * -# @ref NEGEMMConvolutionLayer (executed only in case GEMM is required for the operation)
42  * -# @ref NEWinogradConvolutionLayer (executed only in case Winograd is required for the operation)
43  * -# @ref NEDirectConvolutionLayer (executed only in case Direct Convolution is required for the operation)
44  * -# @ref NEFFTConvolutionLayer (executed only in case FFT is required for the operation)
45  *
46  *
47  * The function selects one of the algorithms mentioned above based on:
48  * - The size of the kernel
49  * - Number of input/output feature maps
50  * - Amount of memory needed
51  *
52  * Generally GEMM-based convolution is executed when neither Winograd nor FFT nor Direct convolution can be performed.
53  *
54  * FP32 Algorithm| Filter Size | Input/Output feature maps |
55  * --------------|----------------------------------------------------|-------------------------------------------|
56  * Winograd | 3x3 1x3 3x1 5x1 1x5 5x5(fast maths) 7x1 1x7 | Input channels is greater than 3 |
57  * FFT | Squared kernels and greater than 9x9 | Input feature maps > Output feature maps |
58  * DirectConv | 9x9 | |
59  * GEMM | Any size | |
60  *
61  * Winograd 5x5 requires fast maths enabled.
62  *
63  * FP16 Algorithm| Filter Size |
64  * --------------|------------------|
65  * Winograd | Not supported |
66  * FFT | Not supported |
67  * DirectConv | 9x9 |
68  * GEMM | Any size |
69  *
70  *
71  */
73 {
74 public:
75  /** Constructor */
76  NEConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
77  /** Prevent instances of this class from being copied (As this class contains pointers) */
78  NEConvolutionLayer(const NEConvolutionLayer &) = delete;
79  /** Prevent instances of this class from being copied (As this class contains pointers) */
81  /** Prevent instances of this class from being moved (As this class contains non movable objects) */
83  /** Prevent instances of this class from being moved (As this class contains non movable objects) */
85  /** Default destructor */
86  ~NEConvolutionLayer() = default;
87  /** Set the input and output tensors.
88  *
89  * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
90  * while every optional dimension from 4 and above represent a batch of inputs.
91  * Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
92  * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
93  * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
94  * Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
95  * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
96  * Data types supported: Same as @p input.
97  * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo.
98  * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
99  * tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input.
100  * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
101  * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
102  * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
103  * available which may introduce a drop of accuracy as well. Default is false
104  * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported
105  */
106  void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo(),
107  const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false, unsigned int num_groups = 1);
108  /** Static function to check if given info will lead to a valid configuration of @ref NEConvolutionLayer
109  *
110  * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
111  * while every optional dimension from 4 and above represent a batch of inputs.
112  * Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
113  * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
114  * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
115  * Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
116  * @param[in] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
117  * Data types supported: Same as @p input.
118  * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo.
119  * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
120  * tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input.
121  * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
122  * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
123  * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
124  * available which may introduce a drop of accuracy as well. Default is false
125  * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported
126  *
127  * @return a status
128  */
129  static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
130  const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false,
131  unsigned int num_groups = 1);
132  /** Static function to check if given info will return the convolution called by @ref NEConvolutionLayer
133  *
134  * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
135  * while every optional dimension from 4 and above represent a batch of inputs.
136  * Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
137  * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
138  * @param[in] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
139  * Data types supported: Same as @p input.
140  * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo.
141  * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights
142  * tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input.
143  * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
144  * @param[in] act_info (Optional) Activation layer information in case of a fused activation.
145  * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
146  * available which may introduce a drop of accuracy as well. Default is false
147  *
148  * @return the Convolution Method Hint
149  */
150  static ConvolutionMethod get_convolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const PadStrideInfo &conv_info,
151  const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false);
152  // Inherited methods overridden:
153  void run() override;
154  void prepare() override;
155 
156 private:
157  std::shared_ptr<IMemoryManager> _memory_manager;
158  std::unique_ptr<IFunction> _function; /**< Function to run */
159 };
160 } // namespace arm_compute
161 #endif /* ARM_COMPUTE_NECONVOLUTIONLAYER_H */
Base class for all functions.
Definition: IFunction.h:30
void run() override
Run the kernels contained in the function.
void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info=WeightsInfo(), const Size2D &dilation=Size2D(1U, 1U), const ActivationLayerInfo &act_info=ActivationLayerInfo(), bool enable_fast_math=false, unsigned int num_groups=1)
Set the input and output tensors.
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
Status class.
Definition: Error.h:52
ConvolutionMethod
Available ConvolutionMethod.
Definition: Types.h:138
NEConvolutionLayer & operator=(const NEConvolutionLayer &)=delete
Prevent instances of this class from being copied (As this class contains pointers) ...
Activation Layer Information class.
Definition: Types.h:1550
Interface for Neon tensor.
Definition: ITensor.h:36
Copyright (c) 2017-2021 Arm Limited.
Convolution Layer Weights Information class.
Definition: Types.h:1765
NEConvolutionLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Constructor.
Basic function to simulate a convolution layer.
const unsigned int num_groups
Definition: Im2Col.cpp:153
Padding and stride information class.
Definition: Types.h:722
~NEConvolutionLayer()=default
Default destructor.
static ConvolutionMethod get_convolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info=WeightsInfo(), const Size2D &dilation=Size2D(1U, 1U), const ActivationLayerInfo &act_info=ActivationLayerInfo(), bool enable_fast_math=false)
Static function to check if given info will return the convolution called by NEConvolutionLayer.
Class for specifying the size of an image or rectangle.
Definition: Size2D.h:34
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info=WeightsInfo(), const Size2D &dilation=Size2D(1U, 1U), const ActivationLayerInfo &act_info=ActivationLayerInfo(), bool enable_fast_math=false, unsigned int num_groups=1)
Static function to check if given info will lead to a valid configuration of NEConvolutionLayer.
void prepare() override
Prepare the function for executing.