xref: /aosp_15_r20/external/ComputeLibrary/arm_compute/runtime/NEON/functions/NEWinogradConvolutionLayer.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2022 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_NEWINOGRADCONVOLUTIONLAYER_H
25 #define ARM_COMPUTE_NEWINOGRADCONVOLUTIONLAYER_H
26 
27 #include "arm_compute/runtime/IFunction.h"
28 
29 #include "arm_compute/core/Types.h"
30 #include "arm_compute/runtime/Tensor.h"
31 
32 #include <memory>
33 
34 namespace arm_compute
35 {
36 // Forward declarations
37 class ITensor;
38 
39 /** Basic function to simulate a convolution layer. This function calls the following kernels:
40  *
41  * -# @ref cpu::CpuWinogradConv2dTransformInputKernel
42  * -# @ref cpu::CpuWinogradConv2dTransformOutputKernel
43  * -# @ref cpu::CpuGemmAssemblyDispatch
44  * -# @ref CPPPermute (three times: weights, input and output)
45  *
46  * @note  Some Winograd configurations (i.e. F(2x2, 5x5), F(4x4, 5x5)) are supported only with enable_fast_math = true
47  */
48 class NEWinogradConvolutionLayer : public IFunction
49 {
50 public:
51     /** Constructor */
52     NEWinogradConvolutionLayer(const std::shared_ptr<IMemoryManager> &memory_manager = nullptr);
53     /** Prevent instances of this class from being copied (As this class contains pointers) */
54     NEWinogradConvolutionLayer(const NEWinogradConvolutionLayer &) = delete;
55     /** Default move constructor */
56     NEWinogradConvolutionLayer(NEWinogradConvolutionLayer &&) = default;
57     /** Prevent instances of this class from being copied (As this class contains pointers) */
58     NEWinogradConvolutionLayer &operator=(const NEWinogradConvolutionLayer &) = delete;
59     /** Default move assignment operator */
60     NEWinogradConvolutionLayer &operator=(NEWinogradConvolutionLayer &&) = default;
61     /** Destructor */
62     ~NEWinogradConvolutionLayer();
63 
64     /** Set the input and output tensors.
65      *
66      * Valid data layouts:
67      * - NHWC
68      * - NCHW
69      *
70      * Valid data type configurations:
71      * |src0           |src1           |src2   |dst            |
72      * |:--------------|:--------------|:------|:--------------|
73      * |F16            |F16            |F16    |F16            |
74      * |F32            |F32            |F32    |F32            |
75      *
76      * @param[in]  input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
77      *                              while every optional dimension from 4 and above represent a batch of inputs.
78      *                              Data types supported: F16/F32.
79      * @param[in]  weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
80      *                              Currently only 3x3 and 5x5 kernels are supported.
81      * @param[in]  biases           Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights.
82      * @param[out] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
83      *                              Data types supported: Same as @p input.
84      * @param[in]  conv_info        Contains padding and stride information described in @ref PadStrideInfo. Currently only unit strides are supported.
85      * @param[in]  act_info         (Optional) Activation layer information in case of a fused activation.
86      * @param[in]  enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
87      *                              available which may introduce a drop of accuracy as well. Default is false
88      */
89     void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info = ActivationLayerInfo(),
90                    bool enable_fast_math = false);
91 
92     // Inherited methods overridden:
93     void run() override;
94     void prepare() override;
95 
96     /** Static function to check if given info will lead to a valid configuration of @ref NEWinogradConvolutionLayer
97      *
98      * Similar to @ref NEWinogradConvolutionLayer::configure()
99      *
100      * @return a status
101      */
102     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
103                            const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false);
104 
105 private:
106     struct Impl;
107     std::unique_ptr<Impl> _impl;
108 };
109 } // namespace arm_compute
110 #endif /* ARM_COMPUTE_NEWINOGRADCONVOLUTIONLAYER_H */
111