xref: /aosp_15_r20/external/ComputeLibrary/arm_compute/runtime/CL/functions/CLConvolutionLayer.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2021 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_CLCONVOLUTIONLAYER_H
25 #define ARM_COMPUTE_CLCONVOLUTIONLAYER_H
26 
27 #include "arm_compute/core/CL/CLCompileContext.h"
28 #include "arm_compute/core/Types.h"
29 #include "arm_compute/core/experimental/IPostOp.h"
30 #include "arm_compute/runtime/CL/CLTensor.h"
31 #include "arm_compute/runtime/IFunction.h"
32 #include "arm_compute/runtime/IMemoryManager.h"
33 
34 #include <memory>
35 
36 namespace arm_compute
37 {
38 class CLCompileContext;
39 class ICLTensor;
40 class ITensorInfo;
41 
42 /** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions:
43  *
44  * -# @ref opencl::ClGemmConv2d
45  * -# @ref opencl::ClWinogradConv2d
46  * -# @ref opencl::ClDirectConv2d
47  * -# @ref CLFFTConvolutionLayer
48  *
49  * The function selects one of the algorithms mentioned above based on:
50  *      - The size of the kernel
51  *      - Number of input/output feature maps
52  *      - Amount of memory needed
53  *
54  * Generally GEMM-based convolution is executed when neither Winograd nor FFT nor Direct convolution can be performed.
55  *
56  * FP32 Algorithm| Filter Size                                                 |   Input/Output feature maps               |
57  * --------------|-------------------------------------------------------------|-------------------------------------------|
58  * Winograd      | 3x3 1x3 3x1 5x1 1x5 5x5(fast maths) 7x1 1x7                 |  Input channels is greater than 3         |
59  * FFT           | Squared kernels and greater than 9x9                        |  Input feature maps > Output feature maps |
60  * DirectConv    | 9x9                                                         |                                           |
61  * GEMM          | Any size                                                    |                                           |
62  *
63  * Winograd 5x5 requires fast maths enabled.
64  *
65  * FP16 Algorithm| Filter Size                |   Input/Output feature maps               |
66  * --------------|----------------------------|-------------------------------------------|
67  * Winograd      | 3x3 1x3 3x1 5x1 1x5 5x5    |  Input channels is greater than 3         |
68  * FFT           | Not supported              |                                           |
69  * DirectConv    | 9x9                        |                                           |
70  * GEMM          | Any size                   |                                           |
71  *
72  * Winograd FP16 requires fast maths enabled.
73  *
74  */
75 class CLConvolutionLayer : public IFunction
76 {
77 public:
78     /** Default constructor */
79     CLConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
80     /** Default Destructor */
81     ~CLConvolutionLayer();
82     /** Prevent instances of this class from being copied (As this class contains pointers) */
83     CLConvolutionLayer(const CLConvolutionLayer &) = delete;
84     /** Default move constructor */
85     CLConvolutionLayer(CLConvolutionLayer &&) = default;
86     /** Prevent instances of this class from being copied (As this class contains pointers) */
87     CLConvolutionLayer &operator=(const CLConvolutionLayer &) = delete;
88     /** Default move assignment operator */
89     CLConvolutionLayer &operator=(CLConvolutionLayer &&) = default;
90     /** Set the input and output tensors.
91      *
92      * Valid data layouts:
93      * - NHWC
94      * - NCHW
95      *
96      * Valid data type configurations:
97      * |src0           |src1               |src2   |dst            |
98      * |:--------------|:------------------|:------|:--------------|
99      * |F16            |F16                |F16    |F16            |
100      * |F32            |F32                |F32    |F32            |
101      * |QASYMM8        |QASYMM8            |S32    |QASYMM8        |
102      * |QASYMM8        |QSYMM8_PER_CHANNEL |S32    |QASYMM8        |
103      * |QASYMM8_SIGNED |QASYMM8_SIGNED     |S32    |QASYMM8_SIGNED |
104      * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32    |QASYMM8_SIGNED |
105      *
106      * @param[in]  input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
107      *                              while every optional dimension from 4 and above represent a batch of inputs.
108      *                              Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
109      * @param[in]  weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
110      *                              Data type supported: Same as @p input, also could be QSYMM8_PER_CHANNEL if input is QASYMM8/QASYMM8_SIGNED.
111      * @param[in]  biases           Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
112      *                              Data type supported: Same as @p input, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
113      * @param[out] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
114      *                              Data types supported: Same as @p input.
115      * @param[in]  conv_info        Contains padding and stride information described in @ref PadStrideInfo.
116      * @param[in]  weights_info     Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input.
117      * @param[in]  dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
118      * @param[in]  act_info         (Optional) Activation layer information in case of a fused activation.
119      * @param[in]  enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
120      *                              available which may introduce a drop of accuracy as well. Default is false
121      * @param[in]  num_groups       (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout
122      * @param[in]  post_ops         (Optional) A sequence of post operations that are performed after the main operation.
123      */
124     void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo(),
125                    const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false, unsigned int num_groups = 1,
126                    const experimental::PostOpList<ICLTensor *> &post_ops = experimental::PostOpList<ICLTensor *> {});
127     /** Set the input and output tensors.
128      *
129      * @param[in]  compile_context  The compile context to be used.
130      * @param[in]  input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
131      *                              while every optional dimension from 4 and above represent a batch of inputs.
132      *                              Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
133      * @param[in]  weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
134      *                              Data type supported: Same as @p input, also could be QSYMM8_PER_CHANNEL if input is QASYMM8/QASYMM8_SIGNED.
135      * @param[in]  biases           Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
136      *                              Data type supported: Same as @p input, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
137      * @param[out] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
138      *                              Data types supported: Same as @p input.
139      * @param[in]  conv_info        Contains padding and stride information described in @ref PadStrideInfo.
140      * @param[in]  weights_info     Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input.
141      * @param[in]  dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
142      * @param[in]  act_info         (Optional) Activation layer information in case of a fused activation.
143      * @param[in]  enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
144      *                              available which may introduce a drop of accuracy as well. Default is false
145      * @param[in]  num_groups       (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout
146      * @param[in]  post_ops         (Optional) A sequence of post operations that are performed after the main operation.
147      */
148     void configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
149                    const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false,
150                    unsigned int num_groups = 1, const experimental::PostOpList<ICLTensor *> &post_ops = experimental::PostOpList<ICLTensor *> {});
151     /** Static function to check if given info will lead to a valid configuration of @ref CLConvolutionLayer
152      *
153      * @param[in] input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
154      *                             while every optional dimension from 4 and above represent a batch of inputs.
155      *                             Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
156      * @param[in] weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
157      *                             Data type supported: Same as @p input, also could be QSYMM8_PER_CHANNEL if input is QASYMM8/QASYMM8_SIGNED.
158      * @param[in] biases           Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
159      *                             Data type supported: Same as @p input, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type.
160      * @param[in] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
161      *                             Data types supported: Same as @p input.
162      * @param[in] conv_info        Contains padding and stride information described in @ref PadStrideInfo.
163      * @param[in] weights_info     Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel.
164      * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
165      * @param[in] act_info         (Optional) Activation layer information in case of a fused activation.
166      * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
167      *                             available which may introduce a drop of accuracy as well. Default is false
168      * @param[in] num_groups       (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout
169      * @param[in] post_ops         (Optional) A sequence of post operations that are performed after the main operation.
170      *
171      * @return a status
172      */
173     static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
174                            const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false,
175                            unsigned int num_groups = 1, const experimental::PostOpList<ITensorInfo *> &post_ops = experimental::PostOpList<ITensorInfo *> {});
176     /** Static function to check if given info will return the convolution called by @ref CLConvolutionLayer
177      *
178      * @param[in] input            Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
179      *                             while every optional dimension from 4 and above represent a batch of inputs.
180      *                             Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
181      * @param[in] weights          Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
182      *                             Data type supported: Same as @p input, also could be QSYMM8_PER_CHANNEL if input is QASYMM8/QASYMM8_SIGNED.
183      * @param[in] output           Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
184      *                             Data types supported: Same as @p input.
185      * @param[in] conv_info        Contains padding and stride information described in @ref PadStrideInfo.
186      * @param[in] weights_info     Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel.
187      * @param[in] act_info         (Optional) Activation layer information in case of a fused activation.
188      * @param[in] gpu_target       Specifies the @p GPUTarget.
189      * @param[in] dilation         (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
190      * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation
191      *                             available which may introduce a drop of accuracy as well. Default is false
192      *
193      * @return the Convolution Method Hint
194      */
195     static ConvolutionMethod get_convolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const PadStrideInfo &conv_info,
196                                                     const WeightsInfo &weights_info, const ActivationLayerInfo &act_info, const GPUTarget gpu_target, const Size2D &dilation = Size2D(1U, 1U), bool enable_fast_math = false);
197     // Inherited methods overridden:
198     void run() override;
199     void prepare() override;
200 
201 private:
202     struct Impl;
203     std::unique_ptr<Impl> _impl;
204 };
205 }
206 #endif /* ARM_COMPUTE_CLCONVOLUTIONLAYER_H */
207