Compute Library
 22.05
CLGEMMDeconvolutionLayer.h
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2019-2021 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_CLGEMMDECONVOLUTIONLAYER_H
25 #define ARM_COMPUTE_CLGEMMDECONVOLUTIONLAYER_H
26 
39 
40 #include <memory>
41 
42 namespace arm_compute
43 {
44 class CLDeconvolutionReshapeOutputKernel;
45 class ICLTensor;
46 /** Function to run the deconvolution layer through a call to GEMM.
47  *
48  * Deconvolution Layer is the backward pass of Convolution Layer. First we transform the input depending on the stride and pad info and then perform a 1x1
49  * convolution pass. Input stride defines how many zeroes we should put between each element of the input, pad is the amount of padding and finally a is a user
50  * specified value where a < stride - 1, that increases the padding top and right of the input image.
51  *
52  * The relation between input to output is as follows:
53  * \f[
54  * width\_output = (width\_input - 1) \cdot stride\_x - 2 \cdot padding\_x + kernel\_x
55  * \f]
56  * \f[
57  * height\_output = (height\_input - 1) \cdot stride\_y - 2 \cdot padding\_y + kernel\_y
58  * \f]
59  *
60  * where:
61  * width_input is the size of the first input dimension.
62  * height_input is the size of the second input dimension.
63  * width_output is the size of the first output dimension.
64  * height_output is the size of the second output dimension.
65  * kernel_x and kernel_y are the convolution sizes in x and y.
66  * stride_x and stride_y is the input stride of the first and second dimension.
67  *
68  * The weights used by Deconvolution are supposed to be the same as the ones used for Convolution.
69  *
70  * This function calls the following OpenCL kernels/functions:
71  *
72  * -# @ref CLGEMMLowpMatrixMultiplyCore
73  * -# @ref CLGEMMLowpOutputStage
74  * -# @ref CLPermute
75  * -# @ref CLPermute
76  * -# @ref CLReshapeLayer
77  * -# @ref CLTranspose
78  * -# @ref CLDeconvolutionReshapeOutputKernel
79  * -# @ref CLSlice
80  */
82 {
83 public:
84  /** Constructor */
85  CLGEMMDeconvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
86  /** Prevent instances of this class from being copied (As this class contains pointers) */
88  /** Default move constructor */
90  /** Prevent instances of this class from being copied (As this class contains pointers) */
92  /** Default move assignment operator */
94  /** Default desctructor */
96  /** Set the input, weights, biases and output tensors.
97  *
98  * Valid data layouts:
99  * - NHWC
100  *
101  * Valid data type configurations:
102  * |src0 |src1 |src2 |dst |
103  * |:--------------|:------------------|:--------|:--------------|
104  * |F16 |F16 |F16 |F16 |
105  * |F32 |F32 |F32 |F32 |
106  * |QASYMM8 |QASYMM8 |S32 |QASYMM8 |
107  * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED |
108  *
109  * @param[in,out] input Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs.
110  * Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. Data layout supported: NHWC
111  * @param[in] weights The 4d weights with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input. Data layout supported: same as @p input.
112  * @param[in] bias (Optional) The biases have one dimension. Data type supported: Same as @p input. Data layout supported: same as @p input.
113  * @param[out] output Output tensor. The output has the same number of dimensions as the @p input. Data layout supported: same as @p input.
114  * @param[in] deconv_info Contains padding and policies to be used in the deconvolution, this is described in @ref PadStrideInfo. This function supports only stride_x = weights.width && stride_y = weights.height. Moreover, padding is not supported.
115  */
116  void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &deconv_info);
117  /** Set the input, weights, biases and output tensors.
118  *
119  * @param[in] compile_context The compile context to be used.
120  * @param[in,out] input Input tensor. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs.
121  * Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. Data layout supported: NHWC
122  * @param[in] weights The 4d weights with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input. Data layout supported: same as @p input.
123  * @param[in] bias (Optional) The biases have one dimension. Data type supported: Same as @p input. Data layout supported: same as @p input.
124  * @param[out] output Output tensor. The output has the same number of dimensions as the @p input. Data layout supported: same as @p input.
125  * @param[in] deconv_info Contains padding and policies to be used in the deconvolution, this is described in @ref PadStrideInfo. This function supports only stride_x = weights.width && stride_y = weights.height. Moreover, padding is not supported.
126  */
127  void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &deconv_info);
128  /** Static function to check if given info will lead to a valid configuration of @ref CLDeconvolutionLayer
129  *
130  * @param[in] input Input tensor info. 3 lower dimensions represent a single input, and an optional 4th dimension for batch of inputs.
131  * Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. Data layout supported: NHWC
132  * @param[in] weights The 4d weights info with dimensions [width, height, IFM, OFM]. Data type supported: Same as @p input. Data layout supported: same as @p input.
133  * @param[in] bias (Optional) The biases have one dimension. Data type supported: Same as @p input. Data layout supported: same as @p input.
134  * @param[in] output Output tensor info. The output has the same number of dimensions as the @p input. Data layout supported: same as @p input.
135  * @param[in] deconv_info Contains padding and policies to be used in the deconvolution, this is described in @ref PadStrideInfo.
136  *
137  * @return a status
138  */
139  static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &deconv_info);
140 
141  // Inherited methods overridden:
142  void run() override;
143  void prepare() override;
144 
145 private:
146  MemoryGroup _memory_group;
147 
148  CLGEMM _mm_gemm;
149  CLGEMMLowpMatrixMultiplyCore _mm_gemmlowp;
150  CLGEMMLowpOutputStage _gemmlowp_output_stage;
151  CLPermute _permute_input_to_nhwc;
152  CLPermute _permute_weights_to_nhwc;
153  CLReshapeLayer _reshape_weights;
154  CLTranspose _transpose_weights;
155  std::unique_ptr<CLDeconvolutionReshapeOutputKernel> _deconv_reshape;
156  CLSlice _slice_gemm;
157 
158  CLTensor _gemmlowp_final;
159  CLTensor _reshaped_weights;
160  CLTensor _reshaped_weights_t;
161  CLTensor _permuted_input;
162  CLTensor _permuted_weights;
163  CLTensor _gemm_output;
164  CLTensor _slice_gemm_input;
165 
166  const ICLTensor *_original_weights;
167  bool _is_prepared;
168  bool _padded_input;
169  bool _is_nchw;
170  bool _is_quantized;
171 };
172 } // namespace arm_compute
173 #endif /* ARM_COMPUTE_CLGEMMDECONVOLUTIONLAYER_H */
void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output, const PadStrideInfo &deconv_info)
Set the input, weights, biases and output tensors.
Base class for all functions.
Definition: IFunction.h:30
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &deconv_info)
Static function to check if given info will lead to a valid configuration of CLDeconvolutionLayer.
Basic function to execute GEMMLowpQuantizeDown kernels on CL.
CLGEMMDeconvolutionLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr)
Constructor.
void run() override
Run the kernels contained in the function.
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
Status class.
Definition: Error.h:52
Basic function to perform tensor slicing.
Definition: CLSlice.h:38
Copyright (c) 2017-2022 Arm Limited.
CLGEMMDeconvolutionLayer & operator=(const CLGEMMDeconvolutionLayer &)=delete
Prevent instances of this class from being copied (As this class contains pointers) ...
Basic function to execute GEMM on OpenCL.
Definition: CLGEMM.h:44
Function to run the deconvolution layer through a call to GEMM.
Basic function to execute an opencl::kernels::ClPermuteKernel.
Definition: CLPermute.h:39
Padding and stride information class.
Definition: Types.h:669
CLCompileContext class.
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
Basic function to execute an opencl::kernels::ClTransposeKernel.
Definition: CLTranspose.h:39
Basic function to execute GEMMLowpMatrixMultiplyCore on OpenCL.
Basic function to run opencl::kernels::ClReshapeKernel.
void prepare() override
Prepare the function for executing.
const int32_t * bias
Basic implementation of the OpenCL tensor interface.
Definition: CLTensor.h:41
~CLGEMMDeconvolutionLayer()
Default desctructor.