• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2017-2020 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_CLWEIGHTSRESHAPEKERNEL_H
25 #define ARM_COMPUTE_CLWEIGHTSRESHAPEKERNEL_H
26 
27 #include "src/core/CL/ICLKernel.h"
28 
29 namespace arm_compute
30 {
31 /** OpenCL kernel to perform reshaping on the weights used by convolution and locally connected layer
32  *
33  * Rearranges each 3-dimensional kernel to a single row leading to a matrix with linearized kernels.
34  * In combination with the @ref CLIm2ColKernel can transform a convolution to a matrix multiplication.
35  *
36  * For example assuming a 3D weight kernel of 3x3 dimensions and depth of 2 we have:
37  * @f[
38  * \left( \begin{array}{ccc}
39  * a000 & a001 & a002 \\
40  * a010 & a011 & a012 \\
41  * a020 & a021 & a022 \\
42  * \end{array} \right)
43  * \left( \begin{array}{ccc}
44  * a100 & a101 & a102 \\
45  * a110 & a111 & a112 \\
46  * a120 & a121 & a122 \\
47  * \end{array} \right)
48  * \rightarrow
49  * \left( \begin{array}{ccccccccc}
50  * a000 & a001 & a002 & a010 & a011 & a012 & a020 & a021 & a022 & a100 & a101 & a102 & a110 & a111 & a112 & a120 & a121 & a122 \\
51  * \end{array} \right)
52  * @f]
53  */
54 class CLWeightsReshapeKernel : public ICLKernel
55 {
56 public:
57     /** Constructor.*/
58     CLWeightsReshapeKernel();
59     /** Prevent instances of this class from being copied (As this class contains pointers) */
60     CLWeightsReshapeKernel(const CLWeightsReshapeKernel &) = delete;
61     /** Prevent instances of this class from being copied (As this class contains pointers) */
62     CLWeightsReshapeKernel &operator=(const CLWeightsReshapeKernel &) = delete;
63     /** Allow instances of this class to be moved */
64     CLWeightsReshapeKernel(CLWeightsReshapeKernel &&) = default;
65     /** Allow instances of this class to be moved */
66     CLWeightsReshapeKernel &operator=(CLWeightsReshapeKernel &&) = default;
67     /** Default destructor */
68     ~CLWeightsReshapeKernel() = default;
69     /** Set the input and output of the kernel.
70      *
71      * @param[in]  input      The input tensor to convert. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] if shared,
72      *                        and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM,  num_patches] if unshared. Data types supported: All
73      * @param[in]  biases     The shared biases tensor to append.  Bias is 1D tensor with dimensions [OFM] if shared and 2D tensor with
74      *                        dimensions [OFM, num_patches] if unshared. Data types supported: F16/F32, for quantized types this must be nullptr.
75      *                        @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
76      * @param[out] output     The output tensor. Should be a 2D Tensor if there are no groups and the weights are not shared; a 3D Tensor otherwise.
77      *                        Data types supported: Same as @p input
78      * @param[in]  num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout
79      *                        Number of groups greater than one are only supported for NCHW data layout, and the number of weights must be a multiple of it.
80      */
81     void configure(const ICLTensor *input, const ICLTensor *biases, ICLTensor *output, unsigned int num_groups = 1);
82     /** Set the input and output of the kernel.
83      *
84      * @param[in]  compile_context The compile context to be used.
85      * @param[in]  input           The input tensor to convert. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] if shared,
86      *                             and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM,  num_patches] if unshared. Data types supported: All
87      * @param[in]  biases          The shared biases tensor to append.  Bias is 1D tensor with dimensions [OFM] if shared and 2D tensor with
88      *                             dimensions [OFM, num_patches] if unshared. Data types supported: F16/F32, for quantized types this must be nullptr.
89      *                             @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
90      * @param[out] output          The output tensor. Should be a 2D Tensor if there are no groups and the weights are not shared; a 3D Tensor otherwise.
91      *                             Data types supported: Same as @p input
92      * @param[in]  num_groups      (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout
93      *                             Number of groups greater than one are only supported for NCHW data layout, and the number of weights must be a multiple of it.
94      */
95     void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *biases, ICLTensor *output, unsigned int num_groups = 1);
96     /** Static function to check if given info will lead to a valid configuration of @ref CLWeightsReshapeKernel
97      *
98      * @param[in] input      The input tensor to convert. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] if shared,
99      *                       and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM,  num_patches] if unshared. Data types supported: All
100      * @param[in] biases     The shared biases tensor to append.  Bias is 1D tensor with dimensions [OFM] if shared and 2D tensor with
101      *                       dimensions [OFM, num_patches] if unshared. Data types supported: F16/F32, for quantized types this must be nullptr.
102      *                       @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types.
103      * @param[in] output     The output tensor. Should be a 2D Tensor if there are no groups and the weights are not shared; a 3D Tensor otherwise.
104      *                       Data types supported: Same as @p input
105      * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout
106      *                       Number of groups greater than one are only supported for NCHW data layout, and the number of weights must be a multiple of it.
107      *
108      * @return a status
109      */
110     static Status validate(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output, unsigned int num_groups = 1);
111 
112     // Inherited methods overridden:
113     void run(const Window &window, cl::CommandQueue &queue) override;
114 
115 private:
116     const ICLTensor *_input;
117     const ICLTensor *_biases;
118     ICLTensor       *_output;
119 };
120 } // namespace arm_compute
121 #endif /*ARM_COMPUTE_CLWEIGHTSRESHAPEKERNEL_H */