• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2018-2020 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_CLWINOGRADOUTPUTTRANSFORMKERNEL_H
25 #define ARM_COMPUTE_CLWINOGRADOUTPUTTRANSFORMKERNEL_H
26 
27 #include "src/core/CL/ICLKernel.h"
28 
29 namespace arm_compute
30 {
31 class ICLTensor;
32 
33 /** Interface for the Winograd output transform kernel. */
34 class CLWinogradOutputTransformKernel : public ICLKernel
35 {
36 public:
37     /** Default constructor */
38     CLWinogradOutputTransformKernel();
39     /** Prevent instances of this class from being copied (As this class contains pointers) */
40     CLWinogradOutputTransformKernel(const CLWinogradOutputTransformKernel &) = delete;
41     /** Prevent instances of this class from being copied (As this class contains pointers) */
42     CLWinogradOutputTransformKernel &operator=(const CLWinogradOutputTransformKernel &) = delete;
43     /** Allow instances of this class to be moved */
44     CLWinogradOutputTransformKernel(CLWinogradOutputTransformKernel &&) = default;
45     /** Allow instances of this class to be moved */
46     CLWinogradOutputTransformKernel &operator=(CLWinogradOutputTransformKernel &&) = default;
47     /** Default destructor */
48     ~CLWinogradOutputTransformKernel() = default;
49     /** Set the input and output tensor.
50      *
51      * @note Winograd output transform supports the following configurations for NCWH data layout
52      *       F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3),
53      *                                   F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3),
54      *                                   F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5)
55      *
56      * @note Winograd output transform supports the following configurations for NHWC data layout
57      *       F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3),
58      *                                   F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5)
59      *
60      *       Strides: only unit strides
61      *
62      * @param[in]  input         Source tensor with shape [C, N, K, batches]. Data types supported: F16/F32.
63      * @param[in]  bias          Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input
64      * @param[out] output        The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_output_transform_shape. Data types supported: Same as @p input
65      * @param[in]  winograd_info Contains Winograd's information described in @ref WinogradInfo
66      * @param[in]  act_info      (Optional) Activation layer information in case of a fused activation.
67      */
68     void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info = ActivationLayerInfo());
69     /** Set the input and output tensor.
70      *
71      * @note Winograd output transform supports the following configurations for NCWH data layout
72      *       F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3),
73      *                                   F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3),
74      *                                   F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5)
75      *
76      * @note Winograd output transform supports the following configurations for NHWC data layout
77      *       F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3),
78      *                                   F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5)
79      *
80      *       Strides: only unit strides
81      *
82      * @param[in]  compile_context The compile context to be used.
83      * @param[in]  input           Source tensor with shape [C, N, K, batches]. Data types supported: F16/F32.
84      * @param[in]  bias            Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input
85      * @param[out] output          The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_output_transform_shape. Data types supported: Same as @p input
86      * @param[in]  winograd_info   Contains Winograd's information described in @ref WinogradInfo
87      * @param[in]  act_info        (Optional) Activation layer information in case of a fused activation.
88      */
89     void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const WinogradInfo &winograd_info,
90                    const ActivationLayerInfo &act_info = ActivationLayerInfo());
91 
92     /** Static function to check if given info will lead to a valid configuration of @ref CLWinogradOutputTransformKernel
93      *
94      * @note Winograd output transform supports the following configurations for NCWH data layout
95      *       F(output tile, kernel size):F(2x2, 3x3), F(2x1, 3x1), F(1x2, 1x3),
96      *                                   F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3),
97      *                                   F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5)
98      *
99      * @note Winograd output transform supports the following configurations for NHWC data layout
100      *       F(output tile, kernel size):F(4x4, 3x3), F(4x1, 3x1), F(1x4, 1x3),
101      *                                   F(4x4, 5x5), F(4x1, 5x1), F(1x4, 1x5)
102      *
103      *       Strides: only unit strides
104      *
105      * @param[in]  input         Source tensor with shape [C, N, K, batches]. Data types supported: F16/F32.
106      * @param[in]  bias          Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. It can be a nullptr. Data type supported: as @p input
107      * @param[out] output        The output tensor. The shape for this tensor can be calculated using the utility function @p compute_winograd_output_transform_shape. Data types supported: Same as @p input
108      * @param[in]  winograd_info Contains Winograd's information described in @ref WinogradInfo
109      * @param[in]  act_info      (Optional) Activation layer information in case of a fused activation @ref ActivationLayerInfo. Only RELU, BOUNDED_RELU, LU_BOUNDED_RELU, LEAKY_RELU and SOFT_RELU supported.
110      *
111      * @return a status
112      */
113     static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const WinogradInfo &winograd_info, const ActivationLayerInfo &act_info = ActivationLayerInfo());
114 
115     // Inherited methods overridden:
116     void run(const Window &window, cl::CommandQueue &queue) override;
117 
118 private:
119     using WinogradKey = std::pair<std::pair<int, int>, std::pair<int, int>>;
120 
121     const ICLTensor *_input;
122     const ICLTensor *_bias;
123     ICLTensor       *_output;
124     bool             _is_nhwc;
125 };
126 } // namespace arm_compute
127 #endif /*ARM_COMPUTE_CLWINOGRADOUTPUTTRANSFORMKERNEL_H */
128