xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/kernels/CLFuseBatchNormalizationKernel.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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_CLFUSEBATCHNORMALIZATIONKERNEL_H
25 #define ARM_COMPUTE_CLFUSEBATCHNORMALIZATIONKERNEL_H
26 
27 #include "src/core/CL/ICLKernel.h"
28 
29 namespace arm_compute
30 {
31 // Forward declarations
32 class ICLTensor;
33 
34 /** OpenCL kernel to fuse the batch normalization node to a preceding convolution node */
35 class CLFuseBatchNormalizationKernel : public ICLKernel
36 {
37 public:
38     /** Default constructor */
39     CLFuseBatchNormalizationKernel();
40     /** Prevent instances of this class from being copied (As this class contains pointers) */
41     CLFuseBatchNormalizationKernel(const CLFuseBatchNormalizationKernel &) = delete;
42     /** Prevent instances of this class from being copied (As this class contains pointers) */
43     CLFuseBatchNormalizationKernel &operator=(const CLFuseBatchNormalizationKernel &) = delete;
44     /** Allow instances of this class to be moved */
45     CLFuseBatchNormalizationKernel(CLFuseBatchNormalizationKernel &&) = default;
46     /** Allow instances of this class to be moved */
47     CLFuseBatchNormalizationKernel &operator=(CLFuseBatchNormalizationKernel &&) = default;
48     /** Default destructor */
49     ~CLFuseBatchNormalizationKernel() = default;
50     /** Set the source, destination of the kernel
51      *
52      * @param[in]  input_weights Input weights tensor for convolution or depthwise convolution layer. Data type supported: F16/F32. Data layout supported: NCHW, NHWC
53      * @param[in]  bn_mean       Batch normalization layer mean tensor. Same as @p input_weights
54      * @param[in]  bn_var        Batch normalization layer variance tensor. Same as @p input_weights
55      * @param[out] fused_weights Output fused weights tensor. It can be a nullptr in case of in-place computation. Same as @p input_weights
56      * @param[out] fused_bias    Output fused bias tensor. It can be a nullptr in case of in-place computation and input_bias != nullptr. Same as @p input_weights
57      * @param[in]  input_bias    (Optional) Input bias tensor for convolution or depthwise convolution layer. It can be a nullptr in case the bias tensor is not required. Same as @p input_weights
58      * @param[in]  bn_beta       (Optional) Batch normalization layer beta tensor. It can be a nullptr in case the beta tensor is not required. Same as @p input_weights
59      *                           @note if nullptr, bn_beta is set to 0.0
60      * @param[in]  bn_gamma      (Optional) Batch normalization layer gamma tensor. It can be a nullptr in case the gamma tensor is not required. Same as @p input_weights
61      *                           @note if nullptr, bn_gamma is set to 1.0
62      * @param[in]  epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
63      * @param[in]  fbn_type      (Optional) Fused batch normalization type. Defaults to CONVOLUTION.
64      */
65     void configure(const ICLTensor *input_weights, const ICLTensor *bn_mean, const ICLTensor *bn_var, ICLTensor *fused_weights, ICLTensor *fused_bias,
66                    const ICLTensor *input_bias = nullptr, const ICLTensor *bn_beta = nullptr, const ICLTensor *bn_gamma = nullptr,
67                    float epsilon = 0.001f, FuseBatchNormalizationType fbn_type = FuseBatchNormalizationType::CONVOLUTION);
68     /** Set the source, destination of the kernel
69      *
70      * @param[in]  compile_context The compile context to be used.
71      * @param[in]  input_weights   Input weights tensor for convolution or depthwise convolution layer. Data type supported: F16/F32. Data layout supported: NCHW, NHWC
72      * @param[in]  bn_mean         Batch normalization layer mean tensor. Same as @p input_weights
73      * @param[in]  bn_var          Batch normalization layer variance tensor. Same as @p input_weights
74      * @param[out] fused_weights   Output fused weights tensor. It can be a nullptr in case of in-place computation. Same as @p input_weights
75      * @param[out] fused_bias      Output fused bias tensor. It can be a nullptr in case of in-place computation and input_bias != nullptr. Same as @p input_weights
76      * @param[in]  input_bias      (Optional) Input bias tensor for convolution or depthwise convolution layer. It can be a nullptr in case the bias tensor is not required. Same as @p input_weights
77      * @param[in]  bn_beta         (Optional) Batch normalization layer beta tensor. It can be a nullptr in case the beta tensor is not required. Same as @p input_weights
78      *                             @note if nullptr, bn_beta is set to 0.0
79      * @param[in]  bn_gamma        (Optional) Batch normalization layer gamma tensor. It can be a nullptr in case the gamma tensor is not required. Same as @p input_weights
80      *                             @note if nullptr, bn_gamma is set to 1.0
81      * @param[in]  epsilon         (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
82      * @param[in]  fbn_type        (Optional) Fused batch normalization type. Defaults to CONVOLUTION.
83      */
84     void configure(const CLCompileContext &compile_context, const ICLTensor *input_weights, const ICLTensor *bn_mean, const ICLTensor *bn_var, ICLTensor *fused_weights, ICLTensor *fused_bias,
85                    const ICLTensor *input_bias = nullptr, const ICLTensor *bn_beta = nullptr, const ICLTensor *bn_gamma = nullptr,
86                    float epsilon = 0.001f, FuseBatchNormalizationType fbn_type = FuseBatchNormalizationType::CONVOLUTION);
87     /** Static function to check if given info will lead to a valid configuration of @ref CLFuseBatchNormalizationKernel
88      *
89      * @param[in] input_weights Input weights tensor info for convolution or depthwise convolution layer. Data type supported: F16/F32. Data layout supported: NCHW, NHWC
90      * @param[in] bn_mean       Batch normalization layer mean tensor info. Same as @p input_weights
91      * @param[in] bn_var        Batch normalization layer variance tensor info. Same as @p input_weights
92      * @param[in] fused_weights Output fused weights tensor info. It can be a nullptr in case of in-place computation. Same as @p input_weights
93      * @param[in] fused_bias    Output fused bias tensor info. It can be a nullptr in case of in-place computation and input_bias != nullptr. Same as @p input_weights
94      * @param[in] input_bias    (Optional) Input bias tensor info for convolution or depthwise convolution layer. It can be a nullptr in case the bias tensor is not required. Same as @p input_weights
95      * @param[in] bn_beta       (Optional) Batch normalization layer beta tensor info. It can be a nullptr in case the beta tensor is not required. Same as @p input_weights
96      *                          @note if nullptr, bn_beta is set to 0.0
97      * @param[in] bn_gamma      (Optional) Batch normalization layer gamma tensor info. It can be a nullptr in case the gamma tensor is not required. Same as @p input_weights
98      *                          @note if nullptr, bn_gamma is set to 1.0
99      * @param[in] epsilon       (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f.
100      * @param[in] fbn_type      (Optional) Fused batch normalization type. Defaults to CONVOLUTION.
101      *
102      * @return a status
103      */
104     static Status validate(const ITensorInfo *input_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
105                            const ITensorInfo *fused_weights, const ITensorInfo *fused_bias,
106                            const ITensorInfo *input_bias = nullptr, const ITensorInfo *bn_beta = nullptr, const ITensorInfo *bn_gamma = nullptr,
107                            float epsilon = 0.001f, FuseBatchNormalizationType fbn_type = FuseBatchNormalizationType::CONVOLUTION);
108 
109     // Inherited methods overridden:
110     void run(const Window &window, cl::CommandQueue &queue) override;
111 
112 private:
113     const ICLTensor *_input_weights;
114     const ICLTensor *_input_bias;
115     const ICLTensor *_bn_mean;
116     const ICLTensor *_bn_var;
117     const ICLTensor *_bn_gamma;
118     const ICLTensor *_bn_beta;
119     ICLTensor       *_fused_weights;
120     ICLTensor       *_fused_bias;
121     float            _epsilon;
122     bool             _run_in_place_weights;
123     bool             _run_in_place_bias;
124 };
125 } // namespace arm_compute
126 #endif /*ARM_COMPUTE_CLFUSEBATCHNORMALIZATIONKERNEL_H */
127