aboutsummaryrefslogtreecommitdiff
path: root/arm_compute/core/NEON/kernels/NEBatchNormalizationLayerKernel.h
blob: 092b1d951415091c3ec8c7b80aefe8885e77f865 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
/*
 * Copyright (c) 2017-2018 ARM Limited.
 *
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to
 * deal in the Software without restriction, including without limitation the
 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
 * sell copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */
#ifndef __ARM_COMPUTE_NEBATCHNORMALIZATIONLAYERKERNEL_H__
#define __ARM_COMPUTE_NEBATCHNORMALIZATIONLAYERKERNEL_H__

#include "arm_compute/core/NEON/INEKernel.h"

namespace arm_compute
{
class ITensor;

/** Interface for the batch normalization layer kernel.
 */
class NEBatchNormalizationLayerKernel : public INEKernel
{
public:
    const char *name() const override
    {
        return "NEBatchNormalizationLayerKernel";
    }
    /** Default constructor */
    NEBatchNormalizationLayerKernel();
    /** Prevent instances of this class from being copied (As this class contains pointers) */
    NEBatchNormalizationLayerKernel(const NEBatchNormalizationLayerKernel &) = delete;
    /** Prevent instances of this class from being copied (As this class contains pointers) */
    NEBatchNormalizationLayerKernel &operator=(const NEBatchNormalizationLayerKernel &) = delete;
    /** Default Move Constructor. */
    NEBatchNormalizationLayerKernel(NEBatchNormalizationLayerKernel &&) = default;
    /** Default move assignment operator */
    NEBatchNormalizationLayerKernel &operator=(NEBatchNormalizationLayerKernel &&) = default;
    /** Default destructor */
    ~NEBatchNormalizationLayerKernel() = default;
    /** Set the input and output tensors.
     *
     * @note If the output tensor is a nullptr, the batch normalization function will be performed in-place
     *
     * @param[in, out] input    Source tensor. In case of @p output tensor = nullptr, this tensor will store the result.
     *                          3 lower dimensions represent a single input with dimensions [width, height, FM].
     *                          The rest are optional and used for representing batches. Data types supported: F16/F32.
     * @param[out]     output   Destination tensor. Output will have the same number of dimensions as input. Data type supported: same as @p input
     * @param[in]      mean     Mean values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
     * @param[in]      var      Variance values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
     * @param[in]      beta     (Optional) Beta values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for beta is 0. Data types supported: Same as @p input
     * @param[in]      gamma    (Optional) Gamma values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for gamma is 1. Data types supported: Same as @p input
     * @param[in]      epsilon  (Optional) Small value to avoid division with zero. Default value is 0.001f.
     * @param[in]      act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
     */
    void configure(ITensor *input, ITensor *output, const ITensor *mean, const ITensor *var, const ITensor *beta = nullptr, const ITensor *gamma = nullptr, float epsilon = 0.001f,
                   ActivationLayerInfo act_info = ActivationLayerInfo());
    /** Static function to check if given info will lead to a valid configuration of @ref NEBatchNormalizationLayerKernel
     *
     * @param[in] input    Source tensor info. In case of @p output tensor = nullptr, this tensor will store the result.
     *                     3 lower dimensions represent a single input with dimensions [width, height, FM].
     *                     The rest are optional and used for representing batches. Data types supported: F16/F32.
     * @param[in] output   Destination tensor info. Output will have the same number of dimensions as input. Data type supported: same as @p input
     * @param[in] mean     Mean values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
     * @param[in] var      Variance values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input
     * @param[in] beta     (Optional) Beta values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for beta is 0. Data types supported: Same as @p input
     * @param[in] gamma    (Optional) Gamma values tensor info. 1 dimension with size equal to the feature maps [FM]. If not provided, default value for gamma is 1. Data types supported: Same as @p input
     * @param[in] epsilon  (Optional) Small value to avoid division with zero. Default value is 0.001f.
     * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported.
     *
     * @return a status
     */
    static Status validate(const ITensorInfo *input, const ITensorInfo *output,
                           const ITensorInfo *mean, const ITensorInfo *var,
                           const ITensorInfo *beta = nullptr, const ITensorInfo *gamma = nullptr,
                           float epsilon = 0.001f, ActivationLayerInfo act_info = ActivationLayerInfo());

    // Inherited methods overridden:
    void run(const Window &window, const ThreadInfo &info) override;

private:
    /** Configure execution function in case of non-fused activation **/
    void configure_non_fused();
    /** Configure execution function in case of fused activation **/
    void configure_fused();

    /** Template function to run batch normalization on fp16
     *
     * @tparam fused_activation Boolean that flags if its a fused activation or not
     *
     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
     */
    template <bool fused_activation, typename F>
    void batch_normalization_fp16_nchw(const Window &window);
    /** Template function to run batch normalization on fp16 on tensors with NHWC format
     *
     * @tparam fused_activation Boolean that flags if its a fused activation or not
     *
     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
     */
    template <bool fused_activation, typename F>
    void batch_normalization_fp16_nhwc(const Window &window);
    /** Template function to run batch normalization on fp32
     *
     * @tparam fused_activation Boolean that flags if its a fused activation or not
     * @tparam F                Activation function functor to run
     *
     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
     */
    template <bool fused_activation, typename F>
    void batch_normalization_fp32_nchw(const Window &window);
    /** Template function to run batch normalization on fp32 on tensors with NHWC format
     *
     * @tparam fused_activation Boolean that flags if its a fused activation or not
     * @tparam F                Activation function functor to run
     *
     * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
     */
    template <bool fused_activation, typename F>
    void batch_normalization_fp32_nhwc(const Window &window);
    /** Common signature for all the batch normalization functions
     *
     * @param[in] window Region on which to execute the kernel.
     */
    using BatchNormFunctionPtr = void (NEBatchNormalizationLayerKernel::*)(const Window &window);

private:
    BatchNormFunctionPtr _func;
    ITensor             *_input;
    ITensor             *_output;
    const ITensor       *_mean;
    const ITensor       *_var;
    const ITensor       *_gamma;
    const ITensor       *_beta;
    float                _epsilon;
    ActivationLayerInfo  _act_info;
};
} // namespace arm_compute
#endif /*__ARM_COMPUTE_NEBATCHNORMALIZATIONLAYERKERNEL_H__ */