xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2020-2021 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_NEQLSTMLAYERNORMALIZATIONKERNEL_H
25 #define ARM_COMPUTE_NEQLSTMLAYERNORMALIZATIONKERNEL_H
26 
27 #include "src/core/NEON/INEKernel.h"
28 #include <functional>
29 
30 namespace arm_compute
31 {
32 class ITensor;
33 
34 /** Kernel to perform layer normalization for QLSTM. */
35 class NEQLSTMLayerNormalizationKernel : public INEKernel
36 {
37 public:
name()38     const char *name() const override
39     {
40         return "NEQLSTMLayerNormalizationKernel";
41     }
42     /** Default constructor */
43     NEQLSTMLayerNormalizationKernel() = default;
44     /** Prevent instances of this class from being copied (As this class contains pointers) */
45     NEQLSTMLayerNormalizationKernel(const NEQLSTMLayerNormalizationKernel &) = delete;
46     /** Prevent instances of this class from being copied (As this class contains pointers) */
47     NEQLSTMLayerNormalizationKernel &operator=(const NEQLSTMLayerNormalizationKernel &) = delete;
48     /** Default Move Constructor. */
49     NEQLSTMLayerNormalizationKernel(NEQLSTMLayerNormalizationKernel &&) = default;
50     /** Default move assignment operator */
51     NEQLSTMLayerNormalizationKernel &operator=(NEQLSTMLayerNormalizationKernel &&) = default;
52     /** Default destructor */
53     ~NEQLSTMLayerNormalizationKernel() = default;
54 
55     /** Set the input and output tensors.
56      *
57      * @param[in]  input  Source tensor. Data types supported: QSYMM16.
58      * @param[out] output Destination tensor. Data types supported: Same as @p input.
59      * @param[in]  weight Weight tensor. Data types supported: Same as @p input.
60      * @param[in]  bias   Bias tensor. Data types supported: S32
61      */
62     void configure(const ITensor *input, ITensor *output, const ITensor *weight, const ITensor *bias);
63     /** Static function to check if given info will lead to a valid configuration of @ref NEQLSTMLayerNormalizationKernel
64      *
65      * @param[in] input  Source tensor info. Data types supported: QSYMM16.
66      * @param[in] output Destination tensor info. Data types supported: Same as @p input.
67      * @param[in] weight Weight tensor info. Data types supported: Same as @p input.
68      * @param[in] bias   Bias tensor info. Data types supported: S32
69      *
70      * @return a status
71      */
72     static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias);
73     // Inherited methods overridden:
74     void run(const Window &window, const ThreadInfo &info) override;
75 
76 private:
77     // constants
78     static constexpr uint32_t max_input_dimension{ 2 };  /**< The maximum input dimension supported */
79     static constexpr uint32_t max_weight_dimension{ 1 }; /**< The maximum weight dimension supported */
80     static constexpr uint32_t max_bias_dimension{ 1 };   /**< The maximum bias dimension supported */
81     static constexpr uint32_t vector_size_byte{ 16 };    /**< Computation vector size in byte */
82 
83     using ComputeFuncType = std::function<void(NEQLSTMLayerNormalizationKernel &)>;
84 
85     ComputeFuncType _fn{}; /**< Function pointer to computation function */
86 
87     const ITensor *_input
88     {
89         nullptr
90     }; /**< Input tensor */
91     const ITensor *_weight
92     {
93         nullptr
94     }; /**< Weight tensor */
95     const ITensor *_bias
96     {
97         nullptr
98     };                           /**< Bias tensor */
99     ITensor *_output{ nullptr }; /**< Output tensor */
100 
101     int32_t _output_multiplier{}; /**< Multiplier for output values */
102     int32_t _output_shift{};      /**< Shift value for output values */
103 
104     int32_t _window_start_x{}; /**< The beginning of x-axis iteration */
105     int32_t _window_end_x{};   /**< The end of x-axis iteration */
106     int32_t _window_step_x{};  /**< The size of x-axis iteration's step */
107 
108     Window _inout_window{};  /**< Window for input and output tensor */
109     Window _weight_window{}; /**< Window for weight and bias tensor */
110 
111     /** Function to configure initial windows for destination of computation
112      *
113      * @param[in] Target destination tensor to use for output window
114      *
115      * @return configured window
116      */
117     Window configure_window(ITensor *target);
118     // Function to compute for data type QSYMM16
119     void compute_qsymm16();
120     /** Function to compute summation and summation of squared input of the given input pointer
121      *
122      * @param[in] Input_ptr pointer to input array
123      *
124      */
125     std::pair<int64_t, int64_t> sum_qsymm16(const int16_t *input_ptr);
126     /** Function to normalize values using computed mean and standard deviation
127      *
128      * @param[in] input_ptr     Pointer to input array
129      * @param[in] output_ptr    Pointer to output array
130      * @param[in] weight_ptr    Pointer to weight array
131      * @param[in] bias_ptr      Pointer to bias array
132      * @param[in] mean          Mean value
133      * @param[in] inv_std_mul   Quantized multiplier for standard deviation
134      * @param[in] inv_std_shift Shift for standard deviation
135      *
136      */
137     void normalize_qasymm16(const int16_t *input_ptr,
138                             int16_t       *output_ptr,
139                             const int16_t *weight_ptr,
140                             const int32_t *bias_ptr,
141                             int32_t mean, int32_t inv_std_mul, int32_t inv_std_shift);
142     /** Function to compute output quantization information */
143     QuantizationInfo compute_output_qinfo();
144 };
145 } // namespace arm_compute
146 #endif /* ARM_COMPUTE_NEQLSTMLAYERNORMALIZATIONKERNEL_H */
147