• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2019-2020 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_NEGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H
25 #define ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H
26 
27 #include "src/core/NEON/INEKernel.h"
28 
29 namespace arm_compute
30 {
31 class ITensor;
32 
33 /** NEON kernel used to quantize down the int32 accumulator values of GEMMLowp to QSYMM16
34  *
35  * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QSYMM16 value.
36  * The following computations will be performed by the kernel:
37  *
38  *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
39  *  -# Add bias to final result if bias tensor is not a nullptr
40  *  -# Round to nearest division by a power-of-two using result_shift
41  *  -# Clamp the value between the specified min and max bounds
42  *  -# Clamp the resulting int32 values to the [-32768, 32767] range and cast to QSYMM16.
43  *
44  */
45 class NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel : public INEKernel
46 {
47 public:
name()48     const char *name() const override
49     {
50         return "NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel";
51     }
52     /** Constructor */
53     NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel();
54     /** Prevent instances of this class from being copied (As this class contains pointers)*/
55     NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel(const NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &) = delete;
56     /** Prevent instances of this class from being copied (As this class contains pointers)*/
57     NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &operator=(const NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &) = delete;
58     /** Allow instances of this class to be moved */
59     NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &&) = default;
60     /** Allow instances of this class to be moved */
61     NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &operator=(NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel &&) = default;
62     /** Default destructor */
63     ~NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel() = default;
64     /** Initialise the kernel's input and output.
65      *
66      * @param[in]  input                        Input tensor. Data type supported: S32
67      * @param[in]  bias                         Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required.
68      *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
69      * @param[out] output                       Output tensor. Data type supported: Data type supported: QSYMM16
70      * @param[in]  result_fixedpoint_multiplier Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
71      * @param[in]  result_shift                 Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
72      * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QSYMM16. Defaults to 0.
73      * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QSYMM16.
74      *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to 0.
75      */
76     void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int min = 0, int max = 0);
77     /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
78      *
79      * @param[in] input  Input tensor info. Data type supported: S32
80      * @param[in] bias   Biases tensor info. Only shared biases supported and it can be a nullptr if the biases addition is not required.
81      *                   Biases are 1D tensor info with dimensions [OFM]. Data type supported: Same as @p input.
82      * @param[in] output Output tensor info. Data type supported: Data type supported: QSYMM16
83      * @param[in] min    (Optional) Min value used to saturate down the output result before converting back to QSYMM16. Defaults to 0.
84      * @param[in] max    (Optional) Max value used to saturate up the output result before converting back to QSYMM16,
85      *                            Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to 0.
86      *
87      * @return a status
88      */
89     static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0);
90 
91     // Inherited methods overridden:
92     void run(const Window &window, const ThreadInfo &info) override;
93 
94 private:
95     /** Template function to run the NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
96      *
97      * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
98      */
99     template <bool is_bounded_relu>
100     void run(const Window &window);
101 
102     /** Common signature for all the specialised NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel functions
103      *
104      * @param[in] window Region on which to execute the kernel.
105      */
106     using QuantizeDownFunctionPtr = void (NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::*)(const Window &window);
107 
108     QuantizeDownFunctionPtr _func;
109     const ITensor          *_input;
110     const ITensor          *_bias;
111     ITensor                *_output;
112     int                     _result_fixedpoint_multiplier;
113     int                     _result_shift;
114     int                     _min;
115     int                     _max;
116 };
117 } // namespace arm_compute
118 #endif /* ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOINT16SCALEBYFIXEDPOINTKERNEL_H */
119