1 /*
2  * Copyright (c) 2019-2022 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_CPU_GEMMLOWP_QUANTIZEDOWN_INT32TOINT16_SCALEBYFIXEDPOINT_KERNEL_H
25 #define ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWN_INT32TOINT16_SCALEBYFIXEDPOINT_KERNEL_H
26 
27 #include "arm_compute/core/KernelDescriptors.h"
28 #include "src/core/common/Macros.h"
29 #include "src/cpu/ICpuKernel.h"
30 
31 namespace arm_compute
32 {
33 // Forward declaration
34 class ITensor;
35 namespace cpu
36 {
37 namespace kernels
38 {
39 /** Kernel used to quantize down the int32 accumulator values of GEMMLowp to QSYMM16
40  *
41  * This kernel takes a final int32 accumulator value (the output of @ref CpuGemmLowpMatrixMultiplyKernel), and processes it to obtain the final QSYMM16 value.
42  * The following computations will be performed by the kernel:
43  *
44  *  -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier
45  *  -# Add bias to final result if bias tensor is not a nullptr
46  *  -# Round to nearest division by a power-of-two using result_shift
47  *  -# Clamp the value between the specified min and max bounds
48  *  -# Clamp the resulting int32 values to the [-32768, 32767] range and cast to QSYMM16.
49  *
50  */
51 class CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel : public ICpuKernel<CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel>
52 {
53 public:
54     CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel() = default;
55     ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel);
56     /** Initialise the kernel's input and output.
57      *
58      * @param[in]  src                          Input tensor info. Data type supported: S32
59      * @param[in]  bias                         Biases tensor info. Only shared biases supported and it can be a nullptr if the biases addition is not required.
60      *                                          Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
61      * @param[out] dst                          Output tensor info. Data type supported: Data type supported: QSYMM16
62      * @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
63      * @param[in]  result_shift                 Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
64      * @param[in]  min                          (Optional) Min value used to saturate down the output result before converting back to QSYMM16. Defaults to 0.
65      * @param[in]  max                          (Optional) Max value used to saturate up the output result before converting back to QSYMM16.
66      *                                          Along with @p min, this value can be used to implement "rectified linear unit" activation functions. Defaults to 0.
67      */
68     void configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift, int min = 0, int max = 0);
69     /** Static function to check if given info will lead to a valid configuration
70      *
71      * Similar to CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure()
72      *
73      * @return a status
74      */
75     static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min = 0, int max = 0);
76 
77     // Inherited methods overridden:
78     void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
79     const char *name() const override;
80 
81 private:
82     /** Template function to run the CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel
83      *
84      * @param[in]  src    Input tensor info
85      * @param[in]  bias   Bias tensor info
86      * @param[out] dst    Output tensor info
87      * @param[in]  window Region on which to execute the kernel. (Must be a valid region of the window returned by window()).
88      */
89     template <bool is_bounded_relu>
90     void run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
91 
92     /** Common signature for all the specialised CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel functions
93      *
94      * @param[in]  src    Input tensor info
95      * @param[in]  bias   Bias tensor info
96      * @param[out] dst    Output tensor info
97      * @param[in]  window Region on which to execute the kernel.
98      */
99     using QuantizeDownFunctionPtr = void (CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::*)(
100                                         const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window);
101 
102     QuantizeDownFunctionPtr _func{ nullptr };
103     int                     _result_fixedpoint_multiplier{ 0 };
104     int                     _result_shift{ 0 };
105     int                     _min{ 0 };
106     int                     _max{ 0 };
107 };
108 } // namespace kernels
109 } // namespace cpu
110 } // namespace arm_compute
111 #endif /* ARM_COMPUTE_CPU_GEMMLOWP_QUANTIZEDOWN_INT32TOINT16_SCALEBYFIXEDPOINT_KERNEL_H */
112