1 /* 2 * Copyright (c) 2017-2020 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_NEGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFIXEDPOINTKERNEL_H 25 #define ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFIXEDPOINTKERNEL_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 QASYMM8 34 * 35 * This kernel takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 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 * -# Add offset to each result 42 * -# Clamp the value between the specified min and max bounds 43 * -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8. 44 * 45 */ 46 class NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel : public INEKernel 47 { 48 public: name()49 const char *name() const override 50 { 51 return "NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel"; 52 } 53 /** Constructor */ 54 NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel(); 55 /** Prevent instances of this class from being copied (As this class contains pointers)*/ 56 NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel(const NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel &) = delete; 57 /** Prevent instances of this class from being copied (As this class contains pointers)*/ 58 NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel &operator=(const NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel &) = delete; 59 /** Allow instances of this class to be moved */ 60 NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel(NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel &&) = default; 61 /** Allow instances of this class to be moved */ 62 NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel &operator=(NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel &&) = default; 63 /** Default destructor */ 64 ~NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel() = default; 65 /** Initialise the kernel's input and output. 66 * 67 * @param[in] input Input tensor. Data type supported: S32 68 * @param[in] bias Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required. 69 * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input. 70 * @param[out] output Output tensor. Data type supported: Data type supported: QASYMM8 71 * @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 72 * @param[in] result_shift Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication 73 * @param[in] result_offset_after_shift Offset to be applied to result before converting it back to QASYMM8 74 * @param[in] min (Optional) Min value used to saturate down the output result before converting back to QASYMM8 75 * @param[in] max (Optional) Max value used to saturate up the output result before converting back to QASYMM8, 76 * Along with @p min, this value can be used to implement "rectified linear unit" activation functions 77 */ 78 void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift, int min = 0, int max = 0); 79 /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel 80 * 81 * @param[in] input Input tensor. Data type supported: S32 82 * @param[in] bias Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required. 83 * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input. 84 * @param[in] output Output tensor. Data type supported: Data type supported: QASYMM8 85 * @param[in] min (Optional) Min value used to saturate down the output result before converting back to QASYMM8 86 * @param[in] max (Optional) Max value used to saturate up the output result before converting back to QASYMM8, 87 * Along with @p min, this value can be used to implement "rectified linear unit" activation functions 88 * 89 * @return a status 90 */ 91 static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0); 92 93 // Inherited methods overridden: 94 void run(const Window &window, const ThreadInfo &info) override; 95 96 private: 97 /** Template function to run the NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel 98 * 99 * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). 100 */ 101 template <bool is_bounded_relu> 102 void run(const Window &window); 103 104 /** Common signature for all the specialised NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel functions 105 * 106 * @param[in] window Region on which to execute the kernel. 107 */ 108 using QuantizeDownFunctionPtr = void (NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::*)(const Window &window); 109 110 QuantizeDownFunctionPtr _func; 111 const ITensor *_input; 112 const ITensor *_bias; 113 ITensor *_output; 114 int _result_fixedpoint_multiplier; 115 int _result_shift; 116 int _result_offset_after_shift; 117 int _min; 118 int _max; 119 }; 120 } // namespace arm_compute 121 #endif /* ARM_COMPUTE_NEGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFIXEDPOINTKERNEL_H */ 122