Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2017-2020 Arm Limited. |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 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 | */ |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 24 | #ifndef ARM_COMPUTE_NEGEMMLOWREDUCTIONKERNEL_H |
| 25 | #define ARM_COMPUTE_NEGEMMLOWREDUCTIONKERNEL_H |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 26 | |
| 27 | #include "arm_compute/core/NEON/INEKernel.h" |
| 28 | |
| 29 | namespace arm_compute |
| 30 | { |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 31 | // Forward declarations |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 32 | class ITensor; |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 33 | struct GEMMLowpReductionKernelInfo; |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 34 | |
| 35 | /** Common interface for all NEON reduction kernels */ |
| 36 | class INEGEMMLowpReductionKernel : public INEKernel |
| 37 | { |
| 38 | public: |
| 39 | /** Constructor */ |
| 40 | INEGEMMLowpReductionKernel(); |
| 41 | /** Prevent instances of this class from being copied (As this class contains pointers)*/ |
| 42 | INEGEMMLowpReductionKernel(const INEGEMMLowpReductionKernel &) = delete; |
| 43 | /** Prevent instances of this class from being copied (As this class contains pointers)*/ |
| 44 | INEGEMMLowpReductionKernel &operator=(const INEGEMMLowpReductionKernel &) = delete; |
| 45 | /** Allow instances of this class to be moved */ |
| 46 | INEGEMMLowpReductionKernel(INEGEMMLowpReductionKernel &&) = default; |
| 47 | /** Allow instances of this class to be moved */ |
| 48 | INEGEMMLowpReductionKernel &operator=(INEGEMMLowpReductionKernel &&) = default; |
| 49 | |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 50 | /** Initialise the kernel's input and output. |
| 51 | * |
Michele Di Giorgio | f9b595a | 2020-07-03 13:34:52 +0100 | [diff] [blame] | 52 | * @param[in] input Input tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 53 | * @param[out] output Output row-vector of sums of all the entries in each row/col of input tensor. Data type supported: S32 |
| 54 | * @param[in] info Kernel metadata: |
| 55 | * - k Number of matrix columns/rows depending on the type of reduction. |
| 56 | * - is_reshaped True if the matrix has been reshaped. |
| 57 | * - scalar Scalar value to multiply each reduced column/row by. |
| 58 | * - mul_byscalar True if each reduced column/row must be multiplied by a scalar value. |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 59 | */ |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 60 | virtual void configure(const ITensor *input, ITensor *output, const GEMMLowpReductionKernelInfo &info) = 0; |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 61 | |
| 62 | protected: |
| 63 | const ITensor *_input; |
| 64 | ITensor *_output; |
| 65 | int32_t _k; |
| 66 | bool _is_reshaped; |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 67 | int32_t _scalar; |
| 68 | bool _mul_by_scalar; |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 69 | }; |
| 70 | |
| 71 | /** NEON kernel used to compute the row-vectors of sums of all the entries in each row of Matrix A. |
| 72 | * |
| 73 | * @note This stage is needed to handle the offset of matrix product |
| 74 | * https://github.com/google/gemmlowp/blob/master/doc/low-precision.md |
| 75 | */ |
| 76 | class NEGEMMLowpMatrixAReductionKernel : public INEGEMMLowpReductionKernel |
| 77 | { |
| 78 | public: |
Anthony Barbier | e8a4983 | 2018-01-18 10:04:05 +0000 | [diff] [blame] | 79 | const char *name() const override |
| 80 | { |
| 81 | return "NEGEMMLowpMatrixAReductionKernel"; |
| 82 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 83 | /** Initialise the kernel's input and output. |
| 84 | * |
Michele Di Giorgio | f9b595a | 2020-07-03 13:34:52 +0100 | [diff] [blame] | 85 | * @param[in] mtx_a Input tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 86 | * @param[out] vector_sum_row Output row-vector of sums of all the entries in each row of mtx_a. Data type supported: S32 |
| 87 | * @param[in] info Kernel metadata: |
| 88 | * - k (num_mtx_a_cols) Number of matrix A columns |
| 89 | * - is_reshaped (is_interleaved4x4) True if the matrix A has been interleaved4x4 |
| 90 | * - scalar Scalar value to multiply each reduced row by. |
| 91 | * - mul_byscalar True if each reduced column must be multiplied by a scalar value. |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 92 | */ |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 93 | void configure(const ITensor *mtx_a, ITensor *vector_sum_row, const GEMMLowpReductionKernelInfo &info) override; |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 94 | /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpMatrixAReductionKernel |
| 95 | * |
Michele Di Giorgio | f9b595a | 2020-07-03 13:34:52 +0100 | [diff] [blame] | 96 | * @param[in] mtx_a Input tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 97 | * @param[in] vector_sum_row Output row-vector of sums of all the entries in each row of mtx_a. Data type supported: S32 |
| 98 | * @param[in] info Kernel metadata: |
| 99 | * - k (num_mtx_a_cols) Number of matrix A columns |
| 100 | * - is_reshaped (is_interleaved4x4) True if the matrix A has been interleaved4x4 |
| 101 | * - scalar Scalar value to multiply each reduced row by. |
| 102 | * - mul_byscalar True if each reduced column must be multiplied by a scalar value. |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 103 | * |
Georgios Pinitas | 631c41a | 2017-12-06 11:53:03 +0000 | [diff] [blame] | 104 | * @return a status |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 105 | */ |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 106 | static Status validate(const ITensorInfo *mtx_a, const ITensorInfo *vector_sum_row, const GEMMLowpReductionKernelInfo &info); |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 107 | |
| 108 | // Inherited methods overridden: |
| 109 | void run(const Window &window, const ThreadInfo &info) override; |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 110 | |
| 111 | private: |
| 112 | /** Execution of the reduction kernel specialized on the input type |
| 113 | * |
| 114 | * @param[in] window Execution window |
| 115 | */ |
| 116 | template <typename T> |
| 117 | void run_internal(const Window &window); |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 118 | }; |
| 119 | |
| 120 | /** NEON kernel used to compute the row-vectors of sums of all the entries in each column of Matrix B. |
| 121 | * |
| 122 | * @note This stage is needed to handle the offset of matrix product |
| 123 | * https://github.com/google/gemmlowp/blob/master/doc/low-precision.md |
| 124 | */ |
| 125 | class NEGEMMLowpMatrixBReductionKernel : public INEGEMMLowpReductionKernel |
| 126 | { |
| 127 | public: |
Anthony Barbier | e8a4983 | 2018-01-18 10:04:05 +0000 | [diff] [blame] | 128 | const char *name() const override |
| 129 | { |
| 130 | return "NEGEMMLowpMatrixBReductionKernel"; |
| 131 | } |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 132 | /** Initialise the kernel's input and output. |
| 133 | * |
Michele Di Giorgio | f9b595a | 2020-07-03 13:34:52 +0100 | [diff] [blame] | 134 | * @param[in] mtx_b Input tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 135 | * @param[out] vector_sum_col Output row-vector of sums of all the entries in each column of mtx_b. Data type supported: S32 |
| 136 | * @param[in] info Kernel metadata: |
| 137 | * - k (num_mtx_b_rows) Number of matrix B rows. |
| 138 | * - is_reshaped (is_transposed1xW) True if the input tensor is transposed 1xW. |
| 139 | * - scalar Scalar value to multiply each reduced row by. |
| 140 | * - mul_byscalar True if each reduced row must be multiplied by a scalar value. |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 141 | */ |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 142 | void configure(const ITensor *mtx_b, ITensor *vector_sum_col, const GEMMLowpReductionKernelInfo &info) override; |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 143 | /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpMatrixBReductionKernel |
| 144 | * |
Michele Di Giorgio | f9b595a | 2020-07-03 13:34:52 +0100 | [diff] [blame] | 145 | * @param[in] mtx_b Input tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 146 | * @param[in] vector_sum_col Output row-vector of sums of all the entries in each column of mtx_b. Data type supported: S32 |
| 147 | * @param[in] info Kernel metadata: |
| 148 | * - k (num_mtx_b_rows) Number of matrix B rows. |
| 149 | * - is_reshaped (is_transposed1xW) True if the input tensor is transposed 1xW. |
| 150 | * - scalar Scalar value to multiply each reduced row by. |
| 151 | * - mul_byscalar True if each reduced row must be multiplied by a scalar value. |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 152 | * |
Georgios Pinitas | 631c41a | 2017-12-06 11:53:03 +0000 | [diff] [blame] | 153 | * @return a status |
Georgios Pinitas | a3b1b46 | 2017-11-16 19:24:39 +0000 | [diff] [blame] | 154 | */ |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 155 | static Status validate(const ITensorInfo *mtx_b, const ITensorInfo *vector_sum_col, const GEMMLowpReductionKernelInfo &info); |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 156 | |
| 157 | // Inherited methods overridden: |
| 158 | void run(const Window &window, const ThreadInfo &info) override; |
Georgios Pinitas | dbdea0d | 2019-10-16 19:21:40 +0100 | [diff] [blame] | 159 | |
| 160 | private: |
| 161 | /** Execution of the reduction kernel specialized on the input type |
| 162 | * |
| 163 | * @param[in] window Execution window |
| 164 | * @param[in] info Thread-related information |
| 165 | */ |
| 166 | template <typename T> |
| 167 | void run_internal(const Window &window, const ThreadInfo &info); |
Gian Marco Iodice | ab18212 | 2017-10-09 15:05:40 +0100 | [diff] [blame] | 168 | }; |
| 169 | } // namespace arm_compute |
| 170 | |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 171 | #endif /* ARM_COMPUTE_NEGEMMLOWREDUCTIONKERNEL_H */ |