| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // | ||
| 2 | // SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 3 | // | ||
| 4 | // SPDX-License-Identifier: Apache-2.0 | ||
| 5 | // | ||
| 6 | |||
| 7 | #ifndef KLEIDIAI_BENCHMARK_MATMUL_MATMUL_RUNNER_HPP | ||
| 8 | #define KLEIDIAI_BENCHMARK_MATMUL_MATMUL_RUNNER_HPP | ||
| 9 | |||
| 10 | #include <cfloat> | ||
| 11 | #include <cstddef> | ||
| 12 | #include <cstdint> | ||
| 13 | #include <test/common/data_type.hpp> | ||
| 14 | |||
| 15 | #include "kai/kai_common.h" | ||
| 16 | #include "matmul_interface.hpp" | ||
| 17 | |||
| 18 | namespace kai::benchmark { | ||
| 19 | |||
| 20 | using DataType = test::DataType; | ||
| 21 | |||
| 22 | /// Runner for the matrix multiplication micro-kernel. | ||
| 23 | /// | ||
| 24 | /// Prepares and executes the run method of the micro-kernel. | ||
| 25 | /// | ||
| 26 | /// @tparam MatMulInterface Interface of the matrix multiplication micro-kernel. | ||
| 27 | template <typename MatMulInterface> | ||
| 28 | class MatMulRunner { | ||
| 29 | public: | ||
| 30 | /// Constructs a MatMulRunner object. | ||
| 31 | /// | ||
| 32 | /// @param matmul_interface Abstraction containing the micro-kernel to run. | ||
| 33 | /// @param dst_type Output type of the micro-kernel. Required for the micro-kernel to make certain assumptions | ||
| 34 | /// internally about the stride of the data. | ||
| 35 | 166 | MatMulRunner(const MatMulInterface& matmul_interface, const DataType dst_type) : | |
| 36 | 166 | matmul_interface_(matmul_interface), dst_type_(dst_type) { | |
| 37 | 166 | } | |
| 38 | |||
| 39 | /// Sets the M, N and K dimensions to describe the operand and result matrices. | ||
| 40 | /// | ||
| 41 | /// @param m Rows in a non-transposed LHS and DST matrix. | ||
| 42 | /// @param n Columns in a non-transposed RHS and DST matrix. | ||
| 43 | /// @param k Columns in a non-transposed LHS matrix, and rows in a non-transposed RHS matrix. | ||
| 44 | 83 | void set_mnk(const size_t m, const size_t n, const size_t k) { | |
| 45 | 83 | m_ = m; | |
| 46 | 83 | n_ = n; | |
| 47 | 83 | k_ = k; | |
| 48 | |||
| 49 | 83 | lhs_stride_ = k_ * data_type_size_in_bits(dst_type_) / 8; | |
| 50 | 83 | dst_stride_row_ = n_ * data_type_size_in_bits(dst_type_) / 8; | |
| 51 | 83 | dst_stride_col_ = data_type_size_in_bits(dst_type_) / 8; | |
| 52 | 83 | } | |
| 53 | |||
| 54 | /// Sets the block size to use. | ||
| 55 | /// | ||
| 56 | /// @param bl Block size. Used for micro-kernels with dynamic blockwise quantization. | ||
| 57 | 83 | void set_bl(const size_t bl) { | |
| 58 | 83 | bl_ = bl; | |
| 59 | 83 | } | |
| 60 | |||
| 61 | /// Runs the matrix multiplication micro-kernel. | ||
| 62 | /// | ||
| 63 | /// @param lhs Buffer containing LHS matrix data. | ||
| 64 | /// @param rhs Buffer containing RHS matrix data. | ||
| 65 | /// @param dst Destination buffer to write to. | ||
| 66 | void run(const void* lhs, const void* rhs, void* dst); | ||
| 67 | |||
| 68 | private: | ||
| 69 | MatMulInterface matmul_interface_ = {}; | ||
| 70 | |||
| 71 | DataType dst_type_ = DataType::FP32; | ||
| 72 | |||
| 73 | 83 | size_t m_ = 1; | |
| 74 | 83 | size_t n_ = 1; | |
| 75 | 83 | size_t k_ = 1; | |
| 76 | 83 | size_t bl_ = 32; | |
| 77 | |||
| 78 | 83 | size_t lhs_stride_ = 1; | |
| 79 | 83 | size_t dst_stride_row_ = 1; | |
| 80 | 83 | size_t dst_stride_col_ = 1; | |
| 81 | }; | ||
| 82 | |||
| 83 | /// Runs the matrix multiplication micro-kernel. | ||
| 84 | /// | ||
| 85 | /// @param lhs Buffer containing LHS matrix data. | ||
| 86 | /// @param rhs Buffer containing RHS matrix data. | ||
| 87 | /// @param dst Destination buffer to write to. | ||
| 88 | template <typename MatMulInterface> | ||
| 89 | 18 | void MatMulRunner<MatMulInterface>::run(const void* lhs, const void* rhs, void* dst) { | |
| 90 | 36 | matmul_interface_.run_matmul( | |
| 91 | 18 | m_, n_, k_, // | |
| 92 | 18 | lhs, rhs, dst, // | |
| 93 | 18 | dst_stride_row_, dst_stride_col_, // | |
| 94 | -FLT_MAX, FLT_MAX // | ||
| 95 | ); | ||
| 96 | 18 | } | |
| 97 | |||
| 98 | /// Runs the matrix multiplication micro-kernel. Specialized on the strided LHS interface. | ||
| 99 | /// | ||
| 100 | /// @param lhs Buffer containing LHS matrix data. | ||
| 101 | /// @param rhs Buffer containing RHS matrix data. | ||
| 102 | /// @param dst Destination buffer to write to. | ||
| 103 | template <> | ||
| 104 | 11 | inline void MatMulRunner<MatMulStridedLhsInterface>::run(const void* lhs, const void* rhs, void* dst) { | |
| 105 | 22 | matmul_interface_.run_matmul( | |
| 106 | 11 | m_, n_, k_, // | |
| 107 | 11 | lhs, lhs_stride_, rhs, dst, // | |
| 108 | 11 | dst_stride_row_, dst_stride_col_, // | |
| 109 | -FLT_MAX, FLT_MAX // | ||
| 110 | ); | ||
| 111 | 11 | } | |
| 112 | |||
| 113 | /// Runs the matrix multiplication micro-kernel. Specialized on the interface with a floating point destination buffer. | ||
| 114 | /// | ||
| 115 | /// @param lhs Buffer containing LHS matrix data. | ||
| 116 | /// @param rhs Buffer containing RHS matrix data. | ||
| 117 | /// @param dst Destination buffer to write to. | ||
| 118 | template <> | ||
| 119 | 19 | inline void MatMulRunner<MatMulFloatInterface>::run(const void* lhs, const void* rhs, void* dst) { | |
| 120 | 38 | matmul_interface_.run_matmul( | |
| 121 | 19 | m_, n_, k_, // | |
| 122 | 19 | lhs, rhs, static_cast<float*>(dst), // | |
| 123 | 19 | dst_stride_row_, dst_stride_col_, // | |
| 124 | -FLT_MAX, FLT_MAX // | ||
| 125 | ); | ||
| 126 | 19 | } | |
| 127 | |||
| 128 | /// Runs the matrix multiplication micro-kernel. Specialized on the static quantization interface. | ||
| 129 | /// | ||
| 130 | /// @param lhs Buffer containing LHS matrix data. | ||
| 131 | /// @param rhs Buffer containing RHS matrix data. | ||
| 132 | /// @param dst Destination buffer to write to. | ||
| 133 | template <> | ||
| 134 | 3 | inline void MatMulRunner<MatMulStaticQuantInterface>::run(const void* lhs, const void* rhs, void* dst) { | |
| 135 | 3 | constexpr kai_matmul_requantize32_params params = {INT8_MIN, INT8_MAX, 0}; | |
| 136 | 6 | matmul_interface_.run_matmul( | |
| 137 | 3 | m_, n_, k_, // | |
| 138 | 3 | lhs, rhs, dst, // | |
| 139 | 3 | dst_stride_row_, dst_stride_col_, // | |
| 140 | ¶ms // | ||
| 141 | ); | ||
| 142 | 3 | } | |
| 143 | |||
| 144 | /// Runs the matrix multiplication micro-kernel. Specialized on the dynamic blockwise quantization interface with | ||
| 145 | /// generic destination buffer. | ||
| 146 | /// | ||
| 147 | /// @param lhs Buffer containing LHS matrix data. | ||
| 148 | /// @param rhs Buffer containing RHS matrix data. | ||
| 149 | /// @param dst Destination buffer to write to. | ||
| 150 | template <> | ||
| 151 | 8 | inline void MatMulRunner<MatMulBlockwiseDynamicQuantGenericDstInterface>::run( | |
| 152 | const void* lhs, const void* rhs, void* dst) { | ||
| 153 | 16 | matmul_interface_.run_matmul( | |
| 154 | 8 | m_, n_, k_, bl_, // | |
| 155 | 8 | lhs, rhs, dst, // | |
| 156 | 8 | dst_stride_row_, dst_stride_col_, // | |
| 157 | -FLT_MAX, FLT_MAX // | ||
| 158 | ); | ||
| 159 | 8 | } | |
| 160 | |||
| 161 | /// Runs the matrix multiplication micro-kernel. Specialized on the dynamic blockwise quantization interface. | ||
| 162 | /// | ||
| 163 | /// @param lhs Buffer containing LHS matrix data. | ||
| 164 | /// @param rhs Buffer containing RHS matrix data. | ||
| 165 | /// @param dst Destination buffer to write to. | ||
| 166 | template <> | ||
| 167 | 24 | inline void MatMulRunner<MatMulBlockwiseDynamicQuantInterface>::run(const void* lhs, const void* rhs, void* dst) { | |
| 168 | 48 | matmul_interface_.run_matmul( | |
| 169 | 24 | m_, n_, k_, bl_, // | |
| 170 | 24 | lhs, rhs, static_cast<float*>(dst), // | |
| 171 | 24 | dst_stride_row_, dst_stride_col_, // | |
| 172 | -FLT_MAX, FLT_MAX // | ||
| 173 | ); | ||
| 174 | 24 | } | |
| 175 | |||
| 176 | } // namespace kai::benchmark | ||
| 177 | |||
| 178 | #endif // KLEIDIAI_BENCHMARK_MATMUL_MATMUL_RUNNER_HPP | ||
| 179 |