test/reference/quantize.hpp
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // | ||
| 2 | // SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 3 | // | ||
| 4 | // SPDX-License-Identifier: Apache-2.0 | ||
| 5 | // | ||
| 6 | |||
| 7 | #pragma once | ||
| 8 | |||
| 9 | #include <cstddef> | ||
| 10 | #include <cstdint> | ||
| 11 | #include <tuple> | ||
| 12 | |||
| 13 | #include "test/common/buffer.hpp" | ||
| 14 | #include "test/common/data_format.hpp" | ||
| 15 | #include "test/common/data_type.hpp" | ||
| 16 | |||
| 17 | namespace kai::test { | ||
| 18 | |||
| 19 | /// Quantization info. | ||
| 20 | struct QuantizationInfo { | ||
| 21 | size_t quant_width{0}; ///< Number of columns in each quantization block. | ||
| 22 | DataType dst_type{DataType::UNKNOWN}; ///< Data type of the output matrix. | ||
| 23 | DataType scale_type{DataType::UNKNOWN}; ///< Data type of the quantization scales. | ||
| 24 | DataType zero_point_type{ | ||
| 25 | DataType::UNKNOWN}; ///< Data type of the quantization zero points (only for asymmetric quantization). | ||
| 26 | }; | ||
| 27 | |||
| 28 | /// Quantization result buffers. | ||
| 29 | 78170 | struct QuantizationOutputs { | |
| 30 | 78170 | Buffer scales{}; ///< Quantization scales. | |
| 31 | 78170 | Buffer zero_points{}; ///< Quantization zero points. | |
| 32 | }; | ||
| 33 | |||
| 34 | template <typename FloatType, typename IntType, typename ZeroPointType> | ||
| 35 | IntType quantize_asymmetric(FloatType value, FloatType scale, ZeroPointType zero_point); | ||
| 36 | |||
| 37 | /// Quantizes each block of the matrix using symmetric quantization method. | ||
| 38 | /// | ||
| 39 | /// The input matrix is divided into quantization blocks of the same size. | ||
| 40 | /// | ||
| 41 | /// The height of the block does not affect the behavior of this function hence it is omitted | ||
| 42 | /// from the function arguments and the figures below. | ||
| 43 | /// | ||
| 44 | /// The quantization scale matrix can be calculated using | ||
| 45 | /// @ref compute_symmetric_per_block_quantization_info function. | ||
| 46 | /// | ||
| 47 | /// The input matrix and the quantization scale matrix: | ||
| 48 | /// | ||
| 49 | /// ``` | ||
| 50 | /// Floating-point data Scale | ||
| 51 | /// | ||
| 52 | /// Quantization blocks -------+ | ||
| 53 | /// | | | ||
| 54 | /// | | | ||
| 55 | /// v v | ||
| 56 | /// +-----------------+-----------------+----- ... +-----+-----+-- ... | ||
| 57 | /// | f00 f01 f02 f03 | f04 f05 f06 f07 | ........ | s00 | s01 | ..... | ||
| 58 | /// | f10 f11 f12 f13 | f14 f15 f16 f17 | ........ | s10 | s11 | ..... | ||
| 59 | /// | f20 f21 f22 f23 | f24 f25 f26 f27 | ........ | s20 | s21 | ..... | ||
| 60 | /// | f30 f31 f32 f33 | f34 f35 f36 f37 | ........ | s30 | s31 | ..... | ||
| 61 | /// | ............... | ............... | ........ | ... | ... | ..... | ||
| 62 | /// : ............... : ............... : ........ : ... : ... : ..... | ||
| 63 | /// ``` | ||
| 64 | /// | ||
| 65 | /// Each row of the quantization block is quantized individually. | ||
| 66 | /// | ||
| 67 | /// ``` | ||
| 68 | /// Floating-point data Scale Quantized data | ||
| 69 | /// +-----------------+ +-----+ +-----------------+ | ||
| 70 | /// | f00 f01 f02 f03 | | s00 | -------> | q00 q01 q02 q03 | | ||
| 71 | /// | f10 f11 f12 f13 | | s10 | -------> | q10 q11 q12 q13 | | ||
| 72 | /// | f20 f21 f22 f23 | | s20 | -------> | q20 q21 q22 q23 | | ||
| 73 | /// | f30 f31 f32 f33 | | s30 | -------> | q30 q31 q32 q33 | | ||
| 74 | /// | ............... | | ... | -------> | ............... | | ||
| 75 | /// : ............... : : ... : : ............... : | ||
| 76 | /// ``` | ||
| 77 | /// | ||
| 78 | /// The computed quantized data matrix: | ||
| 79 | /// | ||
| 80 | /// ``` | ||
| 81 | /// +-----------------+-----------------+----- ... | ||
| 82 | /// | q00 q01 q02 q03 | q04 q05 q06 q07 | ........ | ||
| 83 | /// | q10 q11 q12 q13 | q14 q15 q16 q17 | ........ | ||
| 84 | /// | q20 q21 q22 q23 | q24 q25 q26 q27 | ........ | ||
| 85 | /// | q30 q31 q32 q33 | q34 q35 q36 q37 | ........ | ||
| 86 | /// | ............... | ............... | ........ | ||
| 87 | /// : ............... : ............... : ........ | ||
| 88 | /// ``` | ||
| 89 | /// | ||
| 90 | /// @tparam SrcType The data type of the input data (must be floating-point). | ||
| 91 | /// @tparam DstType The data type of the output data (must be integer). | ||
| 92 | /// @tparam ScaleType The data type of the quantization scales (must be floating-point). | ||
| 93 | /// | ||
| 94 | /// @param[in] src The input matrix. | ||
| 95 | /// @param[in] scales The quantization scale matrix. | ||
| 96 | /// @param[in] height The number of rows. | ||
| 97 | /// @param[in] width The number of columns. | ||
| 98 | /// @param[in] quant_width The number of columns of the quantization block. | ||
| 99 | /// | ||
| 100 | /// @return The quantized data matrix. | ||
| 101 | template <typename SrcType, typename DstType, typename ScaleType> | ||
| 102 | Buffer quantize_symmetric_per_block( | ||
| 103 | const void* src, const void* scales, size_t height, size_t width, size_t quant_width); | ||
| 104 | |||
| 105 | /// Computes the quantization information using asymmetric per-block quantization method. | ||
| 106 | /// | ||
| 107 | /// The input matrix is divided into quantization blocks of the same size. | ||
| 108 | /// | ||
| 109 | /// The height of the block does not affect the behavior of this function hence it is omitted | ||
| 110 | /// from the function arguments and the figures below. | ||
| 111 | /// | ||
| 112 | /// ``` | ||
| 113 | /// Quantization blocks -------+ | ||
| 114 | /// | | | ||
| 115 | /// | | | ||
| 116 | /// v v | ||
| 117 | /// +-----------------+-----------------+----- ... | ||
| 118 | /// | f00 f01 f02 f03 | f04 f05 f06 f07 | ........ | ||
| 119 | /// | f10 f11 f12 f13 | f14 f15 f16 f17 | ........ | ||
| 120 | /// | f20 f21 f22 f23 | f24 f25 f26 f27 | ........ | ||
| 121 | /// | f30 f31 f32 f33 | f34 f35 f36 f37 | ........ | ||
| 122 | /// | ............... | ............... | ........ | ||
| 123 | /// : ............... : ............... : ........ | ||
| 124 | /// ``` | ||
| 125 | /// | ||
| 126 | /// Each row of the quantization block is quantized individually. | ||
| 127 | /// | ||
| 128 | /// ``` | ||
| 129 | /// Floating-point data Scale Zero point | ||
| 130 | /// +-----------------+ +-----+ +-----+ | ||
| 131 | /// | f00 f01 f02 f03 | -------> | s00 | | z00 | | ||
| 132 | /// | f10 f11 f12 f13 | -------> | s10 | | z10 | | ||
| 133 | /// | f20 f21 f22 f23 | -------> | s20 | | z20 | | ||
| 134 | /// | f30 f31 f32 f33 | -------> | s30 | | z30 | | ||
| 135 | /// | ............... | | ... | | ... | | ||
| 136 | /// : ............... : : ... : : ... : | ||
| 137 | /// ``` | ||
| 138 | /// | ||
| 139 | /// The computed quantization scales and zero points matrices: | ||
| 140 | /// | ||
| 141 | /// ``` | ||
| 142 | /// Quantization scale matrix: | ||
| 143 | /// | ||
| 144 | /// +-----+-----+-- ... | ||
| 145 | /// | s00 | s01 | ..... | ||
| 146 | /// | s10 | s11 | ..... | ||
| 147 | /// | s20 | s21 | ..... | ||
| 148 | /// | s30 | s31 | ..... | ||
| 149 | /// | ... | ... | ..... | ||
| 150 | /// : ... : ... : ..... | ||
| 151 | /// ``` | ||
| 152 | /// | ||
| 153 | /// Quantization zero point matrix: | ||
| 154 | /// | ||
| 155 | /// +-----+-----+-- ... | ||
| 156 | /// | z00 | z01 | ..... | ||
| 157 | /// | z10 | z11 | ..... | ||
| 158 | /// | z20 | z21 | ..... | ||
| 159 | /// | z30 | z31 | ..... | ||
| 160 | /// | ... | ... | ..... | ||
| 161 | /// : ... : ... : ..... | ||
| 162 | /// ``` | ||
| 163 | /// | ||
| 164 | /// @tparam SrcType The data type of the input data (must be floating-point). | ||
| 165 | /// @tparam DstType The data type of the output data (must be integer). | ||
| 166 | /// @tparam ScaleType The data type of the quantization scales (must be floating-point). | ||
| 167 | /// @tparam ZeroPointType The data type of the quantization zero points (must be integer). | ||
| 168 | /// | ||
| 169 | /// @param[in] src The input matrix. | ||
| 170 | /// @param[in] height The number of rows. | ||
| 171 | /// @param[in] width The number of columns. | ||
| 172 | /// @param[in] quant_width The number of columns of the quantization block. | ||
| 173 | /// | ||
| 174 | /// @return The quantization scale matrix and the quantization zero point matrix. | ||
| 175 | template <typename SrcType, typename DstType, typename ScaleType, typename ZeroPointType> | ||
| 176 | std::tuple<Buffer, Buffer> compute_asymmetric_per_block_quantization_info( | ||
| 177 | const void* src, size_t height, size_t width, size_t quant_width); | ||
| 178 | |||
| 179 | /// Quantizes each block of the matrix using asymmetric quantization method. | ||
| 180 | /// | ||
| 181 | /// The input matrix is divided into quantization blocks of the same size. | ||
| 182 | /// | ||
| 183 | /// The height of the block does not affect the behavior of this function hence it is omitted | ||
| 184 | /// from the function arguments and the figures below. | ||
| 185 | /// | ||
| 186 | /// The quantization scale and zero point matrix can be calculated using | ||
| 187 | /// @ref compute_asymmetric_per_block_quantization_info function. | ||
| 188 | /// | ||
| 189 | /// The input matrix, quantization scale matrix and quantization zero matrix: | ||
| 190 | /// | ||
| 191 | /// ``` | ||
| 192 | /// Floating-point data Scale Zero point | ||
| 193 | /// | ||
| 194 | /// Quantization blocks -------+ | ||
| 195 | /// | | | ||
| 196 | /// | | | ||
| 197 | /// v v | ||
| 198 | /// +-----------------+-----------------+----- ... +-----+-----+-- ... +-----+-----+-- ... | ||
| 199 | /// | f00 f01 f02 f03 | f04 f05 f06 f07 | ........ | s00 | s01 | ..... | z00 | z01 | ..... | ||
| 200 | /// | f10 f11 f12 f13 | f14 f15 f16 f17 | ........ | s10 | s11 | ..... | z10 | z11 | ..... | ||
| 201 | /// | f20 f21 f22 f23 | f24 f25 f26 f27 | ........ | s20 | s21 | ..... | z20 | z21 | ..... | ||
| 202 | /// | f30 f31 f32 f33 | f34 f35 f36 f37 | ........ | s30 | s31 | ..... | z30 | z31 | ..... | ||
| 203 | /// | ............... | ............... | ........ | ... | ... | ..... | ... | ... | ..... | ||
| 204 | /// : ............... : ............... : ........ : ... : ... : ..... | ... | ... | ..... | ||
| 205 | /// ``` | ||
| 206 | /// | ||
| 207 | /// Each row of the quantization block is quantized individually. | ||
| 208 | /// | ||
| 209 | /// ``` | ||
| 210 | /// Floating-point data Scale Zero point Quantized data | ||
| 211 | /// +-----------------+ +-----+ +-----+ +-----------------+ | ||
| 212 | /// | f00 f01 f02 f03 | | s00 | | z00 | -------> | q00 q01 q02 q03 | | ||
| 213 | /// | f10 f11 f12 f13 | | s10 | | z10 | -------> | q10 q11 q12 q13 | | ||
| 214 | /// | f20 f21 f22 f23 | | s20 | | z20 | -------> | q20 q21 q22 q23 | | ||
| 215 | /// | f30 f31 f32 f33 | | s30 | | z30 | -------> | q30 q31 q32 q33 | | ||
| 216 | /// | ............... | | ... | | ... | -------> | ............... | | ||
| 217 | /// : ............... : : ... : : ... : : ............... : | ||
| 218 | /// ``` | ||
| 219 | /// | ||
| 220 | /// The computed quantized data matrix: | ||
| 221 | /// | ||
| 222 | /// ``` | ||
| 223 | /// +-----------------+-----------------+----- ... | ||
| 224 | /// | q00 q01 q02 q03 | q04 q05 q06 q07 | ........ | ||
| 225 | /// | q10 q11 q12 q13 | q14 q15 q16 q17 | ........ | ||
| 226 | /// | q20 q21 q22 q23 | q24 q25 q26 q27 | ........ | ||
| 227 | /// | q30 q31 q32 q33 | q34 q35 q36 q37 | ........ | ||
| 228 | /// | ............... | ............... | ........ | ||
| 229 | /// : ............... : ............... : ........ | ||
| 230 | /// ``` | ||
| 231 | /// | ||
| 232 | /// @tparam SrcType The data type of the input data (must be floating-point). | ||
| 233 | /// @tparam DstType The data type of the output data (must be integer). | ||
| 234 | /// @tparam ScaleType The data type of the quantization scales (must be floating-point). | ||
| 235 | /// @tparam ZeroPointType The data type of the quantization zero points (must be integer). | ||
| 236 | /// | ||
| 237 | /// @param[in] src The input matrix. | ||
| 238 | /// @param[in] scales The quantization scale matrix. | ||
| 239 | /// @param[in] zero_points The quantization zero point matrix. | ||
| 240 | /// @param[in] height The number of rows. | ||
| 241 | /// @param[in] width The number of columns. | ||
| 242 | /// @param[in] quant_width The number of columns of the quantization block. | ||
| 243 | /// | ||
| 244 | /// @return The quantized data matrix. | ||
| 245 | template <typename SrcType, typename DstType, typename ScaleType, typename ZeroPointType> | ||
| 246 | Buffer quantize_asymmetric_per_block( | ||
| 247 | const void* src, const void* scales, const void* zero_points, size_t height, size_t width, size_t quant_width); | ||
| 248 | |||
| 249 | /// Quantizes the input matrix using the options specified in the quantization info. | ||
| 250 | /// | ||
| 251 | /// @param[in] src The input matrix. | ||
| 252 | /// @param[in] src_type The data type of the input data (must be floating-point). | ||
| 253 | /// @param[in] height The number of rows. | ||
| 254 | /// @param[in] width The number of columns. | ||
| 255 | /// @param[in] qinfo The quantization information. | ||
| 256 | /// | ||
| 257 | /// @return Quantized values and QuantizationOutputs containing scales and (optionally) zero_point data. | ||
| 258 | std::tuple<Buffer, QuantizationOutputs> quantize_dynamic( | ||
| 259 | const void* src, DataType src_type, size_t height, size_t width, const QuantizationInfo& qinfo); | ||
| 260 | } // namespace kai::test | ||
| 261 |