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 <functional> |
12 |
|
|
|
13 |
|
|
#include "test/common/data_type.hpp" |
14 |
|
|
|
15 |
|
|
namespace kai::test { |
16 |
|
|
|
17 |
|
|
/// Data format. |
18 |
|
|
class DataFormat { |
19 |
|
|
public: |
20 |
|
|
/// Packing format. |
21 |
|
|
enum class PackFormat : uint32_t { |
22 |
|
|
NONE, ///< No quantization information is included. |
23 |
|
|
BIAS_PER_ROW, ///< Per-row bias. |
24 |
|
|
QUANTIZE_PER_ROW, ///< Per-row quantization. |
25 |
|
|
}; |
26 |
|
|
|
27 |
|
|
/// Creates a new data format. |
28 |
|
|
/// |
29 |
|
|
/// @param[in] data_type Data type of data value. |
30 |
|
|
/// @param[in] block_height Block height. |
31 |
|
|
/// @param[in] block_width Block width. |
32 |
|
|
/// @param[in] pack_format Packing format. |
33 |
|
|
/// @param[in] zero_point_dt Data type of zero point value. |
34 |
|
|
/// @param[in] scale_dt Data type of scale value. |
35 |
|
|
/// @param[in] subblock_height Sub-block height. |
36 |
|
|
/// @param[in] subblock_width Sub-block width. |
37 |
|
|
DataFormat( |
38 |
|
|
DataType data_type = DataType::UNKNOWN, size_t block_height = 0, size_t block_width = 0, |
39 |
|
|
PackFormat pack_format = PackFormat::NONE, DataType zero_point_dt = DataType::UNKNOWN, |
40 |
|
|
DataType scale_dt = DataType::UNKNOWN, size_t subblock_height = 0, size_t subblock_width = 0) noexcept; |
41 |
|
|
|
42 |
|
|
/// Equality operator. |
43 |
|
|
[[nodiscard]] bool operator==(const DataFormat& rhs) const; |
44 |
|
|
|
45 |
|
|
/// Unequality operator. |
46 |
|
|
[[nodiscard]] bool operator!=(const DataFormat& rhs) const; |
47 |
|
|
|
48 |
|
|
/// Gets the packing format. |
49 |
|
|
[[nodiscard]] PackFormat pack_format() const; |
50 |
|
|
|
51 |
|
|
/// Gets the data type of data value. |
52 |
|
|
[[nodiscard]] DataType data_type() const; |
53 |
|
|
|
54 |
|
|
/// Gets the data type of scale value. |
55 |
|
|
[[nodiscard]] DataType scale_data_type() const; |
56 |
|
|
|
57 |
|
|
/// Gets the data type of zero point value. |
58 |
|
|
[[nodiscard]] DataType zero_point_data_type() const; |
59 |
|
|
|
60 |
|
|
/// Gets a value indicating whether this format has no blocking or packing information. |
61 |
|
|
[[nodiscard]] bool is_raw() const; |
62 |
|
|
|
63 |
|
|
/// Gets the block height. |
64 |
|
|
[[nodiscard]] size_t block_height() const; |
65 |
|
|
|
66 |
|
|
/// Gets the block width. |
67 |
|
|
[[nodiscard]] size_t block_width() const; |
68 |
|
|
|
69 |
|
|
/// Gets the sub-block height. |
70 |
|
|
[[nodiscard]] size_t subblock_height() const; |
71 |
|
|
|
72 |
|
|
/// Gets the sub-block width. |
73 |
|
|
[[nodiscard]] size_t subblock_width() const; |
74 |
|
|
|
75 |
|
|
/// Gets the block height given the full height of the matrix. |
76 |
|
|
/// |
77 |
|
|
/// @param[in] full_height Height of the full matrix. |
78 |
|
|
/// |
79 |
|
|
/// @return The block height. |
80 |
|
|
[[nodiscard]] size_t actual_block_height(size_t full_height) const; |
81 |
|
|
|
82 |
|
|
/// Gets the block width given the full width of the matrix. |
83 |
|
|
/// |
84 |
|
|
/// @param[in] full_width Width of the full matrix. |
85 |
|
|
/// |
86 |
|
|
/// @return The block width. |
87 |
|
|
[[nodiscard]] size_t actual_block_width(size_t full_width) const; |
88 |
|
|
|
89 |
|
|
/// Gets the sub-block height given the full height of the matrix. |
90 |
|
|
/// |
91 |
|
|
/// @param[in] full_height Height of the full matrix. |
92 |
|
|
/// |
93 |
|
|
/// @return The sub-block height. |
94 |
|
|
[[nodiscard]] size_t actual_subblock_height(size_t full_height) const; |
95 |
|
|
|
96 |
|
|
/// Gets the sub-block width given the full width of the matrix. |
97 |
|
|
/// |
98 |
|
|
/// @param[in] full_width Width of the full matrix. |
99 |
|
|
/// |
100 |
|
|
/// @return The sub-block width. |
101 |
|
|
[[nodiscard]] size_t actual_subblock_width(size_t full_width) const; |
102 |
|
|
|
103 |
|
|
/// Gets the scheduling block height. |
104 |
|
|
/// |
105 |
|
|
/// @param[in] full_height Height of the full matrix. |
106 |
|
|
/// |
107 |
|
|
/// @return The block height for scheduling purpose. |
108 |
|
|
[[nodiscard]] size_t scheduler_block_height(size_t full_height) const; |
109 |
|
|
|
110 |
|
|
/// Gets the scheduling block width. |
111 |
|
|
/// |
112 |
|
|
/// @param[in] full_width Width of the full matrix. |
113 |
|
|
/// |
114 |
|
|
/// @return The block width for scheduling purpose. |
115 |
|
|
[[nodiscard]] size_t scheduler_block_width(size_t full_width) const; |
116 |
|
|
|
117 |
|
|
/// Gets the row stride in bytes given the data is stored continuously without any gap in the memory. |
118 |
|
|
/// |
119 |
|
|
/// In case of per-row bias or quantization, the row stride is the number of bytes from one row group |
120 |
|
|
/// to the next. One row group consists of `block_height` rows. |
121 |
|
|
/// |
122 |
|
|
/// @param[in] width Width of the full matrix. |
123 |
|
|
/// |
124 |
|
|
/// @return The default row stride in bytes of the matrix. |
125 |
|
|
[[nodiscard]] uintptr_t default_row_stride(size_t width) const; |
126 |
|
|
|
127 |
|
|
/// Gets the offsets in bytes in the data buffer given the data is stored continuously |
128 |
|
|
/// without any gap in the memory. |
129 |
|
|
/// |
130 |
|
|
/// @param[in] row Row coordinate. |
131 |
|
|
/// @param[in] col Colum coordinate. |
132 |
|
|
/// @param[in] width Width of the full matrix. |
133 |
|
|
/// |
134 |
|
|
/// @return The default offset in bytes. |
135 |
|
|
[[nodiscard]] uintptr_t default_offset_in_bytes(size_t row, size_t col, size_t width) const; |
136 |
|
|
|
137 |
|
|
/// Gets the size in bytes of the matrix given the data is stored continuously without any gap in the memory. |
138 |
|
|
/// |
139 |
|
|
/// @param[in] height Height of the full matrix. |
140 |
|
|
/// @param[in] width Width of the full matrix. |
141 |
|
|
/// |
142 |
|
|
/// @return The size in bytes of the matrix. |
143 |
|
|
[[nodiscard]] size_t default_size_in_bytes(size_t height, size_t width) const; |
144 |
|
|
|
145 |
|
|
/// Hash functor |
146 |
|
|
struct Hash { |
147 |
|
|
size_t operator()(const DataFormat& format) const; |
148 |
|
|
}; |
149 |
|
|
|
150 |
|
|
private: |
151 |
|
|
DataType _data_type; |
152 |
|
|
PackFormat _pack_format; |
153 |
|
|
DataType _scale_dt; |
154 |
|
|
DataType _zero_point_dt; |
155 |
|
|
size_t _block_height; |
156 |
|
|
size_t _block_width; |
157 |
|
|
size_t _subblock_height; |
158 |
|
|
size_t _subblock_width; |
159 |
|
|
}; |
160 |
|
|
|
161 |
|
|
} // namespace kai::test |
162 |
|
|
|
163 |
|
|
template <> |
164 |
|
|
struct std::hash<kai::test::DataFormat> { |
165 |
|
3729 |
size_t operator()(const kai::test::DataFormat& df) const { |
166 |
|
3729 |
return kai::test::DataFormat::Hash{}(df); |
167 |
|
|
} |
168 |
|
|
}; |
169 |
|
|
|
170 |
|
|
template <> |
171 |
|
|
struct std::hash<kai::test::DataFormat::PackFormat> { |
172 |
|
120417 |
size_t operator()(const kai::test::DataFormat::PackFormat& pf) const { |
173 |
|
|
using PF = std::underlying_type_t<kai::test::DataFormat::PackFormat>; |
174 |
|
120417 |
return std::hash<PF>{}(static_cast<PF>(pf)); |
175 |
|
|
} |
176 |
|
|
}; |
177 |
|
|
|