2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2023 The Stockfish developers (see AUTHORS file)
5 Stockfish is free software: you can redistribute it and/or modify
6 it under the terms of the GNU General Public License as published by
7 the Free Software Foundation, either version 3 of the License, or
8 (at your option) any later version.
10 Stockfish is distributed in the hope that it will be useful,
11 but WITHOUT ANY WARRANTY; without even the implied warranty of
12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 GNU General Public License for more details.
15 You should have received a copy of the GNU General Public License
16 along with this program. If not, see <http://www.gnu.org/licenses/>.
19 // Constants used in NNUE evaluation function
21 #ifndef NNUE_COMMON_H_INCLUDED
22 #define NNUE_COMMON_H_INCLUDED
29 #include <type_traits>
34 #include <immintrin.h>
36 #elif defined(USE_SSE41)
37 #include <smmintrin.h>
39 #elif defined(USE_SSSE3)
40 #include <tmmintrin.h>
42 #elif defined(USE_SSE2)
43 #include <emmintrin.h>
45 #elif defined(USE_NEON)
49 namespace Stockfish::Eval::NNUE {
51 // Version of the evaluation file
52 constexpr std::uint32_t Version = 0x7AF32F20u;
54 // Constant used in evaluation value calculation
55 constexpr int OutputScale = 16;
56 constexpr int WeightScaleBits = 6;
58 // Size of cache line (in bytes)
59 constexpr std::size_t CacheLineSize = 64;
61 constexpr const char Leb128MagicString[] = "COMPRESSED_LEB128";
62 constexpr const std::size_t Leb128MagicStringSize = sizeof(Leb128MagicString) - 1;
64 // SIMD width (in bytes)
66 constexpr std::size_t SimdWidth = 32;
68 #elif defined(USE_SSE2)
69 constexpr std::size_t SimdWidth = 16;
71 #elif defined(USE_NEON)
72 constexpr std::size_t SimdWidth = 16;
75 constexpr std::size_t MaxSimdWidth = 32;
77 // Type of input feature after conversion
78 using TransformedFeatureType = std::uint8_t;
79 using IndexType = std::uint32_t;
81 // Round n up to be a multiple of base
82 template<typename IntType>
83 constexpr IntType ceil_to_multiple(IntType n, IntType base) {
84 return (n + base - 1) / base * base;
88 // read_little_endian() is our utility to read an integer (signed or unsigned, any size)
89 // from a stream in little-endian order. We swap the byte order after the read if
90 // necessary to return a result with the byte ordering of the compiling machine.
91 template<typename IntType>
92 inline IntType read_little_endian(std::istream& stream) {
96 stream.read(reinterpret_cast<char*>(&result), sizeof(IntType));
99 std::uint8_t u[sizeof(IntType)];
100 std::make_unsigned_t<IntType> v = 0;
102 stream.read(reinterpret_cast<char*>(u), sizeof(IntType));
103 for (std::size_t i = 0; i < sizeof(IntType); ++i)
104 v = (v << 8) | u[sizeof(IntType) - i - 1];
106 std::memcpy(&result, &v, sizeof(IntType));
113 // write_little_endian() is our utility to write an integer (signed or unsigned, any size)
114 // to a stream in little-endian order. We swap the byte order before the write if
115 // necessary to always write in little endian order, independently of the byte
116 // ordering of the compiling machine.
117 template<typename IntType>
118 inline void write_little_endian(std::ostream& stream, IntType value) {
121 stream.write(reinterpret_cast<const char*>(&value), sizeof(IntType));
124 std::uint8_t u[sizeof(IntType)];
125 std::make_unsigned_t<IntType> v = value;
128 // if constexpr to silence the warning about shift by 8
129 if constexpr (sizeof(IntType) > 1)
131 for (; i + 1 < sizeof(IntType); ++i)
133 u[i] = (std::uint8_t) v;
137 u[i] = (std::uint8_t) v;
139 stream.write(reinterpret_cast<char*>(u), sizeof(IntType));
144 // read_little_endian(s, out, N) : read integers in bulk from a little indian stream.
145 // This reads N integers from stream s and put them in array out.
146 template<typename IntType>
147 inline void read_little_endian(std::istream& stream, IntType* out, std::size_t count) {
149 stream.read(reinterpret_cast<char*>(out), sizeof(IntType) * count);
151 for (std::size_t i = 0; i < count; ++i)
152 out[i] = read_little_endian<IntType>(stream);
156 // write_little_endian(s, values, N) : write integers in bulk to a little indian stream.
157 // This takes N integers from array values and writes them on stream s.
158 template<typename IntType>
159 inline void write_little_endian(std::ostream& stream, const IntType* values, std::size_t count) {
161 stream.write(reinterpret_cast<const char*>(values), sizeof(IntType) * count);
163 for (std::size_t i = 0; i < count; ++i)
164 write_little_endian<IntType>(stream, values[i]);
168 // read_leb_128(s, out, N) : read N signed integers from the stream s, putting them in
169 // the array out. The stream is assumed to be compressed using the signed LEB128 format.
170 // See https://en.wikipedia.org/wiki/LEB128 for a description of the compression scheme.
171 template<typename IntType>
172 inline void read_leb_128(std::istream& stream, IntType* out, std::size_t count) {
174 // Check the presence of our LEB128 magic string
175 char leb128MagicString[Leb128MagicStringSize];
176 stream.read(leb128MagicString, Leb128MagicStringSize);
177 assert(strncmp(Leb128MagicString, leb128MagicString, Leb128MagicStringSize) == 0);
179 static_assert(std::is_signed_v<IntType>, "Not implemented for unsigned types");
181 const std::uint32_t BUF_SIZE = 4096;
182 std::uint8_t buf[BUF_SIZE];
184 auto bytes_left = read_little_endian<std::uint32_t>(stream);
186 std::uint32_t buf_pos = BUF_SIZE;
187 for (std::size_t i = 0; i < count; ++i)
193 if (buf_pos == BUF_SIZE)
195 stream.read(reinterpret_cast<char*>(buf), std::min(bytes_left, BUF_SIZE));
199 std::uint8_t byte = buf[buf_pos++];
201 result |= (byte & 0x7f) << shift;
204 if ((byte & 0x80) == 0)
206 out[i] = (sizeof(IntType) * 8 <= shift || (byte & 0x40) == 0)
208 : result | ~((1 << shift) - 1);
211 } while (shift < sizeof(IntType) * 8);
214 assert(bytes_left == 0);
218 // write_leb_128(s, values, N) : write signed integers to a stream with LEB128 compression.
219 // This takes N integers from array values, compress them with the LEB128 algorithm and
220 // writes the result on the stream s.
221 // See https://en.wikipedia.org/wiki/LEB128 for a description of the compression scheme.
222 template<typename IntType>
223 inline void write_leb_128(std::ostream& stream, const IntType* values, std::size_t count) {
225 // Write our LEB128 magic string
226 stream.write(Leb128MagicString, Leb128MagicStringSize);
228 static_assert(std::is_signed_v<IntType>, "Not implemented for unsigned types");
230 std::uint32_t byte_count = 0;
231 for (std::size_t i = 0; i < count; ++i)
233 IntType value = values[i];
240 } while ((byte & 0x40) == 0 ? value != 0 : value != -1);
243 write_little_endian(stream, byte_count);
245 const std::uint32_t BUF_SIZE = 4096;
246 std::uint8_t buf[BUF_SIZE];
247 std::uint32_t buf_pos = 0;
252 stream.write(reinterpret_cast<char*>(buf), buf_pos);
257 auto write = [&](std::uint8_t byte) {
258 buf[buf_pos++] = byte;
259 if (buf_pos == BUF_SIZE)
263 for (std::size_t i = 0; i < count; ++i)
265 IntType value = values[i];
268 std::uint8_t byte = value & 0x7f;
270 if ((byte & 0x40) == 0 ? value == 0 : value == -1)
282 } // namespace Stockfish::Eval::NNUE
284 #endif // #ifndef NNUE_COMMON_H_INCLUDED