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_MMX)
48 #elif defined(USE_NEON)
52 namespace Stockfish::Eval::NNUE {
54 // Version of the evaluation file
55 constexpr std::uint32_t Version = 0x7AF32F20u;
57 // Constant used in evaluation value calculation
58 constexpr int OutputScale = 16;
59 constexpr int WeightScaleBits = 6;
61 // Size of cache line (in bytes)
62 constexpr std::size_t CacheLineSize = 64;
64 constexpr const char Leb128MagicString[] = "COMPRESSED_LEB128";
65 constexpr const std::size_t Leb128MagicStringSize = sizeof(Leb128MagicString) - 1;
67 // SIMD width (in bytes)
69 constexpr std::size_t SimdWidth = 32;
71 #elif defined(USE_SSE2)
72 constexpr std::size_t SimdWidth = 16;
74 #elif defined(USE_MMX)
75 constexpr std::size_t SimdWidth = 8;
77 #elif defined(USE_NEON)
78 constexpr std::size_t SimdWidth = 16;
81 constexpr std::size_t MaxSimdWidth = 32;
83 // Type of input feature after conversion
84 using TransformedFeatureType = std::uint8_t;
85 using IndexType = std::uint32_t;
87 // Round n up to be a multiple of base
88 template <typename IntType>
89 constexpr IntType ceil_to_multiple(IntType n, IntType base) {
90 return (n + base - 1) / base * base;
94 // read_little_endian() is our utility to read an integer (signed or unsigned, any size)
95 // from a stream in little-endian order. We swap the byte order after the read if
96 // necessary to return a result with the byte ordering of the compiling machine.
97 template <typename IntType>
98 inline IntType read_little_endian(std::istream& stream) {
102 stream.read(reinterpret_cast<char*>(&result), sizeof(IntType));
105 std::uint8_t u[sizeof(IntType)];
106 typename std::make_unsigned<IntType>::type v = 0;
108 stream.read(reinterpret_cast<char*>(u), sizeof(IntType));
109 for (std::size_t i = 0; i < sizeof(IntType); ++i)
110 v = (v << 8) | u[sizeof(IntType) - i - 1];
112 std::memcpy(&result, &v, sizeof(IntType));
119 // write_little_endian() is our utility to write an integer (signed or unsigned, any size)
120 // to a stream in little-endian order. We swap the byte order before the write if
121 // necessary to always write in little endian order, independently of the byte
122 // ordering of the compiling machine.
123 template <typename IntType>
124 inline void write_little_endian(std::ostream& stream, IntType value) {
127 stream.write(reinterpret_cast<const char*>(&value), sizeof(IntType));
130 std::uint8_t u[sizeof(IntType)];
131 typename std::make_unsigned<IntType>::type v = value;
134 // if constexpr to silence the warning about shift by 8
135 if constexpr (sizeof(IntType) > 1)
137 for (; i + 1 < sizeof(IntType); ++i)
139 u[i] = (std::uint8_t)v;
143 u[i] = (std::uint8_t)v;
145 stream.write(reinterpret_cast<char*>(u), sizeof(IntType));
150 // read_little_endian(s, out, N) : read integers in bulk from a little indian stream.
151 // This reads N integers from stream s and put them in array out.
152 template <typename IntType>
153 inline void read_little_endian(std::istream& stream, IntType* out, std::size_t count) {
155 stream.read(reinterpret_cast<char*>(out), sizeof(IntType) * count);
157 for (std::size_t i = 0; i < count; ++i)
158 out[i] = read_little_endian<IntType>(stream);
162 // write_little_endian(s, values, N) : write integers in bulk to a little indian stream.
163 // This takes N integers from array values and writes them on stream s.
164 template <typename IntType>
165 inline void write_little_endian(std::ostream& stream, const IntType* values, std::size_t count) {
167 stream.write(reinterpret_cast<const char*>(values), sizeof(IntType) * count);
169 for (std::size_t i = 0; i < count; ++i)
170 write_little_endian<IntType>(stream, values[i]);
174 // read_leb_128(s, out, N) : read N signed integers from the stream s, putting them in
175 // the array out. The stream is assumed to be compressed using the signed LEB128 format.
176 // See https://en.wikipedia.org/wiki/LEB128 for a description of the compression scheme.
177 template <typename IntType>
178 inline void read_leb_128(std::istream& stream, IntType* out, std::size_t count) {
180 // Check the presence of our LEB128 magic string
181 char leb128MagicString[Leb128MagicStringSize];
182 stream.read(leb128MagicString, Leb128MagicStringSize);
183 assert(strncmp(Leb128MagicString, leb128MagicString, Leb128MagicStringSize) == 0);
185 static_assert(std::is_signed_v<IntType>, "Not implemented for unsigned types");
187 const std::uint32_t BUF_SIZE = 4096;
188 std::uint8_t buf[BUF_SIZE];
190 auto bytes_left = read_little_endian<std::uint32_t>(stream);
192 std::uint32_t buf_pos = BUF_SIZE;
193 for (std::size_t i = 0; i < count; ++i)
199 if (buf_pos == BUF_SIZE)
201 stream.read(reinterpret_cast<char*>(buf), std::min(bytes_left, BUF_SIZE));
205 std::uint8_t byte = buf[buf_pos++];
207 result |= (byte & 0x7f) << shift;
210 if ((byte & 0x80) == 0)
212 out[i] = (sizeof(IntType) * 8 <= shift || (byte & 0x40) == 0) ? result
213 : result | ~((1 << shift) - 1);
217 while (shift < sizeof(IntType) * 8);
220 assert(bytes_left == 0);
224 // write_leb_128(s, values, N) : write signed integers to a stream with LEB128 compression.
225 // This takes N integers from array values, compress them with the LEB128 algorithm and
226 // writes the result on the stream s.
227 // See https://en.wikipedia.org/wiki/LEB128 for a description of the compression scheme.
228 template <typename IntType>
229 inline void write_leb_128(std::ostream& stream, const IntType* values, std::size_t count) {
231 // Write our LEB128 magic string
232 stream.write(Leb128MagicString, Leb128MagicStringSize);
234 static_assert(std::is_signed_v<IntType>, "Not implemented for unsigned types");
236 std::uint32_t byte_count = 0;
237 for (std::size_t i = 0; i < count; ++i)
239 IntType value = values[i];
247 while ((byte & 0x40) == 0 ? value != 0 : value != -1);
250 write_little_endian(stream, byte_count);
252 const std::uint32_t BUF_SIZE = 4096;
253 std::uint8_t buf[BUF_SIZE];
254 std::uint32_t buf_pos = 0;
259 stream.write(reinterpret_cast<char*>(buf), buf_pos);
264 auto write = [&](std::uint8_t byte) {
265 buf[buf_pos++] = byte;
266 if (buf_pos == BUF_SIZE)
270 for (std::size_t i = 0; i < count; ++i)
272 IntType value = values[i];
275 std::uint8_t byte = value & 0x7f;
277 if ((byte & 0x40) == 0 ? value == 0 : value == -1)
289 } // namespace Stockfish::Eval::NNUE
291 #endif // #ifndef NNUE_COMMON_H_INCLUDED