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 // Code for calculating NNUE evaluation function
26 #include <string_view>
28 #include "../evaluate.h"
29 #include "../position.h"
33 #include "evaluate_nnue.h"
35 namespace Stockfish::Eval::NNUE {
37 // Input feature converter
38 LargePagePtr<FeatureTransformer> featureTransformer;
40 // Evaluation function
41 AlignedPtr<Network> network[LayerStacks];
43 // Evaluation function file name
45 std::string netDescription;
49 // Initialize the evaluation function parameters
51 void initialize(AlignedPtr<T>& pointer) {
53 pointer.reset(reinterpret_cast<T*>(std_aligned_alloc(alignof(T), sizeof(T))));
54 std::memset(pointer.get(), 0, sizeof(T));
58 void initialize(LargePagePtr<T>& pointer) {
60 static_assert(alignof(T) <= 4096, "aligned_large_pages_alloc() may fail for such a big alignment requirement of T");
61 pointer.reset(reinterpret_cast<T*>(aligned_large_pages_alloc(sizeof(T))));
62 std::memset(pointer.get(), 0, sizeof(T));
65 // Read evaluation function parameters
67 bool read_parameters(std::istream& stream, T& reference) {
70 header = read_little_endian<std::uint32_t>(stream);
71 if (!stream || header != T::get_hash_value()) return false;
72 return reference.read_parameters(stream);
75 // Write evaluation function parameters
77 bool write_parameters(std::ostream& stream, const T& reference) {
79 write_little_endian<std::uint32_t>(stream, T::get_hash_value());
80 return reference.write_parameters(stream);
85 // Initialize the evaluation function parameters
86 static void initialize() {
88 Detail::initialize(featureTransformer);
89 for (std::size_t i = 0; i < LayerStacks; ++i)
90 Detail::initialize(network[i]);
93 // Read network header
94 static bool read_header(std::istream& stream, std::uint32_t* hashValue, std::string* desc)
96 std::uint32_t version, size;
98 version = read_little_endian<std::uint32_t>(stream);
99 *hashValue = read_little_endian<std::uint32_t>(stream);
100 size = read_little_endian<std::uint32_t>(stream);
101 if (!stream || version != Version) return false;
103 stream.read(&(*desc)[0], size);
104 return !stream.fail();
107 // Write network header
108 static bool write_header(std::ostream& stream, std::uint32_t hashValue, const std::string& desc)
110 write_little_endian<std::uint32_t>(stream, Version);
111 write_little_endian<std::uint32_t>(stream, hashValue);
112 write_little_endian<std::uint32_t>(stream, (std::uint32_t)desc.size());
113 stream.write(&desc[0], desc.size());
114 return !stream.fail();
117 // Read network parameters
118 static bool read_parameters(std::istream& stream) {
120 std::uint32_t hashValue;
121 if (!read_header(stream, &hashValue, &netDescription)) return false;
122 if (hashValue != HashValue) return false;
123 if (!Detail::read_parameters(stream, *featureTransformer)) return false;
124 for (std::size_t i = 0; i < LayerStacks; ++i)
125 if (!Detail::read_parameters(stream, *(network[i]))) return false;
126 return stream && stream.peek() == std::ios::traits_type::eof();
129 // Write network parameters
130 static bool write_parameters(std::ostream& stream) {
132 if (!write_header(stream, HashValue, netDescription)) return false;
133 if (!Detail::write_parameters(stream, *featureTransformer)) return false;
134 for (std::size_t i = 0; i < LayerStacks; ++i)
135 if (!Detail::write_parameters(stream, *(network[i]))) return false;
139 void hint_common_parent_position(const Position& pos) {
141 featureTransformer->hint_common_access(pos);
144 // Evaluation function. Perform differential calculation.
145 Value evaluate(const Position& pos, bool adjusted, int* complexity) {
147 // We manually align the arrays on the stack because with gcc < 9.3
148 // overaligning stack variables with alignas() doesn't work correctly.
150 constexpr uint64_t alignment = CacheLineSize;
151 int delta = 24 - pos.non_pawn_material() / 9560;
153 #if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
154 TransformedFeatureType transformedFeaturesUnaligned[
155 FeatureTransformer::BufferSize + alignment / sizeof(TransformedFeatureType)];
157 auto* transformedFeatures = align_ptr_up<alignment>(&transformedFeaturesUnaligned[0]);
160 TransformedFeatureType transformedFeatures[FeatureTransformer::BufferSize];
163 ASSERT_ALIGNED(transformedFeatures, alignment);
165 const int bucket = (pos.count<ALL_PIECES>() - 1) / 4;
166 const auto psqt = featureTransformer->transform(pos, transformedFeatures, bucket);
167 const auto positional = network[bucket]->propagate(transformedFeatures);
170 *complexity = abs(psqt - positional) / OutputScale;
172 // Give more value to positional evaluation when adjusted flag is set
174 return static_cast<Value>(((1024 - delta) * psqt + (1024 + delta) * positional) / (1024 * OutputScale));
176 return static_cast<Value>((psqt + positional) / OutputScale);
179 struct NnueEvalTrace {
180 static_assert(LayerStacks == PSQTBuckets);
182 Value psqt[LayerStacks];
183 Value positional[LayerStacks];
184 std::size_t correctBucket;
187 static NnueEvalTrace trace_evaluate(const Position& pos) {
189 // We manually align the arrays on the stack because with gcc < 9.3
190 // overaligning stack variables with alignas() doesn't work correctly.
192 constexpr uint64_t alignment = CacheLineSize;
194 #if defined(ALIGNAS_ON_STACK_VARIABLES_BROKEN)
195 TransformedFeatureType transformedFeaturesUnaligned[
196 FeatureTransformer::BufferSize + alignment / sizeof(TransformedFeatureType)];
198 auto* transformedFeatures = align_ptr_up<alignment>(&transformedFeaturesUnaligned[0]);
201 TransformedFeatureType transformedFeatures[FeatureTransformer::BufferSize];
204 ASSERT_ALIGNED(transformedFeatures, alignment);
207 t.correctBucket = (pos.count<ALL_PIECES>() - 1) / 4;
208 for (IndexType bucket = 0; bucket < LayerStacks; ++bucket) {
209 const auto materialist = featureTransformer->transform(pos, transformedFeatures, bucket);
210 const auto positional = network[bucket]->propagate(transformedFeatures);
212 t.psqt[bucket] = static_cast<Value>( materialist / OutputScale );
213 t.positional[bucket] = static_cast<Value>( positional / OutputScale );
219 constexpr std::string_view PieceToChar(" PNBRQK pnbrqk");
222 // format_cp_compact() converts a Value into (centi)pawns and writes it in a buffer.
223 // The buffer must have capacity for at least 5 chars.
224 static void format_cp_compact(Value v, char* buffer) {
226 buffer[0] = (v < 0 ? '-' : v > 0 ? '+' : ' ');
228 int cp = std::abs(100 * v / UCI::NormalizeToPawnValue);
231 buffer[1] = '0' + cp / 10000; cp %= 10000;
232 buffer[2] = '0' + cp / 1000; cp %= 1000;
233 buffer[3] = '0' + cp / 100;
238 buffer[1] = '0' + cp / 1000; cp %= 1000;
239 buffer[2] = '0' + cp / 100; cp %= 100;
241 buffer[4] = '0' + cp / 10;
245 buffer[1] = '0' + cp / 100; cp %= 100;
247 buffer[3] = '0' + cp / 10; cp %= 10;
248 buffer[4] = '0' + cp / 1;
253 // format_cp_aligned_dot() converts a Value into (centi)pawns, always keeping two decimals.
254 static void format_cp_aligned_dot(Value v, std::stringstream &stream) {
255 const double cp = 1.0 * std::abs(int(v)) / UCI::NormalizeToPawnValue;
257 stream << (v < 0 ? '-' : v > 0 ? '+' : ' ')
258 << std::setiosflags(std::ios::fixed)
260 << std::setprecision(2)
265 // trace() returns a string with the value of each piece on a board,
266 // and a table for (PSQT, Layers) values bucket by bucket.
268 std::string trace(Position& pos) {
270 std::stringstream ss;
272 char board[3*8+1][8*8+2];
273 std::memset(board, ' ', sizeof(board));
274 for (int row = 0; row < 3*8+1; ++row)
275 board[row][8*8+1] = '\0';
277 // A lambda to output one box of the board
278 auto writeSquare = [&board](File file, Rank rank, Piece pc, Value value) {
280 const int x = ((int)file) * 8;
281 const int y = (7 - (int)rank) * 3;
282 for (int i = 1; i < 8; ++i)
283 board[y][x+i] = board[y+3][x+i] = '-';
284 for (int i = 1; i < 3; ++i)
285 board[y+i][x] = board[y+i][x+8] = '|';
286 board[y][x] = board[y][x+8] = board[y+3][x+8] = board[y+3][x] = '+';
288 board[y+1][x+4] = PieceToChar[pc];
289 if (value != VALUE_NONE)
290 format_cp_compact(value, &board[y+2][x+2]);
293 // We estimate the value of each piece by doing a differential evaluation from
294 // the current base eval, simulating the removal of the piece from its square.
295 Value base = evaluate(pos);
296 base = pos.side_to_move() == WHITE ? base : -base;
298 for (File f = FILE_A; f <= FILE_H; ++f)
299 for (Rank r = RANK_1; r <= RANK_8; ++r)
301 Square sq = make_square(f, r);
302 Piece pc = pos.piece_on(sq);
303 Value v = VALUE_NONE;
305 if (pc != NO_PIECE && type_of(pc) != KING)
307 auto st = pos.state();
309 pos.remove_piece(sq);
310 st->accumulator.computed[WHITE] = false;
311 st->accumulator.computed[BLACK] = false;
313 Value eval = evaluate(pos);
314 eval = pos.side_to_move() == WHITE ? eval : -eval;
317 pos.put_piece(pc, sq);
318 st->accumulator.computed[WHITE] = false;
319 st->accumulator.computed[BLACK] = false;
322 writeSquare(f, r, pc, v);
325 ss << " NNUE derived piece values:\n";
326 for (int row = 0; row < 3*8+1; ++row)
327 ss << board[row] << '\n';
330 auto t = trace_evaluate(pos);
332 ss << " NNUE network contributions "
333 << (pos.side_to_move() == WHITE ? "(White to move)" : "(Black to move)") << std::endl
334 << "+------------+------------+------------+------------+\n"
335 << "| Bucket | Material | Positional | Total |\n"
336 << "| | (PSQT) | (Layers) | |\n"
337 << "+------------+------------+------------+------------+\n";
339 for (std::size_t bucket = 0; bucket < LayerStacks; ++bucket)
341 ss << "| " << bucket << " ";
342 ss << " | "; format_cp_aligned_dot(t.psqt[bucket], ss); ss << " "
343 << " | "; format_cp_aligned_dot(t.positional[bucket], ss); ss << " "
344 << " | "; format_cp_aligned_dot(t.psqt[bucket] + t.positional[bucket], ss); ss << " "
346 if (bucket == t.correctBucket)
347 ss << " <-- this bucket is used";
351 ss << "+------------+------------+------------+------------+\n";
357 // Load eval, from a file stream or a memory stream
358 bool load_eval(std::string name, std::istream& stream) {
362 return read_parameters(stream);
365 // Save eval, to a file stream or a memory stream
366 bool save_eval(std::ostream& stream) {
368 if (fileName.empty())
371 return write_parameters(stream);
374 /// Save eval, to a file given by its name
375 bool save_eval(const std::optional<std::string>& filename) {
377 std::string actualFilename;
380 if (filename.has_value())
381 actualFilename = filename.value();
384 if (currentEvalFileName != EvalFileDefaultName)
386 msg = "Failed to export a net. A non-embedded net can only be saved if the filename is specified";
388 sync_cout << msg << sync_endl;
391 actualFilename = EvalFileDefaultName;
394 std::ofstream stream(actualFilename, std::ios_base::binary);
395 bool saved = save_eval(stream);
397 msg = saved ? "Network saved successfully to " + actualFilename
398 : "Failed to export a net";
400 sync_cout << msg << sync_endl;
405 } // namespace Stockfish::Eval::NNUE