diff --git a/src/bitboard.cpp b/src/bitboard.cpp index 32c626d47..c842ca127 100644 --- a/src/bitboard.cpp +++ b/src/bitboard.cpp @@ -124,8 +124,14 @@ Bitboard sliding_attack(PieceType pt, Square sq, Bitboard occupied) { for (Direction d : (pt == ROOK ? RookDirections : BishopDirections)) { Square s = sq; - while (safe_destination(s, d) && !(occupied & s)) + while (safe_destination(s, d)) + { attacks |= (s += d); + if (occupied & s) + { + break; + } + } } return attacks; diff --git a/src/evaluate.cpp b/src/evaluate.cpp index 5be7e7a1b..ca09aaf9e 100644 --- a/src/evaluate.cpp +++ b/src/evaluate.cpp @@ -44,6 +44,10 @@ int Eval::simple_eval(const Position& pos, Color c) { + (pos.non_pawn_material(c) - pos.non_pawn_material(~c)); } +bool Eval::use_smallnet(const Position& pos) { + int simpleEval = simple_eval(pos, pos.side_to_move()); + return std::abs(simpleEval) > 1018 + 5 * pos.count(); +} // Evaluate is the evaluator for the outer world. It returns a static evaluation // of the position from the point of view of the side to move. @@ -55,33 +59,33 @@ Value Eval::evaluate(const Eval::NNUE::Networks& networks, assert(!pos.checkers()); int simpleEval = simple_eval(pos, pos.side_to_move()); - bool smallNet = std::abs(simpleEval) > SmallNetThreshold; + bool smallNet = use_smallnet(pos); int nnueComplexity; int v; Value nnue = smallNet ? networks.small.evaluate(pos, &caches.small, true, &nnueComplexity) : networks.big.evaluate(pos, &caches.big, true, &nnueComplexity); - const auto adjustEval = [&](int nnueDiv, int pawnCountConstant, int pawnCountMul, - int npmConstant, int evalDiv, int shufflingConstant) { - // Blend optimism and eval with nnue complexity and material imbalance - optimism += optimism * (nnueComplexity + std::abs(simpleEval - nnue)) / 584; - nnue -= nnue * (nnueComplexity * 5 / 3) / nnueDiv; + if (smallNet && nnue * simpleEval < 0) + { + nnue = networks.big.evaluate(pos, &caches.big, true, &nnueComplexity); + smallNet = false; + } - int npm = pos.non_pawn_material() / 64; - v = (nnue * (npm + pawnCountConstant + pawnCountMul * pos.count()) - + optimism * (npmConstant + npm)) - / evalDiv; + // Blend optimism and eval with nnue complexity and material imbalance + optimism += optimism * (nnueComplexity + std::abs(simpleEval - nnue)) / 620; + nnue -= nnue * (nnueComplexity * 5 / 3) / 32082; - // Damp down the evaluation linearly when shuffling - int shuffling = pos.rule50_count(); - v = v * (shufflingConstant - shuffling) / 207; - }; + v = (nnue + * (32961 + 381 * pos.count() + 349 * pos.count() + + 392 * pos.count() + 649 * pos.count() + 1211 * pos.count()) + + optimism + * (4835 + 136 * pos.count() + 375 * pos.count() + + 403 * pos.count() + 628 * pos.count() + 1124 * pos.count())) + / 36860; - if (!smallNet) - adjustEval(32395, 942, 11, 139, 1058, 178); - else - adjustEval(32793, 944, 9, 140, 1067, 206); + // Damp down the evaluation linearly when shuffling + v = v * (204 - pos.rule50_count()) / 208; // Guarantee evaluation does not hit the tablebase range v = std::clamp(v, VALUE_TB_LOSS_IN_MAX_PLY + 1, VALUE_TB_WIN_IN_MAX_PLY - 1); diff --git a/src/evaluate.h b/src/evaluate.h index 2d244ff67..4b3e91acf 100644 --- a/src/evaluate.h +++ b/src/evaluate.h @@ -29,13 +29,11 @@ class Position; namespace Eval { -constexpr inline int SmallNetThreshold = 1274; - // The default net name MUST follow the format nn-[SHA256 first 12 digits].nnue // for the build process (profile-build and fishtest) to work. Do not change the // name of the macro or the location where this macro is defined, as it is used // in the Makefile/Fishtest. -#define EvalFileDefaultNameBig "nn-ae6a388e4a1a.nnue" +#define EvalFileDefaultNameBig "nn-c721dfca8cd3.nnue" #define EvalFileDefaultNameSmall "nn-baff1ede1f90.nnue" namespace NNUE { @@ -46,6 +44,7 @@ struct AccumulatorCaches; std::string trace(Position& pos, const Eval::NNUE::Networks& networks); int simple_eval(const Position& pos, Color c); +bool use_smallnet(const Position& pos); Value evaluate(const NNUE::Networks& networks, const Position& pos, Eval::NNUE::AccumulatorCaches& caches, diff --git a/src/nnue/layers/clipped_relu.h b/src/nnue/layers/clipped_relu.h index 813234c59..2ee378ad8 100644 --- a/src/nnue/layers/clipped_relu.h +++ b/src/nnue/layers/clipped_relu.h @@ -65,41 +65,37 @@ class ClippedReLU { if constexpr (InputDimensions % SimdWidth == 0) { constexpr IndexType NumChunks = InputDimensions / SimdWidth; - const __m256i Zero = _mm256_setzero_si256(); const __m256i Offsets = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0); const auto in = reinterpret_cast(input); const auto out = reinterpret_cast<__m256i*>(output); for (IndexType i = 0; i < NumChunks; ++i) { const __m256i words0 = - _mm256_srai_epi16(_mm256_packs_epi32(_mm256_load_si256(&in[i * 4 + 0]), - _mm256_load_si256(&in[i * 4 + 1])), + _mm256_srli_epi16(_mm256_packus_epi32(_mm256_load_si256(&in[i * 4 + 0]), + _mm256_load_si256(&in[i * 4 + 1])), WeightScaleBits); const __m256i words1 = - _mm256_srai_epi16(_mm256_packs_epi32(_mm256_load_si256(&in[i * 4 + 2]), - _mm256_load_si256(&in[i * 4 + 3])), + _mm256_srli_epi16(_mm256_packus_epi32(_mm256_load_si256(&in[i * 4 + 2]), + _mm256_load_si256(&in[i * 4 + 3])), WeightScaleBits); - _mm256_store_si256( - &out[i], _mm256_permutevar8x32_epi32( - _mm256_max_epi8(_mm256_packs_epi16(words0, words1), Zero), Offsets)); + _mm256_store_si256(&out[i], _mm256_permutevar8x32_epi32( + _mm256_packs_epi16(words0, words1), Offsets)); } } else { constexpr IndexType NumChunks = InputDimensions / (SimdWidth / 2); - const __m128i Zero = _mm_setzero_si128(); const auto in = reinterpret_cast(input); const auto out = reinterpret_cast<__m128i*>(output); for (IndexType i = 0; i < NumChunks; ++i) { - const __m128i words0 = _mm_srai_epi16( - _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 0]), _mm_load_si128(&in[i * 4 + 1])), + const __m128i words0 = _mm_srli_epi16( + _mm_packus_epi32(_mm_load_si128(&in[i * 4 + 0]), _mm_load_si128(&in[i * 4 + 1])), WeightScaleBits); - const __m128i words1 = _mm_srai_epi16( - _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 2]), _mm_load_si128(&in[i * 4 + 3])), + const __m128i words1 = _mm_srli_epi16( + _mm_packus_epi32(_mm_load_si128(&in[i * 4 + 2]), _mm_load_si128(&in[i * 4 + 3])), WeightScaleBits); - const __m128i packedbytes = _mm_packs_epi16(words0, words1); - _mm_store_si128(&out[i], _mm_max_epi8(packedbytes, Zero)); + _mm_store_si128(&out[i], _mm_packs_epi16(words0, words1)); } } constexpr IndexType Start = InputDimensions % SimdWidth == 0 @@ -109,9 +105,7 @@ class ClippedReLU { #elif defined(USE_SSE2) constexpr IndexType NumChunks = InputDimensions / SimdWidth; - #ifdef USE_SSE41 - const __m128i Zero = _mm_setzero_si128(); - #else + #ifndef USE_SSE41 const __m128i k0x80s = _mm_set1_epi8(-128); #endif @@ -119,6 +113,15 @@ class ClippedReLU { const auto out = reinterpret_cast<__m128i*>(output); for (IndexType i = 0; i < NumChunks; ++i) { + #if defined(USE_SSE41) + const __m128i words0 = _mm_srli_epi16( + _mm_packus_epi32(_mm_load_si128(&in[i * 4 + 0]), _mm_load_si128(&in[i * 4 + 1])), + WeightScaleBits); + const __m128i words1 = _mm_srli_epi16( + _mm_packus_epi32(_mm_load_si128(&in[i * 4 + 2]), _mm_load_si128(&in[i * 4 + 3])), + WeightScaleBits); + _mm_store_si128(&out[i], _mm_packs_epi16(words0, words1)); + #else const __m128i words0 = _mm_srai_epi16( _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 0]), _mm_load_si128(&in[i * 4 + 1])), WeightScaleBits); @@ -126,15 +129,8 @@ class ClippedReLU { _mm_packs_epi32(_mm_load_si128(&in[i * 4 + 2]), _mm_load_si128(&in[i * 4 + 3])), WeightScaleBits); const __m128i packedbytes = _mm_packs_epi16(words0, words1); - _mm_store_si128(&out[i], - - #ifdef USE_SSE41 - _mm_max_epi8(packedbytes, Zero) - #else - _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s) + _mm_store_si128(&out[i], _mm_subs_epi8(_mm_adds_epi8(packedbytes, k0x80s), k0x80s)); #endif - - ); } constexpr IndexType Start = NumChunks * SimdWidth; diff --git a/src/nnue/nnue_feature_transformer.h b/src/nnue/nnue_feature_transformer.h index 2b3248162..2492ccff1 100644 --- a/src/nnue/nnue_feature_transformer.h +++ b/src/nnue/nnue_feature_transformer.h @@ -662,7 +662,10 @@ class FeatureTransformer { for (IndexType j = 0; j < HalfDimensions / TileHeight; ++j) { + auto accTile = + reinterpret_cast(&accumulator.accumulation[Perspective][j * TileHeight]); auto entryTile = reinterpret_cast(&entry.accumulation[j * TileHeight]); + for (IndexType k = 0; k < NumRegs; ++k) acc[k] = entryTile[k]; @@ -677,7 +680,7 @@ class FeatureTransformer { auto columnA = reinterpret_cast(&weights[offsetA]); for (unsigned k = 0; k < NumRegs; ++k) - acc[k] = vec_add_16(vec_sub_16(acc[k], columnR[k]), columnA[k]); + acc[k] = vec_add_16(acc[k], vec_sub_16(columnA[k], columnR[k])); } for (; i < int(removed.size()); ++i) { @@ -700,12 +703,17 @@ class FeatureTransformer { for (IndexType k = 0; k < NumRegs; k++) vec_store(&entryTile[k], acc[k]); + for (IndexType k = 0; k < NumRegs; k++) + vec_store(&accTile[k], acc[k]); } for (IndexType j = 0; j < PSQTBuckets / PsqtTileHeight; ++j) { + auto accTilePsqt = reinterpret_cast( + &accumulator.psqtAccumulation[Perspective][j * PsqtTileHeight]); auto entryTilePsqt = reinterpret_cast(&entry.psqtAccumulation[j * PsqtTileHeight]); + for (std::size_t k = 0; k < NumPsqtRegs; ++k) psqt[k] = entryTilePsqt[k]; @@ -730,6 +738,8 @@ class FeatureTransformer { for (std::size_t k = 0; k < NumPsqtRegs; ++k) vec_store_psqt(&entryTilePsqt[k], psqt[k]); + for (std::size_t k = 0; k < NumPsqtRegs; ++k) + vec_store_psqt(&accTilePsqt[k], psqt[k]); } #else @@ -753,8 +763,6 @@ class FeatureTransformer { entry.psqtAccumulation[k] += psqtWeights[index * PSQTBuckets + k]; } -#endif - // The accumulator of the refresh entry has been updated. // Now copy its content to the actual accumulator we were refreshing @@ -763,6 +771,7 @@ class FeatureTransformer { std::memcpy(accumulator.psqtAccumulation[Perspective], entry.psqtAccumulation, sizeof(int32_t) * PSQTBuckets); +#endif for (Color c : {WHITE, BLACK}) entry.byColorBB[c] = pos.pieces(c); diff --git a/src/nnue/nnue_misc.cpp b/src/nnue/nnue_misc.cpp index bf73a58bf..a13c717c3 100644 --- a/src/nnue/nnue_misc.cpp +++ b/src/nnue/nnue_misc.cpp @@ -45,9 +45,7 @@ constexpr std::string_view PieceToChar(" PNBRQK pnbrqk"); void hint_common_parent_position(const Position& pos, const Networks& networks, AccumulatorCaches& caches) { - - int simpleEvalAbs = std::abs(simple_eval(pos, pos.side_to_move())); - if (simpleEvalAbs > Eval::SmallNetThreshold) + if (Eval::use_smallnet(pos)) networks.small.hint_common_access(pos, &caches.small); else networks.big.hint_common_access(pos, &caches.big); diff --git a/src/search.cpp b/src/search.cpp index e79b25f0b..b58506784 100644 --- a/src/search.cpp +++ b/src/search.cpp @@ -43,6 +43,7 @@ #include "thread.h" #include "timeman.h" #include "tt.h" +#include "types.h" #include "uci.h" #include "ucioption.h" @@ -60,9 +61,9 @@ static constexpr double EvalLevel[10] = {0.981, 0.956, 0.895, 0.949, 0.913, // Futility margin Value futility_margin(Depth d, bool noTtCutNode, bool improving, bool oppWorsening) { - Value futilityMult = 126 - 46 * noTtCutNode; - Value improvingDeduction = 58 * improving * futilityMult / 32; - Value worseningDeduction = (323 + 52 * improving) * oppWorsening * futilityMult / 1024; + Value futilityMult = 127 - 48 * noTtCutNode; + Value improvingDeduction = 65 * improving * futilityMult / 32; + Value worseningDeduction = 334 * oppWorsening * futilityMult / 1024; return futilityMult * d - improvingDeduction - worseningDeduction; } @@ -74,15 +75,15 @@ constexpr int futility_move_count(bool improving, Depth depth) { // Add correctionHistory value to raw staticEval and guarantee evaluation does not hit the tablebase range Value to_corrected_static_eval(Value v, const Worker& w, const Position& pos) { auto cv = w.correctionHistory[pos.side_to_move()][pawn_structure_index(pos)]; - v += cv * std::abs(cv) / 7350; + v += cv * std::abs(cv) / 6047; return std::clamp(v, VALUE_TB_LOSS_IN_MAX_PLY + 1, VALUE_TB_WIN_IN_MAX_PLY - 1); } // History and stats update bonus, based on depth -int stat_bonus(Depth d) { return std::clamp(208 * d - 297, 16, 1406); } +int stat_bonus(Depth d) { return std::clamp(187 * d - 288, 17, 1548); } // History and stats update malus, based on depth -int stat_malus(Depth d) { return (d < 4 ? 520 * d - 312 : 1479); } +int stat_malus(Depth d) { return (d < 4 ? 630 * d - 281 : 1741); } // Add a small random component to draw evaluations to avoid 3-fold blindness Value value_draw(size_t nodes) { return VALUE_DRAW - 1 + Value(nodes & 0x2); } @@ -158,7 +159,8 @@ void Search::Worker::start_searching() { return; } - main_manager()->tm.init(limits, rootPos.side_to_move(), rootPos.game_ply(), options); + main_manager()->tm.init(limits, rootPos.side_to_move(), rootPos.game_ply(), options, + main_manager()->originalPly); tt.new_search(); if (rootMoves.empty()) @@ -352,12 +354,12 @@ void Search::Worker::iterative_deepening() { // Reset aspiration window starting size Value avg = rootMoves[pvIdx].averageScore; - delta = 10 + avg * avg / 9530; + delta = 10 + avg * avg / 9828; alpha = std::max(avg - delta, -VALUE_INFINITE); beta = std::min(avg + delta, VALUE_INFINITE); // Adjust optimism based on root move's averageScore (~4 Elo) - optimism[us] = 119 * avg / (std::abs(avg) + 88); + optimism[us] = 116 * avg / (std::abs(avg) + 84); optimism[~us] = -optimism[us]; // Start with a small aspiration window and, in the case of a fail @@ -543,7 +545,7 @@ void Search::Worker::clear() { counterMoves.fill(Move::none()); mainHistory.fill(0); captureHistory.fill(0); - pawnHistory.fill(0); + pawnHistory.fill(-1300); correctionHistory.fill(0); for (bool inCheck : {false, true}) @@ -553,7 +555,7 @@ void Search::Worker::clear() { h->fill(-60); for (size_t i = 1; i < reductions.size(); ++i) - reductions[i] = int((18.93 + std::log(size_t(options["Threads"])) / 2) * std::log(i)); + reductions[i] = int((21.69 + std::log(size_t(options["Threads"])) / 2) * std::log(i)); refreshTable.clear(networks); } @@ -786,7 +788,7 @@ Value Search::Worker::search( // Use static evaluation difference to improve quiet move ordering (~9 Elo) if (((ss - 1)->currentMove).is_ok() && !(ss - 1)->inCheck && !priorCapture) { - int bonus = std::clamp(-13 * int((ss - 1)->staticEval + ss->staticEval), -1796, 1526); + int bonus = std::clamp(-11 * int((ss - 1)->staticEval + ss->staticEval), -1729, 1517); bonus = bonus > 0 ? 2 * bonus : bonus / 2; thisThread->mainHistory[~us][((ss - 1)->currentMove).from_to()] << bonus; if (type_of(pos.piece_on(prevSq)) != PAWN && ((ss - 1)->currentMove).type_of() != PROMOTION) @@ -808,8 +810,7 @@ Value Search::Worker::search( // Step 7. Razoring (~1 Elo) // If eval is really low check with qsearch if it can exceed alpha, if it can't, // return a fail low. - // Adjust razor margin according to cutoffCnt. (~1 Elo) - if (eval < alpha - 433 - (302 - 141 * ((ss + 1)->cutoffCnt > 3)) * depth * depth) + if (eval < alpha - 474 - 324 * depth * depth) { value = qsearch(pos, ss, alpha - 1, alpha); if (value < alpha) @@ -820,21 +821,21 @@ Value Search::Worker::search( // The depth condition is important for mate finding. if (!ss->ttPv && depth < 11 && eval - futility_margin(depth, cutNode && !ss->ttHit, improving, opponentWorsening) - - (ss - 1)->statScore / 254 + - (ss - 1)->statScore / 252 >= beta && eval >= beta && eval < VALUE_TB_WIN_IN_MAX_PLY && (!ttMove || ttCapture)) return beta > VALUE_TB_LOSS_IN_MAX_PLY ? (eval + beta) / 2 : eval; // Step 9. Null move search with verification search (~35 Elo) - if (!PvNode && (ss - 1)->currentMove != Move::null() && (ss - 1)->statScore < 16993 - && eval >= beta && ss->staticEval >= beta - 19 * depth + 326 && !excludedMove + if (!PvNode && (ss - 1)->currentMove != Move::null() && (ss - 1)->statScore < 15246 + && eval >= beta && ss->staticEval >= beta - 21 * depth + 366 && !excludedMove && pos.non_pawn_material(us) && ss->ply >= thisThread->nmpMinPly && beta > VALUE_TB_LOSS_IN_MAX_PLY) { assert(eval - beta >= 0); // Null move dynamic reduction based on depth and eval - Depth R = std::min(int(eval - beta) / 134, 6) + depth / 3 + 4; + Depth R = std::min(int(eval - beta) / 152, 6) + depth / 3 + 5; ss->currentMove = Move::null(); ss->continuationHistory = &thisThread->continuationHistory[0][0][NO_PIECE][0]; @@ -849,7 +850,16 @@ Value Search::Worker::search( if (nullValue >= beta && nullValue < VALUE_TB_WIN_IN_MAX_PLY) { if (thisThread->nmpMinPly || depth < 16) + { + if (nullValue >= ss->staticEval) + { + auto bonus = std::min(int(nullValue - ss->staticEval) * depth / 32, + CORRECTION_HISTORY_LIMIT / 16); + thisThread->correctionHistory[us][pawn_structure_index(pos)] + << bonus; + } return nullValue; + } assert(!thisThread->nmpMinPly); // Recursive verification is not allowed @@ -871,9 +881,12 @@ Value Search::Worker::search( if (PvNode && !ttMove) depth -= 3; + if (!PvNode && ss->ttHit && (tte->bound() & BOUND_UPPER) && ttValue > alpha + 5 * depth) + depth--; + // Use qsearch if depth <= 0. if (depth <= 0) - return qsearch(pos, ss, alpha, beta); + return qsearch(pos, ss, alpha, beta); // For cutNodes without a ttMove, we decrease depth by 2 if depth is high enough. if (cutNode && depth >= 8 && (!ttMove || tte->bound() == BOUND_UPPER)) @@ -882,7 +895,7 @@ Value Search::Worker::search( // Step 11. ProbCut (~10 Elo) // If we have a good enough capture (or queen promotion) and a reduced search returns a value // much above beta, we can (almost) safely prune the previous move. - probCutBeta = beta + 159 - 66 * improving; + probCutBeta = beta + 176 - 65 * improving; if ( !PvNode && depth > 3 && std::abs(beta) < VALUE_TB_WIN_IN_MAX_PLY @@ -939,7 +952,7 @@ Value Search::Worker::search( moves_loop: // When in check, search starts here // Step 12. A small Probcut idea, when we are in check (~4 Elo) - probCutBeta = beta + 420; + probCutBeta = beta + 440; if (ss->inCheck && !PvNode && ttCapture && (tte->bound() & BOUND_LOWER) && tte->depth() >= depth - 4 && ttValue >= probCutBeta && std::abs(ttValue) < VALUE_TB_WIN_IN_MAX_PLY && std::abs(beta) < VALUE_TB_WIN_IN_MAX_PLY) @@ -1009,8 +1022,7 @@ moves_loop: // When in check, search starts here if (!rootNode && pos.non_pawn_material(us) && bestValue > VALUE_TB_LOSS_IN_MAX_PLY) { // Skip quiet moves if movecount exceeds our FutilityMoveCount threshold (~8 Elo) - if (!moveCountPruning) - moveCountPruning = moveCount >= futility_move_count(improving, depth); + moveCountPruning = moveCount >= futility_move_count(improving, depth); // Reduced depth of the next LMR search int lmrDepth = newDepth - r; @@ -1024,15 +1036,15 @@ moves_loop: // When in check, search starts here // Futility pruning for captures (~2 Elo) if (!givesCheck && lmrDepth < 7 && !ss->inCheck) { - Value futilityValue = ss->staticEval + 295 + 280 * lmrDepth + Value futilityValue = ss->staticEval + 276 + 256 * lmrDepth + PieceValue[capturedPiece] + captHist / 7; if (futilityValue <= alpha) continue; } // SEE based pruning for captures and checks (~11 Elo) - int seeHist = std::clamp(captHist / 32, -197 * depth, 196 * depth); - if (!pos.see_ge(move, -186 * depth - seeHist)) + int seeHist = std::clamp(captHist / 32, -177 * depth, 175 * depth); + if (!pos.see_ge(move, -183 * depth - seeHist)) continue; } else @@ -1040,22 +1052,21 @@ moves_loop: // When in check, search starts here int history = (*contHist[0])[movedPiece][move.to_sq()] + (*contHist[1])[movedPiece][move.to_sq()] - + (*contHist[3])[movedPiece][move.to_sq()] / 2 + thisThread->pawnHistory[pawn_structure_index(pos)][movedPiece][move.to_sq()]; // Continuation history based pruning (~2 Elo) - if (lmrDepth < 6 && history < -4081 * depth) + if (lmrDepth < 6 && history < -4076 * depth) continue; history += 2 * thisThread->mainHistory[us][move.from_to()]; - lmrDepth += history / 4768; + lmrDepth += history / 4401; Value futilityValue = - ss->staticEval + (bestValue < ss->staticEval - 52 ? 134 : 54) + 142 * lmrDepth; + ss->staticEval + (bestValue < ss->staticEval - 53 ? 151 : 57) + 140 * lmrDepth; // Futility pruning: parent node (~13 Elo) - if (!ss->inCheck && lmrDepth < 13 && futilityValue <= alpha) + if (!ss->inCheck && lmrDepth < 10 && futilityValue <= alpha) { if (bestValue <= futilityValue && std::abs(bestValue) < VALUE_TB_WIN_IN_MAX_PLY && futilityValue < VALUE_TB_WIN_IN_MAX_PLY) @@ -1066,7 +1077,7 @@ moves_loop: // When in check, search starts here lmrDepth = std::max(lmrDepth, 0); // Prune moves with negative SEE (~4 Elo) - if (!pos.see_ge(move, -28 * lmrDepth * lmrDepth)) + if (!pos.see_ge(move, -26 * lmrDepth * lmrDepth)) continue; } } @@ -1086,11 +1097,11 @@ moves_loop: // When in check, search starts here // so changing them requires tests at these types of time controls. // Recursive singular search is avoided. if (!rootNode && move == ttMove && !excludedMove - && depth >= 4 - (thisThread->completedDepth > 32) + ss->ttPv + && depth >= 4 - (thisThread->completedDepth > 35) + ss->ttPv && std::abs(ttValue) < VALUE_TB_WIN_IN_MAX_PLY && (tte->bound() & BOUND_LOWER) && tte->depth() >= depth - 3) { - Value singularBeta = ttValue - (65 + 52 * (ss->ttPv && !PvNode)) * depth / 63; + Value singularBeta = ttValue - (57 + 50 * (ss->ttPv && !PvNode)) * depth / 64; Depth singularDepth = newDepth / 2; ss->excludedMove = move; @@ -1100,16 +1111,16 @@ moves_loop: // When in check, search starts here if (value < singularBeta) { - int doubleMargin = 251 * PvNode - 241 * !ttCapture; + int doubleMargin = 298 * PvNode - 209 * !ttCapture; int tripleMargin = - 135 + 234 * PvNode - 248 * !ttCapture + 124 * (ss->ttPv || !ttCapture); - int quadMargin = 447 + 354 * PvNode - 300 * !ttCapture + 206 * ss->ttPv; + 117 + 252 * PvNode - 270 * !ttCapture + 111 * (ss->ttPv || !ttCapture); + int quadMargin = 471 + 343 * PvNode - 281 * !ttCapture + 217 * ss->ttPv; extension = 1 + (value < singularBeta - doubleMargin) + (value < singularBeta - tripleMargin) + (value < singularBeta - quadMargin); - depth += ((!PvNode) && (depth < 14)); + depth += ((!PvNode) && (depth < 15)); } // Multi-cut pruning @@ -1138,17 +1149,13 @@ moves_loop: // When in check, search starts here // If we are on a cutNode but the ttMove is not assumed to fail high over current beta (~1 Elo) else if (cutNode) extension = -2; - - // If the ttMove is assumed to fail low over the value of the reduced search (~1 Elo) - else if (ttValue <= value) - extension = -1; } // Extension for capturing the previous moved piece (~0 Elo on STC, ~1 Elo on LTC) else if (PvNode && move == ttMove && move.to_sq() == prevSq && thisThread->captureHistory[movedPiece][move.to_sq()] [type_of(pos.piece_on(move.to_sq()))] - > 4016) + > 3748) extension = 1; } @@ -1199,11 +1206,10 @@ moves_loop: // When in check, search starts here ss->statScore = 2 * thisThread->mainHistory[us][move.from_to()] + (*contHist[0])[movedPiece][move.to_sq()] - + (*contHist[1])[movedPiece][move.to_sq()] - + (*contHist[3])[movedPiece][move.to_sq()] - 5078; + + (*contHist[1])[movedPiece][move.to_sq()] - 5266; // Decrease/increase reduction for moves with a good/bad history (~8 Elo) - r -= ss->statScore / (17662 - std::min(depth, 16) * 105); + r -= ss->statScore / (14519 - std::min(depth, 15) * 103); // Step 17. Late moves reduction / extension (LMR, ~117 Elo) if (depth >= 2 && moveCount > 1 + rootNode) @@ -1340,7 +1346,7 @@ moves_loop: // When in check, search starts here else { // Reduce other moves if we have found at least one score improvement (~2 Elo) - if (depth > 2 && depth < 13 && beta < 15868 && value > -14630) + if (depth > 2 && depth < 13 && std::abs(value) < VALUE_TB_WIN_IN_MAX_PLY) depth -= 2; assert(depth > 0); @@ -1383,13 +1389,18 @@ moves_loop: // When in check, search starts here // Bonus for prior countermove that caused the fail low else if (!priorCapture && prevSq != SQ_NONE) { - int bonus = (depth > 5) + (PvNode || cutNode) + ((ss - 1)->statScore < -14455) - + ((ss - 1)->moveCount > 10) + (!ss->inCheck && bestValue <= ss->staticEval - 130) - + (!(ss - 1)->inCheck && bestValue <= -(ss - 1)->staticEval - 77); + int bonus = (depth > 4) + (depth > 5) + (PvNode || cutNode) + ((ss - 1)->statScore < -13241) + + ((ss - 1)->moveCount > 10) + (!ss->inCheck && bestValue <= ss->staticEval - 127) + + (!(ss - 1)->inCheck && bestValue <= -(ss - 1)->staticEval - 74); update_continuation_histories(ss - 1, pos.piece_on(prevSq), prevSq, stat_bonus(depth) * bonus); thisThread->mainHistory[~us][((ss - 1)->currentMove).from_to()] << stat_bonus(depth) * bonus / 2; + + + if (type_of(pos.piece_on(prevSq)) != PAWN && ((ss - 1)->currentMove).type_of() != PROMOTION) + thisThread->pawnHistory[pawn_structure_index(pos)][pos.piece_on(prevSq)][prevSq] + << stat_bonus(depth) * bonus * 4; } if (PvNode) @@ -1535,6 +1546,8 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta, // Stand pat. Return immediately if static value is at least beta if (bestValue >= beta) { + if (std::abs(bestValue) < VALUE_TB_WIN_IN_MAX_PLY && !PvNode) + bestValue = (3 * bestValue + beta) / 4; if (!ss->ttHit) Cluster::save(tt, threads, thisThread, tte, posKey, value_to_tt(bestValue, ss->ply), false, BOUND_LOWER, DEPTH_NONE, Move::none(), unadjustedStaticEval, @@ -1546,7 +1559,7 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta, if (bestValue > alpha) alpha = bestValue; - futilityBase = ss->staticEval + 270; + futilityBase = ss->staticEval + 264; } const PieceToHistory* contHist[] = {(ss - 1)->continuationHistory, @@ -1560,8 +1573,6 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta, MovePicker mp(pos, ttMove, depth, &thisThread->mainHistory, &thisThread->captureHistory, contHist, &thisThread->pawnHistory); - int quietCheckEvasions = 0; - // Step 5. Loop through all pseudo-legal moves until no moves remain // or a beta cutoff occurs. while ((move = mp.next_move()) != Move::none()) @@ -1614,23 +1625,17 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta, } } - // We prune after the second quiet check evasion move, where being 'in check' is - // implicitly checked through the counter, and being a 'quiet move' apart from - // being a tt move is assumed after an increment because captures are pushed ahead. - if (quietCheckEvasions > 1) - break; - // Continuation history based pruning (~3 Elo) if (!capture && (*contHist[0])[pos.moved_piece(move)][move.to_sq()] + (*contHist[1])[pos.moved_piece(move)][move.to_sq()] + thisThread->pawnHistory[pawn_structure_index(pos)][pos.moved_piece(move)] [move.to_sq()] - <= 4000) + <= 4348) continue; // Do not search moves with bad enough SEE values (~5 Elo) - if (!pos.see_ge(move, -69)) + if (!pos.see_ge(move, -63)) continue; } @@ -1643,8 +1648,6 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta, &thisThread ->continuationHistory[ss->inCheck][capture][pos.moved_piece(move)][move.to_sq()]; - quietCheckEvasions += !capture && ss->inCheck; - // Step 7. Make and search the move thisThread->nodes.fetch_add(1, std::memory_order_relaxed); pos.do_move(move, st, givesCheck); @@ -1698,7 +1701,7 @@ Value Search::Worker::qsearch(Position& pos, Stack* ss, Value alpha, Value beta, Depth Search::Worker::reduction(bool i, Depth d, int mn, int delta) { int reductionScale = reductions[d] * reductions[mn]; - return (reductionScale + 1318 - delta * 760 / rootDelta) / 1024 + (!i && reductionScale > 1066); + return (reductionScale + 1147 - delta * 755 / rootDelta) / 1024 + (!i && reductionScale > 1125); } // elapsed() returns the time elapsed since the search started. If the @@ -1839,6 +1842,8 @@ void update_all_stats(const Position& pos, // by moves at ply -1, -2, -3, -4, and -6 with current move. void update_continuation_histories(Stack* ss, Piece pc, Square to, int bonus) { + bonus = bonus * 45 / 64; + for (int i : {1, 2, 3, 4, 6}) { // Only update the first 2 continuation histories if we are in check @@ -1876,7 +1881,7 @@ void update_quiet_histories( update_continuation_histories(ss, pos.moved_piece(move), move.to_sq(), bonus); int pIndex = pawn_structure_index(pos); - workerThread.pawnHistory[pIndex][pos.moved_piece(move)][move.to_sq()] << bonus; + workerThread.pawnHistory[pIndex][pos.moved_piece(move)][move.to_sq()] << bonus / 2; } // Updates move sorting heuristics diff --git a/src/search.h b/src/search.h index 8ef1ac4b9..509771412 100644 --- a/src/search.h +++ b/src/search.h @@ -220,6 +220,7 @@ class SearchManager: public ISearchManager { Depth depth) const; Stockfish::TimeManagement tm; + int originalPly; int callsCnt; std::atomic_bool ponder; diff --git a/src/thread.cpp b/src/thread.cpp index 18ce50b9a..9ae40165f 100644 --- a/src/thread.cpp +++ b/src/thread.cpp @@ -172,6 +172,7 @@ void ThreadPool::clear() { main_manager()->callsCnt = 0; main_manager()->bestPreviousScore = VALUE_INFINITE; main_manager()->bestPreviousAverageScore = VALUE_INFINITE; + main_manager()->originalPly = -1; main_manager()->previousTimeReduction = 1.0; main_manager()->tm.clear(); } diff --git a/src/timeman.cpp b/src/timeman.cpp index 4feb329b3..f389e082d 100644 --- a/src/timeman.cpp +++ b/src/timeman.cpp @@ -44,10 +44,8 @@ void TimeManagement::advance_nodes_time(std::int64_t nodes) { // the bounds of time allowed for the current game ply. We currently support: // 1) x basetime (+ z increment) // 2) x moves in y seconds (+ z increment) -void TimeManagement::init(Search::LimitsType& limits, - Color us, - int ply, - const OptionsMap& options) { +void TimeManagement::init( + Search::LimitsType& limits, Color us, int ply, const OptionsMap& options, int& originalPly) { TimePoint npmsec = TimePoint(options["nodestime"]); // If we have no time, we don't need to fully initialize TM. @@ -58,6 +56,9 @@ void TimeManagement::init(Search::LimitsType& limits, if (limits.time[us] == 0) return; + if (originalPly == -1) + originalPly = ply; + TimePoint moveOverhead = TimePoint(options["Move Overhead"]); // optScale is a percentage of available time to use for the current move. @@ -106,6 +107,8 @@ void TimeManagement::init(Search::LimitsType& limits, { // Use extra time with larger increments double optExtra = scaledInc < 500 ? 1.0 : 1.13; + if (ply - originalPly < 2) + optExtra *= 0.95; // Calculate time constants based on current time left. double logTimeInSec = std::log10(scaledTime / 1000.0); diff --git a/src/timeman.h b/src/timeman.h index d2eb984f7..5e0b8369c 100644 --- a/src/timeman.h +++ b/src/timeman.h @@ -36,7 +36,8 @@ struct LimitsType; // the maximum available time, the game move number, and other parameters. class TimeManagement { public: - void init(Search::LimitsType& limits, Color us, int ply, const OptionsMap& options); + void init( + Search::LimitsType& limits, Color us, int ply, const OptionsMap& options, int& originalPly); TimePoint optimum() const; TimePoint maximum() const; diff --git a/src/tt.cpp b/src/tt.cpp index 4885a781a..cb46fc8a9 100644 --- a/src/tt.cpp +++ b/src/tt.cpp @@ -35,7 +35,7 @@ namespace Stockfish { void TTEntry::save( Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev, uint8_t generation8) { - // Preserve any existing move for the same position + // Preserve the old ttmove if we don't have a new one if (m || uint16_t(k) != key16) move16 = m;