/// Types
enum NodeType { NonPV, PV };
+ // Set to true to force running with one thread.
+ // Used for debugging SMP code.
+ const bool FakeSplit = false;
+
// ThreadsManager class is used to handle all the threads related stuff in search,
// init, starting, parking and, the most important, launching a slave thread at a
// split point are what this class does. All the access to shared thread data is
void wake_sleeping_threads();
void put_threads_to_sleep();
void idle_loop(int threadID, SplitPoint* sp);
- bool split(const Position& pos, SearchStack* ss, int ply, Value* alpha, const Value beta, Value* bestValue,
- Depth depth, bool mateThreat, int* moves, MovePicker* mp, int master, bool pvNode);
+
+ template <bool Fake>
+ void split(const Position& pos, SearchStack* ss, int ply, Value* alpha, const Value beta, Value* bestValue,
+ Depth depth, bool mateThreat, int* moveCount, MovePicker* mp, int master, bool pvNode);
private:
friend void poll();
Value search(Position& pos, SearchStack ss[], Value alpha, Value beta, Depth depth, int ply, bool allowNullmove, int threadID, Move excludedMove = MOVE_NONE);
template <NodeType PvNode>
- Depth extension(const Position& pos, Move m, bool captureOrPromotion, bool moveIsCheck, bool singleEvasion, bool mateThreat, bool* dangerous);
-
Value qsearch(Position& pos, SearchStack ss[], Value alpha, Value beta, Depth depth, int ply, int threadID);
+
+ template <NodeType PvNode>
void sp_search(SplitPoint* sp, int threadID);
- void sp_search_pv(SplitPoint* sp, int threadID);
+
+ template <NodeType PvNode>
+ Depth extension(const Position& pos, Move m, bool captureOrPromotion, bool moveIsCheck, bool singleEvasion, bool mateThreat, bool* dangerous);
+
void init_node(SearchStack ss[], int ply, int threadID);
void update_pv(SearchStack ss[], int ply);
void sp_update_pv(SearchStack* pss, SearchStack ss[], int ply);
beta = *betaPtr;
isCheck = pos.is_check();
- // Step 1. Initialize node and poll (omitted at root, but I can see no good reason for this, FIXME)
- // Step 2. Check for aborted search (omitted at root, because we do not initialize root node)
+ // Step 1. Initialize node and poll (omitted at root, init_ss_array() has already initialized root node)
+ // Step 2. Check for aborted search (omitted at root)
// Step 3. Mate distance pruning (omitted at root)
// Step 4. Transposition table lookup (omitted at root)
// At root we do this only to get reference value for child nodes
if (!isCheck)
ss[0].eval = evaluate(pos, ei, 0);
- else
- ss[0].eval = VALUE_NONE; // HACK because we do not initialize root node
// Step 6. Razoring (omitted at root)
// Step 7. Static null move pruning (omitted at root)
assert(alpha >= -VALUE_INFINITE && alpha <= VALUE_INFINITE);
assert(beta > alpha && beta <= VALUE_INFINITE);
+ assert(PvNode || alpha == beta - 1);
assert(ply >= 0 && ply < PLY_MAX);
assert(threadID >= 0 && threadID < TM.active_threads());
oldAlpha = alpha;
if (depth < OnePly)
- return qsearch(pos, ss, alpha, beta, Depth(0), ply, threadID);
+ return qsearch<PvNode>(pos, ss, alpha, beta, Depth(0), ply, threadID);
// Step 1. Initialize node and poll
// Polling can abort search.
&& !pos.has_pawn_on_7th(pos.side_to_move()))
{
Value rbeta = beta - razor_margin(depth);
- Value v = qsearch(pos, ss, rbeta-1, rbeta, Depth(0), ply, threadID);
+ Value v = qsearch<NonPV>(pos, ss, rbeta-1, rbeta, Depth(0), ply, threadID);
if (v < rbeta)
// Logically we should return (v + razor_margin(depth)), but
// surprisingly this did slightly weaker in tests.
continue;
// Value based pruning
- Depth predictedDepth = newDepth - reduction<NonPV>(depth, moveCount); // FIXME We illogically ignore reduction condition depth >= 3*OnePly
+ // We illogically ignore reduction condition depth >= 3*OnePly for predicted depth,
+ // but fixing this made program slightly weaker.
+ Depth predictedDepth = newDepth - reduction<NonPV>(depth, moveCount);
futilityValueScaled = ss[ply].eval + futility_margin(predictedDepth, moveCount)
+ H.gain(pos.piece_on(move_from(move)), move_to(move));
bestValue = value;
if (value > alpha)
{
- alpha = value;
+ if (PvNode && value < beta) // This guarantees that always: alpha < beta
+ alpha = value;
+
update_pv(ss, ply);
+
if (value == value_mate_in(ply + 1))
ss[ply].mateKiller = move;
}
&& Iteration <= 99
&& TM.available_thread_exists(threadID)
&& !AbortSearch
- && !TM.thread_should_stop(threadID)
- && TM.split(pos, ss, ply, &alpha, beta, &bestValue,
- depth, mateThreat, &moveCount, &mp, threadID, PvNode))
- break;
+ && !TM.thread_should_stop(threadID))
+ TM.split<FakeSplit>(pos, ss, ply, &alpha, beta, &bestValue, depth,
+ mateThreat, &moveCount, &mp, threadID, PvNode);
}
// Step 19. Check for mate and stalemate
// search function when the remaining depth is zero (or, to be more precise,
// less than OnePly).
+ template <NodeType PvNode>
Value qsearch(Position& pos, SearchStack ss[], Value alpha, Value beta,
Depth depth, int ply, int threadID) {
assert(alpha >= -VALUE_INFINITE && alpha <= VALUE_INFINITE);
assert(beta >= -VALUE_INFINITE && beta <= VALUE_INFINITE);
+ assert(PvNode || alpha == beta - 1);
assert(depth <= 0);
assert(ply >= 0 && ply < PLY_MAX);
assert(threadID >= 0 && threadID < TM.active_threads());
bool isCheck, enoughMaterial, moveIsCheck, evasionPrunable;
const TTEntry* tte = NULL;
int moveCount = 0;
- bool pvNode = (beta - alpha != 1);
Value oldAlpha = alpha;
// Initialize, and make an early exit in case of an aborted search,
tte = TT.retrieve(pos.get_key());
ttMove = (tte ? tte->move() : MOVE_NONE);
- if (!pvNode && tte && ok_to_use_TT(tte, depth, beta, ply))
+ if (!PvNode && tte && ok_to_use_TT(tte, depth, beta, ply))
{
assert(tte->type() != VALUE_TYPE_EVAL);
ss[ply].currentMove = move;
// Futility pruning
- if ( enoughMaterial
+ if ( !PvNode
+ && enoughMaterial
&& !isCheck
- && !pvNode
&& !moveIsCheck
&& move != ttMove
&& !move_is_promotion(move)
&& !pos.can_castle(pos.side_to_move());
// Don't search moves with negative SEE values
- if ( (!isCheck || evasionPrunable)
- && !pvNode
+ if ( !PvNode
+ && (!isCheck || evasionPrunable)
&& move != ttMove
&& !move_is_promotion(move)
&& pos.see_sign(move) < 0)
// Make and search the move
pos.do_move(move, st, ci, moveIsCheck);
- value = -qsearch(pos, ss, -beta, -alpha, depth-OnePly, ply+1, threadID);
+ value = -qsearch<PvNode>(pos, ss, -beta, -alpha, depth-OnePly, ply+1, threadID);
pos.undo_move(move);
assert(value > -VALUE_INFINITE && value < VALUE_INFINITE);
// also don't need to store anything to the hash table here: This is taken
// care of after we return from the split point.
+ template <NodeType PvNode>
void sp_search(SplitPoint* sp, int threadID) {
assert(threadID >= 0 && threadID < TM.active_threads());
StateInfo st;
Move move;
Depth ext, newDepth;
- Value value, futilityValueScaled;
+ Value value;
+ Value futilityValueScaled; // NonPV specific
bool isCheck, moveIsCheck, captureOrPromotion, dangerous;
int moveCount;
value = -VALUE_INFINITE;
lock_grab(&(sp->lock));
while ( sp->bestValue < sp->beta
- && !TM.thread_should_stop(threadID)
- && (move = sp->mp->get_next_move()) != MOVE_NONE)
+ && (move = sp->mp->get_next_move()) != MOVE_NONE
+ && !TM.thread_should_stop(threadID))
{
- moveCount = ++sp->moves;
+ moveCount = ++sp->moveCount;
lock_release(&(sp->lock));
assert(move_is_ok(move));
captureOrPromotion = pos.move_is_capture_or_promotion(move);
// Step 11. Decide the new search depth
- ext = extension<NonPV>(pos, move, captureOrPromotion, moveIsCheck, false, sp->mateThreat, &dangerous);
+ ext = extension<PvNode>(pos, move, captureOrPromotion, moveIsCheck, false, sp->mateThreat, &dangerous);
newDepth = sp->depth - OnePly + ext;
// Update current move
ss[sp->ply].currentMove = move;
- // Step 12. Futility pruning
- if ( !isCheck
+ // Step 12. Futility pruning (is omitted in PV nodes)
+ if ( !PvNode
+ && !isCheck
&& !dangerous
&& !captureOrPromotion
&& !move_is_castle(move))
&& !move_is_castle(move)
&& !move_is_killer(move, ss[sp->ply]))
{
- ss[sp->ply].reduction = reduction<NonPV>(sp->depth, moveCount);
- if (ss[sp->ply].reduction)
- {
- value = -search<NonPV>(pos, ss, -(sp->alpha+1), -(sp->alpha), newDepth-ss[sp->ply].reduction, sp->ply+1, true, threadID);
- doFullDepthSearch = (value >= sp->beta && !TM.thread_should_stop(threadID));
- }
- }
-
- // Step 15. Full depth search
- if (doFullDepthSearch)
- {
- ss[sp->ply].reduction = Depth(0);
- value = -search<NonPV>(pos, ss, -(sp->alpha+1), -(sp->alpha), newDepth, sp->ply+1, true, threadID);
- }
-
- // Step 16. Undo move
- pos.undo_move(move);
-
- assert(value > -VALUE_INFINITE && value < VALUE_INFINITE);
-
- // Step 17. Check for new best move
- lock_grab(&(sp->lock));
-
- if (value > sp->bestValue && !TM.thread_should_stop(threadID))
- {
- sp->bestValue = value;
- if (sp->bestValue >= sp->beta)
- {
- sp->stopRequest = true;
- sp_update_pv(sp->parentSstack, ss, sp->ply);
- }
- }
- }
-
- /* Here we have the lock still grabbed */
-
- sp->slaves[threadID] = 0;
- sp->cpus--;
-
- lock_release(&(sp->lock));
- }
-
-
- // sp_search_pv() is used to search from a PV split point. This function
- // is called by each thread working at the split point. It is similar to
- // the normal search_pv() function, but simpler. Because we have already
- // probed the hash table and searched the first move before splitting, we
- // don't have to repeat all this work in sp_search_pv(). We also don't
- // need to store anything to the hash table here: This is taken care of
- // after we return from the split point.
-
- void sp_search_pv(SplitPoint* sp, int threadID) {
-
- assert(threadID >= 0 && threadID < TM.active_threads());
- assert(TM.active_threads() > 1);
-
- StateInfo st;
- Move move;
- Depth ext, newDepth;
- Value value;
- bool moveIsCheck, captureOrPromotion, dangerous;
- int moveCount;
- value = -VALUE_INFINITE;
-
- Position pos(*sp->pos);
- CheckInfo ci(pos);
- SearchStack* ss = sp->sstack[threadID];
-
- // Step 10. Loop through moves
- // Loop through all legal moves until no moves remain or a beta cutoff occurs
- lock_grab(&(sp->lock));
-
- while ( sp->alpha < sp->beta
- && !TM.thread_should_stop(threadID)
- && (move = sp->mp->get_next_move()) != MOVE_NONE)
- {
- moveCount = ++sp->moves;
- lock_release(&(sp->lock));
-
- assert(move_is_ok(move));
-
- moveIsCheck = pos.move_is_check(move, ci);
- captureOrPromotion = pos.move_is_capture_or_promotion(move);
-
- // Step 11. Decide the new search depth
- ext = extension<PV>(pos, move, captureOrPromotion, moveIsCheck, false, sp->mateThreat, &dangerous);
- newDepth = sp->depth - OnePly + ext;
-
- // Update current move
- ss[sp->ply].currentMove = move;
-
- // Step 12. Futility pruning (is omitted in PV nodes)
-
- // Step 13. Make the move
- pos.do_move(move, st, ci, moveIsCheck);
-
- // Step 14. Reduced search
- // if the move fails high will be re-searched at full depth.
- bool doFullDepthSearch = true;
-
- if ( !dangerous
- && !captureOrPromotion
- && !move_is_castle(move)
- && !move_is_killer(move, ss[sp->ply]))
- {
- ss[sp->ply].reduction = reduction<PV>(sp->depth, moveCount);
+ ss[sp->ply].reduction = reduction<PvNode>(sp->depth, moveCount);
if (ss[sp->ply].reduction)
{
Value localAlpha = sp->alpha;
value = -search<NonPV>(pos, ss, -(localAlpha+1), -localAlpha, newDepth-ss[sp->ply].reduction, sp->ply+1, true, threadID);
- doFullDepthSearch = (value > localAlpha && !TM.thread_should_stop(threadID));
+ doFullDepthSearch = (value > localAlpha);
}
}
// Step 15. Full depth search
if (doFullDepthSearch)
{
- Value localAlpha = sp->alpha;
ss[sp->ply].reduction = Depth(0);
+ Value localAlpha = sp->alpha;
value = -search<NonPV>(pos, ss, -(localAlpha+1), -localAlpha, newDepth, sp->ply+1, true, threadID);
- if (value > localAlpha && value < sp->beta && !TM.thread_should_stop(threadID))
- {
- // If another thread has failed high then sp->alpha has been increased
- // to be higher or equal then beta, if so, avoid to start a PV search.
- localAlpha = sp->alpha;
- if (localAlpha < sp->beta)
- value = -search<PV>(pos, ss, -sp->beta, -localAlpha, newDepth, sp->ply+1, false, threadID);
- }
+ if (PvNode && value > localAlpha && value < sp->beta)
+ value = -search<PV>(pos, ss, -sp->beta, -sp->alpha, newDepth, sp->ply+1, false, threadID);
}
// Step 16. Undo move
if (value > sp->bestValue && !TM.thread_should_stop(threadID))
{
sp->bestValue = value;
- if (value > sp->alpha)
+
+ if (sp->bestValue > sp->alpha)
{
- // Ask threads to stop before to modify sp->alpha
- if (value >= sp->beta)
+ if (!PvNode || value >= sp->beta)
sp->stopRequest = true;
- sp->alpha = value;
+ if (PvNode && value < sp->beta) // This guarantees that always: sp->alpha < sp->beta
+ sp->alpha = value;
sp_update_pv(sp->parentSstack, ss, sp->ply);
- if (value == value_mate_in(sp->ply + 1))
- ss[sp->ply].mateKiller = move;
}
}
}
/* Here we have the lock still grabbed */
sp->slaves[threadID] = 0;
- sp->cpus--;
lock_release(&(sp->lock));
}
-
// init_node() is called at the beginning of all the search functions
// (search() qsearch(), and so on) and initializes the
// search stack object corresponding to the current node. Once every
ss[ply + 2].initKillers();
}
-
// update_pv() is called whenever a search returns a value > alpha.
// It updates the PV in the SearchStack object corresponding to the
// current node.
threads[threadID].state = THREAD_SEARCHING;
if (threads[threadID].splitPoint->pvNode)
- sp_search_pv(threads[threadID].splitPoint, threadID);
+ sp_search<PV>(threads[threadID].splitPoint, threadID);
else
- sp_search(threads[threadID].splitPoint, threadID);
+ sp_search<NonPV>(threads[threadID].splitPoint, threadID);
assert(threads[threadID].state == THREAD_SEARCHING);
threads[threadID].state = THREAD_AVAILABLE;
}
- // If this thread is the master of a split point and all threads have
+ // If this thread is the master of a split point and all slaves have
// finished their work at this split point, return from the idle loop.
- if (sp && sp->cpus == 0)
+ int i = 0;
+ for ( ; sp && i < ActiveThreads && !sp->slaves[i]; i++) {}
+
+ if (i == ActiveThreads)
{
- // Because sp->cpus is decremented under lock protection,
- // be sure sp->lock has been released before to proceed.
+ // Because sp->slaves[] is reset under lock protection,
+ // be sure sp->lock has been released before to return.
lock_grab(&(sp->lock));
lock_release(&(sp->lock));
// split() does the actual work of distributing the work at a node between
- // several threads at PV nodes. If it does not succeed in splitting the
+ // several available threads. If it does not succeed in splitting the
// node (because no idle threads are available, or because we have no unused
- // split point objects), the function immediately returns false. If
- // splitting is possible, a SplitPoint object is initialized with all the
- // data that must be copied to the helper threads (the current position and
- // search stack, alpha, beta, the search depth, etc.), and we tell our
- // helper threads that they have been assigned work. This will cause them
- // to instantly leave their idle loops and call sp_search_pv(). When all
- // threads have returned from sp_search_pv (or, equivalently, when
- // splitPoint->cpus becomes 0), split() returns true.
-
- bool ThreadsManager::split(const Position& p, SearchStack* sstck, int ply,
- Value* alpha, const Value beta, Value* bestValue,
- Depth depth, bool mateThreat, int* moves, MovePicker* mp, int master, bool pvNode) {
-
+ // split point objects), the function immediately returns. If splitting is
+ // possible, a SplitPoint object is initialized with all the data that must be
+ // copied to the helper threads and we tell our helper threads that they have
+ // been assigned work. This will cause them to instantly leave their idle loops
+ // and call sp_search(). When all threads have returned from sp_search() then
+ // split() returns.
+
+ template <bool Fake>
+ void ThreadsManager::split(const Position& p, SearchStack* sstck, int ply, Value* alpha,
+ const Value beta, Value* bestValue, Depth depth, bool mateThreat,
+ int* moveCount, MovePicker* mp, int master, bool pvNode) {
assert(p.is_ok());
assert(sstck != NULL);
assert(ply >= 0 && ply < PLY_MAX);
assert(*bestValue >= -VALUE_INFINITE);
- assert( ( pvNode && *bestValue <= *alpha)
- || (!pvNode && *bestValue < beta ));
- assert(!pvNode || *alpha < beta);
+ assert(*bestValue <= *alpha);
+ assert(*alpha < beta);
assert(beta <= VALUE_INFINITE);
assert(depth > Depth(0));
assert(master >= 0 && master < ActiveThreads);
assert(ActiveThreads > 1);
- SplitPoint* splitPoint;
-
lock_grab(&MPLock);
// If no other thread is available to help us, or if we have too many
|| threads[master].activeSplitPoints >= ACTIVE_SPLIT_POINTS_MAX)
{
lock_release(&MPLock);
- return false;
+ return;
}
// Pick the next available split point object from the split point stack
- splitPoint = &SplitPointStack[master][threads[master].activeSplitPoints];
+ SplitPoint* splitPoint = &SplitPointStack[master][threads[master].activeSplitPoints];
// Initialize the split point object
splitPoint->parent = threads[master].splitPoint;
splitPoint->beta = beta;
splitPoint->pvNode = pvNode;
splitPoint->bestValue = *bestValue;
- splitPoint->master = master;
splitPoint->mp = mp;
- splitPoint->moves = *moves;
- splitPoint->cpus = 1;
+ splitPoint->moveCount = *moveCount;
splitPoint->pos = &p;
splitPoint->parentSstack = sstck;
for (int i = 0; i < ActiveThreads; i++)
// If we are here it means we are not available
assert(threads[master].state != THREAD_AVAILABLE);
+ int workersCnt = 1; // At least the master is included
+
// Allocate available threads setting state to THREAD_BOOKED
- for (int i = 0; i < ActiveThreads && splitPoint->cpus < MaxThreadsPerSplitPoint; i++)
+ for (int i = 0; !Fake && i < ActiveThreads && workersCnt < MaxThreadsPerSplitPoint; i++)
if (thread_is_available(i, master))
{
threads[i].state = THREAD_BOOKED;
threads[i].splitPoint = splitPoint;
splitPoint->slaves[i] = 1;
- splitPoint->cpus++;
+ workersCnt++;
}
- assert(splitPoint->cpus > 1);
+ assert(Fake || workersCnt > 1);
// We can release the lock because slave threads are already booked and master is not available
lock_release(&MPLock);
// which it will instantly launch a search, because its state is
// THREAD_WORKISWAITING. We send the split point as a second parameter to the
// idle loop, which means that the main thread will return from the idle
- // loop when all threads have finished their work at this split point
- // (i.e. when splitPoint->cpus == 0).
+ // loop when all threads have finished their work at this split point.
idle_loop(master, splitPoint);
// We have returned from the idle loop, which means that all threads are
threads[master].splitPoint = splitPoint->parent;
lock_release(&MPLock);
- return true;
}
init_ss_array(ss);
pos.do_move(cur->move, st);
moves[count].move = cur->move;
- moves[count].score = -qsearch(pos, ss, -VALUE_INFINITE, VALUE_INFINITE, Depth(0), 1, 0);
+ moves[count].score = -qsearch<PV>(pos, ss, -VALUE_INFINITE, VALUE_INFINITE, Depth(0), 1, 0);
moves[count].pv[0] = cur->move;
moves[count].pv[1] = MOVE_NONE;
pos.undo_move(cur->move);
projects / stockfish / blobdiff