namespace { extern "C" {
// start_routine() is the C function which is called when a new thread
- // is launched. It simply calls idle_loop() with the supplied threadID.
+ // is launched. It simply calls idle_loop() of the supplied threadID.
// There are two versions of this function; one for POSIX threads and
// one for Windows threads.
DWORD WINAPI start_routine(LPVOID threadID) {
- Threads.idle_loop(*(int*)threadID, NULL);
+ Threads[*(int*)threadID].idle_loop(NULL);
return 0;
}
void* start_routine(void* threadID) {
- Threads.idle_loop(*(int*)threadID, NULL);
+ Threads[*(int*)threadID].idle_loop(NULL);
return NULL;
}
maxThreadsPerSplitPoint = Options["Maximum Number of Threads per Split Point"].value<int>();
minimumSplitDepth = Options["Minimum Split Depth"].value<int>() * ONE_PLY;
useSleepingThreads = Options["Use Sleeping Threads"].value<bool>();
- activeThreads = Options["Threads"].value<int>();
+
+ set_size(Options["Threads"].value<int>());
+}
+
+
+// set_size() changes the number of active threads and raises do_sleep flag for
+// all the unused threads that will go immediately to sleep.
+
+void ThreadsManager::set_size(int cnt) {
+
+ assert(cnt > 0 && cnt <= MAX_THREADS);
+
+ activeThreads = cnt;
+
+ for (int i = 0; i < MAX_THREADS; i++)
+ if (i < activeThreads)
+ {
+ // Dynamically allocate pawn and material hash tables according to the
+ // number of active threads. This avoids preallocating memory for all
+ // possible threads if only few are used as, for instance, on mobile
+ // devices where memory is scarce and allocating for MAX_THREADS could
+ // even result in a crash.
+ threads[i].pawnTable.init();
+ threads[i].materialTable.init();
+
+ threads[i].do_sleep = false;
+ }
+ else
+ threads[i].do_sleep = true;
}
void ThreadsManager::init() {
- int threadID[MAX_THREADS];
-
- // This flag is needed to properly end the threads when program exits
- allThreadsShouldExit = false;
-
- // Threads will sent to sleep as soon as created, only main thread is kept alive
- activeThreads = 1;
+ // Threads will go to sleep as soon as created, only main thread is kept alive
+ set_size(1);
threads[0].state = Thread::SEARCHING;
+ threads[0].threadID = 0;
- // Allocate pawn and material hash tables for main thread
- init_hash_tables();
-
- lock_init(&mpLock);
+ // Initialize threads lock, used when allocating slaves during splitting
+ lock_init(&threadsLock);
- // Initialize thread and split point locks
+ // Initialize sleep and split point locks
for (int i = 0; i < MAX_THREADS; i++)
{
lock_init(&threads[i].sleepLock);
// Create and startup all the threads but the main that is already running
for (int i = 1; i < MAX_THREADS; i++)
{
- threads[i].state = Thread::INITIALIZING;
- threadID[i] = i;
+ threads[i].state = Thread::AVAILABLE;
+ threads[i].threadID = i;
#if defined(_MSC_VER)
- bool ok = (CreateThread(NULL, 0, start_routine, (LPVOID)&threadID[i], 0, NULL) != NULL);
+ threads[i].handle = CreateThread(NULL, 0, start_routine, (LPVOID)&threads[i].threadID, 0, NULL);
+ bool ok = (threads[i].handle != NULL);
#else
- pthread_t pthreadID;
- bool ok = (pthread_create(&pthreadID, NULL, start_routine, (void*)&threadID[i]) == 0);
- pthread_detach(pthreadID);
+ bool ok = (pthread_create(&threads[i].handle, NULL, start_routine, (void*)&threads[i].threadID) == 0);
#endif
+
if (!ok)
{
std::cout << "Failed to create thread number " << i << std::endl;
::exit(EXIT_FAILURE);
}
-
- // Wait until the thread has finished launching and is gone to sleep
- while (threads[i].state == Thread::INITIALIZING) {}
}
}
-// exit() is called to cleanly exit the threads when the program finishes
+// exit() is called to cleanly terminate the threads when the program finishes
void ThreadsManager::exit() {
- // Force the woken up threads to exit idle_loop() and hence terminate
- allThreadsShouldExit = true;
-
for (int i = 0; i < MAX_THREADS; i++)
{
- // Wake up all the threads and waits for termination
+ // Wake up all the slave threads and wait for termination
if (i != 0)
{
+ threads[i].do_terminate = true;
threads[i].wake_up();
- while (threads[i].state != Thread::TERMINATED) {}
+
+#if defined(_MSC_VER)
+ WaitForSingleObject(threads[i].handle, 0);
+ CloseHandle(threads[i].handle);
+#else
+ pthread_join(threads[i].handle, NULL);
+ pthread_detach(threads[i].handle);
+#endif
}
- // Now we can safely destroy the locks and wait conditions
+ // Now we can safely destroy locks and wait conditions
lock_destroy(&threads[i].sleepLock);
cond_destroy(&threads[i].sleepCond);
lock_destroy(&(threads[i].splitPoints[j].lock));
}
- lock_destroy(&mpLock);
-}
-
-
-// init_hash_tables() dynamically allocates pawn and material hash tables
-// according to the number of active threads. This avoids preallocating
-// memory for all possible threads if only few are used as, for instance,
-// on mobile devices where memory is scarce and allocating for MAX_THREADS
-// threads could even result in a crash.
-
-void ThreadsManager::init_hash_tables() {
-
- for (int i = 0; i < activeThreads; i++)
- {
- threads[i].pawnTable.init();
- threads[i].materialTable.init();
- }
+ lock_destroy(&threadsLock);
}
// call search().When all threads have returned from search() then split() returns.
template <bool Fake>
-void ThreadsManager::split(Position& pos, SearchStack* ss, Value* alpha, const Value beta,
- Value* bestValue, Depth depth, Move threatMove,
- int moveCount, MovePicker* mp, int nodeType) {
+Value ThreadsManager::split(Position& pos, SearchStack* ss, Value alpha, Value beta,
+ Value bestValue, Depth depth, Move threatMove,
+ int moveCount, MovePicker* mp, int nodeType) {
assert(pos.is_ok());
- assert(*bestValue >= -VALUE_INFINITE);
- assert(*bestValue <= *alpha);
- assert(*alpha < beta);
+ assert(bestValue >= -VALUE_INFINITE);
+ assert(bestValue <= alpha);
+ assert(alpha < beta);
assert(beta <= VALUE_INFINITE);
assert(depth > DEPTH_ZERO);
assert(pos.thread() >= 0 && pos.thread() < activeThreads);
int i, master = pos.thread();
Thread& masterThread = threads[master];
- lock_grab(&mpLock);
-
- // If no other thread is available to help us, or if we have too many
- // active split points, don't split.
- if ( !available_slave_exists(master)
- || masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS)
- {
- lock_release(&mpLock);
- return;
- }
+ // If we already have too many active split points, don't split
+ if (masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS)
+ return bestValue;
// Pick the next available split point object from the split point stack
- SplitPoint& splitPoint = masterThread.splitPoints[masterThread.activeSplitPoints++];
+ SplitPoint* sp = masterThread.splitPoints + masterThread.activeSplitPoints;
// Initialize the split point object
- splitPoint.parent = masterThread.splitPoint;
- splitPoint.master = master;
- splitPoint.is_betaCutoff = false;
- splitPoint.depth = depth;
- splitPoint.threatMove = threatMove;
- splitPoint.alpha = *alpha;
- splitPoint.beta = beta;
- splitPoint.nodeType = nodeType;
- splitPoint.bestValue = *bestValue;
- splitPoint.mp = mp;
- splitPoint.moveCount = moveCount;
- splitPoint.pos = &pos;
- splitPoint.nodes = 0;
- splitPoint.ss = ss;
+ sp->parent = masterThread.splitPoint;
+ sp->master = master;
+ sp->is_betaCutoff = false;
+ sp->depth = depth;
+ sp->threatMove = threatMove;
+ sp->alpha = alpha;
+ sp->beta = beta;
+ sp->nodeType = nodeType;
+ sp->bestValue = bestValue;
+ sp->mp = mp;
+ sp->moveCount = moveCount;
+ sp->pos = &pos;
+ sp->nodes = 0;
+ sp->ss = ss;
for (i = 0; i < activeThreads; i++)
- splitPoint.is_slave[i] = false;
-
- masterThread.splitPoint = &splitPoint;
+ sp->is_slave[i] = false;
// If we are here it means we are not available
- assert(masterThread.state != Thread::AVAILABLE);
+ assert(masterThread.state == Thread::SEARCHING);
int workersCnt = 1; // At least the master is included
- // Allocate available threads setting state to THREAD_BOOKED
+ // Try to allocate available threads and ask them to start searching setting
+ // the state to Thread::WORKISWAITING, this must be done under lock protection
+ // to avoid concurrent allocation of the same slave by another master.
+ lock_grab(&threadsLock);
+
for (i = 0; !Fake && i < activeThreads && workersCnt < maxThreadsPerSplitPoint; i++)
if (i != master && threads[i].is_available_to(master))
{
- threads[i].state = Thread::BOOKED;
- threads[i].splitPoint = &splitPoint;
- splitPoint.is_slave[i] = true;
workersCnt++;
- }
+ sp->is_slave[i] = true;
+ threads[i].splitPoint = sp;
- assert(Fake || workersCnt > 1);
+ // This makes the slave to exit from idle_loop()
+ threads[i].state = Thread::WORKISWAITING;
- // We can release the lock because slave threads are already booked and master is not available
- lock_release(&mpLock);
+ if (useSleepingThreads)
+ threads[i].wake_up();
+ }
- // Tell the threads that they have work to do. This will make them leave
- // their idle loop.
- for (i = 0; i < activeThreads; i++)
- if (i == master || splitPoint.is_slave[i])
- {
- assert(i == master || threads[i].state == Thread::BOOKED);
+ lock_release(&threadsLock);
- threads[i].state = Thread::WORKISWAITING; // This makes the slave to exit from idle_loop()
+ // We failed to allocate even one slave, return
+ if (!Fake && workersCnt == 1)
+ return bestValue;
- if (useSleepingThreads && i != master)
- threads[i].wake_up();
- }
+ masterThread.splitPoint = sp;
+ masterThread.activeSplitPoints++;
+ masterThread.state = Thread::WORKISWAITING;
// Everything is set up. The master thread enters the idle loop, from
// 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
+ // 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.
- idle_loop(master, &splitPoint);
+ masterThread.idle_loop(sp);
+
+ // In helpful master concept a master can help only a sub-tree, and
+ // because here is all finished is not possible master is booked.
+ assert(masterThread.state == Thread::AVAILABLE);
// We have returned from the idle loop, which means that all threads are
- // finished. Update alpha and bestValue, and return.
- lock_grab(&mpLock);
+ // finished. Note that changing state and decreasing activeSplitPoints is done
+ // under lock protection to avoid a race with Thread::is_available_to().
+ lock_grab(&threadsLock);
- *alpha = splitPoint.alpha;
- *bestValue = splitPoint.bestValue;
+ masterThread.state = Thread::SEARCHING;
masterThread.activeSplitPoints--;
- masterThread.splitPoint = splitPoint.parent;
- pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes);
- lock_release(&mpLock);
+ lock_release(&threadsLock);
+
+ masterThread.splitPoint = sp->parent;
+ pos.set_nodes_searched(pos.nodes_searched() + sp->nodes);
+
+ return sp->bestValue;
}
// Explicit template instantiations
-template void ThreadsManager::split<false>(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, int);
-template void ThreadsManager::split<true>(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, int);
+template Value ThreadsManager::split<false>(Position&, SearchStack*, Value, Value, Value, Depth, Move, int, MovePicker*, int);
+template Value ThreadsManager::split<true>(Position&, SearchStack*, Value, Value, Value, Depth, Move, int, MovePicker*, int);