kate/BadFish
1
0
Fork 0
mirror of https://github.com/sockspls/badfish synced 2025-04-30 00:33:09 +00:00
Code Issues Releases Wiki Activity

Move pawn and material tables under Thread class

Browse source 
This change allows to remove some quite a bit of code
and seems the natural thing to do.

Introduced file thread.cpp to move away from search.cpp a lot
of threads related stuff.

No functional change.

Signed-off-by: Marco Costalba <mcostalba@gmail.com>
This commit is contained in:
Marco Costalba 2011-04-24 09:20:03 +01:00
parent c9d7e99de6
commit fecefbb99c
13 changed files with 585 additions and 618 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

7
src/Makefile
View file

@ -31,10 +31,9 @@ BINDIR = $(PREFIX)/bin
PGOBENCH = ./$(EXE) bench 32 1 10 default depth
### Object files
OBJS = bitboard.o pawns.o material.o endgame.o evaluate.o main.o \
misc.o move.o movegen.o movepick.o search.o position.o \
tt.o uci.o ucioption.o book.o bitbase.o benchmark.o timeman.o
OBJS = benchmark.o bitbase.o bitboard.o book.o endgame.o evaluate.o main.o \
material.o misc.o move.o movegen.o movepick.o pawns.o position.o \
search.o thread.o timeman.o tt.o uci.o ucioption.o
### ==========================================================================
### Section 2. High-level Configuration

54
src/evaluate.cpp
View file

@ -232,13 +232,6 @@ namespace {
PASSED = 12, UNSTOPPABLE = 13, SPACE = 14, TOTAL = 15
};
// Pawn and material hash tables, indexed by the current thread id.
// We use per-thread tables so that once we get a pointer to an entry
// its life time is unlimited and we don't have to care about someone
// changing the entry under our feet.
MaterialInfoTable* MaterialTable[MAX_THREADS];
PawnInfoTable* PawnTable[MAX_THREADS];
// Function prototypes
template<bool HasPopCnt, bool Trace>
Value do_evaluate(const Position& pos, Value& margin);
@ -271,16 +264,6 @@ namespace {
}
/// prefetchTables() is called in do_move() to prefetch pawn and material
/// hash tables data that will be needed shortly after in evaluation.
void prefetchTables(Key pKey, Key mKey, int threadID) {
PawnTable[threadID]->prefetch(pKey);
MaterialTable[threadID]->prefetch(mKey);
}
/// evaluate() is the main evaluation function. It always computes two
/// values, an endgame score and a middle game score, and interpolates
/// between them based on the remaining material.
@ -320,7 +303,7 @@ Value do_evaluate(const Position& pos, Value& margin) {
margins[WHITE] = margins[BLACK] = VALUE_ZERO;
// Probe the material hash table
MaterialInfo* mi = MaterialTable[pos.thread()]->get_material_info(pos);
MaterialInfo* mi = ThreadsMgr[pos.thread()].materialTable.get_material_info(pos);
bonus += mi->material_value();
// If we have a specialized evaluation function for the current material
@ -332,7 +315,7 @@ Value do_evaluate(const Position& pos, Value& margin) {
}
// Probe the pawn hash table
ei.pi = PawnTable[pos.thread()]->get_pawn_info(pos);
ei.pi = ThreadsMgr[pos.thread()].pawnTable.get_pawn_info(pos);
bonus += apply_weight(ei.pi->pawns_value(), Weights[PawnStructure]);
// Initialize attack and king safety bitboards
@ -433,39 +416,6 @@ Value do_evaluate(const Position& pos, Value& margin) {
} // namespace
/// init_eval() initializes various tables used by the evaluation function
void init_eval(int threads) {
assert(threads <= MAX_THREADS);
for (int i = 0; i < MAX_THREADS; i++)
{
if (i >= threads)
{
delete PawnTable[i];
delete MaterialTable[i];
PawnTable[i] = NULL;
MaterialTable[i] = NULL;
continue;
}
if (!PawnTable[i])
PawnTable[i] = new PawnInfoTable();
if (!MaterialTable[i])
MaterialTable[i] = new MaterialInfoTable();
}
}
/// quit_eval() releases heap-allocated memory at program termination
void quit_eval() {
init_eval(0);
}
/// read_weights() reads evaluation weights from the corresponding UCI parameters
void read_evaluation_uci_options(Color us) {

2
src/evaluate.h
View file

@ -26,8 +26,6 @@ class Position;
extern Value evaluate(const Position& pos, Value& margin);
extern std::string trace_evaluate(const Position& pos);
extern void init_eval(int threads);
extern void quit_eval();
extern void read_evaluation_uci_options(Color sideToMove);
#endif // !defined(EVALUATE_H_INCLUDED)

8
src/main.cpp
View file

@ -28,6 +28,7 @@
#include "evaluate.h"
#include "position.h"
#include "thread.h"
#include "search.h"
#include "ucioption.h"
#ifdef USE_CALLGRIND
@ -52,9 +53,9 @@ int main(int argc, char* argv[]) {
init_bitboards();
Position::init_zobrist();
Position::init_piece_square_tables();
init_eval(1);
init_kpk_bitbase();
init_threads();
init_search();
ThreadsMgr.init_threads();
#ifdef USE_CALLGRIND
CALLGRIND_START_INSTRUMENTATION;
@ -81,7 +82,6 @@ int main(int argc, char* argv[]) {
<< "[limit = 12] [fen positions file = default] "
<< "[depth, time, perft or node limited = depth]" << endl;
exit_threads();
quit_eval();
ThreadsMgr.exit_threads();
return 0;
}

2
src/material.cpp
View file

@ -85,7 +85,7 @@ namespace {
/// MaterialInfoTable c'tor and d'tor allocate and free the space for Endgames
MaterialInfoTable::MaterialInfoTable() { funcs = new Endgames(); }
void MaterialInfoTable::init() { Base::init(); funcs = new Endgames(); }
MaterialInfoTable::~MaterialInfoTable() { delete funcs; }

2
src/material.h
View file

@ -65,8 +65,8 @@ private:
class MaterialInfoTable : public SimpleHash<MaterialInfo, MaterialTableSize> {
public:
MaterialInfoTable();
~MaterialInfoTable();
void init();
MaterialInfo* get_material_info(const Position& pos) const;
static Phase game_phase(const Position& pos);

1
src/misc.h
View file

@ -29,7 +29,6 @@ extern int get_system_time();
extern int cpu_count();
extern int input_available();
extern void prefetch(char* addr);
extern void prefetchTables(Key pKey, Key mKey, int threadID);
extern void dbg_hit_on(bool b);
extern void dbg_hit_on_c(bool c, bool b);

4
src/position.cpp
View file

@ -29,6 +29,7 @@
#include "position.h"
#include "psqtab.h"
#include "rkiss.h"
#include "thread.h"
#include "tt.h"
#include "ucioption.h"
@ -1047,7 +1048,8 @@ void Position::do_move(Move m, StateInfo& newSt, const CheckInfo& ci, bool moveI
}
// Prefetch pawn and material hash tables
prefetchTables(st->pawnKey, st->materialKey, threadID);
ThreadsMgr[threadID].pawnTable.prefetch(st->pawnKey);
ThreadsMgr[threadID].materialTable.prefetch(st->materialKey);
// Update incremental scores
st->value += pst_delta(piece, from, to);

729
src/search.cpp
View file

@ -32,7 +32,6 @@
#include "move.h"
#include "movegen.h"
#include "movepick.h"
#include "lock.h"
#include "search.h"
#include "timeman.h"
#include "thread.h"
@ -54,44 +53,6 @@ namespace {
const bool Slidings[18] = { 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1 };
inline bool piece_is_slider(Piece p) { return Slidings[p]; }
// ThreadsManager class is used to handle all the threads related stuff like init,
// starting, parking and, the most important, launching a slave thread at a split
// point. All the access to shared thread data is done through this class.
class ThreadsManager {
/* As long as the single ThreadsManager object is defined as a global we don't
need to explicitly initialize to zero its data members because variables with
static storage duration are automatically set to zero before enter main()
*/
public:
Thread& operator[](int threadID) { return threads[threadID]; }
void init_threads();
void exit_threads();
int min_split_depth() const { return minimumSplitDepth; }
int active_threads() const { return activeThreads; }
void set_active_threads(int cnt) { activeThreads = cnt; }
void read_uci_options();
bool available_thread_exists(int master) const;
bool thread_is_available(int slave, int master) const;
bool cutoff_at_splitpoint(int threadID) const;
void idle_loop(int threadID, SplitPoint* sp);
template <bool Fake>
void split(Position& pos, SearchStack* ss, Value* alpha, const Value beta, Value* bestValue,
Depth depth, Move threatMove, int moveCount, MovePicker* mp, bool pvNode);
private:
Lock mpLock;
Depth minimumSplitDepth;
int maxThreadsPerSplitPoint;
bool useSleepingThreads;
int activeThreads;
volatile bool allThreadsShouldExit;
Thread threads[MAX_THREADS];
};
// RootMove struct is used for moves at the root of the tree. For each root
// move, we store two scores, a node count, and a PV (really a refutation
// in the case of moves which fail low). Value pv_score is normally set at
@ -203,27 +164,29 @@ namespace {
const Value FutilityMarginQS = Value(0x80);
// Futility lookup tables (initialized at startup) and their access functions
Value FutilityMarginsMatrix[16][64]; // [depth][moveNumber]
int FutilityMoveCountArray[32]; // [depth]
Value FutilityMargins[16][64]; // [depth][moveNumber]
int FutilityMoveCounts[32]; // [depth]
inline Value futility_margin(Depth d, int mn) {
return d < 7 * ONE_PLY ? FutilityMarginsMatrix[Max(d, 1)][Min(mn, 63)]
: 2 * VALUE_INFINITE;
return d < 7 * ONE_PLY ? FutilityMargins[Max(d, 1)][Min(mn, 63)]
: 2 * VALUE_INFINITE;
}
inline int futility_move_count(Depth d) {
return d < 16 * ONE_PLY ? FutilityMoveCountArray[d] : MAX_MOVES;
return d < 16 * ONE_PLY ? FutilityMoveCounts[d] : MAX_MOVES;
}
// Step 14. Reduced search
// Reduction lookup tables (initialized at startup) and their access function
int8_t ReductionMatrix[2][64][64]; // [pv][depth][moveNumber]
int8_t Reductions[2][64][64]; // [pv][depth][moveNumber]
template <NodeType PV>
inline Depth reduction(Depth d, int mn) { return (Depth) ReductionMatrix[PV][Min(d / ONE_PLY, 63)][Min(mn, 63)]; }
template <NodeType PV> inline Depth reduction(Depth d, int mn) {
return (Depth) Reductions[PV][Min(d / ONE_PLY, 63)][Min(mn, 63)];
}
// Easy move margin. An easy move candidate must be at least this much
// better than the second best move.
@ -254,9 +217,6 @@ namespace {
bool SkillLevelEnabled;
RKISS RK;
// Multi-threads manager
ThreadsManager ThreadsMgr;
// Node counters, used only by thread[0] but try to keep in different cache
// lines (64 bytes each) from the heavy multi-thread read accessed variables.
bool SendSearchedNodes;
@ -304,12 +264,6 @@ namespace {
void poll(const Position& pos);
void wait_for_stop_or_ponderhit();
#if !defined(_MSC_VER)
void* init_thread(void* threadID);
#else
DWORD WINAPI init_thread(LPVOID threadID);
#endif
// MovePickerExt is an extended MovePicker class used to choose at compile time
// the proper move source according to the type of node.
@ -378,10 +332,9 @@ namespace {
} // namespace
/// init_threads() is called during startup. It initializes various lookup tables
/// and creates and launches search threads.
/// init_search() is called during startup to initialize various lookup tables
void init_threads() {
void init_search() {
int d; // depth (ONE_PLY == 2)
int hd; // half depth (ONE_PLY == 1)
@ -392,27 +345,20 @@ void init_threads() {
{
double pvRed = log(double(hd)) * log(double(mc)) / 3.0;
double nonPVRed = 0.33 + log(double(hd)) * log(double(mc)) / 2.25;
ReductionMatrix[PV][hd][mc] = (int8_t) ( pvRed >= 1.0 ? floor( pvRed * int(ONE_PLY)) : 0);
ReductionMatrix[NonPV][hd][mc] = (int8_t) (nonPVRed >= 1.0 ? floor(nonPVRed * int(ONE_PLY)) : 0);
Reductions[PV][hd][mc] = (int8_t) ( pvRed >= 1.0 ? floor( pvRed * int(ONE_PLY)) : 0);
Reductions[NonPV][hd][mc] = (int8_t) (nonPVRed >= 1.0 ? floor(nonPVRed * int(ONE_PLY)) : 0);
}
// Init futility margins array
for (d = 1; d < 16; d++) for (mc = 0; mc < 64; mc++)
FutilityMarginsMatrix[d][mc] = Value(112 * int(log(double(d * d) / 2) / log(2.0) + 1.001) - 8 * mc + 45);
FutilityMargins[d][mc] = Value(112 * int(log(double(d * d) / 2) / log(2.0) + 1.001) - 8 * mc + 45);
// Init futility move count array
for (d = 0; d < 32; d++)
FutilityMoveCountArray[d] = int(3.001 + 0.25 * pow(d, 2.0));
// Create and startup threads
ThreadsMgr.init_threads();
FutilityMoveCounts[d] = int(3.001 + 0.25 * pow(d, 2.0));
}
/// exit_threads() is a trampoline to access ThreadsMgr from outside of current file
void exit_threads() { ThreadsMgr.exit_threads(); }
/// perft() is our utility to verify move generation. All the legal moves up to
/// given depth are generated and counted and the sum returned.
@ -489,6 +435,7 @@ bool think(Position& pos, const SearchLimits& limits, Move searchMoves[]) {
UCIMultiPV = Options["MultiPV"].value<int>();
SkillLevel = Options["Skill level"].value<int>();
ThreadsMgr.read_uci_options();
read_evaluation_uci_options(pos.side_to_move());
if (Options["Clear Hash"].value<bool>())
@ -503,10 +450,6 @@ bool think(Position& pos, const SearchLimits& limits, Move searchMoves[]) {
SkillLevelEnabled = (SkillLevel < 20);
MultiPV = (SkillLevelEnabled ? Max(UCIMultiPV, 4) : UCIMultiPV);
// Set the number of active threads
ThreadsMgr.read_uci_options();
init_eval(ThreadsMgr.active_threads());
// Wake up needed threads and reset maxPly counter
for (int i = 0; i < ThreadsMgr.active_threads(); i++)
{
@ -1972,417 +1915,48 @@ split_point_start: // At split points actual search starts from here
}
// init_thread() is the function which is called when a new thread is
// launched. It simply calls the idle_loop() function with the supplied
// threadID. There are two versions of this function; one for POSIX
// threads and one for Windows threads.
// When playing with strength handicap choose best move among the MultiPV set
// using a statistical rule dependent on SkillLevel. Idea by Heinz van Saanen.
void do_skill_level(Move* best, Move* ponder) {
#if !defined(_MSC_VER)
assert(MultiPV > 1);
void* init_thread(void* threadID) {
// Rml list is already sorted by pv_score in descending order
int s;
int max_s = -VALUE_INFINITE;
int size = Min(MultiPV, (int)Rml.size());
int max = Rml[0].pv_score;
int var = Min(max - Rml[size - 1].pv_score, PawnValueMidgame);
int wk = 120 - 2 * SkillLevel;
ThreadsMgr.idle_loop(*(int*)threadID, NULL);
return NULL;
}
// PRNG sequence should be non deterministic
for (int i = abs(get_system_time() % 50); i > 0; i--)
RK.rand<unsigned>();
#else
DWORD WINAPI init_thread(LPVOID threadID) {
ThreadsMgr.idle_loop(*(int*)threadID, NULL);
return 0;
}
#endif
/// The ThreadsManager class
// read_uci_options() updates number of active threads and other internal
// parameters according to the UCI options values. It is called before
// to start a new search.
void ThreadsManager::read_uci_options() {
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>();
}
// idle_loop() is where the threads are parked when they have no work to do.
// The parameter 'sp', if non-NULL, is a pointer to an active SplitPoint
// object for which the current thread is the master.
void ThreadsManager::idle_loop(int threadID, SplitPoint* sp) {
assert(threadID >= 0 && threadID < MAX_THREADS);
int i;
bool allFinished;
while (true)
// Choose best move. For each move's score we add two terms both dependent
// on wk, one deterministic and bigger for weaker moves, and one random,
// then we choose the move with the resulting highest score.
for (int i = 0; i < size; i++)
{
// Slave threads can exit as soon as AllThreadsShouldExit raises,
// master should exit as last one.
if (allThreadsShouldExit)
s = Rml[i].pv_score;
// Don't allow crazy blunders even at very low skills
if (i > 0 && Rml[i-1].pv_score > s + EasyMoveMargin)
break;
// This is our magical formula
s += ((max - s) * wk + var * (RK.rand<unsigned>() % wk)) / 128;
if (s > max_s)
{
assert(!sp);
threads[threadID].state = THREAD_TERMINATED;
return;
}
// If we are not thinking, wait for a condition to be signaled
// instead of wasting CPU time polling for work.
while ( threadID >= activeThreads
|| threads[threadID].state == THREAD_INITIALIZING
|| (useSleepingThreads && threads[threadID].state == THREAD_AVAILABLE))
{
assert(!sp || useSleepingThreads);
assert(threadID != 0 || useSleepingThreads);
if (threads[threadID].state == THREAD_INITIALIZING)
threads[threadID].state = THREAD_AVAILABLE;
// Grab the lock to avoid races with Thread::wake_up()
lock_grab(&threads[threadID].sleepLock);
// If we are master and all slaves have finished do not go to sleep
for (i = 0; sp && i < activeThreads && !sp->slaves[i]; i++) {}
allFinished = (i == activeThreads);
if (allFinished || allThreadsShouldExit)
{
lock_release(&threads[threadID].sleepLock);
break;
}
// Do sleep here after retesting sleep conditions
if (threadID >= activeThreads || threads[threadID].state == THREAD_AVAILABLE)
cond_wait(&threads[threadID].sleepCond, &threads[threadID].sleepLock);
lock_release(&threads[threadID].sleepLock);
}
// If this thread has been assigned work, launch a search
if (threads[threadID].state == THREAD_WORKISWAITING)
{
assert(!allThreadsShouldExit);
threads[threadID].state = THREAD_SEARCHING;
// Copy split point position and search stack and call search()
// with SplitPoint template parameter set to true.
SearchStack ss[PLY_MAX_PLUS_2];
SplitPoint* tsp = threads[threadID].splitPoint;
Position pos(*tsp->pos, threadID);
memcpy(ss, tsp->ss - 1, 4 * sizeof(SearchStack));
(ss+1)->sp = tsp;
if (tsp->pvNode)
search<PV, true, false>(pos, ss+1, tsp->alpha, tsp->beta, tsp->depth);
else
search<NonPV, true, false>(pos, ss+1, tsp->alpha, tsp->beta, tsp->depth);
assert(threads[threadID].state == THREAD_SEARCHING);
threads[threadID].state = THREAD_AVAILABLE;
// Wake up master thread so to allow it to return from the idle loop in
// case we are the last slave of the split point.
if ( useSleepingThreads
&& threadID != tsp->master
&& threads[tsp->master].state == THREAD_AVAILABLE)
threads[tsp->master].wake_up();
}
// 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.
for (i = 0; sp && i < activeThreads && !sp->slaves[i]; i++) {}
allFinished = (i == activeThreads);
if (allFinished)
{
// 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));
// 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(threads[threadID].state == THREAD_AVAILABLE);
threads[threadID].state = THREAD_SEARCHING;
return;
max_s = s;
*best = Rml[i].pv[0];
*ponder = Rml[i].pv[1];
}
}
}
// init_threads() is called during startup. Initializes locks and condition
// variables and launches all threads sending them immediately to sleep.
void ThreadsManager::init_threads() {
int i, arg[MAX_THREADS];
bool ok;
// 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;
lock_init(&mpLock);
for (i = 0; i < MAX_THREADS; i++)
{
// Initialize thread and split point locks
lock_init(&threads[i].sleepLock);
cond_init(&threads[i].sleepCond);
for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
lock_init(&(threads[i].splitPoints[j].lock));
// All threads but first should be set to THREAD_INITIALIZING
threads[i].state = (i == 0 ? THREAD_SEARCHING : THREAD_INITIALIZING);
}
// Create and startup the threads
for (i = 1; i < MAX_THREADS; i++)
{
arg[i] = i;
#if !defined(_MSC_VER)
pthread_t pthread[1];
ok = (pthread_create(pthread, NULL, init_thread, (void*)(&arg[i])) == 0);
pthread_detach(pthread[0]);
#else
ok = (CreateThread(NULL, 0, init_thread, (LPVOID)(&arg[i]), 0, NULL) != NULL);
#endif
if (!ok)
{
cout << "Failed to create thread number " << i << endl;
exit(EXIT_FAILURE);
}
// Wait until the thread has finished launching and is gone to sleep
while (threads[i].state == THREAD_INITIALIZING) {}
}
}
// exit_threads() is called when the program exits. It makes all the
// helper threads exit cleanly.
void ThreadsManager::exit_threads() {
// 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
if (i != 0)
{
threads[i].wake_up();
while (threads[i].state != THREAD_TERMINATED) {}
}
// Now we can safely destroy the locks and wait conditions
lock_destroy(&threads[i].sleepLock);
cond_destroy(&threads[i].sleepCond);
for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
lock_destroy(&(threads[i].splitPoints[j].lock));
}
lock_destroy(&mpLock);
}
// cutoff_at_splitpoint() checks whether a beta cutoff has occurred in
// the thread's currently active split point, or in some ancestor of
// the current split point.
bool ThreadsManager::cutoff_at_splitpoint(int threadID) const {
assert(threadID >= 0 && threadID < activeThreads);
SplitPoint* sp = threads[threadID].splitPoint;
for ( ; sp && !sp->betaCutoff; sp = sp->parent) {}
return sp != NULL;
}
// thread_is_available() checks whether the thread with threadID "slave" is
// available to help the thread with threadID "master" at a split point. An
// obvious requirement is that "slave" must be idle. With more than two
// threads, this is not by itself sufficient: If "slave" is the master of
// some active split point, it is only available as a slave to the other
// threads which are busy searching the split point at the top of "slave"'s
// split point stack (the "helpful master concept" in YBWC terminology).
bool ThreadsManager::thread_is_available(int slave, int master) const {
assert(slave >= 0 && slave < activeThreads);
assert(master >= 0 && master < activeThreads);
assert(activeThreads > 1);
if (threads[slave].state != THREAD_AVAILABLE || slave == master)
return false;
// Make a local copy to be sure doesn't change under our feet
int localActiveSplitPoints = threads[slave].activeSplitPoints;
// No active split points means that the thread is available as
// a slave for any other thread.
if (localActiveSplitPoints == 0 || activeThreads == 2)
return true;
// Apply the "helpful master" concept if possible. Use localActiveSplitPoints
// that is known to be > 0, instead of threads[slave].activeSplitPoints that
// could have been set to 0 by another thread leading to an out of bound access.
if (threads[slave].splitPoints[localActiveSplitPoints - 1].slaves[master])
return true;
return false;
}
// available_thread_exists() tries to find an idle thread which is available as
// a slave for the thread with threadID "master".
bool ThreadsManager::available_thread_exists(int master) const {
assert(master >= 0 && master < activeThreads);
assert(activeThreads > 1);
for (int i = 0; i < activeThreads; i++)
if (thread_is_available(i, master))
return true;
return false;
}
// split() does the actual work of distributing the work at a node between
// 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. 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 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, bool pvNode) {
assert(pos.is_ok());
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);
assert(activeThreads > 1);
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_thread_exists(master)
|| masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS)
{
lock_release(&mpLock);
return;
}
// Pick the next available split point object from the split point stack
SplitPoint& splitPoint = masterThread.splitPoints[masterThread.activeSplitPoints++];
// Initialize the split point object
splitPoint.parent = masterThread.splitPoint;
splitPoint.master = master;
splitPoint.betaCutoff = false;
splitPoint.depth = depth;
splitPoint.threatMove = threatMove;
splitPoint.alpha = *alpha;
splitPoint.beta = beta;
splitPoint.pvNode = pvNode;
splitPoint.bestValue = *bestValue;
splitPoint.mp = mp;
splitPoint.moveCount = moveCount;
splitPoint.pos = &pos;
splitPoint.nodes = 0;
splitPoint.ss = ss;
for (i = 0; i < activeThreads; i++)
splitPoint.slaves[i] = 0;
masterThread.splitPoint = &splitPoint;
// If we are here it means we are not available
assert(masterThread.state != THREAD_AVAILABLE);
int workersCnt = 1; // At least the master is included
// Allocate available threads setting state to THREAD_BOOKED
for (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;
workersCnt++;
}
assert(Fake || workersCnt > 1);
// We can release the lock because slave threads are already booked and master is not available
lock_release(&mpLock);
// 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.slaves[i])
{
assert(i == master || threads[i].state == THREAD_BOOKED);
threads[i].state = THREAD_WORKISWAITING; // This makes the slave to exit from idle_loop()
if (useSleepingThreads && i != master)
threads[i].wake_up();
}
// 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
// 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
// finished. Update alpha and bestValue, and return.
lock_grab(&mpLock);
*alpha = splitPoint.alpha;
*bestValue = splitPoint.bestValue;
masterThread.activeSplitPoints--;
masterThread.splitPoint = splitPoint.parent;
pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes);
lock_release(&mpLock);
}
/// RootMove and RootMoveList method's definitions
RootMove::RootMove() {
@ -2406,6 +1980,33 @@ split_point_start: // At split points actual search starts from here
return *this;
}
void RootMoveList::init(Position& pos, Move searchMoves[]) {
MoveStack mlist[MAX_MOVES];
Move* sm;
clear();
bestMoveChanges = 0;
// Generate all legal moves and add them to RootMoveList
MoveStack* last = generate<MV_LEGAL>(pos, mlist);
for (MoveStack* cur = mlist; cur != last; cur++)
{
// If we have a searchMoves[] list then verify cur->move
// is in the list before to add it.
for (sm = searchMoves; *sm && *sm != cur->move; sm++) {}
if (searchMoves[0] && *sm != cur->move)
continue;
RootMove rm;
rm.pv[0] = cur->move;
rm.pv[1] = MOVE_NONE;
rm.pv_score = -VALUE_INFINITE;
push_back(rm);
}
}
// extract_pv_from_tt() builds a PV by adding moves from the transposition table.
// We consider also failing high nodes and not only VALUE_TYPE_EXACT nodes. This
// allow to always have a ponder move even when we fail high at root and also a
@ -2487,74 +2088,114 @@ split_point_start: // At split points actual search starts from here
return s.str();
}
void RootMoveList::init(Position& pos, Move searchMoves[]) {
MoveStack mlist[MAX_MOVES];
Move* sm;
clear();
bestMoveChanges = 0;
// Generate all legal moves and add them to RootMoveList
MoveStack* last = generate<MV_LEGAL>(pos, mlist);
for (MoveStack* cur = mlist; cur != last; cur++)
{
// If we have a searchMoves[] list then verify cur->move
// is in the list before to add it.
for (sm = searchMoves; *sm && *sm != cur->move; sm++) {}
if (searchMoves[0] && *sm != cur->move)
continue;
RootMove rm;
rm.pv[0] = cur->move;
rm.pv[1] = MOVE_NONE;
rm.pv_score = -VALUE_INFINITE;
push_back(rm);
}
}
// When playing with strength handicap choose best move among the MultiPV set
// using a statistical rule dependent on SkillLevel. Idea by Heinz van Saanen.
void do_skill_level(Move* best, Move* ponder) {
assert(MultiPV > 1);
// Rml list is already sorted by pv_score in descending order
int s;
int max_s = -VALUE_INFINITE;
int size = Min(MultiPV, (int)Rml.size());
int max = Rml[0].pv_score;
int var = Min(max - Rml[size - 1].pv_score, PawnValueMidgame);
int wk = 120 - 2 * SkillLevel;
// PRNG sequence should be non deterministic
for (int i = abs(get_system_time() % 50); i > 0; i--)
RK.rand<unsigned>();
// Choose best move. For each move's score we add two terms both dependent
// on wk, one deterministic and bigger for weaker moves, and one random,
// then we choose the move with the resulting highest score.
for (int i = 0; i < size; i++)
{
s = Rml[i].pv_score;
// Don't allow crazy blunders even at very low skills
if (i > 0 && Rml[i-1].pv_score > s + EasyMoveMargin)
break;
// This is our magical formula
s += ((max - s) * wk + var * (RK.rand<unsigned>() % wk)) / 128;
if (s > max_s)
{
max_s = s;
*best = Rml[i].pv[0];
*ponder = Rml[i].pv[1];
}
}
}
} // namespace
// ThreadsManager::idle_loop() is where the threads are parked when they have no work
// to do. The parameter 'sp', if non-NULL, is a pointer to an active SplitPoint
// object for which the current thread is the master.
void ThreadsManager::idle_loop(int threadID, SplitPoint* sp) {
assert(threadID >= 0 && threadID < MAX_THREADS);
int i;
bool allFinished;
while (true)
{
// Slave threads can exit as soon as AllThreadsShouldExit raises,
// master should exit as last one.
if (allThreadsShouldExit)
{
assert(!sp);
threads[threadID].state = THREAD_TERMINATED;
return;
}
// If we are not thinking, wait for a condition to be signaled
// instead of wasting CPU time polling for work.
while ( threadID >= activeThreads
|| threads[threadID].state == THREAD_INITIALIZING
|| (useSleepingThreads && threads[threadID].state == THREAD_AVAILABLE))
{
assert(!sp || useSleepingThreads);
assert(threadID != 0 || useSleepingThreads);
if (threads[threadID].state == THREAD_INITIALIZING)
threads[threadID].state = THREAD_AVAILABLE;
// Grab the lock to avoid races with Thread::wake_up()
lock_grab(&threads[threadID].sleepLock);
// If we are master and all slaves have finished do not go to sleep
for (i = 0; sp && i < activeThreads && !sp->slaves[i]; i++) {}
allFinished = (i == activeThreads);
if (allFinished || allThreadsShouldExit)
{
lock_release(&threads[threadID].sleepLock);
break;
}
// Do sleep here after retesting sleep conditions
if (threadID >= activeThreads || threads[threadID].state == THREAD_AVAILABLE)
cond_wait(&threads[threadID].sleepCond, &threads[threadID].sleepLock);
lock_release(&threads[threadID].sleepLock);
}
// If this thread has been assigned work, launch a search
if (threads[threadID].state == THREAD_WORKISWAITING)
{
assert(!allThreadsShouldExit);
threads[threadID].state = THREAD_SEARCHING;
// Copy split point position and search stack and call search()
// with SplitPoint template parameter set to true.
SearchStack ss[PLY_MAX_PLUS_2];
SplitPoint* tsp = threads[threadID].splitPoint;
Position pos(*tsp->pos, threadID);
memcpy(ss, tsp->ss - 1, 4 * sizeof(SearchStack));
(ss+1)->sp = tsp;
if (tsp->pvNode)
search<PV, true, false>(pos, ss+1, tsp->alpha, tsp->beta, tsp->depth);
else
search<NonPV, true, false>(pos, ss+1, tsp->alpha, tsp->beta, tsp->depth);
assert(threads[threadID].state == THREAD_SEARCHING);
threads[threadID].state = THREAD_AVAILABLE;
// Wake up master thread so to allow it to return from the idle loop in
// case we are the last slave of the split point.
if ( useSleepingThreads
&& threadID != tsp->master
&& threads[tsp->master].state == THREAD_AVAILABLE)
threads[tsp->master].wake_up();
}
// 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.
for (i = 0; sp && i < activeThreads && !sp->slaves[i]; i++) {}
allFinished = (i == activeThreads);
if (allFinished)
{
// 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));
// 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(threads[threadID].state == THREAD_AVAILABLE);
threads[threadID].state = THREAD_SEARCHING;
return;
}
}
}

3
src/search.h
View file

@ -63,8 +63,7 @@ struct SearchLimits {
bool infinite, ponder;
};
extern void init_threads();
extern void exit_threads();
extern void init_search();
extern int64_t perft(Position& pos, Depth depth);
extern bool think(Position& pos, const SearchLimits& limits, Move searchMoves[]);

333
src/thread.cpp Normal file
View file

@ -0,0 +1,333 @@
/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
Copyright (C) 2004-2008 Tord Romstad (Glaurung author)
Copyright (C) 2008-2010 Marco Costalba, Joona Kiiski, Tord Romstad
Stockfish is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
Stockfish is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
#include <iostream>
#include "thread.h"
#include "ucioption.h"
ThreadsManager ThreadsMgr; // Global object definition
namespace {
// init_thread() is the function which is called when a new thread is
// launched. It simply calls the idle_loop() function with the supplied
// threadID. There are two versions of this function; one for POSIX
// threads and one for Windows threads.
#if !defined(_MSC_VER)
void* init_thread(void* threadID) {
ThreadsMgr.idle_loop(*(int*)threadID, NULL);
return NULL;
}
#else
DWORD WINAPI init_thread(LPVOID threadID) {
ThreadsMgr.idle_loop(*(int*)threadID, NULL);
return 0;
}
#endif
}
// read_uci_options() updates number of active threads and other internal
// parameters according to the UCI options values. It is called before
// to start a new search.
void ThreadsManager::read_uci_options() {
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>();
}
// init_threads() is called during startup. Initializes locks and condition
// variables and launches all threads sending them immediately to sleep.
void ThreadsManager::init_threads() {
int i, arg[MAX_THREADS];
bool ok;
// 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;
lock_init(&mpLock);
for (i = 0; i < MAX_THREADS; i++)
{
// Initialize thread and split point locks
lock_init(&threads[i].sleepLock);
cond_init(&threads[i].sleepCond);
for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
lock_init(&(threads[i].splitPoints[j].lock));
// All threads but first should be set to THREAD_INITIALIZING
threads[i].state = (i == 0 ? THREAD_SEARCHING : THREAD_INITIALIZING);
// Not in Threads c'tor to avoid global initialization order issues
threads[i].pawnTable.init();
threads[i].materialTable.init();
}
// Create and startup the threads
for (i = 1; i < MAX_THREADS; i++)
{
arg[i] = i;
#if !defined(_MSC_VER)
pthread_t pthread[1];
ok = (pthread_create(pthread, NULL, init_thread, (void*)(&arg[i])) == 0);
pthread_detach(pthread[0]);
#else
ok = (CreateThread(NULL, 0, init_thread, (LPVOID)(&arg[i]), 0, NULL) != NULL);
#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_threads() is called when the program exits. It makes all the
// helper threads exit cleanly.
void ThreadsManager::exit_threads() {
// 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
if (i != 0)
{
threads[i].wake_up();
while (threads[i].state != THREAD_TERMINATED) {}
}
// Now we can safely destroy the locks and wait conditions
lock_destroy(&threads[i].sleepLock);
cond_destroy(&threads[i].sleepCond);
for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
lock_destroy(&(threads[i].splitPoints[j].lock));
}
lock_destroy(&mpLock);
}
// cutoff_at_splitpoint() checks whether a beta cutoff has occurred in
// the thread's currently active split point, or in some ancestor of
// the current split point.
bool ThreadsManager::cutoff_at_splitpoint(int threadID) const {
assert(threadID >= 0 && threadID < activeThreads);
SplitPoint* sp = threads[threadID].splitPoint;
for ( ; sp && !sp->betaCutoff; sp = sp->parent) {}
return sp != NULL;
}
// thread_is_available() checks whether the thread with threadID "slave" is
// available to help the thread with threadID "master" at a split point. An
// obvious requirement is that "slave" must be idle. With more than two
// threads, this is not by itself sufficient: If "slave" is the master of
// some active split point, it is only available as a slave to the other
// threads which are busy searching the split point at the top of "slave"'s
// split point stack (the "helpful master concept" in YBWC terminology).
bool ThreadsManager::thread_is_available(int slave, int master) const {
assert(slave >= 0 && slave < activeThreads);
assert(master >= 0 && master < activeThreads);
assert(activeThreads > 1);
if (threads[slave].state != THREAD_AVAILABLE || slave == master)
return false;
// Make a local copy to be sure doesn't change under our feet
int localActiveSplitPoints = threads[slave].activeSplitPoints;
// No active split points means that the thread is available as
// a slave for any other thread.
if (localActiveSplitPoints == 0 || activeThreads == 2)
return true;
// Apply the "helpful master" concept if possible. Use localActiveSplitPoints
// that is known to be > 0, instead of threads[slave].activeSplitPoints that
// could have been set to 0 by another thread leading to an out of bound access.
if (threads[slave].splitPoints[localActiveSplitPoints - 1].slaves[master])
return true;
return false;
}
// available_thread_exists() tries to find an idle thread which is available as
// a slave for the thread with threadID "master".
bool ThreadsManager::available_thread_exists(int master) const {
assert(master >= 0 && master < activeThreads);
assert(activeThreads > 1);
for (int i = 0; i < activeThreads; i++)
if (thread_is_available(i, master))
return true;
return false;
}
// split() does the actual work of distributing the work at a node between
// 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. 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 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, bool pvNode) {
assert(pos.is_ok());
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);
assert(activeThreads > 1);
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_thread_exists(master)
|| masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS)
{
lock_release(&mpLock);
return;
}
// Pick the next available split point object from the split point stack
SplitPoint& splitPoint = masterThread.splitPoints[masterThread.activeSplitPoints++];
// Initialize the split point object
splitPoint.parent = masterThread.splitPoint;
splitPoint.master = master;
splitPoint.betaCutoff = false;
splitPoint.depth = depth;
splitPoint.threatMove = threatMove;
splitPoint.alpha = *alpha;
splitPoint.beta = beta;
splitPoint.pvNode = pvNode;
splitPoint.bestValue = *bestValue;
splitPoint.mp = mp;
splitPoint.moveCount = moveCount;
splitPoint.pos = &pos;
splitPoint.nodes = 0;
splitPoint.ss = ss;
for (i = 0; i < activeThreads; i++)
splitPoint.slaves[i] = 0;
masterThread.splitPoint = &splitPoint;
// If we are here it means we are not available
assert(masterThread.state != THREAD_AVAILABLE);
int workersCnt = 1; // At least the master is included
// Allocate available threads setting state to THREAD_BOOKED
for (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;
workersCnt++;
}
assert(Fake || workersCnt > 1);
// We can release the lock because slave threads are already booked and master is not available
lock_release(&mpLock);
// 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.slaves[i])
{
assert(i == master || threads[i].state == THREAD_BOOKED);
threads[i].state = THREAD_WORKISWAITING; // This makes the slave to exit from idle_loop()
if (useSleepingThreads && i != master)
threads[i].wake_up();
}
// 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
// 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
// finished. Update alpha and bestValue, and return.
lock_grab(&mpLock);
*alpha = splitPoint.alpha;
*bestValue = splitPoint.bestValue;
masterThread.activeSplitPoints--;
masterThread.splitPoint = splitPoint.parent;
pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes);
lock_release(&mpLock);
}
// Explicit template instantiations
template void ThreadsManager::split<0>(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, bool);
template void ThreadsManager::split<1>(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, bool);

50
src/thread.h
View file

@ -23,9 +23,10 @@
#include <cstring>
#include "lock.h"
#include "material.h"
#include "movepick.h"
#include "pawns.h"
#include "position.h"
#include "search.h"
const int MAX_THREADS = 32;
const int MAX_ACTIVE_SPLIT_POINTS = 8;
@ -67,7 +68,14 @@ enum ThreadState
THREAD_TERMINATED // we are quitting and thread is terminated
};
// We use per-thread Pawn and material hash tables so that once we get a
// pointer to an entry its life time is unlimited and we don't have to
// care about someone changing the entry under our feet.
struct Thread {
MaterialInfoTable materialTable;
PawnInfoTable pawnTable;
int maxPly;
Lock sleepLock;
WaitCondition sleepCond;
@ -83,4 +91,44 @@ struct Thread {
}
};
// ThreadsManager class is used to handle all the threads related stuff like init,
// starting, parking and, the most important, launching a slave thread at a split
// point. All the access to shared thread data is done through this class.
class ThreadsManager {
/* As long as the single ThreadsManager object is defined as a global we don't
need to explicitly initialize to zero its data members because variables with
static storage duration are automatically set to zero before enter main()
*/
public:
Thread& operator[](int threadID) { return threads[threadID]; }
void init_threads();
void exit_threads();
int min_split_depth() const { return minimumSplitDepth; }
int active_threads() const { return activeThreads; }
void set_active_threads(int cnt) { activeThreads = cnt; }
void read_uci_options();
bool available_thread_exists(int master) const;
bool thread_is_available(int slave, int master) const;
bool cutoff_at_splitpoint(int threadID) const;
void idle_loop(int threadID, SplitPoint* sp);
template <bool Fake>
void split(Position& pos, SearchStack* ss, Value* alpha, const Value beta, Value* bestValue,
Depth depth, Move threatMove, int moveCount, MovePicker* mp, bool pvNode);
private:
Lock mpLock;
Depth minimumSplitDepth;
int maxThreadsPerSplitPoint;
bool useSleepingThreads;
int activeThreads;
volatile bool allThreadsShouldExit;
Thread threads[MAX_THREADS];
};
extern ThreadsManager ThreadsMgr;
#endif // !defined(THREAD_H_INCLUDED)

8
src/tt.h
View file

@ -142,13 +142,11 @@ inline void TranspositionTable::refresh(const TTEntry* tte) const {
/// Without cluster concept or overwrite policy.
template<class Entry, int HashSize>
class SimpleHash {
struct SimpleHash {
SimpleHash(const SimpleHash&);
SimpleHash& operator=(const SimpleHash&);
typedef SimpleHash<Entry, HashSize> Base;
public:
SimpleHash() {
void init() {
entries = new (std::nothrow) Entry[HashSize];
if (!entries)