multithread.c

/*
 * multithread.c
 *
 * by Jon Kinsey, 2008
 *
 * Multithreaded operations
 *
 *
 * This program is free software; you can redistribute it and/or modify
 * it under the terms of version 3 or later of the GNU General Public License as
 * published by the Free Software Foundation.
 *
 * This program 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, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 * 
 * $Id$
 */

#include "config.h"
#if USE_MULTITHREAD
#ifdef WIN32
#include <process.h>
#endif
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <glib.h>
#if USE_GTK
#include <gtk/gtk.h>
#include <gtkgame.h>
#endif

#include "multithread.h"
#include "speed.h"
#include "rollout.h"
#include "util.h"

#define UI_UPDATETIME 250

#if defined(DEBUG_MULTITHREADED) && defined(WIN32)
unsigned int mainThreadID;
#endif

#ifdef GLIB_THREADS
typedef struct _ManualEvent {
#if GLIB_CHECK_VERSION (2,32,0)
    GCond cond;
#else
    GCond *cond;
#endif
    int signalled;
} *ManualEvent;
typedef GPrivate *TLSItem;
#if GLIB_CHECK_VERSION (2,32,0)
typedef GMutex Mutex;
static GMutex condMutex;        /* Extra mutex needed for waiting */
#else
typedef GMutex *Mutex;
GMutex *condMutex = NULL;       /* Extra mutex needed for waiting */
#endif
GAsyncQueue *async_queue = NULL;        /* Needed for async waiting */
#else
typedef HANDLE ManualEvent;
typedef DWORD TLSItem;
typedef HANDLE Event;
typedef HANDLE Mutex;
#endif

typedef struct _ThreadData {
    ManualEvent activity;
    TLSItem tlsItem;
    Mutex queueLock;
    Mutex multiLock;
    ManualEvent syncStart;
    ManualEvent syncEnd;

    int addedTasks;
    int doneTasks;
    int totalTasks;

    GList *tasks;
    int result;

    int closingThreads;
    unsigned int numThreads;
} ThreadData;

ThreadData td;

#ifdef GLIB_THREADS

#if GLIB_CHECK_VERSION (2,32,0)
/* Dynamic allocation of GPrivate is deprecated */
static GPrivate private_item;

static void
TLSCreate(TLSItem * pItem)
{
    *pItem = &private_item;
}
#else
static void
TLSCreate(TLSItem * pItem)
{
    *pItem = g_private_new(g_free);
}
#endif

static void
TLSFree(TLSItem UNUSED(pItem))
{                               /* Done automaticaly by glib */
}

static void
TLSSetValue(TLSItem pItem, int value)
{
    int *pNew = (int *) malloc(sizeof(int));
    *pNew = value;
    g_private_set(pItem, (gpointer) pNew);
}

#define TLSGet(item) *((int*)g_private_get(item))

static void
InitManualEvent(ManualEvent * pME)
{
    ManualEvent pNewME = malloc(sizeof(*pNewME));
#if GLIB_CHECK_VERSION (2,32,0)
    g_cond_init(&pNewME->cond);
#else
    pNewME->cond = g_cond_new();
#endif
    pNewME->signalled = FALSE;
    *pME = pNewME;
}

static void
FreeManualEvent(ManualEvent ME)
{
#if GLIB_CHECK_VERSION (2,32,0)
    g_cond_clear(&ME->cond);
#else
    g_cond_free(ME->cond);
#endif
    free(ME);
}

static void
WaitForManualEvent(ManualEvent ME)
{
#if GLIB_CHECK_VERSION (2,32,0)
    gint64 end_time;
#else
    GTimeVal tv;
#endif
    multi_debug("wait for manual event locks");
#if GLIB_CHECK_VERSION (2,32,0)
    g_mutex_lock(&condMutex);
    end_time = g_get_monotonic_time() + 10 * G_TIME_SPAN_SECOND;
#else
    g_mutex_lock(condMutex);
#endif
    while (!ME->signalled) {
        multi_debug("waiting for manual event");
#if GLIB_CHECK_VERSION (2,32,0)
        if (!g_cond_wait_until(&ME->cond, &condMutex, end_time))
#else
        g_get_current_time(&tv);
        g_time_val_add(&tv, 10 * 1000 * 1000);
        if (g_cond_timed_wait(ME->cond, condMutex, &tv))
#endif
            break;
        else {
            multi_debug("still waiting for manual event");
        }
    }

#if GLIB_CHECK_VERSION (2,32,0)
    g_mutex_unlock(&condMutex);
#else
    g_mutex_unlock(condMutex);
#endif
    multi_debug("wait for manual event unlocks");
}

static void
ResetManualEvent(ManualEvent ME)
{
    multi_debug("reset manual event locks");
#if GLIB_CHECK_VERSION (2,32,0)
    g_mutex_lock(&condMutex);
    ME->signalled = FALSE;
    g_mutex_unlock(&condMutex);
#else
    g_mutex_lock(condMutex);
    ME->signalled = FALSE;
    g_mutex_unlock(condMutex);
#endif
    multi_debug("reset manual event unlocks");
}

static void
SetManualEvent(ManualEvent ME)
{
    multi_debug("reset manual event locks");
#if GLIB_CHECK_VERSION (2,32,0)
    g_mutex_lock(&condMutex);
    ME->signalled = TRUE;
    g_cond_broadcast(&ME->cond);
    g_mutex_unlock(&condMutex);
#else
    g_mutex_lock(condMutex);
    ME->signalled = TRUE;
    g_cond_broadcast(ME->cond);
    g_mutex_unlock(condMutex);
#endif
    multi_debug("reset manual event unlocks");
}

#if GLIB_CHECK_VERSION (2,32,0)
static void
InitMutex(Mutex * pMutex)
{
    g_mutex_init(pMutex);
}

static void
FreeMutex(Mutex mutex)
{
    g_mutex_clear(&mutex);
}
#else
static void
InitMutex(Mutex * pMutex)
{
    *pMutex = g_mutex_new();
}

static void
FreeMutex(Mutex mutex)
{
    g_mutex_free(mutex);
}
#endif

static void
Mutex_Lock(Mutex mutex, const char *reason)
{
#ifdef DEBUG_MULTITHREADED
    multi_debug(reason);
#else
    (void) reason;
#endif
#if GLIB_CHECK_VERSION (2,32,0)
    g_mutex_lock(&mutex);
#else
    g_mutex_lock(mutex);
#endif
}

static void
Mutex_Release(Mutex mutex)
{
#ifdef DEBUG_MULTITHREADED
    multi_debug("Releasing lock");
#endif
#if GLIB_CHECK_VERSION (2,32,0)
    g_mutex_unlock(&mutex);
#else
    g_mutex_unlock(mutex);
#endif
}

#else                           /* win32 */

static void
TLSCreate(TLSItem * ppItem)
{
    *ppItem = TlsAlloc();
    if (*ppItem == TLS_OUT_OF_INDEXES)
        PrintSystemError("calling TlsAlloc");
}

static void
TLSFree(TLSItem pItem)
{
    free(TlsGetValue(pItem));
    TlsFree(pItem);
}

static void
TLSSetValue(TLSItem pItem, int value)
{
    int *pNew = (int *) malloc(sizeof(int));
    *pNew = value;
    if (TlsSetValue(pItem, pNew) == 0)
        PrintSystemError("calling TLSSetValue");
}

#define TLSGet(item) *((int*)TlsGetValue(item))

static void
InitManualEvent(ManualEvent * pME)
{
    *pME = CreateEvent(NULL, TRUE, FALSE, NULL);
    if (*pME == NULL)
        PrintSystemError("creating manual event");
}

static void
FreeManualEvent(ManualEvent ME)
{
    CloseHandle(ME);
}

#define WaitForManualEvent(ME) WaitForSingleObject(ME, INFINITE)
#define ResetManualEvent(ME) ResetEvent(ME)
#define SetManualEvent(ME) SetEvent(ME)

static void
InitMutex(Mutex * pMutex)
{
    *pMutex = CreateMutex(NULL, FALSE, NULL);
    if (*pMutex == NULL)
        PrintSystemError("creating mutex");
}

static void
FreeMutex(Mutex mutex)
{
    CloseHandle(mutex);
}

#ifdef DEBUG_MULTITHREADED
void
Mutex_Lock(Mutex mutex, const char *reason)
{
    if (WaitForSingleObject(mutex, 0) == WAIT_OBJECT_0) {       /* Got mutex */
        multi_debug("%s: lock acquired", reason);
    } else {
        multi_debug("%s: waiting on lock", reason);
        WaitForSingleObject(mutex, INFINITE);
        multi_debug("lock relinquished");
    }
}

void
Mutex_Release(Mutex mutex)
{
    multi_debug("Releasing lock");
    ReleaseMutex(mutex);
}
#else
#define Mutex_Lock(mutex, reason) WaitForSingleObject(mutex, INFINITE)
#define Mutex_Release(mutex) ReleaseMutex(mutex)
#endif

#endif

extern unsigned int
MT_GetNumThreads(void)
{
    return td.numThreads;
}

static void
CloseThread(void *UNUSED(unused))
{
    g_assert(td.closingThreads);
    MT_SafeInc(&td.result);
}

static void
MT_CloseThreads(void)
{
    td.closingThreads = TRUE;
    mt_add_tasks(td.numThreads, CloseThread, NULL, NULL);
    if (MT_WaitForTasks(NULL, 0, FALSE) != (int) td.numThreads)
        g_print("Error closing threads!\n");
}

static void
MT_TaskDone(Task * pt)
{
    MT_SafeInc(&td.doneTasks);

    if (pt) {
        free(pt->pLinkedTask);
        free(pt);
    }
}

static Task *
MT_GetTask(void)
{
    Task *task = NULL;

    Mutex_Lock(td.queueLock, "get task");

    if (g_list_length(td.tasks) > 0) {
        task = (Task *) g_list_first(td.tasks)->data;
        td.tasks = g_list_delete_link(td.tasks, g_list_first(td.tasks));
        if (g_list_length(td.tasks) == 0) {
            ResetManualEvent(td.activity);
        }
    }

    multi_debug("get task: release");
    Mutex_Release(td.queueLock);

    return task;
}

extern void
MT_AbortTasks(void)
{
    Task *task;
    /* Remove tasks from list */
    while ((task = MT_GetTask()) != NULL)
        MT_TaskDone(task);

    td.result = -1;
}

#ifdef GLIB_THREADS
static gpointer
MT_WorkerThreadFunction(void *id)
#else
static void
MT_WorkerThreadFunction(void *id)
#endif
{
    /* why do we need this align ? - because of a gcc bug */
#if __GNUC__ && defined(WIN32)
    /* Align stack pointer on 16 byte boundary so SSE variables work correctly */
    int align_offset;
    asm __volatile__("andl $-16, %%esp":::"%esp");
    align_offset = ((int) (&align_offset)) % 16;
#endif
    {
        int *pID = (int *) id;
        TLSSetValue(td.tlsItem, *pID);
        free(pID);
        MT_SafeInc(&td.result);
        MT_TaskDone(NULL);      /* Thread created */
        do {
            Task *task;
            WaitForManualEvent(td.activity);
            task = MT_GetTask();
            if (task) {
                task->fun(task->data);
                MT_TaskDone(task);
            }
        } while (!td.closingThreads);

#ifdef GLIB_THREADS
#if __GNUC__ && defined(WIN32)
        /* De-align stack pointer to avoid crash on exit */
        asm __volatile__("addl %0, %%esp"::"r"(align_offset):"%esp");
#endif
        return NULL;
#endif
    }
}

static gboolean
WaitingForThreads(gpointer UNUSED(unused))
{                               /* Unlikely to be called */
    multi_debug("Waiting for threads to be created!");
    return FALSE;
}

static void
MT_CreateThreads(void)
{
    unsigned int i;
    multi_debug("CreateThreads()");
    td.result = 0;
    td.closingThreads = FALSE;
    for (i = 0; i < td.numThreads; i++) {
        int *pID = (int *) malloc(sizeof(int));
        *pID = i;
#ifdef GLIB_THREADS
#if GLIB_CHECK_VERSION (2,32,0)
        if (!g_thread_try_new("Worker", MT_WorkerThreadFunction, pID, NULL))
#else
        if (!g_thread_create(MT_WorkerThreadFunction, pID, FALSE, NULL))
#endif
#else
        if (_beginthread(MT_WorkerThreadFunction, 0, pID) == 0)
#endif
            printf("Failed to create thread\n");
    }
    td.addedTasks = td.numThreads;
    /* Wait for all the threads to be created (timeout after 1 second) */
    if (MT_WaitForTasks(WaitingForThreads, 1000, FALSE) != (int) td.numThreads)
        g_print("Error creating threads!\n");
}

void
MT_SetNumThreads(unsigned int num)
{
    if (num != td.numThreads) {
        if (td.numThreads != 0)
            MT_CloseThreads();
        td.numThreads = num;
        MT_CreateThreads();
        if (num == 1) {         /* No locking in evals */
            EvaluatePosition = EvaluatePositionNoLocking;
            GeneralCubeDecisionE = GeneralCubeDecisionENoLocking;
            GeneralEvaluationE = GeneralEvaluationENoLocking;
            ScoreMove = ScoreMoveNoLocking;
            FindBestMove = FindBestMoveNoLocking;
            FindnSaveBestMoves = FindnSaveBestMovesNoLocking;
            BasicCubefulRollout = BasicCubefulRolloutNoLocking;
        } else {                /* Locking version of evals */
            EvaluatePosition = EvaluatePositionWithLocking;
            GeneralCubeDecisionE = GeneralCubeDecisionEWithLocking;
            GeneralEvaluationE = GeneralEvaluationEWithLocking;
            ScoreMove = ScoreMoveWithLocking;
            FindBestMove = FindBestMoveWithLocking;
            FindnSaveBestMoves = FindnSaveBestMovesWithLocking;
            BasicCubefulRollout = BasicCubefulRolloutWithLocking;
        }
    }
}

extern void
MT_InitThreads(void)
{
#if !GLIB_CHECK_VERSION (2,32,0)
    if (!g_thread_supported())
        g_thread_init(NULL);
    g_assert(g_thread_supported());
#endif

    td.tasks = NULL;
    td.doneTasks = td.addedTasks = 0;
    td.totalTasks = -1;
    InitManualEvent(&td.activity);
    TLSCreate(&td.tlsItem);
    TLSSetValue(td.tlsItem, 0); /* Main thread shares id 0 */
#if defined(DEBUG_MULTITHREADED) && defined(WIN32)
    mainThreadID = GetCurrentThreadId();
#endif
    InitMutex(&td.multiLock);
    InitMutex(&td.queueLock);
    InitManualEvent(&td.syncStart);
    InitManualEvent(&td.syncEnd);
#if defined(GLIB_THREADS) && !GLIB_CHECK_VERSION (2,32,0)
    if (condMutex == NULL)
        condMutex = g_mutex_new();
#endif
    td.numThreads = 0;
}

extern void
MT_StartThreads(void)
{
    if (td.numThreads == 0) {
        /* We could set it to something else (the number of cores or some
         * fraction of that ?) but it is probably not a good idea to hog
         * a lot of resources by default.
         */
        td.numThreads = 1;

        MT_CreateThreads();
    }
}

void
MT_AddTask(Task * pt, gboolean lock)
{
    if (lock) {
        Mutex_Lock(td.queueLock, "add task");
    }
    if (td.addedTasks == 0)
        td.result = 0;          /* Reset result for new tasks */
    td.addedTasks++;
    td.tasks = g_list_append(td.tasks, pt);
    if (g_list_length(td.tasks) == 1) { /* New tasks */
        SetManualEvent(td.activity);
    }
    if (lock) {
        multi_debug("add task: release");
        Mutex_Release(td.queueLock);
    }
}

void
mt_add_tasks(unsigned int num_tasks, AsyncFun pFun, void *taskData, gpointer linked)
{
    unsigned int i;
    {
#ifdef DEBUG_MULTITHREADED
        char buf[20];
        sprintf(buf, "add %u tasks", num_tasks);
        Mutex_Lock(td.queueLock, buf);
#else
        Mutex_Lock(td.queueLock, NULL);
#endif
    }
    for (i = 0; i < num_tasks; i++) {
        Task *pt = (Task *) malloc(sizeof(Task));
        pt->fun = pFun;
        pt->data = taskData;
        pt->pLinkedTask = linked;
        MT_AddTask(pt, FALSE);
    }
    multi_debug("add many release: lock");
    Mutex_Release(td.queueLock);
}

static gboolean
WaitForAllTasks(int time)
{
    int j = 0;
    while (td.doneTasks != td.totalTasks) {
        if (j == 10)
            return FALSE;       /* Not done yet */

        j++;
        g_usleep(100 * time);
    }
    return TRUE;
}

int
MT_WaitForTasks(gboolean(*pCallback) (gpointer), int callbackTime, int autosave)
{
    int callbackLoops = callbackTime / UI_UPDATETIME;
    int waits = 0;
    int polltime = callbackLoops ? UI_UPDATETIME : callbackTime;
    guint as_source = 0;

    /* Set total tasks to wait for */
    td.totalTasks = td.addedTasks;
#if USE_GTK
    GTKSuspendInput();
#endif

    if (autosave)
        as_source = g_timeout_add(nAutoSaveTime * 60000, save_autosave, NULL);
    multi_debug("Waiting for all tasks");
    while (!WaitForAllTasks(polltime)) {
        waits++;
        if (pCallback && waits >= callbackLoops) {
            waits = 0;
            pCallback(NULL);
        }
        ProcessEvents();
    }
    if (autosave) {
        g_source_remove(as_source);
        save_autosave(NULL);
    }
    multi_debug("Done waiting for all tasks");

    td.doneTasks = td.addedTasks = 0;
    td.totalTasks = -1;

#if USE_GTK
    GTKResumeInput();
#endif
    return td.result;
}

extern void
MT_SetResultFailed(void)
{
    td.result = -1;
}

extern void
MT_Close(void)
{
    MT_CloseThreads();

    FreeManualEvent(td.activity);
    FreeMutex(td.multiLock);
    FreeMutex(td.queueLock);

    FreeManualEvent(td.syncStart);
    FreeManualEvent(td.syncEnd);
    TLSFree(td.tlsItem);
}

extern int
MT_GetThreadID(void)
{
    return TLSGet(td.tlsItem);
}

extern void
MT_Exclusive(void)
{
    Mutex_Lock(td.multiLock, "Exclusive lock");
}

extern void
MT_Release(void)
{
    Mutex_Release(td.multiLock);
}

extern int
MT_GetDoneTasks(void)
{
    return td.doneTasks;
}

/* Code below used in calibrate to try and get a resonable figure for multiple threads */

static double start;            /* used for timekeeping */

extern void
MT_SyncInit(void)
{
    ResetManualEvent(td.syncStart);
    ResetManualEvent(td.syncEnd);
}

extern void
MT_SyncStart(void)
{
    static int count = 0;

    /* Wait for all threads to get here */
    if (MT_SafeIncValue(&count) == (int) td.numThreads) {
        count--;
        start = get_time();
        SetManualEvent(td.syncStart);
    } else {
        WaitForManualEvent(td.syncStart);
        if (MT_SafeDecCheck(&count))
            ResetManualEvent(td.syncStart);
    }
}

extern double
MT_SyncEnd(void)
{
    static int count = 0;

    /* Wait for all threads to get here */
    if (MT_SafeIncValue(&count) == (int) td.numThreads) {
        const double now = get_time();
        count--;
        SetManualEvent(td.syncEnd);
        return now - start;
    } else {
        WaitForManualEvent(td.syncEnd);
        if (MT_SafeDecCheck(&count))
            ResetManualEvent(td.syncEnd);
        return 0;
    }
}

#ifdef DEBUG_MULTITHREADED
void
multi_debug(const char *str, ...)
{
    int id;
    va_list vl;
    char tn[5];
    char buf[1024];

    /* Sync output so order makes some sense */
#ifdef WIN32
    WaitForSingleObject(td.multiLock, INFINITE);
#endif
    /* With glib threads, locking here doesn't seem to work :
     * GLib (gthread-posix.c): Unexpected error from C library during 'pthread_mutex_lock': Resource deadlock avoided.  Aborting.
     * Timestamp the output instead.
     * Maybe the g_get_*_time() calls will sync output as well anyway... */

    va_start(vl, str);
    vsprintf(buf, str, vl);

    id = MT_GetThreadID();
#if defined(WIN32)
    if (id == 0 && GetCurrentThreadId() == mainThreadID)
#else
    if (id == 0)
#endif
        strcpy(tn, "MT");
    else
        sprintf(tn, "T%d", id + 1);

#ifdef GLIB_THREADS
#if GLIB_CHECK_VERSION (2,28,0)
    printf("%" G_GINT64_FORMAT " %s: %s\n", g_get_monotonic_time(), tn, buf);
#else
    {
        GTimeVal tv;

        g_get_current_time(&tv);
        printf("%lu %s: %s\n", 1000000UL * tv.tv_sec + tv.tv_usec, tn, buf);
    }
#endif
#else
    printf("%s: %s\n", tn, buf);
#endif

#ifdef WIN32
    ReleaseMutex(td.multiLock);
#endif
}
#endif

#else
#include "multithread.h"
#include <stdlib.h>
#if USE_GTK
#include <gtk/gtk.h>
#include <gtkgame.h>
#endif

#define UI_UPDATETIME 250

typedef struct _ThreadData {
    int doneTasks;
    GList *tasks;
    int result;
} ThreadData;
ThreadData td;

int asyncRet;
void
MT_AddTask(Task * pt, gboolean lock)
{
    td.result = 0;              /* Reset result for new tasks */
    td.tasks = g_list_append(td.tasks, pt);
}

void
mt_add_tasks(unsigned int num_tasks, AsyncFun pFun, void *taskData, gpointer linked)
{
    unsigned int i;
    for (i = 0; i < num_tasks; i++) {
        Task *pt = (Task *) malloc(sizeof(Task));
        pt->fun = pFun;
        pt->data = taskData;
        pt->pLinkedTask = linked;
        MT_AddTask(pt, FALSE);
    }
}

extern int
MT_GetDoneTasks(void)
{
    return td.doneTasks;
}

int
MT_WaitForTasks(gboolean(*pCallback) (gpointer), int callbackTime, int autosave)
{
    GList *member;
    guint cb_source = 0;
    guint as_source = 0;
    td.doneTasks = 0;

#if USE_GTK
    GTKSuspendInput();
#endif

    multi_debug("Waiting for all tasks");

    pCallback(NULL);
    cb_source = g_timeout_add(1000, pCallback, NULL);
    if (autosave)
        as_source = g_timeout_add(nAutoSaveTime * 60000, save_autosave, NULL);
    for (member = g_list_first(td.tasks); member; member = member->next, td.doneTasks++) {
        Task *task = member->data;
        task->fun(task->data);
        free(task->pLinkedTask);
        free(task);
        ProcessEvents();
    }
    g_list_free(td.tasks);
    if (autosave) {
        g_source_remove(as_source);
        save_autosave(NULL);
    }
    td.tasks = NULL;

#if USE_GTK
    GTKResumeInput();
#endif
    return td.result;
}

extern void
MT_AbortTasks(void)
{
    td.result = -1;
}
#endif