UT_Thread.h

Go to the documentation of this file.

/*
 * PROPRIETARY INFORMATION.  This software is proprietary to
 * Side Effects Software Inc., and is not to be reproduced,
 * transmitted, or disclosed in any way without written permission.
 *
 * NAME:        UT_Thread.h ( UT Library, C++)
 *
 * COMMENTS:    Generic thread class.
 *      The owner of the thread can do things like:
 *
 *      killThread()    - Stop execution of thread
 *      waitThread()    - Wait until thread finishes execution
 *      suspendThread() - Suspend execution of thread
 *      restartThread() - Restart a stopped thread
 *
 *  TODO:  It might be nice to have a way to get the exit status of a thread.
 */

#ifndef __UT_Thread__
#define __UT_Thread__

#include "UT_API.h"
#include "UT_Array.h"
#include "UT_Assert.h"
#include "UT_StringHolder.h"
#include "UT_UniquePtr.h"

#include <SYS/SYS_Deprecated.h>
#include <SYS/SYS_SequentialThreadIndex.h>
#include <SYS/SYS_Types.h>

#include <stdlib.h>

#include <thread>
#include <tuple>

#if defined(WIN32)
#   include <intrin.h>
    typedef int         ut_thread_id_t;
#elif defined(USE_PTHREADS)
#   include <sched.h>
#   include <pthread.h>
    typedef pthread_t   ut_thread_id_t;
#else
    #error Unsupported Platform for UT_Thread
#endif

#define UT_INVALID_THREAD_ID ((ut_thread_id_t)0)

// some stack size defines
#define UT_THREAD_DEFAULT_STACK_SIZE    (8U*1024U*1024U)
#define UT_THREAD_SMALL_STACK_SIZE      (1U*1024U*1024U)

typedef void *(*UTthreadFunc)(void*);

// forward declarations
class UT_TaskScope;

class UT_API UT_Thread
{
public:
    // The destructor will wait until the thread is idle before it completes
    // If you wish to kill the thread, call killThread() first.
    virtual ~UT_Thread();

    UT_Thread(const UT_Thread &) = delete;
    UT_Thread &operator=(const UT_Thread &) = delete;

    // This enum specifies the current state for a persistent thread.  The
    // thread will typically be running or idle.  If the thread is idle, it's
    // behaviour will be determined by the SpinState.
    enum State
    {
        ThreadIdle,
        ThreadRunning
    };

    // // The thread status determines how the thread will behave once the
    // callback function is completed:
    //    ThreadSingleRun       - The thread cannot be restarted
    //    ThreadLowUsage        - The thread will yield cycles while idle
    //
    enum SpinMode
    {
        ThreadSingleRun,
        ThreadLowUsage,
    };

    /// Allocate a new thread
    /// @param spin_mode    Use ThreadSingleRun to have it exit when the thread
    ///                     callback is finished. Otherwise, ThreadLowUsage
    ///                     will cause the thread to loop back and wait for
    ///                     more startThread() calls to run different thread
    ///                     callbacks in the same thread.
    static UT_Thread    *allocThread(SpinMode spin_mode);

    static int           getNumProcessors();

    /// This is only valid in debug builds
    static int           activeThreadCount();

    /// Reset the number of threads that is used by Houdini. This will reread
    /// the HOUDINI_MAXTHREADS setting.
    /// @note There should be no active tasks when this is called.
    /// @note Only call this from the MAIN THREAD!
    static void          resetNumProcessors();

    // getMyThreadId() is inlined for speed if we're using pthreads.
#if defined(USE_PTHREADS)
    static ut_thread_id_t getMyThreadId() { return pthread_self(); }
#else
    static ut_thread_id_t getMyThreadId();
#endif

    static ut_thread_id_t getMainThreadId();
    static int            getMainSequentialThreadId();
    static inline int     isMainThread()
                          {
                             return getMyThreadId() == getMainThreadId();
                          }

    /// Returns true if the current thread is a UT_Thread.
    /// Returns false if the current thread is either the main thread
    /// or a TBB thread.
    static bool           isUTThreadCurrent();

    /// Returns true iff the current thread is allowed to create more tasks.
    /// This is sometimes disabled, to avoid needing to create a UT_TaskArena
    /// for small cases that won't get much benefit from threading.
    /// This should be checked by anything using tbb::parallel_for,
    /// tbb::parallel_invoke, or anything else creating TBB tasks.
    static bool           isThreadingEnabled();

    /// This is used to disable (false) threading for the current thread,
    /// to avoid needing to create a UT_TaskArena for small cases that won't
    /// get much benefit from threading.  It returns if it was enabled before.
    /// It is also used to re-enable (true) threading for the current thread.
    static bool           setThreadingEnabled(bool will_be_enabled);

    class ThreadingDisabledScope
    {
    public:
        ThreadingDisabledScope()
            : myPreviouslyEnabled(setThreadingEnabled(false))
        {}
        ~ThreadingDisabledScope()
        {
            if (myPreviouslyEnabled)
                setThreadingEnabled(true);
        }
    private:
        const bool myPreviouslyEnabled;
    };

    // CPU pauses the task for a given number of cycles
    static inline void    pause(uint cycles)
                          {
                              for(uint i = 0; i < cycles; i++)
#if defined(USE_PTHREADS)
#if defined(ARM64)
                                  __asm__ __volatile__("yield;");
#else
                                  __asm__ __volatile__("pause;");
#endif
#else
                                  _mm_pause();
#endif
                          }
    // Yields the task to the scheduler.
#if defined(USE_PTHREADS)
    static inline void    yield(bool higher_only=false)
                          {
                              if (higher_only)
                              {
                                  ::sched_yield();
                              }
                              else
                              {
                                  // Sleep for 100ns. That's 10,000,000 sleep 
                                  // cycles a second (in case you don't have a
                                  // calculator :-)
                                  struct timespec ts = {0,100};
                                  ::nanosleep(&ts, 0);
                              }
                          }
#else
    static void         yield(bool higher_only=false);
#endif

    /// This function has been deprecated. Use SYS_SequentialThreadIndex::get()
    /// or SYSgetSTID instead.
    static int SYS_DEPRECATED(12.5) getMySequentialThreadIndex()
                         { return SYS_SequentialThreadIndex::get(); }

    /// Configure the global number of tasks used by the system
    /// - The default value of 0 uses the number of logical cores on the system
    /// - A negative value wraps it from the number of logical cores.
    ///   eg. -1 will use all cores except for 1.
    /// - If the negative value exceeds the number of logical cores, it is
    ///   clamped to a value of 1.
    /// @note Only call this in the main thread when there are no tasks active.
    /// @note This function is NOT thread-safe.
    static void          configureMaxThreads(int maxthreads = 0);

    /// Configure the default stack size for threads
    /// - A value of 0 uses the stack size of the main thread
    /// - A value larger than 0 will use that specific stack size
    /// @note Only call this in the main thread when there are no tasks active.
    /// @note This function is NOT thread-safe.
    static void          configureThreadStackSize(int stacksize);

    /// Return the current running thread's stack size
    static int           getCurrentThreadStackSize();

    /// Sets the current thread to minimum priority according to the rules
    /// of the platform. This function fails if called on a thread that is
    /// not a running UT_Thread.
    /// Returns true if the operation was successful, otherwise returns false.
    static bool          minimizeThisThreadPriority();

#if defined(MBSD)
    /// Sets the quality of service (QoS) class of a thread. This is used by
    /// the macOS scheduler to prioritize certain tasks.
    /// @note Calling this is optional, however if it is called, it  must be
    ///       called before startThread()
    /// @note This method is only available on macOS
    void                 setQoS(qos_class_t qos);

    /// Returns the quality of service (QoS) class of a thread.
    /// @note This method is only available on macOS
    /// @see setQoS()
    qos_class_t          getQoS() const;
#endif

    class UT_API DisableGlobalControl
    {
    public:
        DisableGlobalControl();
        ~DisableGlobalControl();

        DisableGlobalControl(const DisableGlobalControl &) = delete;
        DisableGlobalControl &operator=(const DisableGlobalControl &) = delete;
    };

    // Start the thread running.  If the thread is not in idle state, the
    // thread will wait until it's in idle before starting. If the thread
    // doesn't exist yet, it will be created.
    virtual bool         startThread(UTthreadFunc func, void *data,
                                     int stacksize) = 0;

    // Use the global thread stack size set by configureMaxThreads()
    bool                 startThread(UTthreadFunc func, void *data);

    // This method is called when the thread function is first entered.
    // By default it does nothing but some sub-classes may need this.
    virtual void         threadStarted();

    // This method is called when the thread function is returned from.
    // By default it sets the state to idle.
    virtual void         threadEnded();


    // Some thread architectures have very expensive resources (i.e. sproc()
    // threads).  While these threads spin (are idle), they consume system
    // resources.  This method will let the user know whether the threads are
    // resource hogs (so that if they spin for a long time, they could
    // possibley be cleaned up).
    virtual int          isResourceHog() const;

    // For persistent threads (which get restarted)
    virtual State        getState();
    virtual SpinMode     getSpinMode();
    virtual void         waitForState(State desired) = 0;
    virtual void         setSpinMode(SpinMode spin_mode);

    // Assign a name to the thread. This makes it easier to find in a debugger.
    // This should be called before the thread is started.
    void                 setThreadName(const UT_StringHolder &name);

    // Terminate the thread process
    virtual void         killThread() = 0;

    // If it's possible to perform these tasks, the return code will be 1.  If
    // not, the return code will be 0.
    virtual int          suspendThread() = 0;
    virtual int          restartThread() = 0;

    int                  isActive()
                         { return waitThread(0); }

    /// NOTE: This level doesn't own any data apart from itself.
    virtual int64        getMemoryUsage(bool inclusive) const = 0;

protected:
    // System dependent internal functions.
    //  waitThread() returns 1 if the thread is still active (i.e. exists) and
    //  should return 0 if the thread doesn't exist.  If waitThread detects
    //  that the thread no longer exists, it should do appropriate cleanup.
    virtual int          waitThread(int block=1) = 0;

    // Quick check to see that the thread is really active
    virtual int          isValid();

    // This method can be used to kill an idle process.
    void                 killIdle();

    static void         *threadWrapper(void *data);

    // Internally used to change the state safely.
    virtual void         setState(State state) = 0;

    volatile State       myState;
    SpinMode             mySpinMode;
    UTthreadFunc         myCallback;
    void                *myCBData;

    UT_StringHolder      myName;

    const UT_TaskScope  *myTaskScope;

#if defined(MBSD)
    // The quality of service (QoS) of this thread for the macOS scheduler
    qos_class_t          myQoS;
#endif

    UT_Thread(SpinMode spin_mode);

private:
    friend class UT_SubSystem;

    static void onExit_();
};

// For debugging, the following uses a single thread (i.e. is not
// multi-threaded)
class UT_API UT_NullThread : public UT_Thread
{
public:
             UT_NullThread();
            ~UT_NullThread() override;

             UT_NullThread(const UT_NullThread &) = delete;
             UT_NullThread &operator=(const UT_NullThread &) = delete;

    bool                 startThread(UTthreadFunc func, void *data,
                                     int stacksize) override;
    void                 killThread() override;
    int                  waitThread(int block) override;
    void                 waitForState(State) override;

    int                  suspendThread() override;
    int                  restartThread() override;

    int64                getMemoryUsage(bool inclusive) const override
    {
        int64 mem = inclusive ? sizeof(*this) : 0;
        // NOTE: We don't know how much memory Windows uses,
        //       so we can't count it.
        return mem;
    }

protected:
    void                 setState(State state) override;
};


class UT_API UT_ThreadSet
{
public:
     UT_ThreadSet(int nthreads=-1, int null_thread_if_1_cpu = 0);
    ~UT_ThreadSet();

    UT_ThreadSet(const UT_ThreadSet &) = delete;
    UT_ThreadSet &operator=(const UT_ThreadSet &) = delete;

    void        setFunc(UTthreadFunc func)
                {
                    myFunc = func;
                }
    void        setUserData(void *user_data_array, size_t structlen)
                {
                    myUserData = user_data_array;
                    myUserDataInc = structlen;
                }
    void        setUserData(void *user_data)
                {
                    myUserData = user_data;
                    myUserDataInc = 0;
                }

    void        reuse(UT_Thread::SpinMode spin_mode);
    void        go();
    int         wait(int block=1);

    int                  getNumThreads() const { return myThreadCount; }
    UT_Thread           *getThread(int which);
    UT_Thread           *operator[](int which)
                         {
                             UT_ASSERT_P(which < myThreadCount);
                             return myThreads[which];
                         }

protected:
    int           myThreadCount;
    UT_Thread   **myThreads;
    UTthreadFunc  myFunc;
    void         *myUserData;
    int64         myUserDataInc;
};

class UT_API UT_ThreadFarm
{
public:
    enum AssignmentStyle
    {
        NON_BLOCKING    = 0,    // Only assign thread if one is available
        BLOCKING        = 1,    // Block until a thread is free.
        DYNAMIC         = 2     // If no threads are availble, create a new one.
    };
    
    // similar to UT_ThreadSet, but a bit simpler. Called UT_ThreadFarm
    // because it farms out the next available thread. You also don't need to
    // match the number of data chunks to the number of threads.
    // ie.
    // farm = new UT_ThreadFarm(4);
    // while(!done) {
    //     thread = farm->nextThread();
    //     thread->startThread(entrypoint, mydata);
    // }
    // farm->wait();
    
     UT_ThreadFarm(int nthreads=-1);
    ~UT_ThreadFarm();

    UT_ThreadFarm(const UT_ThreadFarm &) = delete;
    UT_ThreadFarm &operator=(const UT_ThreadFarm &) = delete;

    // waits for the next available thread, (or returns null if none are
    // available and block = 0). thread_index will contain the thread index
    // if you pass it a non-null pointer.
    UT_Thread   *nextThread(int *thread_index =0,
                            AssignmentStyle style = BLOCKING);

    // waits until all threads are finished (or, returns 0 if not finished and
    // block = 0).
    int          wait(int block = 1);

    // deletes threads in the thread farm. if kill=1 the threads are killed before
    // cleanup, otherwise wait(1) is called.
    void         cleanup(int kill = 0);

    int          getEntries() const     { return myThreadCount; }
    UT_Thread   *operator[](int index)
                 {
                     UT_ASSERT_P(index < myThreadCount);
                     return myThreads[index];
                 }
    
protected:
    void         addThreads(int thread_count);
    
    int           myThreadCount;
    UT_Thread   **myThreads;
};

// Gradual backoff when there's thread contention.
class UT_API UT_ThreadBackoff
{
public:
    UT_ThreadBackoff() : myCycles(1) {}

    static const uint   cycles_for_noop = 4;
    static const uint   cycles_for_pause = cycles_for_noop * 4;
    static const uint   cycles_for_yield_higher = cycles_for_pause * 2;
    static const uint   cycles_for_yield_all = cycles_for_yield_higher * 2;

    // Same thresholds as hboost::detail::yield(), but different behaviour
    void wait()
    {
        if (myCycles > cycles_for_yield_all)
        {
            // Yield the thread completely, to any and all comers.
            UT_Thread::yield(false);
            return;
        }
        
        if (myCycles <= cycles_for_noop)
        {
            // Noop.
        }
        else if (myCycles <= cycles_for_pause)
        {
            UT_Thread::pause(myCycles);
        }
        else if (myCycles <= cycles_for_yield_higher)
        {
            UT_Thread::yield(true);
        }
        myCycles += (myCycles+1)>>1;
    }

    void reset()
    {
        myCycles = 1;
    }
    
private:
    uint                myCycles;
};

namespace UT
{
namespace detail
{
class UT_API ThreadInit
{
public:
    ThreadInit();
    ~ThreadInit();

    ThreadInit(const ThreadInit &) = delete;
    ThreadInit &operator=(const ThreadInit &) = delete;

};
} // namespace detail
} // namespace UT

class UT_StdThread : public std::thread
{
public:
    UT_StdThread() = default;
    template <typename Func, typename... Args>
    UT_StdThread(Func &&func, Args &&... args)
        : std::thread(
                  WrapFunctor<Func, Args...>(std::forward<Func>(func)),
                  std::forward<Args>(args)...)
    {
    }

    UT_StdThread(const UT_StdThread&) = delete;
    UT_StdThread& operator=(const UT_StdThread&) = delete;
    UT_StdThread(UT_StdThread&&) = default;
    UT_StdThread& operator=(UT_StdThread&&) = default;

private:
    template <typename Func, typename... Args>
    class WrapFunctor
    {
    public:
        WrapFunctor(Func&& func)
            : myFunc(std::move(func))
        {
        }

        decltype(auto) operator()(Args&&... args) const
        {
            UT::detail::ThreadInit scope;
            return myFunc(std::forward<Args>(args)...);
        }
    private:
        Func myFunc;
    };
};

class UT_StdThreadGroup
{
public:
    using thread_t = UT_StdThread;

    explicit UT_StdThreadGroup(int nthreads = -1)
    {
        if (nthreads < 1)
            nthreads = UT_Thread::getNumProcessors();

        myThreads.setSize(nthreads);
    }

    UT_StdThreadGroup(const UT_StdThreadGroup&) = delete;
    UT_StdThreadGroup& operator=(const UT_StdThreadGroup&) = delete;

    thread_t& get(int idx)
    {
        return myThreads(idx);
    }
    const thread_t& get(int idx) const
    {
        return myThreads(idx);
    }
    thread_t& operator[](int idx)
    {
        return myThreads[idx];
    }
    const thread_t& operator[](int idx) const
    {
        return myThreads[idx];
    }
    bool joinable() const
    {
        for (auto&& t : myThreads)
        {
            if (!t.joinable())
                return false;
        }
        return true;
    }
    bool joinable(int idx) const
    {
        return get(idx).joinable();
    }
    void join()
    {
        for (auto&& t : myThreads)
        {
            if (t.joinable())
                t.join();
        }
    }
private:
    UT_Array<thread_t> myThreads;
};

// This function has been deprecated. Use SYSgetSTID instead.
static inline int SYS_DEPRECATED(12.5)
UTgetSTID()
{
    return SYS_SequentialThreadIndex::get();
}

#endif