tbb_misc.h

Go to the documentation of this file.

/*
    Copyright (c) 2005-2020 Intel Corporation

    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.
    You may obtain a copy of the License at

        http://www.apache.org/licenses/LICENSE-2.0

    Unless required by applicable law or agreed to in writing, software
    distributed under the License is distributed on an "AS IS" BASIS,
    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    See the License for the specific language governing permissions and
    limitations under the License.
*/

#ifndef _TBB_tbb_misc_H
#define _TBB_tbb_misc_H

#include "tbb/tbb_stddef.h"
#include "tbb/tbb_machine.h"
#include "tbb/atomic.h"     // For atomic_xxx definitions

#if __TBB_NUMA_SUPPORT
#include "tbb/info.h"
#endif /*__TBB_NUMA_SUPPORT*/

#if __linux__ || __FreeBSD__
#include <sys/param.h>  // __FreeBSD_version
#if __FreeBSD_version >= 701000
#include <sys/cpuset.h>
#endif
#endif

// Does the operating system have a system call to pin a thread to a set of OS processors?
#define __TBB_OS_AFFINITY_SYSCALL_PRESENT ((__linux__ && !__ANDROID__) || (__FreeBSD_version >= 701000))
// On IBM* Blue Gene* CNK nodes, the affinity API has restrictions that prevent its usability for TBB,
// and also sysconf(_SC_NPROCESSORS_ONLN) already takes process affinity into account.
#define __TBB_USE_OS_AFFINITY_SYSCALL (__TBB_OS_AFFINITY_SYSCALL_PRESENT && !__bg__)

namespace tbb {

namespace internal {

const size_t MByte = 1024*1024;

#if __TBB_WIN8UI_SUPPORT && (_WIN32_WINNT < 0x0A00)
// In Win8UI mode (Windows 8 Store* applications), TBB uses a thread creation API
// that does not allow to specify the stack size.
// Still, the thread stack size value, either explicit or default, is used by the scheduler.
// So here we set the default value to match the platform's default of 1MB.
const size_t ThreadStackSize = 1*MByte;
#else
const size_t ThreadStackSize = (sizeof(uintptr_t) <= 4 ? 2 : 4 )*MByte;
#endif

#ifndef __TBB_HardwareConcurrency

int AvailableHwConcurrency();

#else

inline int AvailableHwConcurrency() {
    int n = __TBB_HardwareConcurrency();
    return n > 0 ? n : 1; // Fail safety strap
}
#endif /* __TBB_HardwareConcurrency */

size_t DefaultSystemPageSize();

#if _WIN32||_WIN64


int NumberOfProcessorGroups();

int FindProcessorGroupIndex ( int processorIndex );

void MoveThreadIntoProcessorGroup( void* hThread, int groupIndex );

#endif /* _WIN32||_WIN64 */

void handle_win_error( int error_code );

void PrintVersion();

void PrintExtraVersionInfo( const char* category, const char* format, ... );

void PrintRMLVersionInfo( void* arg, const char* server_info );

// For TBB compilation only; not to be used in public headers
#if defined(min) || defined(max)
#undef min
#undef max
#endif


template<typename T>
T min ( const T& val1, const T& val2 ) {
    return val1 < val2 ? val1 : val2;
}


template<typename T>
T max ( const T& val1, const T& val2 ) {
    return val1 < val2 ? val2 : val1;
}

template<int > struct int_to_type {};

//------------------------------------------------------------------------
// FastRandom
//------------------------------------------------------------------------

unsigned GetPrime ( unsigned seed );


class FastRandom {
private:
#if __TBB_OLD_PRIMES_RNG
    unsigned x, a;
    static const unsigned c = 1;
#else
    unsigned x, c;
    static const unsigned a = 0x9e3779b1; // a big prime number
#endif //__TBB_OLD_PRIMES_RNG
public:
    unsigned short get() {
        return get(x);
    }
    unsigned short get( unsigned& seed ) {
        unsigned short r = (unsigned short)(seed>>16);
        __TBB_ASSERT(c&1, "c must be odd for big rng period");
        seed = seed*a+c;
        return r;
    }
    FastRandom( void* unique_ptr ) { init(uintptr_t(unique_ptr)); }
    FastRandom( uint32_t seed) { init(seed); }
    FastRandom( uint64_t seed) { init(seed); }
    template <typename T>
    void init( T seed ) {
        init(seed,int_to_type<sizeof(seed)>());
    }
    void init( uint64_t seed , int_to_type<8> ) {
        init(uint32_t((seed>>32)+seed), int_to_type<4>());
    }
    void init( uint32_t seed, int_to_type<4> ) {
#if __TBB_OLD_PRIMES_RNG
        x = seed;
        a = GetPrime( seed );
#else
        // threads use different seeds for unique sequences
        c = (seed|1)*0xba5703f5; // c must be odd, shuffle by a prime number
        x = c^(seed>>1); // also shuffle x for the first get() invocation
#endif
    }
};

//------------------------------------------------------------------------
// Atomic extensions
//------------------------------------------------------------------------


template<typename T1, typename T2, class Pred>
T1 atomic_update ( tbb::atomic<T1>& dst, T2 newValue, Pred compare ) {
    T1 oldValue = dst;
    while ( compare(oldValue, newValue) ) {
        if ( dst.compare_and_swap((T1)newValue, oldValue) == oldValue )
            break;
        oldValue = dst;
    }
    return oldValue;
}

enum do_once_state {
    do_once_uninitialized = 0,  
    do_once_pending,            
    do_once_executed,           
    initialization_complete = do_once_executed  
};


template <typename F>
void atomic_do_once ( const F& initializer, atomic<do_once_state>& state ) {
    // tbb::atomic provides necessary acquire and release fences.
    // The loop in the implementation is necessary to avoid race when thread T2
    // that arrived in the middle of initialization attempt by another thread T1
    // has just made initialization possible.
    // In such a case T2 has to rely on T1 to initialize, but T1 may already be past
    // the point where it can recognize the changed conditions.
    while ( state != do_once_executed ) {
        if( state == do_once_uninitialized ) {
            if( state.compare_and_swap( do_once_pending, do_once_uninitialized ) == do_once_uninitialized ) {
                run_initializer( initializer, state );
                break;
            }
        }
        spin_wait_while_eq( state, do_once_pending );
    }
}

// Run the initializer which can not fail
inline void run_initializer( void (*f)(), atomic<do_once_state>& state ) {
    f();
    state = do_once_executed;
}

// Run the initializer which can require repeated call
inline void run_initializer( bool (*f)(), atomic<do_once_state>& state ) {
    state = f() ? do_once_executed : do_once_uninitialized;
}

#if __TBB_USE_OS_AFFINITY_SYSCALL
  #if __linux__
    typedef cpu_set_t basic_mask_t;
  #elif __FreeBSD_version >= 701000
    typedef cpuset_t basic_mask_t;
  #else
    #error affinity_helper is not implemented in this OS
  #endif
    class affinity_helper : no_copy {
        basic_mask_t* threadMask;
        int is_changed;
    public:
        affinity_helper() : threadMask(NULL), is_changed(0) {}
        ~affinity_helper();
        void protect_affinity_mask( bool restore_process_mask  );
        void dismiss();
    };
    void destroy_process_mask();
#else
    class affinity_helper : no_copy {
    public:
        void protect_affinity_mask( bool ) {}
        void dismiss() {}
    };
    inline void destroy_process_mask(){}
#endif /* __TBB_USE_OS_AFFINITY_SYSCALL */

bool cpu_has_speculation();
bool gcc_rethrow_exception_broken();
void fix_broken_rethrow();

#if __TBB_NUMA_SUPPORT
class binding_handler;

binding_handler* construct_binding_handler(int slot_num);
void destroy_binding_handler(binding_handler* handler_ptr);
void bind_thread_to_node(binding_handler* handler_ptr, int slot_num , int numa_id);
void restore_affinity_mask(binding_handler* handler_ptr, int slot_num);

namespace numa_topology {
    bool is_initialized();
    void initialize();
    void destroy();
}

#endif /*__TBB_NUMA_SUPPORT*/

} // namespace internal
} // namespace tbb

#endif /* _TBB_tbb_misc_H */