source: trunk/user/sort/sort.c @ 629

///////////////////////////////////////////////////////////////////////////////
// File   :  sort.c
// Date   :  November 2013
// Author :  Cesar Fuguet Tortolero <cesar.fuguet-tortolero@lip6.fr>
///////////////////////////////////////////////////////////////////////////////
// This multi-threaded application implement a multi-stage sort application.
// The various stages are separated by synchronisation barriers.
// There is one thread per physical cores. 
// Computation is organised as a binary tree: 
// - All threads execute in parallel a buble sort on a sub-array during the
//   the first stage of parallel sort,
// - The number of participating threads is divided by 2 at each next stage,
//   to make a merge sort, on two subsets of previous stage.
//
//       Number_of_stages = number of barriers = log2(Number_of_threads)
//
// Constraints :
// - It supports up to 1024 cores: x_size, y_size, and ncores must be
//   power of 2 (max 16*16 clusters / max 4 cores per cluster)
// _ The array of values to be sorted (ARRAY_LENGTH) must be power of 2 
//   larger than the number of cores.
///////////////////////////////////////////////////////////////////////////////

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <pthread.h>
#include <almosmkh.h>
#include <hal_macros.h>

#define ARRAY_LENGTH        4096       // number of items
#define MAX_THREADS         1024       // 16 * 16 * 4

#define USE_DQT_BARRIER     1          // use DQT barrier if non zero
#define DISPLAY_ARRAY       0          // display items values before and after
#define DEBUG_MAIN          1          // trace main function
#define DEBUG_SORT          1          // trace sort function
#define INTERACTIVE_MODE    0          // activate idbg() during instrumentation
#define CHECK_RESULT        0          // for debug
#define INSTRUMENTATION     1          // register computation times on file
#define IDBG                0          // activate interactive debug in main

/////////////////////////////////////////////////////////////
// argument for the sort() function (one thread per core)
/////////////////////////////////////////////////////////////

typedef struct
{
    unsigned int threads;       // total number of threads
    unsigned int thread_uid;    // thread user index (0 to threads -1)
    unsigned int main_uid;      // main thread user index
}
args_t;

//////////////////////////////////////////
//      Global variables
//////////////////////////////////////////

int                 array0[ARRAY_LENGTH];    // values to sort
int                 array1[ARRAY_LENGTH];    

pthread_barrier_t   barrier;                 // synchronisation variables

pthread_attr_t      attr[MAX_THREADS];       // thread attributes (one per thread)
args_t              arg[MAX_THREADS];        // sort function arguments (one per thread)

////////////////////////////////////
static void bubbleSort( int * array,
                        unsigned int length,
                        unsigned int init_pos )
{
    unsigned int i;
    unsigned int j;
    int          aux;

    for(i = 0; i < length; i++)
    {
        for(j = init_pos; j < (init_pos + length - i - 1); j++)
        {
            if(array[j] > array[j + 1])
            {
                aux          = array[j + 1];
                array[j + 1] = array[j];
                array[j]     = aux;
            }
        }
    }
}  // end bubbleSort()


///////////////////////////////////
static void merge( const int * src,               // source array
                   int       * dst,               // destination array
                   int         length,            // number of items in a subset
                   int         init_pos_src_a,    // index first item in src subset A
                   int         init_pos_src_b,    // index first item in src subset B
                   int         init_pos_dst )     // index first item in destination
{
    int i;
    int j;
    int k;

    i = 0;
    j = 0;
    k = init_pos_dst;

    while((i < length) || (j < length))
    {
        if((i < length) && (j < length))
        {
            if(src[init_pos_src_a + i] < src[init_pos_src_b + j])
            {
                dst[k++] = src[init_pos_src_a + i];
                i++;
            }
            else
            {
                dst[k++] = src[init_pos_src_b + j];
                j++;
            }
        }
        else if(i < length)
        {
            dst[k++] = src[init_pos_src_a + i];
            i++;
        }
        else
        {
            dst[k++] = src[init_pos_src_b + j];
            j++;
        }
    }
}  // end merge()

//////////////////////////////////////
static void sort( const args_t * ptr )
{
    unsigned int       i;
    unsigned long long cycle;
    unsigned int       cxy;
    unsigned int       lid;

    int              * src_array  = NULL;
    int              * dst_array  = NULL;

    // get core coordinates an date
    get_core( &cxy , &lid );
    get_cycle( &cycle );

    unsigned int  thread_uid = ptr->thread_uid;
    unsigned int  threads    = ptr->threads;
    unsigned int  main_uid   = ptr->main_uid;

#if DISPLAY_ARRAY
unsigned int n;
if( thread_uid == main_uid )
{
    printf("\n*** array before sort\n");
    for( n=0; n<ARRAY_LENGTH; n++) printf("array[%d] = %d\n", n , array0[n] );
}
#endif

    /////////////////////////////////
    pthread_barrier_wait( &barrier ); 

#if DEBUG_SORT
if( thread_uid == 0 )
printf("\n[sort] thread[%d] exit barrier 0\n", thread_uid );
#endif

    unsigned int  items      = ARRAY_LENGTH / threads;
    unsigned int  stages     = __builtin_ctz( threads ) + 1;

#if DEBUG_SORT
if( thread_uid == 0 )
printf("\n[sort] thread[%d] : start\n", thread_uid );
#endif

    bubbleSort( array0, items, items * thread_uid );

#if DEBUG_SORT
if( thread_uid == 0 )
printf("\n[sort] thread[%d] : stage 0 completed\n", thread_uid );
#endif

    /////////////////////////////////
    pthread_barrier_wait( &barrier ); 

#if DEBUG_SORT
if( thread_uid == 0 )
printf("\n[sort] thread[%d] exit barrier 0\n", thread_uid );
#endif

#if DISPLAY_ARRAY
if( thread_uid == main_uid )
{
    printf("\n*** array after bubble sort\n");
    for( n=0; n<ARRAY_LENGTH; n++) printf("array[%d] = %d\n", n , array0[n] );
}
#endif

    // the number of threads contributing to sort is divided by 2
    // and the number of items is multiplied by 2 at each next stage
    for ( i = 1 ; i < stages ; i++ )
    {
        if((i % 2) == 1)               // odd stage 
        {
            src_array = array0;
            dst_array = array1;
        }
        else                           // even stage
        {
            src_array = array1;
            dst_array = array0;
        }

        if( (thread_uid & ((1<<i)-1)) == 0 )
        {

#if DEBUG_SORT
if( thread_uid == 0 )
printf("\n[sort] thread[%d] : stage %d start\n", thread_uid , i );
#endif
            merge( src_array, 
                   dst_array,
                   items << (i-1),
                   items * thread_uid,
                   items * (thread_uid + (1 << (i-1))),
                   items * thread_uid );

#if DEBUG_SORT
if( thread_uid == 0 )
printf("\n[sort] thread[%d] : stage %d completed\n", thread_uid , i );
#endif
        }

        /////////////////////////////////
        pthread_barrier_wait( &barrier );

#if DEBUG_SORT
if( thread_uid == 0 )
printf("\n[sort] thread[%d] exit barrier %d\n", thread_uid , i );
#endif

#if DISPLAY_ARRAY
if( thread_uid == main_uid )
{
    printf("\n*** array after merge %d\n", i );
    for( n=0; n<ARRAY_LENGTH; n++) printf("array[%d] = %d\n", n , dst_array[n] );
}
#endif

    }  // en for stages 

    // all threads but the main thread exit
    if( thread_uid != main_uid ) pthread_exit( NULL );

} // end sort()


/////////////////
void main( void )
{
    int                    error;
    unsigned int           x_size;             // number of rows
    unsigned int           y_size;             // number of columns
    unsigned int           ncores;             // number of cores per cluster
    unsigned int           total_threads;      // total number of threads
    unsigned int           thread_uid;         // user defined thread index
    unsigned int           main_cxy;           // cluster identifier for main
    unsigned int           main_x;             // X coordinate for main thread
    unsigned int           main_y;             // Y coordinate for main thread
    unsigned int           main_lid;           // core local index for main thread
    unsigned int           main_uid;           // thread user index for main thread
    unsigned int           x;                  // X coordinate for a thread
    unsigned int           y;                  // Y coordinate for a thread
    unsigned int           lid;                // core local index for a thread
    unsigned int           n;                  // index in array to sort
    pthread_t              trdid;              // kernel allocated thread index (unused)
    pthread_barrierattr_t  barrier_attr;       // barrier attributes

    unsigned long long     start_cycle;
    unsigned long long     seq_end_cycle;
    unsigned long long     para_end_cycle;

    /////////////////////////
    get_cycle( &start_cycle );
 
    // compute number of threads (one thread per core)
    get_config( &x_size , &y_size , &ncores );
    total_threads = x_size * y_size * ncores;

    // get core coordinates and user index for the main thread
    get_core( &main_cxy , & main_lid );
    main_x   = HAL_X_FROM_CXY( main_cxy );
    main_y   = HAL_Y_FROM_CXY( main_cxy );
    main_uid = (((main_x * y_size) + main_y) * ncores) + main_lid; 

    // checks number of threads
    if ( (total_threads != 1)   && (total_threads != 2)   && (total_threads != 4)   && 
         (total_threads != 8)   && (total_threads != 16 ) && (total_threads != 32)  && 
         (total_threads != 64)  && (total_threads != 128) && (total_threads != 256) && 
         (total_threads != 512) && (total_threads != 1024) )
    {
        printf("\n[sort error] number of cores must be power of 2\n");
        exit( 0 );
    }

    // check array size
    if ( ARRAY_LENGTH % total_threads) 
    {
        printf("\n[sort error] array size must be multiple of number of threads\n");
        exit( 0 );
    }

    printf("\n[sort] main starts / %d threads / %d items / pid %x / cycle %d\n",
    total_threads, ARRAY_LENGTH, getpid(), (unsigned int)start_cycle );

    // initialize barrier
    if( USE_DQT_BARRIER )
    {
        barrier_attr.x_size   = x_size; 
        barrier_attr.y_size   = y_size;
        barrier_attr.nthreads = ncores;
        error = pthread_barrier_init( &barrier, &barrier_attr , total_threads );
    }
    else // use SIMPLE_BARRIER
    {
        error = pthread_barrier_init( &barrier, NULL , total_threads );
    }

    if( error )
    {
        printf("\n[sort error] cannot initialise barrier\n" );
        exit( 0 );
    }

#if DEBUG_MAIN
if( USE_DQT_BARRIER ) printf("\n[sort] main completes DQT barrier init\n");
else                  printf("\n[sort] main completes simple barrier init\n");
#endif

    // Array to sort initialization
    for ( n = 0 ; n < ARRAY_LENGTH ; n++ )
    {
        array0[n] = ARRAY_LENGTH - n - 1;
    }

#if DEBUG_MAIN
printf("\n[sort] main completes array init\n");
#endif

    // launch other threads to execute sort() function
    // on cores other than the core running the main thread
    for ( x=0 ; x<x_size ; x++ )
    {
        for ( y=0 ; y<y_size ; y++ )
        {
            for ( lid=0 ; lid<ncores ; lid++ )
            {
                thread_uid = (((x * y_size) + y) * ncores) + lid;

                // set sort arguments for all threads
                arg[thread_uid].threads      = total_threads;
                arg[thread_uid].thread_uid   = thread_uid;
                arg[thread_uid].main_uid     = main_uid;

                // set thread attributes for all threads
                attr[thread_uid].attributes = PT_ATTR_DETACH          |
                                              PT_ATTR_CLUSTER_DEFINED | 
                                              PT_ATTR_CORE_DEFINED;
                attr[thread_uid].cxy        = HAL_CXY_FROM_XY( x , y );
                attr[thread_uid].lid        = lid;

                if( thread_uid != main_uid )
                {
                    if ( pthread_create( &trdid,              // not used because no join
                                         &attr[thread_uid],   // thread attributes 
                                         &sort,               // entry function 
                                         &arg[thread_uid] ) ) // sort arguments
                    {
                        printf("\n[sort error] main cannot create thread %x \n", thread_uid );
                        exit( 0 );
                    }
                    else
                    {
#if DEBUG_MAIN
printf("\n[sort] main created thread %x \n", thread_uid );
#endif
                    }
                }
            }
        }
    }
    
    ///////////////////////////
    get_cycle( &seq_end_cycle );

#if DEBUG_MAIN
printf("\n[sort] main completes sequencial init at cycle %d\n",
(unsigned int)seq_end_cycle );
#endif

#if INTERACTIVE_MODE
idbg();
#endif
    
    // the main thread run also the sort() function
    sort( &arg[main_uid] );

    ////////////////////////////
    get_cycle( &para_end_cycle );

    printf("\n[sort] main completes parallel sort at cycle %d\n", 
    (unsigned int)para_end_cycle );

    // destroy barrier
    pthread_barrier_destroy( &barrier );

#if INTERACTIVE_MODE
idbg();
#endif

#if CHECK_RESULT
    int    success = 1;
    int *  res_array = ( (total_threads ==   2) ||
                         (total_threads ==   8) || 
                         (total_threads ==  32) || 
                         (total_threads == 128) || 
                         (total_threads == 512) ) ? array1 : array0;

    for( n=0 ; n<(ARRAY_LENGTH-2) ; n++ )
    {
        if ( res_array[n] > res_array[n+1] )
        {
            printf("\n[sort] array[%d] = %d > array[%d] = %d\n",
            n , res_array[n] , n+1 , res_array[n+1] );
            success = 0;
            break;
        }
    }

    if ( success ) printf("\n[sort] success\n");
    else           printf("\n[sort] failure\n");
#endif

#if INSTRUMENTATION
    char               name[64];
    char               path[128];
    unsigned long long instru_cycle;

    // build file name 
    if( USE_DQT_BARRIER )
    snprintf( name , 64 , "sort_dqt_%d_%d_%d", ARRAY_LENGTH, x_size * y_size, ncores );
    else
    snprintf( name , 64 , "sort_smp_%d_%d_%d", ARRAY_LENGTH, x_size * y_size, ncores );

    // build file pathname
    snprintf( path , 128 , "home/%s" , name );

    // compute results
    unsigned int sequencial = (unsigned int)(seq_end_cycle - start_cycle);
    unsigned int parallel   = (unsigned int)(para_end_cycle - seq_end_cycle);

    // display results on process terminal
    printf("\n----- %s -----\n"
           " - sequencial : %d cycles\n"
           " - parallel   : %d cycles\n", 
           name, sequencial, parallel );

    // open file
    get_cycle( &instru_cycle );
    FILE * stream = fopen( path , NULL );

    if( stream == NULL )
    {
        printf("\n[sort error] cannot open instrumentation file <%s>\n", path );
        exit(0);
    }

    printf("\n[sort] file <%s> open at cycle %d\n", path, (unsigned int)instru_cycle );

#if IDBG
idbg();
#endif

    // register results to file
    get_cycle( &instru_cycle );
    int ret = fprintf( stream , "\n----- %s -----\n"
                                " - sequencial : %d cycles\n"
                                " - parallel   : %d cycles\n", name, sequencial, parallel );
    if( ret < 0 )
    {
        printf("\n[sort error] cannot write to instrumentation file <%s>\n", path );
        exit(0);
    }

    printf("\n[sort] file <%s> written at cycle %d\n", path, (unsigned int)instru_cycle );

#if IDBG
idbg();
#endif

    // close instrumentation file
    get_cycle( &instru_cycle );
    ret = fclose( stream );

    if( ret )
    {
        printf("\n[sort error] cannot close instrumentation file <%s>\n", path );
        exit(0);
    }

    printf("\n[sort] file <%s> closed at cycle %d\n", path, (unsigned int)instru_cycle );

#endif

    exit( 0 );

}  // end main()

/* 
vim: tabstop=4 : shiftwidth=4 : expandtab
*/