XeonPhi上でOpenMPを使って行列積

micmatrix.c

#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <omp.h>

int main( void )
{
    // l * m行列とm * n行列積
    int i;
    int j;
    int k;
    int l;
    int m;
    int n;
    double start_time;
    double stop_time;

    float *matA;
    float *matB;
    float *matAns;

    // 初期設定
    l = 10000;
    m = 10000;
    n = 10000;

    // 行列確保
    matA   = (float *)malloc( sizeof(float) * l * m );
    matB   = (float *)malloc( sizeof(float) * m * n );
    matAns = (float *)malloc( sizeof(float) * l * n ); 

    // 行列の値初期化
    for ( i = 0; i < l; i++ ) {
        for ( j = 0; j < m; j++ ) {
            matA[i * m + j] = (float)rand() / RAND_MAX;
        }
    }
    
    for ( i = 0; i < m; i++ ) {
        for ( j = 0; j < n; j++ ) {
            matB[i * n + j] = (float)rand() / RAND_MAX;
        }
    }
    
    for ( i = 0; i < l; i++ ) {
        for ( j = 0; j < n; j++ ) {
            matAns[i * n + j] = 0;
        }
    }

    // 時間測定開始
    start_time = omp_get_wtime();

    // 行列積
    
    // in/outの指定でデバイス側へ/からのコピーを指示する
    #pragma offload target( mic ) \
        in( matA : length( l * m ) ) \
        in( matB : length( m * n ) ) \
        inout( matAns : length( l * n ) )
    {
        #pragma omp parallel for private( j, k ) shared( matA, matB, matAns, l, m, n )
        for ( i = 0; i < l; i++ ) {
            for ( j = 0; j < n; j++ ) {
                for ( k = 0; k < m; k++ ) {
                    matAns[i * m + j] += matA[i * m + k] * matB[k * n + j];
                }
            }
        }
    }
    
    // 終了
    stop_time = omp_get_wtime();
    
    // 時間出力
    printf( "elapsed %f[sec]\n", stop_time - start_time );
    
    // 出力
    /*
    puts( "matrix A" );
    for ( i = 0; i < l; i++ ) {
        for ( j = 0; j < m; j++ ) {
            printf( "%3.3e ", matA[i * m + j] );
        }
        
        puts( "" );
    }
    
    puts( "matrix B" );
    for ( i = 0; i < m; i++ ) {
        for ( j = 0; j < n; j++ ) {
            printf( "%3.3e ", matB[i * n + j] );
        }
        
        puts( "" );
    }
    
    puts( "matrix Ans" );
    for ( i = 0; i < l; i++ ) {
        for ( j = 0; j < n; j++ ) {
            printf( "%3.3e ", matAns[i * n + j] );
        }
        
        puts( "" );
    }
    */

    return 0;
}