インテル® Xeon Phi™ プロセッサー/コプロセッサーの実行性能の比較

#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>
#include "mkl.h"

void do_work(double C[], double A[], double B[], int N)
{
    int i, j, k;
    double alpha = 1.0, beta = 1.0;

#pragma omp parallel for
    for( i=0; i<N; i++ ){
      for( j=0; j<N; j++ ){
        C[i*N + j] = 0.0;
        A[i*N + j] = j&(1<<10) - 512;
        B[i*N + j] = (j%2) ? 1.0 : -1.0;
      }
    }
    cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
        N, N, N, alpha, A, N, B, N, beta, C, N);
}

int main(void)
{
    long i;
    int N = 6000;　//Matrix Size
    double *A, *B, *C;
    struct timeval start, end;
    long mtime, seconds, useconds;

    A = (double *)mkl_malloc(sizeof(double)*N*N, 64);
    B = (double *)mkl_malloc(sizeof(double)*N*N, 64);
    C = (double *)mkl_malloc(sizeof(double)*N*N, 64);

    printf("\nMatrix size: [%d,%d]\n\n",N,N);
     
    for( i=0; i<6; i++)  {
      gettimeofday(&start, NULL);
      do_work(C, A, B, N);
      gettimeofday(&end, NULL);

      seconds  = end.tv_sec  - start.tv_sec;
      useconds = end.tv_usec - start.tv_usec;
      mtime = ((seconds) * 1000 + useconds/1000.0) + 0.5;
      printf("Time taken to run Matmul: %ld 
        milliseconds (%lf)\n", mtime, C[N*N-1]);

      fflush(stdout);
    }
    mkl_free(A);
    mkl_free(B);
    mkl_free(C);
    return 0;
}

import numpy as np 
import time  
N = 6000

print ('Initializing...')

A = np.matrix(np.random.random((N, N)), dtype=np.double) 
B = np.matrix(np.random.random((N, N)), dtype=np.double)

print ('\nMatrix size: [{0}, {1}]\n'.format(N, N))
   
for i in range(1,6): 
    start = time.time()
    C = np.dot(A, B)
    end = time.time()

    tm = int((end - start) * 1000)
    print ('Run numpy.dot {0} milliseconds'.format(tm))

#include<stdio.h>
#include<stdlib.h>
#include <sys/time.h>
#include <malloc.h>
#include <hbwmalloc.h>

#include "mkl.h"

#define SIZE 15000

main()
{
    long i, j;
    double *A;
    double *x, *y;

    long m = SIZE, n = SIZE;
    int ONE = 1;
    double alpha = 1, beta = 1;

    struct timeval start, end;
    long mtime, seconds, useconds;

    int retA = posix_memalign((void*) &A, 64, sizeof(double)*(n * m));
    int retB = posix_memalign((void*) &x, 64, sizeof(double)*n);
    int retC = posix_memalign((void*) &y, 64, sizeof(double)*n);

    printf("\nMatrix size\n");
    printf("A : [%d,%d]\n",n,m);
    printf("x : [%d]\n",n); 
    printf("y : [%d]\n",n);
    printf("\n");

    gettimeofday(&start, NULL);

#pragma omp parallel for
    for (i = 0; i < m; i++){
        x[i] = 1.0;
        y[i] = 1.0;
        for (j = 0; j < n; j++) {
            A[i*n + j] = j%4 ;
        }
    }

    cblas_dgemv(CblasRowMajor,CblasNoTrans, 
        m, n, alpha, A, m, x, ONE, beta, y, ONE);

    gettimeofday(&end, NULL);

    seconds = end.tv_sec - start.tv_sec;
    useconds = end.tv_usec - start.tv_usec;
    mtime = ((seconds) * 1000 + useconds / 1000.0) + 0.5;

    printf("Time taken to run dgemv: %ld milliseconds\n", mtime);
    for (j = 0; j<5; j++) {
        printf("%f ", y[j]);
    }
    printf("\n");

    free(A);
    free(x);
    free(y);

    printf("\n");
}