my code takes lot of time to execute and returns incorrect result

[ankit_m_]

Hello, 

I am new to programming with MIC cards. I am trying to run a very simple program but it appears that it is taking a long time to offload the data over to the MIC card and also the final output seems to be incorrect, can anyone help me figure out my mistake, please. 

#include <iostream>
#include <memory>
#include "omp.h"
#include <malloc.h>

using namespace std; 

int main()
{
    int xx=100000; 
    int yy=10000;
    
    unsigned long long size = xx*yy; 
    cout << " Simulate Data" << endl; 
    cout << "data size " << size*4 << endl; 
    
    int* aa = (int*) malloc(sizeof(int)*size); 
    for(unsigned long long ii=0; ii < xx*yy; ++ii)
    {
        aa[ii] =1; 
    }
    
    cout << " start offload " << endl; 
    unsigned long long dim = xx*yy; 
    #pragma offload target(mic:0) \
    in(aa:length(dim)) 
    {
        #pragma omp parallel for 
        for (unsigned long long ii; ii < xx*yy; ++ii)
        {
            aa[ii] *= 2; 
        }
    }
    
    cout << " offload end " << endl;
    cout << " Result  " << aa[10] <<"  " << aa[1000] << endl; 
    free(aa);  
    
    return 0;     
}

Thank you

Sincerely, 

AM

[jimdempseyatthecove]

You want to use

#pragma offload target(mic:0) \
    inout(aa:length(dim)) 
   

Note, the first offload has the overhead of transferring the code and initializing the MIC's OpenMP thread pool. Try:

for(int I=0; I<4; ++I) {

cout << " start offload " << endl; 
 double t0 = omp_get_wtime(); 
  unsigned long long dim = xx*yy; 
    #pragma offload target(mic:0) \
    in(aa:length(dim)) 
    {
        #pragma omp parallel for 
        for (unsigned long long ii; ii < xx*yy; ++ii)
        {
            aa[ii] *= 2; 
        }
    }
    double t1 = omp_get_wtime();
    cout << " offload end  " << t1 - t0 << endl;
    cout << " Result  " << aa[10] <<"  " << aa[1000] << endl; 
} // for
 

Jim Dempsey

[ankit_m_]

Thank you very much for your prompt reply Jim. I really appreciate all your help. Now, my program is running correctly however the offload is still too slow. 

[jimdempseyatthecove]

Please note that the code within your offloaded section is trivial.

Read (vector), multiply (vector), write (vector)

That is all it is doing (other than a little loop overhead)

Your offload code should be performing more work to recover the time to pass the data into and out of the MIC.

Choose something like a textbook matrix multiply as a sample code.

Jim Dempsey