Révision 282 Pi/C/OpenACC/Pi_OpenACC.c

Pi_OpenACC.c (revision 282)
1 1 
//
2 2 
// Estimation of Pi using Monte Carlo exploration process
3 3 
// Cecill v2 Emmanuel QUEMENER <emmanuel.quemener@gmail.com>
4 
// Exploit OpenACC on Nvidia GPU
5 
// module load
6 
// icpc -std=c99 -O3 -o Pi_OpenACC Pi_OpenACC.c -lm
4 
// Tested on GCC-10
5 
// gcc -fopenacc -foffload=nvptx-none -foffload="-O3 -misa=sm_35" Pi_OpenACC.c
7 6 
//
8 7 

  		





  9
  8
  
    
#include <math.h>


  		





  ......




  139
  138
  
    
  LENGTH IterationsEach=((iterations%ParallelRate)==0)?iterations/ParallelRate:iterations/ParallelRate+1;


  		





  140
  139
  
    

  		





  141
  140
  
    
  inside=(LENGTH*)malloc(sizeof(LENGTH)*ParallelRate);


  		





  142
  
  
    
  


  		





  143
  
  
    
#if _OPENACC


  		





  144
  
  
    
  acc_init(acc_device_nvidia);


  		





  145
  
  
    
#endif


  		





  146
  141
  
    

  		





  147
  142
  
    
  gettimeofday(&tv1, &tz);


  		





  148
  
  
    
#pragma omp parallel for shared(ParallelRate,inside)


  		





  149
  
  
    
#pragma acc kernels loop


  		





  150
  
  
    
  for (int i=0 ; i<ParallelRate; i++) {


  		





  151
  
  
    
    inside[i]=MainLoopGlobal(IterationsEach,seed_w+i,seed_z+i);


  		





  152
  
  
    
  }


  		





  153
  143
  
    

  		





  
  144
  
    
#pragma acc kernels copy(inside[0:ParallelRate])


  		





  
  145
  
    
  {


  		





  
  146
  
    
    #pragma acc independant


  		





  
  147
  
    
    for (int i=0 ; i<ParallelRate; i++) {


  		





  
  148
  
    
      inside[i]=MainLoopGlobal(IterationsEach,seed_w+i,seed_z+i);


  		





  
  149
  
    
    }


  		





  
  150
  
    
  }


  		





  154
  151
  
    
  gettimeofday(&tv2, &tz);


  		





  155
  152
  
    
  


  		





  156
  153
  
    
  for (int i=0 ; i<ParallelRate; i++) {


  		





  ......




  163
  160
  
    
			  (tv2.tv_usec-tv1.tv_usec))/1000000;


  		





  164
  161
  
    

  		





  165
  162
  
    
  double itops=(double)(ParallelRate*IterationsEach)/elapsed;


  		





  166
  
  
    
  


  		





  167
  
  
    
  printf("ParallelRate %i\nElapsed Time %.2f\nItops %.0f\n",ParallelRate,elapsed,itops);


  		





  168
  163
  
    

  		





  
  164
  
    
  printf("Inside/Total %ld %ld\nParallelRate %i\nElapsed Time %.2f\nItops %.0f\nLogItops %.2f\n",insides,ParallelRate*IterationsEach,ParallelRate,elapsed,itops,log10(itops));


  		





  
  165
  
    

  		





  169
  166
  
    
  free(inside);


  		





  170
  167
  
    
  return(insides);


  		





  171
  168
  
    
}


  		





  ......




  200
  197
  
    

  		





  201
  198
  
    
  LENGTH total=((iterations%ParallelRate)==0)?iterations:(iterations/ParallelRate+1)*ParallelRate;


  		





  202
  199
  
    

  		





  203
  
  
    
  printf("Inside/Total %ld %ld\nPi estimation %f\n\n",insides,total,(4.*(float)insides/total));


  		





  204
  
  
    
   


  		





  
  200
  
    
  printf("Pi estimation %f\n\n",total,(4.*(float)insides/total));


  		





  
  201
  
    

  		





  
  202
  
    

  		





  205
  203
  
    
}


  		












Formats disponibles :
  Unified diff