這是我第一次發佈,所以我會爲我的新手錯誤appologise。請原諒,並非所有的變量名都是英文的。我的問題如下:我使用cygwin gcc編譯器工具鏈在Visual Studio 2010和eclipse中使用openMP編寫了代碼。在視覺上,我得到了加速,但在日食中,我減慢了兩倍於序列版本的速度。請有人能解釋我做錯了嗎?簡而言之,我只是模擬從3D矢量數組複製到雙數組以便通過MPI發送時的加速。OpenMP幫助代碼
#include <omp.h>
#include <time.h>
#include <stdio.h>
#include <vector>
const int NUMAR_FORME=10;
const int NUMAR_SECUNDE_SIMULATE=60; //number of buffers
const int dimensiuni_forme[10]={100,200,300,400,500,600,700,800,900,10000}; //size of each buffer
//-------- the buffers, cuurently only worker_buffer and buff is used
std::vector<std::vector<std::vector<double> > > worker_buffer;
std::vector<std::vector<double> > send_buffer,corect;
double **buff;
double **worker_buffer1;
long i,j,k,l;
int flag=0;
int numarator=0; //number of tests runed
clock_t start;
start=clock();
worker_buffer.resize(1);
buff = new double* [2];
int de_scris=0; //this tells me in which buffer to store, nou I alternate buff[0], buff[1], buff[0], buff[1]
worker_buffer[0].resize(NUMAR_SECUNDE_SIMULATE);
for(i=0;i<NUMAR_SECUNDE_SIMULATE;i++)
worker_buffer[0][i].resize(dimensiuni_forme[9]);
while(numarator<60)
{
if(numarator!=0)
delete [] buff[de_scris];
if(numarator!=0)
de_scris=(de_scris+1)%2;
long limita;
limita=NUMAR_SECUNDE_SIMULATE*dimensiuni_forme[9]*3; //3-comes from the fact that I will have a 3D vector structure
buff[de_scris]= new double [limita];
for(i=0;i<NUMAR_SECUNDE_SIMULATE;i++)
{ for(j=0;j<dimensiuni_forme[9];j++)
{
worker_buffer[0][i][j]=(i*dimensiuni_forme[9]+j)*3;
buff[de_scris][(i*dimensiuni_forme[9]+j)*3]=worker_buffer[0][i][j];
buff[de_scris][(i*dimensiuni_forme[9]+j)*3+1]=worker_buffer[0][i][j]+0.5;
buff[de_scris][(i*dimensiuni_forme[9]+j)*3+2]=worker_buffer[0][i][j]+0.75;
}
}
numarator++;
}
start=clock()-start;
printf("TICKS TOTAL %ld \n",start);
bool ad=true;
long nr;
for(i=0;i<NUMAR_SECUNDE_SIMULATE*dimensiuni_forme[9]*3;i++)
{
if(i%3==0)
nr=i;
if(i%3==0 && buff[de_scris][i]!=i)
ad=false;
else
if(i%3==1 &&buff[de_scris][i]!=(nr+0.5))
ad=false;
else
if(i%3==2 && buff[de_scris][i]!=(nr+0.75))
ad=false;
}
if(ad==false)
printf("not correct \n");
start=clock();
numarator=0;
//parallel version
while(numarator<60)
{
if(numarator!=0)
delete [] buff[de_scris];
long index, limita,id;
omp_set_num_threads(2);
if(numarator!=0)
de_scris=(de_scris+1)%2;
limita=NUMAR_SECUNDE_SIMULATE*dimensiuni_forme[9]*3; //3-
buff[de_scris]= new double [limita];
#pragma omp parallel shared(worker_buffer,limita,buff) private(index,id)
{
printf("intram cu %d threaduri \n", omp_get_num_threads());
id=omp_get_thread_num();
//index=id;
for(index=id;(index*3)<limita;index+=omp_get_num_threads())
{
buff[de_scris][index*3]=worker_buffer[0][index/dimensiuni_forme[9]][index%dimensiuni_forme[9]]; //aici va veni send_buff[index].x
buff[de_scris][index*3+1]=buff[de_scris][index*3]+0.5;
buff[de_scris][index*3+2]=buff[de_scris][index*3]+0.75;
}
// index+=omp_get_num_threads();
}//end parallel zone
numarator++;
}
start=clock()-start;
printf("TICKS TOTAL %ld \n",start);
ad=true;
//testing for correctness
for(i=0;i<NUMAR_SECUNDE_SIMULATE*dimensiuni_forme[9]*3;i++)
{
if(i%3==0)
nr=i;
if(i%3==0 && buff[de_scris][i]!=i)
ad=false;
else
if(i%3==1 &&buff[de_scris][i]!=(nr+0.5))
ad=false;
else
if(i%3==2 && buff[de_scris][i]!=(nr+0.75))
ad=false;
}
if(ad==false)
printf("not correct \n");
return 0;
}
你測量的時間只能用在你的代碼的時鐘功能?如果是這樣,你應該知道時鐘返回使用的CPU時間,這意味着如果你使用2個驅動程序,時鐘()返回的時間將增加兩倍的實時。如果visucal C++以同樣的方式實現clock()函數,我不是100%確定的,可能是你測量不同時間的原因。 – Haatschii 2012-03-24 11:36:59
我試過使用time()和clock_gettime()的結果是一樣的 – Ray 2012-03-25 07:43:25