Trying to know why the OpenMP code does not parallelise

I just started learning how to use OpenMP. I am trying to figure out why the following code does not run in parallel with Visual Studio 2008. It compiles and runs fine. However it uses only one core on my quad core machine. This is part of the code that I am trying to port to a MATLAB mex function. Any pointer is appreciated.

#pragma omp parallel for default(shared) private(dz, t, v, ts_count) reduction(+: sum_v)
for(t = 0; t<T; t++)
{
    dz = aRNG->randn();
    v += mrdt* (tv - v) +
         vv_v_sqrt_dt * dz +
         vv_vv_v_dt*(dz*dz - 1.);

    sum_v += v;
    if(t == ts_count-1)
    {
        int_v->at_w(k++) = sum_v/(double)(t+1);
        ts_count += ts;
    }
}

---

The v variable is computed using the v value of the previous iteration

  for(t = 0; t<T; t++) {
     ...
     v += ... ( tv - v ) ....
     ...
  }

You cannot do that, it breaks the parallelism. The loop must be able to be run in any sequence, or with differents parallel chunks at once, with no side effects. From a first glance, it doesnt look like you can parallelize this kind of loop.

---

I'm not too sure, it's been a long time since I used OpenMP, but you have the loop variable t set as private. Is that what you want? Isn't that the parallelization variable?

---

One possibility is that you are using the "sum_v" variable. Since you are performing a reduction, the runtime probably expects to only accumulate values in there, and access it 'normally' only after the loop is done.

链接地址: http://www.djcxy.com/p/44134.html

上一篇: 在django和unicode中使用python日志时出现问题

下一篇: 试图了解为什么OpenMP代码不平行