Multi-threads slower than single threads? - opencv

Question

I am trying to find out how to speed up a certain part of my code. I've got three float variables: var1, var2, var3.
in sequential mode ...
double start1, end1, t1;
start1= (double)cvGetTickCount();
var1= tester->predict(videocapture, params1, image);
var2= tester->predict(videocapture, params2, image);
var3= tester->predict(videocapture, params3, image);
end1= (double)cvGetTickCount();
t1= (end1-start1)/((double)cvGetTickFrequency()*1000.);
printf( "Time1 = %g ms\n", t1 );
it seems to be slightly faster than parallel threads ...
double start2, end2, t2;
start2= (double)cvGetTickCount();
mp_set_dynamic(0); // Explicitly enable/disable dynamic teams
omp_set_num_threads(3); // Use 3 threads for all consecutive parallel regions
#pragma omp parallel num_threads(3)
{
#pragma omp sections //nowait
{
#pragma omp section
{
#pragma omp critical
{
var1= tester->predict(videocapture, params1, image);
}
}
#pragma omp section
{
#pragma omp critical
{
var2= tester->predict(videocapture, params2, image);
}
}
#pragma omp section
{
#pragma omp critical
{
var3= tester->predict(videocapture, params3, image);
}
}
}
}
}
end2= (double)cvGetTickCount();
t2= (end2-start2)/((double)cvGetTickFrequency()*1000.);
printf( "Time2 = %g ms\n", t2 );
Can someone please help me speed up the process of finding those three variables and tell me what I am doing wrong

Answer 1

Look at the specification of the critical pragma here.
There are three crucial points here:
1) If you don't give the critical pragma a name, it's mapped to some unspecified name, and the same name is used for all unnamed critical sections.
2) Only one thread at a time can be in any critical section with a given name.
3) When a thread encounters a critical section, it waits there until no other threads are active in it before moving into it.
Thus your code tells all your threads to wait while they all take turns executing one item of your code at a time. This is equivalent to doing the operations serially, except that it requires the overhead of generating the threads in the first place (and whatever overhead occurs due to the waiting, which might be busy-waiting).