[aerosayan]

If you got 8 threads, you use as many threads as possible.
However, OpenMP only allows control on a per-loop basis.


I have this bit of section in my solver,
I calculate rho_inverse, a, a_inverse, PSV from CSV, pressure etc :

    //
    //             --- prepare commonly used data ---
    //

    // calculate rho inverse
    //
    inv_arrays(ri_, un1_, 1.0, ncells);

    // calculate x velocity, y velocity, energy
    //
    mul_arrays(vx_, un2_, ri_, ncells);
    mul_arrays(vy_, un3_, ri_, ncells);
    mul_arrays(e_ , un4_, ri_, ncells);

    // calculate pressure
    //
    calc_pressure(p_, un1_, e_, vx_, vy_, 0.4, ncells);

    // calculate the local speed of sound for all cells at once
    //
    calc_spdsound(a_, p_, ri_, 1.4, ncells);

    // calculate the reciprocal of the local speed of sound for all cells
    //
    inv_arrays(ai_, a_, 1.0, ncells);

However, all of this is serial. If I got 4 threads, I could calculate rho_inverse first serially, then use 3 threads to simultaneously calculate the x,y,e values, then calculate pressure, a, a_inverse serially.

The speedup would come from calculating the x,y,e values in parallel, instead of doing them serially one after another.

This kind of optimization opportunities are present throughout the code. However, OpenMP provides limited control.

There's OpenMP sections, and that's what I'm currently using, but they're very limited.

Using OpenMP sections, I can hardcode instructions to calculate x,y,e in parallel. However, since they're hardcoded, they're not scalable when we have more threads to use, and more equations to solve in parallel.

Will probably keep using OpenMP sections, until I find a better alternative.


If anyone knows of a better solution, kindly let me know.