Collapsing Multiple parallel_for's

[tony_az]

I have the following pseudo-code that I'd like to optimize:

for (int i = 0; i < outerLoopCount; i++ {

parallel_for(blocked_range(0, count0, grainsize0),

*parallelForClass0);

SerialFunction0();

parallel_for(blocked_range(0, count1, grainsize1),

*parallelForClass1);

SerialFunction1();

}

The outer loop must remain serial. I'd like to combine the 2 parallel_for's into 1 call, to minimize the overhead of setting up the two parallel_for's on every iteration of the outer loop.

o Is it possible to combine the two parallel_for's so that they are only called once per outer loop iteration?

o Is it possible to combine all of the loops so that initialization is performed only once for the whole loop structure (outer loop included)?

[RafSchietekat]

You seem to be asking for the impossible, and it's not clear what overhead you suspect and why.

Have you tried using affiniity_partitioner with both loops (create instances outside of serial loop, use across serial-loop iterations, with each parallel_for having a separate partitioner)?