Conditional parallelization

Conditional parallelization

Ritratto di Seunghwa Kang
Seunghwa Kang
Ven, 04/01/2013 - 20:30

Hello,

Does TBB provide a mechanism for conditional parallelization?

For example, assume the following code, and I want to parallelize this loop only when loopCount is larger than 1000 (if not, execute sequentially to avoid unnecessary parallelization overhead).

for( int i = 0 ; i < loopCount ; i++ ) { /* do something */ }

I want to do

if( loopCount > 1000 ) { parallel_for(...) { ... } } else { for(...) { ... } }

without replicating the loop body. It will be very nice if I have something like "parallel_for_if" :-)

Thank you very much,

RSS In cima
3 post / 0 new
Ultimo contenuto
Per informazioni complete sulle ottimizzazioni del compilatore, consultare l'Avviso sull'ottimizzazione
Ritratto di Raf Schietekat
Raf Schietekat
Ven, 04/01/2013 - 23:31

That's what grainsize is for (a parameter in the input Range). You want recursive parallelisation down to chunks that can be executed efficiently using serial optimisations like loop unrolling and vectorization, but that should work whether you start out with a number of elements that will never be subdivided or a really large number. Grainsize will be obeyed even with the (default) auto_partitioner, so you can try if it really improves matters (quite likely it won't).

For a more complete picture, grainsize doesn't apply to tbb::parallel_sort, which will in fact check the number of elements and go straight to std::sort for a small-enough collection (or something to that effect), but sorting is inherently quite different from applying a map pattern, and each has to be considered on its own.

In cima
Ritratto di Raf Schietekat
Raf Schietekat
Mar, 12/02/2013 - 13:24

Continued in "parallel_sort".

In cima

Accedere per lasciare un commento.