concurrent memory writes

[Eri_Rubin]

Yes well some times you have to do this, i have to do it 300 times in my solver. And since the alternative is launching 300 kernels, for a total compute time of less then 8ms any alternative that would cost less would be welcome. even if it removes the thread interleaving at that specific point.

Even if you could do a global sync, among all threads it would in effect serialize all ~10,000 threads you are running and remove any benefit from thread interleaving of the data-parallel architecture.

[snapback]377955[/snapback]