Closed sadiqj closed 2 years ago
I think this issue has been addressed for parallel_for
and parallel_for_reduce
.
parallel_scan
could probably still need work.
Addressed for parallel_scan
too now (cf. f8cea3b251080e464eacb83e27cf6c12f397e2e8).
In the case where a task pool has no domains in it, we can have a fast path in the
parallel_*
functions which uses serial constructs and avoids the overhead of queuing/dequeuing tasks.