Fix kernel timer: use after free (#2006)

* KernelTimer: Fix Cost Use-After-Free

AMReX' ParallelFor is async, the free of the cost function occurs too
early can can lead to use-after-free's from the kernel.

* CMake: WarpX_GPUCLOCK on CUDA/HIP

Make conditional on supported backends.

* Add addition to global cost array

As we abstracted this from the original research prototypes
(https://doi.org/10.5281/zenodo.4708449) to merge this to mainline in
1406, we forgot to add the addition to `*costs` at this point.

Measurements were done with the code linked in the zenodo archive.

* Cleaning: Move m_cost assignment more in

Not too important, but with a default nullptr now possible.

* synchronize -> streamSynchronize

Co-authored-by: Weiqun Zhang <WeiqunZhang@lbl.gov>

0 files changed, 0 insertions, 0 deletions