hipLaunchKernelGGL can launch kernel multiple times

[jszuppe]

hipLaunchKernelGGL can launch the same kernel multiple times. Example: multiple_kernel_runs.zip

When the situation occurs:

1. There is a header library bar with bar::bar_impl() function that launches a kernel.
2. bar::bar_impl is used in another header library foo to implement FooClass::foo().
3. There is dynamic lib foobar with foobar() func which calls FooClass::foo() function.
4. Main application both uses foobar() and directly FooClass::foo(). Only one call to those functions launches the kernel. In that call instead of one launch, kernel is run twice.

We suspect that probably in here should be a break after hipModuleLaunchKernel().

See ROCm/rocPRIM#13

[AlexVlx]

Thank you for reporting this (and for providing a repro); orthogonally to this issue, the current code has a logical issue (my fault) in that it'd needlessly iterate through the vector even after having found a valid kernel.

[mangupta]

#575 has been merged. Closing this issue.