space-time-stack: allow demangled names

[romintomasetti]

As promised in #218, this PR extracts the part relevant to demangling names.

1. The demangling function is now shared in common/utils/demangle.hpp.
2. The space time stack tool now demangles the names.

This is how the report looks like with the changes:

9: <average time> <percent of total time> <percent time in Kokkos> <percent MPI imbalance> <remainder> <kernels per second> <number of calls> <name> [type]
9: ===================
9: |-> 1.49e-02 sec 43.8% 100.0% 0.0% ------ 1 Tester<Kokkos::OpenMP, 5ul> [for]
9: |-> 1.48e-02 sec 43.5% 100.0% 0.0% ------ 2 named kernel [for]
9: |-> 4.15e-03 sec 12.2% 100.0% 0.0% ------ 1 Tester<Kokkos::OpenMP, 5ul>/Tester<Kokkos::OpenMP, 5ul>::TagUnnamed [for]

and how it looks without the changes:

<average time> <percent of total time> <percent time in Kokkos> <percent MPI imbalance> <remainder> <kernels per second> <number of calls> <name> [type]
===================
|-> 1.79e-02 sec 59.4% 100.0% 0.0% ------ 2 named kernel [for]
|-> 7.90e-03 sec 26.2% 100.0% 0.0% ------ 1 6TesterIN6Kokkos6OpenMPELm5EE [for]
|-> 4.20e-03 sec 13.9% 100.0% 0.0% ------ 1 6TesterIN6Kokkos6OpenMPELm5EE/N6TesterIN6Kokkos6OpenMPELm5EE10TagUnnamedE [for]

[romintomasetti]

@vlkale @dalg24 @masterleinad

[masterleinad]

Please make sure that the pull request title and description reflect all the changes you make here.

[masterleinad]

You have to fix the indentation as well.

[masterleinad]

Fine with me.

[dalg24]

This is silly, please update find_index and take the trailing argument as a string const ref.

[romintomasetti]

Indeed, and done.

By the way, I think find_index is useless if the variable kernelInfo, which is a std::vector, is changed to being a std::set whose comparator compares the time (see what compareKernelPerformanceInfo does).

I think the code would be much easier to read.

If you agree I can make the change.

[romintomasetti]

@dalg24 I guess you'll want to make the change in another PR 😉

[dalg24]

Please post a before/after tool output snippet in the description.

[romintomasetti]

@dalg24 @masterleinad

I added a test for the space time stack tool that checks how kernels appear in the report.

So I had to modify slightly the code such that Kokkos Tools can also parse unnamed kernels launched with a work tag. See the changes for more details.

[romintomasetti]

@dalg24 @masterleinad Could you please review this one one last time ? 🔍 Thanks!

[dalg24]

Although obviously adding tests in an improvement but you had almost reached consensus and expanding the scope of your PR will delay merging.

[dalg24]

Why would you want to template the whole struct?
Get rid of apply(), move the test to the constructor that you template on ExecutionSpace. Also there is no point on making the number of iteration a template parameter.

[romintomasetti]

• I think I templated the struct over the execution_space type to check what the tool would output when looking at kernels defined in templated classes. If you think it is unnecessary, I can get rid of the execution_space template.
• For the number of iterations, I think I wanted to have a compile-time range, but I guess it is useless for this test. I'll use a range from 0 to 1 now 😉

[vlkale]

@romintomasetti Sorry to not get back sooner - I was away and on vacation + holiday since this time until late last week, and I am still catching up. The idea of demangle.hpp from your initial post and the current code and CMake in your PR makes sense and looks good to me. My top-level comment is whether you want to test this demangling functionality for a couple of other tools, e.g., kernel-logger, simple-kernel-timer? You put this in the common subdirectory, so am I correct that it can be used for those too?

Edit: I see that you have used simple-kernel-timer, so that's great!

Right now, I think
@dalg24 (and @masterleinad) have a better handle on this than me. The PR looks good to me overall. I am good with approving merging this if they are, in order to move things forward.

[masterleinad]

I would prefer to give the functor with a tag a different name. Otherwise, this looks good to me.