Explicit Template Instantiations about datatransferkit

[sslattery]

Trying this for classes with partial specializations but can't seem to get it right.

[sslattery]

Assigning this to Damien. We can work on this together to speed up recompilation

[sslattery]

Ross Bartlett has add Trilinos ETI to the DTK configure. This has dramatically improved recompilation as we no longer need to rebuild all of the Tpetra objects.

[sslattery]

Remove all template types in classic unit tests that don't match internal types.

[sslattery]

We should have an ETI configure option that automatically turns on the right Tpetra instantiations. Several people have emailed with build errors regarding this.

Also need to revisit point cloud operator ETI.

[sslattery]

Looking at this first for point cloud operators as part of speeding up builds. Here is the first time on my laptop on 4 corse for building just DTK with the classic adapters (no other mesh adapters): 9m13.835s

Will report times after ETI added to point cloud operators

[sslattery]

Added ETI for a lot of point cloud and build times increased a bunch: 15m3.224s

What happened?

[dalg24]

weird

[Rombur]

Is it possible that you instantiate too many templates? Maybe you are instantiating templates that were not created before. Something else that you can do if one file takes a lot longer than the others is to split the instantiations so that several processors can work on it at the same time.

[rppawlo]

Did you remember to hide the definitions when ETI is enabled (except for the .cpp file)? Otherwise you are just adding more builds of the same object. That's the only way I could see the times increasing with ETI.

[sslattery]

I did include the definitions in the CMake header list. Could this be the issue then?

[sslattery]

OK tried turning off more of the ETI and am getting this:

/Users/uy7/software/Trilinos/packages/tpetra/core/src/Tpetra_ConfigDefs.hpp:168:4: error: "Tpetra: No global ordinal types in the set {int, long long, long, unsigned long, unsigned} have been enabled."

error "Tpetra: No global ordinal types in the set {int, long long, long, unsigned long, unsigned} have been enabled."

However, we have unsigned long explicitly set...

[Rombur]

There is something weird with global ordinal and unsigned / unsigned long Try again with long long and see if it works

[sslattery]

I let it turn on LO=int and GO=int on its own. I'm sure that will work and this is just a bug in their default types from that Tpetra_ConfigDefs message. This should have been caught at configuration.

[sslattery]

Looking at Tpetra_ConfigDefs.hpp:164 I don't think that macro is getting set without input from the user configure

[sslattery]

Pull request #48 added ETI to the point cloud operators. We are planning future work for templating the code further and using ETI to manage that. Therefore I think we are done for now but we should review. Going to assign to @dalg24 for the review.