Tpetra: unnecessary UVM ETI

[jhux2]

Is this usage of UVM a problem? This doesn't appear to be ifdef guarded.

Trilinos/packages/tpetra/core/src/Tpetra_Details_FixedHashTable_Cuda.cpp

Lines 58 to 66 in 004eee5

	typedef Kokkos::Device<Kokkos::Cuda, Kokkos::CudaUVMSpace> cuda_uvm_device_type;
	#define TPETRA_DETAILS_FIXEDHASHTABLE_INSTANT_CUDA( LO, GO ) \
	TPETRA_DETAILS_FIXEDHASHTABLE_INSTANT( LO, GO, cuda_device_type )
	TPETRA_INSTANTIATE_LG( TPETRA_DETAILS_FIXEDHASHTABLE_INSTANT_CUDA )
	#define TPETRA_DETAILS_FIXEDHASHTABLE_INSTANT_CUDA_UVM( LO, GO ) \
	TPETRA_DETAILS_FIXEDHASHTABLE_INSTANT( LO, GO, cuda_uvm_device_type )

@trilinos/tpetra

[jhux2]

This is only instantiating for UVM, so perhaps it's not so bad.

[brian-kelley]

It seems like we never use FixedHashTable with the <Cuda, CudaUVMSpace>, regardless of the Tpetra_ENABLE_CUDA_UVM setting. There's no major harm in instantiating for it, but it adds some library size and build time.

[csiefer2]

Harmless, but should be removed.

[jhux2]

FHT instantiated on `<Cuda, CudaUVM> is needed even to build Tpetra's own tests. I'm closing this as invalid/wontfix.