KokkosBlas::dot only uses cuBLAS with CudaSpace, not with CudaUVMSpace

[mhoemmen]

kokkos-kernels/src/impl/tpls/KokkosBlas1_dot_tpl_spec_avail.hpp

Line 104 in 4ee5f3c

KOKKOSBLAS1_DOT_TPL_SPEC_AVAIL_CUBLAS( double, Kokkos::LayoutLeft, Kokkos::CudaSpace)

@kyungjoo-kim @jjellio

[mhoemmen]

See #397 .

[jjellio]

Its worth pointing that dot via kokkos is faster than the the library calls in most cases I've seen. I imagine the reason is that that is effectively a bandwidth race and the extra indirection of a library call introduces a tiny penalty. Perhaps, it would be better to provide CMake options that allow users to override the default choice. And leave the default as calling Kokkos Kernels. Atleast that way it is explicitly stated when Kokkos is used vs TPL, and provides a mechanism to change it.

[mhoemmen]

PLEASE NO MORE CMAKE OPTIONS PLEASE

If KokkosBlas::dot is faster than we should call it. If it's not faster, then we should fix it.