Add CudaUVMSpace specializations for cuBLAS IAMAX and SCAL

[vqd8a]

Requested by GEMMA for integrating GEMMA and ADELUS.

[srajama1]

Oh man, why do they need UVM when I am trying to remove it in other places.

[crtrott]

That's for ETI? Because what else would be the specialization?

[vqd8a]

It is for cuBLAS TPL. Kokkos-kernels' IAMAX and SCAL have no CudaUVMSpace specializations when calling cuBLAS TPL. We only have CudaSpace specializations. For example:

kokkos-kernels/src/impl/tpls/KokkosBlas1_iamax_tpl_spec_decl.hpp

Line 524 in ad36ed6

KOKKOSBLAS1_ZIAMAX_TPL_SPEC_DECL_CUBLAS( unsigned long, Kokkos::LayoutLeft, Kokkos::CudaSpace, true)

When CudaUVMSpace is used, the fall-back implementations would be called, which might be slower than cuBLAS.
I did a simple fix in #759 by adding CudaUVMSpace specializations to the TPL files: KokkosBlas1_iamax_tpl_spec_decl.hpp, KokkosBlas1_iamax_tpl_spec_avail.hpp, and KokkosBlas1_scal_tpl_spec_decl.hpp, KokkosBlas1_scal_tpl_spec_avail.hpp

[crtrott]

why isn't this type-erased before it this that (i.e. use Device<Cuda,AnonymousSpace>)?

[vqd8a]

Does it require re-write the whole TPLs? I just followed what was done in #397 and #399 for GEMM for a quick fix.

[vqd8a]

As pointed by @ndellingwood, there is an open issue #144 about AnonymousSpace but it was backlogged. I feel it is somehow beyond the scope of my PR. But should using AnonymousSpace be reprioritized? @srajama1 @crtrott

[crtrott]

yeah its fine not to do it in your PR. but generally it might be worthwhile to look at that.

[vqd8a]

Thanks @crtrott

[vqd8a]

I will look at that.