Consider simplifying how dense matrix is accessed in HiOp interface about hiop HOT 4 CLOSED

I understand now how dereferencing those pointers is not portable. We can change to single pointers. Changes are throughout all the code tiers, from interface to linear solver backend, but not complicated at all. @pelesh @ashermancinelli

from hiop.

for the double** M structure you're referring to, the pointer to the (contiguous) block is M[0]

I kinda like double** because allows the user do "natural" double indexing, M[i[j]...

from hiop.

I agree the 'natural' indexing is nice, but when operating on device pointers, we often want to do something like:

rm.copy(JacD[0], Md_->local_buffer(), ns_*nd_*sizeof(double));

where JacD is a ptr to matrix buffer. This won't work however, since we're dereferencing device data on host. These calls will work for now (in UVM), but for using device memory in FY21 I think this will become a nontrivial stumbling block.

from hiop.

addressed by #128

from hiop.