lrkeefe wrote: I am confused by your statement about there being a big gap between some of the bands. The finite-difference coefficient matrix in this case has no gaps between the diagonals, they are the six diagonals above and below the main diagonal.
Are you confusing storage and the actual matrix? If you have n
points per line in your domain, then the diagonals above the main are at distance 1,2,3,n-1,n,n+1,2n, if I guess correctly the discretization stencil you're using. You may be storing them contiguously, but if you'd make a picture of the matrix, there would be big gaps.
Or maybe I'm misunderstanding you completely.
At this point I would be interested in seeing "research" grade algorithms for banded -matrix matrix multiply. Where is this development work occurring?
I don't think there is much development. Mail Iain Duff (I.Duff@rl.ac.uk
)who was involved in sparse blas, and ask him. Even then, he probably only knows about sparse-times-dense.