Tpetra::CrsMatrixMultiplyOp assigns MultiVectors; should deep_copy instead
Created by: mhoemmen
@trilinos/tpetra @vbrunini
We fixed this in Tpetra::CrsMatrix::applyNonTranspose
a while back; now we need to fix this in CrsMatrixMultiplyOp.
Motivation and Context
An application needs to do mixed-precision (float matrix, double vectors) sparse matrix-vector multiply with Tpetra objects. They roll that themselves now, but it's slow, because it reallocates temporary vectors on each call. CrsMatrixMultiplyOp optimizes that.