You say Householder, but I would say you could be "chasing the subdiagonal" with Givens Rotations You are using a full array but you only need three diagonals of R, and if you save Q as n-1 angles you don't even need a dense Q
This begs the question of why you are doing QR on a symmetric tridiagonal? hopefully not a step for eigenvalues because chasing the bulge is the way to go for that. On Tuesday, December 24, 2013 7:55:23 PM UTC-5, Ivar Nesje wrote: > > There is a SubArray implementation, sub(), but it has some performance > issues related to indexing that should be fixed before it becomes default > for slicing. > > For small arrays I would guess that hardcoded multiplication is faster > than calling a BLAS. > > You can find info about you julia by calling versioninfo() > >
