Re: [eigen] GivensQR

[ Thread Index | Date Index | More Archives ]

2009/9/8 Andrea Arteaga <yo.eres@xxxxxxxxx>:
> In principle my Givens decomposition was designed for small-sized square (or
> quasi-square) matrices. Your benchmarks show that this kind of decomposition
> is actually not optimal for big matrices, due to the algorithm's limits. I
> think however that this decomposition could be faster as Householder if the
> code were a little optimized. We can begin that by using the Jacobi module.

ok, I just did that porting.

it needs profiling. I'm using Gael's new makeGivens function which
while very safe, might incur some overhead. in any case from now it'll
be a breeze to adjust.

In the process I've lost a tiny optimization that you made, where o1
and o2 were computed in-place to avoid the copy. We can get that back
if we decide that m_C becomes an array of PlanarRotation's. But
anyway, I highly doubt that it makes any performance difference,
because that spares 2 float copies that are dwarfed by the cost of the

one thing that bugs me is how you copy the matrices Q and R to
"reduced" matrices, and store both reduced and full matrices. It's
very unlikely that the user might need both, and even if he did, he
would only need the full matrix and would deduce the reduced matrix
from it.


Mail converted by MHonArc 2.6.19+