[Lapack] Proposed interfaces for extra-precise refinement.
From: Jason Riedy
Date: Tue, 18 Apr 2006 13:26:34 -0700
And "Xiaoye S. Li" writes:
- * PARAMS(ITHRESH_PARAM = 3) : Total number of residual computations
- * allowed for refinement.
- * Default: 10 for double-precision, 5 for single-precision
- I think we have decided to remove single-precision (or working-precision)
- refinement. So the default should be 10 always?
Good catch, thanks. So the default should just be 10.
(And to everyone else: We've removed working-precision
refinement from the SGESVXX interface. It's still in
SGESVX (one X).)
- For least squares refinement, it seems we will need two separate
- ERR_BNDS arrays, one for X, another for residual R.
I believe the least-squares interface is still up in the
air. If we return bounds on the residual, it'll suffice
to add two to n_norms and use those slots. These are just
different scalings on the combined [x';r'] vector.
- Also, there are several reciprocal condition numbers floating around:
I think we're waiting on the data to know which actually
hold appropriate, non-redundant information.