And I write:
- Appended are the proposed interfaces for extra-precise
- refinement in single precision.
- Still in flux:
- 1) How to query for the workspace size. [...]
- 2) The actual workspace size. [...]
Due to local comments, we're also re-examining:
3) which routines are responsible for calculating
4) exactly how we compute BERR, and
5) exactly how we compute the pivot growth RPVGRW.
The former may change RCOND in SGERFSX to an output
variable. And the latter may change the comments
describing BERR and RPVGRW.
So we're still kicking around the interface a bit.
The general essence will be the same, so comments are
still appreciated. If you're waiting for the exact,
complete, perfect, final proposal, it ain't there yet.
Major pieces that need consideration:
*) Using PARAMS(NPARAMS) to pass optional arguments
and query defaults.
*) How the error bounds and associated quantities
If anyone wants to follow along in code, the new
interface branch is available at:
SGERFSX will have the new interface first, but it's
not entirely in place.
And the separate branch that uses F95 features and
implements a reverse communication / state machine
version is at:
The F95 features are limited to the *.f95 sources.
I've only tested with gfortran, Sun/Fujitsu's f95,
and Intel's ifort.
(My git repos are also available; ping me if you want
to follow those. Yozo maintains the stable, known to