| 1 |
Hi gang. |
| 2 |
|
| 3 |
Courtezy of Derek Dolney (not a dev) we have an improvement proposed towards |
| 4 |
some of the sci libs. But first a bit of explanation. |
| 5 |
(#30453 has all the details) |
| 6 |
|
| 7 |
BLAS is a Basic Linear Algebra System - is a famous and really basic library. |
| 8 |
Not surprisingly, there are a couple of implementations - first a reference |
| 9 |
one, completely in fortran and called just that - blas. Another often used |
| 10 |
one is atlas (and there are more, for example gsl comes with cblas library as |
| 11 |
part of it). |
| 12 |
|
| 13 |
These libraries can be substituted for one another in majority of cases. As |
| 14 |
such it would be nice to have a virtual/blas and correspondingly blas-config |
| 15 |
utility. |
| 16 |
|
| 17 |
Lacpack is another library (Linear Algebra PACKage) that provides more usefull |
| 18 |
features, It depends on blas (either blas or atlas will do), so is one of the |
| 19 |
would be users of virtual/blas and also has a reference and optimized (atlas |
| 20 |
based) implementation. It is also usefull to be able to have both installed |
| 21 |
together (to do chacks of validity of optimized runs for example). As such |
| 22 |
this would require another virtual: virtual/lapack. |
| 23 |
|
| 24 |
After discussing this with Derek we converged on the following naming scheme |
| 25 |
(does not require approval really, but may be interesting for blas/lapack |
| 26 |
users): |
| 27 |
blas-reference and blas-atlas, plus lapack-reference and lapack-atlas |
| 28 |
which should make sense for all blas/lapack users (now is your time to object |
| 29 |
if it doesn't ;)). |
| 30 |
|
| 31 |
Now on to the last part: ifc use flag: |
| 32 |
So, to complicate things a bit more ;), these libraries (their fortran parts) |
| 33 |
can be compiled with either g77 of ifc (Intel fortran compiler - unrestricted |
| 34 |
download btw, as opposed to icc). However, as these are quite different apps |
| 35 |
name and ionvocation wise, yet another virtual is inappropriate here (and |
| 36 |
wouldn't belong to sci herd legacy anyway), but a use flag is just right. |
| 37 |
|
| 38 |
So, hereby I am proposing above described extensions to blas-related lisb that |
| 39 |
require two new virtuals: blas and lapack and new use flag ifc. |
| 40 |
|
| 41 |
George |
| 42 |
|
| 43 |
|
| 44 |
|
| 45 |
-- |
| 46 |
gentoo-dev@g.o mailing list |
Archives