LinearAlgebra patch
Has anyone tried my patch to LinearAlgebra.py and lapack_litemodule.c? I've found major speedups (by a factor of 3-5) in all operations involving an svd or a eigenanalysis of a symmetric matrix. Paul's initial comment was that since it was not self-contained (it required linking external blas and lapack libs) he could not consider putting it in. I have rectified that now by including new versions of f2c_lite.c, blas_lite.c and d(z)lapack_lite.c that support the newer, faster lapack routines.
I would appreciate any reports on how it works (or doesn't) on other platforms - I've tried it in on MacOS X and Solaris.
Grab it from the 'Patches' section of the sourceforge project page.
-Jeff
-- Jeffrey S. Whitaker Phone : (303)497-6313 Meteorologist FAX : (303)497-6449 NOAA/OAR/CDC R/CDC1 Email : jsw@cdc.noaa.gov 325 Broadway Web : www.cdc.noaa.gov/~jsw Boulder, CO, USA 80303-3328 Office : Skaggs Research Cntr 1D-124
Jeff Whitaker wrote:
Paul's initial comment was that since it was not self-contained (it required linking external blas and lapack libs) he could not consider putting it in. I have rectified that now by including new versions of f2c_lite.c, blas_lite.c and d(z)lapack_lite.c that support the newer, faster lapack routines.
I don't know what it would take, but I would LOVE it if NumPy used a highly optimised BLAS wherever possible. I have no idea whether there is a good BLAS available for every major platform that has a licence that allows it to be included with NumPy, but if so, I think it should be included. I know this is more effor than just including a generic one, but having optimum performace "out of the box" would be a great thing.
just my $.02, which is worth about that, as I am in littel position to help...
-Chris
Chris: The ATLAS project (http://www.netlib.org/ATLAS) provides a mechanism for creating optimized BLAS/LAPACK libs on almost any (unixish) platform. The tuning process is very complicated. It would not be feasible to include this in the Numpy distribution - however the mechanism does exist to link these libs when compiling numpy. Perhaps a more realistic goal would be to modify the setup.py script to automatically detect ATLAS and use those libs if they are installed.
-Jeff
On Thu, 30 Aug 2001, Chris Barker wrote:
Jeff Whitaker wrote:
Paul's initial comment was that since it was not self-contained (it required linking external blas and lapack libs) he could not consider putting it in. I have rectified that now by including new versions of f2c_lite.c, blas_lite.c and d(z)lapack_lite.c that support the newer, faster lapack routines.
I don't know what it would take, but I would LOVE it if NumPy used a highly optimised BLAS wherever possible. I have no idea whether there is a good BLAS available for every major platform that has a licence that allows it to be included with NumPy, but if so, I think it should be included. I know this is more effor than just including a generic one, but having optimum performace "out of the box" would be a great thing.
just my $.02, which is worth about that, as I am in littel position to help...
-Chris
I'd love to have Python and Numpy optimized for the Athlon. I will have to wait for gcc3.0 to get absorbed into FreeBSD. (Mandrake linux has a beta that uses it) It turns out that mpg encoding runs at 10x the rate on my Athlon DDR machine compared to my laptop due to the highly Athlon optimized "GoGo" encoding program. Rob.
Chris Barker wrote:
Jeff Whitaker wrote:
Paul's initial comment was that since it was not self-contained (it required linking external blas and lapack libs) he could not consider putting it in. I have rectified that now by including new versions of f2c_lite.c, blas_lite.c and d(z)lapack_lite.c that support the newer, faster lapack routines.
I don't know what it would take, but I would LOVE it if NumPy used a highly optimised BLAS wherever possible. I have no idea whether there is a good BLAS available for every major platform that has a licence that allows it to be included with NumPy, but if so, I think it should be included. I know this is more effor than just including a generic one, but having optimum performace "out of the box" would be a great thing.
just my $.02, which is worth about that, as I am in littel position to help...
-Chris
-- Christopher Barker, Ph.D. ChrisHBarker@home.net --- --- --- http://members.home.net/barkerlohmann ---@@ -----@@ -----@@ ------@@@ ------@@@ ------@@@ Oil Spill Modeling ------ @ ------ @ ------ @ Water Resources Engineering ------- --------- -------- Coastal and Fluvial Hydrodynamics --------------------------------------
Numpy-discussion mailing list Numpy-discussion@lists.sourceforge.net http://lists.sourceforge.net/lists/listinfo/numpy-discussion
participants (3)
-
Chris Barker -
Jeff Whitaker -
Rob