Boost logo

Ublas :

From: Jesse Perla (jlp400_at_[hidden])
Date: 2008-08-12 10:21:18


Would you say that clapack is the most fully bound for the "basic"
unstructured, dense routines with lots of function overloading? I don't
completely understand all of the complexity between the traits, bindings,
lapack, etc. but I cannot find an implementation for any matrix inverses in
bindings::lapack? (I know... people should try not to take inverses, but
they do all the time...)
I have identified the following core functions that I want to have "easy" to
call for general matrices: transpose, matrix inverse, cholesky,
non-structured eigenvalues, solve non-structured linear system, SV decomp,
 * From what I can tell, the lapack bindings don't support the "getrf/getri"
LAPACK functions. Am I missing something here? I also don't see bindings
to cholesky as well. However I see great bindings to the different types of
eigenvalue problems and linear systems
 * In CLAPACK, I see the cholesky as well as matrix inversion functions.
 * I suppose I could write something like
http://www.crystalclearsoftware.com/cgi-bin/boost_wiki/wiki.pl?LU_Matrix_Inversionto
get through the basic inversion problem, but what about cholesky?
(and
on this note, does anyone have an implementation using just ublas that I can
call as a function: matrix B = InvertMatrix(A) (with all templates correct,
etc.))

Thanks,
Jesse

On Tue, Aug 12, 2008 at 5:52 AM, Thomas Klimpel <Thomas.Klimpel_at_[hidden]
> wrote:

> Hello Jesse,
>
> I'm building und testing [boost][numeric][bindings] on some different
> platforms, with different blas and lapack libraries. This is an ongoing
> effort to improve the regression tests of the [boost][numeric][bindings]
> library that is not yet finished and will not be finished soon.
>
> I have tested on Win32 (XP) with clapack-3.1.1 (
> http://www.netlib.org/clapack/LIB_WINDOWS/prebuilt_libraries_windows.html)
> and MKL 10.0.x. on VC8 and VC9. I have to use the linker option
> "/NODEFAULTLIB:LIBCMT" and "/NODEFAULTLIB:LIBCMTD" in order to avoid link
> problems when I use clapack. I use "mkl_intel_c_dll.lib", "mkl_core_dll.lib"
> and "mkl_sequential_dll.lib" as linker input file in case of MKL. This works
> fine for me and is orders of magnitude faster than clapack.
>
> Testing with ATLAS on win32 is still on my "ToDo" list, but I tested it on
> linux32 and linux64 and it worked fine.
>
> Conclusion: Both clapack and MKL should work fine on Win32 (XP). I can't
> tell whether they will also work fine on Vista.
>
> Question: is "ipo: error #11035: ..." a linker error? Or is it a runtime
> error or even a compile error?
>
>
> > At this point I am willing to accept whatever linear algebra library has
> the easiest installation steps for this environment.
>
> Then you should probably start with clapack (note that the above link
> points to prebuild libraries), but try to move on to something more
> efficient (like MKL) when you feel comfortable enough.
>
> Regards,
> Thomas
>
>
> From: ublas-bounces_at_[hidden] [mailto:ublas-bounces_at_[hidden]]
> On Behalf Of Jesse Perla
> Sent: Dienstag, 12. August 2008 06:48
> To: Boost::ublas Mailing List
> Subject: [ublas] Help with LAPACK or ATLAS bindings to boost
>
> I am a library user who wants to get a working ublas, lapack system up and
> running as easily as possible on windows.
>
> Does anyone have a thorough step-by-step of how to get bindings installed
> and a test project linked? I see lots of issues posted, but no list of
> detailed setup steps other than mentioning that this is possible (though
> noone says that they have linked to intel MKL on windows?)
>
> I am having a hell of a time trying to get the latest bindings linking with
> the lapack and ublas and intel MKL's lapack. I am using Windows Vista with
> Visual Studio 2008 and Intel compiler 10.1.20. The code compiles fine (no
> surprise since it is just the 3-4 lines of test code I believe everyone
> uses), but when linking or setting up libraries, bizarre things happen. The
> latest is: "ipo: error #11035: Fatal error cannot open libmmdd" when I add
> the MKL library folder to the Intel C++'s library folders... Even the
> trusty windows "reboot and reinstall everything to a folder without any
> spaces" trick didn't work.
>
> If noone has a good list of setup/linking/library steps for Intel MKL, can
> they suggest another? Atlas looks scary to me since it talks about using
> cygwin compiling and I can't figure out if I can simply link it back into my
> intel compiler on windows. But if there are binaries available or simple
> setup steps, that would really help me out. At this point I am willing to
> accept whatever linear algebra library has the easiest installation steps
> for this environment.
>
> Sorry if this may be an Intel compiler on windows specific problem, but I
> am at wit's end.
>
> Thanks,
> -Jesse
>