Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
Francis wrote: Thank you for your effort. I guess garnumpy reflects the idea in this Pylab discussion: http://www.scipy.org/PyLab Again I get errors in libblas/lapack related to gfortran (local variable problems). I replaced the libblas.a and the liblaplack.a by the ones of sage. And started make install again. It seems to work until it tries to configure/install umfpack. If you install blas/lapack from sage, it kind of defeats the whole purpose of garnumpy. The goal is to have a unified set of options to build. It is likely that sage uses different options than the ones from garnumpy. If you use garnumpy, you should use it for everything: it is an all for nothing, intended to replace broken distributions/workstations where you don't have admin rights. When I updated a bit garnumpy yesterday, I tested it on CENTOS 5, which is the distribution you are using, right ? xerbla.f:(.text+0xd7): undefined reference to `_g95_stop_blank' collect2: ld returned 1 exit status make[4]: *** [umfpack_di_demo] Error 1 Yes, that's expected: you are using blas from sage, this cannot work. Perhaps it possible to skip the installation of the umfpack since I probably won't need it or it requires the other libblas.a. Just wildly guessing here. garnumpy by default does not install umfpack (scipy only requires blas/lapack, and by default, garnumpy only install what is needed; you can configure it to build umfpack, atlas, whatever, but by default, only the minimal stuff), so I am not sure why you have umfpack errors. cheers, David ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
David Cournapeau wrote: If you install blas/lapack from sage, it kind of defeats the whole purpose of garnumpy. The goal is to have a unified set of options to build. It is likely that sage uses different options than the ones from garnumpy. If you use garnumpy, you should use it for everything: it is an all for nothing should read all or nothing. David ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
On Sep 17, 2:37 am, David Cournapeau [EMAIL PROTECTED] wrote: Francis wrote: To be frank, I am bit of noob, but not completely. I use the gfortran make.inc to build lapack. The make.inc file to build and install atlas is using gfortran. I used the --fcompiler=gfortran flag. I don't think there is much else I can do. Yes there is :) You should use --fcompiler=gnu95, not gfortran. That would explain the problem you had, actually. There is no problem being a noob: compiling numpy itself is easy, but unfortunately, its dependencies are not, and are almost systematically badly packaged by distributions. To make the situation worse, there are still two widely used gnu fortran compilers, which are ABI incompatible... Even a noob can follow simple instructions. :) Tried gnu95 first and then gfortran but both result in the same error. Unfortunately there is no EPD installer for Centos 5 version. There was a blog how to install it, but that look a lot more complicated than what I am trying now. I will try to build sage and look if I can build numpy against the atlas libraries in sage. Otherwise I use sage and remove all the packages that are not relevant for me, e.g. maxima, gap, singular, etc. I am really sorry for all this trouble. The last solution you could try, with a bit old packages is my own repository, with Centos packages: http://download.opensuse.org/repositories/home:/ashigabou/CentOS_5/ Install the numpy and scipy packages from there. There is also a src.rpm for atlas, which enables you to build atlas rpm automatically optimized for your CPU using one command (look at build atlas from ashigabou repository onwww.scipy.org/Installing_Scipy/Linux). I have been trying to update those packages yesterday, but some network-related problems prevented me from. I can't open rpm files so I opened numpy with archive manager and extracted Numpy to site-packages. Unfortunately importing numpy results in ImportError: undefined symbol: _PyUnicodeUCS4_IsWhitespace. Suggested solution found in wiki is to hard-install numpy. Ah ... it is like running in circles. Is it possible to use refblas and lapack libraries in the ashigabou rep. to hard-install numpy. Not sure what the site.cfg should look like than. ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
Francis wrote: I can't open rpm files so I opened numpy with archive manager and extracted Numpy to site-packages. Can't you ask your admin to install rpm packages ? Uncompressing the rpm and putting them into random places is unlikely to work. Unfortunately importing numpy results in ImportError: undefined symbol: _PyUnicodeUCS4_IsWhitespace. Yes, because you built your python yourself. That's why you should not do that: once you built your own python, you are effectively using a different ABI for every python package packaged for your distribution unless you take care of using the exact same options as RH did. You should not do that unless you really know what you're doing. Suggested solution found in wiki is to hard-install numpy. Ah ... it is like running in circles. Is it possible to use refblas and lapack libraries in the ashigabou rep. Yes, as long as you use gfortran to build numpy/scipy. But I would first ask your admin whether it is ok to install the rpms. That's really how packages should be installed, not using the sources directly. cheers, David ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
In general the IT dep. don't want changes since changes require weeks of testing to make sure that existing applications are not conflicting, etc. I am not sure they would fulfill my request. Especially since it has no priority at all. I am installing Sage math now. Although I would like to avoid using it, it is at this moment the simplest solution and it will do what I want it to do. Thank you for trying to help out. On 17 Sep, 10:16, David Cournapeau [EMAIL PROTECTED] wrote: Francis wrote: I can't open rpm files so I opened numpy with archive manager and extracted Numpy to site-packages. Can't you ask your admin to install rpm packages ? Uncompressing the rpm and putting them into random places is unlikely to work. Unfortunately importing numpy results in ImportError: undefined symbol: _PyUnicodeUCS4_IsWhitespace. Yes, because you built your python yourself. That's why you should not do that: once you built your own python, you are effectively using a different ABI for every python package packaged for your distribution unless you take care of using the exact same options as RH did. You should not do that unless you really know what you're doing. Suggested solution found in wiki is to hard-install numpy. Ah ... it is like running in circles. Is it possible to use refblas and lapack libraries in the ashigabou rep. Yes, as long as you use gfortran to build numpy/scipy. But I would first ask your admin whether it is ok to install the rpms. That's really how packages should be installed, not using the sources directly. cheers, David ___ Numpy-discussion mailing list [EMAIL PROTECTED]://projects.scipy.org/mailman/listinfo/numpy-discussion ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
Francis wrote: In general the IT dep. don't want changes since changes require weeks of testing to make sure that existing applications are not conflicting, etc. I feel your pain. In such case, I ask if it is possible to ask for vmware installation: nowadays, cpu are fast enough such as running linux inside linux is fast, and inside your vmware installation, you install your own things. Still, as a numpy/scipy developer, it is quite frustrating that a user have to install sage, a third-party software, to make numpy/scipy work... cheers, David ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
On Wed, Sep 17, 2008 at 7:13 PM, Francis [EMAIL PROTECTED] wrote: In general the IT dep. don't want changes since changes require weeks of testing to make sure that existing applications are not conflicting, etc. I am not sure they would fulfill my request. Especially since it has no priority at all. Ok, this really bothered that you had to install sage to use numpy. You don't have to do it since you already wasted quite some time, but I updated a small project of mine that I almost forgot: garnumpy. It is a set of makefiles to build numpy, scipy and all its dependencies (blas, lapack, fftw, etc...) automatically: - It download the tarballs for you. - It installs everything in a configurable directory, so that it is self-contained It is there; http://github.com/cournape/garnumpy/ You can get the sources with git or just the tarball generated by github if you don't have/don't want git. Basic usage: set up main_prefix in the file garnumpy.conf.mk (or keep the default value: it installs in $HOME/garnumpyinstall) cd platform/scipy make install source startgarnumpy.sh # You now have your env set up for numpy and scipy python -c import scipy If you want to remove it, just rm -rf the path set up in main_prefix. cheers, David ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
Francis Drossaert wrote: Hi everybody, I am trying to install python2.5/scipy/numpy/sympy/matplotlib locally, because of various reasons. I am not root. Do you really need python 2.5 ? By building your own python, you are forcing yourself to build every package you will need for python, including the dependencies. For matplotlib, it will be painful. Python 2.4 is enough (incidentally, I have to use Centos 5 at some place, on the same architecture). For Blas/Lapack, how did you build them ? (Which version, which makefile, which fortran compiler). However after installing I tried to import numpy in python and received to following error: ImportError: /users/francisd/local/lib/python2.5/site-packages/numpy/linalg/lapack_lite.so: undefined symbol: zgesdd_ What lapack_lite.so does depend on ? (ldd /users/francisd/local/lib/python2.5/site-packages/numpy/linalg/lapack_lite.so) Googling this error, it seems that the error is caused by having a wrong lapack version. I am running Centos 5 (basically Red Hat Enterprise Linux 5) on a x86_64 machine. Unfortunately, it can be many things. Blas/Lapack/Atlas are difficult to build by yourself, there are tens of way to screw up at any point. If you give us the above information, it should be clearer where exactly an error was made cheers, David ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
Do you really need python 2.5 ? By building your own python, you are forcing yourself to build every package you will need for python, including the dependencies. For matplotlib, it will be painful. Python 2.4 is enough (incidentally, I have to use Centos 5 at some place, on the same architecture). I don't really need Python2.5. Python2.4 is actually available on the network but I think I will run into problems using the network one and try using local libraries. I assume one would need to change the paths on a file which I can't change. Anyway I installed Python2.5 without any problems. I guess the problem is combining numpy/scipy with the lapack/blas/atlas package. For Blas/Lapack, how did you build them ? (Which version, which makefile, which fortran compiler). I forgot to include the link: http://www.scipy.org/Installing_SciPy/. I basically followed the instructions mentioned at `building Atlas by hand`. The build did not result in any errors. I have also tried to build it using OPTS = -O2 -fPIC -m64 and NOOPT = -O0 -fPIC -m64 when building the Lapack. However after installing I tried to import numpy in python and received to following error: ImportError: /users/francisd/local/lib/python2.5/site-packages/numpy/linalg/lapack_lite.so: undefined symbol: zgesdd_ What lapack_lite.so does depend on ? (ldd /users/francisd/local/lib/python2.5/site-packages/numpy/linalg/lapack_lite.so) Output: libpthread.so.0 = /lib64/libpthread.so.0 (0x2b4187472000) libc.so.6 = /lib64/libc.so.6 (0x2b418768c000) /lib64/ld-linux-x86-64.so.2 (0x0036b880) Googling this error, it seems that the error is caused by having a wrong lapack version. I am running Centos 5 (basically Red Hat Enterprise Linux 5) on a x86_64 machine. Unfortunately, it can be many things. Blas/Lapack/Atlas are difficult to build by yourself, there are tens of way to screw up at any point. If you give us the above information, it should be clearer where exactly an error was made You make it sound as there is no easy solution. cheers, David ___ Numpy-discussion mailing list [EMAIL PROTECTED]://projects.scipy.org/mailman/listinfo/numpy-discussion ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
Francis wrote: I don't really need Python2.5. Python2.4 is actually available on the network but I think I will run into problems using the network one and try using local libraries. No, it is perfectly supported (I do it all the time, even on my own machines where I have root rights). You use --prefix when installing packages, and you need to make sure that PYTHONPATH is correctly set up, e.g.: python setup.py install --prefix=$HOME/local # will install in $HOME/local/lib/python2.4/site-packages for python2.4 And set PYTHONPATH to: PYTHONPATH=$HOME/local/lib/python2.4/site-packages:$PYTHONPATH. You can check which path you use by doing: python -c import numpy; print numpy.__file__. I assume one would need to change the paths on a file which I can't change. Anyway I installed Python2.5 without any problems. The problem is that when building software, you generally don't know whether you did it right or wrong before much later, by building other softwares on the top of it. For numpy/scipy, it is not too much of a problem, because they only depend on the interpreter, but for matplotlib, you will have to build pygtk/pyqt/etc... Since it does not make a difference to use your own python or the included python, you should use it. I personally use python2.4 on Centos to build numpy/scipy myself. Output: libpthread.so.0 = /lib64/libpthread.so.0 (0x2b4187472000) libc.so.6 = /lib64/libc.so.6 (0x2b418768c000) /lib64/ld-linux-x86-64.so.2 (0x0036b880) For some reason, lapack has not been linked. Can you build numpy from scratch and give us the build.log ? For example, in the source tree: rm -rf ./build python setup.py build build.log cheers, David ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
On 16 Sep, 10:18, David Cournapeau [EMAIL PROTECTED]
wrote:
Francis wrote:
I don't really need Python2.5. Python2.4 is actually available on the
network but I think I will run into problems using the network one and
try using local libraries.
No, it is perfectly supported (I do it all the time, even on my own
machines where I have root rights). You use --prefix when installing
packages, and you need to make sure that PYTHONPATH is correctly set up,
e.g.:
python setup.py install --prefix=$HOME/local # will install in
$HOME/local/lib/python2.4/site-packages for python2.4
And set PYTHONPATH to:
PYTHONPATH=$HOME/local/lib/python2.4/site-packages:$PYTHONPATH.
You can check which path you use by doing: python -c import numpy;
print numpy.__file__.
Okay. That should not be much of a problem.
I assume one would need to change the paths
on a file which I can't change. Anyway I installed Python2.5 without
any problems.
The problem is that when building software, you generally don't know
whether you did it right or wrong before much later, by building other
softwares on the top of it. For numpy/scipy, it is not too much of a
problem, because they only depend on the interpreter, but for
matplotlib, you will have to build pygtk/pyqt/etc... Since it does not
make a difference to use your own python or the included python, you
should use it. I personally use python2.4 on Centos to build numpy/scipy
myself.
Output:
libpthread.so.0 = /lib64/libpthread.so.0 (0x2b4187472000)
libc.so.6 = /lib64/libc.so.6 (0x2b418768c000)
/lib64/ld-linux-x86-64.so.2 (0x0036b880)
For some reason, lapack has not been linked. Can you build numpy from
scratch and give us the build.log ? For example, in the source tree:
rm -rf ./build python setup.py build build.log
cheers,
David
I only included the first and last part which I think are relevant.
F2PY Version 2_5585
blas_opt_info:
blas_mkl_info:
libraries mkl,vml,guide not found in /usr/local/lib
libraries mkl,vml,guide not found in /usr/lib
NOT AVAILABLE
atlas_blas_threads_info:
Setting PTATLAS=ATLAS
Setting PTATLAS=ATLAS
Setting PTATLAS=ATLAS
FOUND:
libraries = ['lapack', 'f77blas', 'cblas', 'atlas']
library_dirs = ['/users/francisd/local/lib/atlas']
language = c
customize GnuFCompiler
Found executable /usr/bin/g77
gnu: no Fortran 90 compiler found
gnu: no Fortran 90 compiler found
customize GnuFCompiler
gnu: no Fortran 90 compiler found
gnu: no Fortran 90 compiler found
customize GnuFCompiler using config
compiling '_configtest.c':
/* This file is generated from numpy/distutils/system_info.py */
void ATL_buildinfo(void);
int main(void) {
ATL_buildinfo();
return 0;
}
C compiler: gcc -pthread -fno-strict-aliasing -DNDEBUG -O2 -g -pipe -
Wall -Wp,-D_FORTIFY_SOURCE=2 -fexceptions -fstack-protector --
param=ssp-buffer-size=4 -m64 -mtune=generic -D_GNU_SOURCE -fPIC -fPIC
compile options: '-c'
gcc: _configtest.c
gcc -pthread _configtest.o -L/users/francisd/local/lib/atlas -llapack -
lf77blas -lcblas -latlas -o _configtest
ATLAS version 3.8.2 built by francisd on Mon Sep 15 15:20:08 BST 2008:
UNAME: Linux red-67 2.6.18-92.el5 #1 SMP Tue Jun 10 18:51:06
EDT 2008 x86_64 x86_64 x86_64 GNU/Linux
INSTFLG : -1 0 -a 1
ARCHDEFS : -DATL_OS_Linux -DATL_ARCH_PIII -DATL_CPUMHZ=2500 -
DATL_SSE3 -DATL_SSE2 -DATL_SSE1 -DATL_USE64BITS -DATL_GAS_x8664
F2CDEFS : -DAdd_ -DF77_INTEGER=int -DStringSunStyle
CACHEEDGE: 262144
F77 : gfortran, version GNU Fortran (GCC) 4.1.2 20071124 (Red
Hat 4.1.2-42)
F77FLAGS : -fomit-frame-pointer -mfpmath=387 -O2 -falign-loops=4 -
fPIC -m64
SMC : gcc, version gcc (GCC) 4.1.2 20071124 (Red Hat 4.1.2-42)
SMCFLAGS : -fomit-frame-pointer -mfpmath=387 -O2 -falign-loops=4 -
fPIC -m64
SKC : gcc, version gcc (GCC) 4.1.2 20071124 (Red Hat 4.1.2-42)
SKCFLAGS : -fomit-frame-pointer -mfpmath=387 -O2 -falign-loops=4 -
fPIC -m64
success!
removing: _configtest.c _configtest.o _configtest
FOUND:
libraries = ['lapack', 'f77blas', 'cblas', 'atlas']
library_dirs = ['/users/francisd/local/lib/atlas']
language = c
define_macros = [('ATLAS_INFO', '\\3.8.2\\')]
lapack_opt_info:
lapack_mkl_info:
mkl_info:
libraries mkl,vml,guide not found in /usr/local/lib
libraries mkl,vml,guide not found in /usr/lib
NOT AVAILABLE
NOT AVAILABLE
atlas_threads_info:
Setting PTATLAS=ATLAS
libraries lapack_atlas not found in /users/francisd/local/lib/atlas
numpy.distutils.system_info.atlas_threads_info
Setting PTATLAS=ATLAS
Setting PTATLAS=ATLAS
FOUND:
libraries = ['lapack', 'lapack', 'f77blas', 'cblas', 'atlas']
library_dirs = ['/users/francisd/local/lib/atlas']
language = f77
customize GnuFCompiler
gnu: no Fortran 90 compiler found
gnu: no Fortran 90 compiler found
customize GnuFCompiler
gnu: no Fortran 90 compiler found
gnu: no Fortran 90 compiler found
customize GnuFCompiler using config
compiling
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
Francis wrote: ATLAS version 3.8.2 built by francisd on Mon Sep 15 15:20:08 BST 2008: UNAME: Linux red-67 2.6.18-92.el5 #1 SMP Tue Jun 10 18:51:06 EDT 2008 x86_64 x86_64 x86_64 GNU/Linux INSTFLG : -1 0 -a 1 ARCHDEFS : -DATL_OS_Linux -DATL_ARCH_PIII -DATL_CPUMHZ=2500 - DATL_SSE3 -DATL_SSE2 -DATL_SSE1 -DATL_USE64BITS -DATL_GAS_x8664 F2CDEFS : -DAdd_ -DF77_INTEGER=int -DStringSunStyle CACHEEDGE: 262144 F77 : gfortran, version GNU Fortran (GCC) 4.1.2 20071124 (Red Hat 4.1.2-42) ^^ You built atlas with gfortran /usr/bin/g77 -g -Wall -g -Wall -shared build/temp.linux-x86_64-2.4/ numpy/linalg/lapack_litemodule.o build/temp.linux-x86_64-2.4/numpy/ linalg/python_xerbla.o -L/users/francisd/local/lib/atlas -llapack - llapack -lf77blas -lcblas -latlas -lg2c -o build/lib.linux-x86_64-2.4/ numpy/linalg/lapack_lite.so ^^ Here you use g77. This cannot work. You have to use the same fortran compiler for *everything*, because g77 and gfortran are not ABI compatible. To build numpy and scipy with gfortran, do: python setup.py build --fcompiler=gnu95 But I still find a bit strange that it did not fail during the link. You may have other errors, cheers, David ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
You are right. I have recompiled everything with gfortran, but now I get the following error when I try to import numpy: undefined symbol: _gfortran_concat_string. I have build a while ago sage-math (where everything is build locally), but I delete it since it was over 1GB and I am only interested in numpy+scipy+sympy+matlibplot. Although I did not test it extensive, importing numpy and scipy was not a problem. Perhaps I could try it again and try to use the libraries they use to install numpy if it is clear how it is done. ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
Francis wrote: You are right. I have recompiled everything with gfortran, but now I get the following error when I try to import numpy: undefined symbol: _gfortran_concat_string. This means something at some point was not built with gfortran, or that you did not use gfortran to build numpy. Which .so file generates this error ? Can you give us the result of ldd on that .so ? Also, are you sure libraries paths are correctly propagated ? In particular, you should have LD_LIBRARY_PATH set up correctly to where you put your bustom-built libraries. I have build a while ago sage-math (where everything is build locally), but I delete it since it was over 1GB and I am only interested in numpy+scipy+sympy+matlibplot. Although I did not test it extensive, importing numpy and scipy was not a problem. Yes, I think they package everything by themselves. I am starting to think that we should do the same for blas/lapack, that's also how R does it by default, and I am so tired of seeing the same errors coming again and again. I think that distributions will never package blas/lapack correctly, it is hopeless. You could also try EPD, cheers, David ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
Hi, On Sep 16, 2008, at 11:13 AM, David Cournapeau wrote: Yes, I think they package everything by themselves. I am starting to think that we should do the same for blas/lapack, that's also how R does it by default, and I am so tired of seeing the same errors coming again and again. I think that distributions will never package blas/lapack correctly, it is hopeless. Just out of curiosity. Humor me for a bit and let's assume that in many cases, the machines that people are building numpy/scipy on will most likely also have some install of R. Perhaps it's a compute server in your lab, or your own machine (for which pre-built versions of R most likely exist anyway). Is it possible to conjure up some special flag people can use during a numpy/scipy build/install so that they are able to link against the packaged libraries in the canonical install of R for their platform, or is that a totally bone-headed idea? -steve ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
On 16 Sep, 16:13, David Cournapeau [EMAIL PROTECTED] wrote: Francis wrote: You are right. I have recompiled everything with gfortran, but now I get the following error when I try to import numpy: undefined symbol: _gfortran_concat_string. This means something at some point was not built with gfortran, or that you did not use gfortran to build numpy. Which .so file generates this error ? Can you give us the result of ldd on that .so ? To be frank, I am bit of noob, but not completely. I use the gfortran make.inc to build lapack. The make.inc file to build and install atlas is using gfortran. I used the --fcompiler=gfortran flag. I don't think there is much else I can do. Also, are you sure libraries paths are correctly propagated ? In particular, you should have LD_LIBRARY_PATH set up correctly to where you put your bustom-built libraries. Being a noob, I just followed the instructions on http://www.scipy.org/Installing_SciPy/Linux. I have build a while ago sage-math (where everything is build locally), but I delete it since it was over 1GB and I am only interested in numpy+scipy+sympy+matlibplot. Although I did not test it extensive, importing numpy and scipy was not a problem. Yes, I think they package everything by themselves. I am starting to think that we should do the same for blas/lapack, that's also how R does it by default, and I am so tired of seeing the same errors coming again and again. I think that distributions will never package blas/lapack correctly, it is hopeless. You could also try EPD, Unfortunately there is no EPD installer for Centos 5 version. There was a blog how to install it, but that look a lot more complicated than what I am trying now. I will try to build sage and look if I can build numpy against the atlas libraries in sage. Otherwise I use sage and remove all the packages that are not relevant for me, e.g. maxima, gap, singular, etc. ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
Francis wrote: To be frank, I am bit of noob, but not completely. I use the gfortran make.inc to build lapack. The make.inc file to build and install atlas is using gfortran. I used the --fcompiler=gfortran flag. I don't think there is much else I can do. Yes there is :) You should use --fcompiler=gnu95, not gfortran. That would explain the problem you had, actually. There is no problem being a noob: compiling numpy itself is easy, but unfortunately, its dependencies are not, and are almost systematically badly packaged by distributions. To make the situation worse, there are still two widely used gnu fortran compilers, which are ABI incompatible... Unfortunately there is no EPD installer for Centos 5 version. There was a blog how to install it, but that look a lot more complicated than what I am trying now. I will try to build sage and look if I can build numpy against the atlas libraries in sage. Otherwise I use sage and remove all the packages that are not relevant for me, e.g. maxima, gap, singular, etc. I am really sorry for all this trouble. The last solution you could try, with a bit old packages is my own repository, with Centos packages: http://download.opensuse.org/repositories/home:/ashigabou/CentOS_5/ Install the numpy and scipy packages from there. There is also a src.rpm for atlas, which enables you to build atlas rpm automatically optimized for your CPU using one command (look at build atlas from ashigabou repository on www.scipy.org/Installing_Scipy/Linux). I have been trying to update those packages yesterday, but some network-related problems prevented me from. cheers, David ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
Re: [Numpy-discussion] building numpy locally but get error: undefined symbol: zgesdd_
David Cournapeau wrote: I am really sorry for all this trouble. The last solution you could try, with a bit old packages is my own repository, with Centos packages: http://download.opensuse.org/repositories/home:/ashigabou/CentOS_5/ I could update numpy to a recent one (1.1.1), but unfortunately, I still cannot test the built packages because of the network problems. Blas and Lapack should work correctly if you build numpy with --fcompiler=gnu95. cheers, David ___ Numpy-discussion mailing list Numpy-discussion@scipy.org http://projects.scipy.org/mailman/listinfo/numpy-discussion
