Evgenii B. Rudnyi, 2008, (c) All rights reserved
UMFPACK is Unsymmetric MultiFrontal direct solver for sparse matrices developed by Prof Timothy A. Davis:
http://www.cise.ufl.edu/research/sparse/umfpack/
The goal of this chapter is consider how to compile UMFPACK under gcc. I will use gcc under Cygwin but the procedure should be the same on Linux. Additionally I will describe how to use the UMFPACK library compiled with gcc under Cygwin from within Microsoft Visual Studio.
The plan is as follows. First we compile UMFPACK without BLAS. This is a good starting point for those who are new to Unix environment. The next step is to compile UMFPACK with the optimised BLAS in order to reach the best efficiency of the library (I will employ ATLAS). The two final steps are for people working on Windows - to compile UMFPACK under MinGW and to use this library from within Microsoft Visual Studio.
UMFPACK depends on AMD and UFConfig. Hence it is necessary to download and
unpack three archives from the UMFPACK site. I will use wget to
get archives and tar/gz to unpack them.
$ wget
http://www.cise.ufl.edu/research/sparse/umfpack/UMFPACK-5.2.0.tar.gz
$ wget
http://www.cise.ufl.edu/research/sparse/UFconfig/UFconfig-3.1.0.tar.gz
$ wget
http://www.cise.ufl.edu/research/sparse/amd/AMD-2.2.0.tar.gz
Let us check that we have the files
$ ls
AMD-2.2.0.tar.gz UFconfig-3.1.0.tar.gz UMFPACK-5.2.0.tar.gz
and then unpack them
$ tar zxvf UMFPACK-5.2.0.tar.gz
Use TAB for name completion. That is, type tar zxvf UM and
press TAB.
$ tar zxvf UFconfig-3.1.0.tar.gz
$ tar zxvf AMD-2.2.0.tar.gz
Explore the directories and files. Short information is in
UMFPACK/README.txt and the documentation is in
UMFPACK/Doc. Pay attention to UMFPACK/Makefile.
This is the main makefile that will be used to build the library.
There are different configuration settings. They are made by editing file
UFconfig/UFconfig.mk. The variables defined in this file will be
included in all makefiles.
Let us start by disabling the use of BLAS. The performance will suffer but
this is the simplest way to start working. In this case, we need just to
specify a C compiler in UFconfig.mk. There are some Fortran
files in the UMFPACK distribution but by default they are not used. For
gcc it suffers to change the next lines
CC = gcc
CFLAGS = -O3
The lines below are not to use BLAS. First it is necessary to comment out variables BLAS and LAPACK
#BLAS = -lblas -lgfortran -lgfortranbegin
#LAPACK = -llapack
Second is to modify
UMFPACK_CONFIG = -DNBLAS
Finally on Windows it is good to add *.exe to the variable
CLEAN = *.o *.obj *.ln *.bb *.bbg *.da *.tcov *.gcov gmon.out *.bak
*.d *.exe
On Windows gcc adds .exe to binaries and this
setting will clean them. Otherwise later on you may need to clean them
manually.
The modified file is UFconfig.mk.noblas. You can just replace
with it UFconfig/UFconfig.mk.
Now we go to the directory UMFPACK and run make
$ cd UMFPACK
$ make
Now make reads Makefile in this directory and executes it. It
builds the AMD library in AMD/Lib, the UMFPACK library in
UMFPACK/Lib, and then demos in UMFPACK/Demo. After
that it runs demos and compares output with the output included with the
library. One sees some differences but they are not essential.
That’s it. Now you have AMD/Lib/libamd.a,
UMFPACK/Lib/libumfpack.a and compiled demos in
UMFPACK/Demo, that you can use as starting points on how to use
UMFPACK in your code.
UMFPACK uses calls to BLAS to reach the maximum efficiency. I will use
ATLAS (you can find precompiled ATLAS with Cygwin at lib.tar.gz) and assume that it is
located at $HOME/lib/atlas
$ ls $HOME/lib/atlas
libatlas.a libblas.a libcblas.a libf77blas.a liblapack.a
If you put it in another location, please modify the path to ATLAS below accordingly.
Let us first change the lines that told UMFPACK not to use BLAS and see
what happens. Change in UFconfig.mk the line with
UMFPACK_CONFIG to
UMFPACK_CONFIG =
or just comment it out. Now we have to delete previously compiled object files
$ make purge
and compile UMFPACK again without -DNBLAS
$ make
What happens is that make compiles the AMD and UMFPACK libraries but fails to link demos. There should be linking errors
gcc -O3 -I../Include -I../../AMD/Include -I../../UFconfig -o
umfpack_di_demo umfpack_di_demo.c ../Lib/libumfpack.a ../../AMD/Lib/libamd.a
-lm
../Lib/libumfpack.a(umf_di_blas3_update.o):umf_blas3_update.c:(.text+0xe8):
undefined reference to `_dtrsm_'
../Lib/libumfpack.a(umf_di_blas3_update.o):umf_blas3_update.c:(.text+0x184):
undefined reference to `_dgemm_'
../Lib/libumfpack.a(umf_di_blas3_update.o):umf_blas3_update.c:(.text+0x1ff):
undefined reference to `_dger_'
../Lib/libumfpack.a(umf_di_local_search.o):umf_local_search.c:(.text+0x4e4):
undefined reference to `_dgemv_'
../Lib/libumfpack.a(umf_di_local_search.o):umf_local_search.c:(.text+0x6b0):
undefined reference to `_dtrsv_'
Here we see a list of functions that now have been inserted in the UMFPACK library and which the linker could not find. However, the libraries AMD and UMFPACK by themselves are done and one can already use them.
There is a useful tool nm that allows us to see the symbols
defined in the libraries. The command
$ nm Lib/libumfpack.a | grep dgemm
U _dgemm_
U _dgemm_
confirms us that the symbol _dgemm_ has been used in the
UMFPACK library but is not defined there (symbol U). On the
other hand, we can see that this function is defined in
libf77blas.a (symbol T)
$ nm $HOME/lib/atlas/libf77blas.a | grep _dgemm_
U _atl_f77wrap_dgemm__
00000000 T _dgemm_
00000000 T _atl_f77wrap_dgemm__
This means that we have to tell make to use ATLAS libraries
when it compiles demos. To this end it is necessary to define correctly the
BLAS variable that we commented out previously
BLAS = -L$(HOME)/lib/atlas -lf77blas -latlas -lg2c
-L defines the path to the libraries and you may need to
modify it appropriately. The path should be absolute or relative to
UMFPACK/Demo, as gcc will be executed in this
directory. UMFPACK uses Fortran interface to BLAS that is defined in
libf77blas.a and the implementation by itself is in
libatlas.a. libf77blas.a has been compiled with
g77 and as such it needs libg2c.a.
Now
$ make
should build demos correctly. Note that this time it does not build the libraries again.
The modified UFconfig.mk set to use ATLAS is here (UFconfig.mk.atlas) but it may be necessary
to change -L in BLAS to the right location.
Cygwin is a tool to port Unix applications to Windows. As such, it
emulates Unix API on Windows and the final application is linked to
cygwin1.dll that implements the interface. One can see it with
cygcheck, for example
$ cygcheck Demo/umfpack_simple.exe
Demo/umfpack_simple.exe
C:\cygwin\bin\cygwin1.dll
C:\WINDOWS\system32\ADVAPI32.DLL
C:\WINDOWS\system32\ntdll.dll
C:\WINDOWS\system32\KERNEL32.dll
C:\WINDOWS\system32\RPCRT4.dll
C:\WINDOWS\system32\Secur32.dll
If we would like to use UMFPACK with Microsoft Visual Studio, we need to
make sure that the compiled libraries do not call Unix API. This is possible
if one uses the flag -mno-cygwin, which forces gcc to use MinGW.
Add -mno-cygwin to CFLAGS in
UFconfig.mk
CFLAGS = -O3 -mno-cygwin
Note that CFLAGS is defined two times and it is necessary to
add this to the second definition or to comment the second definition out.
The file with the change can be found here (UFconfig.mk.mingw).
Now
$ make purge
$ make
and the libraries compiled with -mno-cygwin are ready. Note
that in this case the differences in output files will be much bigger. The
first reason is that with -mno-cygwin the application writes the
end of line as CR LF and in the output supplied with UMFPACK the
end of line is just LF. This can be circumvented with
-b for diff
$ diff -b my_umfpack_di_demo.out umfpack_di_demo.out
Still, there will be more differences as the formatting of numbers is different anyway.
Microsoft Visual C compiler can be called from the command line and this
is possible directly from within the Cygwin environment. To this end, it is
necessary to modify the path and to define several environment variables
required by cl. You can look at Using Microsoft Visual C under
Gygwin to see how I have done it under tsch.
When everything is done correctly, the command cl should
produce something like below
$ cl
Microsoft (R) 32-Bit C/C++-Optimierungscompiler Version 14.00.50727.762
für 80x86
Copyright (C) Microsoft Corporation. Alle Rechte vorbehalten.
Syntax: cl [ Option... ] Dateiname... [ /link Linkeroption...
]
The Microsoft linker recognizes the libraries compiled by gcc
under Cygwin. What we need to do is to just rename them from *.a
to *.lib.
Let us take as an example umfpack_simple.c from
UMPFACK/Demo. Make a directory ms in the directory
where we have started
$ cd ..
$ ls
AMD UFconfig UMFPACK AMD-2.2.0.tar.gz
UFconfig-3.1.0.tar.gz UMFPACK-5.2.0.tar.gz
$ mkdir ms
$ cd ms
and copy this file here. Then copy and rename libamd.a,
libumfpack.a, libf77blas.a, libatlas.a
by changing the extension from *.a to *.lib.
Additionally we need libg2c.a for libf77blas.a and
libgcc.a, as libg2c.a uses some symbols from it.
You should take these two libraries from
/lib/gcc/i686-pc-mingw32/3.4.4 (not from
/lib/gcc/i686-pc-cygwin/3.4.4/).
At the end you should have the next files
$ ls
libamd.lib libf77blas.lib libgcc.lib umfpack_simple.c
libatlas.lib libg2c.lib libumfpack.lib
Now the command
$ cl -MD -I../UMFPACK/Include -I../AMD/Include -I../UFconfig
umfpack_simple.c libumfpack.lib libamd.lib libf77blas.lib libatlas.lib
libg2c.lib libgcc.lib
does the job. It compiles umfpack_simple.c and links the
object file with the libraries. The flags -I tells
cl where to find the headers from UMFPACK. You may need to
modify the path if you have made this directory in another place.
Interestingly enough that cl understands correctly the direct
slash as the sigh for directory. -MD is the regime with which
the MinGW libraries are working best. In this case actually it is possible to
compile without it but there will some warnings from the linker. If you
understand the command, you can make also this settings directly in GUI.
Finally the question how I have found that it is necessary to add
libgcc.lib. The answer is simple. Try to compile without it.
Then you immediately see some symbols missing and simple manipulations with
nm shows that they are in libgcc.a.
By editing makefiles in UMPFACK it is possible to compile AMD and UMFPACK
directly with cl. This should be relatively simple. What will be
more challenging is to compile ATLAS with cl. Alternatively you
can use the optimised Intel BLAS.
Compiling UMFPACK with Microsoft Visual C
How to create MS VC++ project for UMFPACK
Please post your comments, questions, suggestions to the discussion group at http://groups.google.com/group/matrixprogramming.