I am currently using ifort (Intel Fortran compiler). However, in the AMD cluster (which has AMD processors), it seems that Intel parallel studio can not be installed. Therefore I decided to use gfortran (GNU Fortran compiler) instead of ifort. In ifort, my compiler flags were
ftn -heap-arrays -mcmodel=large -g -c
I want to know the similar optimization flags in the GNU compiler.
My GCC version is 4.4.7 in centos 6.6 and the rocks cluster version is 6.2.
Related
I have to run a code. The code includes mex files. For building that code, it required gcc version of 4.4.
Following is the description of system
Windows 10
Visual Studio 2012
Matlab 2017a
And gcc version of my system is:
$gcc --version
gcc --version
gcc (GCC) 7.4.0
I do not know how to downgrade it. I have tried so many steps, but no use.
You'll need to install it first - Please use a previous question on the subject here:
Next - choose to use it on Matlab via :
mex -setup
#mano Here's my output of the mex -setup
I have Matlab2015a 64bit running on Win7 and I want to use Mingw-w64 compiler for mex files but the command *
mex -setup
says
MEX configured to use 'Microsoft Visual C++ 2012 (C)' for C language
compilation.
I have installed Mingw-w6 compiler in
C:\mingw-w64\x86_64-4.9.2-win32-seh-rt_v4-rev4\mingw64\bin
I even had a look at this question, but I could'n solve it.
How can I switch to Mingw gcc compiler?
To see which compilers are supported by your MATLAB version and OS go to
mathworks.com/support/compilers/VERSION/index.html
where VERSION is your version. i.e.
mathworks.com/support/compilers/R2015a/index.html
If a compiler is not there, then you can not use it. In your case, its not there.
I am using 32-bit eclipse CDT Kelper to manage a piece of code, which I could compile on command line but not with eclipse if I use any up-to-date c++11 feature. CDT tells me that "-std=c++11" is unrecognized.
After turning on the verbose option in project properties\settings\tool settings, I found that eclipse CDT somehow chooses an older version of MinGW that comes with Haskell platform 2013.2:
Configured with: ../gcc-4.5.2/configure --enable-languages=c,c++,ada,fortran,objc,obj-c++ --disable-sjlj-exceptions --with-dwarf2 --enable-shared --enable-libgomp --disable-win32-registry --enable-libstdcxx-debug --enable-version-specific-runtime-libs --disable-werror --build=mingw32 --prefix=/mingw
Thread model: win32
gcc version 4.5.2 (GCC)
COLLECT_GCC_OPTIONS='-O0' '-g3' '-Wall' '-c' '-fmessage-length=0' '-v' '-o' 'src\vaomp_bnb.o' '-shared-libgcc' '-mtune=i386' '-march=i386'
c:/haskell platform/2013.2.0.0/mingw/bin/../libexec/gcc/mingw32/4.5.2/cc1plus.exe -quiet -v -iprefix c:\haskell platform\2013.2.0.0\mingw\bin\../lib/gcc/mingw32/4.5.2/ -dD ..\src\vaomp_bnb.cpp -quiet -dumpbase vaomp_bnb.cpp -mtune=i386 -march=i386 -auxbase-strip src\vaomp_bnb.o -g3 -O0 -Wall -version -fmessage-length=0 -o C:\DOCUME~1\ting\LOCALS~1\Temp\ccWNoh7I.s
GNU C++ (GCC) version 4.5.2 (mingw32)
I searched SO, and there is a similar question a half year ago here. But the answer there was about setting environment variables and didn't solve the problem.
In my case, CDT can find a MinGW GCC, but found the wrong one. I have installed tdm-gcc 64 bit with gcc-4.8.1, and Haskell platform. The tdm-gcc has priority and
gcc --version
on both DOS and MSYS shows
gcc.exe (tdm64-2) 4.8.1
I don't know what heuristic does CDT use to find toolchains. My question is, how can I tell CDT to use the gcc in a specific location, e.g. c:/MinGW?
Note, I can't uninstall the Haskell platform version of gcc as I will need HP.
Thanks,
I figured it out. It seems that Eclipse CDT does not take tdm-gcc MinGW 64-bit because eclipse itself is 32-bit. So it picked up the only 32-bit MinGW it can find, which is the old version from HP.
To verify this, I installed java 1.7 64-bit, and downloaded 64-bit version of Eclipse CDT (Kelper). Now, CDT automatically selects the TDM-GCC 64 version of MinGW gcc.
For those in a similar situation, one additional issue is that the 64-bit CDT - MinGW64 combination combination does not build projects. No binary/executable files are generated even for a newly created HelloWorld c++ project (even though compilation is OK). In my case, I changed the build tool in the project properties..\tool chain. from CDT internal builder to either of the other two options (one is Gnu's and another is a long name), and then the project can be successfully made to generate exe files.
I don't want to install all as it consumes much space..
the following is what i'm sure is needed:
gcc, make python..
what else?
On the site, in their documentation on installing...
http://www.shogun-toolbox.org/doc/en/current/installation.html
Requirements
SHOGUN requires the standard linux utils like bash, grep, test, sed,
cut, ldd, uname gcc g++ and cat python (debian package: python2.4 or
python2.5) for the ./configure to work.
Optionally you will need atlas and lapack (debian packages
lapack3-dev, atlas3-headers atlas3-base-dev or atlas3-altivec-dev
atlas3-sse2-dev) installed. Note that atlas/lapack is only supported
under linux (high performance computing should be done under linux
only anyway). In case atlas/lapack is unavailable, don't worry most of
shogun will work without, though slightly slower versions are used.
To
enable Multiple Kernel Learning with CPLEX(tm) just make sure cplex
can be found in the PATH. If it is not found shogun will resort to
GLPK (if version at least 4.29 is found) for 1-norm MKL, p-norm MKL
with p>1 will work nonetheless.
To compile the C# interface you need to have the C# developer files
(debian package mono-utils) installed.
To compile the Lua interface
you need to have the Lua developer files (debian package lua)
installed.
To compile the Java interface you need to have the Java
developer files (debian package openjdk-6-jdk) installed.
To compile
the R interface you need to have the R developer files (debian package
r-base-dev) installed.
To compile the Ruby interface you need to have
the Ruby developer files (debian package ruby) installed.
To compile
the octave interface you need to have the octave developer files
(debian package octave3.0-headers) installed.
To compile the python
interface you need to have numpy version 1.x installed (debian package
python-numpy) installed.
I have gcc 3.4.6 installed on sparc9 and sparc10 systems. I am having incompatibility issues when I try to use shared object(built on sparc9) on sparc10. I am not seeing any issue if I use same shared object on sparc9 system. Snippet from makefile:-
LIB=-L/usr/lib/sparcv9 -L/usr/local/lib/sparcv9 -L/lib/sparcv9
gcc -m64 -shared -fPIC -o myapi.so.1 myapi.o $(LIB) -lc -lstdc++ -luuid
Now following in sparc10:-
ldd -d /home/myapi.so
libstdc++.so.6 => /usr/local/lib/sparcv9/libstdc++.so.6
symbol not found: _ZNSt15basic_streambufIcSt11char_traitsIcEE7seekoffExSt12_Ios_SeekdirSt13_Ios_Openmode (./myapi.so)
I have skipped other dependencies in above output which appear normal.
Then, I ran "nm" against libstdc++.so.6 in /usr/local/lib/sparcv9 for above unreferenced symbol and came with following conclusion:-
V9---> _ZNSt15basic_streambufIcSt11char_traitsIcEE7seekoffExSt12_Ios_SeekdirSt13_Ios_Openmode
V10--->_ZNSt15basic_streambufIcSt11char_traitsIcEE7seekoffElSt12_Ios_SeekdirSt13_Ios_Openmode
If you look carefully, you can see a difference in mangled-name symbol in sparc9 and sparc10... "ExSt12" v/s "ElSt12".
Similarly, if I build myapi.so on sparc10, it fails on sparc9 because of symbol mismatch in V9 and V10.
Can someone please point me to any helpful hints as to if I am missing something? Or do I need to install any package on V9. I am assuming that shared object built on V9 should work on V10.
I would like to add that there is a size of libstdc++.so.6 on V9 is different from V10. V9 has SMCgcc and V10 has both SUNWgccruntime and SMCgcc.
Those symbols demangle as:
std::basic_streambuf<char, std::char_traits<char> >::seekoff(long long, std::_Ios_Seekdir, std::_Ios_Openmode)
and
std::basic_streambuf<char, std::char_traits<char> >::seekoff(long, std::_Ios_Seekdir, std::_Ios_Openmode)
Notice that one uses long and the other uses long long
I suspect that means one was built with 64-bit file offsets (aka large file support) and the other wasn't. That means the two versions of GCC have been built differently (by sunfreeware.com or whoever provided them) and are incompatible.
Looks like you're using incompatible versions of the GNU C++ compiler & standard library on the two OS'es. Unfortunately, there is no standard ABI for C++ and many compilers change symbol mangling conventions incompatibly between major releases (such as Sun C++ compiler 4.x vs. 5.x, or g++ 2.x vs. 3.x vs. 4.x).
The /usr/local/lib path implies you're using the libraries from the SMCgcc packages downloaded from the sunfreeware.com site - make sure you have matching versions on the Solaris 9 and 10 machines. (There is no "sparc 10", so you must mean Solaris 10.)