How to set up mex with gfortran on Mac?

I am using Matlab 2016a on a MacOS 10.11 El Capitan (although I have reached the same problem on Sierra) trying to get mex to play nice with gfortran. I have followed the advice of this answer and have reinstalled the .xml files, but when I try to set up mex I get the following result. Not only does "which" not appear to find gfortran (while in the terminal it does), but it also can't find the "libgfortranbegin.a" folder. Mex does work with C. What should I try next?

mex -setup -v Fortran
Verbose mode is on.
... Looking for compiler 'gfortran' ...
... Looking for environment variable 'DEVELOPER_DIR' ...No.
... Executing command 'which gfortran' ...No.
... Looking for file '/usr/local/bin/gfortran' ...Yes.
... Looking for folder '/usr/local/bin' ...Yes.
... Executing command 'which gfortran' ...No.
... Looking for file '/usr/local/bin/gfortran' ...Yes.
... Executing command '/usr/local/bin/gfortran -print-file-name=libgfortran.dylib' ...Yes ('/usr/local/gfortran/lib/libgfortran.dylib').
... Looking for folder '/usr/local/gfortran/lib' ...Yes.
... Executing command 'which gfortran' ...No.
... Looking for file '/usr/local/bin/gfortran' ...Yes.
... Executing command '/usr/local/bin/gfortran -print-file-name=libgfortranbegin.a' ...Yes ('libgfortranbegin.a').
... Looking for folder 'libgfortranbegin.a' ...No.
Did not find installed compiler 'gfortran'.
... Looking for compiler 'Intel Fortran' ...
... Looking for environment variable 'IFORT_COMPILER14' ...No.
... Looking for environment variable 'IFORT_COMPILER13' ...No.
... Executing command 'which ifort' ...No.
Did not find installed compiler 'Intel Fortran'.
Error using mex
No supported compiler or SDK was found. For options, visit
http://www.mathworks.com/support/compilers/R2016a/maci64.html.

Answer 1

[Walter Roberson]

The issue appears to be that gfortran no longer includes libgfortranbegin.a and so the part of the .xml file that checks for it is no longer correct. Discussion of libgfortranbegin at https://root-forum.cern.ch/t/cant-find-libgfortranbegin-a-when-installing-root-6-06-04/21222

But after you hack gfortran.xml to remove the references to that library, the problem remains that no matter what you do, mex forces -i8 to be added to the end of FFLAGS . That is a problem because -i8 is no longer used as a flag by gfortran.

After a bunch of poking around, I was able to work around these two problems.

Step 1 is to install gfortran for OS-X. I picked mine up at http://coudert.name/software.html

Step 2 is to edit /Applications/MATLAB_R2017b.app/bin/maci64/mexopts/gfortran.xml (or as appropriate for your MATLAB version.) Replacement version is enclosed; unzip and put into that directory.

Note: I made several changes compared to the MATLAB provided version from the Question you linked to. I removed the several references to libgfortranbegin, and I added support for XCode SDK 10.12, and I coded the name of the fortran compiler as gfortrani8 in the FC assignment in the "vars" section. The changes I made to libgfortranbegin and SDK 10.12 should not give anyone any trouble, but the gfortrani8 that I introduced does not exist elsewhere.

Step 3 ... well if you installed from that source I gave, just assume the answer you need here is /usr/local/gfortran/bin and go on to Step 4.

Otherwise, go to a terminal window and in a shell (bourne shell family, not csh or zsh) execute

readlink $(which gfortran)

If it gives you back a non-empty answer then the result you need from this step is the directory of what it returns. For example if it returns /usr/local/gfortran/bin/gfortran then the answer you need is /usr/local/gfortran/bin

(If someone using Linux is trying to get some hints from this message, then readlink might not work for you: in that case

ls -l $(which gfortran)

and take the directory of what is returned.)

What you need here is the name of the directory that the gfortran executables are really installed in, which is probably the same place that gcc is really installed in. This will likely not be /usr/local/bin even though there may be a /usr/local/bin/gfortran -- that probably points off to the real directory.

Step 4: download the file gfortrani8.zip that I enclose, and unzip it. It is a shell script. Now copy that shell script into the directory found through step 3. For example it might become /usr/local/gfortran/bin/gfortrani8 . Then, at the shell, chmod +x the script.

If you received this message when you mex'd:

Error using mex
gfortran: error: unrecognized command line option â-i8â; did you mean â-k8â?

then you need to be sure to install this gfortrani8 script.

If you receive this message:

Error using mex
/usr/local/gfortran/bin/gfortrani8: No such file or directory

then you did successfully install my gfortran.xml but you did not install my gfortrani8 (or did not get it into the right place.)

Step 5: go into MATLAB and

delete(fullfile(prefdir, 'mex_FORTRAN_maci64.xml'))

and do not worry if it says it did not already exist.

Step 6: in MATLAB,

mex -setup fortran

If it tells you the compiler was found, then you should be able to proceed with compiling.

Step 7: In MATLAB you will probably need to,

 setenv('DYLD_LIBRARY_PATH', [matlabroot(),'/bin/maci64:', matlabroot(), '/sys/os/maci64:', getenv('DYLD_LIBRARY_PATH')])

If your tests work, you might want to consider putting that into your startup.m file

Step 8: in at least some cases, when you compile, instead of just

mex FileName.F

you might need to

mex -client engine FileName.F

If you got the message

Error using mex
Undefined symbols for architecture x86_64:
  "_mexfunction_", referenced from:
     -exported_symbol[s_list] command line option
ld: symbol(s) not found for architecture x86_64
clang: error: linker command failed with exit code 1 (use -v to see invocation)

when you mex'd, then this is probably the cure.

Step 9: run your code. If you did use the -client engine option then what was produced will be an external executable that you would execute with the system() or ! commands.

Step 10: curse when your program runs part way and then crashes...

[Zhuqing Yuan]

Thanks for providing the answer. I have the same problem in MATLAB 2017b with Fortran. I have a question about the enclosed files. Could you please tell me what does the gfortrani8 do? Is it a compiler?

[Walter Roberson]

The gfortran8i that I included here is a small shell script that examines the input arguments it is given, and deletes any -i8 option that it finds, and then calls the real gfortran with what remains. I could not find a way back then to get mex to stop inserting -i8 so I instead took the route of intercepting the call to the compiler, removing the -i8, and then calling the real compiler.

I just made a correction to the setenv() I had given before.

Note that MacOS Sierra and MacOS High Sierra might disregard DYLD_LIBRARY_PATH . For those two operating systems, it might be necessary to put symbolic links from /usr/bin to matlabroot's bin/maci64 and sys/os/maci64 . I have not tried that yet so I am not sure of the exact steps.

[Tommaso Pino]

Hello Walter, I used your file xml and your script!

Thank you very much to post it. I would like to point out that a substitution for the old flag -i8 is needed. I suggest to modify the gfortrani8 script from this:

"${0%/*}/gfortran" -cpp "$@"

to this

"${0%/*}/gfortran" -cpp -fdefault-integer-8 "$@"

I needed that to correctly compile mine snopt distribution.

The new file is attached.

To use gfortran in macos 10.13 I also modified the .xml file that you provided. You could find it as attach.

Thanks

[Walter Roberson]

Good point, Tommaso Pino.

[Uday]

Also, instead of checking a bunch of SKDs and having to update the list, would it be easier to replace the search in ISYSROOT and SDKVER with the following?

"<cmdReturns name="xcrun -sdk macosx --show-sdk-path"/>"

This finds whatever SDK is installed (in my case it's 10.13). I tested this out, and mex does seem to find the correct SDK. However, I have other issues mex'ing and not sure it is related to this.

[Emily Jollans]

Thanks for the instruction! I managed to get gfortran working in R2018a on macos 10.14 (mojave) with some additional steps.

In the gfortran.xlm file I added my macOS version to the four places where versions are specified.

...
<dirExists name="$$/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.9.sdk" />
<dirExists name="$$/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.10.sdk" />
<dirExists name="$$/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.11.sdk" />
<dirExists name="$$/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.12.sdk" />
<dirExists name="$$/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.14.sdk" />
<cmdReturns name="find $$ -name MacOSX10.9.sdk" />
<cmdReturns name="find $$ -name MacOSX10.10.sdk" />
<cmdReturns name="find $$ -name MacOSX10.11.sdk" />
<cmdReturns name="find $$ -name MacOSX10.12.sdk" />
<cmdReturns name="find $$ -name MacOSX10.14.sdk" />
 ...

I also had to create the directory /Library/Developer/CommandLineTools/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.14.sdk

and copied to it the MacOSX.sdk directory from /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/ using the bash shell.

cp -r /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk /Library/Developer/CommandLineTools/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.14.sdk

I then copied the fortani8 file to my gfortran directory which (when installed using homebrew) was /usr/local/Cellar/gcc/8.2.0/bin/ for me.

cp ../Downloads/gfortrani8 /usr/local/Cellar/gcc/8.2.0/bin/

Then run the following in MATLAB and it should work:

mex -setup -v FORTRAN

[dpb]

Any reason you couldn't just build the mex file outside the mex utility, Walter? If you can find all the pieces of the make file to know all the libraries and all...this having turned mex into a builtin and not being able to mung it is a real kick in the teeth. TMW ought to be ashamed of themselves in this area.

I found this thread while thinking about what it might take to manage to get gfortran to work on Windows again where it is "unsupported". As another said, just because TMW doesn't want to answer calls doesn't mean they should actively make it essentially impossible for the user...and there seems virtually no rational reason not to support gfortran when they do support gcc and all gfortran is is the front end language parser for gcc.

It's simply unconscionable on their part in my view...

[Vahagn Davtyan]

Hi,

I try to follow Your steps, but when I run:

mex -setup fortran

I get an error:

Error using mex

No supported compiler was found. For options, visit https://www.mathworks.com/support/compilers.

I am using Matlab R2020b and macOS 12.0.1 (Monterey)

Answer 2

[Sanjana Ramakrishnan]

Please note that gfortran is not a supported compiler on MATLAB R2016a on Mac OS. Refer the below link for more information:

https://www.mathworks.com/content/dam/mathworks/mathworks-dot-com/support/sysreq/files/SystemRequirements-Release2016a_SupportedCompilers.pdf

[Robert]

not supported doesn't mean that you go out of your way to not allow us to use it

[dpb]

Amen to that!

[dpb]

I'll also note that just looked to see and the ML download in Apps for the MATLAB Compiler downloads and installs gfortran.

I just compiled an ran successfully the minimal "Hello, World!" program so it's hard to see what's keeping TMW from not allowing mex Fortran.

I've not tried to see if can build mex from command line; I may give that a go if get a little extra time more to dig into the make files to find the TMW libraries to link.

The problem I'm afraid am going to run into again is the poison pill of the embedded version stuff...

Answer 3

[Daming Lou]

Hi

Thanks Walter Roberson! I manage to install the C & Fortran compiler in mac (OS10.13) with Matlab 2016b.

mex -setup fortran 

works. But when I run a code from a project gives me a issue which is listed on your post

Undefined symbols for architecture x86_64:
  "_mxsetfield_", referenced from:
      _mexfunction_ in xxx.o

I have figured it out that the mxsetfield is called in the xxx.o like

        fptr = fortran_to_matlab(fact)
        call mxSetField( plhs(1), ind_i, f2, fptr )

        fptr = fortran_to_matlab(ifact)
        call mxSetField( plhs(1), ind_i, f4, fptr )

        fptr = fortran_to_matlab(cntl)
        call mxSetField( plhs(1), ind_i, f5, fptr )

        fptr = fortran_to_matlab(icntl)
        call mxSetField( plhs(1), ind_i, f6, fptr )

        fptr = fortran_to_matlab(info)
        call mxSetField( plhs(1), ind_i, f7, fptr )

        fptr = fortran_to_matlab(rinfo)
        call mxSetField( plhs(1), ind_i, f8, fptr )

I also checked the internet like crazy, it seems related to "Symbol mexFunction Unresolved or Not Defined" ( https://nl.mathworks.com/help/matlab/matlab_external/symbol-mexfunction-unresolved-or-not-defined.html ).

I am very bad at Fortran and have no clue how I can fix.

Any suggestion will be welcome! Thanks!.

[dpb]

That indicates the linker couldn't find the entry point in the mxlib library.

If this is the only unsatisfied external, is this the only call in the mex file to a mx function? That would seem highly unlikely, but if the problem is that the mxlib object file isn't in the search path and it were the only function trying to link to, then fixing that might resolve the issue....

But, if as is more likely, the problem is that ony this one mx routine didn't link but the others did, then the compiler generated the object name inconsistently with those names exposed in the mx object library. That kind of an issue is going to be much harder to resolve if were to be so.

The fortran include file creates a bunch of macros that create naming conventions for all the functions that are based on the detected compiler and MATLAB version and OS -- there are different naming conventions used depending on those including capitalization and whether or not the underscores are/are not used, etc., etc., that all have to agree with the calling convention used when compiling the libraries.

I don't have Mac and the experiments I outlined above to try to use GFortran on Windows got to be more involved than I had time to continue to pursue so no specifics to offer other than that's what the errror is telling you -- it couldn't find the subject function in the mx library.

Can you successfully compile/run one of the simple Fortran examples like timestwo to see if a base trivial mex function will work?

You'll get more info to help others who may have Mac if you'll rerun the mex command with the "-v" flag and post the whole response...