Building SPAMS
For one of our projects, we wanted to have a nice set of “optimized” software for running a number of different types of sparse reconstruction problems. One nice set of software that runs for both signal reconstruction and for dictionary learning/sparse decomposition is the SPArse Modeling Software (SPAMS) package for Matlab developed by J. Mairal, F. Bach, J. Ponce, G. Sapiro, R. Jenatton, and G. Obozinski with additional support from J. Chieze for the R and Python interfaces.
The package is currently hosted on INRIA’s open platform, where you can download the latest, as of this writing, v. 2.4.
Because the heavy-hitting portions of the system are written in C++, a compilation/building step is required to get SPAMS generate the MEX files Matlab will call when using the package. However, I found that there are a number of issues in getting the package running since the release of this last version on OSX – Mavericks. I wanted to detail the steps I had to go through to get the package to work.
First Attempt
So, after downloading and extracting SPAMS, I went ahead and ran the compile.m Matlab script. There are a few options to be aware of inside the compilation script…
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
clear all;
get_architecture;
%%%%%%%%%%%%% COMPILER CONFIGURATION %%%%%%%%%%%%%%%%
% set up the compiler you want to use. Possible choices are
% - 'mex' (default matlab compiler), this is the easy choice if your matlab
% is correctly configured. Note that this choice might not compatible
% with the option 'use_multithread=true'.
% - 'icc' (intel compiler), usually produces the fastest code, but the
% compiler is not free and not installed by default.
% - 'gcc' (gnu compiler), good choice (for Mac, use gcc >= 4.6 for
% the multi-threaded version, otherwise set use_multithread=false).
% For windows, you need to have cygwin installed.
% - 'open64' (amd compiler), optimized for opteron cpus.
% - 'vs' (visual studio compiler) for windows computers (10.0 or more is recommended)
% for some unknown reason, the performance obtained with vs is poor compared to icc/gcc
compiler='mex';
%%%%%%%%%%%% BLAS/LAPACK CONFIGURATION %%%%%%%%%%%%%%
% set up the blas/lapack library you want to use. Possible choices are
% - builtin: blas/lapack shipped with Matlab,
% same as mex: good choice if matlab is correctly configured.
% - mkl: (intel math kernel library), usually the fastest, but not free.
% - acml: (AMD Core math library), optimized for opteron cpus
% - blas: (netlib version of blas/lapack), free
% - atlas: (atlas version of blas/lapack), free,
% ==> you can also tweak this script to include your favorite blas/lapack library
blas='builtin';
%%%%%%%%%%%% MULTITHREADING CONFIGURATION %%%%%%%%%%%%%%
% set true if you want to use multi-threaded capabilities of the toolbox. You
% need an appropriate compiler for that (intel compiler, most recent gcc, or visual studio pro)
use_multithread=false; % (might not compatible with compiler=mex)
% if the compilation fails on Mac, try the single-threaded version.
% to run the toolbox on a cluster, it can be a good idea to deactivate this
use_64bits_integers=true;
% use this option if you have VERY large arrays/matrices
% this option allows such matrices, but may slightly reduce the speed of the computations.
% if you use the options 'mex' and 'builtin', you can proceed with the compilation by
% typing 'compile' in the matlab shell. Otherwise, you need to set up a few path below.
add_flag='';
% WARNING: on Mac OS mountain lion, you may have to uncomment the line
%add_flag=' -mmacosx-version-min=10.7'
For the initial base compile, I went ahead and left the compiler as mex and turned off the multithreaded support. I just want to get something working, so efficiency is not the chief concern at the moment. Running compile.m gives the following output…
class="highlight">1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
>> compile
compilation of: -I./linalg/ -I./decomp/ -I./prox/ -I./dictLearn/ dictLearn/mex/mexTrainDL.cpp
In file included from dictLearn/mex/mexTrainDL.cpp:36:
In file included from ./linalg/mexutils.h:14:
In file included from /Applications/MATLAB_R2013b.app/extern/include/mex.h:58:
In file included from /Applications/MATLAB_R2013b.app/extern/include/matrix.h:252:
/Applications/MATLAB_R2013b.app/extern/include/tmwtypes.h:831:9: error: unknown type name 'char16_t'
typedef char16_t CHAR16_T;
^
In file included from dictLearn/mex/mexTrainDL.cpp:37:
In file included from ./dictLearn/dicts.h:35:
./decomp/decomp.h:2:1: warning: '/*' within block comment [-Wcomment]
/* Software SPAMS v2.2 - Copyright 2009-2011 Julien Mairal
^
1 warning and 1 error generated.
mex: compile of ' "dictLearn/mex/mexTrainDL.cpp"' failed.
Unable to complete successfully.
Error in compile (line 390)
mex(args{:});
There appear to be some issues with a core Matlab mex-specific header file. That doesn’t seem good. So a little digging around turned up a Stack-Exchange thread on this topic. The core issue seems to be that the latest updates to C++ have not yet propagated to Matlab properly. There are two possible work-arounds:
- Modify your
mexopts.sh settings script. - Modify the offending line in header file, itself (not a great solution).
I went with the first approach. It probably pays to have a healthy fear of messing with the innards of Matlab.
Updating mexopts.sh
This is the most direct approach. All one needs to do is to update the CXXFLAGS flag in the script by adding the additional option -std=c++11. For me, this was l. 150 in mexopts.sh.
So, we change from
class="highlight">1
CXXFLAGS="-fno-common -fexceptions -arch $ARCHS -isysroot $MW_SDKROOT -mmacosx-version-min=$MACOSX_DEPLOYMENT_TARGET"
to
class="highlight">1
CXXFLAGS="-fno-common -fexceptions -arch $ARCHS -isysroot $MW_SDKROOT -mmacosx-version-min=$MACOSX_DEPLOYMENT_TARGET -std=c++11"
Now, re-running compile.m I get
class="highlight">1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
>> compile
... a bunch of warnings and notes ...
compilation of: -I./linalg/ -I./prox/ prox/mex/mexSimpleGroupTree.cpp
prox/mex/mexSimpleGroupTree.cpp:33:22: error: non-constant-expression cannot be narrowed from type
'int' to 'mwSize' (aka 'unsigned long') in initializer list [-Wc++11-narrowing]
mwSize cdims[] = {n};
^
prox/mex/mexSimpleGroupTree.cpp:33:22: note: override this message by inserting an explicit cast
mwSize cdims[] = {n};
^
static_cast<mwSize>( )
1 error generated.
mex: compile of ' "prox/mex/mexSimpleGroupTree.cpp"' failed.
Unable to complete successfully.
Error in compile (line 390)
mex(args{:});
Now, this appears to be an issue with the way narrowing is handled between the previous and the current version of C++. So, lets take mex’s advice and investigate this line.
Modifying prox/mex/mexSimpleGroupTree.cpp
Taking a look at the compiler error messages, we see that we should take a look at l. 32 of prox/mex/mexSimpleGroupTree.cpp. Here, we see that I’ve commented out the previous version and inserted an updated version which uses the suggested type-conversion rather than an assumed narrowing.
class="highlight">1
2
3
4
5
int n = dims[1];
// Old way -- This throws errors for c++11
// mwSize cdims[] = {n};
// Updated way -- Maybe no errors ??
mwSize cdims[] = {static_cast<mwSize>(n)};
Now, we run compile.m…
class="highlight">1
2
3
4
5
6
7
8
9
10
11
12
>> compile
... many warnings and notes...
1 warning generated.
compilation of: -I./linalg/ -I./decomp/ decomp/mex/mexSparseProject.cpp
compilation of: -I./linalg/ linalg/mex/mexBayer.cpp
compilation of: -I./linalg/ -I./prox/ prox/mex/mexGraphOfGroupStruct.cpp
compilation of: -I./linalg/ -I./prox/ prox/mex/mexGroupStructOfString.cpp
compilation of: -I./linalg/ -I./prox/ prox/mex/mexReadGroupStruct.cpp
compilation of: -I./linalg/ -I./prox/ prox/mex/mexSimpleGroupTree.cpp
compilation of: -I./linalg/ -I./prox/ prox/mex/mexTreeOfGroupStruct.cpp
>>
And so we have success :) . At this point, for me, all the SPAMS test scripts in the test_release directory work.
Summary
The basic steps to follow to get SPAMS to build on OSX Maverics:
- Change
CXXFLAGS in mexopts.sh to include the option -std=c++11. - Modify line 32 of
prox/mex/mexSimpleGroupTree.cpp to mwSize cdims[] = {static_cast<mwSize>(n)};.
Eric W. Tramel, Ph.D.
Senior Machine Learning Scientist, Owkin