%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%                                                           %
% A real-time convex optimization                           %
% sparse linear prediction solver                           %
%                                                           %
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

This code are designed around solving the convex optimization 
problem

minimize ||x-X*alpha||_1 + gamma || alpha ||_1

It also includes scipts for compiling, linking, test, timings 
and benchmark procedures.

The algorithmic background, motivation, etc. is described in

REAL-TIME IMPLEMENTATIONS OF SPARSE LINEAR PREDICTION 
              FOR SPEECH PROCESSING
T.L. Jensen, D. Giacobello, M.G. Christensen, S.H. Jensen, M. Moonen
                   ICASSP2013

If you use this implementation then please give reference to the 
above paper.

----- Matlab: 
There is a Matlab/Mex interface to the core implementations in 
C++. This needs to be compiled. The following procedure should 
work on all platforms but linking on a range of platforms can
be difficult.

1. Go to the dir SparseLP/mlib
Of you have yet not used mex, then you need to run 

>>mex -setup 

and possible install a compiler or select one from the presented
list. Then you can run

>> install_slp

It will default linkt against the BLAS/LAPACK libraries that comes
with Matlab. If this does not succeed then, see the output error,
use help install_slp, and possible supply your own BLAS/LAPACK
libraries and the corresponding link command as e.g.
install_slp('-lblas -llapack').
 
2.
Go to the dir SparseLP/test
>>test_cimp
This will solve a small problem. It also provides an example
of how to use the solver. Possible consult the description via

>> help slp
(make sure SparseLP/mlib is in your path) 


----- Reproducibility:
This can be a little difficulty because it requires the use of 
code/executables in several languages and different software 
packages as well as some commandline possibilities. We recommend
to try this first on a unix system

1. Requirements:
   Run Makefiles
	 gcc (GNU C compiler)
	 Matlab
	 CVX (http://cvxr.com/cvx/)

2. First we need to be able to link to BLAS/LAPACK library.
Go to SparseLP/src/Makefile. Here you can set the options
at the top. For most unix based systems you can change the lines

LINK = -lblas -llapack
#LINK = -DMKL $(MKL)

to link to the default system BLAS/LAPACK libraries. We will
recommend that you use vendor specific implemenations since 
they are usually faster.

Then in SparseLP/src run

$ make slp_primal_double
$ ./slp_primal_double

and you should see an output with some iterations, where the numbers
gets smaller and smaller (the algorithm converges). 

Also goto SparseLP/cvxgen5010 and make the cvxgen solver
>> make

3. Generate the test data. Go to SparseLP/data_process
>> form_benchmark_data( 40, 10)
>> form_benchmark_data( 160, 40)
>> form_benchmark_data( 160, 100)

Then go to SparseLP/test and run the script test_all.sh from the
command line. In e.g. unix systems
$ ./test_all.sh

This will probably take some hours. On my machine ~8hrs.

4. The data can be viewed as follows. Go to SparseLP/test
>>plot_benchmark_signal('problem5010')
>>plot_benchmark_signal('problem20040')
>>plot_benchmark_signal('problem260100')