# SPM/Faster SPM

## Contents

## SPM Optimisation[edit]

### Adjusting SPM settings[edit]

A defaults variable called *maxmem* indicates how much memory can be used at the same time during GLM estimation. If your computer has a large amount of RAM, you can increase that memory setting in spm_defaults.m:

defaults.stats.maxmem = 2^30;

- 2^32 = 4GB
- 2^31 = 2GB
- 2^30 = 1GB
- 2^29 = 512MB

In SPM12, there is another defaults variable called *resmem* governing whether temporary files during GLM estimation are stored on disk (*false*) or kept in memory (*true*). If you have enough available RAM, not writing the files to disk with fasten the estimation.

defaults.stats.resmem = true;

### Compiling the MEX files[edit]

The compiled MEX files provided with SPM are built in such a way to be compatible with most platforms and MATLAB versions but you might benefit from compiling them for your exact platform/MATLAB version - some C compilers might also produce better optimised binaries (such as Intel Compilers). See the installation pages for more details on how to recompile SPM MEX files.

## MATLAB Optimisation[edit]

### General statements[edit]

Maximizing MATLAB Performance.

Note that it is not recommended to disable the JAVA Virtual Machine when launching MATLAB (matlab -nojvm). If you don't want to use the MATLAB desktop, you can preferably launch MATLAB with:

matlab -nodesktop

When using Matlab in 'nodesktop' mode, initialise SPM in the following manner to prevent graphics windows from opening:

spm defaults fmri spm_jobman initcfg spm_get_defaults('cmdline',true)

(Substituting 'fmri' for 'pet' or 'eeg' as appropriate.)

### Multithreading[edit]

Recent MATLABs support implicit multiprocessing allowing to run multiple threads on a single machine without any change to the MATLAB code itself: this requires a multiple CPU (multiprocessor or multicore) system. The gain in compute time with SPM is not dramatic though.

See:

- http://www.mathworks.com/access/helpdesk/help/techdoc/matlab_prog/brdo29n-1.html
- http://www.mathworks.com/access/helpdesk/help/techdoc/ref/maxnumcompthreads.html

If you run many MATLAB sessions in parallel to manually distribute your SPM processings, it is recommended to set the number of computational threads to one.

### Updating BLAS/LAPACK[edit]

Install the latest Basic Linear Algebra Subroutines (BLAS)/Linear Algebra PACKage (LAPACK) for your system, as MATLAB doesn't ship the latest version in its releases.

The main choices are between:

- ATLAS - Automatically Tuned Linear Algebra Software (open source)
- MKL - Intel Math Kernel Library
- ACML - AMD Core Math Library
- GotoBLAS - GotoBLAS (Texas Advanced Computing Center)

For more information on how to upgrade your MATLAB libraries with ATLAS/MKL, see:

- http://software.intel.com/en-us/articles/using-intel-mkl-with-matlab/
- http://software.intel.com/en-us/articles/intel-math-kernel-library-intel-mkl-for-windows-using-intel-mkl-in-matlab-executable-mex-files/
- http://imaging.mrc-cbu.cam.ac.uk/imaging/SpmWithPentium4

## Parallel Computing[edit]

### SPM toolboxes[edit]

See pSPM for SPM2.

### General tools[edit]

There is currently work in progress to provide an official parallel/distributed version of SPM.

See also the Sun Grid Engine Project (SGE):

## Using the Graphics Processing Unit (GPU)[edit]

It is possible for MATLAB to take advantage of the GPU (the processor of the graphic card, by opposition to the CPU), to perform some operations, resulting in significant speed improvements. A number of toolboxes are available:

- Parallel Computing Toolbox GPU Computing with MATLAB
- GPUmat GPU toolbox for MATLAB
- AccelerEyes Jacket Software for MATLAB GPU computing

Jacket & SPM Accelerating SPM & fMRI with GPUs for Neuroimaging