Table des matières

Using GPU in MATLAB

In this lab, we will use GPUs with MATLAB to run fftw calculations. First of all we use the parallel toolbox directly from matlab. Then we use the Matlab compiler to stop consuming tokens (only one token is available for the parallel toolbox).

This simple example will be used:

fftgpu.m
function fftgpu()
 
Nx=1024; Ny=1024;
 
[x,y]=ndgrid(-Nx/2:Nx/2-1,-Ny/2:Ny/2-1);
r=x+1i*y;   ph=angle(r);    r=abs(r);
 
gpu=gpuDevice();
 
tic;
% init m in GPU
m=gpuArray(double(exp(-(r/64).^2)));
% run FFT on GPU
m=fftshift(fft(m));
% gather m from GPU to CPU
 
toc;
 
m=gather(m);
 
% save matrix
save('saveM.mat','m');
 
exit();
end
Notice that the code is enclosed in function to be able to use the matlab compiler later

Using Matlab directly

prerequisites

Here is the SGE script:

fftmatlab1.sge
#!/bin/bash -l
 
#$ -N fft_matlabGPU
#$ -o $JOB_NAME.$JOB_ID.out
#$ -q tesla.q  ## request tesla
#$ -l h_vmem=20G ## we request more because of Java
 
module load gpu/cuda/8.0    
module load matlab/r2017b 
 
matlab -nodisplay -r "fftgpu"

submit the job with:

$ qsub fftmatlab1.sge

Using Matlab compiler

We use compiler to avoid using parallel tokens ( 1 token available!)

MATLAB compiler will generate executable and scripts files in the same directory.

here is the SGE script:

fftmatlab2.sge
#!/bin/bash -l
 
#$ -N fft_matlabGPUCompiler
#$ -o $JOB_NAME.$JOB_ID.out
#$ -q tesla.q  ## request tesla
#$ -l h_vmem=20G ## we request more because of Java
 
module load gpu/cuda/8.0    
module load matlab/r2017b 
mcc -m fftgpu.m
./run_fftgpu.sh $MATLAB_HOME

submit the job with:

$ qsub fftmatlab2.sge