MATLAB GPU Tutorial

Run a MATLAB script that performs a regression test comparing the computation time of a non-GPU-based computation with that of a GPU-based computation.

MATLAB Script

% gpu-bench.m

for i = 300:100:700

    fprintf('i = %d\n', i);

    % Generates a random matrix X of size 1000i by 800 and a vector Y of size 1000i by 1

    X = random('unif', 0, 1, 1000*i, 800);

    Beta = [1:801]';

    allones = ones(1000*i, 1);

    % Add a constant term

    X = [allones, X];

    Y = X * Beta + random('norm', 0, 1, 1000*i, 1);

    % Non-GPU computation

    disp('Time for non-GPU:');

    tic;

    Bestimate = inv(X'*X) * (X'*Y);

    Bestimate = inv(X'*X) * (X'*Y);

    Bestimate = inv(X'*X) * (X'*Y);

    Bestimate = inv(X'*X) * (X'*Y);

    toc;

    % Now do it on the GPU

    % Copy the data to the GPU and create gX and gY in the GPU

    disp('Load time for GPU:');

    tic;

    gX = gpuArray(X);

    gY = gpuArray(Y);

    gBestimate = gpuArray(zeros(801, 1));

    toc;

    % GPU computation

    disp('Time for GPU compute:');

    tic;

    gBestimate = inv(gX'*gX) * (gX'*gY);

    gBestimate = inv(gX'*gX) * (gX'*gY);

    gBestimate = inv(gX'*gX) * (gX'*gY);

    gBestimate = inv(gX'*gX) * (gX'*gY);

    toc;

    % Copy estimates back

    disp('Copy estimates back:');

    tic;

    Bestimate = gather(gBestimate);

    toc;

end

The script generates a random matrix X of size 1000i by 800 and a vector Y of size 1000i by 1. 

It then adds a column of ones to X to account for the intercept term in the regression model.

Next, it computes the regression coefficients (Bestimate) using the ordinary least squares (OLS) method for the non-GPU-based computation, and it computes the regression coefficients (gBestimate) using the OLS method on the GPU.

Slurm Script

#!/bin/bash

#

# [ gpu-bench.sbatch ]

#SBATCH --job-name=gpu-bench

#SBATCH --output=gpu-bench.%j.out

#SBATCH --export=ALL

#SBATCH --gpu-bind=closest

#SBATCH --time=00:9:00

#SBATCH --mem=32G

#SBATCH --mail-type=END,FAIL

#SBATCH --partition=gpu

#

# `--partition=gpu` specifies that we want to run it on the GPU node

#

# This script runs a MATLAB program saving the output into a files called:

# gpu-bench_#.out, where the # represents the jobID.  See below.

#

#

#

# You SUBMIT this 'gpu-bench.sbatch' script to the grid with the  sbatch  command:

#

#

#     sbatch  gpu-bench.sbatch

#

# Specify the software and version to use, and run the MATLAB program:

  echo "Running on node: `hostname`"

  echo ""

  module  purge

  module  load    matlab/2019a

  matlab -nojvm < gpu-bench.m

You can run the following command to execute the script:

sbatch gpu-bench.sbatch