From CPU code to GPU

Hello Everybody

I have an heavy code, it's a Split-Step Fourier program, with different external function.

The main vector (the propagating one) is 2^16 to 2^18 point. It currently work as a classic CPU program.

My computer has a CUDADevice with properties: Name: 'Quadro RTX 4000'. So I try to tranlate my program.

Before that I test the speed of the GPU vs CPU with some code I found here. And It's faster, Perfect

To translate my program I do this modification

A=gpuArray(A);
B=gpuArray(B);

And I try to do it for the most variable I have, it's quite long and painfull

I also do it in the function I need to use, to be sure that the most part of the variable are GpuArray

BUT, it's slower...

I never use GPU before, so I guess I am doing something wrong..

Thanks for your time

Martin

Answer 1

[MartinM]

for exemple

clc,clear all
close all
%% CPU
tic
num.n=1*2^18;
num.tspan = 2e-09;
num.dt=num.tspan/(num.n-1);
T = zeros(1, num.n);
for k=1:1:num.n
    T(k)=(k-1)*num.dt-num.tspan/2;
end
toc

%% GPU 
clear all
close all
tic
num.n=gpuArray(1*2^18);
num.tspan =gpuArray(2e-09);
num.dt=gpuArray(num.tspan/(num.n-1));
T = gpuArray(zeros(1, num.n));
for k=1:1:num.n
    T(k)=(k-1)*num.dt-num.tspan/2;
end
toc

the result is

Elapsed time is 0.027964 seconds.

Elapsed time is 47.296786 seconds.

Answer 2

[Joss Knight]

You need to vectorize your code. The GPU is not intended for performing this kind of looping series of operations on scalar variables. For instance use

k = gpuArray(1:num.n);
T=(k-1)*num.dt-num.tspan/2;

to compute every element of T at once.

You do not need to convert every variable to a gpuArray, just your inputs.