Contents

Vectorizing is good.

Matlab is much faster when working with vectors than with loops. [This is not so true of all languages, in C and Julia it makes little difference.]

So whenever we can vectorize code we can make it run much faster.

The following two examples illustrate this difference.

N=10^4;

tic;
A=ones(N,N);
for a_c=1:N
    A(1:N,a_c)=a_c;
    % Matlab lets you put a scalar value into all the elements of a vector.
    % (a scalar on the RHS is about the only case in which LHS and RHS are allowed to be different sizes)
end
toc

tic;
a1=ones(1,N);
a2=1:1:N;
A=a2'*a1;
toc

Elapsed time is 0.260351 seconds.
Elapsed time is 0.081887 seconds.

With gpu it is even better.

Notice that vectorizing things, instead of using loops, often involves making all the operations parallel. This means that it works even better in combination with gpu.

tic;
a1_gpu=ones(1,N,'gpuArray');
a2_gpu=gpuArray.linspace(1,N,N); % I have switched to linspace as not sure how to use 1:1:N to generate on the gpu.
A_gpu=a2_gpu'*a1_gpu;
toc

% This example is simple enough that gpu makes little difference on
% speed, but it is still slightly faster.

Elapsed time is 0.078622 seconds.

When you work with gpus not all commands can be used.

A list of Matlab commands that can be run on gpu is: Matlab: MATLAB Functions with gpuArray Arguments

[Side note: GPU computing in Julia looks likely to be seriously impressive, but for now (August 2017) the corresponding list of functions that have been already adapted for use on gpu is small. For more see: http://julialang.org/blog/2017/03/cudanative http://mikeinnes.github.io/2017/08/24/cudanative.html Once this issue has been resolved, which one imagines it will be over the next year or two, it seems likely that Julia will provide the nicest gpu computing experience of any language.]

Published with MATLAB® R2017a