Speed performance between class, struct and local variable

Here is the illustration of my problem :

% Class
classdef MaClasse < handle
    properties (Access = public)
        vec1 = [];
        vec2 = [];
        vec3 = [];
    end
end

...

    nb = 50000;
randstream = Randstream('mrg32k3a');

% Cas 1
pObj = MaClasse();
pObj.vec1 = randn(randstream,100,1000);
pObj.vec2 = randn(randstream,100,1000);
pObj.vec3 = randn(randstream,100,1000);
tic
for i=1:nb
    pObj.vec1 = pObj.vec1.*pObj.vec2 + pObj.vec3;
end
toc
% ----> 6.80 seconds

% Cas 2
vec1 = randn(randstream,100,1000);
vec2 = randn(randstream,100,1000);
vec3 = randn(randstream,100,1000);
tic
for i=1:nb
    vec1 = vec1.*vec2 + vec3;
end
toc
% ----> 1.61 seconds

% Cas 3
str.vec1 = randn(randstream,100,1000);
str.vec2 = randn(randstream,100,1000);
str.vec3 = randn(randstream,100,1000);
tic
for i=1:nb
    str.vec1 = str.vec1.*str.vec2 + str.vec3;
end
toc
% ----> 6.95 seconds

A difference of factor 4 seems critical to use Matlab with POO (even with struct according to my test)

If someone has a solution but not to say using local variable and then copy into attributs ;-)

Answer 1

[Matt J]

Well Case 1 and Case 3 have more indexing operations, so it makes sense that that overhead should dominate assuming the + and .* operations are sufficiently fast. That seems to be the case, because below we see that once the data size is sufficiently large, the relative differences in performance diminish. It just shows how fast .* and + are for small/medium data sizes!

rs = RandStream('mrg32k3a');

M=5000; N=5000;

% Cas 1
pObj = MaClasse();
pObj.vec1 = randn(rs,M,N);
pObj.vec2 = randn(rs,M,N);
pObj.vec3 = randn(rs,M,N);

str.vec1 = randn(rs,M,N);
str.vec2 = randn(rs,M,N);
str.vec3 = randn(rs,M,N);

vec1 = randn(rs,M,N);
vec2 = randn(rs,M,N);
vec3 = randn(rs,M,N);


timeit( @() pObj.vec1.*pObj.vec2 + pObj.vec3 )

ans = 0.0732

timeit( @() str.vec1.*str.vec2 + str.vec3 )

ans = 0.0726

timeit( @() vec1.*vec2 + vec3 )

ans = 0.0663

[Franck AUBINEAU]

"Well Case 1 and Case 3 have more indexing operations, so it makes sense that that overhead should dominate assuming the + and .* operations are sufficiently fast."

It means that if you're using class in a complex simulation calculation code with Matlab, you have to accept that it will be less efficient several simple assembled code witj local variable unless your data are sufficiently large or adapted to your code.

I thought that, with pre allocation, there won't be such a difference.

[Matt J]

It means that if you're using class in a complex simulation calculation code with Matlab, you have to accept that it will be less efficient

I don't think it means it's less efficient just because it takes longer. It takes longer because you are giving it more work.