MATLAB parfor is slower than for -- any help to speed up of my loop cacluations within KDL function?

Question

Mehdi 2024-1-11

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2069016-matlab-parfor-is-slower-than-for-any-help-to-speed-up-of-my-loop-cacluations-within-kdl-function

评论： Torsten 2024-1-13

II=9;
JJ=9;
M=11;
W=rand(II+1,JJ+1,3,M);
k_d=01e12;
syms x y
for i=1:II+1
    for j=1:JJ+1
        lgij(i,j) = legendreP(i-1, x)*legendreP(j-1, y);
    end
end 
wxy2 = sym('wxy2',[1 M]);
wxy3 = sym('wxy3',[1 M]);
wxy2(1:M) = sym('0');
wxy3(1:M) = sym('0');  
for r=1:M
    for i=1:II+1
        for j=1:JJ+1
            wxy2(r) = W(i, j, 2, r)*lgij(i,j) + wxy2(r); 
            wxy3(r) = W(i, j, 3, r)*lgij(i,j) + wxy3(r);
        end
    end
end 
wxxyy2=simplify(wxy2);
wxxyy3=simplify(wxy3);
for r=1:M
    for o=1:M
        Wijklmo(r,o)= expand(wxxyy2(r)*wxxyy2(o) + wxxyy3(r)*wxxyy3(o)- wxxyy2(r)*wxxyy3(o) - wxxyy2(o)*wxxyy3(r));
    end
end
dt = 3.8655e-08;
T=1e9*dt;
q = zeros(M, floor(taim/dt)+1);
RA=111;
s=1;
for tn=t__0:dt:T
    p_hat(1:M,s) = KDL(wxxyy2, wxxyy3, M, q, RA, k_d, Wijklmo, s)
    s=s+1
end

a

function kdl = KDL( wxxyy2, wxxyy3, M, q, RA, k_d, Wijklmo, s)
xx = sym('xx');
yy = sym('yy');
Wxy2 = sym('0');
Wxy3 = sym('0');
for r=1:M
    Wxy2 = wxxyy2(r)*q(r, s) + Wxy2;
    Wxy3 = wxxyy3(r)*q(r, s) + Wxy3;
end 
w23 = expand(Wxy3-Wxy2);
kdl=zeros(M,1);
xspan = [-1 1];
yspan = [-1 1];
Gy0 = 0;
H(xx,yy) =(0.5*(1+tanh(k_d*RA*w23)));
g = matlabFunction(H,'Vars',[xx yy]);
if w23==0
   kdl(:,1) =0;
else
   tic
   for i=1:M
   %parfor i=1:M    
       ff(i,1)= vpa(Wijklmo(i,:)*q(1:M, s));
       f = matlabFunction(ff(i,1),'Vars',[xx yy]);
       D = @(xx,yy)f(xx,yy).*(g(xx ,yy)>0);
       if w23==0
          kdl(i,1) =0;
       else
       [~,G] = ode45(@(y,Gy)fun(y,Gy,D,xspan),yspan,Gy0,odeset('RelTol',1e-10,'AbsTol',1e-10));
       kdl(i,1) = k_d*real(G(end));
   end
end
toc
end

a

function dGydy = fun(y,Gy,g,xspan)
  % Compute the x-integrals at y = y
  Gx0 = 0;
  [~,Gx] = ode45(@(x,~)g(x,y),xspan,Gx0,odeset('RelTol',1e-10,'AbsTol',1e-10));
  dGydy = Gx(end);
end

4 个评论
显示 2更早的评论隐藏 2更早的评论

Mehdi 2024-1-13

I did this, but did not speed-up. The problem is that parfor is slower than for.

Torsten 2024-1-13

Ok, then I think @Matt J is correct: the gain in speed is overcompensated by the overhead of parallel computing.

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Matt J 2024-1-12

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2069016-matlab-parfor-is-slower-than-for-any-help-to-speed-up-of-my-loop-cacluations-within-kdl-function#answer_1388206

编辑：Matt J 2024-1-12

在 MATLAB Online 中打开

M=11 is a rather small number of iterations. If this is a realistic value, there may not be enough iterative work to make the overhead of parfor worthwhile. It might be better to move the parfor to the outer loop that calls KDL.

Tn=t__0:dt:T; %where are these values used???
parfor s=1:numel(Tn)
    
    p_hat(:,s) = KDL(wxxyy2, wxxyy3, M, q, RA, k_d, Wijklmo, s);
    
end

A few other remarks,

(1) Things like this loop

for r=1:M
    Wxy2 = wxxyy2(r)*q(r, s) + Wxy2;
    Wxy3 = wxxyy3(r)*q(r, s) + Wxy3;
end 

can be replaced with vectorized statements,

 Wxy2 = wxxyy2*q(:, s);
 Wxy3 = wxxyy3*q(:, s);

(2) Also, q(:, s) will probably evaluate faster than q(1:M,s).

3 个评论
显示 1更早的评论隐藏 1更早的评论

Matt J 2024-1-12

编辑：Matt J 2024-1-12

在 MATLAB Online 中打开

You should pre-allocate p_hat

M=11;
N=5;
p_hat=zeros(M,N);
parfor s=1:N
    
    p_hat(:,s) = KDL(s);
    
end
Starting parallel pool (parpool) using the 'Processes' profile ...
Parallel pool using the 'Processes' profile is shutting down.
p_hat
p_hat = 11×5
     1     2     3     4     5
     1     2     3     4     5
     1     2     3     4     5
     1     2     3     4     5
     1     2     3     4     5
     1     2     3     4     5
     1     2     3     4     5
     1     2     3     4     5
     1     2     3     4     5
     1     2     3     4     5
function s=KDL(s)
end

Mehdi 2024-1-12

编辑：Mehdi 2024-1-12

this method does not work for me since in my original problem KDL(s+1)=f(KDL(s)).(because of complexities I have not brought this part of my code here.

请先登录，再进行评论。