Alternative method needed for a few 'for' loops inside each other

Hi,

I have 4 for loops inside each other as seen here:

for A=1:a
    for B=1:b
        for C=1:c
            for D=1:d
                do_something (A,B,C,D)
            end
        end
    end
end

The values of A, B, C, D are dependent on each other so I can't calculate them separately. The issue is that this is extremely slow. Could anyone please help me with an alternative fast method? Thank you very much for your time and help.

Answer 1

[Jan]

No. It is impossible to give a useful suggestion based on this abstract code. The problem is not the nested loops, but the code hidden behind "do_something". The explanation "A, B, C, D are dependent" does not allow to understand the underlying concept.

The profiler helps to identify teh bottlenecks of the program. Use it and post the corresponding code in a less simplified manner. Then a helpful suggestion is possible.

[Mehdi]

Thank you for your reply. Sorry for oversimplification.

Here is the code:

mpd_new = 8000;
mpr1    = 40000;
aicon    = 0.2;
mstr1    = 12000;
g0       = 9.81;
Irp1     = 317;
dt       = 0.1;
for mssf1 = 1300:1400;
    for t_i = 2:0.1:4;
        for alp_max = 1:0.1:3;
            t1       = 0;
            V1_X_cut = eps;
            V1_Z_cut = 0;
            ht1      = 0;
            pcg1     = 90;
            flp1     = 90;
            alp1     = 0;
            m_in1    = mstr1+mpr1+mpd_new;
            m_f1     = m_in1-mpr1;
            Mcrnt1   = m_in1-t1*mssf1;
            tht1     = mssf1*Irp1*g0;
            while Mcrnt1 > m_f1
                Expon1   = exp(aicon*(t_i-t1));
                alp1     = -4*alp_max*Expon1*(1-Expon1);
                flp1     = atan(V1_Z_cut/V1_X_cut)*180/pi;
                pcg1     = flp1+alp1;
                pchrd1   = pcg1*pi/180;
                acc_z1   = thst1*sin(pchrd1)/Mcrnt1;
                acc_x1   = thst1*cos(pchrd1)/Mcrnt1;
                V1_X_cut = V1_X_cut+acc_x1*dt;
                V1_Z_cut = V1_Z_cut+acc_z1*dt;
                V_cr1    = sqrt(V1_X_cut^2 + V1_Z_cut^2);
                ht1      = ht1+V1_Z_cut*dt;
                Mcrnt1   = Mcrnt1-dt*mssf1;
                t1       = t1+dt;
            end
            V_cr1
        end
    end
end

The purpose of this code is to find sets of mssf1 , t_i , and alp_max that would result in V_cr1 that are about 5000. Profiler shows that all the lines in while loop are responsible for the long duration of run.

So I'm wondering if there is a way to increase speed. it's really slow. Any insight would be helpful.

[Mehdi]

Jan,

Did you get any feedback about my question?

[Stephen23]

Why not just vectorize the whole thing? There don't seem to be any slow operations involved, so it is the nested loops that are slowing this down, hence remove the loops.

[Mehdi]

I did that as you see below:

mpd_new = 8000;
mpr1    = 40000;
aicon    = 0.2;
mstr1    = 12000;
g0       = 9.81;
Irp1     = 317;
dt       = 0.1;

mssf1   = 1370:1400; % No. of elements is 31. 
t_i     = 2:0.1:5; % No. of elements is 31.
alp_max = 1:0.1:4; % No. of elements is 31.

t1       = 0;
V1_X_cut = eps;
V1_Z_cut = 0;
ht1      = 0;
pcg1     = 90;
flp1     = 90;
alp1     = 0;
m_in1    = mstr1+mpr1+mpd_new;
m_f1     = m_in1-mpr1;
Mcrnt1   = m_in1-t1.*mssf1;
thst1     = mssf1.*Irp1.*g0;

while Mcrnt1 > m_f1
    Expon1   = exp(aicon.*(t_i-t1));
    alp1     = -4.*alp_max.*Expon1.*(1-Expon1);
    flp1     = atan(V1_Z_cut./V1_X_cut).*180./pi;
    pcg1     = flp1+alp1;
    pchrd1   = pcg1.*pi./180;
    acc_z1   = thst1.*sin(pchrd1)./Mcrnt1;
    acc_x1   = thst1.*cos(pchrd1)./Mcrnt1;
    V1_X_cut = V1_X_cut+acc_x1.*dt;
    V1_Z_cut = V1_Z_cut+acc_z1.*dt;
    V_cr1    = sqrt(V1_X_cut.^2 + V1_Z_cut.^2);
    ht1      = ht1+V1_Z_cut.*dt;
    Mcrnt1   = Mcrnt1-dt.*mssf1;
    t1       = t1+dt;
end
V_cr1

But apparently, the while loop runs only once. Obviously it should run the while loop 31*31*31=29791 times for each element of each vector. Any idea how this should be implemented?