How to parallelise sections of my for loop

Hello,

I am trying to think of a way to make my for loop quicker. It has occurred to me that I could parallelise the commands in the loops and then bring them back together at the end. Is there such a thing in matlab or am I overcomplicating everything?

My MWE is:

for kx = 1:length(w) 
    for ky = 1:length(w) %These could be linearised
        for kkx = 1:length(w)
            for kky = 1:length(w)
                for freq_sign_ansatz = 1:2 
                    for freq_sign_test = 1:2
                        [pqf, v1qf, v2qf, pw1f, pw2f, v1w1f, v1w2f, v2w1f, v2w2f] = Convection_basis(kx,ky,kkx,kky,w, freq_ansatz, freq_test);
                        pq_freq(freq_sign_ansatz, freq_sign_test) = pqf(kx,ky,kkx,kky,freq_ansatz, freq_test);
                        v1w1_freq(freq_sign_ansatz, freq_sign_test) = v1w1f(kx,ky,kkx,kky,freq_ansatz, freq_test);
                        v1w2_freq(freq_sign_ansatz, freq_sign_test) = v1w2f(kx,ky,kkx,kky,freq_ansatz, freq_test);
                        v2w1_freq(freq_sign_ansatz, freq_sign_test) = v2w1f(kx,ky,kkx,kky,freq_ansatz, freq_test);
                        v2w2_freq(freq_sign_ansatz, freq_sign_test) = v2w2f(kx,ky,kkx,kky,freq_ansatz, freq_test);
                        pw1_freq(freq_sign_ansatz, freq_sign_test) = pw1f(kx,ky,kkx,kky,freq_ansatz, freq_test);
                        pw2_freq(freq_sign_ansatz, freq_sign_test) = pw2f(kx,ky,kkx,kky,freq_ansatz, freq_test);
                        v1q_freq(freq_sign_ansatz, freq_sign_test) = v1qf(kx,ky,kkx,kky,freq_ansatz, freq_test);
                        v2q_freq(freq_sign_ansatz, freq_sign_test) = v2qf(kx,ky,kkx,kky,freq_ansatz, freq_test);
                     end
                end 
                index = [(kx - 1) * Length_Wavenumbers + ky, (kkx - 1) * Length_Wavenumbers + kky];
                if w(kkx) == w(kx) && w(kx) == 0 && (w(kky) == w(ky) && w(kky) == 0) && zero_counter == 0
                    zero_counter = 1;
                    zero_entries(zero_counter) = {index};
                end
                pq{index(1), index(2)} = pq_freq; %pd_dxw
                v1w1{index(1), index(2)} = v1w1_freq; %pd_dxw
                v1w2{index(1), index(2)} = v1w2_freq; %pd_dxw
                v2w1{index(1), index(2)} = v2w1_freq; %pd_dxw
                v2w2{index(1), index(2)} = v2w2_freq; %pd_dxw
                
                pw1{index(1), index(2)} = pw1_freq; %pd_dxw
                pw2{index(1), index(2)} = pw2_freq; %pd_dxw
                v1q{index(1), index(2)} = v1q_freq; %vd_dxq
                v2q{index(1), index(2)} = v2q_freq; %vd_dxq
             end
        end
    end
end

Clearly here I could linearise some of the for loops through combinations of kx,ky,kkx,kky but I was interested whether there is a way of parallelising it as each pw1 etc are square matrices which are independent of each other before I combine them into a cell.

I am new to the way of parallelising things.

Answer 1

[Gaurav Garg]

Hi,

Rather than using 4-6 for loops, you can try making a single for loop traversing through all the elements.

For example -

for i = 1:2
    for j = 1:2
        %do anything
    end
end

can be rewritten as (using parfor loop)

parfor i=1:4
    [r,c] = ind2sub([2 2],i)
    disp(r);
    disp(c);
end

Here, ind2sub function has been used to convert the single iterator into r and c which are the iterators i and j in the first case.

Similarly you can rewrite your loops as (not tested)-

parfor x = 1:length(w)*length(w)*length(w)*length(w)
    [kx, ky, kkx, kky] = ind2sub([length(w) length(w) length(w) length(w)],x);
end

[KieranSQ]

Thank you! I will give this a go :)