Is there a way to improve the random sampling speed of dtmc/simulate in Matlab?

I am using the simulate function (inside the Econometrics Toolbox of Matlab) to generate random walk for a discrete time markov chain (dtmc).

However, I need to simulate a random walk with the length of 18million steps and I realize the the bottleneck is mainly caused by the existing implementation which samples one value at a time.

I read that it would be faster if I can vectorize the randsample(V, 1, true, P) call but I couldn't quite figure out the exact translation of the loop implementation to a singular randomsample call. P is my transition probability matrix with non-negative values and the sum of each row is 1. All runs start with the same initial state which is the 1st state in the X0 with the probability of 1 (and the rest of the values in the X0 are 0).

Below is the code for the standard (loop) implementation of the sampling:

%x number of steps are created (x number of walk)
X = zeros(1+numSteps,numSims);

for j = 1:numSims
    simState = find(X0~=0,1);
    X(1,j) = simState;
    
    X0(simState) = X0(simState)-1;

    for i = 2:(1+numSteps)
        u = rand;
        simState = find(u < cumsum(P(simState,:)),1);
        X(i,j) = simState;

    end

end

I tried randsample(X, 1, true, P(simState,:)); but I get the

W must have length 18000000.

error. In addition, I am not exactly sure how I can actually make sure the first element of the generated random walk is fixed and it is the first state and I just need to generate the random walk from 2nd step and on.