How to write correctly a multiclass perceptron

I am learning Pattern Recognition and Machine Learning and I was required to write a multiclass perceptron like this:

there are vectors: v_i = [v_i0, v_i1, v_i2, ..., v_im], i = (1, 2, ..., n)

write them to augmented style: x_i = [v_i0, v_i1, v_i2, ..., v_im, 1]

initial value w_i(0) = [0 for _ in range(m)], C

for iteration ITER:    for i in range(n):    step = k = ITER * n + i    use x_i as learning example:    calculate d_j = w_j(k) * x_i, j = (1, 2, ..., n)    if d_i(k) > d_j(k) is TRUE for any (j != i):        w_j(k+1) = w_j(k)    else if exists l such that d_i(k) <= d_l(k):        w_i(k+1) = w_i(k) + C * x_i        w_l(k+1) = w_l(k) - C * x_i        w_j(k+1) = w_j(k) for any (j != i,l)

Then I have a question that:

if there are N vectors w_l(k) such that d_i(k) <= d_l(k)

should I do w_i(k+1) = w_i(k) + C * x_i once only when all comparations between d_i and d_j is over so that w_i(k+1) = w_i(k) + C * x_i

like this:

        PUNISH = 0        for _j in range(1,n):            j = (i + _j) % n            if d[i] > d[j]:                COUNT += 1            else:                w[j] -= C * x[i]                PUNISH += 1        if PUNISH > 0:            w[i] += C * x[i]

or I do w_i(k+1) = w_i(k) + C * x_i each time d_i(k) <= d_l(k) so that w_i(k+1) = w_i(k) + N * C * x_i

        for _j in range(1,n):            j = (i + _j) % n            if d[i] > d[j]:                COUNT += 1            else:                w[i] += C * x[i]                w[j] -= C * x[i]