OpenMP critical section matrix multiplication

Question

First of all I know the example is not a noticeable improvement from a single threaded execution. I am learning the basic of OpenMP and I dont know why it doesnt work as I expected.

So, I am parallelizing the traditional matrix multiplication algorithm and finding what loops can be parallelized and which not. The algorithm is:

for (i = 0; i < SIZE; i++)
    for (j = 0; j < SIZE; j++)
        for (k = 0; k < SIZE; k = k + 1)
            C[i][j] = C[i][j] + A[i][k] * B[k][j];

I know that putting a parallel for on the k loop can cause unexpected results so I want to use a critical section to fix it like this:

for (i = 0; i < SIZE; i++)
    for (j = 0; j < SIZE; j++)
        #pragma omp parallel for shared(C)
        for (k = 0; k < SIZE; k = k + 1)
        {
            int tmp = A[i][k] * B[k][j];
            #pragma omp critical 
            {
                C[i][j] += tmp;
            }
        }
}

But when comparing results with the code below correct gives me 0. D is the matrix without omp. I know there are better ways to do this with reduction but I am learning about critical sections and want to know why this doesnt work.

int correct = 1;
    for (i = 0; i < SIZE && correct; i++)
        for (j = 0; j < SIZE && correct; j++)
            if (C[i][j] != D[i][j])
                correct = 0;

Answer 1

There are two fairly basic bugs in my code.

• The tmp var is int when it should be double.
• After changing tmp to double. The comparison between two doubles it is not associative and an epsilon should be taken in consideration when comparing.