How can i avoid a large calculation error when i try to optimize my Gauss–Seidel program?

Question

#include <cstdint>
#include <cstdlib>
#include <cstring>
#include <omp.h>
#include <algorithm>

/**
 * Pipeline-parallel implementation of Gauss-Seidel method
 * Maintains original function signature while implementing wavefront parallelism
 */
void my_seidel(int n, int T, double *A)
{
    // Determine number of threads available
    int num_threads = omp_get_max_threads();
    omp_set_num_threads(num_threads);
    
    for (int t = 0; t < T; t++) {
        // Number of diagonals in an (n-2)x(n-2) interior grid is 2*(n-2)-1
        const int num_diagonals = 2 * (n - 2) - 1;
        
        // Process diagonals in sequence
        for (int d = 0; d < num_diagonals; d++) {
            // Determine start and end points for this diagonal
            int start_i = std::max(1, d - (n - 3));
            int end_i = std::min(n - 2, d + 1);
            
            // Process cells along this diagonal in parallel
            #pragma omp parallel for schedule(dynamic)
            for (int i = start_i; i <= end_i; i++) {
                int j = d - (i - 1) + 1;
                if (j >= 1 && j < n - 1) {
                    // Use the most up-to-date values (already computed in this iteration)
                    A[i * n + j] = (A[(i - 1) * n + j - 1] + A[(i - 1) * n + j] + A[(i - 1) * n + j + 1]
                                + A[i * n + j - 1] + A[i * n + j] + A[i * n + j + 1]
                                + A[(i + 1) * n + j - 1] + A[(i + 1) * n + j] + A[(i + 1) * n + j + 1])
                                / 9.0;
                }
            }
            // Implicit barrier at the end of parallel region ensures all
            // updates in this diagonal are complete before proceeding
        }
    }
}

I found that once I used OMP to parallelize the process, the calculation went wrong. I have tried a lot of things to avoid the error but they didn't work. Could anyone give some suggestions to optimize the code without losing accuracy?