openmp spmv CSR vectorization

I've produced this kernel

struct csrFormat {
    int M, N;
    int *IRP;
    int *JA;
    double *AS;
};

struct vector {
    int dim;
    double *val;
};


double csrSIMDReduction(struct csrFormat *csr, struct vector *vec, struct vector *res, int nz){
    int M = csr->M;
    int nonzerosPerRow = nz/M;
    int chunk_size = 10000/nonzerosPerRow;
    
    double *val = res->val;
    
    struct timeval start,end;
    
    gettimeofday(&start, NULL);
    
    //IRP is the row_pointer array
    //JA is the column indices array
    //AS is the values array

    int i,j,tmp;
    #pragma omp parallel 
    {
        #pragma omp for private(i, j, tmp) schedule(dynamic, chunk_size)
        for (i=0; i<M; i++)
        {
            double result = 0.0;
            #pragma omp simd reduction(+ : result)
            for (j = csr->IRP[i]; j < csr->IRP[i+1]; j++)
            {
                tmp = csr->JA[j];
                result += csr->AS[j] * vec->val[tmp];
            }
            val[i] = result;
        }
    }
    
    gettimeofday(&end, NULL);
    
    res->dim = M;
    
    
    //printf("%ld.%06ld\n", start.tv_sec, start.tv_usec);
    //printf("%ld.%06ld\n", end.tv_sec, end.tv_usec);
    long t = (end.tv_sec - start.tv_sec)*1000000.0 + end.tv_usec - start.tv_usec;

    return (double) t;
}

but i'm having a doubt: how does the vectorization works?

I mean, from what i've understood, the inner cicle is runned running multiple iterations together, but how can a single thread do this?

EDIT: code updated.