int mscore_c_omp( SV* str ) {
    STRLEN len;
    char* buf = SvPVbyte( str, len );
    len --;
    
    int nparts = omp_get_num_procs();
    if ( nparts > 4 ) nparts = 4;
    int psize = len / nparts;
    
    int acc_a[ nparts ];
    int cum_a[ nparts ];
    int max_a[ nparts ];
    
    #pragma omp parallel for schedule( static, 1 ) \
                             num_threads( nparts )
    for ( int j = 0; j < nparts; j ++ ) {
        int lo = j * psize;
        int hi = ( j == nparts - 1 ) ? len : lo + psize;
        
        int acc = 0;
        int cum = 0;
        int max = -1;
        
        for ( int i = lo; i < hi; i ++ ) {
            int one = buf[ i ] - '0';
            acc += one;
            cum += one ? -1 : 1;
            if ( cum > max ) max = cum;
        }
        acc_a[ j ] = acc;
        cum_a[ j ] = cum;
        max_a[ j ] = max;
    }
    
    int acc = acc_a[ 0 ]; 
    int max = max_a[ 0 ];
    
    for ( int j = 1; j < nparts; j ++ ) {
        acc += acc_a[ j ];
        
        cum_a[ j ] += cum_a[ j - 1 ];
        max_a[ j ] += cum_a[ j - 1 ];
        
        if ( max_a[ j ] > max ) max = max_a[ j ];
    }
    
    return acc + max + ( buf[ len ] == '1' );
}