concurrent_vector is not working inside parallel_for ( PPL )

Question

there is a sample working code below ( parallel_for using Parallel Pattern Library ( ppl ) ). The main problem in here is sqr < concurrent_vector > stored values changing in every execution, but it should not be!

I used < concurrent_vector > for random access why it's not working?

#include <iostream>
#include <ppl.h>
#include <concurrent_vector.h>

using namespace std;
using namespace concurrency;

const int a = 10, b = 30;

critical_section cs;

int main() {

    concurrent_vector< int > labels( a * b );

    concurrent_vector< int > sqr( 5 );

    // filling label vector
    for ( int y = 0; y < b; y++ ) {
        for ( int x = 0; x < a; x++ ) {

            if( x<2 && y>3 )
                labels[ a * y + x ] = 1;
            else if( x<30 && y<5 )
                labels[ a * y + x ] = 2;
            else if( x>5 && y>10 )
                labels[ a * y + x ] = 3;
            else if( x>2 && y>20 )
                labels[ a * y + x ] = 4;
        }
    }

    // printing
    for ( int y = 0; y < b; y++ ) {
        for ( int x = 0; x < a; x++ ) {

            cout << labels[ a * y + x ] << ", ";
        }
        cout << endl;
    }

    parallel_for ( 0, b, [ & ]( int y ) {
        for ( int x = 0; x < a; x++ ) {

            //cs.lock();  // when i used it's working but slow
            int i = labels[ a * y + x ];
            //cs.unlock();

            if ( i < 0 ) continue;

            sqr[ i ] ++;
        }
    } );

    for( int i=0; i<5; i++ )
        cout << sqr[i] << ", ";
    cout << "" << endl;

    system ("pause");

    return 0;
}

Answer 1

Using task_group::wait method should be faster (as you don't have to lock/unlock every time) and it may work as you expect.

This method blocks the current task until the tasks of another task group have completed their work.

See MSDN: Parallel Tasks.

Update: I have run some timing tests and seems that this is not a solution (besides both fail on large data inputs on my Dual-Core). This can be a bug of "design' in concurrent_vector" as in Intel's TBB - tbb::concurrent_vector returns wrong size