Распараллеливание программы с библиотекой TBB

const int N = 25000000;
float a[N];
 
 
void fill(int n){
    a[n] = sin((double)n);
}
 
template<class T>
void merge(T a[], long lb, long split, long ub) {
    long pos1 = lb;
 
    long pos2 = split + 1; 
    long pos3 = 0;
 
    T *temp = new T[ub - lb + 1];
 
    while (pos1 <= split && pos2 <= ub) {
        if (a[pos1] < a[pos2])
            temp[pos3++] = a[pos1++];
        else
            temp[pos3++] = a[pos2++];
    }
 
    while (pos2 <= ub)   
        temp[pos3++] = a[pos2++];
    while (pos1 <= split)  
        temp[pos3++] = a[pos1++];
 
    for (pos3 = 0; pos3 < ub - lb + 1; pos3++)
        a[lb + pos3] = temp[pos3];
 
    delete []temp;
}
 
 
 
template<class T>
struct SortParallelMerge : public task {
    long m_lb;
    long m_ub;
    T *m_arr;
    
    SortParallelMerge(T  arr[], long lb, long ub) :
        m_lb(lb), m_ub(ub), m_arr(arr)
    {
 
    }
    task * execute() {
        long split = (m_lb + m_ub) / 2;
        if (m_lb < m_ub) {                
            SortParallelMerge sort_left = *new(allocate_child()) SortParallelMerge(m_arr, m_lb, split);
            SortParallelMerge sort_right = *new(allocate_child())SortParallelMerge(m_arr, split + 1, m_ub);
            set_ref_count(3);
            spawn(sort_left);  
            spawn_and_wait_for_all(sort_right); 
            merge(m_arr, m_lb, split, m_ub);    
        }
    }
 
};
 
void mergeParallelSort()
{
    task_scheduler_init init;
    tick_count parallel_start = tick_count::now();
    int t;
    int f;
    SortParallelMerge<float> & sortParal = *new(task::allocate_root())SortParallelMerge<float>(a, t, f);
     
    task::spawn_root_and_wait(sortParal); 
    tick_count parallel_end = tick_count::now();
    printf("Time of parallel sort =  %f\n", (parallel_end - parallel_start).seconds());
}
 
void FillArray()
{
    tick_count parallel_start = tick_count::now();
    parallel_for<size_t>(1, N, 1, fill);
    tick_count parallel_end = tick_count::now();
    printf("Fill array parallel sort =  %f\n", (parallel_end - parallel_start).seconds());
}
 
 
int main()
{
    FillArray();
    mergeParallelSort();
    
}

#include <string>
#include <sstream>
#include <iostream>
#include <tbb/task_scheduler_init.h>
#include <tbb/blocked_range.h>
#include <math.h>
#include <tbb/parallel_sort.h>
#include <tbb/parallel_for.h>
#include <time.h>
#include <iomanip>      // std::setprecision
#include <tbb/tbb.h>
#include <assert.h>
#include <algorithm>
using namespace tbb;
 
const int N = 25000000;
float a[N];
 
 
void fill(int n){
    a[n] = sin((double)n);
}
 
template<class T>
void merge(T a[], long lb, long split, long ub) {
    long pos1 = lb;
 
    long pos2 = split + 1; 
    long pos3 = 0;
 
    T *temp = new T[ub - lb + 1];
 
    while (pos1 <= split && pos2 <= ub) {
        if (a[pos1] < a[pos2])
            temp[pos3++] = a[pos1++];
        else
            temp[pos3++] = a[pos2++];
    }
 
    while (pos2 <= ub)   
        temp[pos3++] = a[pos2++];
    while (pos1 <= split)  
        temp[pos3++] = a[pos1++];
 
    for (pos3 = 0; pos3 < ub - lb + 1; pos3++)
        a[lb + pos3] = temp[pos3];
 
    delete []temp;
}
 
 
 
template<class T>
struct SortParallelMerge : public task {
    long m_lb;
    long m_ub;
    T *m_arr;
    
    SortParallelMerge(T  arr[], long lb, long ub) :
        m_lb(lb), m_ub(ub), m_arr(arr)
    {
 
    }
    task * execute() {
        long split = (m_lb + m_ub) / 2;
        if (m_lb < m_ub) {                
            SortParallelMerge sort_left = *new(allocate_child()) SortParallelMerge(m_arr, m_lb, split);
            SortParallelMerge sort_right = *new(allocate_child())SortParallelMerge(m_arr, split + 1, m_ub);
            set_ref_count(3);
            spawn(sort_left);  
            spawn_and_wait_for_all(sort_right); 
            merge(m_arr, m_lb, split, m_ub);    
        }
    }
 
};
 
void mergeParallelSort()
{
    task_scheduler_init init;
    tick_count parallel_start = tick_count::now();
    int t;
    int f;
    SortParallelMerge<float> & sortParal = *new(task::allocate_root())SortParallelMerge<float>(a, t, f);
     
    task::spawn_root_and_wait(sortParal); 
    tick_count parallel_end = tick_count::now();
    printf("Time of parallel sort =  %f\n", (parallel_end - parallel_start).seconds());
}
 
void FillArray()
{
    tick_count parallel_start = tick_count::now();
    parallel_for<size_t>(1, N, 1, fill);
    tick_count parallel_end = tick_count::now();
    printf("Fill array parallel sort =  %f\n", (parallel_end - parallel_start).seconds());
}
 
 
int main()
{
    FillArray();
    mergeParallelSort();
    
}