Dynamically Readjustable Arrays and OpenMP - c++

Question

I have a function that uses realloc to dynamically adjust the memory of a 1D array as the initial size of the array cannot be predetermined .I want to parallelize this code by dividing the task across multiple threads whereby each thread would work on a smaller 1D array that they would dynamically readjust according to the memory required. As part of the process each thread also produces a private variable that would contain the final size of the small array.
In Openmp I want to access the private copy of these arrays (through the master thread) and put all the small arrays together to obtain a final array based on the size of the arrays computed by each thread.
Is it possible ??

Answer 1

That can be done with a dynamic array such as std::vector. For example let's assume that you have an array called data with values between zero and one and n elements and you want to select out values greater that 0.5 and store them in a new array vec. You could do exactly what you want like this
double data[n];
for(int i=0; i<n; i++) data[i] = 1.0*rand()/RAND_MAX;
std::vector<double> vec;
#pragma omp parallel
{
std::vector<double> vec_private;
#pragma omp for nowait
for(int i=0; i<n; i++) {
if(data[i]>0.5) vec_private.push_back(data[i]);
}
#pragma omp critcal
vec.insert(vec.end(), vec_private.begin(), vec_private.end());
}
To do this without a critical section requires a bit more work. It requires saving an array of the size for each array and them doing a cumulative sum (aka prefix sum) on that array in a single section. Once we have the cumulative sum we can use it to merge the arrays in parallel.
int *sizea;
#pragma omp parallel
{
int nthreads = omp_get_num_threads();
#pragma omp single
{
sizea = new int [nthreads+1];
sizea[0] = 0;
}
std::vector<double> vec_private;
#pragma omp for schedule(static) nowait
for(int i=0; i<n; i++) {
if(data[i]>0.5) vec_private.push_back(data[i]);
}
sizea[omp_get_thread_num()+1] = vec_private.size();
#pragma omp barrier
#pragma omp single
{
int size = 0;
for(int i=0; i<nthreads+1; i++) {
size += sizea[i];
sizea[i] = size;
}
vec.resize(size);
}
std::copy(vec_private.begin(), vec_private.end(), vec.begin()+sizea[omp_get_thread_num()]);
}
delete[] sizea;