Is there any standard way to pre-initialize the PPL thread pool? The problem is: PPL creates its thread pool at runtime when e.g. parallel_for() is executing. This costs a little performance during the first run, due to creation of additional threads.
For the clarification of the problem, here is an example:
#pragma once
#include <vector>
#include <ppl.h>
using namespace std;
using namespace Concurrency;
// Use this define for experiments.
#define PPL_INIT_METHOD 2
int main()
{
#if (PPL_INIT_METHOD == 0) // experiment 1: initialize default scheduler
CurrentScheduler::Create(SchedulerPolicy());
// After this call only one additional thread is created
#elif (PPL_INIT_METHOD == 1) // experiment 2: initialize custom scheduler
SchedulerPolicy my_policy(3,
MinConcurrency, 12,
MaxConcurrency, 12,
ContextPriority, THREAD_PRIORITY_NORMAL);
Scheduler* my_scheduler = Scheduler::Create(my_policy);
// After this call only one additional thread is created
my_scheduler->Attach();
assert(my_scheduler->GetNumberOfVirtualProcessors() == 12);
// Still same number of threads (= 2)
// cleanup stuff ...
#else // experiment 3: execute dummy parallel_for()
std::vector<int> v(1024*1024, 42);
parallel_for(0u, v.size(), [&v](size_t i) { v[i] += 1; }, static_partitioner());
// After this call all 12 threads are created and ready for more work!
#endif
// Do real work now!!!
return 0;
}
