TThreadExecutor.cxx

Go to the documentation of this file.

#include "ROOT/TThreadExecutor.hxx"
#include "ROpaqueTaskArena.hxx"
#if !defined(_MSC_VER)
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wshadow"
#endif
#include "tbb/tbb.h"
#define TBB_PREVIEW_GLOBAL_CONTROL 1 // required for TBB versions preceding 2019_U4
#include "tbb/global_control.h"
#if !defined(_MSC_VER)
#pragma GCC diagnostic pop
#endif
 
//////////////////////////////////////////////////////////////////////////
///
/// \class ROOT::TThreadExecutor
/// \ingroup Parallelism
/// \brief This class provides a simple interface to execute the same task
/// multiple times in parallel threads, possibly with different arguments every
/// time.
///
/// ### ROOT::TThreadExecutor::Map
/// This class inherits its interfaces from ROOT::TExecutorCRTP\n, adapting them for multithreaded
/// parallelism and extends them supporting:
/// * Parallel `Foreach` operations.
/// * Custom task granularity and partial reduction, by specifying reduction function
/// and the number of chunks as extra parameters for the Map call. This is specially useful
/// to reduce the size of intermediate results when dealing with a sizeable number of elements
/// in the input data.
///
/// The two possible usages of the Map method are:\n
/// * Map(F func, unsigned nTimes): func is executed nTimes with no arguments
/// * Map(F func, T& args): func is executed on each element of the collection of arguments args
///
/// For either signature, func is executed as many times as needed by a pool of
/// nThreads threads, where nThreads typically defaults to the number of cores.\n
/// A collection containing the result of each execution is returned.\n
/// **Note:** the user is responsible for the deletion of any object that might
/// be created upon execution of func, returned objects included: ROOT::TThreadExecutor never
/// deletes what it returns, it simply forgets it.\n
///
/// \param func
/// \parblock
/// a callable object, such as a lambda expression, an std::function, a
/// functor object or a function that takes zero arguments (for the first signature)
/// or one (for the second signature).
/// \endparblock
/// \param args
/// \parblock
/// a standard vector, a ROOT::TSeq of integer type or an initializer list for the second signature.
/// An integer only for the first.
/// \endparblock
/// **Note:** in cases where the function to be executed takes more than
/// zero/one argument but all are fixed except zero/one, the function can be wrapped
/// in a lambda or via std::bind to give it the right signature.\n
///
/// #### Return value:
/// An std::vector. The elements in the container
/// will be the objects returned by func.
///
///
/// #### Examples:
///
/// ~~~{.cpp}
/// root[] ROOT::TThreadExecutor pool; auto hists = pool.Map(CreateHisto, 10);
/// root[] ROOT::TThreadExecutor pool(2); auto squares = pool.Map([](int a) { return a*a; }, {1,2,3});
/// ~~~
///
/// ### ROOT::TThreadExecutor::MapReduce
/// This set of methods behaves exactly like Map, but takes an additional
/// function as a third argument. This function is applied to the set of
/// objects returned by the corresponding Map execution to "squash" them
/// into a single object. This function should be independent of the size of
/// the vector returned by Map due to optimization of the number of chunks.
///
/// If this function is a binary operator, the "squashing" will be performed in parallel.
/// This is exclusive to ROOT::TThreadExecutor and not any other ROOT::TExecutorCRTP-derived classes.\n
///
/// An integer can be passed as the fourth argument indicating the number of chunks we want to divide our work in.
/// This may be useful to avoid the overhead introduced when running really short tasks.
///
/// #### Examples:
/// ~~~{.cpp}
/// root[] ROOT::TThreadExecutor pool; auto ten = pool.MapReduce([]() { return 1; }, 10, [](const std::vector<int> &v) { return std::accumulate(v.begin(), v.end(), 0); })
/// root[] ROOT::TThreadExecutor pool; auto hist = pool.MapReduce(CreateAndFillHists, 10, PoolUtils::ReduceObjects);
/// ~~~
///
//////////////////////////////////////////////////////////////////////////
 
/*
VERY IMPORTANT NOTE ABOUT WORK ISOLATION
 
We enclose the parallel_for and parallel_reduce invocations in a
task_arena::isolate because we want to prevent a thread to start executing an
outer task when the task it's running spawned subtasks, e.g. with a parallel_for,
and is waiting on inner tasks to be completed.
 
While this change has a negligible performance impact, it has benefits for
several applications, for example big parallelised HEP frameworks and
RDataFrame analyses.
- For HEP Frameworks, without work isolation, it can happen that a huge
framework task is pulled by a yielding ROOT task.
This causes to delay the processing of the event which is interrupted by the
long task.
For example, work isolation avoids that during the wait due to the parallel
flushing of baskets, a very long simulation task is pulled in by the idle task.
- For RDataFrame analyses we want to guarantee that each entry is processed from
the beginning to the end without TBB interrupting it to pull in other work items.
As a corollary, the usage of ROOT (or TBB in work isolation mode) in actions
and transformations guarantee that each entry is processed from the beginning to
the end without being interrupted by the processing of outer tasks.
*/
 
namespace ROOT {
namespace Internal {
 
/// A helper function to implement the TThreadExecutor::ParallelReduce methods
template<typename T>
static T ParallelReduceHelper(const std::vector<T> &objs, const std::function<T(T a, T b)> &redfunc)
{
   using BRange_t = tbb::blocked_range<decltype(objs.begin())>;
 
   auto pred = [redfunc](BRange_t const & range, T init) {
      return std::accumulate(range.begin(), range.end(), init, redfunc);
   };
 
   BRange_t objRange(objs.begin(), objs.end());
 
   return tbb::this_task_arena::isolate([&] {
      return tbb::parallel_reduce(objRange, T{}, pred, redfunc);
   });
 
}
 
} // End NS Internal
 
//////////////////////////////////////////////////////////////////////////
/// \brief Class constructor.
/// If the scheduler is active (e.g. because another TThreadExecutor is in flight, or ROOT::EnableImplicitMT() was
/// called), work with the current pool of threads.
/// If not, initialize the pool of threads, spawning nThreads. nThreads' default value, 0, initializes the
/// pool with as many logical threads as are available in the system (see NLogicalCores in RTaskArenaWrapper.cxx).
///
/// At construction time, TThreadExecutor automatically enables ROOT's thread-safety locks as per calling
/// ROOT::EnableThreadSafety().
TThreadExecutor::TThreadExecutor(UInt_t nThreads)
{
   fTaskArenaW = ROOT::Internal::GetGlobalTaskArena(nThreads);
}
 
//////////////////////////////////////////////////////////////////////////
/// \brief Execute a function in parallel over the indices of a loop.
///
/// \param start Start index of the loop.
/// \param end End index of the loop.
/// \param step Step size of the loop.
/// \param f function to execute.
void TThreadExecutor::ParallelFor(unsigned int start, unsigned int end, unsigned step,
                                  const std::function<void(unsigned int i)> &f)
{
   if (GetPoolSize() > tbb::global_control::active_value(tbb::global_control::max_allowed_parallelism)) {
      Warning("TThreadExecutor::ParallelFor",
              "tbb::global_control is limiting the number of parallel workers."
              " Proceeding with %zu threads this time",
              tbb::global_control::active_value(tbb::global_control::max_allowed_parallelism));
   }
   fTaskArenaW->Access().execute([&] {
      tbb::this_task_arena::isolate([&] {
         tbb::parallel_for(start, end, step, f);
      });
   });
}
 
//////////////////////////////////////////////////////////////////////////
/// \brief "Reduce" in parallel an std::vector<double> into a single double value
///
/// \param objs A vector of elements to combine.
/// \param redfunc Reduction function to combine the elements of the vector `objs`.
/// \return A value result of combining the vector elements into a single object of the same type.
double TThreadExecutor::ParallelReduce(const std::vector<double> &objs,
                                       const std::function<double(double a, double b)> &redfunc)
{
   if (GetPoolSize() > tbb::global_control::active_value(tbb::global_control::max_allowed_parallelism)) {
      Warning("TThreadExecutor::ParallelReduce",
              "tbb::global_control is limiting the number of parallel workers."
              " Proceeding with %zu threads this time",
              tbb::global_control::active_value(tbb::global_control::max_allowed_parallelism));
   }
   return fTaskArenaW->Access().execute([&] { return ROOT::Internal::ParallelReduceHelper<double>(objs, redfunc); });
}
 
//////////////////////////////////////////////////////////////////////////
/// \brief "Reduce" in parallel an std::vector<float> into a single float value
///
/// \param objs A vector of elements to combine.
/// \param redfunc Reduction function to combine the elements of the vector `objs`.
/// \return A value result of combining the vector elements into a single object of the same type.
float TThreadExecutor::ParallelReduce(const std::vector<float> &objs,
                                      const std::function<float(float a, float b)> &redfunc)
{
   if (GetPoolSize() > tbb::global_control::active_value(tbb::global_control::max_allowed_parallelism)) {
      Warning("TThreadExecutor::ParallelReduce",
              "tbb::global_control is limiting the number of parallel workers."
              " Proceeding with %zu threads this time",
              tbb::global_control::active_value(tbb::global_control::max_allowed_parallelism));
   }
   return fTaskArenaW->Access().execute([&] { return ROOT::Internal::ParallelReduceHelper<float>(objs, redfunc); });
}
 
//////////////////////////////////////////////////////////////////////////
/// \brief Returns the number of worker threads in the task arena.
/// \return the number of worker threads assigned to the task arena.
unsigned TThreadExecutor::GetPoolSize() const
{
   return fTaskArenaW->TaskArenaSize();
}
 
} // namespace ROOT