work_stream.h

Go to the documentation of this file.

// ---------------------------------------------------------------------
//
// Copyright (C) 2008 - 2020 by the deal.II authors
//
// This file is part of the deal.II library.
//
// The deal.II library is free software; you can use it, redistribute
// it, and/or modify it under the terms of the GNU Lesser General
// Public License as published by the Free Software Foundation; either
// version 2.1 of the License, or (at your option) any later version.
// The full text of the license can be found in the file LICENSE.md at
// the top level directory of deal.II.
//
// ---------------------------------------------------------------------
 
#ifndef dealii_work_stream_h
#  define dealii_work_stream_h
 
 
#  include <deal.II/base/config.h>
 
#  include <deal.II/base/graph_coloring.h>
#  include <deal.II/base/iterator_range.h>
#  include <deal.II/base/multithread_info.h>
#  include <deal.II/base/parallel.h>
#  include <deal.II/base/template_constraints.h>
#  include <deal.II/base/thread_local_storage.h>
#  include <deal.II/base/thread_management.h>
 
#  ifdef DEAL_II_WITH_TBB
#    include <tbb/pipeline.h>
#  endif
 
#  include <functional>
#  include <iterator>
#  include <memory>
#  include <utility>
#  include <vector>
 
DEAL_II_NAMESPACE_OPEN
 
 
 
namespace WorkStream
{
  namespace internal
  {
#  ifdef DEAL_II_WITH_TBB
    namespace tbb_no_coloring
    {
      template <typename Iterator, typename ScratchData, typename CopyData>
      class IteratorRangeToItemStream : public tbb::filter
      {
      public:
        struct ItemType
        {
          struct ScratchDataObject
          {
            std::unique_ptr<ScratchData> scratch_data;
            bool                         currently_in_use;
 
            ScratchDataObject()
              : currently_in_use(false)
            {}
 
            ScratchDataObject(ScratchData *p, const bool in_use)
              : scratch_data(p)
              , currently_in_use(in_use)
            {}
 
            // Provide a copy constructor that actually doesn't copy the
            // internal state. This makes handling ScratchAndCopyDataObjects
            // easier to handle with STL containers.
            ScratchDataObject(const ScratchDataObject &)
              : currently_in_use(false)
            {}
 
            ScratchDataObject(ScratchDataObject &&o) noexcept = default;
          };
 
 
          using ScratchDataList = std::list<ScratchDataObject>;
 
          std::vector<Iterator> work_items;
 
          std::vector<CopyData> copy_datas;
 
          unsigned int n_items;
 
          Threads::ThreadLocalStorage<ScratchDataList> *scratch_data;
 
          const ScratchData *sample_scratch_data;
 
          bool currently_in_use;
 
 
          ItemType()
            : n_items(0)
            , scratch_data(nullptr)
            , sample_scratch_data(nullptr)
            , currently_in_use(false)
          {}
        };
 
 
        IteratorRangeToItemStream(const Iterator &   begin,
                                  const Iterator &   end,
                                  const unsigned int buffer_size,
                                  const unsigned int chunk_size,
                                  const ScratchData &sample_scratch_data,
                                  const CopyData &   sample_copy_data)
          : tbb::filter(/*is_serial=*/true)
          , remaining_iterator_range(begin, end)
          , item_buffer(buffer_size)
          , sample_scratch_data(sample_scratch_data)
          , chunk_size(chunk_size)
        {
          // initialize the elements of the ring buffer
          for (unsigned int element = 0; element < item_buffer.size();
               ++element)
            {
              Assert(item_buffer[element].n_items == 0, ExcInternalError());
 
              item_buffer[element].work_items.resize(
                chunk_size, remaining_iterator_range.second);
              item_buffer[element].scratch_data        = &thread_local_scratch;
              item_buffer[element].sample_scratch_data = &sample_scratch_data;
              item_buffer[element].copy_datas.resize(chunk_size,
                                                     sample_copy_data);
              item_buffer[element].currently_in_use = false;
            }
        }
 
 
        virtual void *
        operator()(void *) override
        {
          // find first unused item. we know that there must be one
          // because we have set the maximal number of tokens in flight
          // and have set the ring buffer to have exactly this size. so
          // if this function is called, we know that less than the
          // maximal number of items in currently in flight
          //
          // note that we need not lock access to this array since
          // the current stage is run sequentially and we can therefore
          // enter the following block only once at any given time.
          // thus, there can be no race condition between checking that
          // a flag is false and setting it to true. (there may be
          // another thread where we release items and set 'false'
          // flags to 'true', but that too does not produce any
          // problems)
          ItemType *current_item = nullptr;
          for (unsigned int i = 0; i < item_buffer.size(); ++i)
            if (item_buffer[i].currently_in_use == false)
              {
                item_buffer[i].currently_in_use = true;
                current_item                    = &item_buffer[i];
                break;
              }
          Assert(current_item != nullptr,
                 ExcMessage("This can't be. There must be a free item!"));
 
          // initialize the next item. it may
          // consist of at most chunk_size
          // elements
          current_item->n_items = 0;
          while ((remaining_iterator_range.first !=
                  remaining_iterator_range.second) &&
                 (current_item->n_items < chunk_size))
            {
              current_item->work_items[current_item->n_items] =
                remaining_iterator_range.first;
 
              ++remaining_iterator_range.first;
              ++current_item->n_items;
            }
 
          if (current_item->n_items == 0)
            // there were no items
            // left. terminate the pipeline
            return nullptr;
          else
            return current_item;
        }
 
      private:
        std::pair<Iterator, Iterator> remaining_iterator_range;
 
        std::vector<ItemType> item_buffer;
 
        Threads::ThreadLocalStorage<typename ItemType::ScratchDataList>
          thread_local_scratch;
 
        const ScratchData &sample_scratch_data;
 
        const unsigned int chunk_size;
      };
 
 
 
      template <typename Iterator, typename ScratchData, typename CopyData>
      class TBBWorker : public tbb::filter
      {
      public:
        TBBWorker(
          const std::function<void(const Iterator &, ScratchData &, CopyData &)>
            &  worker,
          bool copier_exist = true)
          : tbb::filter(/* is_serial= */ false)
          , worker(worker)
          , copier_exist(copier_exist)
        {}
 
 
        void *
        operator()(void *item) override
        {
          // first unpack the current item
          using ItemType =
            typename IteratorRangeToItemStream<Iterator,
                                               ScratchData,
                                               CopyData>::ItemType;
 
          ItemType *current_item = static_cast<ItemType *>(item);
 
          // we need to find an unused scratch data object in the list that
          // corresponds to the current thread and then mark it as used. if
          // we can't find one, create one
          //
          // as discussed in the discussion of the documentation of the
          // IteratorRangeToItemStream::scratch_data variable, there is no
          // need to synchronize access to this variable using a mutex
          // as long as we have no yield-point in between. this means that
          // we can't take an iterator into the list now and expect it to
          // still be valid after calling the worker, but we at least do
          // not have to lock the following section
          ScratchData *scratch_data = nullptr;
          {
            typename ItemType::ScratchDataList &scratch_data_list =
              current_item->scratch_data->get();
 
            // see if there is an unused object. if so, grab it and mark
            // it as used
            for (typename ItemType::ScratchDataList::iterator p =
                   scratch_data_list.begin();
                 p != scratch_data_list.end();
                 ++p)
              if (p->currently_in_use == false)
                {
                  scratch_data        = p->scratch_data.get();
                  p->currently_in_use = true;
                  break;
                }
 
            // if no object was found, create one and mark it as used
            if (scratch_data == nullptr)
              {
                scratch_data =
                  new ScratchData(*current_item->sample_scratch_data);
 
                typename ItemType::ScratchDataList::value_type
                  new_scratch_object(scratch_data, true);
                scratch_data_list.push_back(std::move(new_scratch_object));
              }
          }
 
          // then call the worker function on each element of the chunk we were
          // given. since these worker functions are called on separate threads,
          // nothing good can happen if they throw an exception and we are best
          // off catching it and showing an error message
          for (unsigned int i = 0; i < current_item->n_items; ++i)
            {
              try
                {
                  if (worker)
                    worker(current_item->work_items[i],
                           *scratch_data,
                           current_item->copy_datas[i]);
                }
              catch (const std::exception &exc)
                {
                  Threads::internal::handle_std_exception(exc);
                }
              catch (...)
                {
                  Threads::internal::handle_unknown_exception();
                }
            }
 
          // finally mark the scratch object as unused again. as above, there
          // is no need to lock anything here since the object we work on
          // is thread-local
          {
            typename ItemType::ScratchDataList &scratch_data_list =
              current_item->scratch_data->get();
 
            for (typename ItemType::ScratchDataList::iterator p =
                   scratch_data_list.begin();
                 p != scratch_data_list.end();
                 ++p)
              if (p->scratch_data.get() == scratch_data)
                {
                  Assert(p->currently_in_use == true, ExcInternalError());
                  p->currently_in_use = false;
                }
          }
 
          // if there is no copier, mark current item as usable again
          if (copier_exist == false)
            current_item->currently_in_use = false;
 
 
          // then return the original pointer
          // to the now modified object
          return item;
        }
 
 
      private:
        const std::function<void(const Iterator &, ScratchData &, CopyData &)>
          worker;
 
        bool copier_exist;
      };
 
 
 
      template <typename Iterator, typename ScratchData, typename CopyData>
      class TBBCopier : public tbb::filter
      {
      public:
        TBBCopier(const std::function<void(const CopyData &)> &copier)
          : tbb::filter(/*is_serial=*/true)
          , copier(copier)
        {}
 
 
        void *
        operator()(void *item) override
        {
          // first unpack the current item
          using ItemType =
            typename IteratorRangeToItemStream<Iterator,
                                               ScratchData,
                                               CopyData>::ItemType;
 
          ItemType *current_item = static_cast<ItemType *>(item);
 
          // initiate copying data. for the same reasons as in the worker class
          // above, catch exceptions rather than letting it propagate into
          // unknown territories
          for (unsigned int i = 0; i < current_item->n_items; ++i)
            {
              try
                {
                  if (copier)
                    copier(current_item->copy_datas[i]);
                }
              catch (const std::exception &exc)
                {
                  Threads::internal::handle_std_exception(exc);
                }
              catch (...)
                {
                  Threads::internal::handle_unknown_exception();
                }
            }
 
          // mark current item as usable again
          current_item->currently_in_use = false;
 
 
          // return an invalid item since we are at the end of the
          // pipeline
          return nullptr;
        }
 
 
      private:
        const std::function<void(const CopyData &)> copier;
      };
 
      template <typename Worker,
                typename Copier,
                typename Iterator,
                typename ScratchData,
                typename CopyData>
      void
      run(const Iterator &                         begin,
          const typename identity<Iterator>::type &end,
          Worker                                   worker,
          Copier                                   copier,
          const ScratchData &                      sample_scratch_data,
          const CopyData &                         sample_copy_data,
          const unsigned int                       queue_length,
          const unsigned int                       chunk_size)
      {
        // create the three stages of the pipeline
        IteratorRangeToItemStream<Iterator, ScratchData, CopyData>
          iterator_range_to_item_stream(begin,
                                        end,
                                        queue_length,
                                        chunk_size,
                                        sample_scratch_data,
                                        sample_copy_data);
 
        TBBWorker<Iterator, ScratchData, CopyData> worker_filter(worker);
        TBBCopier<Iterator, ScratchData, CopyData> copier_filter(copier);
 
        // now create a pipeline from these stages
        tbb::pipeline assembly_line;
        assembly_line.add_filter(iterator_range_to_item_stream);
        assembly_line.add_filter(worker_filter);
        assembly_line.add_filter(copier_filter);
 
        // and run it
        assembly_line.run(queue_length);
 
        assembly_line.clear();
      }
 
    }    // namespace tbb_no_coloring
#  endif // DEAL_II_WITH_TBB
 
 
    namespace sequential
    {
      template <typename Worker,
                typename Copier,
                typename Iterator,
                typename ScratchData,
                typename CopyData>
      void
      run(const Iterator &                         begin,
          const typename identity<Iterator>::type &end,
          Worker                                   worker,
          Copier                                   copier,
          const ScratchData &                      sample_scratch_data,
          const CopyData &                         sample_copy_data)
      {
        // need to copy the sample since it is marked const
        ScratchData scratch_data = sample_scratch_data;
        CopyData    copy_data    = sample_copy_data; // NOLINT
 
        // Optimization: Check if the functions are not the zero function. To
        // check zero-ness, create a C++ function out of it:
        const bool have_worker =
          (static_cast<const std::function<
             void(const Iterator &, ScratchData &, CopyData &)> &>(worker)) !=
          nullptr;
        const bool have_copier =
          (static_cast<const std::function<void(const CopyData &)> &>(
            copier)) != nullptr;
 
        // Finally loop over all items and perform the necessary work:
        for (Iterator i = begin; i != end; ++i)
          {
            if (have_worker)
              worker(i, scratch_data, copy_data);
            if (have_copier)
              copier(copy_data);
          }
      }
 
 
 
      template <typename Worker,
                typename Copier,
                typename Iterator,
                typename ScratchData,
                typename CopyData>
      void
      run(const std::vector<std::vector<Iterator>> &colored_iterators,
          Worker                                    worker,
          Copier                                    copier,
          const ScratchData &                       sample_scratch_data,
          const CopyData &                          sample_copy_data)
      {
        // need to copy the sample since it is marked const
        ScratchData scratch_data = sample_scratch_data;
        CopyData    copy_data    = sample_copy_data; // NOLINT
 
        // Optimization: Check if the functions are not the zero function. To
        // check zero-ness, create a C++ function out of it:
        const bool have_worker =
          (static_cast<const std::function<
             void(const Iterator &, ScratchData &, CopyData &)> &>(worker)) !=
          nullptr;
        const bool have_copier =
          (static_cast<const std::function<void(const CopyData &)> &>(
            copier)) != nullptr;
 
        // Finally loop over all items and perform the necessary work:
        for (unsigned int color = 0; color < colored_iterators.size(); ++color)
          if (colored_iterators[color].size() > 0)
            for (auto &it : colored_iterators[color])
              {
                if (have_worker)
                  worker(it, scratch_data, copy_data);
                if (have_copier)
                  copier(copy_data);
              }
      }
 
    } // namespace sequential
 
 
 
#  ifdef DEAL_II_WITH_TBB
    namespace tbb_colored
    {
      template <typename Iterator, typename ScratchData, typename CopyData>
      struct ScratchAndCopyDataObjects
      {
        std::unique_ptr<ScratchData> scratch_data;
        std::unique_ptr<CopyData>    copy_data;
        bool                         currently_in_use;
 
        ScratchAndCopyDataObjects()
          : currently_in_use(false)
        {}
 
        ScratchAndCopyDataObjects(std::unique_ptr<ScratchData> &&p,
                                  std::unique_ptr<CopyData> &&   q,
                                  const bool                     in_use)
          : scratch_data(std::move(p))
          , copy_data(std::move(q))
          , currently_in_use(in_use)
        {}
 
        // Provide a copy constructor that actually doesn't copy the
        // internal state. This makes handling ScratchAndCopyDataObjects
        // easier to handle with STL containers.
        ScratchAndCopyDataObjects(const ScratchAndCopyDataObjects &)
          : currently_in_use(false)
        {}
      };
 
 
 
      template <typename Iterator, typename ScratchData, typename CopyData>
      class WorkerAndCopier
      {
      public:
        WorkerAndCopier(
          const std::function<void(const Iterator &, ScratchData &, CopyData &)>
            &                                          worker,
          const std::function<void(const CopyData &)> &copier,
          const ScratchData &                          sample_scratch_data,
          const CopyData &                             sample_copy_data)
          : worker(worker)
          , copier(copier)
          , sample_scratch_data(sample_scratch_data)
          , sample_copy_data(sample_copy_data)
        {}
 
 
        void
        operator()(const tbb::blocked_range<
                   typename std::vector<Iterator>::const_iterator> &range)
        {
          // we need to find an unused scratch and corresponding copy
          // data object in the list that corresponds to the current
          // thread and then mark it as used. If we can't find one,
          // create one as discussed in the discussion of the documentation
          // of the IteratorRangeToItemStream::scratch_data variable,
          // there is no need to synchronize access to this variable
          // using a mutex as long as we have no yield-point in between.
          // This means that we can't take an iterator into the list
          // now and expect it to still be valid after calling the worker,
          // but we at least do not have to lock the following section.
          ScratchData *scratch_data = nullptr;
          CopyData *   copy_data    = nullptr;
          {
            ScratchAndCopyDataList &scratch_and_copy_data_list = data.get();
 
            // see if there is an unused object. if so, grab it and mark
            // it as used
            for (typename ScratchAndCopyDataList::iterator p =
                   scratch_and_copy_data_list.begin();
                 p != scratch_and_copy_data_list.end();
                 ++p)
              if (p->currently_in_use == false)
                {
                  scratch_data        = p->scratch_data.get();
                  copy_data           = p->copy_data.get();
                  p->currently_in_use = true;
                  break;
                }
 
            // if no element in the list was found, create one and mark it as
            // used
            if (scratch_data == nullptr)
              {
                Assert(copy_data == nullptr, ExcInternalError());
 
                scratch_and_copy_data_list.emplace_back(
                  std::make_unique<ScratchData>(sample_scratch_data),
                  std::make_unique<CopyData>(sample_copy_data),
                  true);
                scratch_data =
                  scratch_and_copy_data_list.back().scratch_data.get();
                copy_data = scratch_and_copy_data_list.back().copy_data.get();
              }
          }
 
          // then call the worker and copier functions on each
          // element of the chunk we were given.
          for (typename std::vector<Iterator>::const_iterator p = range.begin();
               p != range.end();
               ++p)
            {
              try
                {
                  if (worker)
                    worker(*p, *scratch_data, *copy_data);
                  if (copier)
                    copier(*copy_data);
                }
              catch (const std::exception &exc)
                {
                  Threads::internal::handle_std_exception(exc);
                }
              catch (...)
                {
                  Threads::internal::handle_unknown_exception();
                }
            }
 
          // finally mark the scratch object as unused again. as above, there
          // is no need to lock anything here since the object we work on
          // is thread-local
          {
            ScratchAndCopyDataList &scratch_and_copy_data_list = data.get();
 
            for (typename ScratchAndCopyDataList::iterator p =
                   scratch_and_copy_data_list.begin();
                 p != scratch_and_copy_data_list.end();
                 ++p)
              if (p->scratch_data.get() == scratch_data)
                {
                  Assert(p->currently_in_use == true, ExcInternalError());
                  p->currently_in_use = false;
                }
          }
        }
 
      private:
        using ScratchAndCopyDataObjects = typename tbb_colored::
          ScratchAndCopyDataObjects<Iterator, ScratchData, CopyData>;
 
        using ScratchAndCopyDataList = std::list<ScratchAndCopyDataObjects>;
 
        Threads::ThreadLocalStorage<ScratchAndCopyDataList> data;
 
        const std::function<void(const Iterator &, ScratchData &, CopyData &)>
          worker;
 
        const std::function<void(const CopyData &)> copier;
 
        const ScratchData &sample_scratch_data;
        const CopyData &   sample_copy_data;
      };
 
      template <typename Worker,
                typename Copier,
                typename Iterator,
                typename ScratchData,
                typename CopyData>
      void
      run(const std::vector<std::vector<Iterator>> &colored_iterators,
          Worker                                    worker,
          Copier                                    copier,
          const ScratchData &                       sample_scratch_data,
          const CopyData &                          sample_copy_data,
          const unsigned int                        chunk_size)
      {
        // loop over the various colors of what we're given
        for (unsigned int color = 0; color < colored_iterators.size(); ++color)
          if (colored_iterators[color].size() > 0)
            {
              using WorkerAndCopier = internal::tbb_colored::
                WorkerAndCopier<Iterator, ScratchData, CopyData>;
 
              using RangeType = typename std::vector<Iterator>::const_iterator;
 
              WorkerAndCopier worker_and_copier(worker,
                                                copier,
                                                sample_scratch_data,
                                                sample_copy_data);
 
              parallel::internal::parallel_for(
                colored_iterators[color].begin(),
                colored_iterators[color].end(),
                [&worker_and_copier](
                  const tbb::blocked_range<
                    typename std::vector<Iterator>::const_iterator> &range) {
                  worker_and_copier(range);
                },
                chunk_size);
            }
      }
 
    }    // namespace tbb_colored
#  endif // DEAL_II_WITH_TBB
 
 
  } // namespace internal
 
 
 
  template <typename Worker,
            typename Copier,
            typename Iterator,
            typename ScratchData,
            typename CopyData>
  void
  run(const std::vector<std::vector<Iterator>> &colored_iterators,
      Worker                                    worker,
      Copier                                    copier,
      const ScratchData &                       sample_scratch_data,
      const CopyData &                          sample_copy_data,
      const unsigned int queue_length = 2 * MultithreadInfo::n_threads(),
      const unsigned int chunk_size   = 8);
 
 
  template <typename Worker,
            typename Copier,
            typename Iterator,
            typename ScratchData,
            typename CopyData>
  void
  run(const Iterator &                         begin,
      const typename identity<Iterator>::type &end,
      Worker                                   worker,
      Copier                                   copier,
      const ScratchData &                      sample_scratch_data,
      const CopyData &                         sample_copy_data,
      const unsigned int queue_length = 2 * MultithreadInfo::n_threads(),
      const unsigned int chunk_size   = 8)
  {
    Assert(queue_length > 0,
           ExcMessage("The queue length must be at least one, and preferably "
                      "larger than the number of processors on this system."));
    (void)queue_length; // removes -Wunused-parameter warning in optimized mode
    Assert(chunk_size > 0, ExcMessage("The chunk_size must be at least one."));
    (void)chunk_size; // removes -Wunused-parameter warning in optimized mode
 
    // If no work then skip. (only use operator!= for iterators since we may
    // not have an equality comparison operator)
    if (!(begin != end))
      return;
 
    if (MultithreadInfo::n_threads() > 1)
      {
#  ifdef DEAL_II_WITH_TBB
        if (static_cast<const std::function<void(const CopyData &)> &>(copier))
          {
            // If we have a copier, run the algorithm:
            internal::tbb_no_coloring::run(begin,
                                           end,
                                           worker,
                                           copier,
                                           sample_scratch_data,
                                           sample_copy_data,
                                           queue_length,
                                           chunk_size);
          }
        else
          {
            // There is no copier function. in this case, we have an
            // embarrassingly parallel problem where we can
            // essentially apply parallel_for. because parallel_for
            // requires subdividing the range for which operator- is
            // necessary between iterators, it is often inefficient to
            // apply it directly to cell ranges and similar iterator
            // types for which operator- is expensive or, in fact,
            // nonexistent. rather, in that case, we simply copy the
            // iterators into a large array and use operator- on
            // iterators to this array of iterators.
            //
            // instead of duplicating code, this is essentially the
            // same situation we have in the colored implementation below, so we
            // just defer to that place
            std::vector<std::vector<Iterator>> all_iterators(1);
            for (Iterator p = begin; p != end; ++p)
              all_iterators[0].push_back(p);
 
            run(all_iterators,
                worker,
                copier,
                sample_scratch_data,
                sample_copy_data,
                queue_length,
                chunk_size);
          }
 
        // exit this function to not run the sequential version below:
        return;
#  endif
      }
 
    // no TBB installed or we are requested to run sequentially:
    internal::sequential::run(
      begin, end, worker, copier, sample_scratch_data, sample_copy_data);
  }
 
 
 
  template <typename Worker,
            typename Copier,
            typename IteratorRangeType,
            typename ScratchData,
            typename CopyData,
            typename = typename std::enable_if<
              has_begin_and_end<IteratorRangeType>::value>::type>
  void
  run(IteratorRangeType  iterator_range,
      Worker             worker,
      Copier             copier,
      const ScratchData &sample_scratch_data,
      const CopyData &   sample_copy_data,
      const unsigned int queue_length = 2 * MultithreadInfo::n_threads(),
      const unsigned int chunk_size   = 8)
  {
    // Call the function above
    run(iterator_range.begin(),
        iterator_range.end(),
        worker,
        copier,
        sample_scratch_data,
        sample_copy_data,
        queue_length,
        chunk_size);
  }
 
 
 
  template <typename Worker,
            typename Copier,
            typename Iterator,
            typename ScratchData,
            typename CopyData>
  void
  run(const IteratorRange<Iterator> &iterator_range,
      Worker                         worker,
      Copier                         copier,
      const ScratchData &            sample_scratch_data,
      const CopyData &               sample_copy_data,
      const unsigned int queue_length = 2 * MultithreadInfo::n_threads(),
      const unsigned int chunk_size   = 8)
  {
    // Call the function above
    run(iterator_range.begin(),
        iterator_range.end(),
        worker,
        copier,
        sample_scratch_data,
        sample_copy_data,
        queue_length,
        chunk_size);
  }
 
 
 
  template <typename Worker,
            typename Copier,
            typename Iterator,
            typename ScratchData,
            typename CopyData>
  void
  run(const std::vector<std::vector<Iterator>> &colored_iterators,
      Worker                                    worker,
      Copier                                    copier,
      const ScratchData &                       sample_scratch_data,
      const CopyData &                          sample_copy_data,
      const unsigned int                        queue_length,
      const unsigned int                        chunk_size)
  {
    Assert(queue_length > 0,
           ExcMessage("The queue length must be at least one, and preferably "
                      "larger than the number of processors on this system."));
    (void)queue_length; // removes -Wunused-parameter warning in optimized mode
    Assert(chunk_size > 0, ExcMessage("The chunk_size must be at least one."));
    (void)chunk_size; // removes -Wunused-parameter warning in optimized mode
 
 
    if (MultithreadInfo::n_threads() > 1)
      {
#  ifdef DEAL_II_WITH_TBB
        internal::tbb_colored::run(colored_iterators,
                                   worker,
                                   copier,
                                   sample_scratch_data,
                                   sample_copy_data,
                                   chunk_size);
 
        // exit this function to not run the sequential version below:
        return;
#  endif
      }
 
    // run all colors sequentially:
    {
      internal::sequential::run(colored_iterators,
                                worker,
                                copier,
                                sample_scratch_data,
                                sample_copy_data);
    }
  }
 
 
 
  template <typename MainClass,
            typename Iterator,
            typename ScratchData,
            typename CopyData>
  void
  run(const Iterator &                         begin,
      const typename identity<Iterator>::type &end,
      MainClass &                              main_object,
      void (MainClass::*worker)(const Iterator &, ScratchData &, CopyData &),
      void (MainClass::*copier)(const CopyData &),
      const ScratchData &sample_scratch_data,
      const CopyData &   sample_copy_data,
      const unsigned int queue_length = 2 * MultithreadInfo::n_threads(),
      const unsigned int chunk_size   = 8)
  {
    // forward to the other function
    run(begin,
        end,
        [&main_object, worker](const Iterator &iterator,
                               ScratchData &   scratch_data,
                               CopyData &      copy_data) {
          (main_object.*worker)(iterator, scratch_data, copy_data);
        },
        [&main_object, copier](const CopyData &copy_data) {
          (main_object.*copier)(copy_data);
        },
        sample_scratch_data,
        sample_copy_data,
        queue_length,
        chunk_size);
  }
 
 
  template <typename MainClass,
            typename Iterator,
            typename ScratchData,
            typename CopyData>
  void
  run(const IteratorOverIterators<Iterator> &                         begin,
      const IteratorOverIterators<typename identity<Iterator>::type> &end,
      MainClass &main_object,
      void (MainClass::*worker)(const Iterator &, ScratchData &, CopyData &),
      void (MainClass::*copier)(const CopyData &),
      const ScratchData &sample_scratch_data,
      const CopyData &   sample_copy_data,
      const unsigned int queue_length = 2 * MultithreadInfo::n_threads(),
      const unsigned int chunk_size   = 8)
  {
    // forward to the other function
    run(begin,
        end,
        [&main_object, worker](const Iterator &iterator,
                               ScratchData &   scratch_data,
                               CopyData &      copy_data) {
          (main_object.*worker)(iterator, scratch_data, copy_data);
        },
        [&main_object, copier](const CopyData &copy_data) {
          (main_object.*copier)(copy_data);
        },
        sample_scratch_data,
        sample_copy_data,
        queue_length,
        chunk_size);
  }
 
 
 
  template <typename MainClass,
            typename IteratorRangeType,
            typename ScratchData,
            typename CopyData,
            typename = typename std::enable_if<
              has_begin_and_end<IteratorRangeType>::value>::type>
  void
  run(IteratorRangeType iterator_range,
      MainClass &       main_object,
      void (MainClass::*worker)(
        const typename identity<IteratorRangeType>::type::iterator &,
        ScratchData &,
        CopyData &),
      void (MainClass::*copier)(const CopyData &),
      const ScratchData &sample_scratch_data,
      const CopyData &   sample_copy_data,
      const unsigned int queue_length = 2 * MultithreadInfo::n_threads(),
      const unsigned int chunk_size   = 8)
  {
    // Call the function above
    run(std::begin(iterator_range),
        std::end(iterator_range),
        main_object,
        worker,
        copier,
        sample_scratch_data,
        sample_copy_data,
        queue_length,
        chunk_size);
  }
 
 
 
  template <typename MainClass,
            typename Iterator,
            typename ScratchData,
            typename CopyData>
  void
  run(IteratorRange<Iterator> iterator_range,
      MainClass &             main_object,
      void (MainClass::*worker)(const Iterator &, ScratchData &, CopyData &),
      void (MainClass::*copier)(const CopyData &),
      const ScratchData &sample_scratch_data,
      const CopyData &   sample_copy_data,
      const unsigned int queue_length = 2 * MultithreadInfo::n_threads(),
      const unsigned int chunk_size   = 8)
  {
    // Call the function above
    run(std::begin(iterator_range),
        std::end(iterator_range),
        main_object,
        worker,
        copier,
        sample_scratch_data,
        sample_copy_data,
        queue_length,
        chunk_size);
  }
 
} // namespace WorkStream
 
 
 
DEAL_II_NAMESPACE_CLOSE
 
 
 
//----------------------------   work_stream.h     ---------------------------
// end of #ifndef dealii_work_stream_h
#endif
//----------------------------   work_stream.h     ---------------------------