Tpetra_Details_allReduceView.hpp

Go to the documentation of this file.

// @HEADER
// *****************************************************************************
//          Tpetra: Templated Linear Algebra Services Package
//
// Copyright 2008 NTESS and the Tpetra contributors.
// SPDX-License-Identifier: BSD-3-Clause
// *****************************************************************************
// @HEADER

#ifndef TPETRA_DETAILS_ALLREDUCEVIEW_HPP
#define TPETRA_DETAILS_ALLREDUCEVIEW_HPP

#include "Tpetra_Details_Behavior.hpp"
#include "Tpetra_Details_isInterComm.hpp"
#include "Kokkos_Core.hpp"
#include "Teuchos_CommHelpers.hpp"
#include "Tpetra_Details_temporaryViewUtils.hpp"
#include <limits>
#include <type_traits>


namespace Tpetra {
namespace Details {

template <typename InputViewType, typename OutputViewType>
static void
allReduceRawContiguous(const OutputViewType& output,
                       const InputViewType& input,
                       const Teuchos::Comm<int>& comm) {
  using Teuchos::outArg;
  using Teuchos::REDUCE_SUM;
  using Teuchos::reduceAll;
  using ValueType = typename InputViewType::non_const_value_type;
  size_t count    = input.span();
  TEUCHOS_ASSERT(count <= size_t(INT_MAX));
  if (isInterComm(comm) && input.data() == output.data()) {
    // Can't do in-place collective on an intercomm,
    // so use a separate copy as the input.
    typename InputViewType::array_layout layout(input.extent(0), input.extent(1), input.extent(2), input.extent(3), input.extent(4), input.extent(5), input.extent(6), input.extent(7));
    Kokkos::View<typename InputViewType::non_const_data_type, typename InputViewType::array_layout, typename InputViewType::device_type>
        tempInput(Kokkos::ViewAllocateWithoutInitializing("tempInput"), layout);
    // DEEP_COPY REVIEW - This could be either DEVICE-TO-DEVICE or HOST-TO-HOST
    // Either way, MPI is called right afterwards, meaning we'd need a sync on device
    Kokkos::deep_copy(tempInput, input);
    reduceAll<int, ValueType>(comm, REDUCE_SUM, static_cast<int>(count),
                              tempInput.data(), output.data());
  } else
    reduceAll<int, ValueType>(comm, REDUCE_SUM, static_cast<int>(count),
                              input.data(), output.data());
}

template <class InputViewType, class OutputViewType>
static void
allReduceView(const OutputViewType& output,
              const InputViewType& input,
              const Teuchos::Comm<int>& comm) {
  // using execution_space = typename OutputViewType::execution_space;
  const bool viewsAlias = output.data() == input.data();
  if (comm.getSize() == 1) {
    if (!viewsAlias) {
      // InputViewType and OutputViewType can't be AnonymousSpace
      // Views, because deep_copy needs to know their memory spaces.
      // DEEP_COPY REVIEW - NOT TESTED
      Kokkos::deep_copy(output, input);
    }
    return;
  }

  // we must ensure MPI can handle the pointers we pass it
  // if GPUAware, we are done
  // otherwise, if the views use GPUs, then we should copy them
  using Layout = typename TempView::UnifiedContiguousLayout<InputViewType, OutputViewType>::type;
  // if one or both is already in the correct layout, toLayout returns the same view
  auto inputContig  = TempView::toLayout<InputViewType, Layout>(input);
  auto outputContig = TempView::toLayout<InputViewType, Layout>(output);
  if (Tpetra::Details::Behavior::assumeMpiIsGPUAware()) {
    allReduceRawContiguous(outputContig, inputContig, comm);

  } else {
    auto inputMPI  = TempView::toMPISafe<decltype(inputContig), false>(inputContig);
    auto outputMPI = TempView::toMPISafe<decltype(outputContig), false>(outputContig);
    allReduceRawContiguous(outputMPI, inputMPI, comm);
    // DEEP_COPY REVIEW - Could be either
    Kokkos::deep_copy(outputContig, outputMPI);
  }
  // DEEP_COPY REVIEW - Could be either
  Kokkos::deep_copy(output, outputContig);
}

}  // namespace Details
}  // namespace Tpetra

#endif  // TPETRA_DETAILS_ALLREDUCEVIEW_HPP