AcceleratorMemoryCopier.h

// -*- tab-width: 2; indent-tabs-mode: nil; coding: utf-8-with-signature -*-
//-----------------------------------------------------------------------------
// Copyright 2000-2025 CEA (www.cea.fr) IFPEN (www.ifpenergiesnouvelles.com)
// See the top-level COPYRIGHT file for details.
// SPDX-License-Identifier: Apache-2.0
//-----------------------------------------------------------------------------
/*---------------------------------------------------------------------------*/
/* AcceleratorMemoryCopier.h                                   (C) 2000-2025 */
/*                                                                           */
/* Implémentation sur accélérateurs des fonctions de copie mémoire.          */
/*---------------------------------------------------------------------------*/
#ifndef ARCCORE_ACCELERATOR_INTERNAL_ACCELERATORMEMORYCOPIER_H
#define ARCCORE_ACCELERATOR_INTERNAL_ACCELERATORMEMORYCOPIER_H
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
#include "arccore/base/Ref.h"
#include "arccore/base/FixedArray.h"
#include "arccore/base/NotSupportedException.h"
 
#include "arccore/common/accelerator/RunQueue.h"
#include "arccore/common/internal/SpecificMemoryCopyList.h"
 
#include "arccore/accelerator/RunCommandLoop.h"
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
namespace Arcane::Accelerator::impl
{
 
using IndexedMemoryCopyArgs = Arcane::impl::IndexedMemoryCopyArgs;
using IndexedMultiMemoryCopyArgs = Arcane::impl::IndexedMultiMemoryCopyArgs;
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
template <typename DataType, typename Extent>
class AcceleratorSpecificMemoryCopy
: public Arcane::impl::SpecificMemoryCopyBase<DataType, Extent>
{
  using BaseClass = Arcane::impl::SpecificMemoryCopyBase<DataType, Extent>;
  using BaseClass::_toTrueType;
 
 public:
 
  using BaseClass::m_extent;
 
 public:
 
  void copyFrom(const IndexedMemoryCopyArgs& args) override
  {
    _copyFrom(args.m_queue, args.m_indexes, _toTrueType(args.m_source), _toTrueType(args.m_destination));
  }
 
  void copyTo(const IndexedMemoryCopyArgs& args) override
  {
    _copyTo(args.m_queue, args.m_indexes, _toTrueType(args.m_source), _toTrueType(args.m_destination));
  }
 
  void fill(const IndexedMemoryCopyArgs& args) override
  {
    _fill(args.m_queue, args.m_indexes, _toTrueType(args.m_source), _toTrueType(args.m_destination));
  }
 
  void copyFrom(const IndexedMultiMemoryCopyArgs& args) override
  {
    _copyFrom(args.m_queue, args.m_indexes, args.m_multi_memory, _toTrueType(args.m_source_buffer));
  }
 
  void copyTo(const IndexedMultiMemoryCopyArgs& args) override
  {
    _copyTo(args.m_queue, args.m_indexes, args.m_const_multi_memory, _toTrueType(args.m_destination_buffer));
  }
 
  void fill(const IndexedMultiMemoryCopyArgs& args) override
  {
    _fill(args.m_queue, args.m_indexes, args.m_multi_memory, _toTrueType(args.m_source_buffer));
  }
 
 public:
 
  void _copyFrom(const RunQueue* queue, SmallSpan<const Int32> indexes,
                 Span<const DataType> source, Span<DataType> destination)
  {
    ARCCORE_CHECK_POINTER(queue);
 
    ARCCORE_CHECK_ACCESSIBLE_POINTER(queue, indexes.data());
    ARCCORE_CHECK_ACCESSIBLE_POINTER(queue, source.data());
    ARCCORE_CHECK_ACCESSIBLE_POINTER(queue, destination.data());
 
    Int32 nb_index = indexes.size();
    const auto extent = m_extent;
 
    auto command = makeCommand(queue);
    command << RUNCOMMAND_LOOP1(iter, nb_index)
    {
      Int32 i = iter;
      Int64 zindex = i * extent.size();
      Int64 zci = indexes[i] * extent.size();
      for (Int32 z = 0; z < extent.v; ++z)
        destination[zindex + z] = source[zci + z];
    };
  }
 
  void _copyFrom(const RunQueue* queue, SmallSpan<const Int32> indexes, SmallSpan<Span<std::byte>> multi_views,
                 Span<const DataType> source)
  {
    ARCCORE_CHECK_POINTER(queue);
    if (arccoreIsCheck()) {
      ARCCORE_CHECK_ACCESSIBLE_POINTER_ALWAYS(queue, indexes.data());
      ARCCORE_CHECK_ACCESSIBLE_POINTER_ALWAYS(queue, source.data());
      ARCCORE_CHECK_ACCESSIBLE_POINTER_ALWAYS(queue, multi_views.data());
      // Idéalement il faudrait tester les valeurs des éléments de multi_views
      // mais si on fait cela on peut potentiellement faire des transferts
      // entre l'accélérateur et le CPU.
    }
    const Int32 nb_index = indexes.size() / 2;
    const auto extent = m_extent;
 
    auto command = makeCommand(queue);
    command << RUNCOMMAND_LOOP1(iter, nb_index)
    {
      auto [i] = iter();
      Int32 index0 = indexes[i * 2];
      Int64 index1 = indexes[(i * 2) + 1];
      Span<std::byte> orig_view_bytes = multi_views[index0];
      auto* orig_view_data = reinterpret_cast<DataType*>(orig_view_bytes.data());
      // Utilise un span pour tester les débordements de tableau mais on
      // pourrait directement utiliser 'orig_view_data' pour plus de performances
      Span<DataType> orig_view = { orig_view_data, orig_view_bytes.size() / (Int64)sizeof(DataType) };
      Int64 zci = index1 * extent.v;
      Int64 z_index = i * extent.size();
      for (Int32 z = 0, n = extent.v; z < n; ++z)
        orig_view[zci + z] = source[z_index + z];
    };
  }

  void _fill(const RunQueue* queue, SmallSpan<const Int32> indexes, Span<const DataType> source,
             Span<DataType> destination)
  {
    ARCCORE_CHECK_POINTER(queue);
 
    ARCCORE_CHECK_ACCESSIBLE_POINTER(queue, indexes.data());
    ARCCORE_CHECK_ACCESSIBLE_POINTER(queue, destination.data());
    ARCCORE_CHECK_ACCESSIBLE_POINTER(eExecutionPolicy::Sequential, source.data());
 
    Int32 nb_index = indexes.size();
    const auto extent = m_extent;
    constexpr Int32 max_size = 24;
 
    // Pour l'instant on limite la taille de DataType en dur.
    // A terme, il faudrait allouer sur le device et désallouer en fin
    // d'exécution (via cudaMallocAsync/cudaFreeAsync pour gérer l'asynchronisme)
    if (extent.v > max_size)
      ARCCORE_THROW(NotSupportedException, "sizeof(type) is too big (v={0} max={1})",
                    sizeof(DataType) * extent.v, sizeof(DataType) * max_size);
    FixedArray<DataType, max_size> local_source;
    for (Int32 z = 0; z < extent.v; ++z)
      local_source[z] = source[z];
    for (Int32 z = extent.v; z < max_size; ++z)
      local_source[z] = {};
 
    auto command = makeCommand(queue);
    // Si \a nb_index vaut 0, on remplit tous les éléments
    if (nb_index == 0) {
      Int32 nb_value = CheckedConvert::toInt32(destination.size() / extent.v);
      command << RUNCOMMAND_LOOP1(iter, nb_value)
      {
        auto [i] = iter();
        Int64 zci = i * extent.size();
        for (Int32 z = 0; z < extent.v; ++z)
          destination[zci + z] = local_source[z];
      };
    }
    else {
      command << RUNCOMMAND_LOOP1(iter, nb_index)
      {
        auto [i] = iter();
        Int64 zci = indexes[i] * extent.size();
        for (Int32 z = 0; z < extent.v; ++z)
          destination[zci + z] = local_source[z];
      };
    }
  }
 
  void _fill(const RunQueue* queue, SmallSpan<const Int32> indexes, SmallSpan<Span<std::byte>> multi_views,
             Span<const DataType> source)
  {
    ARCCORE_CHECK_POINTER(queue);
 
    if (arccoreIsCheck()) {
      ARCCORE_CHECK_ACCESSIBLE_POINTER_ALWAYS(queue, indexes.data());
      ARCCORE_CHECK_ACCESSIBLE_POINTER_ALWAYS(eExecutionPolicy::Sequential, source.data());
      ARCCORE_CHECK_ACCESSIBLE_POINTER_ALWAYS(queue, multi_views.data());
      // Idéalement il faudrait tester les valeurs des éléments de multi_views
      // mais si on fait cela on peut potentiellement faire des transferts
      // entre l'accélérateur et le CPU.
    }
    const Int32 nb_index = indexes.size() / 2;
    const auto extent = m_extent;
    constexpr Int32 max_size = 24;
 
    // Pour l'instant on limite la taille de DataType en dur.
    // A terme, il faudrait allouer sur le device et désallouer en fin
    // d'exécution (via cudaMallocAsync/cudaFreeAsync pour gérer l'asynchronisme)
    if (extent.v > max_size)
      ARCCORE_THROW(NotSupportedException, "sizeof(type) is too big (v={0} max={1})",
                    sizeof(DataType) * extent.v, sizeof(DataType) * max_size);
    FixedArray<DataType, max_size> local_source;
    for (Int32 z = 0; z < extent.v; ++z)
      local_source[z] = source[z];
    for (Int32 z = extent.v; z < max_size; ++z)
      local_source[z] = {};
 
    if (nb_index == 0) {
      // Remplit toutes les valeurs du tableau avec la source.
      // Comme le nombre d'éléments de la deuxième dimension dépend de la première,
      // on utilise un noyau par dimension.
      RunQueue q(*queue);
      RunQueue::ScopedAsync sc(&q);
      const Int32 nb_dim1 = multi_views.size();
      for (Int32 zz = 0; zz < nb_dim1; ++zz) {
        Span<DataType> orig_view = Arccore::asSpan<DataType>(multi_views[zz]);
        Int32 nb_value = CheckedConvert::toInt32(orig_view.size());
        auto command = makeCommand(queue);
        command << RUNCOMMAND_LOOP1(iter, nb_value)
        {
          auto [i] = iter();
          orig_view[i] = local_source[i % extent.v];
        };
      }
    }
    else {
      auto command = makeCommand(queue);
      command << RUNCOMMAND_LOOP1(iter, nb_index)
      {
        auto [i] = iter();
        Int32 index0 = indexes[i * 2];
        Int64 index1 = indexes[(i * 2) + 1];
        Span<std::byte> orig_view_bytes = multi_views[index0];
        auto* orig_view_data = reinterpret_cast<DataType*>(orig_view_bytes.data());
        // Utilise un span pour tester les débordements de tableau mais on
        // pourrait directement utiliser 'orig_view_data' pour plus de performances
        Span<DataType> orig_view = { orig_view_data, orig_view_bytes.size() / (Int64)sizeof(DataType) };
        Int64 zci = index1 * extent.v;
        for (Int32 z = 0, n = extent.v; z < n; ++z)
          orig_view[zci + z] = local_source[z];
      };
    }
  }
 
  void _copyTo(const RunQueue* queue, SmallSpan<const Int32> indexes, Span<const DataType> source,
               Span<DataType> destination)
  {
    ARCCORE_CHECK_POINTER(queue);
 
    ARCCORE_CHECK_ACCESSIBLE_POINTER(queue, indexes.data());
    ARCCORE_CHECK_ACCESSIBLE_POINTER(queue, source.data());
    ARCCORE_CHECK_ACCESSIBLE_POINTER(queue, destination.data());
 
    Int32 nb_index = indexes.size();
    const auto extent = m_extent;
 
    auto command = makeCommand(queue);
    command << RUNCOMMAND_LOOP1(iter, nb_index)
    {
      auto [i] = iter();
      Int64 zindex = i * extent.size();
      Int64 zci = indexes[i] * extent.v;
      for (Int32 z = 0; z < extent.v; ++z)
        destination[zci + z] = source[zindex + z];
    };
  }
  void _copyTo(const RunQueue* queue, SmallSpan<const Int32> indexes, SmallSpan<const Span<const std::byte>> multi_views,
               Span<DataType> destination)
  {
    ARCCORE_CHECK_POINTER(queue);
 
    if (arccoreIsCheck()) {
      ARCCORE_CHECK_ACCESSIBLE_POINTER_ALWAYS(queue, indexes.data());
      ARCCORE_CHECK_ACCESSIBLE_POINTER_ALWAYS(queue, destination.data());
      ARCCORE_CHECK_ACCESSIBLE_POINTER_ALWAYS(queue, multi_views.data());
      // Idéalement il faudrait tester les valeurs des éléments de multi_views
      // mais si on fait cela on peut potentiellement faire des transferts
      // entre l'accélérateur et le CPU.
    }
 
    const Int32 nb_index = indexes.size() / 2;
    const auto extent = m_extent;
 
    auto command = makeCommand(queue);
    command << RUNCOMMAND_LOOP1(iter, nb_index)
    {
      auto [i] = iter();
      Int32 index0 = indexes[i * 2];
      Int64 index1 = indexes[(i * 2) + 1];
      Span<const std::byte> orig_view_bytes = multi_views[index0];
      auto* orig_view_data = reinterpret_cast<const DataType*>(orig_view_bytes.data());
      // Utilise un span pour tester les débordements de tableau mais on
      // pourrait directement utiliser 'orig_view_data' pour plus de performances
      Span<const DataType> orig_view = { orig_view_data, orig_view_bytes.size() / (Int64)sizeof(DataType) };
      Int64 zci = index1 * extent.v;
      Int64 z_index = i * extent.size();
      for (Int32 z = 0, n = extent.v; z < n; ++z)
        destination[z_index + z] = orig_view[zci + z];
    };
  }
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
class AcceleratorIndexedCopyTraits
{
 public:
 
  using InterfaceType = Arcane::impl::ISpecificMemoryCopy;
  template <typename DataType, typename Extent> using SpecificType = AcceleratorSpecificMemoryCopy<DataType, Extent>;
  using RefType = Arcane::impl::SpecificMemoryCopyRef<AcceleratorIndexedCopyTraits>;
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
class AcceleratorSpecificMemoryCopyList
: public Arcane::impl::SpecificMemoryCopyList<AcceleratorIndexedCopyTraits>
{
 public:
 
  AcceleratorSpecificMemoryCopyList();
 
  void addExplicitTemplate1();
  void addExplicitTemplate2();
  void addExplicitTemplate3();
  void addExplicitTemplate4();
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
} // namespace Arcane::Accelerator::impl
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
#endif