CooperativeWorkGroupLoopRange.h

// -*- tab-width: 2; indent-tabs-mode: nil; coding: utf-8-with-signature -*-
//-----------------------------------------------------------------------------
// Copyright 2000-2026 CEA (www.cea.fr) IFPEN (www.ifpenergiesnouvelles.com)
// See the top-level COPYRIGHT file for details.
// SPDX-License-Identifier: Apache-2.0
//-----------------------------------------------------------------------------
/*---------------------------------------------------------------------------*/
/* CooperativeWorkGroupLoopRange.h                             (C) 2000-2026 */
/*                                                                           */
/* Boucle pour le parallélisme hiérarchique coopératif.                      */
/*---------------------------------------------------------------------------*/
#ifndef ARCCORE_ACCELERATOR_COOPERATIVEWORKGROUPLOOPRANGE_H
#define ARCCORE_ACCELERATOR_COOPERATIVEWORKGROUPLOOPRANGE_H
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
#include "arccore/accelerator/WorkGroupLoopRange.h"
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
namespace Arcane::Accelerator
{
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
class CooperativeHostWorkItemGrid
{
  template<typename T> friend class CooperativeWorkGroupLoopContext;
 
 private:

  explicit CooperativeHostWorkItemGrid(Int32 nb_block, Impl::ThreadGridSynchronizer* syncer)
  : m_nb_block(nb_block)
  , m_syncer(syncer)
  {}
 
 public:

  Int32 nbBlock() const { return m_nb_block; }

  void barrier()
  {
    if (m_syncer)
      m_syncer->sync();
  }
 
 private:
 
  Int32 m_nb_block = 0;
  Impl::ThreadGridSynchronizer* m_syncer = nullptr;
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
#if defined(ARCCORE_COMPILING_CUDA_OR_HIP)
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
class CooperativeDeviceWorkItemGrid
{
  template <typename T> friend class CooperativeWorkGroupLoopContext;
 
 private:

  __device__ CooperativeDeviceWorkItemGrid()
  : m_grid_group(cooperative_groups::this_grid())
  {}
 
 public:

  __device__ Int32 nbBlock() const { return m_grid_group.group_dim().x; }

  __device__ void barrier() { m_grid_group.sync(); }
 
 private:
 
  cooperative_groups::grid_group m_grid_group;
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
#endif
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
template <typename IndexType_>
class CooperativeWorkGroupLoopContext
: public WorkGroupLoopContextBase<IndexType_>
{
  // Pour accéder aux constructeurs
  friend class CooperativeWorkGroupLoopRange<IndexType_>;
  friend Impl::WorkGroupSequentialForHelper;
  friend Impl::WorkGroupLoopContextBuilder;
  using BaseClass = WorkGroupLoopContextBase<IndexType_>;
 
 public:
 
  using IndexType = IndexType_;
 
 private:

  constexpr CooperativeWorkGroupLoopContext(IndexType loop_index, Int32 group_index,
                                            Int32 group_size, Int32 nb_active_item,
                                            IndexType total_size, Int32 nb_block, Impl::ThreadGridSynchronizer* syncer)
  : BaseClass(loop_index, group_index, group_size, nb_active_item, total_size)
  , m_nb_block(nb_block)
  , m_syncer(syncer)
  {
  }
 
  // Ce constructeur n'est utilisé que sur le device
  // Il ne fait rien car les valeurs utiles sont récupérées via cooperative_groups::this_thread_block()
  explicit constexpr ARCCORE_DEVICE CooperativeWorkGroupLoopContext(IndexType total_size)
  : BaseClass(total_size)
  {}
 
 public:
 
#if defined(ARCCORE_DEVICE_CODE) && !defined(ARCCORE_COMPILING_SYCL)
  __device__ CooperativeDeviceWorkItemGrid grid() const { return CooperativeDeviceWorkItemGrid{}; }
#else
  CooperativeHostWorkItemGrid grid() const { return CooperativeHostWorkItemGrid(m_nb_block, m_syncer); }
#endif
 
 private:
 
  Int32 m_nb_block = 0;
  Impl::ThreadGridSynchronizer* m_syncer = nullptr;
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
/*
 * Implémentation pour SYCL.
 */
#if defined(ARCCORE_COMPILING_SYCL)

class SyclCooperativeDeviceWorkItemGrid
{
  template <typename T> friend class SyclCooperativeWorkGroupLoopContext;
 
 private:
 
  explicit SyclCooperativeDeviceWorkItemGrid(sycl::nd_item<1> n)
  : m_nd_item(n)
  {
  }
 
 public:

  Int32 nbBlock() const { return static_cast<Int32>(m_nd_item.get_group_range(0)); }

  void barrier() { /* Not Yet Implemented */ }
 
 private:
 
  sycl::nd_item<1> m_nd_item;
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
template <typename IndexType_>
class SyclCooperativeWorkGroupLoopContext
: public SyclWorkGroupLoopContextBase<IndexType_>
{
  friend CooperativeWorkGroupLoopRange<IndexType_>;
  friend Impl::WorkGroupLoopContextBuilder;
 
 public:
 
  using IndexType = IndexType_;
 
 private:
 
  // Ce constructeur n'est utilisé que sur le device
  explicit SyclCooperativeWorkGroupLoopContext(sycl::nd_item<1> nd_item, IndexType total_size)
  : SyclWorkGroupLoopContextBase<IndexType_>(nd_item, total_size)
  {
  }
 
 public:

  SyclCooperativeDeviceWorkItemGrid grid() const
  {
    return SyclCooperativeDeviceWorkItemGrid(this->m_nd_item);
  }
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
#endif // ARCCORE_COMPILING_SYCL
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
template <typename IndexType_>
class CooperativeWorkGroupLoopRange
: public WorkGroupLoopRangeBase<true, IndexType_>
{
 public:
 
  using LoopIndexType = CooperativeWorkGroupLoopContext<IndexType_>;
  using IndexType = IndexType_;
 
 public:
 
  CooperativeWorkGroupLoopRange() = default;
  explicit CooperativeWorkGroupLoopRange(IndexType total_nb_element)
  : WorkGroupLoopRangeBase<true, IndexType_>(total_nb_element)
  {}
 
 public:
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
} // namespace Arcane::Accelerator
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
#endif
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/