GenericFilterer.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
//-----------------------------------------------------------------------------
/*---------------------------------------------------------------------------*/
/* GenericFilterer.h                                           (C) 2000-2025 */
/*                                                                           */
/* Algorithme de filtrage.                                                   */
/*---------------------------------------------------------------------------*/
#ifndef ARCANE_ACCELERATOR_GENERICFILTERER_H
#define ARCANE_ACCELERATOR_GENERICFILTERER_H
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
#include "arcane/utils/ArrayView.h"
#include "arcane/utils/FatalErrorException.h"
#include "arcane/utils/NumArray.h"
#include "arcane/utils/TraceInfo.h"
 
#include "arcane/accelerator/core/RunQueue.h"
 
#include "arcane/accelerator/AcceleratorGlobal.h"
#include "arcane/accelerator/CommonUtils.h"
#include "arcane/accelerator/RunCommandLaunchInfo.h"
#include "arcane/accelerator/RunCommandLoop.h"
#include "arcane/accelerator/ScanImpl.h"
#include "arcane/accelerator/MultiThreadAlgo.h"
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
namespace Arcane::Accelerator::impl
{
//#define ARCANE_USE_SCAN_ONEDPL
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
class ARCANE_ACCELERATOR_EXPORT GenericFilteringBase
{
  template <typename DataType, typename FlagType, typename OutputDataType>
  friend class GenericFilteringFlag;
  friend class GenericFilteringIf;
  friend class SyclGenericFilteringImpl;
 
 public:
 
 
 protected:
 
  GenericFilteringBase();
 
 protected:
 
  Int32 _nbOutputElement();
  void _allocate();
  void _allocateTemporaryStorage(size_t size);
  int* _getDeviceNbOutPointer();
  void _copyDeviceNbOutToHostNbOut();
  void _setCalled();
  bool _checkEmpty(Int32 nb_value);
 
 protected:
 
  RunQueue m_queue;
  // Mémoire de travail pour l'algorithme de filtrage.
  GenericDeviceStorage m_algo_storage;
  DeviceStorage<int> m_device_nb_out_storage;
  NumArray<Int32, MDDim1> m_host_nb_out_storage;
  bool m_use_direct_host_storage = true;
 
  bool m_is_already_called = false;
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
#if defined(ARCANE_COMPILING_SYCL)
class SyclGenericFilteringImpl
{
 public:
 
  template <typename SelectLambda, typename InputIterator, typename OutputIterator>
  static void apply(GenericFilteringBase& s, Int32 nb_item, InputIterator input_iter,
                    OutputIterator output_iter, SelectLambda select_lambda)
  {
    RunQueue queue = s.m_queue;
    using DataType = std::iterator_traits<OutputIterator>::value_type;
#if defined(ARCANE_USE_SCAN_ONEDPL) && defined(__INTEL_LLVM_COMPILER)
    sycl::queue true_queue = AcceleratorUtils::toSyclNativeStream(queue);
    auto policy = oneapi::dpl::execution::make_device_policy(true_queue);
    auto out_iter = oneapi::dpl::copy_if(policy, input_iter, input_iter + nb_item, output_iter, select_lambda);
    Int32 nb_output = out_iter - output_iter;
    s.m_host_nb_out_storage[0] = nb_output;
#else
    NumArray<Int32, MDDim1> scan_input_data(nb_item);
    NumArray<Int32, MDDim1> scan_output_data(nb_item);
    SmallSpan<Int32> in_scan_data = scan_input_data.to1DSmallSpan();
    SmallSpan<Int32> out_scan_data = scan_output_data.to1DSmallSpan();
    {
      auto command = makeCommand(queue);
      command << RUNCOMMAND_LOOP1(iter, nb_item)
      {
        auto [i] = iter();
        in_scan_data[i] = select_lambda(input_iter[i]) ? 1 : 0;
      };
    }
    queue.barrier();
    SyclScanner<false /*is_exclusive*/, Int32, ScannerSumOperator<Int32>> scanner;
    scanner.doScan(queue, in_scan_data, out_scan_data, 0);
    // La valeur de 'out_data' pour le dernier élément (nb_item-1) contient la taille du filtre
    Int32 nb_output = out_scan_data[nb_item - 1];
    s.m_host_nb_out_storage[0] = nb_output;
 
    const bool do_verbose = false;
    if (do_verbose && nb_item < 1500)
      for (int i = 0; i < nb_item; ++i) {
        std::cout << "out_data i=" << i << " out_data=" << out_scan_data[i]
                  << " in_data=" << in_scan_data[i] << " value=" << input_iter[i] << "\n ";
      }
    // Copie depuis 'out_data' vers 'in_data' les indices correspondant au filtre
    // Comme 'output_iter' et 'input_iter' peuvent se chevaucher, il
    // faut faire une copie intermédiaire
    // TODO: détecter cela et ne faire la copie que si nécessaire.
    NumArray<DataType,MDDim1> out_copy(eMemoryRessource::Device);
    out_copy.resize(nb_output);
    auto out_copy_view = out_copy.to1DSpan();
    {
      auto command = makeCommand(queue);
      command << RUNCOMMAND_LOOP1(iter, nb_item)
      {
        auto [i] = iter();
        if (in_scan_data[i] == 1)
          out_copy_view[out_scan_data[i] - 1] = input_iter[i];
      };
    }
    {
      auto command = makeCommand(queue);
      command << RUNCOMMAND_LOOP1(iter, nb_output)
      {
        auto [i] = iter();
        output_iter[i] = out_copy_view[i];
      };
    }
    // Obligatoire à cause de 'out_copy'. On pourra le supprimer avec une
    // allocation temporaire.
    queue.barrier();
#endif
  }
};
#endif
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
template <typename DataType, typename FlagType, typename OutputDataType>
class GenericFilteringFlag
{
 public:
 
  void apply(GenericFilteringBase& s, SmallSpan<const DataType> input,
             SmallSpan<OutputDataType> output, SmallSpan<const FlagType> flag)
  {
    const Int32 nb_item = input.size();
    if (output.size() != nb_item)
      ARCANE_FATAL("Sizes are not equals: input={0} output={1}", nb_item, output.size());
    [[maybe_unused]] const DataType* input_data = input.data();
    [[maybe_unused]] DataType* output_data = output.data();
    [[maybe_unused]] const FlagType* flag_data = flag.data();
    eExecutionPolicy exec_policy = eExecutionPolicy::Sequential;
    RunQueue queue = s.m_queue;
    exec_policy = queue.executionPolicy();
    switch (exec_policy) {
#if defined(ARCANE_COMPILING_CUDA)
    case eExecutionPolicy::CUDA: {
      size_t temp_storage_size = 0;
      cudaStream_t stream = AcceleratorUtils::toCudaNativeStream(queue);
      // Premier appel pour connaitre la taille pour l'allocation
      int* nb_out_ptr = nullptr;
      ARCANE_CHECK_CUDA(::cub::DeviceSelect::Flagged(nullptr, temp_storage_size,
                                                     input_data, flag_data, output_data, nb_out_ptr, nb_item, stream));
 
      s._allocateTemporaryStorage(temp_storage_size);
      nb_out_ptr = s._getDeviceNbOutPointer();
      ARCANE_CHECK_CUDA(::cub::DeviceSelect::Flagged(s.m_algo_storage.address(), temp_storage_size,
                                                     input_data, flag_data, output_data, nb_out_ptr, nb_item, stream));
      s._copyDeviceNbOutToHostNbOut();
    } break;
#endif
#if defined(ARCANE_COMPILING_HIP)
    case eExecutionPolicy::HIP: {
      size_t temp_storage_size = 0;
      // Premier appel pour connaitre la taille pour l'allocation
      hipStream_t stream = AcceleratorUtils::toHipNativeStream(queue);
      int* nb_out_ptr = nullptr;
      ARCANE_CHECK_HIP(rocprim::select(nullptr, temp_storage_size, input_data, flag_data, output_data,
                                       nb_out_ptr, nb_item, stream));
 
      s._allocateTemporaryStorage(temp_storage_size);
      nb_out_ptr = s._getDeviceNbOutPointer();
 
      ARCANE_CHECK_HIP(rocprim::select(s.m_algo_storage.address(), temp_storage_size, input_data, flag_data, output_data,
                                       nb_out_ptr, nb_item, stream));
      s._copyDeviceNbOutToHostNbOut();
    } break;
#endif
#if defined(ARCANE_COMPILING_SYCL)
    case eExecutionPolicy::SYCL: {
      impl::IndexIterator iter2(0);
      auto filter_lambda = [=](Int32 input_index) -> bool { return flag[input_index] != 0; };
      auto setter_lambda = [=](Int32 input_index, Int32 output_index) { output[output_index] = input[input_index]; };
      impl::SetterLambdaIterator<decltype(setter_lambda)> out(setter_lambda);
      SyclGenericFilteringImpl::apply(s, nb_item, iter2, out, filter_lambda);
    } break;
#endif
    case eExecutionPolicy::Thread:
      // Pas encore implémenté en multi-thread
      [[fallthrough]];
    case eExecutionPolicy::Sequential: {
      Int32 index = 0;
      for (Int32 i = 0; i < nb_item; ++i) {
        if (flag[i] != 0) {
          output[index] = input[i];
          ++index;
        }
      }
      s.m_host_nb_out_storage[0] = index;
    } break;
    default:
      ARCANE_FATAL(getBadPolicyMessage(exec_policy));
    }
  }
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
class GenericFilteringIf
{
 public:
 
  template <bool InPlace, typename SelectLambda, typename InputIterator, typename OutputIterator>
  void apply(GenericFilteringBase& s, Int32 nb_item, InputIterator input_iter, OutputIterator output_iter,
             const SelectLambda& select_lambda, const TraceInfo& trace_info)
  {
    eExecutionPolicy exec_policy = eExecutionPolicy::Sequential;
    RunQueue queue = s.m_queue;
    exec_policy = queue.executionPolicy();
    RunCommand command = makeCommand(queue);
    command << trace_info;
    impl::RunCommandLaunchInfo launch_info(command, nb_item);
    launch_info.beginExecute();
    switch (exec_policy) {
#if defined(ARCANE_COMPILING_CUDA)
    case eExecutionPolicy::CUDA: {
      size_t temp_storage_size = 0;
      cudaStream_t stream = impl::CudaUtils::toNativeStream(queue);
      // Premier appel pour connaitre la taille pour l'allocation
      int* nb_out_ptr = nullptr;
      if constexpr (InPlace)
        ARCANE_CHECK_CUDA(::cub::DeviceSelect::If(nullptr, temp_storage_size,
                                                  input_iter, nb_out_ptr, nb_item,
                                                  select_lambda, stream));
      else
        ARCANE_CHECK_CUDA(::cub::DeviceSelect::If(nullptr, temp_storage_size,
                                                  input_iter, output_iter, nb_out_ptr, nb_item,
                                                  select_lambda, stream));
 
      s._allocateTemporaryStorage(temp_storage_size);
      nb_out_ptr = s._getDeviceNbOutPointer();
      if constexpr (InPlace)
        ARCANE_CHECK_CUDA(::cub::DeviceSelect::If(s.m_algo_storage.address(), temp_storage_size,
                                                  input_iter, nb_out_ptr, nb_item,
                                                  select_lambda, stream));
      else
        ARCANE_CHECK_CUDA(::cub::DeviceSelect::If(s.m_algo_storage.address(), temp_storage_size,
                                                  input_iter, output_iter, nb_out_ptr, nb_item,
                                                  select_lambda, stream));
 
      s._copyDeviceNbOutToHostNbOut();
    } break;
#endif
#if defined(ARCANE_COMPILING_HIP)
    case eExecutionPolicy::HIP: {
      size_t temp_storage_size = 0;
      // Premier appel pour connaitre la taille pour l'allocation
      hipStream_t stream = impl::HipUtils::toNativeStream(queue);
      int* nb_out_ptr = nullptr;
      // NOTE: il n'y a pas de version spécifique de 'select' en-place.
      // A priori il est possible que \a input_iter et \a output_iter
      // aient la même valeur.
      ARCANE_CHECK_HIP(rocprim::select(nullptr, temp_storage_size, input_iter, output_iter,
                                       nb_out_ptr, nb_item, select_lambda, stream));
      s._allocateTemporaryStorage(temp_storage_size);
      nb_out_ptr = s._getDeviceNbOutPointer();
      ARCANE_CHECK_HIP(rocprim::select(s.m_algo_storage.address(), temp_storage_size, input_iter, output_iter,
                                       nb_out_ptr, nb_item, select_lambda, 0));
      s._copyDeviceNbOutToHostNbOut();
    } break;
#endif
#if defined(ARCANE_COMPILING_SYCL)
    case eExecutionPolicy::SYCL: {
      SyclGenericFilteringImpl::apply(s, nb_item, input_iter, output_iter, select_lambda);
    } break;
#endif
    case eExecutionPolicy::Thread:
      if (nb_item > 500) {
        MultiThreadAlgo scanner;
        Int32 v = scanner.doFilter<InPlace>(launch_info.loopRunInfo(), nb_item, input_iter, output_iter, select_lambda);
        s.m_host_nb_out_storage[0] = v;
        break;
      }
      [[fallthrough]];
    case eExecutionPolicy::Sequential: {
      Int32 index = 0;
      for (Int32 i = 0; i < nb_item; ++i) {
        if (select_lambda(*input_iter)) {
          *output_iter = *input_iter;
          ++index;
          ++output_iter;
        }
        ++input_iter;
      }
      s.m_host_nb_out_storage[0] = index;
    } break;
    default:
      ARCANE_FATAL(getBadPolicyMessage(exec_policy));
    }
    launch_info.endExecute();
  }
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
} // namespace Arcane::Accelerator::impl
 
namespace Arcane::Accelerator
{
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
class GenericFilterer
: private impl::GenericFilteringBase
{
 
 public:
 
  ARCANE_DEPRECATED_REASON("Y2024: Use GenericFilterer(const RunQueue&) instead")
  explicit GenericFilterer(RunQueue* queue)
  {
    ARCANE_CHECK_POINTER(queue);
    m_queue = *queue;
    _allocate();
  }
 
  explicit GenericFilterer(const RunQueue& queue)
  {
    m_queue = queue;
    _allocate();
  }
 
 public:
 
  template <typename InputDataType, typename OutputDataType, typename FlagType>
  void apply(SmallSpan<const InputDataType> input, SmallSpan<OutputDataType> output, SmallSpan<const FlagType> flag)
  {
    const Int32 nb_value = input.size();
    if (output.size() != nb_value)
      ARCANE_FATAL("Sizes are not equals: input={0} output={1}", nb_value, output.size());
    if (flag.size() != nb_value)
      ARCANE_FATAL("Sizes are not equals: input={0} flag={1}", nb_value, flag.size());
 
    if (_checkEmpty(nb_value))
      return;
    _setCalled();
    impl::GenericFilteringBase* base_ptr = this;
    impl::GenericFilteringFlag<InputDataType, FlagType, OutputDataType> gf;
    gf.apply(*base_ptr, input, output, flag);
  }
 
  template <typename DataType, typename SelectLambda>
  void applyIf(SmallSpan<const DataType> input, SmallSpan<DataType> output,
               const SelectLambda& select_lambda, const TraceInfo& trace_info = TraceInfo())
  {
    const Int32 nb_value = input.size();
    if (output.size() != nb_value)
      ARCANE_FATAL("Sizes are not equals: input={0} output={1}", nb_value, output.size());
    if (input.data() == output.data())
      ARCANE_FATAL("Input and Output are the same. Use in place overload instead");
    if (_checkEmpty(nb_value))
      return;
    _setCalled();
    impl::GenericFilteringBase* base_ptr = this;
    impl::GenericFilteringIf gf;
    gf.apply<false>(*base_ptr, nb_value, input.data(), output.data(), select_lambda, trace_info);
  }
 
  template <typename DataType, typename SelectLambda>
  void applyIf(SmallSpan<DataType> input_output, const SelectLambda& select_lambda,
               const TraceInfo& trace_info = TraceInfo())
  {
    const Int32 nb_value = input_output.size();
    if (_checkEmpty(nb_value))
      return;
    _setCalled();
    impl::GenericFilteringBase* base_ptr = this;
    impl::GenericFilteringIf gf;
    gf.apply<true>(*base_ptr, nb_value, input_output.data(), input_output.data(), select_lambda, trace_info);
  }
 
  template <typename InputIterator, typename OutputIterator, typename SelectLambda>
  void applyIf(Int32 nb_value, InputIterator input_iter, OutputIterator output_iter,
               const SelectLambda& select_lambda, const TraceInfo& trace_info = TraceInfo())
  {
    if (_checkEmpty(nb_value))
      return;
    _setCalled();
    impl::GenericFilteringBase* base_ptr = this;
    impl::GenericFilteringIf gf;
    gf.apply<false>(*base_ptr, nb_value, input_iter, output_iter, select_lambda, trace_info);
  }
 
  template <typename SelectLambda, typename SetterLambda>
  void applyWithIndex(Int32 nb_value, const SelectLambda& select_lambda,
                      const SetterLambda& setter_lambda, const TraceInfo& trace_info = TraceInfo())
  {
    if (_checkEmpty(nb_value))
      return;
    _setCalled();
    impl::GenericFilteringBase* base_ptr = this;
    impl::GenericFilteringIf gf;
    impl::IndexIterator input_iter;
    impl::SetterLambdaIterator<SetterLambda> out(setter_lambda);
    gf.apply<false>(*base_ptr, nb_value, input_iter, out, select_lambda, trace_info);
  }
 
  Int32 nbOutputElement()
  {
    return _nbOutputElement();
  }
};
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
} // namespace Arcane::Accelerator
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
 
#endif
 
/*---------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/