handler.hpp

Go to the documentation of this file.

//==-------- handler.hpp --- SYCL command group handler --------------------==//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
 
#pragma once
 
#include <sycl/access/access.hpp>
#include <sycl/accessor.hpp>
#include <sycl/context.hpp>
#include <sycl/detail/cg.hpp>
#include <sycl/detail/cg_types.hpp>
#include <sycl/detail/cl.h>
#include <sycl/detail/export.hpp>
#include <sycl/detail/handler_proxy.hpp>
#include <sycl/detail/os_util.hpp>
#include <sycl/event.hpp>
#include <sycl/id.hpp>
#include <sycl/interop_handle.hpp>
#include <sycl/item.hpp>
#include <sycl/kernel.hpp>
#include <sycl/kernel_bundle.hpp>
#include <sycl/kernel_handler.hpp>
#include <sycl/nd_item.hpp>
#include <sycl/nd_range.hpp>
#include <sycl/property_list.hpp>
#include <sycl/sampler.hpp>
#include <sycl/stl.hpp>
 
#include <functional>
#include <limits>
#include <memory>
#include <tuple>
#include <type_traits>
 
// SYCL_LANGUAGE_VERSION is 4 digit year followed by 2 digit revision
#if !SYCL_LANGUAGE_VERSION || SYCL_LANGUAGE_VERSION < 202001
#define __SYCL_NONCONST_FUNCTOR__
#endif
 
template <typename DataT, int Dimensions, sycl::access::mode AccessMode,
          sycl::access::target AccessTarget,
          sycl::access::placeholder IsPlaceholder>
class __fill;
 
template <typename T> class __usmfill;
 
template <typename T_Src, typename T_Dst, int Dims,
          sycl::access::mode AccessMode, sycl::access::target AccessTarget,
          sycl::access::placeholder IsPlaceholder>
class __copyAcc2Ptr;
 
template <typename T_Src, typename T_Dst, int Dims,
          sycl::access::mode AccessMode, sycl::access::target AccessTarget,
          sycl::access::placeholder IsPlaceholder>
class __copyPtr2Acc;
 
template <typename T_Src, int Dims_Src, sycl::access::mode AccessMode_Src,
          sycl::access::target AccessTarget_Src, typename T_Dst, int Dims_Dst,
          sycl::access::mode AccessMode_Dst,
          sycl::access::target AccessTarget_Dst,
          sycl::access::placeholder IsPlaceholder_Src,
          sycl::access::placeholder IsPlaceholder_Dst>
class __copyAcc2Acc;
 
// For unit testing purposes
class MockHandler;
 
namespace sycl {
__SYCL_INLINE_VER_NAMESPACE(_V1) {
 
// Forward declaration
 
class handler;
template <typename T, int Dimensions, typename AllocatorT, typename Enable>
class buffer;
namespace detail {
 
class handler_impl;
class kernel_impl;
class queue_impl;
class stream_impl;
template <typename DataT, int Dimensions, access::mode AccessMode,
          access::target AccessTarget, access::placeholder IsPlaceholder>
class image_accessor;
template <typename RetType, typename Func, typename Arg>
static Arg member_ptr_helper(RetType (Func::*)(Arg) const);
 
// Non-const version of the above template to match functors whose 'operator()'
// is declared w/o the 'const' qualifier.
template <typename RetType, typename Func, typename Arg>
static Arg member_ptr_helper(RetType (Func::*)(Arg));
 
// template <typename RetType, typename Func>
// static void member_ptr_helper(RetType (Func::*)() const);
 
// template <typename RetType, typename Func>
// static void member_ptr_helper(RetType (Func::*)());
 
template <typename F, typename SuggestedArgType>
decltype(member_ptr_helper(&F::operator())) argument_helper(int);
 
template <typename F, typename SuggestedArgType>
SuggestedArgType argument_helper(...);
 
template <typename F, typename SuggestedArgType>
using lambda_arg_type = decltype(argument_helper<F, SuggestedArgType>(0));
 
// Used when parallel_for range is rounded-up.
template <typename Name> class __pf_kernel_wrapper;
 
template <typename Type> struct get_kernel_wrapper_name_t {
  using name = __pf_kernel_wrapper<Type>;
};
 
__SYCL_EXPORT device getDeviceFromHandler(handler &);
 
#if __SYCL_ID_QUERIES_FIT_IN_INT__
template <typename T> struct NotIntMsg;
 
template <int Dims> struct NotIntMsg<range<Dims>> {
  constexpr static const char *Msg =
      "Provided range is out of integer limits. Pass "
      "`-fno-sycl-id-queries-fit-in-int' to disable range check.";
};
 
template <int Dims> struct NotIntMsg<id<Dims>> {
  constexpr static const char *Msg =
      "Provided offset is out of integer limits. Pass "
      "`-fno-sycl-id-queries-fit-in-int' to disable offset check.";
};
#endif
 
#if __SYCL_ID_QUERIES_FIT_IN_INT__
template <typename T, typename ValT>
typename detail::enable_if_t<std::is_same<ValT, size_t>::value ||
                             std::is_same<ValT, unsigned long long>::value>
checkValueRangeImpl(ValT V) {
  static constexpr size_t Limit =
      static_cast<size_t>((std::numeric_limits<int>::max)());
  if (V > Limit)
    throw runtime_error(NotIntMsg<T>::Msg, PI_ERROR_INVALID_VALUE);
}
#endif
 
template <int Dims, typename T>
typename detail::enable_if_t<std::is_same<T, range<Dims>>::value ||
                             std::is_same<T, id<Dims>>::value>
checkValueRange(const T &V) {
#if __SYCL_ID_QUERIES_FIT_IN_INT__
  for (size_t Dim = 0; Dim < Dims; ++Dim)
    checkValueRangeImpl<T>(V[Dim]);
 
  {
    unsigned long long Product = 1;
    for (size_t Dim = 0; Dim < Dims; ++Dim) {
      Product *= V[Dim];
      // check value now to prevent product overflow in the end
      checkValueRangeImpl<T>(Product);
    }
  }
#else
  (void)V;
#endif
}
 
template <int Dims>
void checkValueRange(const range<Dims> &R, const id<Dims> &O) {
#if __SYCL_ID_QUERIES_FIT_IN_INT__
  checkValueRange<Dims>(R);
  checkValueRange<Dims>(O);
 
  for (size_t Dim = 0; Dim < Dims; ++Dim) {
    unsigned long long Sum = R[Dim] + O[Dim];
 
    checkValueRangeImpl<range<Dims>>(Sum);
  }
#else
  (void)R;
  (void)O;
#endif
}
 
template <int Dims, typename T>
typename detail::enable_if_t<std::is_same<T, nd_range<Dims>>::value>
checkValueRange(const T &V) {
#if __SYCL_ID_QUERIES_FIT_IN_INT__
  checkValueRange<Dims>(V.get_global_range());
  checkValueRange<Dims>(V.get_local_range());
  checkValueRange<Dims>(V.get_offset());
 
  checkValueRange<Dims>(V.get_global_range(), V.get_offset());
#else
  (void)V;
#endif
}
 
template <typename TransformedArgType, int Dims, typename KernelType>
class RoundedRangeKernel {
public:
  RoundedRangeKernel(range<Dims> NumWorkItems, KernelType KernelFunc)
      : NumWorkItems(NumWorkItems), KernelFunc(KernelFunc) {}
 
  void operator()(TransformedArgType Arg) const {
    if (Arg[0] >= NumWorkItems[0])
      return;
    Arg.set_allowed_range(NumWorkItems);
    KernelFunc(Arg);
  }
 
private:
  range<Dims> NumWorkItems;
  KernelType KernelFunc;
};
 
template <typename TransformedArgType, int Dims, typename KernelType>
class RoundedRangeKernelWithKH {
public:
  RoundedRangeKernelWithKH(range<Dims> NumWorkItems, KernelType KernelFunc)
      : NumWorkItems(NumWorkItems), KernelFunc(KernelFunc) {}
 
  void operator()(TransformedArgType Arg, kernel_handler KH) const {
    if (Arg[0] >= NumWorkItems[0])
      return;
    Arg.set_allowed_range(NumWorkItems);
    KernelFunc(Arg, KH);
  }
 
private:
  range<Dims> NumWorkItems;
  KernelType KernelFunc;
};
 
template <typename T, class BinaryOperation, int Dims, size_t Extent,
          typename RedOutVar>
class reduction_impl_algo;
 
using sycl::detail::enable_if_t;
using sycl::detail::queue_impl;
 
// Kernels with single reduction
 
template <typename KernelName, typename KernelType, int Dims, class Reduction>
bool reduCGFuncForRange(handler &CGH, KernelType KernelFunc,
                        const range<Dims> &Range, size_t MaxWGSize,
                        uint32_t NumConcurrentWorkGroups, Reduction &Redu);
 
template <typename KernelName, typename KernelType, int Dims, class Reduction>
void reduCGFuncAtomic64(handler &CGH, KernelType KernelFunc,
                        const nd_range<Dims> &Range, Reduction &Redu);
 
template <typename KernelName, typename KernelType, int Dims, class Reduction>
void reduCGFunc(handler &CGH, KernelType KernelFunc,
                const nd_range<Dims> &Range, Reduction &Redu);
 
// Kernels with multiple reductions
 
// sycl::nd_range version
template <typename KernelName, typename KernelType, int Dims,
          typename... Reductions, size_t... Is>
void reduCGFuncMulti(handler &CGH, KernelType KernelFunc,
                     const nd_range<Dims> &Range,
                     std::tuple<Reductions...> &ReduTuple,
                     std::index_sequence<Is...>);
 
template <typename KernelName, typename KernelType, class Reduction>
size_t reduAuxCGFunc(handler &CGH, size_t NWorkItems, size_t MaxWGSize,
                     Reduction &Redu);
 
template <typename KernelName, typename KernelType, typename... Reductions,
          size_t... Is>
size_t reduAuxCGFunc(handler &CGH, size_t NWorkItems, size_t MaxWGSize,
                     std::tuple<Reductions...> &ReduTuple,
                     std::index_sequence<Is...>);
 
template <typename KernelName, class Reduction>
std::enable_if_t<!Reduction::is_usm>
reduSaveFinalResultToUserMem(handler &CGH, Reduction &Redu);
 
template <typename KernelName, class Reduction>
std::enable_if_t<Reduction::is_usm>
reduSaveFinalResultToUserMem(handler &CGH, Reduction &Redu);
 
template <typename... Reduction, size_t... Is>
std::shared_ptr<event>
reduSaveFinalResultToUserMem(std::shared_ptr<detail::queue_impl> Queue,
                             bool IsHost, std::tuple<Reduction...> &ReduTuple,
                             std::index_sequence<Is...>);
 
__SYCL_EXPORT uint32_t
reduGetMaxNumConcurrentWorkGroups(std::shared_ptr<queue_impl> Queue);
 
__SYCL_EXPORT size_t reduGetMaxWGSize(std::shared_ptr<queue_impl> Queue,
                                      size_t LocalMemBytesPerWorkItem);
 
__SYCL_EXPORT size_t reduGetPreferredWGSize(std::shared_ptr<queue_impl> &Queue,
                                            size_t LocalMemBytesPerWorkItem);
 
template <typename... ReductionT, size_t... Is>
size_t reduGetMemPerWorkItem(std::tuple<ReductionT...> &ReduTuple,
                             std::index_sequence<Is...>);
 
template <typename TupleT, std::size_t... Is>
std::tuple<std::tuple_element_t<Is, TupleT>...>
tuple_select_elements(TupleT Tuple, std::index_sequence<Is...>);
 
template <typename FirstT, typename... RestT> struct AreAllButLastReductions;
 
template <class FunctorTy>
event withAuxHandler(std::shared_ptr<detail::queue_impl> Queue, bool IsHost,
                     FunctorTy Func);
} // namespace detail
 
class __SYCL_EXPORT handler {
private:
  handler(std::shared_ptr<detail::queue_impl> Queue, bool IsHost);
 
  handler(std::shared_ptr<detail::queue_impl> Queue,
          std::shared_ptr<detail::queue_impl> PrimaryQueue,
          std::shared_ptr<detail::queue_impl> SecondaryQueue, bool IsHost);
 
  template <typename T, typename F = typename detail::remove_const_t<
                            typename detail::remove_reference_t<T>>>
  F *storePlainArg(T &&Arg) {
    MArgsStorage.emplace_back(sizeof(T));
    auto Storage = reinterpret_cast<F *>(MArgsStorage.back().data());
    *Storage = Arg;
    return Storage;
  }
 
  void setType(detail::CG::CGTYPE Type) { MCGType = Type; }
 
  detail::CG::CGTYPE getType() { return MCGType; }
 
  void throwIfActionIsCreated() {
    if (detail::CG::None != getType())
      throw sycl::runtime_error("Attempt to set multiple actions for the "
                                "command group. Command group must consist of "
                                "a single kernel or explicit memory operation.",
                                PI_ERROR_INVALID_OPERATION);
  }
 
  void
  extractArgsAndReqsFromLambda(char *LambdaPtr, size_t KernelArgsNum,
                               const detail::kernel_param_desc_t *KernelArgs,
                               bool IsESIMD);
 
  void extractArgsAndReqs();
 
  void processArg(void *Ptr, const detail::kernel_param_kind_t &Kind,
                  const int Size, const size_t Index, size_t &IndexShift,
                  bool IsKernelCreatedFromSource, bool IsESIMD);
 
  std::string getKernelName();
 
  template <typename LambdaNameT> bool lambdaAndKernelHaveEqualName() {
    // TODO It is unclear a kernel and a lambda/functor must to be equal or not
    // for parallel_for with sycl::kernel and lambda/functor together
    // Now if they are equal we extract argumets from lambda/functor for the
    // kernel. Else it is necessary use set_atg(s) for resolve the order and
    // values of arguments for the kernel.
    assert(MKernel && "MKernel is not initialized");
    const std::string LambdaName = detail::KernelInfo<LambdaNameT>::getName();
    const std::string KernelName = getKernelName();
    return LambdaName == KernelName;
  }
 
  void saveCodeLoc(detail::code_location CodeLoc) { MCodeLoc = CodeLoc; }
 
  event finalize();
 
  void addStream(const std::shared_ptr<detail::stream_impl> &Stream) {
    MStreamStorage.push_back(Stream);
  }
 
  template <class FunctorTy>
  event withAuxHandler(std::shared_ptr<detail::queue_impl> Queue,
                       FunctorTy Func) {
    handler AuxHandler(Queue, MIsHost);
    AuxHandler.saveCodeLoc(MCodeLoc);
    Func(AuxHandler);
    return AuxHandler.finalize();
  }
 
  template <class FunctorTy>
  friend event detail::withAuxHandler(std::shared_ptr<detail::queue_impl> Queue,
                                      bool IsHost, FunctorTy Func);
 
  void addReduction(const std::shared_ptr<const void> &ReduObj);
 
  ~handler() = default;
 
  // TODO: Private and unusued. Remove when ABI break is allowed.
  bool is_host() { return MIsHost; }
 
#ifdef __SYCL_DEVICE_ONLY__
  // In device compilation accessor isn't inherited from AccessorBaseHost, so
  // can't detect by it. Since we don't expect it to be ever called in device
  // execution, just use blind void *.
  void associateWithHandler(void *AccBase, access::target AccTarget);
#else
  void associateWithHandler(detail::AccessorBaseHost *AccBase,
                            access::target AccTarget);
#endif
 
  // Recursively calls itself until arguments pack is fully processed.
  // The version for regular(standard layout) argument.
  template <typename T, typename... Ts>
  void setArgsHelper(int ArgIndex, T &&Arg, Ts &&...Args) {
    set_arg(ArgIndex, std::move(Arg));
    setArgsHelper(++ArgIndex, std::move(Args)...);
  }
 
  void setArgsHelper(int) {}
 
  // setArgHelper for local accessor argument.
  template <typename DataT, int Dims, access::mode AccessMode,
            access::placeholder IsPlaceholder>
  void setArgHelper(int ArgIndex,
                    accessor<DataT, Dims, AccessMode, access::target::local,
                             IsPlaceholder> &&Arg) {
    detail::LocalAccessorBaseHost *LocalAccBase =
        (detail::LocalAccessorBaseHost *)&Arg;
    detail::LocalAccessorImplPtr LocalAccImpl =
        detail::getSyclObjImpl(*LocalAccBase);
    detail::LocalAccessorImplHost *Req = LocalAccImpl.get();
    MLocalAccStorage.push_back(std::move(LocalAccImpl));
    MArgs.emplace_back(detail::kernel_param_kind_t::kind_accessor, Req,
                       static_cast<int>(access::target::local), ArgIndex);
  }
 
  // setArgHelper for non local accessor argument.
  template <typename DataT, int Dims, access::mode AccessMode,
            access::target AccessTarget, access::placeholder IsPlaceholder>
  typename detail::enable_if_t<AccessTarget != access::target::local, void>
  setArgHelper(
      int ArgIndex,
      accessor<DataT, Dims, AccessMode, AccessTarget, IsPlaceholder> &&Arg) {
    detail::AccessorBaseHost *AccBase = (detail::AccessorBaseHost *)&Arg;
    detail::AccessorImplPtr AccImpl = detail::getSyclObjImpl(*AccBase);
    detail::AccessorImplHost *Req = AccImpl.get();
    // Add accessor to the list of requirements.
    MRequirements.push_back(Req);
    // Store copy of the accessor.
    MAccStorage.push_back(std::move(AccImpl));
    // Add accessor to the list of arguments.
    MArgs.emplace_back(detail::kernel_param_kind_t::kind_accessor, Req,
                       static_cast<int>(AccessTarget), ArgIndex);
  }
 
  template <typename T> void setArgHelper(int ArgIndex, T &&Arg) {
    auto StoredArg = static_cast<void *>(storePlainArg(Arg));
 
    if (!std::is_same<cl_mem, T>::value && std::is_pointer<T>::value) {
      MArgs.emplace_back(detail::kernel_param_kind_t::kind_pointer, StoredArg,
                         sizeof(T), ArgIndex);
    } else {
      MArgs.emplace_back(detail::kernel_param_kind_t::kind_std_layout,
                         StoredArg, sizeof(T), ArgIndex);
    }
  }
 
  void setArgHelper(int ArgIndex, sampler &&Arg) {
    auto StoredArg = static_cast<void *>(storePlainArg(Arg));
    MArgs.emplace_back(detail::kernel_param_kind_t::kind_sampler, StoredArg,
                       sizeof(sampler), ArgIndex);
  }
 
  // TODO: Unusued. Remove when ABI break is allowed.
  void verifyKernelInvoc(const kernel &Kernel) {
    std::ignore = Kernel;
    return;
  }
 
  /* The kernel passed to StoreLambda can take an id, an item or an nd_item as
   * its argument. Since esimd plugin directly invokes the kernel (doesn’t use
   * piKernelSetArg), the kernel argument type must be known to the plugin.
   * However, passing kernel argument type to the plugin requires changing ABI
   * in HostKernel class. To overcome this problem, helpers below wrap the
   * “original” kernel with a functor that always takes an nd_item as argument.
   * A functor is used instead of a lambda because extractArgsAndReqsFromLambda
   * needs access to the “original” kernel and keeps references to its internal
   * data, i.e. the kernel passed as argument cannot be local in scope. The
   * functor itself is again encapsulated in a std::function since functor’s
   * type is unknown to the plugin.
   */
 
  // For 'id, item w/wo offset, nd_item' kernel arguments
  template <class KernelType, class NormalizedKernelType, int Dims>
  KernelType *ResetHostKernelHelper(const KernelType &KernelFunc) {
    NormalizedKernelType NormalizedKernel(KernelFunc);
    auto NormalizedKernelFunc =
        std::function<void(const sycl::nd_item<Dims> &)>(NormalizedKernel);
    auto HostKernelPtr =
        new detail::HostKernel<decltype(NormalizedKernelFunc),
                               sycl::nd_item<Dims>, Dims>(NormalizedKernelFunc);
    MHostKernel.reset(HostKernelPtr);
    return &HostKernelPtr->MKernel.template target<NormalizedKernelType>()
                ->MKernelFunc;
  }
 
  // For 'sycl::id<Dims>' kernel argument
  template <class KernelType, typename ArgT, int Dims>
  typename std::enable_if<std::is_same<ArgT, sycl::id<Dims>>::value,
                          KernelType *>::type
  ResetHostKernel(const KernelType &KernelFunc) {
    struct NormalizedKernelType {
      KernelType MKernelFunc;
      NormalizedKernelType(const KernelType &KernelFunc)
          : MKernelFunc(KernelFunc) {}
      void operator()(const nd_item<Dims> &Arg) {
        detail::runKernelWithArg(MKernelFunc, Arg.get_global_id());
      }
    };
    return ResetHostKernelHelper<KernelType, struct NormalizedKernelType, Dims>(
        KernelFunc);
  }
 
  // For 'sycl::nd_item<Dims>' kernel argument
  template <class KernelType, typename ArgT, int Dims>
  typename std::enable_if<std::is_same<ArgT, sycl::nd_item<Dims>>::value,
                          KernelType *>::type
  ResetHostKernel(const KernelType &KernelFunc) {
    struct NormalizedKernelType {
      KernelType MKernelFunc;
      NormalizedKernelType(const KernelType &KernelFunc)
          : MKernelFunc(KernelFunc) {}
      void operator()(const nd_item<Dims> &Arg) {
        detail::runKernelWithArg(MKernelFunc, Arg);
      }
    };
    return ResetHostKernelHelper<KernelType, struct NormalizedKernelType, Dims>(
        KernelFunc);
  }
 
  // For 'sycl::item<Dims, without_offset>' kernel argument
  template <class KernelType, typename ArgT, int Dims>
  typename std::enable_if<std::is_same<ArgT, sycl::item<Dims, false>>::value,
                          KernelType *>::type
  ResetHostKernel(const KernelType &KernelFunc) {
    struct NormalizedKernelType {
      KernelType MKernelFunc;
      NormalizedKernelType(const KernelType &KernelFunc)
          : MKernelFunc(KernelFunc) {}
      void operator()(const nd_item<Dims> &Arg) {
        sycl::item<Dims, false> Item = detail::Builder::createItem<Dims, false>(
            Arg.get_global_range(), Arg.get_global_id());
        detail::runKernelWithArg(MKernelFunc, Item);
      }
    };
    return ResetHostKernelHelper<KernelType, struct NormalizedKernelType, Dims>(
        KernelFunc);
  }
 
  // For 'sycl::item<Dims, with_offset>' kernel argument
  template <class KernelType, typename ArgT, int Dims>
  typename std::enable_if<std::is_same<ArgT, sycl::item<Dims, true>>::value,
                          KernelType *>::type
  ResetHostKernel(const KernelType &KernelFunc) {
    struct NormalizedKernelType {
      KernelType MKernelFunc;
      NormalizedKernelType(const KernelType &KernelFunc)
          : MKernelFunc(KernelFunc) {}
      void operator()(const nd_item<Dims> &Arg) {
        sycl::item<Dims, true> Item = detail::Builder::createItem<Dims, true>(
            Arg.get_global_range(), Arg.get_global_id(), Arg.get_offset());
        detail::runKernelWithArg(MKernelFunc, Item);
      }
    };
    return ResetHostKernelHelper<KernelType, struct NormalizedKernelType, Dims>(
        KernelFunc);
  }
 
  // For 'void' kernel argument (single_task)
  template <class KernelType, typename ArgT, int Dims>
  typename std::enable_if_t<std::is_same<ArgT, void>::value, KernelType *>
  ResetHostKernel(const KernelType &KernelFunc) {
    struct NormalizedKernelType {
      KernelType MKernelFunc;
      NormalizedKernelType(const KernelType &KernelFunc)
          : MKernelFunc(KernelFunc) {}
      void operator()(const nd_item<Dims> &Arg) {
        (void)Arg;
        detail::runKernelWithoutArg(MKernelFunc);
      }
    };
    return ResetHostKernelHelper<KernelType, struct NormalizedKernelType, Dims>(
        KernelFunc);
  }
 
  // For 'sycl::group<Dims>' kernel argument
  // 'wrapper'-based approach using 'NormalizedKernelType' struct is not used
  // for 'void(sycl::group<Dims>)' since 'void(sycl::group<Dims>)' is not
  // supported in ESIMD.
  template <class KernelType, typename ArgT, int Dims>
  typename std::enable_if<std::is_same<ArgT, sycl::group<Dims>>::value,
                          KernelType *>::type
  ResetHostKernel(const KernelType &KernelFunc) {
    MHostKernel.reset(
        new detail::HostKernel<KernelType, ArgT, Dims>(KernelFunc));
    return (KernelType *)(MHostKernel->getPtr());
  }
 
  void verifyUsedKernelBundle(const std::string &KernelName);
 
  template <typename KernelName, typename KernelType, int Dims,
            typename LambdaArgType>
  void StoreLambda(KernelType KernelFunc) {
    using KI = detail::KernelInfo<KernelName>;
 
    constexpr bool IsCallableWithKernelHandler =
        detail::KernelLambdaHasKernelHandlerArgT<KernelType,
                                                 LambdaArgType>::value;
 
    if (IsCallableWithKernelHandler && MIsHost) {
      throw sycl::feature_not_supported(
          "kernel_handler is not yet supported by host device.",
          PI_ERROR_INVALID_OPERATION);
    }
 
    KernelType *KernelPtr =
        ResetHostKernel<KernelType, LambdaArgType, Dims>(KernelFunc);
 
    using KI = sycl::detail::KernelInfo<KernelName>;
    constexpr bool KernelHasName =
        KI::getName() != nullptr && KI::getName()[0] != '\0';
 
    // Some host compilers may have different captures from Clang. Currently
    // there is no stable way of handling this when extracting the captures, so
    // a static assert is made to fail for incompatible kernel lambdas.
    static_assert(
        !KernelHasName || sizeof(KernelFunc) == KI::getKernelSize(),
        "Unexpected kernel lambda size. This can be caused by an "
        "external host compiler producing a lambda with an "
        "unexpected layout. This is a limitation of the compiler."
        "In many cases the difference is related to capturing constexpr "
        "variables. In such cases removing constexpr specifier aligns the "
        "captures between the host compiler and the device compiler."
        "\n"
        "In case of MSVC, passing "
        "-fsycl-host-compiler-options='/std:c++latest' "
        "might also help.");
 
    // Empty name indicates that the compilation happens without integration
    // header, so don't perform things that require it.
    if (KernelHasName) {
      // TODO support ESIMD in no-integration-header case too.
      MArgs.clear();
      extractArgsAndReqsFromLambda(reinterpret_cast<char *>(KernelPtr),
                                   KI::getNumParams(), &KI::getParamDesc(0),
                                   KI::isESIMD());
      MKernelName = KI::getName();
      MOSModuleHandle = detail::OSUtil::getOSModuleHandle(KI::getName());
    } else {
      // In case w/o the integration header it is necessary to process
      // accessors from the list(which are associated with this handler) as
      // arguments.
      MArgs = std::move(MAssociatedAccesors);
    }
 
    // If the kernel lambda is callable with a kernel_handler argument, manifest
    // the associated kernel handler.
    if (IsCallableWithKernelHandler) {
      getOrInsertHandlerKernelBundle(/*Insert=*/true);
    }
  }
 
  template <int Dims_Src, int Dims_Dst>
  static bool IsCopyingRectRegionAvailable(const range<Dims_Src> Src,
                                           const range<Dims_Dst> Dst) {
    if (Dims_Src > Dims_Dst)
      return false;
    for (size_t I = 0; I < Dims_Src; ++I)
      if (Src[I] > Dst[I])
        return false;
    return true;
  }
 
  template <typename TSrc, int DimSrc, access::mode ModeSrc,
            access::target TargetSrc, typename TDst, int DimDst,
            access::mode ModeDst, access::target TargetDst,
            access::placeholder IsPHSrc, access::placeholder IsPHDst>
  detail::enable_if_t<(DimSrc > 0) && (DimDst > 0), bool>
  copyAccToAccHelper(accessor<TSrc, DimSrc, ModeSrc, TargetSrc, IsPHSrc> Src,
                     accessor<TDst, DimDst, ModeDst, TargetDst, IsPHDst> Dst) {
    if (!MIsHost &&
        IsCopyingRectRegionAvailable(Src.get_range(), Dst.get_range()))
      return false;
 
    range<1> LinearizedRange(Src.size());
    parallel_for<
        class __copyAcc2Acc<TSrc, DimSrc, ModeSrc, TargetSrc, TDst, DimDst,
                            ModeDst, TargetDst, IsPHSrc, IsPHDst>>(
        LinearizedRange, [=](id<1> Id) {
          size_t Index = Id[0];
          id<DimSrc> SrcId = detail::getDelinearizedId(Src.get_range(), Index);
          id<DimDst> DstId = detail::getDelinearizedId(Dst.get_range(), Index);
          Dst[DstId] = Src[SrcId];
        });
    return true;
  }
 
  template <typename TSrc, int DimSrc, access::mode ModeSrc,
            access::target TargetSrc, typename TDst, int DimDst,
            access::mode ModeDst, access::target TargetDst,
            access::placeholder IsPHSrc, access::placeholder IsPHDst>
  detail::enable_if_t<DimSrc == 0 || DimDst == 0, bool>
  copyAccToAccHelper(accessor<TSrc, DimSrc, ModeSrc, TargetSrc, IsPHSrc> Src,
                     accessor<TDst, DimDst, ModeDst, TargetDst, IsPHDst> Dst) {
    if (!MIsHost)
      return false;
 
    single_task<
        class __copyAcc2Acc<TSrc, DimSrc, ModeSrc, TargetSrc, TDst, DimDst,
                            ModeDst, TargetDst, IsPHSrc, IsPHDst>>(
        [=]() { *(Dst.get_pointer()) = *(Src.get_pointer()); });
    return true;
  }
 
#ifndef __SYCL_DEVICE_ONLY__
  template <typename TSrc, typename TDst, int Dim, access::mode AccMode,
            access::target AccTarget, access::placeholder IsPH>
  detail::enable_if_t<(Dim > 0)>
  copyAccToPtrHost(accessor<TSrc, Dim, AccMode, AccTarget, IsPH> Src,
                   TDst *Dst) {
    range<Dim> Range = Src.get_range();
    parallel_for<
        class __copyAcc2Ptr<TSrc, TDst, Dim, AccMode, AccTarget, IsPH>>(
        Range, [=](id<Dim> Index) {
          const size_t LinearIndex = detail::getLinearIndex(Index, Range);
          using TSrcNonConst = typename detail::remove_const_t<TSrc>;
          (reinterpret_cast<TSrcNonConst *>(Dst))[LinearIndex] = Src[Index];
        });
  }
 
  template <typename TSrc, typename TDst, int Dim, access::mode AccMode,
            access::target AccTarget, access::placeholder IsPH>
  detail::enable_if_t<Dim == 0>
  copyAccToPtrHost(accessor<TSrc, Dim, AccMode, AccTarget, IsPH> Src,
                   TDst *Dst) {
    single_task<class __copyAcc2Ptr<TSrc, TDst, Dim, AccMode, AccTarget, IsPH>>(
        [=]() {
          using TSrcNonConst = typename detail::remove_const_t<TSrc>;
          *(reinterpret_cast<TSrcNonConst *>(Dst)) = *(Src.get_pointer());
        });
  }
 
  template <typename TSrc, typename TDst, int Dim, access::mode AccMode,
            access::target AccTarget, access::placeholder IsPH>
  detail::enable_if_t<(Dim > 0)>
  copyPtrToAccHost(TSrc *Src,
                   accessor<TDst, Dim, AccMode, AccTarget, IsPH> Dst) {
    range<Dim> Range = Dst.get_range();
    parallel_for<
        class __copyPtr2Acc<TSrc, TDst, Dim, AccMode, AccTarget, IsPH>>(
        Range, [=](id<Dim> Index) {
          const size_t LinearIndex = detail::getLinearIndex(Index, Range);
          Dst[Index] = (reinterpret_cast<const TDst *>(Src))[LinearIndex];
        });
  }
 
  template <typename TSrc, typename TDst, int Dim, access::mode AccMode,
            access::target AccTarget, access::placeholder IsPH>
  detail::enable_if_t<Dim == 0>
  copyPtrToAccHost(TSrc *Src,
                   accessor<TDst, Dim, AccMode, AccTarget, IsPH> Dst) {
    single_task<class __copyPtr2Acc<TSrc, TDst, Dim, AccMode, AccTarget, IsPH>>(
        [=]() {
          *(Dst.get_pointer()) = *(reinterpret_cast<const TDst *>(Src));
        });
  }
#endif // __SYCL_DEVICE_ONLY__
 
  constexpr static bool isConstOrGlobal(access::target AccessTarget) {
    return AccessTarget == access::target::device ||
           AccessTarget == access::target::constant_buffer;
  }
 
  constexpr static bool isImageOrImageArray(access::target AccessTarget) {
    return AccessTarget == access::target::image ||
           AccessTarget == access::target::image_array;
  }
 
  constexpr static bool
  isValidTargetForExplicitOp(access::target AccessTarget) {
    return isConstOrGlobal(AccessTarget) || isImageOrImageArray(AccessTarget);
  }
 
  constexpr static bool isValidModeForSourceAccessor(access::mode AccessMode) {
    return AccessMode == access::mode::read ||
           AccessMode == access::mode::read_write;
  }
 
  constexpr static bool
  isValidModeForDestinationAccessor(access::mode AccessMode) {
    return AccessMode == access::mode::write ||
           AccessMode == access::mode::read_write ||
           AccessMode == access::mode::discard_write ||
           AccessMode == access::mode::discard_read_write;
  }
 
  template <int Dims, typename LambdaArgType> struct TransformUserItemType {
    using type = typename std::conditional<
        std::is_convertible<nd_item<Dims>, LambdaArgType>::value, nd_item<Dims>,
        typename std::conditional<
            std::is_convertible<item<Dims>, LambdaArgType>::value, item<Dims>,
            LambdaArgType>::type>::type;
  };
 
  template <typename KernelName, typename KernelType, int Dims>
  void parallel_for_lambda_impl(range<Dims> NumWorkItems,
                                KernelType KernelFunc) {
    throwIfActionIsCreated();
    using LambdaArgType = sycl::detail::lambda_arg_type<KernelType, item<Dims>>;
 
    // If 1D kernel argument is an integral type, convert it to sycl::item<1>
    // If user type is convertible from sycl::item/sycl::nd_item, use
    // sycl::item/sycl::nd_item to transport item information
    using TransformedArgType = typename std::conditional<
        std::is_integral<LambdaArgType>::value && Dims == 1, item<Dims>,
        typename TransformUserItemType<Dims, LambdaArgType>::type>::type;
 
    using NameT =
        typename detail::get_kernel_name_t<KernelName, KernelType>::name;
 
    verifyUsedKernelBundle(detail::KernelInfo<NameT>::getName());
 
    // Range rounding can be disabled by the user.
    // Range rounding is not done on the host device.
    // Range rounding is supported only for newer SYCL standards.
#if !defined(__SYCL_DISABLE_PARALLEL_FOR_RANGE_ROUNDING__) &&                  \
    !defined(DPCPP_HOST_DEVICE_OPENMP) &&                                      \
    !defined(DPCPP_HOST_DEVICE_PERF_NATIVE) && SYCL_LANGUAGE_VERSION >= 202001
    // Range should be a multiple of this for reasonable performance.
    size_t MinFactorX = 16;
    // Range should be a multiple of this for improved performance.
    size_t GoodFactorX = 32;
    // Range should be at least this to make rounding worthwhile.
    size_t MinRangeX = 1024;
 
    // Check if rounding parameters have been set through environment:
    // SYCL_PARALLEL_FOR_RANGE_ROUNDING_PARAMS=MinRound:PreferredRound:MinRange
    this->GetRangeRoundingSettings(MinFactorX, GoodFactorX, MinRangeX);
 
    // Disable the rounding-up optimizations under these conditions:
    // 1. The env var SYCL_DISABLE_PARALLEL_FOR_RANGE_ROUNDING is set.
    // 2. The kernel is provided via an interoperability method.
    // 3. The range is already a multiple of the rounding factor.
    //
    // Cases 2 and 3 could be supported with extra effort.
    // As an optimization for the common case it is an
    // implementation choice to not support those scenarios.
    // Note that "this_item" is a free function, i.e. not tied to any
    // specific id or item. When concurrent parallel_fors are executing
    // on a device it is difficult to tell which parallel_for the call is
    // being made from. One could replicate portions of the
    // call-graph to make this_item calls kernel-specific but this is
    // not considered worthwhile.
 
    // Get the kernel name to check condition 2.
    std::string KName = typeid(NameT *).name();
    using KI = detail::KernelInfo<KernelName>;
    bool DisableRounding =
        this->DisableRangeRounding() ||
        (KI::getName() == nullptr || KI::getName()[0] == '\0');
 
    // Perform range rounding if rounding-up is enabled
    // and there are sufficient work-items to need rounding
    // and the user-specified range is not a multiple of a "good" value.
    if (!DisableRounding && (NumWorkItems[0] >= MinRangeX) &&
        (NumWorkItems[0] % MinFactorX != 0)) {
      // It is sufficient to round up just the first dimension.
      // Multiplying the rounded-up value of the first dimension
      // by the values of the remaining dimensions (if any)
      // will yield a rounded-up value for the total range.
      size_t NewValX =
          ((NumWorkItems[0] + GoodFactorX - 1) / GoodFactorX) * GoodFactorX;
      if (this->RangeRoundingTrace())
        std::cout << "parallel_for range adjusted from " << NumWorkItems[0]
                  << " to " << NewValX << std::endl;
 
      using NameWT = typename detail::get_kernel_wrapper_name_t<NameT>::name;
      auto Wrapper =
          getRangeRoundedKernelLambda<NameWT, TransformedArgType, Dims>(
              KernelFunc, NumWorkItems);
 
      using KName = std::conditional_t<std::is_same<KernelType, NameT>::value,
                                       decltype(Wrapper), NameWT>;
 
      range<Dims> AdjustedRange = NumWorkItems;
      AdjustedRange.set_range_dim0(NewValX);
      kernel_parallel_for_wrapper<KName, TransformedArgType>(Wrapper);
#ifndef __SYCL_DEVICE_ONLY__
      detail::checkValueRange<Dims>(AdjustedRange);
      MNDRDesc.set(std::move(AdjustedRange));
      StoreLambda<KName, decltype(Wrapper), Dims, TransformedArgType>(
          std::move(Wrapper));
      setType(detail::CG::Kernel);
#endif
    } else
#endif // !__SYCL_DISABLE_PARALLEL_FOR_RANGE_ROUNDING__ &&
       // !DPCPP_HOST_DEVICE_OPENMP && !DPCPP_HOST_DEVICE_PERF_NATIVE &&
       // SYCL_LANGUAGE_VERSION >= 202001
    {
      (void)NumWorkItems;
      kernel_parallel_for_wrapper<NameT, TransformedArgType>(KernelFunc);
#ifndef __SYCL_DEVICE_ONLY__
      detail::checkValueRange<Dims>(NumWorkItems);
      MNDRDesc.set(std::move(NumWorkItems));
      StoreLambda<NameT, KernelType, Dims, TransformedArgType>(
          std::move(KernelFunc));
      setType(detail::CG::Kernel);
#endif
    }
  }
 
  template <int Dims>
  void parallel_for_impl(range<Dims> NumWorkItems, kernel Kernel) {
    throwIfActionIsCreated();
    MKernel = detail::getSyclObjImpl(std::move(Kernel));
    detail::checkValueRange<Dims>(NumWorkItems);
    MNDRDesc.set(std::move(NumWorkItems));
    setType(detail::CG::Kernel);
    extractArgsAndReqs();
    MKernelName = getKernelName();
  }
 
#ifdef SYCL_LANGUAGE_VERSION
#define __SYCL_KERNEL_ATTR__ [[clang::sycl_kernel]]
#else
#define __SYCL_KERNEL_ATTR__
#endif
  // NOTE: the name of this function - "kernel_single_task" - is used by the
  // Front End to determine kernel invocation kind.
  template <typename KernelName, typename KernelType>
  __SYCL_KERNEL_ATTR__ void
#ifdef __SYCL_NONCONST_FUNCTOR__
  kernel_single_task(KernelType KernelFunc) {
#else
  kernel_single_task(const KernelType &KernelFunc) {
#endif
#ifdef __SYCL_DEVICE_ONLY__
    KernelFunc();
#else
    (void)KernelFunc;
#endif
  }
 
  // NOTE: the name of this function - "kernel_single_task" - is used by the
  // Front End to determine kernel invocation kind.
  template <typename KernelName, typename KernelType>
  __SYCL_KERNEL_ATTR__ void
#ifdef __SYCL_NONCONST_FUNCTOR__
  kernel_single_task(KernelType KernelFunc, kernel_handler KH) {
#else
  kernel_single_task(const KernelType &KernelFunc, kernel_handler KH) {
#endif
#ifdef __SYCL_DEVICE_ONLY__
    KernelFunc(KH);
#else
    (void)KernelFunc;
    (void)KH;
#endif
  }
 
  // NOTE: the name of these functions - "kernel_parallel_for" - are used by the
  // Front End to determine kernel invocation kind.
  template <typename KernelName, typename ElementType, typename KernelType>
  __SYCL_KERNEL_ATTR__ void
#ifdef __SYCL_NONCONST_FUNCTOR__
  kernel_parallel_for(KernelType KernelFunc) {
#else
  kernel_parallel_for(const KernelType &KernelFunc) {
#endif
#ifdef __SYCL_DEVICE_ONLY__
    KernelFunc(detail::Builder::getElement(detail::declptr<ElementType>()));
#else
    (void)KernelFunc;
#endif
  }
 
  // NOTE: the name of these functions - "kernel_parallel_for" - are used by the
  // Front End to determine kernel invocation kind.
  template <typename KernelName, typename ElementType, typename KernelType>
  __SYCL_KERNEL_ATTR__ void
#ifdef __SYCL_NONCONST_FUNCTOR__
  kernel_parallel_for(KernelType KernelFunc, kernel_handler KH) {
#else
  kernel_parallel_for(const KernelType &KernelFunc, kernel_handler KH) {
#endif
#ifdef __SYCL_DEVICE_ONLY__
    KernelFunc(detail::Builder::getElement(detail::declptr<ElementType>()), KH);
#else
    (void)KernelFunc;
    (void)KH;
#endif
  }
 
  // NOTE: the name of this function - "kernel_parallel_for_work_group" - is
  // used by the Front End to determine kernel invocation kind.
  template <typename KernelName, typename ElementType, typename KernelType>
  __SYCL_KERNEL_ATTR__ void
#ifdef __SYCL_NONCONST_FUNCTOR__
  kernel_parallel_for_work_group(KernelType KernelFunc) {
#else
  kernel_parallel_for_work_group(const KernelType &KernelFunc) {
#endif
#ifdef __SYCL_DEVICE_ONLY__
    KernelFunc(detail::Builder::getElement(detail::declptr<ElementType>()));
#else
    (void)KernelFunc;
#endif
  }
 
  // NOTE: the name of this function - "kernel_parallel_for_work_group" - is
  // used by the Front End to determine kernel invocation kind.
  template <typename KernelName, typename ElementType, typename KernelType>
  __SYCL_KERNEL_ATTR__ void
#ifdef __SYCL_NONCONST_FUNCTOR__
  kernel_parallel_for_work_group(KernelType KernelFunc, kernel_handler KH) {
#else
  kernel_parallel_for_work_group(const KernelType &KernelFunc,
                                 kernel_handler KH) {
#endif
#ifdef __SYCL_DEVICE_ONLY__
    KernelFunc(detail::Builder::getElement(detail::declptr<ElementType>()), KH);
#else
    (void)KernelFunc;
    (void)KH;
#endif
  }
 
  // Wrappers for kernel_*** functions above with and without support of
  // additional kernel_handler argument.
 
  // NOTE: to support kernel_handler argument in kernel lambdas, only
  // kernel_***_wrapper functions must be called in this code
 
  // Wrappers for kernel_single_task(...)
 
  template <typename KernelName, typename KernelType>
  std::enable_if_t<detail::KernelLambdaHasKernelHandlerArgT<KernelType>::value>
#ifdef __SYCL_NONCONST_FUNCTOR__
  kernel_single_task_wrapper(KernelType KernelFunc) {
#else
  kernel_single_task_wrapper(const KernelType &KernelFunc) {
#endif
#ifdef __SYCL_DEVICE_ONLY__
    detail::CheckDeviceCopyable<KernelType>();
#endif // __SYCL_DEVICE_ONLY__
    kernel_handler KH;
    kernel_single_task<KernelName>(KernelFunc, KH);
  }
 
  template <typename KernelName, typename KernelType>
  std::enable_if_t<!detail::KernelLambdaHasKernelHandlerArgT<KernelType>::value>
#ifdef __SYCL_NONCONST_FUNCTOR__
  kernel_single_task_wrapper(KernelType KernelFunc) {
#else
  kernel_single_task_wrapper(const KernelType &KernelFunc) {
#endif
#ifdef __SYCL_DEVICE_ONLY__
    detail::CheckDeviceCopyable<KernelType>();
#endif // __SYCL_DEVICE_ONLY__
    kernel_single_task<KernelName>(KernelFunc);
  }
 
  // Wrappers for kernel_parallel_for(...)
 
  template <typename KernelName, typename ElementType, typename KernelType>
  std::enable_if_t<
      detail::KernelLambdaHasKernelHandlerArgT<KernelType, ElementType>::value>
#ifdef __SYCL_NONCONST_FUNCTOR__
  kernel_parallel_for_wrapper(KernelType KernelFunc) {
#else
  kernel_parallel_for_wrapper(const KernelType &KernelFunc) {
#endif
#ifdef __SYCL_DEVICE_ONLY__
    detail::CheckDeviceCopyable<KernelType>();
#endif // __SYCL_DEVICE_ONLY__
    kernel_handler KH;
    kernel_parallel_for<KernelName, ElementType>(KernelFunc, KH);
  }
 
  template <typename KernelName, typename ElementType, typename KernelType>
  std::enable_if_t<
      !detail::KernelLambdaHasKernelHandlerArgT<KernelType, ElementType>::value>
#ifdef __SYCL_NONCONST_FUNCTOR__
  kernel_parallel_for_wrapper(KernelType KernelFunc) {
#else
  kernel_parallel_for_wrapper(const KernelType &KernelFunc) {
#endif
#ifdef __SYCL_DEVICE_ONLY__
    detail::CheckDeviceCopyable<KernelType>();
#endif // __SYCL_DEVICE_ONLY__
    kernel_parallel_for<KernelName, ElementType>(KernelFunc);
  }
 
  // Wrappers for kernel_parallel_for_work_group(...)
 
  template <typename KernelName, typename ElementType, typename KernelType>
  std::enable_if_t<
      detail::KernelLambdaHasKernelHandlerArgT<KernelType, ElementType>::value>
#ifdef __SYCL_NONCONST_FUNCTOR__
  kernel_parallel_for_work_group_wrapper(KernelType KernelFunc) {
#else
  kernel_parallel_for_work_group_wrapper(const KernelType &KernelFunc) {
#endif
#ifdef __SYCL_DEVICE_ONLY__
    detail::CheckDeviceCopyable<KernelType>();
#endif // __SYCL_DEVICE_ONLY__
    kernel_handler KH;
    kernel_parallel_for_work_group<KernelName, ElementType>(KernelFunc, KH);
  }
 
  template <typename KernelName, typename ElementType, typename KernelType>
  std::enable_if_t<
      !detail::KernelLambdaHasKernelHandlerArgT<KernelType, ElementType>::value>
#ifdef __SYCL_NONCONST_FUNCTOR__
  kernel_parallel_for_work_group_wrapper(KernelType KernelFunc) {
#else
  kernel_parallel_for_work_group_wrapper(const KernelType &KernelFunc) {
#endif
#ifdef __SYCL_DEVICE_ONLY__
    detail::CheckDeviceCopyable<KernelType>();
#endif // __SYCL_DEVICE_ONLY__
    kernel_parallel_for_work_group<KernelName, ElementType>(KernelFunc);
  }
 
  void setStateExplicitKernelBundle();
  void setStateSpecConstSet();
  bool isStateExplicitKernelBundle() const;
 
  std::shared_ptr<detail::kernel_bundle_impl>
  getOrInsertHandlerKernelBundle(bool Insert) const;
 
  void setHandlerKernelBundle(kernel Kernel);
 
  void setHandlerKernelBundle(
      const std::shared_ptr<detail::kernel_bundle_impl> &NewKernelBundleImpPtr);
 
  template <typename FuncT>
  detail::enable_if_t<
      detail::check_fn_signature<detail::remove_reference_t<FuncT>,
                                 void()>::value ||
      detail::check_fn_signature<detail::remove_reference_t<FuncT>,
                                 void(interop_handle)>::value>
  host_task_impl(FuncT &&Func) {
    throwIfActionIsCreated();
 
    MNDRDesc.set(range<1>(1));
    MArgs = std::move(MAssociatedAccesors);
 
    MHostTask.reset(new detail::HostTask(std::move(Func)));
 
    setType(detail::CG::CodeplayHostTask);
  }
 
public:
  handler(const handler &) = delete;
  handler(handler &&) = delete;
  handler &operator=(const handler &) = delete;
  handler &operator=(handler &&) = delete;
 
#if __cplusplus >= 201703L
  template <auto &SpecName>
  void set_specialization_constant(
      typename std::remove_reference_t<decltype(SpecName)>::value_type Value) {
 
    setStateSpecConstSet();
 
    std::shared_ptr<detail::kernel_bundle_impl> KernelBundleImplPtr =
        getOrInsertHandlerKernelBundle(/*Insert=*/true);
 
    detail::createSyclObjFromImpl<kernel_bundle<bundle_state::input>>(
        KernelBundleImplPtr)
        .set_specialization_constant<SpecName>(Value);
  }
 
  template <auto &SpecName>
  typename std::remove_reference_t<decltype(SpecName)>::value_type
  get_specialization_constant() const {
 
    if (isStateExplicitKernelBundle())
      throw sycl::exception(make_error_code(errc::invalid),
                            "Specialization constants cannot be read after "
                            "explicitly setting the used kernel bundle");
 
    std::shared_ptr<detail::kernel_bundle_impl> KernelBundleImplPtr =
        getOrInsertHandlerKernelBundle(/*Insert=*/true);
 
    return detail::createSyclObjFromImpl<kernel_bundle<bundle_state::input>>(
               KernelBundleImplPtr)
        .get_specialization_constant<SpecName>();
  }
 
#endif
 
  void
  use_kernel_bundle(const kernel_bundle<bundle_state::executable> &ExecBundle);
 
  template <typename DataT, int Dims, access::mode AccMode,
            access::target AccTarget>
  void
  require(accessor<DataT, Dims, AccMode, AccTarget, access::placeholder::true_t>
              Acc) {
    associateWithHandler(&Acc, AccTarget);
  }
 
  void depends_on(event Event);
 
  void depends_on(const std::vector<event> &Events);
 
  template <typename T>
  using remove_cv_ref_t =
      typename detail::remove_cv_t<detail::remove_reference_t<T>>;
 
  template <typename U, typename T>
  using is_same_type = std::is_same<remove_cv_ref_t<U>, remove_cv_ref_t<T>>;
 
  template <typename T> struct ShouldEnableSetArg {
    static constexpr bool value =
        std::is_trivially_copyable<detail::remove_reference_t<T>>::value
#if SYCL_LANGUAGE_VERSION && SYCL_LANGUAGE_VERSION <= 201707
            && std::is_standard_layout<detail::remove_reference_t<T>>::value
#endif
        || is_same_type<sampler, T>::value // Sampler
        || (!is_same_type<cl_mem, T>::value &&
            std::is_pointer<remove_cv_ref_t<T>>::value) // USM
        || is_same_type<cl_mem, T>::value;              // Interop
  };
 
  template <typename T>
  typename detail::enable_if_t<ShouldEnableSetArg<T>::value, void>
  set_arg(int ArgIndex, T &&Arg) {
    setArgHelper(ArgIndex, std::move(Arg));
  }
 
  template <typename DataT, int Dims, access::mode AccessMode,
            access::target AccessTarget, access::placeholder IsPlaceholder>
  void
  set_arg(int ArgIndex,
          accessor<DataT, Dims, AccessMode, AccessTarget, IsPlaceholder> Arg) {
    setArgHelper(ArgIndex, std::move(Arg));
  }
 
  template <typename DataT, int Dims>
  void set_arg(int ArgIndex, local_accessor<DataT, Dims> Arg) {
    setArgHelper(ArgIndex, std::move(Arg));
  }
 
  template <typename... Ts> void set_args(Ts &&...Args) {
    setArgsHelper(0, std::move(Args)...);
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType>
#ifdef __SYCL_NONCONST_FUNCTOR__
  void single_task(KernelType KernelFunc) {
#else
  void single_task(const KernelType &KernelFunc) {
#endif
    throwIfActionIsCreated();
    using NameT =
        typename detail::get_kernel_name_t<KernelName, KernelType>::name;
    verifyUsedKernelBundle(detail::KernelInfo<NameT>::getName());
    kernel_single_task_wrapper<NameT>(KernelFunc);
#ifndef __SYCL_DEVICE_ONLY__
    // No need to check if range is out of INT_MAX limits as it's compile-time
    // known constant.
    MNDRDesc.set(range<1>{1});
 
    StoreLambda<NameT, KernelType, /*Dims*/ 1, void>(KernelFunc);
    setType(detail::CG::Kernel);
#endif
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType>
#ifdef __SYCL_NONCONST_FUNCTOR__
  void parallel_for(range<1> NumWorkItems, KernelType KernelFunc) {
#else
  void parallel_for(range<1> NumWorkItems, const KernelType &KernelFunc) {
#endif
    parallel_for_lambda_impl<KernelName>(NumWorkItems, std::move(KernelFunc));
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType>
#ifdef __SYCL_NONCONST_FUNCTOR__
  void parallel_for(range<2> NumWorkItems, KernelType KernelFunc) {
#else
  void parallel_for(range<2> NumWorkItems, const KernelType &KernelFunc) {
#endif
    parallel_for_lambda_impl<KernelName>(NumWorkItems, std::move(KernelFunc));
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType>
#ifdef __SYCL_NONCONST_FUNCTOR__
  void parallel_for(range<3> NumWorkItems, KernelType KernelFunc) {
#else
  void parallel_for(range<3> NumWorkItems, const KernelType &KernelFunc) {
#endif
    parallel_for_lambda_impl<KernelName>(NumWorkItems, std::move(KernelFunc));
  }
 
  template <typename FuncT>
  __SYCL_DEPRECATED(
      "run_on_host_intel() is deprecated, use host_task() instead")
  void run_on_host_intel(FuncT Func) {
    throwIfActionIsCreated();
    // No need to check if range is out of INT_MAX limits as it's compile-time
    // known constant
    MNDRDesc.set(range<1>{1});
 
    MArgs = std::move(MAssociatedAccesors);
    MHostKernel.reset(new detail::HostKernel<FuncT, void, 1>(std::move(Func)));
    setType(detail::CG::RunOnHostIntel);
  }
 
  template <typename FuncT>
  detail::enable_if_t<
      detail::check_fn_signature<detail::remove_reference_t<FuncT>,
                                 void()>::value ||
      detail::check_fn_signature<detail::remove_reference_t<FuncT>,
                                 void(interop_handle)>::value>
  host_task(FuncT &&Func) {
    host_task_impl(Func);
  }
 
// replace _KERNELFUNCPARAM(KernelFunc) with   KernelType KernelFunc
//                                     or     const KernelType &KernelFunc
#ifdef __SYCL_NONCONST_FUNCTOR__
#define _KERNELFUNCPARAM(a) KernelType a
#else
#define _KERNELFUNCPARAM(a) const KernelType &a
#endif
 
  template <typename KernelName = detail::auto_name, typename KernelType,
            int Dims>
  __SYCL2020_DEPRECATED("offsets are deprecated in SYCL2020")
  void parallel_for(range<Dims> NumWorkItems, id<Dims> WorkItemOffset,
                    _KERNELFUNCPARAM(KernelFunc)) {
    throwIfActionIsCreated();
    using NameT =
        typename detail::get_kernel_name_t<KernelName, KernelType>::name;
    verifyUsedKernelBundle(detail::KernelInfo<NameT>::getName());
    using LambdaArgType = sycl::detail::lambda_arg_type<KernelType, item<Dims>>;
    (void)NumWorkItems;
    (void)WorkItemOffset;
    kernel_parallel_for_wrapper<NameT, LambdaArgType>(KernelFunc);
#ifndef __SYCL_DEVICE_ONLY__
    detail::checkValueRange<Dims>(NumWorkItems, WorkItemOffset);
    MNDRDesc.set(std::move(NumWorkItems), std::move(WorkItemOffset));
    StoreLambda<NameT, KernelType, Dims, LambdaArgType>(std::move(KernelFunc));
    setType(detail::CG::Kernel);
#endif
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType,
            int Dims>
  void parallel_for(nd_range<Dims> ExecutionRange,
                    _KERNELFUNCPARAM(KernelFunc)) {
    throwIfActionIsCreated();
    using NameT =
        typename detail::get_kernel_name_t<KernelName, KernelType>::name;
    verifyUsedKernelBundle(detail::KernelInfo<NameT>::getName());
    using LambdaArgType =
        sycl::detail::lambda_arg_type<KernelType, nd_item<Dims>>;
    // If user type is convertible from sycl::item/sycl::nd_item, use
    // sycl::item/sycl::nd_item to transport item information
    using TransformedArgType =
        typename TransformUserItemType<Dims, LambdaArgType>::type;
    (void)ExecutionRange;
    kernel_parallel_for_wrapper<NameT, TransformedArgType>(KernelFunc);
#ifndef __SYCL_DEVICE_ONLY__
    detail::checkValueRange<Dims>(ExecutionRange);
    MNDRDesc.set(std::move(ExecutionRange));
    StoreLambda<NameT, KernelType, Dims, TransformedArgType>(
        std::move(KernelFunc));
    setType(detail::CG::Kernel);
#endif
  }
 
// "if constexpr" simplifies implementation/increases readability in comparison
// with SFINAE-based approach.
#if __cplusplus >= 201703L
  template <typename KernelName = detail::auto_name, typename KernelType,
            int Dims, typename Reduction>
  void parallel_for(range<Dims> Range, Reduction Redu,
                    _KERNELFUNCPARAM(KernelFunc)) {
    std::shared_ptr<detail::queue_impl> QueueCopy = MQueue;
 
    // Before running the kernels, check that device has enough local memory
    // to hold local arrays required for the tree-reduction algorithm.
    constexpr bool IsTreeReduction =
        !Reduction::has_fast_reduce && !Reduction::has_fast_atomics;
    size_t OneElemSize =
        IsTreeReduction ? sizeof(typename Reduction::result_type) : 0;
    uint32_t NumConcurrentWorkGroups =
#ifdef __SYCL_REDUCTION_NUM_CONCURRENT_WORKGROUPS
        __SYCL_REDUCTION_NUM_CONCURRENT_WORKGROUPS;
#else
        detail::reduGetMaxNumConcurrentWorkGroups(MQueue);
#endif
    // TODO: currently the preferred work group size is determined for the given
    // queue/device, while it is safer to use queries to the kernel pre-compiled
    // for the device.
    size_t PrefWGSize = detail::reduGetPreferredWGSize(MQueue, OneElemSize);
    if (detail::reduCGFuncForRange<KernelName>(*this, KernelFunc, Range,
                                               PrefWGSize,
                                               NumConcurrentWorkGroups, Redu)) {
      this->finalize();
      MLastEvent = withAuxHandler(QueueCopy, [&](handler &CopyHandler) {
        detail::reduSaveFinalResultToUserMem<KernelName>(CopyHandler, Redu);
      });
    }
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType,
            int Dims, typename Reduction>
  void parallel_for(nd_range<Dims> Range, Reduction Redu,
                    _KERNELFUNCPARAM(KernelFunc)) {
    if constexpr (!Reduction::has_fast_atomics &&
                  !Reduction::has_float64_atomics) {
      // The most basic implementation.
      parallel_for_impl<KernelName>(Range, Redu, KernelFunc);
      return;
    } else { // Can't "early" return for "if constexpr".
      std::shared_ptr<detail::queue_impl> QueueCopy = MQueue;
      if constexpr (Reduction::has_float64_atomics) {
        device D = detail::getDeviceFromHandler(*this);
 
        if (D.has(aspect::atomic64)) {
 
          detail::reduCGFuncAtomic64<KernelName>(*this, KernelFunc, Range,
                                                 Redu);
        } else {
          // Resort to basic implementation as well.
          parallel_for_impl<KernelName>(Range, Redu, KernelFunc);
          return;
        }
      } else {
        // Use fast sycl::atomic operations to update reduction variable at the
        // end of each work-group work.
        detail::reduCGFunc<KernelName>(*this, KernelFunc, Range, Redu);
      }
      // If the reduction variable must be initialized with the identity value
      // before the kernel run, then an additional working accessor is created,
      // initialized with the identity value and used in the kernel. That
      // working accessor is then copied to user's accessor or USM pointer after
      // the kernel run.
      // For USM pointers without initialize_to_identity properties the same
      // scheme with working accessor is used as re-using user's USM pointer in
      // the kernel would require creation of another variant of user's kernel,
      // which does not seem efficient.
      if (Reduction::is_usm || Redu.initializeToIdentity()) {
        this->finalize();
        MLastEvent = withAuxHandler(QueueCopy, [&](handler &CopyHandler) {
          detail::reduSaveFinalResultToUserMem<KernelName>(CopyHandler, Redu);
        });
      }
    }
  }
 
  template <typename KernelName, typename KernelType, int Dims,
            typename Reduction>
  void parallel_for_impl(nd_range<Dims> Range, Reduction Redu,
                         KernelType KernelFunc) {
    // This parallel_for() is lowered to the following sequence:
    // 1) Call a kernel that a) call user's lambda function and b) performs
    //    one iteration of reduction, storing the partial reductions/sums
    //    to either a newly created global buffer or to user's reduction
    //    accessor. So, if the original 'Range' has totally
    //    N1 elements and work-group size is W, then after the first iteration
    //    there will be N2 partial sums where N2 = N1 / W.
    //    If (N2 == 1) then the partial sum is written to user's accessor.
    //    Otherwise, a new global buffer is created and partial sums are written
    //    to it.
    // 2) Call an aux kernel (if necessary, i.e. if N2 > 1) as many times as
    //    necessary to reduce all partial sums into one final sum.
 
    // Before running the kernels, check that device has enough local memory
    // to hold local arrays that may be required for the reduction algorithm.
    // TODO: If the work-group-size is limited by the local memory, then
    // a special version of the main kernel may be created. The one that would
    // not use local accessors, which means it would not do the reduction in
    // the main kernel, but simply generate Range.get_global_range.size() number
    // of partial sums, leaving the reduction work to the additional/aux
    // kernels.
    constexpr bool HFR = Reduction::has_fast_reduce;
    size_t OneElemSize = HFR ? 0 : sizeof(typename Reduction::result_type);
    // TODO: currently the maximal work group size is determined for the given
    // queue/device, while it may be safer to use queries to the kernel compiled
    // for the device.
    size_t MaxWGSize = detail::reduGetMaxWGSize(MQueue, OneElemSize);
    if (Range.get_local_range().size() > MaxWGSize)
      throw sycl::runtime_error("The implementation handling parallel_for with"
                                " reduction requires work group size not bigger"
                                " than " +
                                    std::to_string(MaxWGSize),
                                PI_ERROR_INVALID_WORK_GROUP_SIZE);
 
    // 1. Call the kernel that includes user's lambda function.
    detail::reduCGFunc<KernelName>(*this, KernelFunc, Range, Redu);
    std::shared_ptr<detail::queue_impl> QueueCopy = MQueue;
    this->finalize();
 
    // 2. Run the additional kernel as many times as needed to reduce
    // all partial sums into one scalar.
 
    // TODO: Create a special slow/sequential version of the kernel that would
    // handle the reduction instead of reporting an assert below.
    if (MaxWGSize <= 1)
      throw sycl::runtime_error("The implementation handling parallel_for with "
                                "reduction requires the maximal work group "
                                "size to be greater than 1 to converge. "
                                "The maximal work group size depends on the "
                                "device and the size of the objects passed to "
                                "the reduction.",
                                PI_ERROR_INVALID_WORK_GROUP_SIZE);
    size_t NWorkItems = Range.get_group_range().size();
    while (NWorkItems > 1) {
      MLastEvent = withAuxHandler(QueueCopy, [&](handler &AuxHandler) {
        NWorkItems = detail::reduAuxCGFunc<KernelName, KernelType>(
            AuxHandler, NWorkItems, MaxWGSize, Redu);
      });
    } // end while (NWorkItems > 1)
 
    if (Reduction::is_usm) {
      MLastEvent = withAuxHandler(QueueCopy, [&](handler &CopyHandler) {
        detail::reduSaveFinalResultToUserMem<KernelName>(CopyHandler, Redu);
      });
    }
  }
 
  // This version of parallel_for may handle one or more reductions packed in
  // \p Rest argument. The last element in \p Rest pack is the kernel function,
  // everything else is reduction(s).
  // TODO: this variant is currently enabled for 2+ reductions only as the
  // versions handling 1 reduction variable are more efficient right now.
  //
  // This is basically a tree reduction where we re-use user's reduction
  // variable instead of creating temporary storage for the last iteration
  // (#WG == 1).
  template <typename KernelName = detail::auto_name, int Dims,
            typename... RestT>
  std::enable_if_t<(sizeof...(RestT) >= 3 &&
                    detail::AreAllButLastReductions<RestT...>::value)>
  parallel_for(nd_range<Dims> Range, RestT... Rest) {
    std::tuple<RestT...> ArgsTuple(Rest...);
    constexpr size_t NumArgs = sizeof...(RestT);
    auto KernelFunc = std::get<NumArgs - 1>(ArgsTuple);
    auto ReduIndices = std::make_index_sequence<NumArgs - 1>();
    auto ReduTuple = detail::tuple_select_elements(ArgsTuple, ReduIndices);
 
    size_t LocalMemPerWorkItem =
        detail::reduGetMemPerWorkItem(ReduTuple, ReduIndices);
    // TODO: currently the maximal work group size is determined for the given
    // queue/device, while it is safer to use queries to the kernel compiled
    // for the device.
    size_t MaxWGSize = detail::reduGetMaxWGSize(MQueue, LocalMemPerWorkItem);
    if (Range.get_local_range().size() > MaxWGSize)
      throw sycl::runtime_error("The implementation handling parallel_for with"
                                " reduction requires work group size not bigger"
                                " than " +
                                    std::to_string(MaxWGSize),
                                PI_ERROR_INVALID_WORK_GROUP_SIZE);
 
    detail::reduCGFuncMulti<KernelName>(*this, KernelFunc, Range, ReduTuple,
                                        ReduIndices);
    std::shared_ptr<detail::queue_impl> QueueCopy = MQueue;
    this->finalize();
 
    size_t NWorkItems = Range.get_group_range().size();
    while (NWorkItems > 1) {
      MLastEvent = withAuxHandler(QueueCopy, [&](handler &AuxHandler) {
        NWorkItems = detail::reduAuxCGFunc<KernelName, decltype(KernelFunc)>(
            AuxHandler, NWorkItems, MaxWGSize, ReduTuple, ReduIndices);
      });
    } // end while (NWorkItems > 1)
  }
#endif // __cplusplus >= 201703L
 
  template <typename KernelName = detail::auto_name, typename KernelType,
            int Dims>
  void parallel_for_work_group(range<Dims> NumWorkGroups,
                               _KERNELFUNCPARAM(KernelFunc)) {
    throwIfActionIsCreated();
    using NameT =
        typename detail::get_kernel_name_t<KernelName, KernelType>::name;
    verifyUsedKernelBundle(detail::KernelInfo<NameT>::getName());
    using LambdaArgType =
        sycl::detail::lambda_arg_type<KernelType, group<Dims>>;
    (void)NumWorkGroups;
    kernel_parallel_for_work_group_wrapper<NameT, LambdaArgType>(KernelFunc);
#ifndef __SYCL_DEVICE_ONLY__
    detail::checkValueRange<Dims>(NumWorkGroups);
    MNDRDesc.setNumWorkGroups(NumWorkGroups);
    StoreLambda<NameT, KernelType, Dims, LambdaArgType>(std::move(KernelFunc));
    setType(detail::CG::Kernel);
#endif // __SYCL_DEVICE_ONLY__
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType,
            int Dims>
  void parallel_for_work_group(range<Dims> NumWorkGroups,
                               range<Dims> WorkGroupSize,
                               _KERNELFUNCPARAM(KernelFunc)) {
    throwIfActionIsCreated();
    using NameT =
        typename detail::get_kernel_name_t<KernelName, KernelType>::name;
    verifyUsedKernelBundle(detail::KernelInfo<NameT>::getName());
    using LambdaArgType =
        sycl::detail::lambda_arg_type<KernelType, group<Dims>>;
    (void)NumWorkGroups;
    (void)WorkGroupSize;
    kernel_parallel_for_work_group_wrapper<NameT, LambdaArgType>(KernelFunc);
#ifndef __SYCL_DEVICE_ONLY__
    nd_range<Dims> ExecRange =
        nd_range<Dims>(NumWorkGroups * WorkGroupSize, WorkGroupSize);
    detail::checkValueRange<Dims>(ExecRange);
    MNDRDesc.set(std::move(ExecRange));
    StoreLambda<NameT, KernelType, Dims, LambdaArgType>(std::move(KernelFunc));
    setType(detail::CG::Kernel);
#endif // __SYCL_DEVICE_ONLY__
  }
 
  void single_task(kernel Kernel) {
    throwIfActionIsCreated();
    // Ignore any set kernel bundles and use the one associated with the kernel
    setHandlerKernelBundle(Kernel);
    // No need to check if range is out of INT_MAX limits as it's compile-time
    // known constant
    MNDRDesc.set(range<1>{1});
    MKernel = detail::getSyclObjImpl(std::move(Kernel));
    setType(detail::CG::Kernel);
    extractArgsAndReqs();
    MKernelName = getKernelName();
  }
 
  void parallel_for(range<1> NumWorkItems, kernel Kernel) {
    parallel_for_impl(NumWorkItems, Kernel);
  }
 
  void parallel_for(range<2> NumWorkItems, kernel Kernel) {
    parallel_for_impl(NumWorkItems, Kernel);
  }
 
  void parallel_for(range<3> NumWorkItems, kernel Kernel) {
    parallel_for_impl(NumWorkItems, Kernel);
  }
 
  template <int Dims>
  __SYCL2020_DEPRECATED("offsets are deprecated in SYCL 2020")
  void parallel_for(range<Dims> NumWorkItems, id<Dims> WorkItemOffset,
                    kernel Kernel) {
    throwIfActionIsCreated();
    MKernel = detail::getSyclObjImpl(std::move(Kernel));
    detail::checkValueRange<Dims>(NumWorkItems, WorkItemOffset);
    MNDRDesc.set(std::move(NumWorkItems), std::move(WorkItemOffset));
    setType(detail::CG::Kernel);
    extractArgsAndReqs();
    MKernelName = getKernelName();
  }
 
  template <int Dims> void parallel_for(nd_range<Dims> NDRange, kernel Kernel) {
    throwIfActionIsCreated();
    MKernel = detail::getSyclObjImpl(std::move(Kernel));
    detail::checkValueRange<Dims>(NDRange);
    MNDRDesc.set(std::move(NDRange));
    setType(detail::CG::Kernel);
    extractArgsAndReqs();
    MKernelName = getKernelName();
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType>
  void single_task(kernel Kernel, _KERNELFUNCPARAM(KernelFunc)) {
    throwIfActionIsCreated();
    // Ignore any set kernel bundles and use the one associated with the kernel
    setHandlerKernelBundle(Kernel);
    using NameT =
        typename detail::get_kernel_name_t<KernelName, KernelType>::name;
    verifyUsedKernelBundle(detail::KernelInfo<NameT>::getName());
    (void)Kernel;
    kernel_single_task<NameT>(KernelFunc);
#ifndef __SYCL_DEVICE_ONLY__
    // No need to check if range is out of INT_MAX limits as it's compile-time
    // known constant
    MNDRDesc.set(range<1>{1});
    MKernel = detail::getSyclObjImpl(std::move(Kernel));
    setType(detail::CG::Kernel);
    if (!MIsHost && !lambdaAndKernelHaveEqualName<NameT>()) {
      extractArgsAndReqs();
      MKernelName = getKernelName();
    } else
      StoreLambda<NameT, KernelType, /*Dims*/ 1, void>(std::move(KernelFunc));
#else
    detail::CheckDeviceCopyable<KernelType>();
#endif
  }
 
  template <typename FuncT>
  __SYCL_DEPRECATED("interop_task() is deprecated, use host_task() instead")
  void interop_task(FuncT Func) {
 
    MInteropTask.reset(new detail::InteropTask(std::move(Func)));
    setType(detail::CG::CodeplayInteropTask);
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType,
            int Dims>
  void parallel_for(kernel Kernel, range<Dims> NumWorkItems,
                    _KERNELFUNCPARAM(KernelFunc)) {
    throwIfActionIsCreated();
    // Ignore any set kernel bundles and use the one associated with the kernel
    setHandlerKernelBundle(Kernel);
    using NameT =
        typename detail::get_kernel_name_t<KernelName, KernelType>::name;
    verifyUsedKernelBundle(detail::KernelInfo<NameT>::getName());
    using LambdaArgType = sycl::detail::lambda_arg_type<KernelType, item<Dims>>;
    (void)Kernel;
    (void)NumWorkItems;
    kernel_parallel_for_wrapper<NameT, LambdaArgType>(KernelFunc);
#ifndef __SYCL_DEVICE_ONLY__
    detail::checkValueRange<Dims>(NumWorkItems);
    MNDRDesc.set(std::move(NumWorkItems));
    MKernel = detail::getSyclObjImpl(std::move(Kernel));
    setType(detail::CG::Kernel);
    if (!MIsHost && !lambdaAndKernelHaveEqualName<NameT>()) {
      extractArgsAndReqs();
      MKernelName = getKernelName();
    } else
      StoreLambda<NameT, KernelType, Dims, LambdaArgType>(
          std::move(KernelFunc));
#endif
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType,
            int Dims>
  __SYCL2020_DEPRECATED("offsets are deprecated in SYCL 2020")
  void parallel_for(kernel Kernel, range<Dims> NumWorkItems,
                    id<Dims> WorkItemOffset, _KERNELFUNCPARAM(KernelFunc)) {
    throwIfActionIsCreated();
    // Ignore any set kernel bundles and use the one associated with the kernel
    setHandlerKernelBundle(Kernel);
    using NameT =
        typename detail::get_kernel_name_t<KernelName, KernelType>::name;
    verifyUsedKernelBundle(detail::KernelInfo<NameT>::getName());
    using LambdaArgType = sycl::detail::lambda_arg_type<KernelType, item<Dims>>;
    (void)Kernel;
    (void)NumWorkItems;
    (void)WorkItemOffset;
    kernel_parallel_for_wrapper<NameT, LambdaArgType>(KernelFunc);
#ifndef __SYCL_DEVICE_ONLY__
    detail::checkValueRange<Dims>(NumWorkItems, WorkItemOffset);
    MNDRDesc.set(std::move(NumWorkItems), std::move(WorkItemOffset));
    MKernel = detail::getSyclObjImpl(std::move(Kernel));
    setType(detail::CG::Kernel);
    if (!MIsHost && !lambdaAndKernelHaveEqualName<NameT>()) {
      extractArgsAndReqs();
      MKernelName = getKernelName();
    } else
      StoreLambda<NameT, KernelType, Dims, LambdaArgType>(
          std::move(KernelFunc));
#endif
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType,
            int Dims>
  void parallel_for(kernel Kernel, nd_range<Dims> NDRange,
                    _KERNELFUNCPARAM(KernelFunc)) {
    throwIfActionIsCreated();
    // Ignore any set kernel bundles and use the one associated with the kernel
    setHandlerKernelBundle(Kernel);
    using NameT =
        typename detail::get_kernel_name_t<KernelName, KernelType>::name;
    verifyUsedKernelBundle(detail::KernelInfo<NameT>::getName());
    using LambdaArgType =
        sycl::detail::lambda_arg_type<KernelType, nd_item<Dims>>;
    (void)Kernel;
    (void)NDRange;
    kernel_parallel_for_wrapper<NameT, LambdaArgType>(KernelFunc);
#ifndef __SYCL_DEVICE_ONLY__
    detail::checkValueRange<Dims>(NDRange);
    MNDRDesc.set(std::move(NDRange));
    MKernel = detail::getSyclObjImpl(std::move(Kernel));
    setType(detail::CG::Kernel);
    if (!MIsHost && !lambdaAndKernelHaveEqualName<NameT>()) {
      extractArgsAndReqs();
      MKernelName = getKernelName();
    } else
      StoreLambda<NameT, KernelType, Dims, LambdaArgType>(
          std::move(KernelFunc));
#endif
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType,
            int Dims>
  void parallel_for_work_group(kernel Kernel, range<Dims> NumWorkGroups,
                               _KERNELFUNCPARAM(KernelFunc)) {
    throwIfActionIsCreated();
    // Ignore any set kernel bundles and use the one associated with the kernel
    setHandlerKernelBundle(Kernel);
    using NameT =
        typename detail::get_kernel_name_t<KernelName, KernelType>::name;
    verifyUsedKernelBundle(detail::KernelInfo<NameT>::getName());
    using LambdaArgType =
        sycl::detail::lambda_arg_type<KernelType, group<Dims>>;
    (void)Kernel;
    (void)NumWorkGroups;
    kernel_parallel_for_work_group_wrapper<NameT, LambdaArgType>(KernelFunc);
#ifndef __SYCL_DEVICE_ONLY__
    detail::checkValueRange<Dims>(NumWorkGroups);
    MNDRDesc.setNumWorkGroups(NumWorkGroups);
    MKernel = detail::getSyclObjImpl(std::move(Kernel));
    StoreLambda<NameT, KernelType, Dims, LambdaArgType>(std::move(KernelFunc));
    setType(detail::CG::Kernel);
#endif // __SYCL_DEVICE_ONLY__
  }
 
  template <typename KernelName = detail::auto_name, typename KernelType,
            int Dims>
  void parallel_for_work_group(kernel Kernel, range<Dims> NumWorkGroups,
                               range<Dims> WorkGroupSize,
                               _KERNELFUNCPARAM(KernelFunc)) {
    throwIfActionIsCreated();
    // Ignore any set kernel bundles and use the one associated with the kernel
    setHandlerKernelBundle(Kernel);
    using NameT =
        typename detail::get_kernel_name_t<KernelName, KernelType>::name;
    verifyUsedKernelBundle(detail::KernelInfo<NameT>::getName());
    using LambdaArgType =
        sycl::detail::lambda_arg_type<KernelType, group<Dims>>;
    (void)Kernel;
    (void)NumWorkGroups;
    (void)WorkGroupSize;
    kernel_parallel_for_work_group_wrapper<NameT, LambdaArgType>(KernelFunc);
#ifndef __SYCL_DEVICE_ONLY__
    nd_range<Dims> ExecRange =
        nd_range<Dims>(NumWorkGroups * WorkGroupSize, WorkGroupSize);
    detail::checkValueRange<Dims>(ExecRange);
    MNDRDesc.set(std::move(ExecRange));
    MKernel = detail::getSyclObjImpl(std::move(Kernel));
    StoreLambda<NameT, KernelType, Dims, LambdaArgType>(std::move(KernelFunc));
    setType(detail::CG::Kernel);
#endif // __SYCL_DEVICE_ONLY__
  }
 
  // Clean up KERNELFUNC macro.
#undef _KERNELFUNCPARAM
 
  // Explicit copy operations API
 
  template <typename T_Src, typename T_Dst, int Dims, access::mode AccessMode,
            access::target AccessTarget,
            access::placeholder IsPlaceholder = access::placeholder::false_t>
  void copy(accessor<T_Src, Dims, AccessMode, AccessTarget, IsPlaceholder> Src,
            std::shared_ptr<T_Dst> Dst) {
    throwIfActionIsCreated();
    static_assert(isValidTargetForExplicitOp(AccessTarget),
                  "Invalid accessor target for the copy method.");
    static_assert(isValidModeForSourceAccessor(AccessMode),
                  "Invalid accessor mode for the copy method.");
    // Make sure data shared_ptr points to is not released until we finish
    // work with it.
    MSharedPtrStorage.push_back(Dst);
    typename std::shared_ptr<T_Dst>::element_type *RawDstPtr = Dst.get();
    copy(Src, RawDstPtr);
  }
 
  template <typename T_Src, typename T_Dst, int Dims, access::mode AccessMode,
            access::target AccessTarget,
            access::placeholder IsPlaceholder = access::placeholder::false_t>
  void
  copy(std::shared_ptr<T_Src> Src,
       accessor<T_Dst, Dims, AccessMode, AccessTarget, IsPlaceholder> Dst) {
    throwIfActionIsCreated();
    static_assert(isValidTargetForExplicitOp(AccessTarget),
                  "Invalid accessor target for the copy method.");
    static_assert(isValidModeForDestinationAccessor(AccessMode),
                  "Invalid accessor mode for the copy method.");
    // Make sure data shared_ptr points to is not released until we finish
    // work with it.
    MSharedPtrStorage.push_back(Src);
    typename std::shared_ptr<T_Src>::element_type *RawSrcPtr = Src.get();
    copy(RawSrcPtr, Dst);
  }
 
  template <typename T_Src, typename T_Dst, int Dims, access::mode AccessMode,
            access::target AccessTarget,
            access::placeholder IsPlaceholder = access::placeholder::false_t>
  void copy(accessor<T_Src, Dims, AccessMode, AccessTarget, IsPlaceholder> Src,
            T_Dst *Dst) {
    throwIfActionIsCreated();
    static_assert(isValidTargetForExplicitOp(AccessTarget),
                  "Invalid accessor target for the copy method.");
    static_assert(isValidModeForSourceAccessor(AccessMode),
                  "Invalid accessor mode for the copy method.");
#ifndef __SYCL_DEVICE_ONLY__
    if (MIsHost) {
      // TODO: Temporary implementation for host. Should be handled by memory
      // manager.
      copyAccToPtrHost(Src, Dst);
      return;
    }
#endif
    setType(detail::CG::CopyAccToPtr);
 
    detail::AccessorBaseHost *AccBase = (detail::AccessorBaseHost *)&Src;
    detail::AccessorImplPtr AccImpl = detail::getSyclObjImpl(*AccBase);
 
    MRequirements.push_back(AccImpl.get());
    MSrcPtr = static_cast<void *>(AccImpl.get());
    MDstPtr = static_cast<void *>(Dst);
    // Store copy of accessor to the local storage to make sure it is alive
    // until we finish
    MAccStorage.push_back(std::move(AccImpl));
  }
 
  template <typename T_Src, typename T_Dst, int Dims, access::mode AccessMode,
            access::target AccessTarget,
            access::placeholder IsPlaceholder = access::placeholder::false_t>
  void
  copy(const T_Src *Src,
       accessor<T_Dst, Dims, AccessMode, AccessTarget, IsPlaceholder> Dst) {
    throwIfActionIsCreated();
    static_assert(isValidTargetForExplicitOp(AccessTarget),
                  "Invalid accessor target for the copy method.");
    static_assert(isValidModeForDestinationAccessor(AccessMode),
                  "Invalid accessor mode for the copy method.");
#ifndef __SYCL_DEVICE_ONLY__
    if (MIsHost) {
      // TODO: Temporary implementation for host. Should be handled by memory
      // manager.
      copyPtrToAccHost(Src, Dst);
      return;
    }
#endif
    setType(detail::CG::CopyPtrToAcc);
 
    detail::AccessorBaseHost *AccBase = (detail::AccessorBaseHost *)&Dst;
    detail::AccessorImplPtr AccImpl = detail::getSyclObjImpl(*AccBase);
 
    MRequirements.push_back(AccImpl.get());
    MSrcPtr = const_cast<T_Src *>(Src);
    MDstPtr = static_cast<void *>(AccImpl.get());
    // Store copy of accessor to the local storage to make sure it is alive
    // until we finish
    MAccStorage.push_back(std::move(AccImpl));
  }
 
  template <
      typename T_Src, int Dims_Src, access::mode AccessMode_Src,
      access::target AccessTarget_Src, typename T_Dst, int Dims_Dst,
      access::mode AccessMode_Dst, access::target AccessTarget_Dst,
      access::placeholder IsPlaceholder_Src = access::placeholder::false_t,
      access::placeholder IsPlaceholder_Dst = access::placeholder::false_t>
  void copy(accessor<T_Src, Dims_Src, AccessMode_Src, AccessTarget_Src,
                     IsPlaceholder_Src>
                Src,
            accessor<T_Dst, Dims_Dst, AccessMode_Dst, AccessTarget_Dst,
                     IsPlaceholder_Dst>
                Dst) {
    throwIfActionIsCreated();
    static_assert(isValidTargetForExplicitOp(AccessTarget_Src),
                  "Invalid source accessor target for the copy method.");
    static_assert(isValidTargetForExplicitOp(AccessTarget_Dst),
                  "Invalid destination accessor target for the copy method.");
    static_assert(isValidModeForSourceAccessor(AccessMode_Src),
                  "Invalid source accessor mode for the copy method.");
    static_assert(isValidModeForDestinationAccessor(AccessMode_Dst),
                  "Invalid destination accessor mode for the copy method.");
    if (Dst.get_size() < Src.get_size())
      throw sycl::invalid_object_error(
          "The destination accessor size is too small to copy the memory into.",
          PI_ERROR_INVALID_OPERATION);
 
    if (copyAccToAccHelper(Src, Dst))
      return;
    setType(detail::CG::CopyAccToAcc);
 
    detail::AccessorBaseHost *AccBaseSrc = (detail::AccessorBaseHost *)&Src;
    detail::AccessorImplPtr AccImplSrc = detail::getSyclObjImpl(*AccBaseSrc);
 
    detail::AccessorBaseHost *AccBaseDst = (detail::AccessorBaseHost *)&Dst;
    detail::AccessorImplPtr AccImplDst = detail::getSyclObjImpl(*AccBaseDst);
 
    MRequirements.push_back(AccImplSrc.get());
    MRequirements.push_back(AccImplDst.get());
    MSrcPtr = AccImplSrc.get();
    MDstPtr = AccImplDst.get();
    // Store copy of accessor to the local storage to make sure it is alive
    // until we finish
    MAccStorage.push_back(std::move(AccImplSrc));
    MAccStorage.push_back(std::move(AccImplDst));
  }
 
  template <typename T, int Dims, access::mode AccessMode,
            access::target AccessTarget,
            access::placeholder IsPlaceholder = access::placeholder::false_t>
  void
  update_host(accessor<T, Dims, AccessMode, AccessTarget, IsPlaceholder> Acc) {
    throwIfActionIsCreated();
    static_assert(isValidTargetForExplicitOp(AccessTarget),
                  "Invalid accessor target for the update_host method.");
    setType(detail::CG::UpdateHost);
 
    detail::AccessorBaseHost *AccBase = (detail::AccessorBaseHost *)&Acc;
    detail::AccessorImplPtr AccImpl = detail::getSyclObjImpl(*AccBase);
 
    MDstPtr = static_cast<void *>(AccImpl.get());
    MRequirements.push_back(AccImpl.get());
    MAccStorage.push_back(std::move(AccImpl));
  }
 
  template <typename T, int Dims, access::mode AccessMode,
            access::target AccessTarget,
            access::placeholder IsPlaceholder = access::placeholder::false_t,
            typename PropertyListT = property_list>
  void
  fill(accessor<T, Dims, AccessMode, AccessTarget, IsPlaceholder, PropertyListT>
           Dst,
       const T &Pattern) {
    throwIfActionIsCreated();
    // TODO add check:T must be an integral scalar value or a SYCL vector type
    static_assert(isValidTargetForExplicitOp(AccessTarget),
                  "Invalid accessor target for the fill method.");
    if (!MIsHost && (((Dims == 1) && isConstOrGlobal(AccessTarget)) ||
                     isImageOrImageArray(AccessTarget))) {
      setType(detail::CG::Fill);
 
      detail::AccessorBaseHost *AccBase = (detail::AccessorBaseHost *)&Dst;
      detail::AccessorImplPtr AccImpl = detail::getSyclObjImpl(*AccBase);
 
      MDstPtr = static_cast<void *>(AccImpl.get());
      MRequirements.push_back(AccImpl.get());
      MAccStorage.push_back(std::move(AccImpl));
 
      MPattern.resize(sizeof(T));
      auto PatternPtr = reinterpret_cast<T *>(MPattern.data());
      *PatternPtr = Pattern;
    } else {
 
      // TODO: Temporary implementation for host. Should be handled by memory
      // manger.
      range<Dims> Range = Dst.get_range();
      parallel_for<
          class __fill<T, Dims, AccessMode, AccessTarget, IsPlaceholder>>(
          Range, [=](id<Dims> Index) { Dst[Index] = Pattern; });
    }
  }
 
  template <typename T> void fill(void *Ptr, const T &Pattern, size_t Count) {
    throwIfActionIsCreated();
    static_assert(std::is_trivially_copyable<T>::value,
                  "Pattern must be trivially copyable");
    parallel_for<class __usmfill<T>>(range<1>(Count), [=](id<1> Index) {
      T *CastedPtr = static_cast<T *>(Ptr);
      CastedPtr[Index] = Pattern;
    });
  }
 
  void ext_oneapi_barrier() {
    throwIfActionIsCreated();
    setType(detail::CG::Barrier);
  }
 
  __SYCL2020_DEPRECATED("use 'ext_oneapi_barrier' instead")
  void barrier() { ext_oneapi_barrier(); }
 
  void ext_oneapi_barrier(const std::vector<event> &WaitList);
 
  __SYCL2020_DEPRECATED("use 'ext_oneapi_barrier' instead")
  void barrier(const std::vector<event> &WaitList);
 
  void memcpy(void *Dest, const void *Src, size_t Count);
 
  template <typename T> void copy(const T *Src, T *Dest, size_t Count) {
    this->memcpy(Dest, Src, Count * sizeof(T));
  }
 
  void memset(void *Dest, int Value, size_t Count);
 
  void prefetch(const void *Ptr, size_t Count);
 
  void mem_advise(const void *Ptr, size_t Length, int Advice);
 
private:
  std::shared_ptr<detail::handler_impl> MImpl;
  std::shared_ptr<detail::queue_impl> MQueue;
  std::vector<std::vector<char>> MArgsStorage;
  std::vector<detail::AccessorImplPtr> MAccStorage;
  std::vector<detail::LocalAccessorImplPtr> MLocalAccStorage;
  std::vector<std::shared_ptr<detail::stream_impl>> MStreamStorage;
  mutable std::vector<std::shared_ptr<const void>> MSharedPtrStorage;
  std::vector<detail::ArgDesc> MArgs;
  std::vector<detail::ArgDesc> MAssociatedAccesors;
  std::vector<detail::AccessorImplHost *> MRequirements;
  detail::NDRDescT MNDRDesc;
  std::string MKernelName;
  std::shared_ptr<detail::kernel_impl> MKernel;
  detail::CG::CGTYPE MCGType = detail::CG::None;
  void *MSrcPtr = nullptr;
  void *MDstPtr = nullptr;
  size_t MLength = 0;
  std::vector<char> MPattern;
  std::unique_ptr<detail::HostKernelBase> MHostKernel;
  std::unique_ptr<detail::HostTask> MHostTask;
  detail::OSModuleHandle MOSModuleHandle = detail::OSUtil::ExeModuleHandle;
  // Storage for a lambda or function when using InteropTasks
  std::unique_ptr<detail::InteropTask> MInteropTask;
  std::vector<detail::EventImplPtr> MEvents;
  std::vector<detail::EventImplPtr> MEventsWaitWithBarrier;
 
  bool MIsHost = false;
 
  detail::code_location MCodeLoc = {};
  bool MIsFinalized = false;
  event MLastEvent;
 
  // Make queue_impl class friend to be able to call finalize method.
  friend class detail::queue_impl;
  // Make accessor class friend to keep the list of associated accessors.
  template <typename DataT, int Dims, access::mode AccMode,
            access::target AccTarget, access::placeholder isPlaceholder,
            typename PropertyListT>
  friend class accessor;
  friend device detail::getDeviceFromHandler(handler &);
 
  template <typename DataT, int Dimensions, access::mode AccessMode,
            access::target AccessTarget, access::placeholder IsPlaceholder>
  friend class detail::image_accessor;
  // Make stream class friend to be able to keep the list of associated streams
  friend class stream;
  friend class detail::stream_impl;
  // Make reduction friends to store buffers and arrays created for it
  // in handler from reduction methods.
  template <typename T, class BinaryOperation, int Dims, size_t Extent,
            typename RedOutVar>
  friend class detail::reduction_impl_algo;
 
#ifndef __SYCL_DEVICE_ONLY__
  friend void detail::associateWithHandler(handler &,
                                           detail::AccessorBaseHost *,
                                           access::target);
#endif
 
  friend class ::MockHandler;
  friend class detail::queue_impl;
 
  bool DisableRangeRounding();
 
  bool RangeRoundingTrace();
 
  void GetRangeRoundingSettings(size_t &MinFactor, size_t &GoodFactor,
                                size_t &MinRange);
 
  template <typename WrapperT, typename TransformedArgType, int Dims,
            typename KernelType,
            std::enable_if_t<detail::KernelLambdaHasKernelHandlerArgT<
                KernelType, TransformedArgType>::value> * = nullptr>
  auto getRangeRoundedKernelLambda(KernelType KernelFunc,
                                   range<Dims> NumWorkItems) {
    return detail::RoundedRangeKernelWithKH<TransformedArgType, Dims,
                                            KernelType>(NumWorkItems,
                                                        KernelFunc);
  }
 
  template <typename WrapperT, typename TransformedArgType, int Dims,
            typename KernelType,
            std::enable_if_t<!detail::KernelLambdaHasKernelHandlerArgT<
                KernelType, TransformedArgType>::value> * = nullptr>
  auto getRangeRoundedKernelLambda(KernelType KernelFunc,
                                   range<Dims> NumWorkItems) {
    return detail::RoundedRangeKernel<TransformedArgType, Dims, KernelType>(
        NumWorkItems, KernelFunc);
  }
};
} // __SYCL_INLINE_VER_NAMESPACE(_V1)
} // namespace sycl