queue_impl.hpp

Go to the documentation of this file.

//==------------------ queue_impl.hpp - SYCL queue -------------------------==//
//
// 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 <detail/config.hpp>
#include <detail/context_impl.hpp>
#include <detail/device_impl.hpp>
#include <detail/event_impl.hpp>
#include <detail/global_handler.hpp>
#include <detail/kernel_impl.hpp>
#include <detail/plugin.hpp>
#include <detail/scheduler/scheduler.hpp>
#include <detail/thread_pool.hpp>
#include <sycl/context.hpp>
#include <sycl/detail/assert_happened.hpp>
#include <sycl/detail/cuda_definitions.hpp>
#include <sycl/device.hpp>
#include <sycl/event.hpp>
#include <sycl/exception.hpp>
#include <sycl/exception_list.hpp>
#include <sycl/handler.hpp>
#include <sycl/properties/context_properties.hpp>
#include <sycl/properties/queue_properties.hpp>
#include <sycl/property_list.hpp>
#include <sycl/stl.hpp>
 
#include <utility>
 
#ifdef XPTI_ENABLE_INSTRUMENTATION
#include "xpti/xpti_trace_framework.hpp"
#include <detail/xpti_registry.hpp>
#endif
 
namespace sycl {
__SYCL_INLINE_VER_NAMESPACE(_V1) {
namespace detail {
 
using ContextImplPtr = std::shared_ptr<detail::context_impl>;
using DeviceImplPtr = std::shared_ptr<detail::device_impl>;
 
static constexpr size_t MaxNumQueues = 256;
 
enum class CUDAContextT : char { primary, custom };
 
constexpr CUDAContextT DefaultContextType = CUDAContextT::custom;
 
enum QueueOrder { Ordered, OOO };
 
class queue_impl {
public:
  // \return a default context for the platform if it includes the device
  // passed and default contexts are enabled, a new context otherwise.
  static ContextImplPtr getDefaultOrNew(const DeviceImplPtr &Device) {
    if (!SYCLConfig<SYCL_ENABLE_DEFAULT_CONTEXTS>::get())
      return detail::getSyclObjImpl(
          context{createSyclObjFromImpl<device>(Device), {}, {}});
 
    ContextImplPtr DefaultContext = detail::getSyclObjImpl(
        Device->get_platform().ext_oneapi_get_default_context());
    if (DefaultContext->isDeviceValid(Device))
      return DefaultContext;
    return detail::getSyclObjImpl(
        context{createSyclObjFromImpl<device>(Device), {}, {}});
  }
  queue_impl(const DeviceImplPtr &Device, const async_handler &AsyncHandler,
             const property_list &PropList)
      : queue_impl(Device, getDefaultOrNew(Device), AsyncHandler, PropList){};
 
  queue_impl(const DeviceImplPtr &Device, const ContextImplPtr &Context,
             const async_handler &AsyncHandler, const property_list &PropList)
      : MDevice(Device), MContext(Context), MAsyncHandler(AsyncHandler),
        MPropList(PropList), MHostQueue(MDevice->is_host()),
        MAssertHappenedBuffer(range<1>{1}),
        MIsInorder(has_property<property::queue::in_order>()),
        MDiscardEvents(
            has_property<ext::oneapi::property::queue::discard_events>()),
        MIsProfilingEnabled(has_property<property::queue::enable_profiling>()),
        MHasDiscardEventsSupport(MDiscardEvents &&
                                 (MHostQueue ? true : MIsInorder)) {
    // We enable XPTI tracing events using the TLS mechanism; if the code
    // location data is available, then the tracing data will be rich.
#if XPTI_ENABLE_INSTRUMENTATION
    XPTIScope PrepareNotify((void *)this,
                            (uint16_t)xpti::trace_point_type_t::queue_create,
                            SYCL_STREAM_NAME, "queue_create");
    // Cache the trace event, stream id and instance IDs for the destructor
    if (xptiTraceEnabled()) {
      MTraceEvent = (void *)PrepareNotify.traceEvent();
      MStreamID = PrepareNotify.streamID();
      MInstanceID = PrepareNotify.instanceID();
    }
    // Add the function to capture meta data for the XPTI trace event
    PrepareNotify.addMetadata([&](auto TEvent) {
      xpti::addMetadata(TEvent, "sycl_context",
                        reinterpret_cast<size_t>(MContext->getHandleRef()));
      if (MDevice) {
        xpti::addMetadata(TEvent, "sycl_device_name", MDevice->getDeviceName());
        xpti::addMetadata(
            TEvent, "sycl_device",
            reinterpret_cast<size_t>(
                MDevice->is_host() ? 0 : MDevice->getHandleRef()));
      }
      xpti::addMetadata(TEvent, "is_inorder", MIsInorder);
    });
    PrepareNotify.notify();
#endif
    if (has_property<property::queue::enable_profiling>()) {
      if (has_property<ext::oneapi::property::queue::discard_events>())
        throw sycl::exception(make_error_code(errc::invalid),
                              "Queue cannot be constructed with both of "
                              "discard_events and enable_profiling.");
      if (!MDevice->has(aspect::queue_profiling))
        throw sycl::exception(make_error_code(errc::feature_not_supported),
                              "Cannot enable profiling, the associated device "
                              "does not have the queue_profiling aspect");
    }
    if (has_property<ext::intel::property::queue::compute_index>()) {
      int Idx = get_property<ext::intel::property::queue::compute_index>()
                    .get_index();
      int NumIndices =
          createSyclObjFromImpl<device>(Device)
              .get_info<ext::intel::info::device::max_compute_queue_indices>();
      if (Idx < 0 || Idx >= NumIndices)
        throw sycl::exception(
            make_error_code(errc::invalid),
            "Queue compute index must be a non-negative number less than "
            "device's number of available compute queue indices.");
    }
    if (!Context->isDeviceValid(Device)) {
      if (!Context->is_host() &&
          Context->getPlugin().getBackend() == backend::opencl)
        throw sycl::invalid_object_error(
            "Queue cannot be constructed with the given context and device "
            "since the device is not a member of the context (descendants of "
            "devices from the context are not supported on OpenCL yet).",
            PI_ERROR_INVALID_DEVICE);
      throw sycl::invalid_object_error(
          "Queue cannot be constructed with the given context and device "
          "since the device is neither a member of the context nor a "
          "descendant of its member.",
          PI_ERROR_INVALID_DEVICE);
    }
    if (!MHostQueue) {
      const QueueOrder QOrder =
          MIsInorder ? QueueOrder::Ordered : QueueOrder::OOO;
      MQueues.push_back(createQueue(QOrder));
      // This section is the second part of the instrumentation that uses the
      // tracepoint information and notifies
    }
  }
 
  queue_impl(RT::PiQueue PiQueue, const ContextImplPtr &Context,
             const async_handler &AsyncHandler)
      : MContext(Context), MAsyncHandler(AsyncHandler), MPropList(),
        MHostQueue(false), MAssertHappenedBuffer(range<1>{1}),
        MIsInorder(has_property<property::queue::in_order>()),
        MDiscardEvents(
            has_property<ext::oneapi::property::queue::discard_events>()),
        MIsProfilingEnabled(has_property<property::queue::enable_profiling>()),
        MHasDiscardEventsSupport(MDiscardEvents &&
                                 (MHostQueue ? true : MIsInorder)) {
    // The following commented section provides a guideline on how to use the
    // TLS enabled mechanism to create a tracepoint and notify using XPTI. This
    // is the prolog section and the epilog section will initiate the
    // notification.
#if XPTI_ENABLE_INSTRUMENTATION
    XPTIScope PrepareNotify((void *)this,
                            (uint16_t)xpti::trace_point_type_t::queue_create,
                            SYCL_STREAM_NAME, "queue_create");
    if (xptiTraceEnabled()) {
      // Cache the trace event, stream id and instance IDs for the destructor
      MTraceEvent = (void *)PrepareNotify.traceEvent();
      MStreamID = PrepareNotify.streamID();
      MInstanceID = PrepareNotify.instanceID();
    }
    // Add the function to capture meta data for the XPTI trace event
    PrepareNotify.addMetadata([&](auto TEvent) {
      xpti::addMetadata(TEvent, "sycl_context",
                        reinterpret_cast<size_t>(MContext->getHandleRef()));
      if (MDevice) {
        xpti::addMetadata(TEvent, "sycl_device_name", MDevice->getDeviceName());
        xpti::addMetadata(
            TEvent, "sycl_device",
            reinterpret_cast<size_t>(
                MDevice->is_host() ? 0 : MDevice->getHandleRef()));
      }
      xpti::addMetadata(TEvent, "is_inorder", MIsInorder);
    });
    PrepareNotify.notify();
#endif
    if (has_property<ext::oneapi::property::queue::discard_events>() &&
        has_property<property::queue::enable_profiling>()) {
      throw sycl::exception(make_error_code(errc::invalid),
                            "Queue cannot be constructed with both of "
                            "discard_events and enable_profiling.");
    }
 
    MQueues.push_back(pi::cast<RT::PiQueue>(PiQueue));
 
    RT::PiDevice DevicePI{};
    const detail::plugin &Plugin = getPlugin();
    // TODO catch an exception and put it to list of asynchronous exceptions
    Plugin.call<PiApiKind::piQueueGetInfo>(
        MQueues[0], PI_QUEUE_INFO_DEVICE, sizeof(DevicePI), &DevicePI, nullptr);
    MDevice = MContext->findMatchingDeviceImpl(DevicePI);
    if (MDevice == nullptr) {
      throw sycl::exception(
          make_error_code(errc::invalid),
          "Device provided by native Queue not found in Context.");
    }
  }
 
  ~queue_impl() {
    // The trace event created in the constructor should be active through the
    // lifetime of the queue object as member variables when ABI breakage is
    // allowed. This example shows MTraceEvent as a member variable.
#if XPTI_ENABLE_INSTRUMENTATION
    if (xptiTraceEnabled()) {
      // Used cached information in member variables
      xptiNotifySubscribers(
          MStreamID, (uint16_t)xpti::trace_point_type_t::queue_destroy, nullptr,
          (xpti::trace_event_data_t *)MTraceEvent, MInstanceID,
          static_cast<const void *>("queue_destroy"));
    }
#endif
    throw_asynchronous();
    if (!MHostQueue) {
      getPlugin().call<PiApiKind::piQueueRelease>(MQueues[0]);
    }
  }
 
  cl_command_queue get() {
    if (MHostQueue) {
      throw invalid_object_error(
          "This instance of queue doesn't support OpenCL interoperability",
          PI_ERROR_INVALID_QUEUE);
    }
    getPlugin().call<PiApiKind::piQueueRetain>(MQueues[0]);
    return pi::cast<cl_command_queue>(MQueues[0]);
  }
 
  context get_context() const {
    return createSyclObjFromImpl<context>(MContext);
  }
 
  const plugin &getPlugin() const { return MContext->getPlugin(); }
 
  const ContextImplPtr &getContextImplPtr() const { return MContext; }
 
  const DeviceImplPtr &getDeviceImplPtr() const { return MDevice; }
 
  device get_device() const { return createSyclObjFromImpl<device>(MDevice); }
 
  bool is_host() const { return MHostQueue; }
 
  bool has_discard_events_support() const { return MHasDiscardEventsSupport; }
 
  bool isInOrder() const { return MIsInorder; }
 
  template <typename Param> typename Param::return_type get_info() const;
 
  using SubmitPostProcessF = std::function<void(bool, bool, event &)>;
 
  event submit(const std::function<void(handler &)> &CGF,
               const std::shared_ptr<queue_impl> &Self,
               const std::shared_ptr<queue_impl> &SecondQueue,
               const detail::code_location &Loc,
               const SubmitPostProcessF *PostProcess = nullptr) {
    try {
      return submit_impl(CGF, Self, Self, SecondQueue, Loc, PostProcess);
    } catch (...) {
      return SecondQueue->submit_impl(CGF, SecondQueue, Self, SecondQueue, Loc,
                                      PostProcess);
    }
  }
 
  event submit(const std::function<void(handler &)> &CGF,
               const std::shared_ptr<queue_impl> &Self,
               const detail::code_location &Loc,
               const SubmitPostProcessF *PostProcess = nullptr) {
    return submit_impl(CGF, Self, Self, nullptr, Loc, PostProcess);
  }
 
  void wait(const detail::code_location &Loc = {});
 
  exception_list getExceptionList() const { return MExceptions; }
 
  void wait_and_throw(const detail::code_location &Loc = {}) {
    wait(Loc);
    throw_asynchronous();
  }
 
  void throw_asynchronous() {
    if (!MAsyncHandler)
      return;
 
    exception_list Exceptions;
    {
      std::lock_guard<std::mutex> Lock(MMutex);
      std::swap(Exceptions, MExceptions);
    }
    // Unlock the mutex before calling user-provided handler to avoid
    // potential deadlock if the same queue is somehow referenced in the
    // handler.
    if (Exceptions.size())
      MAsyncHandler(std::move(Exceptions));
  }
 
  RT::PiQueue createQueue(QueueOrder Order) {
    RT::PiQueueProperties CreationFlags = 0;
 
    if (Order == QueueOrder::OOO) {
      CreationFlags = PI_QUEUE_FLAG_OUT_OF_ORDER_EXEC_MODE_ENABLE;
    }
    if (MPropList.has_property<property::queue::enable_profiling>()) {
      CreationFlags |= PI_QUEUE_FLAG_PROFILING_ENABLE;
    }
    if (MPropList.has_property<
            ext::oneapi::cuda::property::queue::use_default_stream>()) {
      CreationFlags |= __SYCL_PI_CUDA_USE_DEFAULT_STREAM;
    }
    if (MPropList
            .has_property<ext::oneapi::property::queue::discard_events>()) {
      // Pass this flag to the Level Zero plugin to be able to check it from
      // queue property.
      CreationFlags |= PI_EXT_ONEAPI_QUEUE_FLAG_DISCARD_EVENTS;
    }
    // Track that priority settings are not ambiguous.
    bool PrioritySeen = false;
    if (MPropList
            .has_property<ext::oneapi::property::queue::priority_normal>()) {
      // Normal is the default priority, don't pass anything.
      PrioritySeen = true;
    }
    if (MPropList.has_property<ext::oneapi::property::queue::priority_low>()) {
      if (PrioritySeen) {
        throw sycl::exception(
            make_error_code(errc::invalid),
            "Queue cannot be constructed with different priorities.");
      }
      CreationFlags |= PI_EXT_ONEAPI_QUEUE_FLAG_PRIORITY_LOW;
      PrioritySeen = true;
    }
    if (MPropList.has_property<ext::oneapi::property::queue::priority_high>()) {
      if (PrioritySeen) {
        throw sycl::exception(
            make_error_code(errc::invalid),
            "Queue cannot be constructed with different priorities.");
      }
      CreationFlags |= PI_EXT_ONEAPI_QUEUE_FLAG_PRIORITY_HIGH;
      PrioritySeen = true;
    }
    RT::PiQueue Queue{};
    RT::PiContext Context = MContext->getHandleRef();
    RT::PiDevice Device = MDevice->getHandleRef();
    const detail::plugin &Plugin = getPlugin();
 
    assert(Plugin.getBackend() == MDevice->getPlugin().getBackend());
    RT::PiQueueProperties Properties[] = {PI_QUEUE_FLAGS, CreationFlags, 0, 0,
                                          0};
    if (has_property<ext::intel::property::queue::compute_index>()) {
      int Idx = get_property<ext::intel::property::queue::compute_index>()
                    .get_index();
      Properties[2] = PI_QUEUE_COMPUTE_INDEX;
      Properties[3] = static_cast<RT::PiQueueProperties>(Idx);
    }
    RT::PiResult Error = Plugin.call_nocheck<PiApiKind::piextQueueCreate>(
        Context, Device, Properties, &Queue);
 
    // If creating out-of-order queue failed and this property is not
    // supported (for example, on FPGA), it will return
    // PI_ERROR_INVALID_QUEUE_PROPERTIES and will try to create in-order queue.
    if (!MEmulateOOO && Error == PI_ERROR_INVALID_QUEUE_PROPERTIES) {
      MEmulateOOO = true;
      Queue = createQueue(QueueOrder::Ordered);
    } else {
      Plugin.checkPiResult(Error);
    }
 
    return Queue;
  }
 
  RT::PiQueue &getExclusiveQueueHandleRef() {
    RT::PiQueue *PIQ = nullptr;
    bool ReuseQueue = false;
    {
      std::lock_guard<std::mutex> Lock(MMutex);
 
      // To achieve parallelism for FPGA with in order execution model with
      // possibility of two kernels to share data with each other we shall
      // create a queue for every kernel enqueued.
      if (MQueues.size() < MaxNumQueues) {
        MQueues.push_back({});
        PIQ = &MQueues.back();
      } else {
        // If the limit of OpenCL queues is going to be exceeded - take the
        // earliest used queue, wait until it finished and then reuse it.
        PIQ = &MQueues[MNextQueueIdx];
        MNextQueueIdx = (MNextQueueIdx + 1) % MaxNumQueues;
        ReuseQueue = true;
      }
    }
 
    if (!ReuseQueue)
      *PIQ = createQueue(QueueOrder::Ordered);
    else
      getPlugin().call<PiApiKind::piQueueFinish>(*PIQ);
 
    return *PIQ;
  }
 
  RT::PiQueue &getHandleRef() {
    if (!MEmulateOOO)
      return MQueues[0];
 
    return getExclusiveQueueHandleRef();
  }
 
  template <typename propertyT> bool has_property() const noexcept {
    return MPropList.has_property<propertyT>();
  }
 
  template <typename propertyT> propertyT get_property() const {
    return MPropList.get_property<propertyT>();
  }
 
  event memset(const std::shared_ptr<queue_impl> &Self, void *Ptr, int Value,
               size_t Count, const std::vector<event> &DepEvents);
  event memcpy(const std::shared_ptr<queue_impl> &Self, void *Dest,
               const void *Src, size_t Count,
               const std::vector<event> &DepEvents);
  event mem_advise(const std::shared_ptr<queue_impl> &Self, const void *Ptr,
                   size_t Length, pi_mem_advice Advice,
                   const std::vector<event> &DepEvents);
 
  void reportAsyncException(const std::exception_ptr &ExceptionPtr) {
    std::lock_guard<std::mutex> Lock(MMutex);
    MExceptions.PushBack(ExceptionPtr);
  }
 
  ThreadPool &getThreadPool() {
    return GlobalHandler::instance().getHostTaskThreadPool();
  }
 
  pi_native_handle getNative() const;
 
  buffer<AssertHappened, 1> &getAssertHappenedBuffer() {
    return MAssertHappenedBuffer;
  }
 
  void registerStreamServiceEvent(const EventImplPtr &Event) {
    std::lock_guard<std::mutex> Lock(MMutex);
    MStreamsServiceEvents.push_back(Event);
  }
 
  bool ext_oneapi_empty() const;
 
  bool is_in_fusion_mode() {
    return detail::Scheduler::getInstance().isInFusionMode(
        std::hash<typename std::shared_ptr<queue_impl>::element_type *>()(
            this));
  }
 
  event memcpyToDeviceGlobal(const std::shared_ptr<queue_impl> &Self,
                             void *DeviceGlobalPtr, const void *Src,
                             bool IsDeviceImageScope, size_t NumBytes,
                             size_t Offset,
                             const std::vector<event> &DepEvents);
  event memcpyFromDeviceGlobal(const std::shared_ptr<queue_impl> &Self,
                               void *Dest, const void *DeviceGlobalPtr,
                               bool IsDeviceImageScope, size_t NumBytes,
                               size_t Offset,
                               const std::vector<event> &DepEvents);
 
protected:
  // template is needed for proper unit testing
  template <typename HandlerType = handler>
  void finalizeHandler(HandlerType &Handler, const CG::CGTYPE &Type,
                       event &EventRet) {
    if (MIsInorder) {
 
      auto IsExpDepManaged = [](const CG::CGTYPE &Type) {
        return (Type == CG::CGTYPE::CodeplayHostTask ||
                Type == CG::CGTYPE::CodeplayInteropTask);
      };
 
      // Accessing and changing of an event isn't atomic operation.
      // Hence, here is the lock for thread-safety.
      std::lock_guard<std::mutex> Lock{MLastEventMtx};
 
      if (MLastCGType == CG::CGTYPE::None)
        MLastCGType = Type;
      // Also handles case when sync model changes. E.g. Last is host, new is
      // kernel.
      bool NeedSeparateDependencyMgmt =
          IsExpDepManaged(Type) || IsExpDepManaged(MLastCGType);
 
      if (NeedSeparateDependencyMgmt)
        Handler.depends_on(MLastEvent);
 
      EventRet = Handler.finalize();
 
      MLastEvent = EventRet;
      MLastCGType = Type;
    } else
      EventRet = Handler.finalize();
  }
 
protected:
  event submit_impl(const std::function<void(handler &)> &CGF,
                    const std::shared_ptr<queue_impl> &Self,
                    const std::shared_ptr<queue_impl> &PrimaryQueue,
                    const std::shared_ptr<queue_impl> &SecondaryQueue,
                    const detail::code_location &Loc,
                    const SubmitPostProcessF *PostProcess) {
    handler Handler(Self, PrimaryQueue, SecondaryQueue, MHostQueue);
    Handler.saveCodeLoc(Loc);
    CGF(Handler);
 
    // Scheduler will later omit events, that are not required to execute tasks.
    // Host and interop tasks, however, are not submitted to low-level runtimes
    // and require separate dependency management.
    const CG::CGTYPE Type = Handler.getType();
    event Event = detail::createSyclObjFromImpl<event>(
        std::make_shared<detail::event_impl>());
 
    if (PostProcess) {
      bool IsKernel = Type == CG::Kernel;
      bool KernelUsesAssert = false;
 
      if (IsKernel)
        // Kernel only uses assert if it's non interop one
        KernelUsesAssert = !(Handler.MKernel && Handler.MKernel->isInterop()) &&
                           ProgramManager::getInstance().kernelUsesAssert(
                               Handler.MOSModuleHandle, Handler.MKernelName);
 
      finalizeHandler(Handler, Type, Event);
 
      (*PostProcess)(IsKernel, KernelUsesAssert, Event);
    } else
      finalizeHandler(Handler, Type, Event);
 
    addEvent(Event);
    return Event;
  }
 
  // When instrumentation is enabled emits trace event for wait begin and
  // returns the telemetry event generated for the wait
  void *instrumentationProlog(const detail::code_location &CodeLoc,
                              std::string &Name, int32_t StreamID,
                              uint64_t &iid);
  // Uses events generated by the Prolog and emits wait done event
  void instrumentationEpilog(void *TelementryEvent, std::string &Name,
                             int32_t StreamID, uint64_t IId);
 
  void addSharedEvent(const event &Event);
 
  void addEvent(const event &Event);
 
  mutable std::mutex MMutex;
 
  DeviceImplPtr MDevice;
  const ContextImplPtr MContext;
 
  std::vector<std::weak_ptr<event_impl>> MEventsWeak;
 
  std::vector<event> MEventsShared;
  exception_list MExceptions;
  const async_handler MAsyncHandler;
  const property_list MPropList;
 
  std::vector<RT::PiQueue> MQueues;
  size_t MNextQueueIdx = 0;
 
  const bool MHostQueue = false;
  bool MEmulateOOO = false;
 
  // Buffer to store assert failure descriptor
  buffer<AssertHappened, 1> MAssertHappenedBuffer;
 
  // This event is employed for enhanced dependency tracking with in-order queue
  // Access to the event should be guarded with MLastEventMtx
  event MLastEvent;
  mutable std::mutex MLastEventMtx;
  // Used for in-order queues in pair with MLastEvent
  // Host tasks are explicitly synchronized in RT, pi tasks - implicitly by
  // backend. Using type to setup explicit sync between host and pi tasks.
  CG::CGTYPE MLastCGType = CG::CGTYPE::None;
 
  const bool MIsInorder;
 
  std::vector<EventImplPtr> MStreamsServiceEvents;
 
  // All member variable defined here  are needed for the SYCL instrumentation
  // layer. Do not guard these variables below with XPTI_ENABLE_INSTRUMENTATION
  // to ensure we have the same object layout when the macro in the library and
  // SYCL app are not the same.
  void *MTraceEvent = nullptr;
  uint8_t MStreamID;
  uint64_t MInstanceID = 0;
 
public:
  // Queue constructed with the discard_events property
  const bool MDiscardEvents;
  const bool MIsProfilingEnabled;
 
protected:
  // This flag says if we can discard events based on a queue "setup" which will
  // be common for all operations submitted to the queue. This is a must
  // condition for discarding, but even if it's true, in some cases, we won't be
  // able to discard events, because the final decision is made right before the
  // operation itself.
  const bool MHasDiscardEventsSupport;
};
 
} // namespace detail
} // __SYCL_INLINE_VER_NAMESPACE(_V1)
} // namespace sycl