[SYCL] Implement queue flushing (#5052)

According to OpenCL spec:
"To use event objects that refer to commands enqueued in a command-queue as
event objects to wait on by commands enqueued in a different command-queue,
the application must call a clFlush or any blocking commands that perform an
implicit flush of the command-queue where the commands that refer to these
event objects are enqueued."

This patch adds a piQueueFlush function that's lowered to clFinish in OpenCL
and called whenever an unsubmitted cross-queue dependency is about to be used.

Author: sergey-semenov

sycl/include/CL/sycl/detail/pi.def

@@ -45,6 +45,7 @@ _PI_API(piextContextCreateWithNativeHandle)
 _PI_API(piQueueCreate)
 _PI_API(piQueueGetInfo)
 _PI_API(piQueueFinish)
+_PI_API(piQueueFlush)
 _PI_API(piQueueRetain)
 _PI_API(piQueueRelease)
 _PI_API(piextQueueGetNativeHandle)

sycl/include/CL/sycl/detail/pi.h

@@ -40,11 +40,12 @@
 // changes the API version from 3.5 to 4.6.
 // 5.7 Added new context and ownership arguments to
 //   piextEventCreateWithNativeHandle
-// 6.8 Added new ownership argument to piextProgramCreateWithNativeHandle.
+// 6.8 Added new ownership argument to piextProgramCreateWithNativeHandle. Added
+// piQueueFlush function.
 //
 #include "CL/cl.h"
-#define _PI_H_VERSION_MAJOR 5
-#define _PI_H_VERSION_MINOR 7
+#define _PI_H_VERSION_MAJOR 6
+#define _PI_H_VERSION_MINOR 8
 
 #define _PI_STRING_HELPER(a) #a
 #define _PI_CONCAT(a, b) _PI_STRING_HELPER(a.b)
@@ -1083,6 +1084,8 @@ __SYCL_EXPORT pi_result piQueueRelease(pi_queue command_queue);
 
 __SYCL_EXPORT pi_result piQueueFinish(pi_queue command_queue);
 
+__SYCL_EXPORT pi_result piQueueFlush(pi_queue command_queue);
+
 /// Gets the native handle of a PI queue object.
 ///
 /// \param queue is the PI queue to get the native handle of.

sycl/plugins/cuda/pi_cuda.cpp

@@ -2254,6 +2254,11 @@ pi_result cuda_piQueueFinish(pi_queue command_queue) {
   return result;
 }
 
+// There is no CUDA counterpart for queue flushing and we don't run into the
+// same problem of having to flush cross-queue dependencies as some of the
+// other plugins, so it can be left as no-op.
+pi_result cuda_piQueueFlush(pi_queue command_queue) { return PI_SUCCESS; }
+
 /// Gets the native CUDA handle of a PI queue object
 ///
 /// \param[in] queue The PI queue to get the native CUDA object of.
@@ -4886,6 +4891,7 @@ pi_result piPluginInit(pi_plugin *PluginInit) {
   _PI_CL(piQueueCreate, cuda_piQueueCreate)
   _PI_CL(piQueueGetInfo, cuda_piQueueGetInfo)
   _PI_CL(piQueueFinish, cuda_piQueueFinish)
+  _PI_CL(piQueueFlush, cuda_piQueueFlush)
   _PI_CL(piQueueRetain, cuda_piQueueRetain)
   _PI_CL(piQueueRelease, cuda_piQueueRelease)
   _PI_CL(piextQueueGetNativeHandle, cuda_piextQueueGetNativeHandle)

sycl/plugins/hip/pi_hip.cpp

@@ -2202,6 +2202,11 @@ pi_result hip_piQueueFinish(pi_queue command_queue) {
   return result;
 }
 
+// There is no HIP counterpart for queue flushing and we don't run into the
+// same problem of having to flush cross-queue dependencies as some of the
+// other plugins, so it can be left as no-op.
+pi_result hip_piQueueFlush(pi_queue command_queue) { return PI_SUCCESS; }
+
 /// Gets the native HIP handle of a PI queue object
 ///
 /// \param[in] queue The PI queue to get the native HIP object of.
@@ -4820,6 +4825,7 @@ pi_result piPluginInit(pi_plugin *PluginInit) {
   _PI_CL(piQueueCreate, hip_piQueueCreate)
   _PI_CL(piQueueGetInfo, hip_piQueueGetInfo)
   _PI_CL(piQueueFinish, hip_piQueueFinish)
+  _PI_CL(piQueueFlush, hip_piQueueFlush)
   _PI_CL(piQueueRetain, hip_piQueueRetain)
   _PI_CL(piQueueRelease, hip_piQueueRelease)
   _PI_CL(piextQueueGetNativeHandle, hip_piextQueueGetNativeHandle)

sycl/plugins/level_zero/pi_level_zero.cpp

@@ -3029,6 +3029,10 @@ pi_result piQueueFinish(pi_queue Queue) {
   return PI_SUCCESS;
 }
 
+// Flushing cross-queue dependencies is covered by createAndRetainPiZeEventList,
+// so this can be left as a no-op.
+pi_result piQueueFlush(pi_queue Queue) { return PI_SUCCESS; }
+
 pi_result piextQueueGetNativeHandle(pi_queue Queue,
                                     pi_native_handle *NativeHandle) {
   PI_ASSERT(Queue, PI_INVALID_QUEUE);

sycl/plugins/opencl/pi_opencl.cpp

@@ -1424,6 +1424,7 @@ pi_result piPluginInit(pi_plugin *PluginInit) {
   _PI_CL(piQueueCreate, piQueueCreate)
   _PI_CL(piQueueGetInfo, clGetCommandQueueInfo)
   _PI_CL(piQueueFinish, clFinish)
+  _PI_CL(piQueueFlush, clFlush)
   _PI_CL(piQueueRetain, clRetainCommandQueue)
   _PI_CL(piQueueRelease, clReleaseCommandQueue)
   _PI_CL(piextQueueGetNativeHandle, piextQueueGetNativeHandle)

sycl/source/detail/event_impl.cpp

@@ -93,11 +93,12 @@ void event_impl::setContextImpl(const ContextImplPtr &Context) {
   MState = HES_NotComplete;
 }
 
-event_impl::event_impl() : MState(HES_Complete) {}
+event_impl::event_impl() : MIsFlushed(true), MState(HES_Complete) {}
 
 event_impl::event_impl(RT::PiEvent Event, const context &SyclContext)
     : MEvent(Event), MContext(detail::getSyclObjImpl(SyclContext)),
-      MOpenCLInterop(true), MHostEvent(false), MState(HES_Complete) {
+      MOpenCLInterop(true), MHostEvent(false), MIsFlushed(true),
+      MState(HES_Complete) {
 
   if (MContext->is_host()) {
     throw cl::sycl::invalid_parameter_error(
@@ -120,7 +121,7 @@ event_impl::event_impl(RT::PiEvent Event, const context &SyclContext)
   getPlugin().call<PiApiKind::piEventRetain>(MEvent);
 }
 
-event_impl::event_impl(QueueImplPtr Queue) {
+event_impl::event_impl(const QueueImplPtr &Queue) : MQueue{Queue} {
   if (Queue->is_host()) {
     MState.store(HES_NotComplete);
 
@@ -344,6 +345,32 @@ std::vector<EventImplPtr> event_impl::getWaitList() {
   return Result;
 }
 
+void event_impl::flushIfNeeded(const QueueImplPtr &UserQueue) {
+  assert(MEvent != nullptr);
+  if (MIsFlushed)
+    return;
+
+  QueueImplPtr Queue = MQueue.lock();
+  // If the queue has been released, all of the commands have already been
+  // implicitly flushed by piQueueRelease.
+  if (!Queue) {
+    MIsFlushed = true;
+    return;
+  }
+  if (Queue == UserQueue)
+    return;
+
+  // Check if the task for this event has already been submitted.
+  pi_event_status Status = PI_EVENT_QUEUED;
+  getPlugin().call<PiApiKind::piEventGetInfo>(
+      MEvent, PI_EVENT_INFO_COMMAND_EXECUTION_STATUS, sizeof(pi_int32), &Status,
+      nullptr);
+  if (Status == PI_EVENT_QUEUED) {
+    getPlugin().call<PiApiKind::piQueueFlush>(Queue->getHandleRef());
+  }
+  MIsFlushed = true;
+}
+
 void event_impl::cleanupDependencyEvents() {
   std::lock_guard<std::mutex> Lock(MMutex);
   MPreparedDepsEvents.clear();

sycl/source/detail/event_impl.hpp

@@ -44,7 +44,7 @@ class event_impl {
   /// \param Event is a valid instance of plug-in event.
   /// \param SyclContext is an instance of SYCL context.
   event_impl(RT::PiEvent Event, const context &SyclContext);
-  event_impl(QueueImplPtr Queue);
+  event_impl(const QueueImplPtr &Queue);
 
   /// Checks if this event is a SYCL host event.
   ///
@@ -182,6 +182,11 @@ class event_impl {
   /// @return a vector of "immediate" dependencies for this event_impl.
   std::vector<EventImplPtr> getWaitList();
 
+  /// Performs a flush on the queue associated with this event if the user queue
+  /// is different and the task associated with this event hasn't been submitted
+  /// to the device yet.
+  void flushIfNeeded(const QueueImplPtr &UserQueue);
+
   /// Cleans dependencies of this event_impl
   void cleanupDependencyEvents();
 
@@ -200,11 +205,16 @@ class event_impl {
   bool MHostEvent = true;
   std::unique_ptr<HostProfilingInfo> MHostProfilingInfo;
   void *MCommand = nullptr;
+  std::weak_ptr<queue_impl> MQueue;
 
   /// Dependency events prepared for waiting by backend.
   std::vector<EventImplPtr> MPreparedDepsEvents;
   std::vector<EventImplPtr> MPreparedHostDepsEvents;
 
+  /// Indicates that the task associated with this event has been submitted by
+  /// the queue to the device.
+  std::atomic<bool> MIsFlushed = false;
+
   enum HostEventState : int { HES_NotComplete = 0, HES_Complete };
 
   // State of host event. Employed only for host events and event with no

sycl/source/detail/scheduler/commands.cpp

@@ -169,6 +169,13 @@ getPiEvents(const std::vector<EventImplPtr> &EventImpls) {
   return RetPiEvents;
 }
 
+static void flushCrossQueueDeps(const std::vector<EventImplPtr> &EventImpls,
+                                const QueueImplPtr &Queue) {
+  for (auto &EventImpl : EventImpls) {
+    EventImpl->flushIfNeeded(Queue);
+  }
+}
+
 class DispatchHostTask {
   ExecCGCommand *MThisCmd;
   std::vector<interop_handle::ReqToMem> MReqToMem;
@@ -325,6 +332,7 @@ void Command::waitForEvents(QueueImplPtr Queue,
 #endif
 
       std::vector<RT::PiEvent> RawEvents = getPiEvents(EventImpls);
+      flushCrossQueueDeps(EventImpls, getWorkerQueue());
       const detail::plugin &Plugin = Queue->getPlugin();
       Plugin.call<PiApiKind::piEnqueueEventsWait>(
           Queue->getHandleRef(), RawEvents.size(), &RawEvents[0], &Event);
@@ -1073,6 +1081,7 @@ cl_int MapMemObject::enqueueImp() {
   waitForPreparedHostEvents();
   std::vector<EventImplPtr> EventImpls = MPreparedDepsEvents;
   std::vector<RT::PiEvent> RawEvents = getPiEvents(EventImpls);
+  flushCrossQueueDeps(EventImpls, getWorkerQueue());
 
   RT::PiEvent &Event = MEvent->getHandleRef();
   *MDstPtr = MemoryManager::map(
@@ -1150,6 +1159,7 @@ cl_int UnMapMemObject::enqueueImp() {
   waitForPreparedHostEvents();
   std::vector<EventImplPtr> EventImpls = MPreparedDepsEvents;
   std::vector<RT::PiEvent> RawEvents = getPiEvents(EventImpls);
+  flushCrossQueueDeps(EventImpls, getWorkerQueue());
 
   RT::PiEvent &Event = MEvent->getHandleRef();
   MemoryManager::unmap(MDstAllocaCmd->getSYCLMemObj(),
@@ -1250,6 +1260,7 @@ cl_int MemCpyCommand::enqueueImp() {
   RT::PiEvent &Event = MEvent->getHandleRef();
 
   auto RawEvents = getPiEvents(EventImpls);
+  flushCrossQueueDeps(EventImpls, getWorkerQueue());
 
   MemoryManager::copy(
       MSrcAllocaCmd->getSYCLMemObj(), MSrcAllocaCmd->getMemAllocation(),
@@ -1400,6 +1411,7 @@ cl_int MemCpyCommandHost::enqueueImp() {
     return CL_SUCCESS;
   }
 
+  flushCrossQueueDeps(EventImpls, getWorkerQueue());
   MemoryManager::copy(
       MSrcAllocaCmd->getSYCLMemObj(), MSrcAllocaCmd->getMemAllocation(),
       MSrcQueue, MSrcReq.MDims, MSrcReq.MMemoryRange, MSrcReq.MAccessRange,
@@ -1989,6 +2001,7 @@ cl_int ExecCGCommand::enqueueImp() {
     waitForPreparedHostEvents();
   std::vector<EventImplPtr> EventImpls = MPreparedDepsEvents;
   auto RawEvents = getPiEvents(EventImpls);
+  flushCrossQueueDeps(EventImpls, getWorkerQueue());
 
   RT::PiEvent &Event = MEvent->getHandleRef();
 

sycl/test/abi/pi_level_zero_symbol_check.dump

@@ -70,6 +70,7 @@ piProgramRelease
 piProgramRetain
 piQueueCreate
 piQueueFinish
+piQueueFlush
 piQueueGetInfo
 piQueueRelease
 piQueueRetain

sycl/unittests/helpers/CommonRedefinitions.hpp

@@ -114,6 +114,18 @@ inline pi_result redefinedEventsWaitCommon(pi_uint32 num_events,
   return PI_SUCCESS;
 }
 
+inline pi_result redefinedEventGetInfoCommon(pi_event event,
+                                             pi_event_info param_name,
+                                             size_t param_value_size,
+                                             void *param_value,
+                                             size_t *param_value_size_ret) {
+  if (param_name == PI_EVENT_INFO_COMMAND_EXECUTION_STATUS) {
+    auto *status = reinterpret_cast<pi_event_status *>(param_value);
+    *status = PI_EVENT_SUBMITTED;
+  }
+  return PI_SUCCESS;
+}
+
 inline pi_result redefinedEventReleaseCommon(pi_event event) {
   if (event != nullptr)
     delete reinterpret_cast<int *>(event);
@@ -166,6 +178,7 @@ inline void setupDefaultMockAPIs(sycl::unittest::PiMock &Mock) {
   Mock.redefine<PiApiKind::piKernelSetExecInfo>(
       redefinedKernelSetExecInfoCommon);
   Mock.redefine<PiApiKind::piEventsWait>(redefinedEventsWaitCommon);
+  Mock.redefine<PiApiKind::piEventGetInfo>(redefinedEventGetInfoCommon);
   Mock.redefine<PiApiKind::piEventRelease>(redefinedEventReleaseCommon);
   Mock.redefine<PiApiKind::piEnqueueKernelLaunch>(
       redefinedEnqueueKernelLaunchCommon);

sycl/unittests/scheduler/CMakeLists.txt

@@ -15,5 +15,6 @@ add_sycl_unittest(SchedulerTests OBJECT
     InOrderQueueHostTaskDeps.cpp
     AllocaLinking.cpp
     RequiredWGSize.cpp
+    QueueFlushing.cpp
     utils.cpp
 )