Further reduce the number of wait calls for blocked commands

Any blocked command was being added to MPreparedHostDepsEvents which led
to waiting for the dependency event before enqueueing the dependant
command rather than passing the dependency to PI.

Author: sergey-semenov

sycl/source/detail/scheduler/commands.cpp

@@ -486,10 +486,14 @@ Command *Command::processDepEvent(EventImplPtr DepEvent, const DepDesc &Dep) {
   const ContextImplPtr &WorkerContext = WorkerQueue->getContextImplPtr();
 
   // 1. Async work is not supported for host device.
-  // 2. The event handle can be null in case of, for example, alloca command,
-  //    which is currently synchronous, so don't generate OpenCL event.
-  //    Though, this event isn't host one as it's context isn't host one.
-  if (DepEvent->is_host() || DepEvent->getHandleRef() == nullptr) {
+  // 2. Some types of commands do not produce PI events after they are enqueued
+  // (e.g. alloca). Note that we can't check the pi event to make that
+  // distinction since the command might still be unenqueued at this point.
+  bool PiEventExpected = !DepEvent->is_host();
+  if (auto *DepCmd = static_cast<Command *>(DepEvent->getCommand()))
+    PiEventExpected &= DepCmd->producesPiEvent();
+
+  if (!PiEventExpected) {
     // call to waitInternal() is in waitForPreparedHostEvents() as it's called
     // from enqueue process functions
     MPreparedHostDepsEvents.push_back(DepEvent);
@@ -520,6 +524,8 @@ const ContextImplPtr &Command::getWorkerContext() const {
 
 const QueueImplPtr &Command::getWorkerQueue() const { return MQueue; }
 
+bool Command::producesPiEvent() const { return true; }
+
 Command *Command::addDep(DepDesc NewDep) {
   Command *ConnectionCmd = nullptr;
 
@@ -731,6 +737,8 @@ void AllocaCommandBase::emitInstrumentationData() {
 #endif
 }
 
+bool AllocaCommandBase::producesPiEvent() const { return false; }
+
 AllocaCommand::AllocaCommand(QueueImplPtr Queue, Requirement Req,
                              bool InitFromUserData,
                              AllocaCommandBase *LinkedAllocaCmd)
@@ -998,6 +1006,8 @@ void ReleaseCommand::printDot(std::ostream &Stream) const {
   }
 }
 
+bool ReleaseCommand::producesPiEvent() const { return false; }
+
 MapMemObject::MapMemObject(AllocaCommandBase *SrcAllocaCmd, Requirement Req,
                            void **DstPtr, QueueImplPtr Queue,
                            access::mode MapMode)
@@ -1392,6 +1402,8 @@ void EmptyCommand::printDot(std::ostream &Stream) const {
   }
 }
 
+bool EmptyCommand::producesPiEvent() const { return false; }
+
 void MemCpyCommandHost::printDot(std::ostream &Stream) const {
   Stream << "\"" << this << "\" [style=filled, fillcolor=\"#B6A2EB\", label=\"";
 
@@ -2193,6 +2205,10 @@ cl_int ExecCGCommand::enqueueImp() {
   return PI_INVALID_OPERATION;
 }
 
+bool ExecCGCommand::producesPiEvent() const {
+  return MCommandGroup->getType() != CG::CGTYPE::CODEPLAY_HOST_TASK;
+}
+
 } // namespace detail
 } // namespace sycl
 } // __SYCL_INLINE_NAMESPACE(cl)

sycl/source/detail/scheduler/commands.hpp

@@ -189,6 +189,9 @@ class Command {
   /// for memory copy commands.
   virtual const QueueImplPtr &getWorkerQueue() const;
 
+  /// Returns true iff the command produces a PI event on non-host devices.
+  virtual bool producesPiEvent() const;
+
 protected:
   EventImplPtr MEvent;
   QueueImplPtr MQueue;
@@ -306,6 +309,8 @@ class EmptyCommand : public Command {
 
   void emitInstrumentationData() override;
 
+  bool producesPiEvent() const final;
+
 private:
   cl_int enqueueImp() final;
 
@@ -323,6 +328,7 @@ class ReleaseCommand : public Command {
 
   void printDot(std::ostream &Stream) const final;
   void emitInstrumentationData() override;
+  bool producesPiEvent() const final;
 
 private:
   cl_int enqueueImp() final;
@@ -347,6 +353,8 @@ class AllocaCommandBase : public Command {
 
   void emitInstrumentationData() override;
 
+  bool producesPiEvent() const final;
+
   void *MMemAllocation = nullptr;
 
   /// Alloca command linked with current command.
@@ -518,6 +526,8 @@ class ExecCGCommand : public Command {
     MCommandGroup.release();
   }
 
+  bool producesPiEvent() const final;
+
 private:
   cl_int enqueueImp() final;