[SYCL] Cache and reuse events in the Level Zero plugin.

sycl/plugins/level_zero/pi_level_zero.cpp

@@ -99,6 +99,16 @@ static const bool UseMultipleCmdlistBarriers = [] {
   return std::stoi(UseMultipleCmdlistBarriersFlag) > 0;
 }();
 
+// This is an experimental option that allows to disable caching of events in
+// the context.
+static const bool DisableEventsCaching = [] {
+  const char *DisableEventsCachingFlag =
+      std::getenv("SYCL_PI_LEVEL_ZERO_DISABLE_EVENTS_CACHING");
+  if (!DisableEventsCachingFlag)
+    return false;
+  return std::stoi(DisableEventsCachingFlag) != 0;
+}();
+
 // This class encapsulates actions taken along with a call to Level Zero API.
 class ZeCall {
 private:
@@ -468,10 +478,18 @@ _pi_context::getFreeSlotInExistingOrNewPool(ze_event_pool_handle_t &Pool,
   std::list<ze_event_pool_handle_t> *ZePoolCache =
       getZeEventPoolCache(HostVisible, ProfilingEnabled);
 
-  // Remove full pool from the cache.
   if (!ZePoolCache->empty()) {
     if (NumEventsAvailableInEventPool[ZePoolCache->front()] == 0) {
-      ZePoolCache->erase(ZePoolCache->begin());
+      if (DisableEventsCaching) {
+        // Remove full pool from the cache if events caching is disabled.
+        ZePoolCache->erase(ZePoolCache->begin());
+      } else {
+        // If event caching is enabled then we don't destroy events so there is
+        // no need to remove pool from the cache and add it back when it has
+        // available slots. Just keep it in the tail of the cache so that all
+        // pools can be destroyed during context destruction.
+        ZePoolCache->push_front(nullptr);
+      }
     }
   }
   if (ZePoolCache->empty()) {
@@ -868,7 +886,18 @@ pi_result _pi_context::initialize() {
 pi_result _pi_context::finalize() {
   // This function is called when pi_context is deallocated, piContextRelease.
   // There could be some memory that may have not been deallocated.
-  // For example, event pool caches would be still alive.
+  // For example, event and event pool caches would be still alive.
+
+  if (!DisableEventsCaching) {
+    std::scoped_lock Lock(EventCacheMutex);
+    for (auto &EventCache : EventCaches) {
+      for (auto Event : EventCache) {
+        ZE_CALL(zeEventDestroy, (Event->ZeEvent));
+        delete Event;
+      }
+      EventCache.clear();
+    }
+  }
   {
     std::scoped_lock Lock(ZeEventPoolCacheMutex);
     for (auto &ZePoolCache : ZeEventPoolCache) {
@@ -5430,24 +5459,67 @@ _pi_event::getOrCreateHostVisibleEvent(ze_event_handle_t &ZeHostVisibleEvent) {
   return PI_SUCCESS;
 }
 
+pi_result _pi_event::reset() {
+  Queue = nullptr;
+  CleanedUp = false;
+  Completed = false;
+  CommandData = nullptr;
+  CommandType = PI_COMMAND_TYPE_USER;
+  WaitList = {};
+  RefCount.reset(1);
+
+  if (!isHostVisible())
+    HostVisibleEvent = nullptr;
+
+  ZE_CALL(zeEventHostReset, (ZeEvent));
+  return PI_SUCCESS;
+}
+
+pi_event _pi_context::getEventFromCache(bool HostVisible, bool WithProfiling) {
+  std::scoped_lock Lock(EventCacheMutex);
+  auto Cache = getEventCache(HostVisible, WithProfiling);
+  if (Cache->empty())
+    return nullptr;
+
+  auto It = Cache->begin();
+  pi_event Event = *It;
+  Cache->erase(It);
+  return Event;
+}
+
+void _pi_context::addEventToCache(pi_event Event) {
+  std::scoped_lock Lock(EventCacheMutex);
+  auto Cache =
+      getEventCache(Event->isHostVisible(), Event->isProfilingEnabled());
+  Event->reset();
+  Cache->emplace_back(Event);
+}
+
 // Helper function for creating a PI event.
 // The "Queue" argument specifies the PI queue where a command is submitted.
 // The "HostVisible" argument specifies if event needs to be allocated from
 // a host-visible pool.
 //
 static pi_result EventCreate(pi_context Context, pi_queue Queue,
                              bool HostVisible, pi_event *RetEvent) {
-
   bool ProfilingEnabled =
       !Queue || (Queue->Properties & PI_QUEUE_PROFILING_ENABLE) != 0;
 
-  size_t Index = 0;
+  if (auto CachedEvent =
+          Context->getEventFromCache(HostVisible, ProfilingEnabled)) {
+    *RetEvent = CachedEvent;
+    return PI_SUCCESS;
+  }
+
+  ze_event_handle_t ZeEvent;
   ze_event_pool_handle_t ZeEventPool = {};
+
+  size_t Index = 0;
+
   if (auto Res = Context->getFreeSlotInExistingOrNewPool(
           ZeEventPool, Index, HostVisible, ProfilingEnabled))
     return Res;
 
-  ze_event_handle_t ZeEvent;
   ZeStruct<ze_event_desc_t> ZeEventDesc;
   ZeEventDesc.index = Index;
   ZeEventDesc.wait = 0;
@@ -5456,9 +5528,9 @@ static pi_result EventCreate(pi_context Context, pi_queue Queue,
     ZeEventDesc.signal = ZE_EVENT_SCOPE_FLAG_HOST;
   } else {
     //
-    // Set the scope to "device" for every event. This is sufficient for global
-    // device access and peer device access. If needed to be seen on the host
-    // we are doing special handling, see EventsScope options.
+    // Set the scope to "device" for every event. This is sufficient for
+    // global device access and peer device access. If needed to be seen on
+    // the host we are doing special handling, see EventsScope options.
     //
     // TODO: see if "sub-device" (ZE_EVENT_SCOPE_FLAG_SUBDEVICE) can better be
     //       used in some circumstances.
@@ -5819,7 +5891,12 @@ pi_result piEventRelease(pi_event Event) {
     Event->CommandData = nullptr;
   }
   if (Event->OwnZeEvent) {
-    ZE_CALL(zeEventDestroy, (Event->ZeEvent));
+    if (DisableEventsCaching) {
+      ZE_CALL(zeEventDestroy, (Event->ZeEvent));
+      auto Context = Event->Context;
+      if (auto Res = Context->decrementUnreleasedEventsInPool(Event))
+        return Res;
+    }
   }
   // It is possible that host-visible event was never created.
   // In case it was check if that's different from this same event
@@ -5829,18 +5906,19 @@ pi_result piEventRelease(pi_event Event) {
     PI_CALL(piEventRelease(Event->HostVisibleEvent));
   }
 
-  auto Context = Event->Context;
-  if (auto Res = Context->decrementUnreleasedEventsInPool(Event))
-    return Res;
-
   // We intentionally incremented the reference counter when an event is
   // created so that we can avoid pi_queue is released before the associated
   // pi_event is released. Here we have to decrement it so pi_queue
   // can be released successfully.
   if (Event->Queue) {
     PI_CALL(piQueueReleaseInternal(Event->Queue));
   }
-  delete Event;
+
+  if (DisableEventsCaching || !Event->OwnZeEvent) {
+    delete Event;
+  } else {
+    Event->Context->addEventToCache(Event);
+  }
 
   return PI_SUCCESS;
 }

sycl/plugins/level_zero/pi_level_zero.hpp

@@ -276,6 +276,8 @@ template <class T> struct ZeCache : private T {
 struct ReferenceCounter {
   ReferenceCounter(pi_uint32 InitVal) : RefCount{InitVal} {}
 
+  void reset(pi_uint32 InitVal) { RefCount = InitVal; }
+
   // Used when retaining an object.
   void increment() { RefCount++; }
 
@@ -750,6 +752,12 @@ struct _pi_context : _pi_object {
   // when kernel has finished execution.
   std::unordered_map<void *, MemAllocRecord> MemAllocs;
 
+  // Get pi_event from cache.
+  pi_event getEventFromCache(bool HostVisible, bool WithProfiling);
+
+  // Add pi_event to cache.
+  void addEventToCache(pi_event);
+
 private:
   // If context contains one device then return this device.
   // If context contains sub-devices of the same device, then return this parent
@@ -798,6 +806,20 @@ struct _pi_context : _pi_object {
   // Mutex to control operations on event pool caches and the helper maps
   // holding the current pool usage counts.
   pi_mutex ZeEventPoolCacheMutex;
+
+  // Mutex to control operations on event caches.
+  pi_mutex EventCacheMutex;
+
+  // Caches for events.
+  std::vector<std::list<pi_event>> EventCaches{4};
+
+  // Get the cache of events for a provided scope and profiling mode.
+  auto getEventCache(bool HostVisible, bool WithProfiling) {
+    if (HostVisible)
+      return WithProfiling ? &EventCaches[0] : &EventCaches[1];
+    else
+      return WithProfiling ? &EventCaches[2] : &EventCaches[3];
+  }
 };
 
 struct _pi_queue : _pi_object {
@@ -1350,6 +1372,9 @@ struct _pi_event : _pi_object {
   // L0 event (if any) is not guranteed to have been signalled, or
   // being visible to the host at all.
   bool Completed = {false};
+
+  // Reset _pi_event object.
+  pi_result reset();
 };
 
 struct _pi_program : _pi_object {