[SYCL] Improve DPC++ runtime performance running on L0 (intel#3797)

Fixes a performance issue due to intel#3612

This change does two things to fix the performance issue.
The first is in the level zero plugin.  Here the change is to only
close and submit the batch if the event being queried is one that
will be signalled one of the commands in the batch.

The second change is in the sycl run-time itself.  This change
prevents the event cleanup code from querying every single event
that is outstanding in the system.  This is necessary to prevent
the most recent events (which are likely to be in the plug-ins batch)
from being queried.

Both changes are required to fix the performance issue.

Author: kbsmith-intel

sycl/plugins/level_zero/pi_level_zero.cpp

@@ -4111,8 +4111,13 @@ pi_result piEventGetInfo(pi_event Event, pi_event_info ParamName,
       // Lock automatically releases when this goes out of scope.
       std::lock_guard<std::mutex> lock(Event->Queue->PiQueueMutex);
 
-      if (auto Res = Event->Queue->executeOpenCommandList())
-        return Res;
+      // Only do the execute of the open command list if the event that
+      // is being queried and event that is to be signalled by something
+      // currently in that open command list.
+      if (Event->Queue->ZeOpenCommandList == Event->ZeCommandList) {
+        if (auto Res = Event->Queue->executeOpenCommandList())
+          return Res;
+      }
     }
 
     ze_result_t ZeResult;

sycl/source/detail/queue_impl.cpp

@@ -122,14 +122,21 @@ void queue_impl::addSharedEvent(const event &Event) {
   // of them can be released.
   const size_t EventThreshold = 128;
   if (MEventsShared.size() >= EventThreshold) {
+    // Generally, the vector is ordered so that the oldest events are in the
+    // front and the newer events are in the end.  So, search to find the first
+    // event that isn't yet complete.  All the events prior to that can be
+    // erased. This could leave some few events further on that have completed
+    // not yet erased, but that is OK.  This cleanup doesn't have to be perfect.
+    // This also keeps the algorithm linear rather than quadratic because it
+    // doesn't continually recheck things towards the back of the list that
+    // really haven't had time to complete.
     MEventsShared.erase(
-        std::remove_if(
-            MEventsShared.begin(), MEventsShared.end(),
-            [](const event &E) {
-              return E.get_info<info::event::command_execution_status>() ==
+        MEventsShared.begin(),
+        std::find_if(
+            MEventsShared.begin(), MEventsShared.end(), [](const event &E) {
+              return E.get_info<info::event::command_execution_status>() !=
                      info::event_command_status::complete;
-            }),
-        MEventsShared.end());
+            }));
   }
   MEventsShared.push_back(Event);
 }

sycl/unittests/queue/EventClear.cpp

@@ -23,6 +23,8 @@ struct TestCtx {
 
 std::unique_ptr<TestCtx> TestContext;
 
+const int ExpectedEventThreshold = 128;
+
 pi_result redefinedUSMEnqueueMemset(pi_queue queue, void *ptr, pi_int32 value,
                                     size_t count,
                                     pi_uint32 num_events_in_waitlist,
@@ -44,10 +46,16 @@ pi_result redefinedEventGetInfo(pi_event event, pi_event_info param_name,
                                 size_t *param_value_size_ret) {
   EXPECT_EQ(param_name, PI_EVENT_INFO_COMMAND_EXECUTION_STATUS)
       << "Unexpected event info requested";
-  // Report half of events as complete
+  // Report first half of events as complete.
+  // Report second half of events as running.
+  // This is important, because removal algorithm assumes that
+  // events are likely to be removed oldest first, and stops removing
+  // at the first non-completed event.
   static int Counter = 0;
   auto *Result = reinterpret_cast<pi_event_status *>(param_value);
-  *Result = (++Counter % 2 == 0) ? PI_EVENT_COMPLETE : PI_EVENT_RUNNING;
+  *Result = (Counter < (ExpectedEventThreshold / 2)) ? PI_EVENT_COMPLETE
+                                                     : PI_EVENT_RUNNING;
+  Counter++;
   return PI_SUCCESS;
 }
 
@@ -117,7 +125,6 @@ TEST(QueueEventClear, CleanupOnThreshold) {
   queue Q{Ctx, default_selector()};
 
   unsigned char *HostAlloc = (unsigned char *)malloc_host(1, Ctx);
-  const int ExpectedEventThreshold = 128;
   TestContext->EventReferenceCount = ExpectedEventThreshold;
   for (size_t I = 0; I < ExpectedEventThreshold; ++I) {
     Q.memset(HostAlloc, 42, 1).wait();