refactor: add debug flag to synchronize Event before reset

Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>

level_zero/core/source/event/event_impl.inl

@@ -392,6 +392,13 @@ ze_result_t EventImp<TagSizeT>::hostSynchronize(uint64_t timeout) {
 
 template <typename TagSizeT>
 ze_result_t EventImp<TagSizeT>::reset() {
+    if (NEO::DebugManager.flags.SynchronizeEventBeforeReset.get() != -1) {
+        if (NEO::DebugManager.flags.SynchronizeEventBeforeReset.get() == 2 && queryStatus() != ZE_RESULT_SUCCESS) {
+            printf("\nzeEventHostReset: Event %p not ready. Calling zeEventHostSynchronize.", this);
+        }
+
+        hostSynchronize(std::numeric_limits<uint64_t>::max());
+    }
     if (inOrderExecEvent) {
         inOrderExecEvent = false;
         inOrderExecDataAllocation = nullptr;

level_zero/core/test/unit_tests/sources/event/test_event.cpp

@@ -3218,6 +3218,7 @@ HWTEST_F(EventTests, GivenCsrTbxModeWhenEventCreatedAndSignaledThenEventAllocati
 struct MockEventCompletion : public L0::EventImp<uint32_t> {
     using EventImp<uint32_t>::gpuStartTimestamp;
     using EventImp<uint32_t>::gpuEndTimestamp;
+    using EventImp<uint32_t>::hostAddress;
 
     MockEventCompletion(L0::EventPool *eventPool, int index, L0::Device *device) : EventImp(eventPool, index, device, false) {
         auto neoDevice = device->getNEODevice();
@@ -3249,8 +3250,24 @@ struct MockEventCompletion : public L0::EventImp<uint32_t> {
         return EventImp<uint32_t>::hostEventSetValue(eventValue);
     }
 
+    ze_result_t hostSynchronize(uint64_t timeout) override {
+        hostSynchronizeCalled++;
+        return L0::EventImp<uint32_t>::hostSynchronize(timeout);
+    }
+
+    ze_result_t queryStatus() override {
+        if (failOnNextQueryStatus) {
+            failOnNextQueryStatus = false;
+            return ZE_RESULT_NOT_READY;
+        }
+
+        return L0::EventImp<uint32_t>::queryStatus();
+    }
+
     bool shouldHostEventSetValueFail = false;
+    bool failOnNextQueryStatus = false;
     uint32_t assignKernelEventCompletionDataCounter = 0u;
+    uint32_t hostSynchronizeCalled = 0;
 };
 
 TEST_F(EventTests, WhenQueryingStatusAfterHostSignalThenDontAccessMemoryAndReturnSuccess) {
@@ -3261,6 +3278,68 @@ TEST_F(EventTests, WhenQueryingStatusAfterHostSignalThenDontAccessMemoryAndRetur
     EXPECT_EQ(event->assignKernelEventCompletionDataCounter, 0u);
 }
 
+TEST_F(EventTests, givenDebugFlagSetWhenCallingResetThenSynchronizeBeforeReset) {
+    DebugManager.flags.SynchronizeEventBeforeReset.set(1);
+
+    auto event = std::make_unique<MockEventCompletion>(eventPool.get(), 1u, device);
+    event->failOnNextQueryStatus = true;
+
+    *reinterpret_cast<uint32_t *>(event->hostAddress) = Event::STATE_SIGNALED;
+
+    testing::internal::CaptureStdout();
+
+    EXPECT_EQ(0u, event->hostSynchronizeCalled);
+
+    event->reset();
+
+    EXPECT_EQ(1u, event->hostSynchronizeCalled);
+
+    std::string output = testing::internal::GetCapturedStdout();
+    std::string expectedOutput("");
+    EXPECT_EQ(expectedOutput, output);
+}
+
+TEST_F(EventTests, givenDebugFlagSetWhenCallingResetThenPrintLogAndSynchronizeBeforeReset) {
+    DebugManager.flags.SynchronizeEventBeforeReset.set(2);
+
+    auto event = std::make_unique<MockEventCompletion>(eventPool.get(), 1u, device);
+    *reinterpret_cast<uint32_t *>(event->hostAddress) = Event::STATE_SIGNALED;
+
+    {
+        event->failOnNextQueryStatus = false;
+
+        testing::internal::CaptureStdout();
+
+        EXPECT_EQ(0u, event->hostSynchronizeCalled);
+
+        event->reset();
+
+        EXPECT_EQ(1u, event->hostSynchronizeCalled);
+
+        std::string output = testing::internal::GetCapturedStdout();
+        std::string expectedOutput("");
+        EXPECT_EQ(expectedOutput, output);
+    }
+
+    {
+        event->failOnNextQueryStatus = true;
+
+        testing::internal::CaptureStdout();
+
+        EXPECT_EQ(1u, event->hostSynchronizeCalled);
+
+        event->reset();
+
+        EXPECT_EQ(2u, event->hostSynchronizeCalled);
+
+        std::string output = testing::internal::GetCapturedStdout();
+        char expectedStr[128] = {};
+        snprintf(expectedStr, 128, "\nzeEventHostReset: Event %p not ready. Calling zeEventHostSynchronize.", event.get());
+
+        EXPECT_EQ(std::string(expectedStr), output);
+    }
+}
+
 TEST_F(EventTests, whenAppendAdditionalCsrThenStoreUniqueCsr) {
     auto csr1 = reinterpret_cast<NEO::CommandStreamReceiver *>(0x1234);
     auto csr2 = reinterpret_cast<NEO::CommandStreamReceiver *>(0x5678);

shared/source/debug_settings/debug_variables_base.inl

@@ -245,6 +245,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, DebugSetMemoryDiagnosticsDelay, -1, "-1: default
 DECLARE_DEBUG_VARIABLE(int32_t, EnableDeviceStateVerification, -1, "-1: default, 0: disable, 1: enable check of device state before submit on Windows")
 DECLARE_DEBUG_VARIABLE(int32_t, EnableDynamicPostSyncAllocLayout, -1, "-1: default, 0: Keep Timestamp size layout, 1: Use write immediate layout (qword) and switch dynamically to TS for profiling")
 DECLARE_DEBUG_VARIABLE(int32_t, PrintTimestampPacketUsage, -1, "-1: default, 0: Disabled, 1: Print when TSP is allocated, initialized, returned to pool, etc.")
+DECLARE_DEBUG_VARIABLE(int32_t, SynchronizeEventBeforeReset, -1, "-1: default, 0: Disabled, 1: Synchronize Event completion on host before calling reset. 2: Synchronize + print extra logs.")
 DECLARE_DEBUG_VARIABLE(int32_t, TrackNumCsrClientsOnSyncPoints, -1, "-1: default, 0: Disabled, 1: If set, synchronization points like zeEventHostSynchronize will unregister CmdQ from CSR clients")
 
 /*LOGGING FLAGS*/

shared/test/common/test_files/igdrcl.config

@@ -537,5 +537,6 @@ EnableDynamicPostSyncAllocLayout = -1
 PrintTimestampPacketUsage = -1
 TrackNumCsrClientsOnSyncPoints = -1
 CommandListTimestampRefreshIntervalInMilliSec = -1
+SynchronizeEventBeforeReset = -1
 RemoveRestrictionsOnNumberOfThreadsInGpgpuThreadGroup = 0
 # Please don't edit below this line