Add notify parameter to direct submission post sync operations

Related-To: NEO-5845

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>

Author: zzdanowicz

opencl/test/unit_test/command_stream/command_stream_receiver_hw_2_tests.cpp

@@ -1375,7 +1375,7 @@ HWTEST_F(BcsTests, givenBlitterDirectSubmissionEnabledWhenProgrammingBlitterThen
     csr.blitterDirectSubmission = std::make_unique<DirectSubmission>(*pDevice, *csr.osContext);
     csr.recordFlusheBatchBuffer = true;
     DirectSubmission *directSubmission = reinterpret_cast<DirectSubmission *>(csr.blitterDirectSubmission.get());
-    bool initRet = directSubmission->initialize(true);
+    bool initRet = directSubmission->initialize(true, false);
     EXPECT_TRUE(initRet);
 
     cl_int retVal = CL_SUCCESS;
@@ -1418,7 +1418,7 @@ HWTEST_F(BcsTests, givenBlitterDirectSubmissionEnabledWhenFlushTagUpdateThenBatc
     csr.blitterDirectSubmission = std::make_unique<DirectSubmission>(*pDevice, *csr.osContext);
     csr.recordFlusheBatchBuffer = true;
     DirectSubmission *directSubmission = reinterpret_cast<DirectSubmission *>(csr.blitterDirectSubmission.get());
-    bool initRet = directSubmission->initialize(true);
+    bool initRet = directSubmission->initialize(true, false);
     EXPECT_TRUE(initRet);
 
     csr.flushTagUpdate();

opencl/test/unit_test/os_interface/windows/device_command_stream_tests.cpp

@@ -124,12 +124,12 @@ struct MockWddmCsr : public WddmCommandStreamReceiver<GfxFamily> {
             if (!initBlitterDirectSubmission) {
                 directSubmission = std::make_unique<
                     MockWddmDirectSubmission<GfxFamily, RenderDispatcher<GfxFamily>>>(device, osContext);
-                ret = directSubmission->initialize(true);
+                ret = directSubmission->initialize(true, false);
                 this->dispatchMode = DispatchMode::ImmediateDispatch;
             } else {
                 blitterDirectSubmission = std::make_unique<
                     MockWddmDirectSubmission<GfxFamily, BlitterDispatcher<GfxFamily>>>(device, osContext);
-                blitterDirectSubmission->initialize(true);
+                blitterDirectSubmission->initialize(true, false);
             }
         }
         return ret;

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -1431,11 +1431,11 @@ inline bool CommandStreamReceiverHw<GfxFamily>::initDirectSubmission(Device &dev
         if (!this->isAnyDirectSubmissionEnabled()) {
             if (EngineHelpers::isBcs(osContext.getEngineType())) {
                 blitterDirectSubmission = DirectSubmissionHw<GfxFamily, BlitterDispatcher<GfxFamily>>::create(device, osContext);
-                ret = blitterDirectSubmission->initialize(submitOnInit);
+                ret = blitterDirectSubmission->initialize(submitOnInit, this->isUsedNotifyEnableForPostSync());
 
             } else {
                 directSubmission = DirectSubmissionHw<GfxFamily, RenderDispatcher<GfxFamily>>::create(device, osContext);
-                ret = directSubmission->initialize(submitOnInit);
+                ret = directSubmission->initialize(submitOnInit, this->isUsedNotifyEnableForPostSync());
             }
             auto directSubmissionController = executionEnvironment.initializeDirectSubmissionController();
             if (directSubmissionController) {

shared/source/direct_submission/direct_submission_hw.h

@@ -59,7 +59,7 @@ class DirectSubmissionHw {
 
     virtual ~DirectSubmissionHw();
 
-    bool initialize(bool submitOnInit);
+    bool initialize(bool submitOnInit, bool useNotify);
 
     MOCKABLE_VIRTUAL bool stopRingBuffer();
 
@@ -160,5 +160,6 @@ class DirectSubmissionHw {
     bool disableMonitorFence = false;
     bool partitionedMode = false;
     bool partitionConfigSet = true;
+    bool useNotifyForPostSync = false;
 };
 } // namespace NEO

shared/source/direct_submission/direct_submission_hw.inl

@@ -154,7 +154,8 @@ inline void DirectSubmissionHw<GfxFamily, Dispatcher>::cpuCachelineFlush(void *p
 }
 
 template <typename GfxFamily, typename Dispatcher>
-bool DirectSubmissionHw<GfxFamily, Dispatcher>::initialize(bool submitOnInit) {
+bool DirectSubmissionHw<GfxFamily, Dispatcher>::initialize(bool submitOnInit, bool useNotify) {
+    useNotifyForPostSync = useNotify;
     bool ret = allocateResources();
 
     initDiagnostic(submitOnInit);
@@ -222,7 +223,7 @@ bool DirectSubmissionHw<GfxFamily, Dispatcher>::stopRingBuffer() {
     if (disableMonitorFence) {
         TagData currentTagData = {};
         getTagAddressValue(currentTagData);
-        Dispatcher::dispatchMonitorFence(ringCommandStream, currentTagData.tagAddress, currentTagData.tagValue, *hwInfo, false, this->partitionedMode);
+        Dispatcher::dispatchMonitorFence(ringCommandStream, currentTagData.tagAddress, currentTagData.tagValue, *hwInfo, this->useNotifyForPostSync, this->partitionedMode);
     }
     Dispatcher::dispatchStopCommandBuffer(ringCommandStream);
 
@@ -277,7 +278,7 @@ inline void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchSwitchRingBufferS
     if (disableMonitorFence) {
         TagData currentTagData = {};
         getTagAddressValue(currentTagData);
-        Dispatcher::dispatchMonitorFence(ringCommandStream, currentTagData.tagAddress, currentTagData.tagValue, *hwInfo, false, this->partitionedMode);
+        Dispatcher::dispatchMonitorFence(ringCommandStream, currentTagData.tagAddress, currentTagData.tagValue, *hwInfo, this->useNotifyForPostSync, this->partitionedMode);
     }
     Dispatcher::dispatchStartCommandBuffer(ringCommandStream, nextBufferGpuAddress);
 }
@@ -375,7 +376,7 @@ void *DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchWorkloadSection(BatchBu
     if (!disableMonitorFence) {
         TagData currentTagData = {};
         getTagAddressValue(currentTagData);
-        Dispatcher::dispatchMonitorFence(ringCommandStream, currentTagData.tagAddress, currentTagData.tagValue, *hwInfo, false, this->partitionedMode);
+        Dispatcher::dispatchMonitorFence(ringCommandStream, currentTagData.tagAddress, currentTagData.tagValue, *hwInfo, this->useNotifyForPostSync, this->partitionedMode);
     }
 
     dispatchSemaphoreSection(currentQueueWorkCount + 1);

shared/source/direct_submission/dispatchers/render_dispatcher.inl

@@ -35,6 +35,7 @@ inline void RenderDispatcher<GfxFamily>::dispatchMonitorFence(LinearStream &cmdB
     PipeControlArgs args;
     args.dcFlushEnable = MemorySynchronizationCommands<GfxFamily>::isDcFlushAllowed(true, hwInfo);
     args.workloadPartitionOffset = partitionedWorkload;
+    args.notifyEnable = useNotifyEnable;
     MemorySynchronizationCommands<GfxFamily>::addPipeControlAndProgramPostSyncOperation(
         cmdBuffer,
         POST_SYNC_OPERATION::POST_SYNC_OPERATION_WRITE_IMMEDIATE_DATA,

shared/test/common/mocks/mock_direct_submission_hw.h

@@ -63,6 +63,7 @@ struct MockDirectSubmissionHw : public DirectSubmissionHw<GfxFamily, Dispatcher>
     using BaseClass::startRingBuffer;
     using BaseClass::stopRingBuffer;
     using BaseClass::switchRingBuffersAllocations;
+    using BaseClass::useNotifyForPostSync;
     using BaseClass::workloadMode;
     using BaseClass::workloadModeOneExpectedValue;
     using BaseClass::workloadModeOneStoreAddress;

shared/test/common/mocks/windows/mock_wddm_direct_submission.h

@@ -35,6 +35,7 @@ struct MockWddmDirectSubmission : public WddmDirectSubmission<GfxFamily, Dispatc
     using BaseClass::submit;
     using BaseClass::switchRingBuffers;
     using BaseClass::updateTagValue;
+    using BaseClass::useNotifyForPostSync;
     using BaseClass::wddm;
     using BaseClass::WddmDirectSubmission;
     using typename BaseClass::RingBufferUse;