Dont add PIPE_CONTROL between Walkers for TimestampPacket writes

Change-Id: I6216a0926678c7f7a997cb37ce4eb59f914b518e
Signed-off-by: Dunajski, Bartosz <bartosz.dunajski@intel.com>

Author: BartoszDunajski

runtime/command_stream/command_stream_receiver_hw.inl

@@ -175,7 +175,8 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
 
         epiloguePipeControlLocation = ptrOffset(commandStreamTask.getCpuBase(), commandStreamTask.getUsed());
 
-        if (dispatchFlags.outOfOrderExecutionAllowed && !dispatchFlags.dcFlush) {
+        if ((dispatchFlags.outOfOrderExecutionAllowed || DebugManager.flags.EnableTimestampPacket.get()) &&
+            !dispatchFlags.dcFlush) {
             currentPipeControlForNooping = epiloguePipeControlLocation;
         }
 
@@ -348,7 +349,9 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
 
     // Add a PC if we have a dependency on a previous walker to avoid concurrency issues.
     if (taskLevel > this->taskLevel) {
-        addPipeControl(commandStreamCSR, false);
+        if (!DebugManager.flags.EnableTimestampPacket.get()) {
+            addPipeControl(commandStreamCSR, false);
+        }
         this->taskLevel = taskLevel;
         DBG_LOG(LogTaskCounts, __FUNCTION__, "Line: ", __LINE__, "this->taskCount", this->taskCount);
     }
@@ -615,7 +618,8 @@ size_t CommandStreamReceiverHw<GfxFamily>::getRequiredCmdStreamSizeAligned(const
     return alignUp(size, MemoryConstants::cacheLineSize);
 }
 
-template <typename GfxFamily> size_t CommandStreamReceiverHw<GfxFamily>::getRequiredStateBaseAddressSize() const {
+template <typename GfxFamily>
+size_t CommandStreamReceiverHw<GfxFamily>::getRequiredStateBaseAddressSize() const {
     return sizeof(typename GfxFamily::STATE_BASE_ADDRESS) + sizeof(PIPE_CONTROL);
 }
 

unit_tests/command_stream/command_stream_receiver_flush_task_tests.cpp

@@ -297,6 +297,23 @@ HWTEST_F(CommandStreamReceiverFlushTaskTests, higherTaskLevelShouldSendAPipeCont
     EXPECT_NE(cmdList.end(), itorPC);
 }
 
+HWTEST_F(CommandStreamReceiverFlushTaskTests, givenHigherTaskLevelWhenDebugVariableIsEnabledThenDontAddPipeControl) {
+    DebugManagerStateRestore restore;
+    DebugManager.flags.EnableTimestampPacket.set(true);
+    auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
+    commandStreamReceiver.isPreambleSent = true;
+    configureCSRtoNonDirtyState<FamilyType>();
+    commandStreamReceiver.taskLevel = taskLevel;
+    taskLevel++; // submit with higher taskLevel
+
+    flushTask(commandStreamReceiver);
+
+    parseCommands<FamilyType>(commandStreamReceiver.commandStream, 0);
+
+    auto itorPC = find<typename FamilyType::PIPE_CONTROL *>(cmdList.begin(), cmdList.end());
+    EXPECT_EQ(cmdList.end(), itorPC);
+}
+
 HWTEST_F(CommandStreamReceiverFlushTaskTests, whenSamplerCacheFlushNotRequiredThenDontSendPipecontrol) {
     auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
     OCLRT::WorkaroundTable *waTable = nullptr;
@@ -2902,6 +2919,46 @@ HWCMDTEST_F(IGFX_GEN8_CORE, CommandStreamReceiverFlushTaskTests, givenCsrInBatch
     EXPECT_NE(itorPipeControl, itorBatchBufferStartSecond);
 }
 
+HWTEST_F(CommandStreamReceiverFlushTaskTests, givenCsrInBatchingModeAndOoqFlagSetToFalseWhenDebugVariableIsSetThenNoopPipeControl) {
+    DebugManagerStateRestore restore;
+    CommandQueueHw<FamilyType> commandQueue(nullptr, pDevice, 0);
+    auto &commandStream = commandQueue.getCS(4096u);
+
+    auto mockCsr = new MockCsrHw2<FamilyType>(*platformDevices[0], *pDevice->executionEnvironment);
+    pDevice->resetCommandStreamReceiver(mockCsr);
+
+    mockCsr->overrideDispatchPolicy(DispatchMode::BatchedDispatch);
+
+    auto mockedSubmissionsAggregator = new mockSubmissionsAggregator();
+    mockCsr->overrideSubmissionAggregator(mockedSubmissionsAggregator);
+
+    DispatchFlags dispatchFlags;
+    dispatchFlags.guardCommandBufferWithPipeControl = true;
+    dispatchFlags.outOfOrderExecutionAllowed = false;
+
+    auto taskLevelPriorToSubmission = mockCsr->peekTaskLevel();
+
+    DebugManager.flags.EnableTimestampPacket.set(false);
+    mockCsr->flushTask(commandStream, 0, dsh, ioh, ssh, taskLevelPriorToSubmission, dispatchFlags, *pDevice);
+    mockCsr->flushTask(commandStream, 0, dsh, ioh, ssh, taskLevelPriorToSubmission, dispatchFlags, *pDevice);
+
+    auto firstCmdBuffer = mockedSubmissionsAggregator->peekCommandBuffers().peekHead();
+    EXPECT_EQ(nullptr, firstCmdBuffer->pipeControlThatMayBeErasedLocation);
+    auto secondCmdBuffer = firstCmdBuffer->next;
+    EXPECT_EQ(nullptr, secondCmdBuffer->pipeControlThatMayBeErasedLocation);
+
+    mockCsr->flushBatchedSubmissions();
+
+    DebugManager.flags.EnableTimestampPacket.set(true);
+    mockCsr->flushTask(commandStream, 0, dsh, ioh, ssh, taskLevelPriorToSubmission, dispatchFlags, *pDevice);
+    mockCsr->flushTask(commandStream, 0, dsh, ioh, ssh, taskLevelPriorToSubmission, dispatchFlags, *pDevice);
+
+    firstCmdBuffer = mockedSubmissionsAggregator->peekCommandBuffers().peekHead();
+    EXPECT_NE(nullptr, firstCmdBuffer->pipeControlThatMayBeErasedLocation);
+    secondCmdBuffer = firstCmdBuffer->next;
+    EXPECT_NE(nullptr, secondCmdBuffer->pipeControlThatMayBeErasedLocation);
+}
+
 HWCMDTEST_F(IGFX_GEN8_CORE, CommandStreamReceiverFlushTaskTests, givenCsrInBatchingModeWhenPipeControlForNoopAddressIsNullThenPipeControlIsNotNooped) {
     CommandQueueHw<FamilyType> commandQueue(nullptr, pDevice, 0);
     auto &commandStream = commandQueue.getCS(4096u);