Add clearAllDependencies parameter to obtainNewTimestampPacketNodes

The capability to clear Timestamp packet dependencies from command queue.

Related-To: NEO-2747

Change-Id: Id3812539a47b96e23d0b8b17b9b8f54878ee2ef2
Signed-off-by: Milczarek, Slawomir <[email protected]>

Author: smilczar

runtime/command_queue/command_queue.cpp

@@ -551,11 +551,11 @@ void CommandQueue::dispatchAuxTranslation(MultiDispatchInfo &multiDispatchInfo,
     multiDispatchInfo.rbegin()->setPipeControlRequired(true);
 }
 
-void CommandQueue::obtainNewTimestampPacketNodes(size_t numberOfNodes, TimestampPacketContainer &previousNodes) {
+void CommandQueue::obtainNewTimestampPacketNodes(size_t numberOfNodes, TimestampPacketContainer &previousNodes, bool clearAllDependencies) {
     auto allocator = getCommandStreamReceiver().getTimestampPacketAllocator();
 
     previousNodes.swapNodes(*timestampPacketContainer);
-    previousNodes.resolveDependencies(isOOQEnabled() || DebugManager.flags.OmitTimestampPacketDependencies.get());
+    previousNodes.resolveDependencies(clearAllDependencies);
 
     DEBUG_BREAK_IF(timestampPacketContainer->peekNodes().size() > 0);
 

runtime/command_queue/command_queue.h

@@ -444,7 +444,7 @@ class CommandQueue : public BaseObject<_cl_command_queue> {
     MOCKABLE_VIRTUAL void dispatchAuxTranslation(MultiDispatchInfo &multiDispatchInfo, MemObjsForAuxTranslation &memObjsForAuxTranslation,
                                                  AuxTranslationDirection auxTranslationDirection);
 
-    void obtainNewTimestampPacketNodes(size_t numberOfNodes, TimestampPacketContainer &previousNodes);
+    void obtainNewTimestampPacketNodes(size_t numberOfNodes, TimestampPacketContainer &previousNodes, bool clearAllDependencies);
     void processProperties(const cl_queue_properties *properties);
 
     Context *context = nullptr;

runtime/command_queue/enqueue_common.h

@@ -139,6 +139,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
     Kernel *parentKernel = multiDispatchInfo.peekParentKernel();
     auto devQueue = this->getContext().getDefaultDeviceQueue();
     DeviceQueueHw<GfxFamily> *devQueueHw = castToObject<DeviceQueueHw<GfxFamily>>(devQueue);
+    auto clearAllDependencies = isOOQEnabled() || DebugManager.flags.OmitTimestampPacketDependencies.get();
 
     TagNode<HwTimeStamps> *hwTimeStamps = nullptr;
 
@@ -205,11 +206,9 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
     if (getCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
         csrDeps.fillFromEventsRequestAndMakeResident(eventsRequest, getCommandStreamReceiver(), CsrDependencies::DependenciesType::OnCsr);
 
-        if (!multiDispatchInfo.empty()) {
-            obtainNewTimestampPacketNodes(estimateTimestampPacketNodesCount(multiDispatchInfo), previousTimestampPacketNodes);
-            csrDeps.push_back(&previousTimestampPacketNodes);
-        } else if (isCacheFlushCommand(commandType)) {
-            obtainNewTimestampPacketNodes(1, previousTimestampPacketNodes);
+        auto nodesCount = !multiDispatchInfo.empty() ? estimateTimestampPacketNodesCount(multiDispatchInfo) : isCacheFlushCommand(commandType) ? 1u : 0u;
+        if (nodesCount > 0) {
+            obtainNewTimestampPacketNodes(nodesCount, previousTimestampPacketNodes, clearAllDependencies);
             csrDeps.push_back(&previousTimestampPacketNodes);
         }
     }

unit_tests/helpers/timestamp_packet_tests.cpp

@@ -958,7 +958,7 @@ HWTEST_F(TimestampPacketTests, givenAlreadyAssignedNodeWhenEnqueueingNonBlockedT
 
     auto cmdQ = std::make_unique<MockCommandQueueHw<FamilyType>>(context, device.get(), nullptr);
     TimestampPacketContainer previousNodes;
-    cmdQ->obtainNewTimestampPacketNodes(1, previousNodes);
+    cmdQ->obtainNewTimestampPacketNodes(1, previousNodes, false);
     auto firstNode = cmdQ->timestampPacketContainer->peekNodes().at(0);
 
     csr.storeMakeResidentAllocations = true;
@@ -980,7 +980,7 @@ HWTEST_F(TimestampPacketTests, givenAlreadyAssignedNodeWhenEnqueueingBlockedThen
 
     auto cmdQ = clUniquePtr(new MockCommandQueueHw<FamilyType>(context, device.get(), nullptr));
     TimestampPacketContainer previousNodes;
-    cmdQ->obtainNewTimestampPacketNodes(1, previousNodes);
+    cmdQ->obtainNewTimestampPacketNodes(1, previousNodes, false);
     auto firstNode = cmdQ->timestampPacketContainer->peekNodes().at(0);
 
     csr.storeMakeResidentAllocations = true;
@@ -1003,7 +1003,7 @@ HWTEST_F(TimestampPacketTests, givenAlreadyAssignedNodeWhenEnqueueingThenDontKee
 
     MockCommandQueueHw<FamilyType> cmdQ(context, device.get(), nullptr);
     TimestampPacketContainer previousNodes;
-    cmdQ.obtainNewTimestampPacketNodes(1, previousNodes);
+    cmdQ.obtainNewTimestampPacketNodes(1, previousNodes, false);
     auto firstNode = cmdQ.timestampPacketContainer->peekNodes().at(0);
     setTagToReadyState(firstNode);
 
@@ -1034,7 +1034,7 @@ HWTEST_F(TimestampPacketTests, givenAlreadyAssignedNodeWhenEnqueueingThenKeepDep
 
     MockCommandQueueHw<FamilyType> cmdQ(context, device.get(), nullptr);
     TimestampPacketContainer previousNodes;
-    cmdQ.obtainNewTimestampPacketNodes(2, previousNodes);
+    cmdQ.obtainNewTimestampPacketNodes(2, previousNodes, false);
     firstNode.add(cmdQ.timestampPacketContainer->peekNodes().at(0));
     firstNode.add(cmdQ.timestampPacketContainer->peekNodes().at(1));
     auto firstTag0 = firstNode.getNode(0);
@@ -1067,7 +1067,7 @@ HWTEST_F(TimestampPacketTests, givenAlreadyAssignedNodeWhenEnqueueingToOoqThenDo
     cl_queue_properties properties[] = {CL_QUEUE_PROPERTIES, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, 0};
     MockCommandQueueHw<FamilyType> cmdQ(context, device.get(), properties);
     TimestampPacketContainer previousNodes;
-    cmdQ.obtainNewTimestampPacketNodes(1, previousNodes);
+    cmdQ.obtainNewTimestampPacketNodes(1, previousNodes, false);
 
     cmdQ.enqueueKernel(kernel->mockKernel, 1, nullptr, gws, nullptr, 0, nullptr, nullptr);
 
@@ -1097,7 +1097,7 @@ HWTEST_F(TimestampPacketTests, givenAlreadyAssignedNodeWhenEnqueueingWithOmitTim
 
     MockCommandQueueHw<FamilyType> cmdQ(context, device.get(), nullptr);
     TimestampPacketContainer previousNodes;
-    cmdQ.obtainNewTimestampPacketNodes(1, previousNodes);
+    cmdQ.obtainNewTimestampPacketNodes(1, previousNodes, false);
 
     cmdQ.enqueueKernel(kernel->mockKernel, 1, nullptr, gws, nullptr, 0, nullptr, nullptr);
 
@@ -1250,7 +1250,7 @@ TEST_F(TimestampPacketTests, givenDispatchSizeWhenAskingForNewTimestampsThenObta
     EXPECT_EQ(0u, mockCmdQ->timestampPacketContainer->peekNodes().size());
 
     TimestampPacketContainer previousNodes;
-    mockCmdQ->obtainNewTimestampPacketNodes(dispatchSize, previousNodes);
+    mockCmdQ->obtainNewTimestampPacketNodes(dispatchSize, previousNodes, false);
     EXPECT_EQ(dispatchSize, mockCmdQ->timestampPacketContainer->peekNodes().size());
 }