fix: clear all TS packets for combined CB event

Related-To: NEO-13971

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

Author: BartoszDunajski

level_zero/core/source/cmdlist/cmdlist_hw.inl

@@ -4188,7 +4188,7 @@ bool CommandListCoreFamily<gfxCoreFamily>::handleCounterBasedEventOperations(Eve
             auto tag = device->getInOrderTimestampAllocator()->getTag();
 
             this->commandContainer.addToResidencyContainer(tag->getBaseGraphicsAllocation()->getGraphicsAllocation(device->getRootDeviceIndex()));
-            signalEvent->resetInOrderTimestampNode(tag);
+            signalEvent->resetInOrderTimestampNode(tag, this->partitionCount);
         }
     }
 

level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl

@@ -314,7 +314,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(K
             if (!compactEvent || this->asMutable() || !compactEvent->isCounterBased() || compactEvent->isUsingContextEndOffset()) {
                 if (inOrderNonWalkerSignalling) {
                     if (!eventForInOrderExec->getAllocation(this->device) && Event::standaloneInOrderTimestampAllocationEnabled()) {
-                        eventForInOrderExec->resetInOrderTimestampNode(device->getInOrderTimestampAllocator()->getTag());
+                        eventForInOrderExec->resetInOrderTimestampNode(device->getInOrderTimestampAllocator()->getTag(), this->partitionCount);
                     }
                     if (this->asMutable() || !eventForInOrderExec->isCounterBased()) {
                         dispatchEventPostSyncOperation(eventForInOrderExec, nullptr, launchParams.outListCommands, Event::STATE_CLEARED, false, false, false, false, false);

level_zero/core/source/event/event.cpp

@@ -516,7 +516,7 @@ void Event::releaseTempInOrderTimestampNodes() {
 }
 
 ze_result_t Event::destroy() {
-    resetInOrderTimestampNode(nullptr);
+    resetInOrderTimestampNode(nullptr, 0);
     releaseTempInOrderTimestampNodes();
 
     if (isCounterBasedExplicitlyEnabled() && isFromIpcPool) {
@@ -660,13 +660,13 @@ void Event::setReferenceTs(uint64_t currentCpuTimeStamp) {
 }
 
 void Event::unsetInOrderExecInfo() {
-    resetInOrderTimestampNode(nullptr);
+    resetInOrderTimestampNode(nullptr, 0);
     inOrderExecInfo.reset();
     inOrderAllocationOffset = 0;
     inOrderExecSignalValue = 0;
 }
 
-void Event::resetInOrderTimestampNode(NEO::TagNodeBase *newNode) {
+void Event::resetInOrderTimestampNode(NEO::TagNodeBase *newNode, uint32_t partitionCount) {
     bool removeExistingNodes = inOrderIncrementValue == 0 || !newNode;
 
     if (inOrderIncrementValue > 0 && inOrderTimestampNode.size() == inOrderIncrementOperationsCount) {
@@ -687,7 +687,7 @@ void Event::resetInOrderTimestampNode(NEO::TagNodeBase *newNode) {
         inOrderTimestampNode.push_back(newNode);
 
         if (NEO::debugManager.flags.ClearStandaloneInOrderTimestampAllocation.get() != 0) {
-            clearLatestInOrderTimestampData();
+            clearLatestInOrderTimestampData(partitionCount);
         }
     }
 }

level_zero/core/source/event/event.h

@@ -325,7 +325,7 @@ struct Event : _ze_event_handle_t {
 
     void setExternalInterruptId(uint32_t interruptId) { externalInterruptId = interruptId; }
 
-    void resetInOrderTimestampNode(NEO::TagNodeBase *newNode);
+    void resetInOrderTimestampNode(NEO::TagNodeBase *newNode, uint32_t partitionCount);
 
     bool hasInOrderTimestampNode() const { return !inOrderTimestampNode.empty(); }
 
@@ -345,7 +345,7 @@ struct Event : _ze_event_handle_t {
 
     void unsetCmdQueue();
     void releaseTempInOrderTimestampNodes();
-    virtual void clearLatestInOrderTimestampData() = 0;
+    virtual void clearLatestInOrderTimestampData(uint32_t partitionCount) = 0;
 
     EventPool *eventPool = nullptr;
 

level_zero/core/source/event/event_imp.h

@@ -72,7 +72,7 @@ struct EventImp : public Event {
     ze_result_t queryCounterBasedEventStatus();
     void handleSuccessfulHostSynchronization();
     MOCKABLE_VIRTUAL ze_result_t hostEventSetValueTimestamps(State eventState);
-    void clearLatestInOrderTimestampData() override;
+    void clearLatestInOrderTimestampData(uint32_t partitionCount) override;
     MOCKABLE_VIRTUAL void assignKernelEventCompletionData(void *address);
     void setRemainingPackets(TagSizeT eventVal, uint64_t nextPacketGpuVa, void *nextPacketAddress, uint32_t packetsAlreadySet);
     void getSynchronizedKernelTimestamps(ze_synchronized_timestamp_result_ext_t *pSynchronizedTimestampsBuffer,

level_zero/core/source/event/event_impl.inl

@@ -219,15 +219,15 @@ ze_result_t EventImp<TagSizeT>::calculateProfilingData() {
 }
 
 template <typename TagSizeT>
-void EventImp<TagSizeT>::clearLatestInOrderTimestampData() {
+void EventImp<TagSizeT>::clearLatestInOrderTimestampData(uint32_t partitionCount) {
     auto node = inOrderTimestampNode.back();
     auto hostAddress = node->getCpuBase();
     auto deviceAddress = node->getGpuAddress();
 
     const std::array<TagSizeT, 4> data = {Event::STATE_INITIAL, Event::STATE_INITIAL, Event::STATE_INITIAL, Event::STATE_INITIAL};
     constexpr size_t copySize = data.size() * sizeof(TagSizeT);
 
-    for (uint32_t i = 0; i < kernelEventCompletionData[0].getPacketsUsed(); i++) {
+    for (uint32_t i = 0; i < partitionCount; i++) {
         copyDataToEventAlloc(hostAddress, deviceAddress, copySize, data.data());
 
         hostAddress = ptrOffset(hostAddress, singlePacketSize);

level_zero/core/test/unit_tests/fixtures/in_order_cmd_list_fixture.h

@@ -50,7 +50,7 @@ struct InOrderCmdListFixture : public ::Test<ModuleFixture> {
         using EventImp<uint32_t>::inOrderTimestampNode;
 
         void makeCounterBasedInitiallyDisabled(MultiGraphicsAllocation &poolAllocation) {
-            resetInOrderTimestampNode(nullptr);
+            resetInOrderTimestampNode(nullptr, 0);
             counterBasedMode = CounterBasedMode::initiallyDisabled;
             resetCompletionStatus();
             counterBasedFlags = 0;
@@ -60,7 +60,7 @@ struct InOrderCmdListFixture : public ::Test<ModuleFixture> {
         }
 
         void makeCounterBasedImplicitlyDisabled(MultiGraphicsAllocation &poolAllocation) {
-            resetInOrderTimestampNode(nullptr);
+            resetInOrderTimestampNode(nullptr, 0);
             counterBasedMode = CounterBasedMode::implicitlyDisabled;
             resetCompletionStatus();
             counterBasedFlags = 0;

level_zero/core/test/unit_tests/mocks/mock_event.h

@@ -178,7 +178,7 @@ class MockEvent : public ::L0::Event {
     ze_result_t hostEventSetValue(State eventState) override {
         return ZE_RESULT_SUCCESS;
     }
-    void clearLatestInOrderTimestampData() override {}
+    void clearLatestInOrderTimestampData(uint32_t partitionCount) override {}
     uint32_t getPacketsUsedInLastKernel() override { return 1; }
     uint32_t getPacketsInUse() const override { return 1; }
     void resetPackets(bool resetAllPackets) override {}

level_zero/core/test/unit_tests/sources/cmdlist/test_in_order_cmdlist_2.cpp

@@ -2659,6 +2659,52 @@ HWTEST2_F(MultiTileSynchronizedDispatchTests, givenLimitedSyncDispatchWhenAppend
 
 using MultiTileInOrderCmdListTests = MultiTileInOrderCmdListFixture;
 
+HWTEST2_F(MultiTileInOrderCmdListTests, givenExternalSyncStorageWhenCallingAppendThenClearAllTsPackets, IsAtLeastXeHpcCore) {
+    using TagSizeT = typename FamilyType::TimestampPacketType;
+
+    uint64_t counterValue = 4;
+    uint64_t incValue = 2;
+
+    auto devAddress = reinterpret_cast<uint64_t *>(allocDeviceMem(sizeof(uint64_t)));
+    auto eventObj = createExternalSyncStorageEvent(counterValue, incValue, devAddress);
+    eventObj->isTimestampEvent = true;
+    eventObj->setSinglePacketSize(NEO::TimestampPackets<TagSizeT, 1>::getSinglePacketSize());
+
+    auto handle = eventObj->toHandle();
+
+    auto immCmdList = createMultiTileImmCmdList<gfxCoreFamily>();
+
+    auto tag = device->getInOrderTimestampAllocator()->getTag();
+    auto node = static_cast<NEO::TimestampPackets<TagSizeT, 1> *>(tag->getCpuBase());
+
+    constexpr TagSizeT packet00[4] = {12, 13, 14, 15};
+    constexpr TagSizeT packet01[4] = {16, 17, 18, 19};
+
+    node->assignDataToAllTimestamps(0, packet00);
+    node->assignDataToAllTimestamps(1, packet01);
+
+    tag->returnTag();
+
+    immCmdList->appendLaunchKernel(kernel->toHandle(), groupCount, handle, 0, nullptr, launchParams, false);
+    ASSERT_EQ(1u, eventObj->inOrderTimestampNode.size());
+
+    auto node2 = static_cast<NEO::TimestampPackets<TagSizeT, 1> *>(eventObj->inOrderTimestampNode[0]->getCpuBase());
+
+    EXPECT_EQ(node, node2);
+
+    auto expectedValue = static_cast<uint64_t>(Event::STATE_INITIAL);
+    EXPECT_EQ(expectedValue, node2->getContextStartValue(0));
+    EXPECT_EQ(expectedValue, node2->getGlobalStartValue(0));
+    EXPECT_EQ(expectedValue, node2->getContextEndValue(0));
+    EXPECT_EQ(expectedValue, node2->getGlobalEndValue(0));
+    EXPECT_EQ(expectedValue, node2->getContextStartValue(1));
+    EXPECT_EQ(expectedValue, node2->getGlobalStartValue(1));
+    EXPECT_EQ(expectedValue, node2->getContextEndValue(1));
+    EXPECT_EQ(expectedValue, node2->getGlobalEndValue(1));
+
+    context->freeMem(devAddress);
+}
+
 HWTEST2_F(MultiTileInOrderCmdListTests, givenStandaloneEventWhenCallingAppendThenSuccess, IsAtLeastXeHpCore) {
     uint64_t counterValue = 2;
     auto hostAddress = reinterpret_cast<uint64_t *>(allocHostMem(sizeof(uint64_t)));

shared/source/helpers/timestamp_packet.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2024 Intel Corporation
+ * Copyright (C) 2018-2025 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -48,7 +48,7 @@ class TimestampPackets : public TagTypeBase {
         }
     }
 
-    void assignDataToAllTimestamps(uint32_t packetIndex, void *source) {
+    void assignDataToAllTimestamps(uint32_t packetIndex, const void *source) {
         memcpy_s(&packets[packetIndex], sizeof(Packet), source, sizeof(Packet));
     }