fix: copy offload cross engine dependency handling

Related-To: HSD-18043670900

Signed-off-by: Bartosz Dunajski <[email protected]>

Author: BartoszDunajski

level_zero/core/source/cmdlist/cmdlist_hw.h

@@ -484,6 +484,7 @@ struct CommandListCoreFamily : public CommandListImp {
     bool implicitSynchronizedDispatchForCooperativeKernelsAllowed = false;
     bool useAdditionalBlitProperties = false;
     bool isPostImageWriteFlushRequired = false;
+    bool latestFlushIsDualCopyOffload = false;
 };
 
 template <PRODUCT_FAMILY gfxProductFamily>

level_zero/core/source/cmdlist/cmdlist_hw.inl

@@ -3166,7 +3166,8 @@ void CommandListCoreFamily<gfxCoreFamily>::appendWaitOnInOrderDependency(std::sh
                                                                                                    isQwordInOrderCounter(), copyOnlyWait);
 
         } else {
-            auto resolveDependenciesViaPipeControls = !copyOnlyWait && implicitDependency && (this->dcFlushSupport || (!this->heaplessModeEnabled && this->latestOperationHasOptimizedCbEvent));
+            bool crossEngineDependency = (latestFlushIsDualCopyOffload != dualStreamCopyOffloadOperation);
+            auto resolveDependenciesViaPipeControls = !crossEngineDependency && !copyOnlyWait && implicitDependency && (this->dcFlushSupport || (!this->heaplessModeEnabled && this->latestOperationHasOptimizedCbEvent));
 
             if (NEO::debugManager.flags.ResolveDependenciesViaPipeControls.get() != -1) {
                 resolveDependenciesViaPipeControls = NEO::debugManager.flags.ResolveDependenciesViaPipeControls.get();

level_zero/core/source/cmdlist/cmdlist_hw_immediate.h

@@ -268,7 +268,6 @@ struct CommandListCoreFamilyImmediate : public CommandListCoreFamily<gfxCoreFami
     uint64_t relaxedOrderingCounter = 0;
     std::atomic<bool> dependenciesPresent{false};
     bool latestFlushIsHostVisible = false;
-    bool latestFlushIsDualCopyOffload = false;
     bool keepRelaxedOrderingEnabled = false;
 };
 

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

@@ -9,6 +9,7 @@
 #include "shared/source/gmm_helper/gmm_helper.h"
 #include "shared/source/helpers/compiler_product_helper.h"
 #include "shared/source/memory_manager/internal_allocation_storage.h"
+#include "shared/test/common/cmd_parse/hw_parse.h"
 #include "shared/test/common/helpers/relaxed_ordering_commands_helper.h"
 #include "shared/test/common/libult/ult_command_stream_receiver.h"
 #include "shared/test/common/mocks/mock_direct_submission_hw.h"
@@ -1061,6 +1062,29 @@ HWTEST2_F(CopyOffloadInOrderTests, givenProfilingEventWithRelaxedOrderingWhenApp
     }
 }
 
+HWCMDTEST_F(IGFX_XE_HP_CORE, CopyOffloadInOrderTests, givenCrossEngineDependencyWhenComputeWorkSubmittedThenUseSemaphore) {
+    debugManager.flags.OverrideCopyOffloadMode.set(CopyOffloadModes::dualStream);
+
+    uint32_t counterOffset = 64;
+
+    auto immCmdList = createImmCmdListWithOffload<FamilyType::gfxCoreFamily>();
+    immCmdList->inOrderExecInfo->setAllocationOffset(counterOffset);
+
+    auto cmdStream = immCmdList->getCmdContainer().getCommandStream();
+
+    immCmdList->appendMemoryCopy(&copyData1, &copyData2, 1, nullptr, 0, nullptr, copyParams);
+    EXPECT_TRUE(immCmdList->latestFlushIsDualCopyOffload);
+
+    auto offset = cmdStream->getUsed();
+
+    immCmdList->appendLaunchKernel(kernel->toHandle(), groupCount, nullptr, 0, nullptr, launchParams);
+
+    auto hwCmds = HardwareParse::parseCommandBuffer<FamilyType>(*cmdStream, offset);
+
+    auto itor = find<typename FamilyType::MI_SEMAPHORE_WAIT *>(hwCmds.begin(), hwCmds.end());
+    EXPECT_NE(hwCmds.end(), itor);
+}
+
 HWTEST2_F(CopyOffloadInOrderTests, givenAtomicSignalingModeWhenUpdatingCounterThenUseCorrectHwCommands, IsAtLeastXe2HpgCore) {
     using MI_ATOMIC = typename FamilyType::MI_ATOMIC;
     using ATOMIC_OPCODES = typename FamilyType::MI_ATOMIC::ATOMIC_OPCODES;

shared/test/common/cmd_parse/hw_parse.h

@@ -218,6 +218,13 @@ struct HardwareParse : NEO::NonCopyableAndNonMovableClass {
         return FamilyType::Parse::getCommandName(cmd);
     }
 
+    template <typename FamilyType>
+    static GenCmdList parseCommandBuffer(const LinearStream &linearStream, size_t offset) {
+        GenCmdList cmds;
+        EXPECT_TRUE(CmdParse<FamilyType>::parseCommandBuffer(cmds, ptrOffset(linearStream.getCpuBase(), offset), linearStream.getUsed() - offset));
+        return cmds;
+    }
+
     // The starting point of parsing commandBuffers.  This is important
     // because as buffers get reused, we only want to parse the deltas.
     LinearStream *previousCS = nullptr;