Add EngineUsageHint debug variable

Related-To: NEO-6219

Signed-off-by: Filip Hazubski <filip.hazubski@intel.com>

opencl/source/command_queue/command_queue.cpp

@@ -905,6 +905,10 @@ void CommandQueue::processProperties(const cl_queue_properties *properties) {
                 const auto &engine = getDevice().getEngineGroups()[selectedQueueFamilyIndex].engines[selectedQueueIndex];
                 auto engineType = engine.getEngineType();
                 auto engineUsage = engine.getEngineUsage();
+                if ((DebugManager.flags.EngineUsageHint.get() != -1) &&
+                    (getDevice().tryGetEngine(engineType, static_cast<EngineUsage>(DebugManager.flags.EngineUsageHint.get())) != nullptr)) {
+                    engineUsage = static_cast<EngineUsage>(DebugManager.flags.EngineUsageHint.get());
+                }
                 this->overrideEngine(engineType, engineUsage);
                 this->queueCapabilities = getClDevice().getDeviceInfo().queueFamilyProperties[selectedQueueFamilyIndex].capabilities;
                 this->queueFamilyIndex = selectedQueueFamilyIndex;
@@ -929,7 +933,7 @@ void CommandQueue::overrideEngine(aub_stream::EngineType engineType, EngineUsage
         deferredTimestampPackets = std::make_unique<TimestampPacketContainer>();
         isCopyOnly = true;
     } else {
-        gpgpuEngine = &device->getEngine(engineType, EngineUsage::Regular);
+        gpgpuEngine = &device->getEngine(engineType, engineUsage);
     }
 }
 

opencl/test/unit_test/command_queue/command_queue_tests.cpp

@@ -858,6 +858,28 @@ HWTEST_F(CommandQueueTests, givenMultipleCommandQueuesWhenMarkerIsEmittedThenGra
     EXPECT_EQ(commandStreamGraphicsAllocation, commandStreamGraphicsAllocation2);
 }
 
+HWTEST_F(CommandQueueTests, givenEngineUsageHintSetWithInvalidValueWhenCreatingCommandQueueThenReturnSuccess) {
+    DebugManagerStateRestore restore;
+    DebugManager.flags.EngineUsageHint.set(static_cast<int32_t>(EngineUsage::EngineUsageCount));
+
+    auto pDevice = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(defaultHwInfo.get()));
+    MockContext context(pDevice.get());
+
+    cl_int retVal = CL_SUCCESS;
+    cl_queue_properties propertiesCooperativeQueue[] = {CL_QUEUE_FAMILY_INTEL, 0, CL_QUEUE_INDEX_INTEL, 0, 0};
+
+    auto pCmdQ = CommandQueue::create(
+        &context,
+        pDevice.get(),
+        propertiesCooperativeQueue,
+        false,
+        retVal);
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    EXPECT_NE(nullptr, pCmdQ);
+    EXPECT_EQ(EngineUsage::Regular, pCmdQ->getGpgpuEngine().getEngineUsage());
+    delete pCmdQ;
+}
+
 struct WaitForQueueCompletionTests : public ::testing::Test {
     template <typename Family>
     struct MyCmdQueue : public CommandQueueHw<Family> {

opencl/test/unit_test/command_queue/command_queue_tests_pvc_and_later.cpp

@@ -112,6 +112,39 @@ HWTEST2_F(CommandQueuePvcAndLaterTests, givenQueueWithMainBcsIsReleasedWhenNewQu
     EXPECT_EQ(aub_stream::EngineType::ENGINE_BCS2, queue2->getBcsCommandStreamReceiver(aub_stream::ENGINE_BCS2)->getOsContext().getEngineType());
 }
 
+HWTEST2_F(CommandQueuePvcAndLaterTests, givenCooperativeEngineUsageHintAndCcsWhenCreatingCommandQueueThenCreateQueueWithCooperativeEngine, IsAtLeastXeHpcCore) {
+    DebugManagerStateRestore restore;
+    DebugManager.flags.EngineUsageHint.set(static_cast<int32_t>(EngineUsage::Cooperative));
+
+    auto hwInfo = *defaultHwInfo;
+    hwInfo.featureTable.flags.ftrCCSNode = true;
+    hwInfo.gtSystemInfo.CCSInfo.NumberOfCCSEnabled = 4;
+    auto &hwHelper = NEO::HwHelper::get(hwInfo.platform.eRenderCoreFamily);
+    auto &hwInfoConfig = *NEO::HwInfoConfig::get(hwInfo.platform.eProductFamily);
+
+    uint32_t revisions[] = {REVISION_A0, REVISION_B};
+    for (auto &revision : revisions) {
+        auto hwRevId = hwInfoConfig.getHwRevIdFromStepping(revision, hwInfo);
+        hwInfo.platform.usRevId = hwRevId;
+        if (hwRevId == CommonConstants::invalidStepping ||
+            !hwHelper.isCooperativeEngineSupported(hwInfo)) {
+            continue;
+        }
+
+        auto pDevice = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(&hwInfo));
+        MockContext context(pDevice.get());
+        cl_queue_properties propertiesCooperativeQueue[] = {CL_QUEUE_FAMILY_INTEL, 0, CL_QUEUE_INDEX_INTEL, 0, 0};
+        propertiesCooperativeQueue[1] = pDevice->getDevice().getEngineGroupIndexFromEngineGroupType(EngineGroupType::Compute);
+
+        for (size_t i = 0; i < 4; i++) {
+            propertiesCooperativeQueue[3] = i;
+            auto pCommandQueue = std::make_unique<MockCommandQueueHw<FamilyType>>(&context, pDevice.get(), propertiesCooperativeQueue);
+            EXPECT_EQ(aub_stream::ENGINE_CCS + i, pCommandQueue->gpgpuEngine->osContext->getEngineType());
+            EXPECT_EQ(EngineUsage::Cooperative, pCommandQueue->gpgpuEngine->osContext->getEngineUsage());
+        }
+    }
+}
+
 struct BcsCsrSelectionCommandQueueTests : ::testing::Test {
     void SetUp() override {
         HardwareInfo hwInfo = *::defaultHwInfo;

opencl/test/unit_test/test_files/igdrcl.config

@@ -359,6 +359,7 @@ ProgramGlobalFenceAsKernelInstructionInEUKernel = -1
 DoNotReportTile1BscWaActive = -1
 ForceTile0PlacementForTile1ResourcesWaActive = -1
 ClosEnabled = -1
+EngineUsageHint = -1
 AddStatePrefetchCmdToMemoryPrefetchAPI = -1
 UpdateCrossThreadDataSize = 0
 ForceBcsEngineIndex = -1

shared/source/debug_settings/debug_variables_base.inl

@@ -176,6 +176,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, ProgramGlobalFenceAsKernelInstructionInEUKernel,
 DECLARE_DEBUG_VARIABLE(int32_t, DoNotReportTile1BscWaActive, -1, "-1: default, 0: WA Disabled, 1: WA enabled")
 DECLARE_DEBUG_VARIABLE(int32_t, ForceTile0PlacementForTile1ResourcesWaActive, -1, "-1: default, 0: WA Disabled, 1: WA enabled")
 DECLARE_DEBUG_VARIABLE(int32_t, ClosEnabled, -1, "-1: default, 0: disabled, 1: enabled. Enable CLOS based cache reservation")
+DECLARE_DEBUG_VARIABLE(int32_t, EngineUsageHint, -1, "-1: default, >=0: engine usage value to use when creating command queue on user selected engine")
 DECLARE_DEBUG_VARIABLE(int32_t, ForceBcsEngineIndex, -1, "-1: default, >=0 Copy Engine index")
 
 /*LOGGING FLAGS*/