fix: use lowest priority level as default for primary context

kcencele · Compute-Runtime-Automation · commit 5cf3eeb6154a · 2025-11-04T11:42:35.000+01:00
calculate the lowest and highest software priority in one place

Related-To: HSD-18043767497
Signed-off-by: Katarzyna Cencelewska &lt;katarzyna.cencelewska@intel.com&gt;
diff --git a/level_zero/core/source/device/device_imp.cpp b/level_zero/core/source/device/device_imp.cpp
@@ -1586,11 +1586,8 @@ Device *Device::create(DriverHandle *driverHandle, NEO::Device *neoDevice, bool
     device->populateSubDeviceCopyEngineGroups();
     auto &productHelper = device->getProductHelper();
     device->calculationForDisablingEuFusionWithDpasNeeded = productHelper.isCalculationForDisablingEuFusionWithDpasNeeded(hwInfo);
-
-    auto numPriorities = static_cast<int32_t>(device->getNEODevice()->getGfxCoreHelper().getQueuePriorityLevels());
-
-    device->queuePriorityHigh = -(numPriorities + 1) / 2 + 1;
-    device->queuePriorityLow = (numPriorities) / 2;
+    device->queuePriorityHigh = gfxCoreHelper.getHighestQueuePriorityLevel();
+    device->queuePriorityLow = gfxCoreHelper.getLowestQueuePriorityLevel();
 
     return device;
 }
@@ -1839,7 +1836,7 @@ ze_result_t DeviceImp::getCsrForOrdinalAndIndex(NEO::CommandStreamReceiver **csr
     if (priorityLevel.has_value()) {
         if (priorityLevel.value() <= 0) {
             priority = ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH;
-        } else if (priorityLevel.value() > 0) {
+        } else {
             priority = ZE_COMMAND_QUEUE_PRIORITY_NORMAL;
         }
     }
diff --git a/level_zero/core/test/unit_tests/mocks/mock_device.h b/level_zero/core/test/unit_tests/mocks/mock_device.h
@@ -112,8 +112,6 @@ struct MockDeviceImp : public L0::DeviceImp {
     using Base::inOrderTimestampAllocator;
     using Base::neoDevice;
     using Base::queryPeerAccess;
-    using Base::queuePriorityHigh;
-    using Base::queuePriorityLow;
     using Base::subDeviceCopyEngineGroups;
     using Base::syncDispatchTokenAllocation;
 
diff --git a/level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_6.cpp b/level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_6.cpp
@@ -3406,11 +3406,11 @@ HWTEST2_F(ContextGroupStateBaseAddressGlobalStatelessTest,
     NEO::EngineTypeUsage engineTypeUsage;
     engineTypeUsage.first = hwInfo.capabilityTable.defaultEngineType;
     engineTypeUsage.second = NEO::EngineUsage::regular;
-    auto primaryCsr = neoDevice->getSecondaryEngineCsr(engineTypeUsage, 0, false)->commandStreamReceiver;
+    auto primaryCsr = neoDevice->getSecondaryEngineCsr(engineTypeUsage, std::nullopt, false)->commandStreamReceiver;
     EXPECT_EQ(nullptr, primaryCsr->getOsContext().getPrimaryContext());
     EXPECT_TRUE(primaryCsr->getOsContext().isPartOfContextGroup());
 
-    auto secondaryCsr = neoDevice->getSecondaryEngineCsr(engineTypeUsage, 0, false)->commandStreamReceiver;
+    auto secondaryCsr = neoDevice->getSecondaryEngineCsr(engineTypeUsage, std::nullopt, false)->commandStreamReceiver;
 
     ze_command_queue_desc_t desc = {};
     auto otherCommandQueue = new MockCommandQueueHw<FamilyType::gfxCoreFamily>(device, secondaryCsr, &desc);
@@ -3457,11 +3457,11 @@ HWTEST2_F(ContextGroupStateBaseAddressGlobalStatelessTest,
     NEO::EngineTypeUsage engineTypeUsage;
     engineTypeUsage.first = hwInfo.capabilityTable.defaultEngineType;
     engineTypeUsage.second = NEO::EngineUsage::regular;
-    auto primaryCsr = neoDevice->getSecondaryEngineCsr(engineTypeUsage, 0, false)->commandStreamReceiver;
+    auto primaryCsr = neoDevice->getSecondaryEngineCsr(engineTypeUsage, std::nullopt, false)->commandStreamReceiver;
     EXPECT_EQ(nullptr, primaryCsr->getOsContext().getPrimaryContext());
     EXPECT_TRUE(primaryCsr->getOsContext().isPartOfContextGroup());
 
-    [[maybe_unused]] auto secondaryCsr = neoDevice->getSecondaryEngineCsr(engineTypeUsage, 0, false)->commandStreamReceiver;
+    [[maybe_unused]] auto secondaryCsr = neoDevice->getSecondaryEngineCsr(engineTypeUsage, std::nullopt, false)->commandStreamReceiver;
 
     ze_command_queue_desc_t queueDesc{ZE_STRUCTURE_TYPE_COMMAND_QUEUE_DESC};
     queueDesc.ordinal = 0u;
diff --git a/level_zero/core/test/unit_tests/sources/device/test_l0_device.cpp b/level_zero/core/test/unit_tests/sources/device/test_l0_device.cpp
@@ -4226,6 +4226,7 @@ HWTEST_F(DeviceTest, givenContextGroupSupportedWhenGettingHighPriorityCsrThenCor
     {
         MockExecutionEnvironment *executionEnvironment = new MockExecutionEnvironment{&hwInfo};
         auto *neoMockDevice = NEO::MockDevice::createWithExecutionEnvironment<NEO::MockDevice>(&hwInfo, executionEnvironment, rootDeviceIndex);
+        auto highestPriorityLevel = mockExecutionEnvironment.rootDeviceEnvironments[rootDeviceIndex]->getHelper<GfxCoreHelper>().getHighestQueuePriorityLevel();
         MockDeviceImp deviceImp(neoMockDevice);
 
         NEO::CommandStreamReceiver *highPriorityCsr = nullptr;
@@ -4249,7 +4250,7 @@ HWTEST_F(DeviceTest, givenContextGroupSupportedWhenGettingHighPriorityCsrThenCor
         ASSERT_TRUE(engineGroups[ordinalCopy].engineGroupType == NEO::EngineGroupType::copy);
 
         uint32_t index = 1;
-        auto result = deviceImp.getCsrForOrdinalAndIndex(&highPriorityCsr, ordinal, index, ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH, deviceImp.queuePriorityHigh, false);
+        auto result = deviceImp.getCsrForOrdinalAndIndex(&highPriorityCsr, ordinal, index, ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH, highestPriorityLevel, false);
         EXPECT_EQ(ZE_RESULT_SUCCESS, result);
         ASSERT_NE(nullptr, highPriorityCsr);
 
@@ -4266,7 +4267,7 @@ HWTEST_F(DeviceTest, givenContextGroupSupportedWhenGettingHighPriorityCsrThenCor
         EXPECT_TRUE(highPriorityCsr->getOsContext().isPartOfContextGroup());
         EXPECT_NE(nullptr, highPriorityCsr->getOsContext().getPrimaryContext());
 
-        result = deviceImp.getCsrForOrdinalAndIndex(&highPriorityCsr2, ordinal, index, ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH, deviceImp.queuePriorityHigh, false);
+        result = deviceImp.getCsrForOrdinalAndIndex(&highPriorityCsr2, ordinal, index, ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH, highestPriorityLevel, false);
         EXPECT_EQ(ZE_RESULT_SUCCESS, result);
         ASSERT_NE(nullptr, highPriorityCsr2);
         EXPECT_NE(highPriorityCsr, highPriorityCsr2);
@@ -4275,19 +4276,19 @@ HWTEST_F(DeviceTest, givenContextGroupSupportedWhenGettingHighPriorityCsrThenCor
         EXPECT_TRUE(highPriorityCsr2->getOsContext().isPartOfContextGroup());
 
         index = 100;
-        result = deviceImp.getCsrForOrdinalAndIndex(&highPriorityCsr, ordinal, index, ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH, deviceImp.queuePriorityHigh, false);
+        result = deviceImp.getCsrForOrdinalAndIndex(&highPriorityCsr, ordinal, index, ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH, highestPriorityLevel, false);
         EXPECT_EQ(ZE_RESULT_ERROR_INVALID_ARGUMENT, result);
 
         index = 0;
         ordinal = 100;
-        result = deviceImp.getCsrForOrdinalAndIndex(&highPriorityCsr, ordinal, index, ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH, deviceImp.queuePriorityHigh, false);
+        result = deviceImp.getCsrForOrdinalAndIndex(&highPriorityCsr, ordinal, index, ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH, highestPriorityLevel, false);
         EXPECT_EQ(ZE_RESULT_ERROR_INVALID_ARGUMENT, result);
 
         // When no HP copy engine, then hp csr from group is returned
         NEO::CommandStreamReceiver *bcsEngine = nullptr, *bcsEngine2 = nullptr;
         EXPECT_EQ(nullptr, neoMockDevice->getHpCopyEngine());
 
-        result = deviceImp.getCsrForOrdinalAndIndex(&bcsEngine, ordinalCopy, index, ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH, deviceImp.queuePriorityHigh, false);
+        result = deviceImp.getCsrForOrdinalAndIndex(&bcsEngine, ordinalCopy, index, ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH, highestPriorityLevel, false);
         EXPECT_EQ(ZE_RESULT_SUCCESS, result);
         ASSERT_NE(nullptr, bcsEngine);
         EXPECT_TRUE(bcsEngine->getOsContext().isHighPriority());
@@ -4395,7 +4396,7 @@ HWTEST2_F(DeviceTest, givenHpCopyEngineWhenGettingHighPriorityCsrThenCorrectCsrA
                 ordinal = i;
 
                 uint32_t index = 0;
-                auto result = deviceImp.getCsrForOrdinalAndIndex(&highPriorityCsr, ordinal, index, ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH, 0, false);
+                auto result = deviceImp.getCsrForOrdinalAndIndex(&highPriorityCsr, ordinal, index, ZE_COMMAND_QUEUE_PRIORITY_PRIORITY_HIGH, std::nullopt, false);
                 EXPECT_EQ(ZE_RESULT_SUCCESS, result);
                 ASSERT_NE(nullptr, highPriorityCsr);
 
diff --git a/opencl/source/command_queue/command_queue.cpp b/opencl/source/command_queue/command_queue.cpp
@@ -178,7 +178,7 @@ CommandQueue::~CommandQueue() {
 }
 
 void tryAssignSecondaryEngine(Device &device, EngineControl *&engineControl, EngineTypeUsage engineTypeUsage) {
-    auto newEngine = device.getSecondaryEngineCsr(engineTypeUsage, 0, false);
+    auto newEngine = device.getSecondaryEngineCsr(engineTypeUsage, std::nullopt, false);
     if (newEngine) {
         engineControl = newEngine;
     }
diff --git a/shared/source/debug_settings/debug_variables_base.inl b/shared/source/debug_settings/debug_variables_base.inl
@@ -377,6 +377,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, PrintL0MetricLogs, 0, "L0 Metrics logs mask. 0 -
 DECLARE_DEBUG_VARIABLE(bool, PrintL0SetKernelArg, false, "Print L0 Set Kernel Arg data")
 DECLARE_DEBUG_VARIABLE(bool, LogIndirectDetectionKernelDetails, false, "Log information for indirect detection for each kernel")
 DECLARE_DEBUG_VARIABLE(bool, PrintMclData, false, "Print all parameters used for MCL");
+DECLARE_DEBUG_VARIABLE(bool, PrintSecondaryContextEngineInfo, false, "Print info about used secondary contexts engine");
 
 /*PERFORMANCE FLAGS*/
 DECLARE_DEBUG_VARIABLE(bool, DisableZeroCopyForBuffers, false, "When active all buffer allocations will not share memory with CPU.")
diff --git a/shared/source/device/device.cpp b/shared/source/device/device.cpp
@@ -393,10 +393,6 @@ bool Device::createEngines() {
 
             auto highPriorityContextCount = gfxCoreHelper.getContextGroupHpContextsCount(engineGroupType, hpEngineAvailable);
 
-            if (debugManager.flags.OverrideNumHighPriorityContexts.get() != -1) {
-                highPriorityContextCount = static_cast<uint32_t>(debugManager.flags.OverrideNumHighPriorityContexts.get());
-            }
-
             if (getRootDeviceEnvironment().osInterface && getRootDeviceEnvironment().osInterface->getAggregatedProcessCount() > 1) {
                 const auto numProcesses = getRootDeviceEnvironment().osInterface->getAggregatedProcessCount();
 
@@ -413,6 +409,10 @@ bool Device::createEngines() {
                     gfxCoreHelper.adjustCopyEngineRegularContextCount(engineGroup->engines.size(), contextCount);
                 }
             }
+
+            if (debugManager.flags.OverrideNumHighPriorityContexts.get() != -1) {
+                highPriorityContextCount = static_cast<uint32_t>(debugManager.flags.OverrideNumHighPriorityContexts.get());
+            }
             for (uint32_t engineIndex = 0; engineIndex < static_cast<uint32_t>(engineGroup->engines.size()); engineIndex++) {
                 auto engineType = engineGroup->engines[engineIndex].getEngineType();
 
@@ -565,7 +565,9 @@ bool Device::createEngine(EngineTypeUsage engineTypeUsage) {
     osContext->setContextGroupCount(useContextGroup ? gfxCoreHelper.getContextGroupContextsCount() : 0);
     osContext->setIsPrimaryEngine(isPrimaryEngine);
     osContext->setIsDefaultEngine(isDefaultEngine);
-
+    if (isPrimaryEngine) {
+        osContext->overridePriority(gfxCoreHelper.getLowestQueuePriorityLevel());
+    }
     DEBUG_BREAK_IF(getDeviceBitfield().count() > 1 && !osContext->isRootDevice());
 
     commandStreamReceiver->setupContext(*osContext);
@@ -1336,11 +1338,14 @@ EngineControl *SecondaryContexts::getEngine(EngineUsage usage, std::optional<int
     }
     if (priorityLevel.has_value()) {
         engines[secondaryEngineIndex].osContext->overridePriority(priorityLevel.value());
+    } else {
+        int32_t lowestPriorityLevel = engines[secondaryEngineIndex].commandStreamReceiver->getGfxCoreHelper().getLowestQueuePriorityLevel();
+        engines[secondaryEngineIndex].osContext->overridePriority(lowestPriorityLevel);
     }
+    PRINT_DEBUG_STRING(debugManager.flags.PrintSecondaryContextEngineInfo.get(), stdout, "SecondaryContexts::getEngine-> engineUsage: %d index %d priorityLevel: %d \n", EngineHelpers::engineUsageToString(usage).c_str(), secondaryEngineIndex, engines[secondaryEngineIndex].osContext->getPriorityLevel());
 
     return &engines[secondaryEngineIndex];
 }
-
 void Device::stopDirectSubmissionForCopyEngine() {
     auto internalBcsEngine = getInternalCopyEngine();
     if (internalBcsEngine == nullptr || getHardwareInfo().featureTable.ftrBcsInfo.count() > 1) {
diff --git a/shared/source/helpers/gfx_core_helper.cpp b/shared/source/helpers/gfx_core_helper.cpp
@@ -8,6 +8,7 @@
 #include "shared/source/helpers/gfx_core_helper.h"
 
 #include "shared/source/debug_settings/debug_settings_manager.h"
+#include "shared/source/helpers/basic_math.h"
 #include "shared/source/helpers/constants.h"
 #include "shared/source/helpers/hw_info.h"
 #include "shared/source/os_interface/product_helper.h"
@@ -153,4 +154,12 @@ bool GfxCoreHelper::isWorkaroundRequired(uint32_t lowestSteppingWithBug, uint32_
     return (lowestHwRevIdWithBug <= hwInfo.platform.usRevId && hwInfo.platform.usRevId < hwRevIdWithFix);
 }
 
+int32_t GfxCoreHelper::getHighestQueuePriorityLevel() const {
+    return -(static_cast<int32_t>(Math::divideAndRoundUp(getQueuePriorityLevels(), 2))) + 1;
+}
+
+int32_t GfxCoreHelper::getLowestQueuePriorityLevel() const {
+    return getQueuePriorityLevels() / 2;
+}
+
 } // namespace NEO
diff --git a/shared/source/helpers/gfx_core_helper.h b/shared/source/helpers/gfx_core_helper.h
@@ -218,6 +218,8 @@ class GfxCoreHelper {
 
     virtual bool isExtendedUsmPoolSizeEnabled() const = 0;
     virtual bool crossEngineCacheFlushRequired() const = 0;
+    int32_t getHighestQueuePriorityLevel() const;
+    int32_t getLowestQueuePriorityLevel() const;
 
     virtual ~GfxCoreHelper() = default;
 
diff --git a/shared/source/os_interface/os_context.h b/shared/source/os_interface/os_context.h
@@ -43,7 +43,10 @@ class OsContext : public ReferenceTrackedObject<OsContext> {
     }
 
     bool hasPriorityLevel() const { return priorityLevel.has_value(); }
-    int getPriorityLevel() const { return priorityLevel.value_or(0); }
+    int getPriorityLevel() const {
+        UNRECOVERABLE_IF(!priorityLevel.has_value());
+        return priorityLevel.value();
+    }
     bool isRegular() const { return engineUsage == EngineUsage::regular; }
     bool isLowPriority() const { return engineUsage == EngineUsage::lowPriority; }
     bool isHighPriority() const { return engineUsage == EngineUsage::highPriority; }
@@ -126,7 +129,7 @@ class OsContext : public ReferenceTrackedObject<OsContext> {
     const uint32_t numSupportedDevices;
     aub_stream::EngineType engineType = aub_stream::ENGINE_RCS;
     EngineUsage engineUsage;
-    std::optional<int> priorityLevel;
+    std::optional<int> priorityLevel = std::nullopt;
     const bool rootDevice = false;
     bool defaultContext = false;
     bool directSubmissionActive = false;
diff --git a/shared/test/common/test_files/igdrcl.config b/shared/test/common/test_files/igdrcl.config
@@ -674,6 +674,7 @@ EnableUsmAllocationPoolManager = -1
 ForceTotalWMTPDataSize = -1
 CopyLockedMemoryBeforeWrite = 0
 SplitBcsPerEngineMaxSize = -1
+PrintSecondaryContextEngineInfo = 0
 HostFunctionWorkMode = -1
 EnableUsmPoolResidencyTracking = -1
 # Please don't edit below this line
diff --git a/shared/test/unit_test/command_stream/command_stream_receiver_tests.cpp b/shared/test/unit_test/command_stream/command_stream_receiver_tests.cpp
@@ -5960,11 +5960,11 @@ HWTEST_F(CommandStreamReceiverContextGroupTest, givenSecondaryCsrWhenGettingInte
     primaryCsr->createGlobalStatelessHeap();
 
     for (uint32_t secondaryIndex = 1; secondaryIndex < secondaryEnginesCount; secondaryIndex++) {
-        device->getSecondaryEngineCsr({EngineHelpers::mapCcsIndexToEngineType(ccsIndex), EngineUsage::regular}, 0, false);
+        device->getSecondaryEngineCsr({EngineHelpers::mapCcsIndexToEngineType(ccsIndex), EngineUsage::regular}, std::nullopt, false);
     }
 
     for (uint32_t i = 0; i < secondaryEngines.highPriorityEnginesTotal; i++) {
-        device->getSecondaryEngineCsr({EngineHelpers::mapCcsIndexToEngineType(ccsIndex), EngineUsage::highPriority}, 0, false);
+        device->getSecondaryEngineCsr({EngineHelpers::mapCcsIndexToEngineType(ccsIndex), EngineUsage::highPriority}, std::nullopt, false);
     }
 
     for (uint32_t secondaryIndex = 0; secondaryIndex < secondaryEnginesCount; secondaryIndex++) {
@@ -6013,11 +6013,11 @@ HWTEST_F(CommandStreamReceiverContextGroupTest, givenSecondaryRootCsrWhenGetting
     primaryCsr->createGlobalStatelessHeap();
 
     for (uint32_t secondaryIndex = 1; secondaryIndex < secondaryEnginesCount; secondaryIndex++) {
-        device->getSecondaryEngineCsr({EngineHelpers::mapCcsIndexToEngineType(ccsIndex), EngineUsage::regular}, 0, false);
+        device->getSecondaryEngineCsr({EngineHelpers::mapCcsIndexToEngineType(ccsIndex), EngineUsage::regular}, std::nullopt, false);
     }
 
     for (uint32_t i = 0; i < secondaryEngines.highPriorityEnginesTotal; i++) {
-        device->getSecondaryEngineCsr({EngineHelpers::mapCcsIndexToEngineType(ccsIndex), EngineUsage::highPriority}, 0, false);
+        device->getSecondaryEngineCsr({EngineHelpers::mapCcsIndexToEngineType(ccsIndex), EngineUsage::highPriority}, std::nullopt, false);
     }
 
     for (uint32_t secondaryIndex = 0; secondaryIndex < secondaryEnginesCount; secondaryIndex++) {
@@ -6143,8 +6143,8 @@ HWTEST_F(CommandStreamReceiverContextGroupTest, givenSecondaryCsrsWhenSameResour
     auto device = std::unique_ptr<MockDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(&hwInfo));
 
     const auto ccsIndex = 0;
-    auto &commandStreamReceiver0 = *device->getSecondaryEngineCsr({EngineHelpers::mapCcsIndexToEngineType(ccsIndex), EngineUsage::regular}, 0, false)->commandStreamReceiver;
-    auto &commandStreamReceiver1 = *device->getSecondaryEngineCsr({EngineHelpers::mapCcsIndexToEngineType(ccsIndex), EngineUsage::regular}, 0, false)->commandStreamReceiver;
+    auto &commandStreamReceiver0 = *device->getSecondaryEngineCsr({EngineHelpers::mapCcsIndexToEngineType(ccsIndex), EngineUsage::regular}, std::nullopt, false)->commandStreamReceiver;
+    auto &commandStreamReceiver1 = *device->getSecondaryEngineCsr({EngineHelpers::mapCcsIndexToEngineType(ccsIndex), EngineUsage::regular}, std::nullopt, false)->commandStreamReceiver;
 
     auto csr0ContextId = commandStreamReceiver0.getOsContext().getContextId();
     auto csr1ContextId = commandStreamReceiver1.getOsContext().getContextId();
diff --git a/shared/test/unit_test/device/neo_device_tests.cpp b/shared/test/unit_test/device/neo_device_tests.cpp
diff --git a/shared/test/unit_test/os_interface/os_context_tests.cpp b/shared/test/unit_test/os_interface/os_context_tests.cpp

Original file line number	Diff line number	Diff line change
`@@ -178,7 +178,7 @@ CommandQueue::~CommandQueue() {`
`178`	`178`	`}`
`179`	`179`
`180`	`180`	`void tryAssignSecondaryEngine(Device &device, EngineControl *&engineControl, EngineTypeUsage engineTypeUsage) {`
`181`		`- auto newEngine = device.getSecondaryEngineCsr(engineTypeUsage, 0, false);`
	`181`	`+ auto newEngine = device.getSecondaryEngineCsr(engineTypeUsage, std::nullopt, false);`
`182`	`182`	`if (newEngine) {`
`183`	`183`	`engineControl = newEngine;`
`184`	`184`	`}`