Capability to set low scheduling priority for Wddm context

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

opencl/test/unit_test/os_interface/windows/wddm20_tests.cpp

@@ -665,6 +665,62 @@ TEST_F(Wddm20WithMockGdiDllTestsWithoutWddmInit, givenUseNoRingFlushesKmdModeDeb
     EXPECT_FALSE(!!privateData->NoRingFlushes);
 }
 
+struct WddmContextSchedulingPriorityTests : public Wddm20WithMockGdiDllTestsWithoutWddmInit {
+    void initContext(bool lowPriority) {
+        auto preemptionMode = PreemptionHelper::getDefaultPreemptionMode(*defaultHwInfo);
+        wddmMockInterface = static_cast<WddmMockInterface20 *>(wddm->wddmInterface.release());
+        wddm->init();
+        wddm->wddmInterface.reset(wddmMockInterface);
+
+        auto hwInfo = rootDeviceEnvironment->getHardwareInfo();
+        auto engine = HwHelper::get(defaultHwInfo->platform.eRenderCoreFamily).getGpgpuEngineInstances(*hwInfo)[0];
+
+        auto engineDescriptor = EngineDescriptorHelper::getDefaultDescriptor(engine, preemptionMode);
+        engineDescriptor.engineTypeUsage.second = lowPriority ? EngineUsage::LowPriority : EngineUsage::Regular;
+        osContext = std::make_unique<OsContextWin>(*osInterface->getDriverModel()->as<Wddm>(), 0u, engineDescriptor);
+        osContext->ensureContextInitialized();
+    }
+};
+
+TEST_F(WddmContextSchedulingPriorityTests, givenLowPriorityContextWhenInitializingThenCallSetPriority) {
+    initContext(true);
+
+    auto createContextParams = this->getSetContextSchedulingPriorityDataCallFcn();
+
+    EXPECT_EQ(osContext->getWddmContextHandle(), createContextParams->hContext);
+    EXPECT_EQ(1, createContextParams->Priority);
+}
+
+TEST_F(WddmContextSchedulingPriorityTests, givenLowPriorityContextWhenFailingDuringSetSchedulingPriorityThenThrow) {
+    *this->getFailOnSetContextSchedulingPriorityCallFcn() = true;
+
+    EXPECT_ANY_THROW(initContext(true));
+}
+
+TEST_F(WddmContextSchedulingPriorityTests, givenDebugFlagSetWhenInitializingLowPriorityContextThenSetPriorityValue) {
+    DebugManagerStateRestore dbgRestore;
+
+    constexpr int32_t newPriority = 3;
+
+    DebugManager.flags.ForceWddmLowPriorityContextValue.set(newPriority);
+
+    initContext(true);
+
+    auto createContextParams = this->getSetContextSchedulingPriorityDataCallFcn();
+
+    EXPECT_EQ(osContext->getWddmContextHandle(), createContextParams->hContext);
+    EXPECT_EQ(newPriority, createContextParams->Priority);
+}
+
+TEST_F(WddmContextSchedulingPriorityTests, givenRegularContextWhenInitializingThenDontCallSetPriority) {
+    initContext(false);
+
+    auto createContextParams = this->getSetContextSchedulingPriorityDataCallFcn();
+
+    EXPECT_EQ(0, createContextParams->hContext);
+    EXPECT_EQ(0, createContextParams->Priority);
+}
+
 TEST_F(Wddm20WithMockGdiDllTestsWithoutWddmInit, givenCreateContextCallWhenDriverHintsThenItPointsToOpenCL) {
     init();
     auto createContextParams = this->getCreateContextDataFcn();

opencl/test/unit_test/test_files/igdrcl.config

@@ -419,3 +419,4 @@ UseContextEndOffsetForEventCompletion = -1
 DirectSubmissionInsertExtraMiMemFenceCommands = -1
 DirectSubmissionInsertSfenceInstructionPriorToSubmission = -1
 EnableTimestampWaitForEvents = -1
+ForceWddmLowPriorityContextValue = -1

shared/source/debug_settings/debug_variables_base.inl

@@ -202,6 +202,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, BatchBufferStartPrepatchingWaEnabled, -1, "-1: d
 DECLARE_DEBUG_VARIABLE(int32_t, SetVmAdviseAtomicAttribute, -1, "-1: default - atomic system, 0: atomic none, 1: atomic device, 2: atomic system)")
 DECLARE_DEBUG_VARIABLE(int32_t, ReadBackCommandBufferAllocation, -1, "Read command buffer allocation back on the host side. -1: default, 0 - disabled, 1 - local memory only, 2 - local and system memory")
 DECLARE_DEBUG_VARIABLE(int32_t, UseContextEndOffsetForEventCompletion, -1, "Use Context End or Context Start for event completion signalling. -1: default: platform dependent, 0 - Use Context Start, 1 - Use Context End")
+DECLARE_DEBUG_VARIABLE(int32_t, ForceWddmLowPriorityContextValue, -1, "Force scheduling priority value during Wddm low priority context creation. -1 - default.")
 DECLARE_DEBUG_VARIABLE(bool, DisableScratchPages, false, "Disable scratch pages during VM creations")
 /*LOGGING FLAGS*/
 DECLARE_DEBUG_VARIABLE(int32_t, PrintDriverDiagnostics, -1, "prints driver diagnostics messages to standard output, value corresponds to hint level")

shared/source/os_interface/windows/gdi_interface.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2021 Intel Corporation
+ * Copyright (C) 2018-2022 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -77,6 +77,7 @@ bool Gdi::getAllProcAddresses() {
     registerTrimNotification = gdiDll->getProcAddress("D3DKMTRegisterTrimNotification");
     unregisterTrimNotification = gdiDll->getProcAddress("D3DKMTUnregisterTrimNotification");
     setAllocationPriority = gdiDll->getProcAddress("D3DKMTSetAllocationPriority");
+    setSchedulingPriority = gdiDll->getProcAddress("D3DKMTSetContextSchedulingPriority");
 
     // For debug purposes
     getDeviceState = gdiDll->getProcAddress("D3DKMTGetDeviceState");
@@ -93,7 +94,7 @@ bool Gdi::getAllProcAddresses() {
         && signalSynchronizationObjectFromGpu && createPagingQueue && destroyPagingQueue
         && lock2 && unlock2 && mapGpuVirtualAddress && reserveGpuVirtualAddress
         && freeGpuVirtualAddress && updateGpuVirtualAddress &&submitCommand 
-        && makeResident && evict){
+        && makeResident && evict && setSchedulingPriority){
         if (NEO::OSInterface::requiresSupportForWddmTrimNotification) {
             if(registerTrimNotification && unregisterTrimNotification){
                 return true;

shared/source/os_interface/windows/gdi_interface.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2021 Intel Corporation
+ * Copyright (C) 2018-2022 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -62,6 +62,7 @@ class Gdi {
     ThkWrapper<IN D3DKMT_REGISTERTRIMNOTIFICATION *> registerTrimNotification{};
     ThkWrapper<IN D3DKMT_UNREGISTERTRIMNOTIFICATION *> unregisterTrimNotification{};
     ThkWrapper<IN CONST D3DKMT_SETALLOCATIONPRIORITY *> setAllocationPriority{};
+    ThkWrapper<IN CONST D3DKMT_SETCONTEXTSCHEDULINGPRIORITY *> setSchedulingPriority{};
 
     // HW queue
     ThkWrapper<IN OUT D3DKMT_CREATEHWQUEUE *> createHwQueue{};

shared/source/os_interface/windows/thk_wrapper.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2021 Intel Corporation
+ * Copyright (C) 2018-2022 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -71,6 +71,7 @@ GET_ID(D3DKMT_CREATEHWQUEUE *, SYSTIMER_ID_CREATEHWQUEUE)
 GET_ID(CONST D3DKMT_DESTROYHWQUEUE *, SYSTIMER_ID_DESTROYHWQUEUE)
 GET_ID(CONST D3DKMT_SUBMITCOMMANDTOHWQUEUE *, SYSTIMER_ID_SUBMITCOMMANDTOHWQUEUE)
 GET_ID(CONST D3DKMT_SETALLOCATIONPRIORITY *, SYSTIMER_ID_SETALLOCATIONPRIORITY)
+GET_ID(CONST D3DKMT_SETCONTEXTSCHEDULINGPRIORITY *, SYSTIMER_ID_SETCONTEXTSCHEDULINGPRIORITY)
 
 template <typename Param>
 class ThkWrapper {

shared/source/os_interface/windows/wddm/wddm.cpp

@@ -792,6 +792,25 @@ void Wddm::kmDafLock(D3DKMT_HANDLE handle) {
     kmDafListener->notifyLock(featureTable->flags.ftrKmdDaf, getAdapter(), device, handle, 0, getGdi()->escape);
 }
 
+bool Wddm::setLowPriorityContextParam(D3DKMT_HANDLE contextHandle) {
+    D3DKMT_SETCONTEXTSCHEDULINGPRIORITY contextPriority = {};
+
+    contextPriority.hContext = contextHandle;
+    contextPriority.Priority = 1;
+
+    if (DebugManager.flags.ForceWddmLowPriorityContextValue.get() != -1) {
+        contextPriority.Priority = static_cast<INT>(DebugManager.flags.ForceWddmLowPriorityContextValue.get());
+    }
+
+    auto status = getGdi()->setSchedulingPriority(&contextPriority);
+
+    PRINT_DEBUG_STRING(DebugManager.flags.PrintDebugMessages.get(), stdout,
+                       "\nSet scheduling priority for Wddm context. Status: :%lu, context handle: %u, priority: %d \n",
+                       status, contextHandle, contextPriority.Priority);
+
+    return (status == STATUS_SUCCESS);
+}
+
 bool Wddm::createContext(OsContextWin &osContext) {
     NTSTATUS status = STATUS_UNSUCCESSFUL;
     D3DKMT_CREATECONTEXTVIRTUAL CreateContext = {};
@@ -838,7 +857,15 @@ bool Wddm::createContext(OsContextWin &osContext) {
                        "\nCreated Wddm context. Status: :%lu, engine: %u, contextId: %u, deviceBitfield: %lu \n",
                        status, osContext.getEngineType(), osContext.getContextId(), osContext.getDeviceBitfield().to_ulong());
 
-    return status == STATUS_SUCCESS;
+    if (status != STATUS_SUCCESS) {
+        return false;
+    }
+
+    if (osContext.isLowPriority()) {
+        return setLowPriorityContextParam(osContext.getWddmContextHandle());
+    }
+
+    return true;
 }
 
 bool Wddm::destroyContext(D3DKMT_HANDLE context) {

shared/source/os_interface/windows/wddm/wddm.h

@@ -243,6 +243,7 @@ class Wddm : public DriverModel {
     bool destroyDevice();
     void getDeviceState();
     MOCKABLE_VIRTUAL void createPagingFenceLogger();
+    bool setLowPriorityContextParam(D3DKMT_HANDLE contextHandle);
 
     static GetSystemInfoFcn getSystemInfo;
 

shared/test/common/mock_gdi/gdi32_mock.def

@@ -54,6 +54,7 @@ D3DKMTCreateHwQueue
 D3DKMTDestroyHwQueue
 D3DKMTSubmitCommandToHwQueue
 D3DKMTSetAllocationPriority
+D3DKMTSetContextSchedulingPriority
 MockSetAdapterInfo
 MockSetSizes
 GetMockSizes
@@ -73,6 +74,8 @@ getSubmitCommandToHwQueueData
 getDestroySynchronizationObjectData
 getMonitorFenceCpuFenceAddress
 getCreateSynchronizationObject2FailCall
+getFailOnSetContextSchedulingPriorityCall
+getSetContextSchedulingPriorityDataCall
 getRegisterTrimNotificationFailCall
 getLastPriority
 setAdapterBDF

shared/test/common/mock_gdi/mock_gdi.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2021 Intel Corporation
+ * Copyright (C) 2018-2022 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -122,6 +122,19 @@ NTSTATUS __stdcall D3DKMTCreateContextVirtual(IN D3DKMT_CREATECONTEXTVIRTUAL *cr
     return STATUS_SUCCESS;
 }
 
+static bool failOnSetContextSchedulingPriority = false;
+static D3DKMT_SETCONTEXTSCHEDULINGPRIORITY setContextSchedulingPriorityData = {};
+
+NTSTATUS __stdcall D3DKMTSetContextSchedulingPriority(_In_ CONST D3DKMT_SETCONTEXTSCHEDULINGPRIORITY *setContextSchedulingPriority) {
+    setContextSchedulingPriorityData = *setContextSchedulingPriority;
+
+    if (failOnSetContextSchedulingPriority) {
+        return STATUS_INVALID_PARAMETER;
+    }
+
+    return STATUS_SUCCESS;
+}
+
 NTSTATUS __stdcall D3DKMTDestroyContext(IN CONST D3DKMT_DESTROYCONTEXT *destroyContext) {
     if (destroyContext == nullptr || destroyContext->hContext != CONTEXT_HANDLE) {
         return STATUS_INVALID_PARAMETER;
@@ -567,3 +580,11 @@ uint32_t getLastPriority() {
 void setAdapterBDF(ADAPTER_BDF &adapterBDF) {
     gAdapterBDF = adapterBDF;
 }
+
+bool *getFailOnSetContextSchedulingPriorityCall() {
+    return &failOnSetContextSchedulingPriority;
+}
+
+D3DKMT_SETCONTEXTSCHEDULINGPRIORITY *getSetContextSchedulingPriorityDataCall() {
+    return &setContextSchedulingPriorityData;
+}

shared/test/common/mock_gdi/mock_gdi.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2021 Intel Corporation
+ * Copyright (C) 2018-2022 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -78,6 +78,8 @@ D3DKMT_DESTROYSYNCHRONIZATIONOBJECT *getDestroySynchronizationObjectData();
 void InitGfxPartition();
 VOID *getMonitorFenceCpuFenceAddress();
 bool *getCreateSynchronizationObject2FailCall();
+bool *getFailOnSetContextSchedulingPriorityCall();
+D3DKMT_SETCONTEXTSCHEDULINGPRIORITY *getSetContextSchedulingPriorityDataCall();
 bool *getRegisterTrimNotificationFailCall();
 uint32_t getLastPriority();
 void setAdapterBDF(ADAPTER_BDF &adapterBDF);

shared/test/common/os_interface/windows/gdi_dll_fixture.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2021 Intel Corporation
+ * Copyright (C) 2018-2022 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -47,6 +47,10 @@ struct GdiDllFixture {
             reinterpret_cast<decltype(&getMonitorFenceCpuFenceAddress)>(mockGdiDll->getProcAddress("getMonitorFenceCpuFenceAddress"));
         getCreateSynchronizationObject2FailCallFcn =
             reinterpret_cast<decltype(&getCreateSynchronizationObject2FailCall)>(mockGdiDll->getProcAddress("getCreateSynchronizationObject2FailCall"));
+        getFailOnSetContextSchedulingPriorityCallFcn =
+            reinterpret_cast<decltype(&getFailOnSetContextSchedulingPriorityCall)>(mockGdiDll->getProcAddress("getFailOnSetContextSchedulingPriorityCall"));
+        getSetContextSchedulingPriorityDataCallFcn =
+            reinterpret_cast<decltype(&getSetContextSchedulingPriorityDataCall)>(mockGdiDll->getProcAddress("getSetContextSchedulingPriorityDataCall"));
         getRegisterTrimNotificationFailCallFcn =
             reinterpret_cast<decltype(&getRegisterTrimNotificationFailCall)>(mockGdiDll->getProcAddress("getRegisterTrimNotificationFailCall"));
         getLastPriorityFcn =
@@ -56,6 +60,8 @@ struct GdiDllFixture {
         setMockLastDestroyedResHandleFcn((D3DKMT_HANDLE)0);
         *getDestroySynchronizationObjectDataFcn() = {};
         *getCreateSynchronizationObject2FailCallFcn() = false;
+        *getFailOnSetContextSchedulingPriorityCallFcn() = false;
+        *getSetContextSchedulingPriorityDataCallFcn() = {};
         *getRegisterTrimNotificationFailCallFcn() = false;
     }
 
@@ -66,6 +72,8 @@ struct GdiDllFixture {
         *getDestroySynchronizationObjectDataFcn() = {};
         setMapGpuVaFailConfigFcn(0, 0);
         *getCreateSynchronizationObject2FailCallFcn() = false;
+        *getFailOnSetContextSchedulingPriorityCallFcn() = false;
+        *getSetContextSchedulingPriorityDataCallFcn() = {};
         *getRegisterTrimNotificationFailCallFcn() = false;
     }
 
@@ -89,6 +97,8 @@ struct GdiDllFixture {
     decltype(&getDestroySynchronizationObjectData) getDestroySynchronizationObjectDataFcn = nullptr;
     decltype(&getMonitorFenceCpuFenceAddress) getMonitorFenceCpuFenceAddressFcn = nullptr;
     decltype(&getCreateSynchronizationObject2FailCall) getCreateSynchronizationObject2FailCallFcn = nullptr;
+    decltype(&getFailOnSetContextSchedulingPriorityCall) getFailOnSetContextSchedulingPriorityCallFcn = nullptr;
+    decltype(&getSetContextSchedulingPriorityDataCall) getSetContextSchedulingPriorityDataCallFcn = nullptr;
     decltype(&getRegisterTrimNotificationFailCall) getRegisterTrimNotificationFailCallFcn = nullptr;
     decltype(&getLastPriority) getLastPriorityFcn = nullptr;
     decltype(&setAdapterBDF) setAdapterBDFFcn = nullptr;