Flush task at device init

Resolves: NEO-7642

Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -1349,9 +1349,13 @@ SubmissionStatus CommandStreamReceiverHw<GfxFamily>::flushPipeControl() {
     args.notifyEnable = isUsedNotifyEnableForPostSync();
     args.workloadPartitionOffset = isMultiTileOperationEnabled();
 
-    auto &commandStream = getCS(MemorySynchronizationCommands<GfxFamily>::getSizeForBarrierWithPostSyncOperation(hwInfo, args.tlbInvalidation));
+    auto dispatchSize = MemorySynchronizationCommands<GfxFamily>::getSizeForBarrierWithPostSyncOperation(hwInfo, args.tlbInvalidation) + this->getCmdSizeForPrologue();
+
+    auto &commandStream = getCS(dispatchSize);
     auto commandStreamStart = commandStream.getUsed();
 
+    this->programEnginePrologue(commandStream);
+
     MemorySynchronizationCommands<GfxFamily>::addBarrierWithPostSyncOperation(commandStream,
                                                                               PostSyncMode::ImmediateData,
                                                                               getTagAllocation()->getGpuAddress(),
@@ -1360,6 +1364,7 @@ SubmissionStatus CommandStreamReceiverHw<GfxFamily>::flushPipeControl() {
                                                                               args);
 
     makeResident(*tagAllocation);
+    makeResident(*commandStream.getGraphicsAllocation());
 
     auto submissionStatus = this->flushSmallTask(commandStream, commandStreamStart);
     this->latestFlushedTaskCount = taskCount.load();
@@ -1597,11 +1602,7 @@ void CommandStreamReceiverHw<GfxFamily>::createKernelArgsBufferAllocation() {
 
 template <typename GfxFamily>
 SubmissionStatus CommandStreamReceiverHw<GfxFamily>::initializeDeviceWithFirstSubmission() {
-    auto lock = obtainUniqueOwnership();
-
-    auto &commandStream = getCS(EncodeBatchBufferStartOrEnd<GfxFamily>::getBatchBufferEndSize());
-    auto commandStreamStart = commandStream.getUsed();
-    return this->flushSmallTask(commandStream, commandStreamStart);
+    return flushTagUpdate();
 }
 
 template <typename GfxFamily>

shared/source/device/device.cpp

@@ -381,13 +381,18 @@ bool Device::createEngine(uint32_t deviceCsrIndex, EngineTypeUsage engineTypeUsa
 
     commandStreamReceiver->createKernelArgsBufferAllocation();
 
+    if (preemptionMode == PreemptionMode::MidThread && !commandStreamReceiver->createPreemptionAllocation()) {
+        return false;
+    }
+
     if (isDefaultEngine) {
         bool defaultEngineAlreadySet = (allEngines.size() > defaultEngineIndex) && (allEngines[defaultEngineIndex].getEngineType() == engineType);
 
         if (!defaultEngineAlreadySet) {
             defaultEngineIndex = deviceCsrIndex;
 
-            if (osContext->isDebuggableContext()) {
+            if (osContext->isDebuggableContext() ||
+                this->isInitDeviceWithFirstSubmissionSupported(commandStreamReceiver->getType())) {
                 if (SubmissionStatus::SUCCESS != commandStreamReceiver->initializeDeviceWithFirstSubmission()) {
                     return false;
                 }
@@ -399,15 +404,12 @@ bool Device::createEngine(uint32_t deviceCsrIndex, EngineTypeUsage engineTypeUsa
         defaultBcsEngineIndex = deviceCsrIndex;
     }
 
-    if (preemptionMode == PreemptionMode::MidThread && !commandStreamReceiver->createPreemptionAllocation()) {
-        return false;
-    }
-
     EngineControl engine{commandStreamReceiver.get(), osContext};
     allEngines.push_back(engine);
     if (engineUsage == EngineUsage::Regular) {
         addEngineToEngineGroup(engine);
     }
+
     commandStreamReceivers.push_back(std::move(commandStreamReceiver));
 
     return true;
@@ -440,6 +442,11 @@ bool Device::isBcsSplitSupported() {
     return bcsSplit;
 }
 
+bool Device::isInitDeviceWithFirstSubmissionSupported(CommandStreamReceiverType csrType) {
+    return getProductHelper().isInitDeviceWithFirstSubmissionRequired(getHardwareInfo()) &&
+           Device::isInitDeviceWithFirstSubmissionEnabled(csrType);
+}
+
 double Device::getPlatformHostTimerResolution() const {
     if (getOSTime()) {
         return getOSTime()->getHostTimerResolution();

shared/source/device/device.h

@@ -10,6 +10,7 @@
 #include "shared/source/helpers/engine_control.h"
 #include "shared/source/helpers/engine_node_helper.h"
 #include "shared/source/helpers/non_copyable_or_moveable.h"
+#include "shared/source/helpers/options.h"
 #include "shared/source/os_interface/hw_info_config.h"
 #include "shared/source/os_interface/performance_counters.h"
 #include "shared/source/utilities/reference_tracked_object.h"
@@ -108,7 +109,9 @@ class Device : public ReferenceTrackedObject<Device> {
     RootDeviceEnvironment &getRootDeviceEnvironmentRef() const;
     bool isFullRangeSvm() const;
     static bool isBlitSplitEnabled();
+    static bool isInitDeviceWithFirstSubmissionEnabled(CommandStreamReceiverType csrType);
     bool isBcsSplitSupported();
+    bool isInitDeviceWithFirstSubmissionSupported(CommandStreamReceiverType csrType);
     bool areSharedSystemAllocationsAllowed() const;
     template <typename SpecializedDeviceT>
     void setSpecializedDevice(SpecializedDeviceT *specializedDevice) {

shared/source/dll/device_dll.cpp

@@ -13,4 +13,8 @@ bool Device::isBlitSplitEnabled() {
     return true;
 }
 
+bool Device::isInitDeviceWithFirstSubmissionEnabled(CommandStreamReceiverType csrType) {
+    return csrType == CommandStreamReceiverType::CSR_HW;
+}
+
 } // namespace NEO

shared/source/helpers/completion_stamp.h

@@ -7,6 +7,7 @@
 
 #pragma once
 
+#include "shared/source/command_stream/submission_status.h"
 #include "shared/source/command_stream/task_count_helper.h"
 
 #include <cstdint>
@@ -14,7 +15,6 @@
 
 namespace NEO {
 using FlushStamp = uint64_t;
-enum class SubmissionStatus : uint32_t;
 struct CompletionStamp {
     static TaskCountType getTaskCountFromSubmissionStatusError(SubmissionStatus submissionStatus);
 

shared/source/os_interface/hw_info_config.h

@@ -135,6 +135,7 @@ class ProductHelper {
     virtual bool allowMemoryPrefetch(const HardwareInfo &hwInfo) const = 0;
     virtual bool isBcsReportWaRequired(const HardwareInfo &hwInfo) const = 0;
     virtual bool isBlitSplitEnqueueWARequired(const HardwareInfo &hwInfo) const = 0;
+    virtual bool isInitDeviceWithFirstSubmissionRequired(const HardwareInfo &hwInfo) const = 0;
     virtual bool isBlitCopyRequiredForLocalMemory(const HardwareInfo &hwInfo, const GraphicsAllocation &allocation) const = 0;
     virtual bool isImplicitScalingSupported(const HardwareInfo &hwInfo) const = 0;
     virtual bool isCpuCopyNecessary(const void *ptr, MemoryManager *memoryManager) const = 0;
@@ -280,6 +281,7 @@ class ProductHelperHw : public ProductHelper {
     bool isTimestampWaitSupportedForEvents() const override;
     bool isTilePlacementResourceWaRequired(const HardwareInfo &hwInfo) const override;
     bool isBlitSplitEnqueueWARequired(const HardwareInfo &hwInfo) const override;
+    bool isInitDeviceWithFirstSubmissionRequired(const HardwareInfo &hwInfo) const override;
     bool allowMemoryPrefetch(const HardwareInfo &hwInfo) const override;
     bool isBcsReportWaRequired(const HardwareInfo &hwInfo) const override;
     bool isBlitCopyRequiredForLocalMemory(const HardwareInfo &hwInfo, const GraphicsAllocation &allocation) const override;

shared/source/os_interface/hw_info_config.inl

@@ -469,6 +469,11 @@ bool ProductHelperHw<gfxProduct>::isBlitSplitEnqueueWARequired(const HardwareInf
     return false;
 }
 
+template <PRODUCT_FAMILY gfxProduct>
+bool ProductHelperHw<gfxProduct>::isInitDeviceWithFirstSubmissionRequired(const HardwareInfo &hwInfo) const {
+    return false;
+}
+
 template <PRODUCT_FAMILY gfxProduct>
 bool ProductHelperHw<gfxProduct>::isBlitCopyRequiredForLocalMemory(const HardwareInfo &hwInfo, const GraphicsAllocation &allocation) const {
     return allocation.isAllocatedInLocalMemoryPool() &&

shared/source/xe_hpc_core/pvc/os_agnostic_hw_info_config_pvc.inl

@@ -173,6 +173,11 @@ bool ProductHelperHw<gfxProduct>::isBlitSplitEnqueueWARequired(const HardwareInf
     return true;
 }
 
+template <>
+bool ProductHelperHw<gfxProduct>::isInitDeviceWithFirstSubmissionRequired(const HardwareInfo &hwInfo) const {
+    return true;
+}
+
 template <>
 bool ProductHelperHw<gfxProduct>::isImplicitScalingSupported(const HardwareInfo &hwInfo) const {
     return getSteppingFromHwRevId(hwInfo) >= REVISION_B;

shared/test/common/base_ult_config_listener.cpp

@@ -36,7 +36,7 @@ void NEO::BaseUltConfigListener::OnTestEnd(const ::testing::TestInfo &) {
 
     // Ensure that global state is restored
     UltHwConfig expectedState{};
-    static_assert(sizeof(UltHwConfig) == 13 * sizeof(bool), ""); // Ensure that there is no internal padding
+    static_assert(sizeof(UltHwConfig) == 14 * sizeof(bool), ""); // Ensure that there is no internal padding
     EXPECT_EQ(0, memcmp(&expectedState, &ultHwConfig, sizeof(UltHwConfig)));
 
     EXPECT_EQ(0, memcmp(&referencedHwInfo.platform, &defaultHwInfo->platform, sizeof(PLATFORM)));

shared/test/common/helpers/ult_hw_config.h

@@ -14,6 +14,7 @@ struct UltHwConfig {
     bool forceOsAgnosticMemoryManager = true;
     bool useWaitForTimestamps = false;
     bool useBlitSplit = false;
+    bool useFirstSubmissionInitDevice = false;
 
     bool csrFailInitDirectSubmission = false;
     bool csrBaseCallDirectSubmissionAvailable = false;

shared/test/common/mocks/device_ult.cpp

@@ -7,11 +7,14 @@
 
 #include "shared/source/device/device.h"
 #include "shared/test/common/helpers/ult_hw_config.h"
-
 namespace NEO {
 
 bool Device::isBlitSplitEnabled() {
     return ultHwConfig.useBlitSplit;
 }
 
+bool Device::isInitDeviceWithFirstSubmissionEnabled(CommandStreamReceiverType csrType) {
+    return ultHwConfig.useFirstSubmissionInitDevice;
+}
+
 } // namespace NEO

shared/test/unit_test/device/neo_device_tests.cpp

@@ -673,6 +673,9 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, DeviceTests, givenZexNumberOfCssEnvVariableSetAmbig
 }
 
 HWCMDTEST_F(IGFX_XE_HP_CORE, DeviceTests, givenDebuggableOsContextWhenDeviceCreatesEnginesThenDeviceIsInitializedWithFirstSubmission) {
+    VariableBackup<UltHwConfig> backup(&ultHwConfig);
+    ultHwConfig.useFirstSubmissionInitDevice = true;
+
     auto hwInfo = *defaultHwInfo;
     hardwareInfoSetup[hwInfo.platform.eProductFamily](&hwInfo, true, 0);
 
@@ -687,7 +690,10 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, DeviceTests, givenDebuggableOsContextWhenDeviceCrea
     EXPECT_EQ(1u, csr->peekLatestSentTaskCount());
 }
 
-HWCMDTEST_F(IGFX_XE_HP_CORE, DeviceTests, givenNonDebuggableOsContextWhenDeviceCreatesEnginesThenDeviceIsNotInitializedWithFirstSubmission) {
+HWCMDTEST_F(IGFX_XE_HP_CORE, DeviceTests, whenDeviceCreatesEnginesThenDeviceIsInitializedWithFirstSubmission) {
+    VariableBackup<UltHwConfig> backup(&ultHwConfig);
+    ultHwConfig.useFirstSubmissionInitDevice = true;
+
     auto hwInfo = *defaultHwInfo;
     hardwareInfoSetup[hwInfo.platform.eProductFamily](&hwInfo, true, 0);
 
@@ -698,7 +704,7 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, DeviceTests, givenNonDebuggableOsContextWhenDeviceC
 
     auto device = deviceFactory.rootDevices[0];
     auto csr = device->allEngines[device->defaultEngineIndex].commandStreamReceiver;
-    EXPECT_EQ(0u, csr->peekLatestSentTaskCount());
+    EXPECT_EQ(device->isInitDeviceWithFirstSubmissionSupported(csr->getType()), csr->peekLatestSentTaskCount());
 }
 
 TEST(FailDeviceTest, GivenFailedDeviceWhenCreatingDeviceThenNullIsReturned) {