Fix start and dispatch blitter commands in direct mode

Related-To: NEO-5010

Change-Id: I3d03ef39325adb2beba26a989906381f5eccc4ff
Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>

opencl/source/os_interface/linux/drm_command_stream.inl

@@ -77,6 +77,10 @@ bool DrmCommandStreamReceiver<GfxFamily>::flush(BatchBuffer &batchBuffer, Reside
         memoryOperationsInterface->makeResidentWithinOsContext(this->osContext, ArrayRef<GraphicsAllocation *>(&batchBuffer.commandBufferAllocation, 1), true);
         return this->directSubmission->dispatchCommandBuffer(batchBuffer, *this->flushStamp.get());
     }
+    if (this->blitterDirectSubmission.get()) {
+        memoryOperationsInterface->makeResidentWithinOsContext(this->osContext, ArrayRef<GraphicsAllocation *>(&batchBuffer.commandBufferAllocation, 1), true);
+        return this->blitterDirectSubmission->dispatchCommandBuffer(batchBuffer, *this->flushStamp.get());
+    }
 
     this->flushStamp->setStamp(bb->peekHandle());
     this->flushInternal(batchBuffer, allocationsForResidency);

opencl/source/os_interface/windows/wddm_device_command_stream.inl

@@ -77,6 +77,9 @@ bool WddmCommandStreamReceiver<GfxFamily>::flush(BatchBuffer &batchBuffer, Resid
     if (directSubmission.get()) {
         return directSubmission->dispatchCommandBuffer(batchBuffer, *(flushStamp.get()));
     }
+    if (blitterDirectSubmission.get()) {
+        return blitterDirectSubmission->dispatchCommandBuffer(batchBuffer, *(flushStamp.get()));
+    }
 
     COMMAND_BUFFER_HEADER *pHeader = reinterpret_cast<COMMAND_BUFFER_HEADER *>(commandBufferHeader);
     pHeader->RequiresCoherency = batchBuffer.requiresCoherency;

opencl/test/unit_test/command_stream/command_stream_receiver_hw_1_tests.cpp

@@ -222,8 +222,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenNoBlitterOverrideWhenBlitterNotSuppo
     properties.engineSupported = false;
     properties.submitOnInit = false;
     bool startOnInit = true;
-    EXPECT_FALSE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_BCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_FALSE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_BCS, startOnInit, startInContext));
     EXPECT_FALSE(startOnInit);
+    EXPECT_FALSE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenNoBlitterOverrideWhenBlitterSupportedThenExpectTrueReturned) {
@@ -233,8 +235,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenNoBlitterOverrideWhenBlitterSupporte
     properties.engineSupported = true;
     properties.submitOnInit = true;
     bool startOnInit = false;
-    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_BCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_BCS, startOnInit, startInContext));
     EXPECT_TRUE(startOnInit);
+    EXPECT_FALSE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenBlitterOverrideEnableWhenBlitterNotSupportedThenExpectTrueReturned) {
@@ -246,8 +250,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenBlitterOverrideEnableWhenBlitterNotS
     properties.engineSupported = false;
     properties.submitOnInit = false;
     bool startOnInit = false;
-    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_BCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_BCS, startOnInit, startInContext));
     EXPECT_TRUE(startOnInit);
+    EXPECT_TRUE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenBlitterOverrideEnableAndNoStartWhenBlitterNotSupportedThenExpectTrueReturnedStartOnInitSetToTrue) {
@@ -259,8 +265,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenBlitterOverrideEnableAndNoStartWhenB
     properties.engineSupported = false;
     properties.submitOnInit = true;
     bool startOnInit = true;
-    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_BCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_BCS, startOnInit, startInContext));
     EXPECT_FALSE(startOnInit);
+    EXPECT_TRUE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenBlitterOverrideDisableWhenBlitterSupportedThenExpectFalseReturned) {
@@ -272,8 +280,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenBlitterOverrideDisableWhenBlitterSup
     properties.engineSupported = true;
     properties.submitOnInit = false;
     bool startOnInit = true;
-    EXPECT_FALSE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_BCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_FALSE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_BCS, startOnInit, startInContext));
     EXPECT_FALSE(startOnInit);
+    EXPECT_FALSE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenNoRenderOverrideWhenRenderNotSupportedThenExpectFalseReturned) {
@@ -283,8 +293,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenNoRenderOverrideWhenRenderNotSupport
     properties.engineSupported = false;
     properties.submitOnInit = false;
     bool startOnInit = true;
-    EXPECT_FALSE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_RCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_FALSE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_RCS, startOnInit, startInContext));
     EXPECT_FALSE(startOnInit);
+    EXPECT_FALSE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenNoRenderOverrideWhenRenderSupportedThenExpectTrueReturned) {
@@ -294,8 +306,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenNoRenderOverrideWhenRenderSupportedT
     properties.engineSupported = true;
     properties.submitOnInit = true;
     bool startOnInit = false;
-    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_RCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_RCS, startOnInit, startInContext));
     EXPECT_TRUE(startOnInit);
+    EXPECT_FALSE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenRenderOverrideEnableWhenRenderNotSupportedThenExpectTrueReturned) {
@@ -307,8 +321,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenRenderOverrideEnableWhenRenderNotSup
     properties.engineSupported = false;
     properties.submitOnInit = false;
     bool startOnInit = false;
-    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_RCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_RCS, startOnInit, startInContext));
     EXPECT_TRUE(startOnInit);
+    EXPECT_TRUE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenRenderOverrideEnableAndNoStartWhenRenderNotSupportedThenExpectTrueReturnedAndStartOnInitSetFalse) {
@@ -320,8 +336,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenRenderOverrideEnableAndNoStartWhenRe
     properties.engineSupported = false;
     properties.submitOnInit = true;
     bool startOnInit = true;
-    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_RCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_RCS, startOnInit, startInContext));
     EXPECT_FALSE(startOnInit);
+    EXPECT_TRUE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenRenderOverrideDisableWhenRenderSupportedThenExpectFalseReturned) {
@@ -333,8 +351,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenRenderOverrideDisableWhenRenderSuppo
     properties.engineSupported = true;
     properties.submitOnInit = false;
     bool startOnInit = true;
-    EXPECT_FALSE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_RCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_FALSE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_RCS, startOnInit, startInContext));
     EXPECT_FALSE(startOnInit);
+    EXPECT_FALSE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenNoComputeOverrideWhenComputeNotSupportedThenExpectFalseReturned) {
@@ -344,8 +364,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenNoComputeOverrideWhenComputeNotSuppo
     properties.engineSupported = false;
     properties.submitOnInit = false;
     bool startOnInit = true;
-    EXPECT_FALSE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_CCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_FALSE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_CCS, startOnInit, startInContext));
     EXPECT_FALSE(startOnInit);
+    EXPECT_FALSE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenNoComputeOverrideWhenComputeSupportedThenExpectTrueReturned) {
@@ -355,8 +377,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenNoComputeOverrideWhenComputeSupporte
     properties.engineSupported = true;
     properties.submitOnInit = true;
     bool startOnInit = false;
-    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_CCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_CCS, startOnInit, startInContext));
     EXPECT_TRUE(startOnInit);
+    EXPECT_FALSE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenComputeOverrideEnableWhenComputeNotSupportedThenExpectTrueReturned) {
@@ -368,8 +392,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenComputeOverrideEnableWhenComputeNotS
     properties.engineSupported = false;
     properties.submitOnInit = false;
     bool startOnInit = false;
-    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_CCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_CCS, startOnInit, startInContext));
     EXPECT_TRUE(startOnInit);
+    EXPECT_TRUE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenComputeOverrideEnableAndNoStartWhenComputeNotSupportedThenExpectTrueReturnedAndStartOnInitSetToFalse) {
@@ -381,8 +407,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenComputeOverrideEnableAndNoStartWhenC
     properties.engineSupported = false;
     properties.submitOnInit = true;
     bool startOnInit = true;
-    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_CCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_TRUE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_CCS, startOnInit, startInContext));
     EXPECT_FALSE(startOnInit);
+    EXPECT_TRUE(startInContext);
 }
 
 HWTEST_F(UltCommandStreamReceiverTest, givenComputeOverrideDisableWhenComputeSupportedThenExpectFalseReturned) {
@@ -394,8 +422,10 @@ HWTEST_F(UltCommandStreamReceiverTest, givenComputeOverrideDisableWhenComputeSup
     properties.engineSupported = true;
     properties.submitOnInit = false;
     bool startOnInit = true;
-    EXPECT_FALSE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_CCS, startOnInit));
+    bool startInContext = false;
+    EXPECT_FALSE(commandStreamReceiver.checkDirectSubmissionSupportsEngine(properties, aub_stream::ENGINE_CCS, startOnInit, startInContext));
     EXPECT_FALSE(startOnInit);
+    EXPECT_FALSE(startInContext);
 }
 
 typedef UltCommandStreamReceiverTest CommandStreamReceiverFlushTests;
@@ -574,7 +604,7 @@ HWTEST_F(BcsTests, whenAskingForCmdSizeForMiFlushDwWithMemoryWriteThenReturnCorr
     EXPECT_EQ(miFlushDwSize + additionalSize, totalSize);
 }
 
-HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenExstimatingCommandsSizeThenCalculateForAllAttachedProperites) {
+HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenEstimatingCommandsSizeThenCalculateForAllAttachedProperites) {
     const auto max2DBlitSize = BlitterConstants::maxBlitWidth * BlitterConstants::maxBlitHeight;
     const uint32_t numberOfBlts = 3;
     const size_t bltSize = (3 * max2DBlitSize);
@@ -598,12 +628,41 @@ HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenExstimatingCommandsSizeThenCa
     expectedAlignedSize = alignUp(expectedAlignedSize, MemoryConstants::cacheLineSize);
 
     auto alignedEstimatedSize = BlitCommandsHelper<FamilyType>::estimateBlitCommandsSize(
-        blitPropertiesContainer, false, false, pClDevice->getRootDeviceEnvironment());
+        blitPropertiesContainer, false, false, false, pClDevice->getRootDeviceEnvironment());
 
     EXPECT_EQ(expectedAlignedSize, alignedEstimatedSize);
 }
 
-HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenExstimatingCommandsSizeForWriteReadBufferRectThenCalculateForAllAttachedProperites) {
+HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenDirectsubmissionEnabledEstimatingCommandsSizeThenCalculateForAllAttachedProperites) {
+    const auto max2DBlitSize = BlitterConstants::maxBlitWidth * BlitterConstants::maxBlitHeight;
+    const uint32_t numberOfBlts = 3;
+    const size_t bltSize = (3 * max2DBlitSize);
+    const uint32_t numberOfBlitOperations = 4;
+
+    auto baseSize = EncodeMiFlushDW<FamilyType>::getMiFlushDwCmdSizeForDataWrite() + sizeof(typename FamilyType::MI_BATCH_BUFFER_START);
+    constexpr size_t cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK);
+    auto expectedBlitInstructionsSize = cmdsSizePerBlit * numberOfBlts;
+
+    auto expectedAlignedSize = baseSize + MemorySynchronizationCommands<FamilyType>::getSizeForAdditonalSynchronization(pDevice->getHardwareInfo());
+
+    BlitPropertiesContainer blitPropertiesContainer;
+    for (uint32_t i = 0; i < numberOfBlitOperations; i++) {
+        BlitProperties blitProperties;
+        blitProperties.copySize = {bltSize, 1, 1};
+        blitPropertiesContainer.push_back(blitProperties);
+
+        expectedAlignedSize += expectedBlitInstructionsSize;
+    }
+
+    expectedAlignedSize = alignUp(expectedAlignedSize, MemoryConstants::cacheLineSize);
+
+    auto alignedEstimatedSize = BlitCommandsHelper<FamilyType>::estimateBlitCommandsSize(
+        blitPropertiesContainer, false, false, true, pClDevice->getRootDeviceEnvironment());
+
+    EXPECT_EQ(expectedAlignedSize, alignedEstimatedSize);
+}
+
+HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenEstimatingCommandsSizeForWriteReadBufferRectThenCalculateForAllAttachedProperites) {
     constexpr auto max2DBlitSize = BlitterConstants::maxBlitWidth * BlitterConstants::maxBlitHeight;
     const Vec3<size_t> bltSize = {(3 * max2DBlitSize), 4, 2};
     const size_t numberOfBlts = 3 * bltSize.y * bltSize.z;
@@ -627,7 +686,36 @@ HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenExstimatingCommandsSizeForWri
     expectedAlignedSize = alignUp(expectedAlignedSize, MemoryConstants::cacheLineSize);
 
     auto alignedEstimatedSize = BlitCommandsHelper<FamilyType>::estimateBlitCommandsSize(
-        blitPropertiesContainer, false, false, pClDevice->getRootDeviceEnvironment());
+        blitPropertiesContainer, false, false, false, pClDevice->getRootDeviceEnvironment());
+
+    EXPECT_EQ(expectedAlignedSize, alignedEstimatedSize);
+}
+
+HWTEST_F(BcsTests, givenBlitPropertiesContainerWhenDirectSubmissionEnabledEstimatingCommandsSizeForWriteReadBufferRectThenCalculateForAllAttachedProperites) {
+    constexpr auto max2DBlitSize = BlitterConstants::maxBlitWidth * BlitterConstants::maxBlitHeight;
+    const Vec3<size_t> bltSize = {(3 * max2DBlitSize), 4, 2};
+    const size_t numberOfBlts = 3 * bltSize.y * bltSize.z;
+    const size_t numberOfBlitOperations = 4 * bltSize.y * bltSize.z;
+    const size_t cmdsSizePerBlit = sizeof(typename FamilyType::XY_COPY_BLT) + sizeof(typename FamilyType::MI_ARB_CHECK);
+
+    auto baseSize = EncodeMiFlushDW<FamilyType>::getMiFlushDwCmdSizeForDataWrite() + sizeof(typename FamilyType::MI_BATCH_BUFFER_START);
+    auto expectedBlitInstructionsSize = cmdsSizePerBlit * numberOfBlts;
+
+    auto expectedAlignedSize = baseSize + MemorySynchronizationCommands<FamilyType>::getSizeForAdditonalSynchronization(pDevice->getHardwareInfo());
+
+    BlitPropertiesContainer blitPropertiesContainer;
+    for (uint32_t i = 0; i < numberOfBlitOperations; i++) {
+        BlitProperties blitProperties;
+        blitProperties.copySize = bltSize;
+        blitPropertiesContainer.push_back(blitProperties);
+
+        expectedAlignedSize += expectedBlitInstructionsSize;
+    }
+
+    expectedAlignedSize = alignUp(expectedAlignedSize, MemoryConstants::cacheLineSize);
+
+    auto alignedEstimatedSize = BlitCommandsHelper<FamilyType>::estimateBlitCommandsSize(
+        blitPropertiesContainer, false, false, true, pClDevice->getRootDeviceEnvironment());
 
     EXPECT_EQ(expectedAlignedSize, alignedEstimatedSize);
 }

opencl/test/unit_test/command_stream/command_stream_receiver_hw_2_tests.cpp

@@ -6,6 +6,9 @@
  */
 
 #include "shared/source/command_stream/scratch_space_controller_base.h"
+#include "shared/source/direct_submission/dispatchers/blitter_dispatcher.h"
+#include "shared/test/unit_test/helpers/ult_hw_config.h"
+#include "shared/test/unit_test/mocks/mock_direct_submission_hw.h"
 #include "shared/test/unit_test/utilities/base_object_utils.h"
 
 #include "opencl/source/event/user_event.h"
@@ -66,7 +69,7 @@ HWTEST_F(BcsTests, givenDebugCapabilityWhenEstimatingCommandSizeThenAddAllRequir
     blitPropertiesContainer.push_back(blitProperties);
 
     auto estimatedSize = BlitCommandsHelper<FamilyType>::estimateBlitCommandsSize(
-        blitPropertiesContainer, false, true, pClDevice->getRootDeviceEnvironment());
+        blitPropertiesContainer, false, true, false, pClDevice->getRootDeviceEnvironment());
 
     EXPECT_EQ(expectedSize, estimatedSize);
     EXPECT_FALSE(BlitCommandsHelper<FamilyType>::isCopyRegionPreferred(blitProperties.copySize, pClDevice->getRootDeviceEnvironment()));
@@ -1166,6 +1169,50 @@ HWTEST_F(BcsTests, givenInvalidBlitDirectionWhenConstructPropertiesThenException
     EXPECT_THROW(ClBlitProperties::constructProperties(static_cast<BlitterConstants::BlitDirection>(7), csr, {}), std::exception);
 }
 
+HWTEST_F(BcsTests, givenBlitterDirectSubmissionEnabledWhenProgrammingBlitterThenExpectRingBufferDispatched) {
+    using MI_BATCH_BUFFER_START = typename FamilyType::MI_BATCH_BUFFER_START;
+    using MI_SEMAPHORE_WAIT = typename FamilyType::MI_SEMAPHORE_WAIT;
+
+    VariableBackup<UltHwConfig> backup(&ultHwConfig);
+    ultHwConfig.csrBaseCallBlitterDirectSubmissionAvailable = true;
+
+    auto &csr = pDevice->getUltCommandStreamReceiver<FamilyType>();
+    using DirectSubmission = MockDirectSubmissionHw<FamilyType, BlitterDispatcher<FamilyType>>;
+
+    csr.blitterDirectSubmission = std::make_unique<DirectSubmission>(*pDevice, *csr.osContext);
+    csr.recordFlusheBatchBuffer = true;
+    DirectSubmission *directSubmission = reinterpret_cast<DirectSubmission *>(csr.blitterDirectSubmission.get());
+    bool initRet = directSubmission->initialize(true);
+    EXPECT_TRUE(initRet);
+
+    cl_int retVal = CL_SUCCESS;
+    auto buffer = clUniquePtr<Buffer>(Buffer::create(context.get(), CL_MEM_READ_WRITE, 1, nullptr, retVal));
+    void *hostPtr = reinterpret_cast<void *>(0x12340000);
+    size_t numberNodesPerContainer = 5;
+    auto graphicsAllocation = buffer->getGraphicsAllocation(pDevice->getRootDeviceIndex());
+
+    auto blitProperties = BlitProperties::constructPropertiesForReadWriteBuffer(BlitterConstants::BlitDirection::HostPtrToBuffer,
+                                                                                csr, graphicsAllocation, nullptr, hostPtr,
+                                                                                graphicsAllocation->getGpuAddress(), 0,
+                                                                                0, 0, {1, 1, 1}, 0, 0, 0, 0);
+
+    MockTimestampPacketContainer timestamp0(*csr.getTimestampPacketAllocator(), numberNodesPerContainer);
+    MockTimestampPacketContainer timestamp1(*csr.getTimestampPacketAllocator(), numberNodesPerContainer);
+    blitProperties.csrDependencies.push_back(&timestamp0);
+    blitProperties.csrDependencies.push_back(&timestamp1);
+
+    blitBuffer(&csr, blitProperties, true);
+
+    HardwareParse hwParser;
+    hwParser.parseCommands<FamilyType>(csr.commandStream, 0u);
+    ASSERT_NE(nullptr, csr.latestFlushedBatchBuffer.endCmdPtr);
+
+    MI_BATCH_BUFFER_START *bbStart = hwParser.getCommand<MI_BATCH_BUFFER_START>();
+    ASSERT_NE(nullptr, bbStart);
+    EXPECT_EQ(csr.latestFlushedBatchBuffer.endCmdPtr, bbStart);
+    EXPECT_EQ(0ull, bbStart->getBatchBufferStartAddressGraphicsaddress472());
+}
+
 struct MockScratchSpaceController : ScratchSpaceControllerBase {
     using ScratchSpaceControllerBase::privateScratchAllocation;
     using ScratchSpaceControllerBase::ScratchSpaceControllerBase;

opencl/test/unit_test/command_stream/command_stream_receiver_tests.cpp

@@ -322,8 +322,8 @@ HWTEST_F(CommandStreamReceiverTest, whenDirectSubmissionDisabledThenExpectNoFeat
     EXPECT_FALSE(csr.isBlitterDirectSubmissionEnabled());
 }
 
-struct InitDirectSubmissionTest : public ::testing::Test {
-    void SetUp() override {
+struct InitDirectSubmissionFixture {
+    void SetUp() {
         DebugManager.flags.EnableDirectSubmission.set(1);
         executionEnvironment = new MockExecutionEnvironment();
         DeviceFactory::prepareDeviceEnvironments(*executionEnvironment);
@@ -333,11 +333,15 @@ struct InitDirectSubmissionTest : public ::testing::Test {
         device.reset(new MockDevice(executionEnvironment, 0u));
     }
 
+    void TearDown() {}
+
     DebugManagerStateRestore restore;
     MockExecutionEnvironment *executionEnvironment;
     std::unique_ptr<MockDevice> device;
 };
 
+using InitDirectSubmissionTest = Test<InitDirectSubmissionFixture>;
+
 HWTEST_F(InitDirectSubmissionTest, whenDirectSubmissionEnabledOnRcsThenExpectFeatureAvailable) {
     auto csr = std::make_unique<CommandStreamReceiverHw<FamilyType>>(*device->executionEnvironment, device->getRootDeviceIndex());
     std::unique_ptr<OsContext> osContext(OsContext::create(device->getExecutionEnvironment()->rootDeviceEnvironments[0]->osInterface.get(),
@@ -559,6 +563,27 @@ HWTEST_F(InitDirectSubmissionTest, givenNonDefaultContextContextWhenDirectSubmis
     csr.reset();
 }
 
+HWTEST_F(InitDirectSubmissionTest, GivenBlitterOverrideEnabledWhenBlitterIsNonDefaultContextThenExpectDirectSubmissionStarted) {
+    DebugManager.flags.DirectSubmissionOverrideBlitterSupport.set(1);
+
+    auto csr = std::make_unique<CommandStreamReceiverHw<FamilyType>>(*device->executionEnvironment, device->getRootDeviceIndex());
+    std::unique_ptr<OsContext> osContext(OsContext::create(device->getExecutionEnvironment()->rootDeviceEnvironments[0]->osInterface.get(),
+                                                           0, device->getDeviceBitfield(), aub_stream::ENGINE_BCS, PreemptionMode::ThreadGroup,
+                                                           false, false, false));
+    osContext->setDefaultContext(false);
+
+    auto hwInfo = device->getRootDeviceEnvironment().getMutableHardwareInfo();
+    hwInfo->capabilityTable.directSubmissionEngines.data[aub_stream::ENGINE_BCS].engineSupported = false;
+    hwInfo->capabilityTable.directSubmissionEngines.data[aub_stream::ENGINE_BCS].useNonDefault = false;
+    hwInfo->capabilityTable.directSubmissionEngines.data[aub_stream::ENGINE_BCS].submitOnInit = false;
+
+    bool ret = csr->initDirectSubmission(*device, *osContext.get());
+    EXPECT_TRUE(ret);
+    EXPECT_FALSE(csr->isDirectSubmissionEnabled());
+    EXPECT_TRUE(csr->isBlitterDirectSubmissionEnabled());
+    EXPECT_TRUE(osContext->isDirectSubmissionActive());
+}
+
 HWTEST_F(CommandStreamReceiverTest, whenCsrIsCreatedThenUseTimestampPacketWriteIfPossible) {
     CommandStreamReceiverHw<FamilyType> csr(*pDevice->executionEnvironment, pDevice->getRootDeviceIndex());
     EXPECT_EQ(UnitTestHelper<FamilyType>::isTimestampPacketWriteSupported(), csr.peekTimestampPacketWriteEnabled());

opencl/test/unit_test/libult/ult_command_stream_receiver.h

@@ -29,8 +29,10 @@ class UltCommandStreamReceiver : public CommandStreamReceiverHw<GfxFamily>, publ
     using BaseClass = CommandStreamReceiverHw<GfxFamily>;
 
   public:
+    using BaseClass::blitterDirectSubmission;
     using BaseClass::checkPlatformSupportsGpuIdleImplicitFlush;
     using BaseClass::checkPlatformSupportsNewResourceImplicitFlush;
+    using BaseClass::directSubmission;
     using BaseClass::dshState;
     using BaseClass::getCmdSizeForPrologue;
     using BaseClass::getScratchPatchAddress;

opencl/test/unit_test/mocks/linux/mock_drm_command_stream_receiver.h

@@ -20,6 +20,7 @@ class TestedDrmCommandStreamReceiver : public DrmCommandStreamReceiver<GfxFamily
     using CommandStreamReceiver::useNewResourceImplicitFlush;
     using DrmCommandStreamReceiver<GfxFamily>::residency;
     using CommandStreamReceiverHw<GfxFamily>::directSubmission;
+    using CommandStreamReceiverHw<GfxFamily>::blitterDirectSubmission;
     using CommandStreamReceiverHw<GfxFamily>::CommandStreamReceiver::lastSentSliceCount;
 
     TestedDrmCommandStreamReceiver(gemCloseWorkerMode mode, ExecutionEnvironment &executionEnvironment)

opencl/test/unit_test/os_interface/linux/drm_command_stream_tests.cpp

@@ -1312,11 +1312,39 @@ struct DrmCommandStreamDirectSubmissionTest : public DrmCommandStreamEnhancedTes
     DebugManagerStateRestore restorer;
 };
 
+struct DrmCommandStreamBlitterDirectSubmissionTest : public DrmCommandStreamDirectSubmissionTest {
+    template <typename GfxFamily>
+    void SetUpT() {
+        DebugManager.flags.DirectSubmissionOverrideBlitterSupport.set(1u);
+        DebugManager.flags.DirectSubmissionOverrideRenderSupport.set(0u);
+        DebugManager.flags.DirectSubmissionOverrideComputeSupport.set(0u);
+
+        DrmCommandStreamDirectSubmissionTest::SetUpT<GfxFamily>();
+
+        osContext.reset(OsContext::create(device->getExecutionEnvironment()->rootDeviceEnvironments[0]->osInterface.get(),
+                                          0, device->getDeviceBitfield(), aub_stream::ENGINE_BCS, PreemptionMode::ThreadGroup,
+                                          false, false, false));
+        csr->initDirectSubmission(*device.get(), *osContext.get());
+    }
+
+    template <typename GfxFamily>
+    void TearDownT() {
+        DrmCommandStreamDirectSubmissionTest::TearDownT<GfxFamily>();
+    }
+
+    std::unique_ptr<OsContext> osContext;
+};
+
 template <typename GfxFamily>
 struct MockDrmDirectSubmission : public DrmDirectSubmission<GfxFamily, RenderDispatcher<GfxFamily>> {
     using DrmDirectSubmission<GfxFamily, RenderDispatcher<GfxFamily>>::currentTagData;
 };
 
+template <typename GfxFamily>
+struct MockDrmBlitterDirectSubmission : public DrmDirectSubmission<GfxFamily, BlitterDispatcher<GfxFamily>> {
+    using DrmDirectSubmission<GfxFamily, BlitterDispatcher<GfxFamily>>::currentTagData;
+};
+
 HWTEST_TEMPLATED_F(DrmCommandStreamDirectSubmissionTest, givenEnabledDirectSubmissionWhenFlushThenFlushStampIsNotUpdated) {
     auto &cs = csr->getCS();
     CommandStreamReceiverHw<FamilyType>::addBatchBufferEnd(cs, nullptr);
@@ -1331,6 +1359,7 @@ HWTEST_TEMPLATED_F(DrmCommandStreamDirectSubmissionTest, givenEnabledDirectSubmi
     EXPECT_EQ(csr->obtainCurrentFlushStamp(), flushStamp);
 
     auto directSubmission = static_cast<TestedDrmCommandStreamReceiver<FamilyType> *>(csr)->directSubmission.get();
+    ASSERT_NE(nullptr, directSubmission);
     static_cast<MockDrmDirectSubmission<FamilyType> *>(directSubmission)->currentTagData.tagValue = 0u;
 }
 
@@ -1348,9 +1377,50 @@ HWTEST_TEMPLATED_F(DrmCommandStreamDirectSubmissionTest, givenEnabledDirectSubmi
     EXPECT_EQ(memoryOperationsInterface->isResident(device.get(), *batchBuffer.commandBufferAllocation), MemoryOperationsStatus::SUCCESS);
 
     auto directSubmission = static_cast<TestedDrmCommandStreamReceiver<FamilyType> *>(csr)->directSubmission.get();
+    ASSERT_NE(nullptr, directSubmission);
     static_cast<MockDrmDirectSubmission<FamilyType> *>(directSubmission)->currentTagData.tagValue = 0u;
 }
 
+HWTEST_TEMPLATED_F(DrmCommandStreamBlitterDirectSubmissionTest, givenEnabledDirectSubmissionOnBlitterWhenFlushThenFlushStampIsNotUpdated) {
+    auto &cs = csr->getCS();
+    CommandStreamReceiverHw<FamilyType>::addBatchBufferEnd(cs, nullptr);
+    CommandStreamReceiverHw<FamilyType>::alignToCacheLine(cs);
+    BatchBuffer batchBuffer{cs.getGraphicsAllocation(), 4, 0, nullptr, false, false, QueueThrottle::MEDIUM, QueueSliceCount::defaultSliceCount, cs.getUsed(), &cs, nullptr};
+    uint8_t bbStart[64];
+    batchBuffer.endCmdPtr = &bbStart[0];
+
+    auto flushStamp = csr->obtainCurrentFlushStamp();
+    csr->flush(batchBuffer, csr->getResidencyAllocations());
+
+    EXPECT_EQ(csr->obtainCurrentFlushStamp(), flushStamp);
+
+    auto directSubmission = static_cast<TestedDrmCommandStreamReceiver<FamilyType> *>(csr)->blitterDirectSubmission.get();
+    ASSERT_NE(nullptr, directSubmission);
+    static_cast<MockDrmBlitterDirectSubmission<FamilyType> *>(directSubmission)->currentTagData.tagValue = 0u;
+
+    EXPECT_EQ(nullptr, static_cast<TestedDrmCommandStreamReceiver<FamilyType> *>(csr)->directSubmission.get());
+}
+
+HWTEST_TEMPLATED_F(DrmCommandStreamBlitterDirectSubmissionTest, givenEnabledDirectSubmissionOnBlitterWhenFlushThenCommandBufferAllocationIsResident) {
+    auto &cs = csr->getCS();
+    CommandStreamReceiverHw<FamilyType>::addBatchBufferEnd(cs, nullptr);
+    CommandStreamReceiverHw<FamilyType>::alignToCacheLine(cs);
+    BatchBuffer batchBuffer{cs.getGraphicsAllocation(), 4, 0, nullptr, false, false, QueueThrottle::MEDIUM, QueueSliceCount::defaultSliceCount, cs.getUsed(), &cs, nullptr};
+    uint8_t bbStart[64];
+    batchBuffer.endCmdPtr = &bbStart[0];
+
+    csr->flush(batchBuffer, csr->getResidencyAllocations());
+
+    auto memoryOperationsInterface = executionEnvironment->rootDeviceEnvironments[rootDeviceIndex]->memoryOperationsInterface.get();
+    EXPECT_EQ(memoryOperationsInterface->isResident(device.get(), *batchBuffer.commandBufferAllocation), MemoryOperationsStatus::SUCCESS);
+
+    auto directSubmission = static_cast<TestedDrmCommandStreamReceiver<FamilyType> *>(csr)->blitterDirectSubmission.get();
+    ASSERT_NE(nullptr, directSubmission);
+    static_cast<MockDrmBlitterDirectSubmission<FamilyType> *>(directSubmission)->currentTagData.tagValue = 0u;
+
+    EXPECT_EQ(nullptr, static_cast<TestedDrmCommandStreamReceiver<FamilyType> *>(csr)->directSubmission.get());
+}
+
 HWTEST_TEMPLATED_F(DrmCommandStreamEnhancedTest, CheckDrmFree) {
     auto &cs = csr->getCS();
     auto commandBuffer = static_cast<DrmAllocation *>(cs.getGraphicsAllocation());

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

@@ -123,10 +123,16 @@ struct MockWddmCsr : public WddmCommandStreamReceiver<GfxFamily> {
         }
         bool ret = true;
         if (DebugManager.flags.EnableDirectSubmission.get() == 1) {
-            directSubmission = std::make_unique<
-                MockWddmDirectSubmission<GfxFamily, RenderDispatcher<GfxFamily>>>(device, osContext);
-            ret = directSubmission->initialize(true);
-            this->dispatchMode = DispatchMode::ImmediateDispatch;
+            if (!initBlitterDirectSubmission) {
+                directSubmission = std::make_unique<
+                    MockWddmDirectSubmission<GfxFamily, RenderDispatcher<GfxFamily>>>(device, osContext);
+                ret = directSubmission->initialize(true);
+                this->dispatchMode = DispatchMode::ImmediateDispatch;
+            } else {
+                blitterDirectSubmission = std::make_unique<
+                    MockWddmDirectSubmission<GfxFamily, BlitterDispatcher<GfxFamily>>>(device, osContext);
+                blitterDirectSubmission->initialize(true);
+            }
         }
         return ret;
     }
@@ -135,6 +141,7 @@ struct MockWddmCsr : public WddmCommandStreamReceiver<GfxFamily> {
     std::unique_ptr<CommandBuffer> recordedCommandBuffer = nullptr;
 
     bool callParentInitDirectSubmission = true;
+    bool initBlitterDirectSubmission = false;
 };
 
 class WddmCommandStreamWithMockGdiFixture {
@@ -1042,6 +1049,44 @@ TEST_F(WddmCommandStreamMockGdiTest, givenDirectSubmissionEnabledOnRcsWhenFlushi
     memoryManager->freeGraphicsMemory(commandBuffer);
 }
 
+TEST_F(WddmCommandStreamMockGdiTest, givenDirectSubmissionEnabledOnBcsWhenFlushingCommandBufferThenExpectDirectSubmissionUsed) {
+    using Dispatcher = BlitterDispatcher<DEFAULT_TEST_FAMILY_NAME>;
+    using MockSubmission =
+        MockWddmDirectSubmission<DEFAULT_TEST_FAMILY_NAME, Dispatcher>;
+
+    DebugManager.flags.EnableDirectSubmission.set(1);
+
+    auto hwInfo = device->getRootDeviceEnvironment().getMutableHardwareInfo();
+    hwInfo->capabilityTable.directSubmissionEngines.data[aub_stream::ENGINE_BCS].engineSupported = true;
+
+    std::unique_ptr<OsContext> osContext;
+    osContext.reset(OsContext::create(device->getExecutionEnvironment()->rootDeviceEnvironments[0]->osInterface.get(),
+                                      0, device->getDeviceBitfield(), aub_stream::ENGINE_BCS, PreemptionMode::ThreadGroup,
+                                      false, false, false));
+    csr->callParentInitDirectSubmission = false;
+    csr->initBlitterDirectSubmission = true;
+    bool ret = csr->initDirectSubmission(*device.get(), *osContext.get());
+    EXPECT_TRUE(ret);
+    EXPECT_FALSE(csr->isDirectSubmissionEnabled());
+    EXPECT_TRUE(csr->isBlitterDirectSubmissionEnabled());
+
+    GraphicsAllocation *commandBuffer = memoryManager->allocateGraphicsMemoryWithProperties(MockAllocationProperties{csr->getRootDeviceIndex(), MemoryConstants::pageSize});
+    ASSERT_NE(nullptr, commandBuffer);
+    LinearStream cs(commandBuffer);
+    BatchBuffer batchBuffer{cs.getGraphicsAllocation(), 0, 0,
+                            nullptr, false, false, QueueThrottle::MEDIUM, QueueSliceCount::defaultSliceCount, cs.getUsed(),
+                            &cs, commandBuffer->getUnderlyingBuffer()};
+    csr->flush(batchBuffer, csr->getResidencyAllocations());
+    auto directSubmission = reinterpret_cast<MockSubmission *>(csr->blitterDirectSubmission.get());
+    EXPECT_TRUE(directSubmission->ringStart);
+    size_t actualDispatchSize = directSubmission->ringCommandStream.getUsed();
+    size_t expectedSize = directSubmission->getSizeSemaphoreSection() +
+                          Dispatcher::getSizePreemption() +
+                          directSubmission->getSizeDispatch();
+    EXPECT_EQ(expectedSize, actualDispatchSize);
+    memoryManager->freeGraphicsMemory(commandBuffer);
+}
+
 TEST_F(WddmCommandStreamTest, givenResidencyLoggingAvailableWhenFlushingCommandBufferThenNotifiesResidencyLogger) {
     if (!NEO::wddmResidencyLoggingAvailable) {
         GTEST_SKIP();