Disable RelaxedOrdering if UpdateTagFromWait is disabled

Related-To: NEO-7458

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

level_zero/core/test/unit_tests/xe_hpc_core/test_cmdqueue_xe_hpc_core.cpp

@@ -548,6 +548,7 @@ HWTEST2_F(CommandQueueCommandsXeHpc, givenFlushTaskSubmissionEnabledAndSplitBcsC
 
     DebugManagerStateRestore restorer;
     DebugManager.flags.SplitBcsCopy.set(1);
+    DebugManager.flags.UpdateTaskCountFromWait.set(3);
 
     ze_result_t returnValue;
     auto hwInfo = *NEO::defaultHwInfo;

opencl/test/unit_test/kernel/kernel_tests.cpp

@@ -509,7 +509,7 @@ class CommandStreamReceiverMock : public CommandStreamReceiver {
 
     SubmissionStatus flushTagUpdate() override { return SubmissionStatus::SUCCESS; };
     void updateTagFromWait() override{};
-    bool isUpdateTagFromWaitEnabled() override { return false; };
+    bool isUpdateTagFromWaitEnabled() const override { return false; };
     void createKernelArgsBufferAllocation() override {}
 
     bool isMultiOsContextCapable() const override { return false; }

shared/source/command_stream/command_stream_receiver.h

@@ -250,7 +250,7 @@ class CommandStreamReceiver {
 
     virtual SubmissionStatus flushTagUpdate() = 0;
     virtual void updateTagFromWait() = 0;
-    virtual bool isUpdateTagFromWaitEnabled() = 0;
+    virtual bool isUpdateTagFromWaitEnabled() const = 0;
 
     ScratchSpaceController *getScratchSpaceController() const {
         return scratchSpaceController.get();
@@ -269,7 +269,7 @@ class CommandStreamReceiver {
 
     void startControllingDirectSubmissions();
 
-    bool isAnyDirectSubmissionEnabled() {
+    bool isAnyDirectSubmissionEnabled() const {
         return this->isDirectSubmissionEnabled() || isBlitterDirectSubmissionEnabled();
     }
 

shared/source/command_stream/command_stream_receiver_hw.h

@@ -105,7 +105,7 @@ class CommandStreamReceiverHw : public CommandStreamReceiver {
                                     size_t commandStreamStartTask);
     SubmissionStatus flushHandler(BatchBuffer &batchBuffer, ResidencyContainer &allocationsForResidency);
 
-    bool isUpdateTagFromWaitEnabled() override;
+    bool isUpdateTagFromWaitEnabled() const override;
     void updateTagFromWait() override;
 
     bool isMultiOsContextCapable() const override;

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -1456,7 +1456,7 @@ inline SubmissionStatus CommandStreamReceiverHw<GfxFamily>::flushHandler(BatchBu
 }
 
 template <typename GfxFamily>
-inline bool CommandStreamReceiverHw<GfxFamily>::isUpdateTagFromWaitEnabled() {
+inline bool CommandStreamReceiverHw<GfxFamily>::isUpdateTagFromWaitEnabled() const {
     auto &gfxCoreHelper = getGfxCoreHelper();
     auto enabled = gfxCoreHelper.isUpdateTaskCountFromWaitSupported();
     enabled &= this->isAnyDirectSubmissionEnabled();

shared/source/direct_submission/direct_submission_hw.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2022 Intel Corporation
+ * Copyright (C) 2022-2023 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -16,6 +16,7 @@ DirectSubmissionInputParams::DirectSubmissionInputParams(const CommandStreamRece
     globalFenceAllocation = commandStreamReceiver.getGlobalFenceAllocation();
     workPartitionAllocation = commandStreamReceiver.getWorkPartitionAllocation();
     completionFenceAllocation = commandStreamReceiver.getTagAllocation();
+    isUpdateTagFromWaitEnabled = commandStreamReceiver.isUpdateTagFromWaitEnabled();
 }
 
 } // namespace NEO

shared/source/direct_submission/direct_submission_hw.h

@@ -70,6 +70,7 @@ struct DirectSubmissionInputParams : NonCopyableClass {
     GraphicsAllocation *workPartitionAllocation = nullptr;
     GraphicsAllocation *completionFenceAllocation = nullptr;
     const uint32_t rootDeviceIndex;
+    bool isUpdateTagFromWaitEnabled = false;
 };
 
 template <typename GfxFamily, typename Dispatcher>

shared/source/direct_submission/direct_submission_hw.inl

@@ -85,7 +85,7 @@ DirectSubmissionHw<GfxFamily, Dispatcher>::DirectSubmissionHw(const DirectSubmis
 
     dcFlushRequired = MemorySynchronizationCommands<GfxFamily>::getDcFlushEnable(true, inputParams.rootDeviceEnvironment);
     auto &gfxCoreHelper = inputParams.rootDeviceEnvironment.getHelper<GfxCoreHelper>();
-    relaxedOrderingEnabled = gfxCoreHelper.isRelaxedOrderingSupported();
+    relaxedOrderingEnabled = gfxCoreHelper.isRelaxedOrderingSupported() && inputParams.isUpdateTagFromWaitEnabled;
 
     this->currentRelaxedOrderingQueueSize = RelaxedOrderingHelper::queueSizeMultiplier;
 

shared/test/common/mocks/mock_command_stream_receiver.h

@@ -71,7 +71,7 @@ class MockCommandStreamReceiver : public CommandStreamReceiver {
 
     SubmissionStatus flushTagUpdate() override { return SubmissionStatus::SUCCESS; };
     void updateTagFromWait() override{};
-    bool isUpdateTagFromWaitEnabled() override { return false; };
+    bool isUpdateTagFromWaitEnabled() const override { return false; };
 
     bool isMultiOsContextCapable() const override { return multiOsContextCapable; }
 

shared/test/unit_test/direct_submission/direct_submission_tests_1.cpp

@@ -134,8 +134,8 @@ HWTEST_F(DirectSubmissionTest, givenDirectSubmissionWithoutCompletionFenceAlloca
     EXPECT_EQ(nullptr, directSubmission.completionFenceAllocation);
 
     size_t expectedAllocationsCnt = 3;
-    auto &gfxCoreHelper = pDevice->getGfxCoreHelper();
+
-    if (gfxCoreHelper.isRelaxedOrderingSupported()) {
+    if (directSubmission.isRelaxedOrderingEnabled()) {
         expectedAllocationsCnt += 2;
     }
 
@@ -166,8 +166,7 @@ HWTEST_F(DirectSubmissionTest, givenDirectSubmissionWithCompletionFenceAllocatio
     EXPECT_EQ(&completionFenceAllocation, directSubmission.completionFenceAllocation);
 
     size_t expectedAllocationsCnt = 4;
-    auto &gfxCoreHelper = pDevice->getGfxCoreHelper();
+    if (directSubmission.isRelaxedOrderingEnabled()) {
-    if (gfxCoreHelper.isRelaxedOrderingSupported()) {
         expectedAllocationsCnt += 2;
     }
 

shared/test/unit_test/direct_submission/direct_submission_tests_2.cpp

@@ -1017,6 +1017,7 @@ HWTEST_F(DirectSubmissionDispatchBufferTest, givenDebugFlagSetWhenStoppingRingbu
 struct DirectSubmissionRelaxedOrderingTests : public DirectSubmissionDispatchBufferTest {
     void SetUp() override {
         DebugManager.flags.DirectSubmissionRelaxedOrdering.set(1);
+        DebugManager.flags.UpdateTaskCountFromWait.set(3);
         DirectSubmissionDispatchBufferTest::SetUp();
     }
 
@@ -2704,4 +2705,24 @@ HWTEST2_F(DirectSubmissionRelaxedOrderingTests, givenNumClientsWhenAskingIfRelax
     ultCsr->registerClient();
     EXPECT_EQ(4u, ultCsr->getNumClients());
     EXPECT_TRUE(NEO::RelaxedOrderingHelper::isRelaxedOrderingDispatchAllowed(*ultCsr, 1));
+}
+
+HWTEST2_F(DirectSubmissionRelaxedOrderingTests, givenUpdateTagFromWaitDisabledWhenCreatingDirectSubmissionThenDisableRelaxedOrdering, IsAtLeastXeHpcCore) {
+    DebugManager.flags.DirectSubmissionRelaxedOrdering.set(-1);
+    auto csr = pDevice->getDefaultEngine().commandStreamReceiver;
+
+    DebugManager.flags.UpdateTaskCountFromWait.set(0);
+    {
+        MockDirectSubmissionHw<FamilyType, RenderDispatcher<FamilyType>> directSubmission(*csr);
+
+        EXPECT_FALSE(directSubmission.isRelaxedOrderingEnabled());
+    }
+
+    DebugManager.flags.DirectSubmissionRelaxedOrdering.set(1);
+    {
+
+        MockDirectSubmissionHw<FamilyType, RenderDispatcher<FamilyType>> directSubmission(*csr);
+
+        EXPECT_TRUE(directSubmission.isRelaxedOrderingEnabled());
+    }
 }

shared/test/unit_test/direct_submission/windows/wddm_direct_submission_tests.cpp

@@ -62,9 +62,8 @@ HWTEST_F(WddmDirectSubmissionTest, givenWddmWhenDirectIsInitializedAndStartedThe
     EXPECT_NE(nullptr, wddmDirectSubmission->ringBuffers[1].ringBuffer);
     EXPECT_NE(nullptr, wddmDirectSubmission->semaphores);
 
-    auto &gfxCoreHelper = device->getGfxCoreHelper();
     size_t expectedAllocationsCnt = 3;
-    if (gfxCoreHelper.isRelaxedOrderingSupported()) {
+    if (wddmDirectSubmission->isRelaxedOrderingEnabled()) {
         expectedAllocationsCnt += 2;
     }
 
@@ -99,9 +98,8 @@ HWTEST_F(WddmDirectSubmissionNoPreemptionTest, givenWddmWhenDirectIsInitializedA
     EXPECT_NE(nullptr, wddmDirectSubmission->ringBuffers[1].ringBuffer);
     EXPECT_NE(nullptr, wddmDirectSubmission->semaphores);
 
-    auto &gfxCoreHelper = device->getGfxCoreHelper();
     size_t expectedAllocationsCnt = 3;
-    if (gfxCoreHelper.isRelaxedOrderingSupported()) {
+    if (wddmDirectSubmission->isRelaxedOrderingEnabled()) {
         expectedAllocationsCnt += 2;
     }
 
@@ -142,9 +140,9 @@ HWTEST_F(WddmDirectSubmissionTest, givenWddmWhenAllocateOsResourcesThenExpectRin
 
     bool ret = wddmDirectSubmission.allocateResources();
     EXPECT_TRUE(ret);
-    auto &gfxCoreHelper = device->getGfxCoreHelper();
+
     size_t expectedAllocationsCnt = 3;
-    if (gfxCoreHelper.isRelaxedOrderingSupported()) {
+    if (wddmDirectSubmission.isRelaxedOrderingEnabled()) {
         expectedAllocationsCnt += 2;
     }
 
@@ -186,9 +184,9 @@ HWTEST_F(WddmDirectSubmissionTest, givenWddmWhenAllocateOsResourcesResidencyFail
 
     bool ret = wddmDirectSubmission.allocateResources();
     EXPECT_FALSE(ret);
-    auto &gfxCoreHelper = device->getGfxCoreHelper();
+
     size_t expectedAllocationsCnt = 3;
-    if (gfxCoreHelper.isRelaxedOrderingSupported()) {
+    if (wddmDirectSubmission.isRelaxedOrderingEnabled()) {
         expectedAllocationsCnt += 2;
     }
 

shared/test/unit_test/mocks/windows/mock_wddm_direct_submission.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2020-2022 Intel Corporation
+ * Copyright (C) 2020-2023 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,6 +26,7 @@ struct MockWddmDirectSubmission : public WddmDirectSubmission<GfxFamily, Dispatc
     using BaseClass::handleCompletionFence;
     using BaseClass::handleResidency;
     using BaseClass::isCompleted;
+    using BaseClass::isRelaxedOrderingEnabled;
     using BaseClass::miMemFenceRequired;
     using BaseClass::osContextWin;
     using BaseClass::ringBuffers;