Enable prepatcher DirectSubmission WA

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

Author: BartoszDunajski

opencl/test/unit_test/command_stream/command_stream_receiver_hw_2_tests.cpp

@@ -1405,7 +1405,7 @@ HWTEST_F(BcsTests, givenBlitterDirectSubmissionEnabledWhenProgrammingBlitterThen
     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->getBatchBufferStartAddress());
+    EXPECT_NE(0ull, bbStart->getBatchBufferStartAddress());
 }
 
 HWTEST_F(BcsTests, givenBlitterDirectSubmissionEnabledWhenFlushTagUpdateThenBatchBufferStartIsProgrammed) {

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -94,9 +94,9 @@ inline void CommandStreamReceiverHw<GfxFamily>::addBatchBufferEnd(LinearStream &
 template <typename GfxFamily>
 inline void CommandStreamReceiverHw<GfxFamily>::programEndingCmd(LinearStream &commandStream, Device &device, void **patchLocation, bool directSubmissionEnabled) {
     if (directSubmissionEnabled) {
-        uint64_t startAddress = 0;
-        if (DebugManager.flags.BatchBufferStartPrepatchingWaEnabled.get() == 1) {
-            startAddress = commandStream.getGraphicsAllocation()->getGpuAddress() + commandStream.getUsed();
+        uint64_t startAddress = commandStream.getGraphicsAllocation()->getGpuAddress() + commandStream.getUsed();
+        if (DebugManager.flags.BatchBufferStartPrepatchingWaEnabled.get() == 0) {
+            startAddress = 0;
         }
 
         *patchLocation = commandStream.getSpace(sizeof(MI_BATCH_BUFFER_START));

shared/test/unit_test/direct_submission/direct_submission_tests_1.cpp

@@ -602,9 +602,12 @@ HWTEST_F(DirectSubmissionTest, givenDirectSubmissionAvailableWhenProgrammingEndi
     bool ret = mockCsr->isDirectSubmissionEnabled();
     EXPECT_TRUE(ret);
 
+    MockGraphicsAllocation mockAllocation;
+
     void *location = nullptr;
     uint8_t buffer[128];
     mockCsr->commandStream.replaceBuffer(&buffer[0], 128u);
+    mockCsr->commandStream.replaceGraphicsAllocation(&mockAllocation);
     auto &device = *pDevice;
     mockCsr->programEndingCmd(mockCsr->commandStream, device, &location, ret);
     EXPECT_EQ(sizeof(MI_BATCH_BUFFER_START), mockCsr->commandStream.getUsed());
@@ -615,18 +618,17 @@ HWTEST_F(DirectSubmissionTest, givenDirectSubmissionAvailableWhenProgrammingEndi
                           mockCsr->getCmdSizeForEpilogueCommands(dispatchFlags);
     expectedSize = alignUp(expectedSize, MemoryConstants::cacheLineSize);
     EXPECT_EQ(expectedSize, mockCsr->getCmdSizeForEpilogue(dispatchFlags));
+
+    mockCsr->commandStream.replaceGraphicsAllocation(nullptr);
 }
 
-HWTEST_F(DirectSubmissionTest, givenDebugFlagSetWhenProgrammingEndingCommandThenUseNonZeroBatchBufferStart) {
+HWTEST_F(DirectSubmissionTest, givenDebugFlagSetWhenProgrammingEndingCommandThenUseCorrectBatchBufferStartValue) {
     using MI_BATCH_BUFFER_START = typename FamilyType::MI_BATCH_BUFFER_START;
 
     DebugManagerStateRestore restorer;
-
-    DebugManager.flags.BatchBufferStartPrepatchingWaEnabled.set(1);
-
     MockGraphicsAllocation mockAllocation;
-
     int32_t executionStamp = 0;
+
     std::unique_ptr<MockCsr<FamilyType>> mockCsr = std::make_unique<MockCsr<FamilyType>>(executionStamp, *pDevice->executionEnvironment,
                                                                                          pDevice->getRootDeviceIndex(), pDevice->getDeviceBitfield());
 
@@ -636,25 +638,30 @@ HWTEST_F(DirectSubmissionTest, givenDebugFlagSetWhenProgrammingEndingCommandThen
     EXPECT_TRUE(ret);
 
     void *location = nullptr;
-    uint8_t buffer[128];
-    mockCsr->commandStream.replaceBuffer(&buffer[0], 128u);
-    mockCsr->commandStream.replaceGraphicsAllocation(&mockAllocation);
+
+    uint8_t buffer[256] = {};
+    auto &cmdStream = mockCsr->commandStream;
+    cmdStream.replaceBuffer(&buffer[0], 256);
+    cmdStream.replaceGraphicsAllocation(&mockAllocation);
     auto &device = *pDevice;
-    mockCsr->programEndingCmd(mockCsr->commandStream, device, &location, ret);
-    EXPECT_EQ(sizeof(MI_BATCH_BUFFER_START), mockCsr->commandStream.getUsed());
 
-    DispatchFlags dispatchFlags = DispatchFlagsHelper::createDefaultDispatchFlags();
-    dispatchFlags.epilogueRequired = true;
-    size_t expectedSize = sizeof(MI_BATCH_BUFFER_START) +
-                          mockCsr->getCmdSizeForEpilogueCommands(dispatchFlags);
-    expectedSize = alignUp(expectedSize, MemoryConstants::cacheLineSize);
-    EXPECT_EQ(expectedSize, mockCsr->getCmdSizeForEpilogue(dispatchFlags));
+    for (int32_t value : {-1, 0, 1}) {
+        DebugManager.flags.BatchBufferStartPrepatchingWaEnabled.set(value);
 
-    auto bbStartCmd = reinterpret_cast<MI_BATCH_BUFFER_START *>(mockCsr->commandStream.getCpuBase());
+        auto currectBbStartCmd = reinterpret_cast<MI_BATCH_BUFFER_START *>(cmdStream.getSpace(0));
+        uint64_t expectedGpuVa = cmdStream.getGraphicsAllocation()->getGpuAddress() + cmdStream.getUsed();
 
-    EXPECT_EQ(mockCsr->commandStream.getGraphicsAllocation()->getGpuAddress(), bbStartCmd->getBatchBufferStartAddress());
+        mockCsr->programEndingCmd(cmdStream, device, &location, ret);
+        EncodeNoop<FamilyType>::alignToCacheLine(cmdStream);
 
-    mockCsr->commandStream.replaceGraphicsAllocation(nullptr);
+        if (value == 0) {
+            EXPECT_EQ(0u, currectBbStartCmd->getBatchBufferStartAddress());
+        } else {
+            EXPECT_EQ(expectedGpuVa, currectBbStartCmd->getBatchBufferStartAddress());
+        }
+    }
+
+    cmdStream.replaceGraphicsAllocation(nullptr);
 }
 
 HWTEST_F(DirectSubmissionTest,