performance: immediate flush add jump to batch buffer when preamble is present

Related-To: NEO-7808

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

shared/source/command_stream/command_stream_receiver_hw.h

@@ -252,6 +252,12 @@ class CommandStreamReceiverHw : public CommandStreamReceiver {
     inline void handleImmediateFlushOneTimeContextInitState(ImmediateDispatchFlags &dispatchFlags, ImmediateFlushData &flushData, Device &device);
     inline void dispatchImmediateFlushOneTimeContextInitCommand(ImmediateFlushData &flushData, LinearStream &csrStream, Device &device);
 
+    inline void handleImmediateFlushJumpToImmediate(ImmediateFlushData &flushData);
+    inline void dispatchImmediateFlushJumpToImmediateCommand(LinearStream &immediateCommandStream,
+                                                             size_t immediateCommandStreamStart,
+                                                             ImmediateFlushData &flushData,
+                                                             LinearStream &csrStream);
+
     inline void handleImmediateFlushAllocationsResidency(Device &device);
 
     HeapDirtyState dshState;

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -305,6 +305,8 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushImmediateTask(
     handleImmediateFlushStateBaseAddressState(dispatchFlags, flushData, device);
     handleImmediateFlushOneTimeContextInitState(dispatchFlags, flushData, device);
 
+    handleImmediateFlushJumpToImmediate(flushData);
+
     auto &csrCommandStream = getCS(flushData.estimatedSize);
 
     dispatchImmediateFlushPipelineSelectCommand(flushData, csrCommandStream);
@@ -313,6 +315,8 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushImmediateTask(
     dispatchImmediateFlushStateBaseAddressCommand(flushData, csrCommandStream, device);
     dispatchImmediateFlushOneTimeContextInitCommand(flushData, csrCommandStream, device);
 
+    dispatchImmediateFlushJumpToImmediateCommand(immediateCommandStream, immediateCommandStreamStart, flushData, csrCommandStream);
+
     handleImmediateFlushAllocationsResidency(device);
 
     CompletionStamp completionStamp = {
@@ -2050,4 +2054,25 @@ void CommandStreamReceiverHw<GfxFamily>::handleImmediateFlushAllocationsResidenc
     }
 }
 
+template <typename GfxFamily>
+void CommandStreamReceiverHw<GfxFamily>::handleImmediateFlushJumpToImmediate(ImmediateFlushData &flushData) {
+    if (flushData.estimatedSize > 0) {
+        flushData.estimatedSize += EncodeBatchBufferStartOrEnd<GfxFamily>::getBatchBufferStartSize();
+        flushData.estimatedSize = alignUp(flushData.estimatedSize, MemoryConstants::cacheLineSize);
+    }
+}
+
+template <typename GfxFamily>
+void CommandStreamReceiverHw<GfxFamily>::dispatchImmediateFlushJumpToImmediateCommand(LinearStream &immediateCommandStream,
+                                                                                      size_t immediateCommandStreamStart,
+                                                                                      ImmediateFlushData &flushData,
+                                                                                      LinearStream &csrStream) {
+    if (flushData.estimatedSize > 0) {
+        uint64_t immediateStartAddress = immediateCommandStream.getGpuBase() + immediateCommandStreamStart;
+
+        EncodeBatchBufferStartOrEnd<GfxFamily>::programBatchBufferStart(&csrStream, immediateStartAddress, false, false, false);
+        EncodeNoop<GfxFamily>::alignToCacheLine(csrStream);
+    }
+}
+
 } // namespace NEO

shared/test/common/fixtures/command_stream_receiver_fixture.h

@@ -17,7 +17,7 @@ struct CommandStreamReceiverFixture : public NEO::DeviceFixture {
     void setUp();
     void tearDown();
 
-    static constexpr size_t bufferSize = 256;
+    static constexpr size_t bufferSize = 1024;
     uint8_t cmdBuffer[bufferSize];
     uint8_t dshBuffer[bufferSize];
     uint8_t iohBuffer[bufferSize];

shared/test/unit_test/command_stream/command_stream_receiver_tests.cpp

@@ -3781,3 +3781,35 @@ HWTEST2_F(CommandStreamReceiverHwTest,
 
     EXPECT_TRUE(commandStreamReceiver.isMadeResident(pDevice->getRTMemoryBackedBuffer()));
 }
+
+HWTEST2_F(CommandStreamReceiverHwTest,
+          givenImmediateFlushTaskWhenCsrHasPreambleCommandsThenDispatchIndirectJumpToImmediateBatchBuffer,
+          IsAtLeastXeHpCore) {
+    using MI_BATCH_BUFFER_START = typename FamilyType::MI_BATCH_BUFFER_START;
+    using COMPUTE_WALKER = typename FamilyType::COMPUTE_WALKER;
+
+    auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto startOffset = commandStream.getUsed();
+    *commandStream.getSpaceForCmd<COMPUTE_WALKER>() = FamilyType::cmdInitGpgpuWalker;
+    uint64_t immediateStartAddress = commandStream.getGpuBase() + startOffset;
+
+    commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
+
+    HardwareParse hwParserCsr;
+    hwParserCsr.parseCommands<FamilyType>(commandStreamReceiver.commandStream, 0);
+    auto bbStartCmd = hwParserCsr.getCommand<MI_BATCH_BUFFER_START>();
+    ASSERT_NE(nullptr, bbStartCmd);
+    EXPECT_EQ(immediateStartAddress, bbStartCmd->getBatchBufferStartAddress());
+
+    startOffset = commandStream.getUsed();
+    *commandStream.getSpaceForCmd<COMPUTE_WALKER>() = FamilyType::cmdInitGpgpuWalker;
+
+    size_t usedSize = commandStreamReceiver.commandStream.getUsed();
+    commandStreamReceiver.flushImmediateTask(commandStream, startOffset, immediateFlushTaskFlags, *pDevice);
+
+    hwParserCsr.tearDown();
+    hwParserCsr.parseCommands<FamilyType>(commandStreamReceiver.commandStream, usedSize);
+    bbStartCmd = hwParserCsr.getCommand<MI_BATCH_BUFFER_START>();
+    ASSERT_EQ(nullptr, bbStartCmd);
+}