performance: reuse cmd buffer without dc flush

Related-To: NEO-16348

Signed-off-by: Szymon Morek <szymon.morek@intel.com>

shared/source/command_stream/command_stream_receiver.cpp

@@ -546,6 +546,7 @@ void CommandStreamReceiver::setTagAllocation(GraphicsAllocation *allocation) {
     this->tagAddress = reinterpret_cast<TagAddressType *>(allocation->getUnderlyingBuffer());
     this->debugPauseStateAddress = reinterpret_cast<DebugPauseState *>(
         reinterpret_cast<uint8_t *>(allocation->getUnderlyingBuffer()) + TagAllocationLayout::debugPauseStateAddressOffset);
+    this->ucTagAddress = static_cast<TagAddressType *>(ptrOffset(allocation->getUnderlyingBuffer(), TagAllocationLayout::ucTagAddressOffset));
 }
 
 MultiGraphicsAllocation &CommandStreamReceiver::createMultiAllocationInSystemMemoryPool(AllocationType allocationType) {
@@ -877,14 +878,18 @@ bool CommandStreamReceiver::initializeTagAllocation() {
     }
 
     this->setTagAllocation(tagAllocation);
+
     auto initValue = debugManager.flags.EnableNullHardware.get() ? static_cast<uint32_t>(-1) : initialHardwareTag;
     auto tagAddress = this->tagAddress;
+    auto ucTagAddress = this->ucTagAddress;
     auto completionFence = reinterpret_cast<TaskCountType *>(getCompletionAddress());
     UNRECOVERABLE_IF(!completionFence);
     uint32_t subDevices = static_cast<uint32_t>(this->deviceBitfield.count());
     for (uint32_t i = 0; i < subDevices; i++) {
         *tagAddress = initValue;
         tagAddress = ptrOffset(tagAddress, this->immWritePostSyncWriteOffset);
+        *ucTagAddress = initValue;
+        ucTagAddress = ptrOffset(ucTagAddress, this->immWritePostSyncWriteOffset);
         *completionFence = 0;
         completionFence = ptrOffset(completionFence, this->immWritePostSyncWriteOffset);
     }
@@ -1207,6 +1212,7 @@ TaskCountType CompletionStamp::getTaskCountFromSubmissionStatusError(SubmissionS
 }
 uint64_t CommandStreamReceiver::getBarrierCountGpuAddress() const { return ptrOffset(this->tagAllocation->getGpuAddress(), TagAllocationLayout::barrierCountOffset); }
 uint64_t CommandStreamReceiver::getDebugPauseStateGPUAddress() const { return tagAllocation->getGpuAddress() + TagAllocationLayout::debugPauseStateAddressOffset; }
+uint64_t CommandStreamReceiver::getUcTagGPUAddress() const { return tagAllocation->getGpuAddress() + TagAllocationLayout::ucTagAddressOffset; }
 uint64_t CommandStreamReceiver::getCompletionAddress() const {
     uint64_t completionFenceAddress = castToUint64(const_cast<TagAddressType *>(tagAddress));
     if (completionFenceAddress == 0) {

shared/source/command_stream/command_stream_receiver.h

@@ -166,9 +166,11 @@ class CommandStreamReceiver : NEO::NonCopyableAndNonMovableClass {
     TaskCountType getNextBarrierCount() { return this->barrierCount.fetch_add(1u); }
     TaskCountType peekBarrierCount() const { return this->barrierCount.load(); }
     volatile TagAddressType *getTagAddress() const { return tagAddress; }
+    volatile TagAddressType *getUcTagAddress() const { return ucTagAddress; }
     volatile TagAddressType *getBarrierCountTagAddress() const { return this->barrierCountTagAddress; }
     uint64_t getBarrierCountGpuAddress() const;
     uint64_t getDebugPauseStateGPUAddress() const;
+    uint64_t getUcTagGPUAddress() const;
 
     virtual bool waitForFlushStamp(FlushStamp &flushStampToWait) { return true; }
 
@@ -627,6 +629,7 @@ class CommandStreamReceiver : NEO::NonCopyableAndNonMovableClass {
     uint64_t totalMemoryUsed = 0u;
 
     volatile TagAddressType *tagAddress = nullptr;
+    volatile TagAddressType *ucTagAddress = nullptr;
     volatile TagAddressType *barrierCountTagAddress = nullptr;
     volatile DebugPauseState *debugPauseStateAddress = nullptr;
     SpinLock debugPauseStateLock;
@@ -694,6 +697,7 @@ class CommandStreamReceiver : NEO::NonCopyableAndNonMovableClass {
     bool gsbaStateDirty = true;
     bool bindingTableBaseAddressRequired = false;
     bool heapStorageRequiresRecyclingTag = false;
+    bool ucResourceRequiresTagUpdate = false;
     bool mediaVfeStateDirty = true;
     bool stateComputeModeDirty = true;
     bool btdCommandDirty = true;

shared/source/command_stream/command_stream_receiver_hw.h

@@ -276,6 +276,8 @@ class CommandStreamReceiverHw : public CommandStreamReceiver {
                                               bool areMultipleSubDevicesInContext,
                                               bool setGeneralStateBaseAddress);
 
+    inline void emitTagUpdateWithoutDCFlush(LinearStream &commandStream);
+
     inline void processBarrierWithPostSync(LinearStream &commandStreamTask,
                                            DispatchFlags &dispatchFlags,
                                            bool &levelClosed,

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -442,7 +442,9 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTaskHeapful(
     if (detectInitProgrammingFlagsRequired(dispatchFlags)) {
         initProgrammingFlags();
     }
-
+    if (this->ucResourceRequiresTagUpdate) {
+        this->emitTagUpdateWithoutDCFlush(commandStreamTask);
+    }
     const auto &hwInfo = peekHwInfo();
 
     bool hasStallingCmdsOnTaskStream = false;
@@ -1842,6 +1844,25 @@ inline void CommandStreamReceiverHw<GfxFamily>::programStateBaseAddressCommon(
     }
 }
 
+template <typename GfxFamily>
+inline void CommandStreamReceiverHw<GfxFamily>::emitTagUpdateWithoutDCFlush(LinearStream &commandStream) {
+    auto &rootDeviceEnvironment = this->peekRootDeviceEnvironment();
+    auto address = this->getUcTagGPUAddress();
+
+    PipeControlArgs args = {};
+    args.notifyEnable = isUsedNotifyEnableForPostSync();
+    MemorySynchronizationCommands<GfxFamily>::addBarrierWithPostSyncOperation(
+        commandStream,
+        PostSyncMode::immediateData,
+        address,
+        taskCount + 1,
+        rootDeviceEnvironment,
+        args);
+
+    makeResident(*tagAllocation);
+    this->ucResourceRequiresTagUpdate = false;
+}
+
 template <typename GfxFamily>
 inline void CommandStreamReceiverHw<GfxFamily>::processBarrierWithPostSync(LinearStream &commandStreamTask, DispatchFlags &dispatchFlags, bool &levelClosed, void *&currentPipeControlForNooping, void *&epiloguePipeControlLocation, bool &hasStallingCmdsOnTaskStream, PipeControlArgs &args) {
 

shared/source/command_stream/tag_allocation_layout.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2022-2023 Intel Corporation
+ * Copyright (C) 2022-2025 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -11,6 +11,7 @@
 namespace NEO {
 namespace TagAllocationLayout {
 inline constexpr uint64_t debugPauseStateAddressOffset = MemoryConstants::kiloByte;
+inline constexpr uint64_t ucTagAddressOffset = MemoryConstants::kiloByte + MemoryConstants::cacheLineSize;
 inline constexpr uint64_t completionFenceOffset = 2 * MemoryConstants::kiloByte;
 inline constexpr uint64_t barrierCountOffset = 3 * MemoryConstants::kiloByte;
 } // namespace TagAllocationLayout