Use LogicalStateHelper to encode SystemMemoryFence

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

Author: BartoszDunajski

shared/source/command_container/command_encoder.h

@@ -27,6 +27,7 @@ namespace NEO {
 
 class BindlessHeapsHelper;
 class GmmHelper;
+class LogicalStateHelper;
 class IndirectHeap;
 class Gmm;
 struct HardwareInfo;
@@ -455,7 +456,7 @@ template <typename GfxFamily>
 struct EncodeMemoryFence {
     static size_t getSystemMemoryFenceSize();
 
-    static void encodeSystemMemoryFence(LinearStream &commandStream, const GraphicsAllocation *globalFenceAllocation);
+    static void encodeSystemMemoryFence(LinearStream &commandStream, const GraphicsAllocation *globalFenceAllocation, LogicalStateHelper *logicalStateHelper);
 };
 
 } // namespace NEO

shared/source/command_container/command_encoder.inl

@@ -884,7 +884,7 @@ size_t EncodeMemoryFence<Family>::getSystemMemoryFenceSize() {
 }
 
 template <typename Family>
-void EncodeMemoryFence<Family>::encodeSystemMemoryFence(LinearStream &commandStream, const GraphicsAllocation *globalFenceAllocation) {
+void EncodeMemoryFence<Family>::encodeSystemMemoryFence(LinearStream &commandStream, const GraphicsAllocation *globalFenceAllocation, LogicalStateHelper *logicalStateHelper) {
 }
 
 } // namespace NEO

shared/source/command_container/command_encoder_xe_hpc_core_and_later.inl

@@ -19,7 +19,7 @@ size_t EncodeMemoryFence<Family>::getSystemMemoryFenceSize() {
 }
 
 template <>
-void EncodeMemoryFence<Family>::encodeSystemMemoryFence(LinearStream &commandStream, const GraphicsAllocation *globalFenceAllocation) {
+void EncodeMemoryFence<Family>::encodeSystemMemoryFence(LinearStream &commandStream, const GraphicsAllocation *globalFenceAllocation, LogicalStateHelper *logicalStateHelper) {
     using STATE_SYSTEM_MEM_FENCE_ADDRESS = typename Family::STATE_SYSTEM_MEM_FENCE_ADDRESS;
 
     auto stateSystemFenceAddressSpace = commandStream.getSpaceForCmd<STATE_SYSTEM_MEM_FENCE_ADDRESS>();

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -1053,6 +1053,10 @@ std::optional<uint32_t> CommandStreamReceiverHw<GfxFamily>::flushBcsTask(const B
         pageTableManagerInitialized = pageTableManager->initPageTableManagerRegisters(this);
     }
 
+    if (logicalStateHelper) {
+        logicalStateHelper->writeStreamInline(commandStream);
+    }
+
     for (auto &blitProperties : blitPropertiesContainer) {
         TimestampPacketHelper::programCsrDependenciesForTimestampPacketContainer<GfxFamily>(commandStream, blitProperties.csrDependencies);
         TimestampPacketHelper::programCsrDependenciesForForTaskCountContainer<GfxFamily>(commandStream, blitProperties.csrDependencies);

shared/source/direct_submission/direct_submission_hw.cpp

@@ -12,6 +12,7 @@
 namespace NEO {
 DirectSubmissionInputParams::DirectSubmissionInputParams(const CommandStreamReceiver &commandStreamReceiver) : osContext(commandStreamReceiver.getOsContext()), rootDeviceEnvironment(commandStreamReceiver.peekRootDeviceEnvironment()), rootDeviceIndex(commandStreamReceiver.getRootDeviceIndex()) {
     memoryManager = commandStreamReceiver.getMemoryManager();
+    logicalStateHelper = commandStreamReceiver.getLogicalStateHelper();
     globalFenceAllocation = commandStreamReceiver.getGlobalFenceAllocation();
     workPartitionAllocation = commandStreamReceiver.getWorkPartitionAllocation();
     completionFenceAllocation = commandStreamReceiver.getTagAllocation();

shared/source/direct_submission/direct_submission_hw.h

@@ -54,6 +54,7 @@ struct BatchBuffer;
 class DirectSubmissionDiagnosticsCollector;
 class FlushStampTracker;
 class GraphicsAllocation;
+class LogicalStateHelper;
 struct HardwareInfo;
 class OsContext;
 class MemoryOperationsHandler;
@@ -62,6 +63,7 @@ struct DirectSubmissionInputParams : NonCopyableClass {
     DirectSubmissionInputParams(const CommandStreamReceiver &commandStreamReceiver);
     OsContext &osContext;
     const RootDeviceEnvironment &rootDeviceEnvironment;
+    LogicalStateHelper *logicalStateHelper = nullptr;
     MemoryManager *memoryManager = nullptr;
     const GraphicsAllocation *globalFenceAllocation = nullptr;
     GraphicsAllocation *workPartitionAllocation = nullptr;
@@ -170,6 +172,7 @@ class DirectSubmissionHw {
     OsContext &osContext;
     const uint32_t rootDeviceIndex;
     MemoryManager *memoryManager = nullptr;
+    LogicalStateHelper *logicalStateHelper = nullptr;
     MemoryOperationsHandler *memoryOperationHandler = nullptr;
     const HardwareInfo *hwInfo = nullptr;
     const GraphicsAllocation *globalFenceAllocation = nullptr;

shared/source/direct_submission/direct_submission_hw.inl

@@ -13,6 +13,7 @@
 #include "shared/source/direct_submission/direct_submission_hw.h"
 #include "shared/source/direct_submission/direct_submission_hw_diagnostic_mode.h"
 #include "shared/source/helpers/flush_stamp.h"
+#include "shared/source/helpers/logical_state_helper.h"
 #include "shared/source/helpers/ptr_math.h"
 #include "shared/source/memory_manager/allocation_properties.h"
 #include "shared/source/memory_manager/graphics_allocation.h"
@@ -34,6 +35,7 @@ DirectSubmissionHw<GfxFamily, Dispatcher>::DirectSubmissionHw(const DirectSubmis
     : ringBuffers(RingBufferUse::initialRingBufferCount), osContext(inputParams.osContext), rootDeviceIndex(inputParams.rootDeviceIndex) {
     memoryManager = inputParams.memoryManager;
     globalFenceAllocation = inputParams.globalFenceAllocation;
+    logicalStateHelper = inputParams.logicalStateHelper;
     hwInfo = inputParams.rootDeviceEnvironment.getHardwareInfo();
     memoryOperationHandler = inputParams.rootDeviceEnvironment.memoryOperationsInterface.get();
 
@@ -642,7 +644,11 @@ size_t DirectSubmissionHw<GfxFamily, Dispatcher>::getDiagnosticModeSection() {
 template <typename GfxFamily, typename Dispatcher>
 void DirectSubmissionHw<GfxFamily, Dispatcher>::dispatchSystemMemoryFenceAddress() {
     UNRECOVERABLE_IF(!this->globalFenceAllocation);
-    EncodeMemoryFence<GfxFamily>::encodeSystemMemoryFence(ringCommandStream, this->globalFenceAllocation);
+    EncodeMemoryFence<GfxFamily>::encodeSystemMemoryFence(ringCommandStream, this->globalFenceAllocation, this->logicalStateHelper);
+
+    if (logicalStateHelper) {
+        logicalStateHelper->writeStreamInline(ringCommandStream);
+    }
 }
 
 template <typename GfxFamily, typename Dispatcher>

shared/source/xe_hpc_core/command_stream_receiver_hw_xe_hpc_core.cpp

@@ -34,7 +34,7 @@ template <>
 void CommandStreamReceiverHw<Family>::programEnginePrologue(LinearStream &csr) {
     if (!this->isEnginePrologueSent) {
         if (globalFenceAllocation) {
-            EncodeMemoryFence<Family>::encodeSystemMemoryFence(csr, globalFenceAllocation);
+            EncodeMemoryFence<Family>::encodeSystemMemoryFence(csr, globalFenceAllocation, nullptr);
         }
         this->isEnginePrologueSent = true;
     }

shared/test/unit_test/encoders/test_command_encoder.cpp

@@ -154,7 +154,7 @@ HWTEST_F(CommandEncoderTest, givenPlatformSupportingMiMemFenceWhenEncodingThenPr
 
     size_t size = EncodeMemoryFence<FamilyType>::getSystemMemoryFenceSize();
 
-    EncodeMemoryFence<FamilyType>::encodeSystemMemoryFence(cmdStream, &allocation);
+    EncodeMemoryFence<FamilyType>::encodeSystemMemoryFence(cmdStream, &allocation, nullptr);
 
     if constexpr (FamilyType::isUsingMiMemFence) {
         using STATE_SYSTEM_MEM_FENCE_ADDRESS = typename FamilyType::STATE_SYSTEM_MEM_FENCE_ADDRESS;