Add additional StateBaseAddress cmd wa

Resolves: NEO-5982
Signed-off-by: Sebastian Luzynski <sebastian.jozef.luzynski@intel.com>

Author: SebastianLuzynski

level_zero/core/source/cmdlist/cmdlist_hw.inl

@@ -1998,6 +1998,7 @@ void CommandListCoreFamily<gfxCoreFamily>::programStateBaseAddress(NEO::CommandC
     args.hdcPipelineFlush = true;
     args.textureCacheInvalidationEnable = true;
     NEO::MemorySynchronizationCommands<GfxFamily>::addPipeControl(*commandContainer.getCommandStream(), args);
+    auto &hwInfo = device->getHwInfo();
 
     STATE_BASE_ADDRESS sba;
     NEO::EncodeStateBaseAddress<GfxFamily>::encode(commandContainer, sba);
@@ -2012,6 +2013,8 @@ void CommandListCoreFamily<gfxCoreFamily>::programStateBaseAddress(NEO::CommandC
 
         device->getL0Debugger()->captureStateBaseAddress(commandContainer, sbaAddresses);
     }
+
+    NEO::EncodeStateBaseAddress<GfxFamily>::addStateBaseAddressIfRequired(commandContainer, sba, hwInfo);
 }
 
 template <GFXCORE_FAMILY gfxCoreFamily>

opencl/test/unit_test/os_interface/hw_info_config_tests.cpp

@@ -139,3 +139,10 @@ HWTEST_F(HwInfoConfigTest, givenSamplerStateWhenAdjustSamplerStateThenNothingIsC
 
     EXPECT_EQ(0, memcmp(&initialState, &state, sizeof(SAMPLER_STATE)));
 }
+
+HWTEST_F(HwInfoConfigTest, whenCallingIsAdditionalStateBaseAddressWARequiredThenFalseIsReturned) {
+    auto hwInfoConfig = HwInfoConfig::get(pInHwInfo.platform.eProductFamily);
+    bool ret = hwInfoConfig->isAdditionalStateBaseAddressWARequired(pInHwInfo);
+
+    EXPECT_FALSE(ret);
+}

opencl/test/unit_test/os_interface/linux/hw_info_config_linux_tests.cpp

@@ -88,6 +88,12 @@ void HwInfoConfigHw<IGFX_UNKNOWN>::adjustSamplerState(void *sampler, const Hardw
 
 template <>
 void HwInfoConfigHw<IGFX_UNKNOWN>::convertTimestampsFromOaToCsDomain(uint64_t &timestampData){};
+
+template <>
+bool HwInfoConfigHw<IGFX_UNKNOWN>::isAdditionalStateBaseAddressWARequired(const HardwareInfo &hwInfo) const {
+    return false;
+}
+
 } // namespace NEO
 
 struct DummyHwConfig : HwInfoConfigHw<IGFX_UNKNOWN> {
@@ -539,6 +545,11 @@ TEST_F(HwInfoConfigTestLinuxDummy, givenFailingGttSizeIoctlWhenInitializingHwInf
     EXPECT_EQ(pInHwInfo.capabilityTable.gpuAddressSpace, outHwInfo.capabilityTable.gpuAddressSpace);
 }
 
+HWTEST_F(HwInfoConfigTestLinuxDummy, givenHardwareInfoWhenCallingIsAdditionalStateBaseAddressWARequiredThenFalseIsReturned) {
+    bool ret = hwConfig.isAdditionalStateBaseAddressWARequired(outHwInfo);
+    EXPECT_FALSE(ret);
+}
+
 using HwConfigLinux = ::testing::Test;
 
 HWTEST2_F(HwConfigLinux, GivenDifferentValuesFromTopologyQueryWhenConfiguringHwInfoThenMaxSlicesSupportedSetToAvailableCountInGtSystemInfo, MatchAny) {

opencl/test/unit_test/os_interface/windows/hw_info_config_win_tests.cpp

@@ -67,6 +67,11 @@ void HwInfoConfigHw<IGFX_UNKNOWN>::convertTimestampsFromOaToCsDomain(uint64_t &t
 template <>
 void HwInfoConfigHw<IGFX_UNKNOWN>::adjustSamplerState(void *sampler, const HardwareInfo &hwInfo){};
 
+template <>
+bool HwInfoConfigHw<IGFX_UNKNOWN>::isAdditionalStateBaseAddressWARequired(const HardwareInfo &hwInfo) const {
+    return false;
+}
+
 HwInfoConfigTestWindows::HwInfoConfigTestWindows() {
     this->executionEnvironment = std::make_unique<MockExecutionEnvironment>();
     this->rootDeviceEnvironment = std::make_unique<RootDeviceEnvironment>(*executionEnvironment);
@@ -117,6 +122,11 @@ TEST_F(HwInfoConfigTestWindows, givenInstrumentationForHardwareIsEnabledOrDisabl
     EXPECT_TRUE(outHwInfo.capabilityTable.instrumentationEnabled);
 }
 
+HWTEST_F(HwInfoConfigTestWindows, givenHardwareInfoWhenCallingIsAdditionalStateBaseAddressWARequiredThenFalseIsReturned) {
+    bool ret = hwConfig.isAdditionalStateBaseAddressWARequired(outHwInfo);
+    EXPECT_FALSE(ret);
+}
+
 HWTEST_F(HwInfoConfigTestWindows, givenFtrIaCoherencyFlagWhenConfiguringHwInfoThenSetCoherencySupportCorrectly) {
     HardwareInfo initialHwInfo = *defaultHwInfo;
     auto &hwHelper = HwHelper::get(initialHwInfo.platform.eRenderCoreFamily);

shared/source/command_container/command_encoder.h

@@ -219,6 +219,7 @@ struct EncodeStateBaseAddress {
     using STATE_BASE_ADDRESS = typename GfxFamily::STATE_BASE_ADDRESS;
     static void encode(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd);
     static void encode(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd, uint32_t statelessMocsIndex, bool useGlobalAtomics);
+    static void addStateBaseAddressIfRequired(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd, const HardwareInfo &hwInfo);
 };
 
 template <typename GfxFamily>

shared/source/command_container/command_encoder_bdw_plus.inl

@@ -386,6 +386,9 @@ void EncodeStateBaseAddress<Family>::encode(CommandContainer &container, STATE_B
     EncodeWA<Family>::encodeAdditionalPipelineSelect(*container.getDevice(), *container.getCommandStream(), false);
 }
 
+template <typename Family>
+void EncodeStateBaseAddress<Family>::addStateBaseAddressIfRequired(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd, const HardwareInfo &hwInfo) {}
+
 template <typename Family>
 void EncodeL3State<Family>::encode(CommandContainer &container, bool enableSLM) {
     auto offset = L3CNTLRegisterOffset<Family>::registerOffset;

shared/source/command_container/command_encoder_xehp_plus.inl

@@ -484,6 +484,9 @@ void EncodeStateBaseAddress<Family>::encode(CommandContainer &container, STATE_B
     }
 }
 
+template <typename Family>
+void EncodeStateBaseAddress<Family>::addStateBaseAddressIfRequired(CommandContainer &container, STATE_BASE_ADDRESS &sbaCmd, const HardwareInfo &hwInfo) {}
+
 template <typename Family>
 void EncodeComputeMode<Family>::adjustComputeMode(LinearStream &csr, void *const stateComputeModePtr, StateComputeModeProperties &properties) {
     using STATE_COMPUTE_MODE = typename Family::STATE_COMPUTE_MODE;

shared/source/command_stream/command_stream_receiver_hw.h

@@ -133,6 +133,7 @@ class CommandStreamReceiverHw : public CommandStreamReceiver {
     void programPreamble(LinearStream &csr, Device &device, DispatchFlags &dispatchFlags, uint32_t &newL3Config);
     void programPipelineSelect(LinearStream &csr, PipelineSelectArgs &pipelineSelectArgs);
     void programAdditionalPipelineSelect(LinearStream &csr, PipelineSelectArgs &pipelineSelectArgs, bool is3DPipeline);
+    void programAdditionalStateBaseAddress(LinearStream &csr, typename GfxFamily::STATE_BASE_ADDRESS &cmd, Device &device);
     void programEpilogue(LinearStream &csr, Device &device, void **batchBufferEndLocation, DispatchFlags &dispatchFlags);
     void programEpliogueCommands(LinearStream &csr, const DispatchFlags &dispatchFlags);
     void programMediaSampler(LinearStream &csr, DispatchFlags &dispatchFlags);

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -422,6 +422,8 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
             dispatchFlags.areMultipleSubDevicesInContext);
         *pCmd = cmd;
 
+        programAdditionalStateBaseAddress(commandStreamCSR, cmd, device);
+
         if (sshDirty) {
             bindingTableBaseAddressRequired = true;
         }
@@ -1314,6 +1316,9 @@ inline void CommandStreamReceiverHw<GfxFamily>::programAdditionalPipelineSelect(
     }
 }
 
+template <typename GfxFamily>
+inline void CommandStreamReceiverHw<GfxFamily>::programAdditionalStateBaseAddress(LinearStream &csr, typename GfxFamily::STATE_BASE_ADDRESS &cmd, Device &device) {}
+
 template <typename GfxFamily>
 inline bool CommandStreamReceiverHw<GfxFamily>::isComputeModeNeeded() const {
     return false;

shared/source/os_interface/hw_info_config.h

@@ -38,6 +38,7 @@ class HwInfoConfig {
     virtual uint64_t getSharedSystemMemCapabilities() = 0;
     virtual void convertTimestampsFromOaToCsDomain(uint64_t &timestampData) = 0;
     virtual uint32_t getDeviceMemoryMaxClkRate(const HardwareInfo *hwInfo) = 0;
+    virtual bool isAdditionalStateBaseAddressWARequired(const HardwareInfo &hwInfo) const = 0;
     uint32_t threadsPerEu;
 };
 
@@ -59,6 +60,7 @@ class HwInfoConfigHw : public HwInfoConfig {
     uint64_t getSharedSystemMemCapabilities() override;
     void convertTimestampsFromOaToCsDomain(uint64_t &timestampData) override;
     uint32_t getDeviceMemoryMaxClkRate(const HardwareInfo *hwInfo) override;
+    bool isAdditionalStateBaseAddressWARequired(const HardwareInfo &hwInfo) const override;
 
   protected:
     HwInfoConfigHw() = default;