refactor: change additional walker fields encoder 4/n

- move post sync system fence into dedicated encoder

Related-To: NEO-12639

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

opencl/source/command_queue/hardware_interface_xehp_and_later.inl

@@ -148,6 +148,7 @@ inline void HardwareInterface<GfxFamily>::programWalker(
     EncodeWalkerArgs encodeWalkerArgs{kernel.getExecutionType(), requiredSystemFence, kernelInfo.kernelDescriptor, kernelAttributes.walkOrder, kernelAttributes.additionalSize, maxFrontEndThreads};
 
     EncodeDispatchKernel<GfxFamily>::template encodeAdditionalWalkerFields<WalkerType>(rootDeviceEnvironment, walkerCmd, encodeWalkerArgs);
+    EncodeDispatchKernel<GfxFamily>::template encodeWalkerPostSyncFields<WalkerType>(walkerCmd, encodeWalkerArgs);
     EncodeDispatchKernel<GfxFamily>::template overrideDefaultValues<WalkerType, InterfaceDescriptorType>(walkerCmd, *interfaceDescriptor);
 
     auto devices = queueCsr.getOsContext().getDeviceBitfield();

shared/source/command_container/command_encoder.h

@@ -229,6 +229,8 @@ struct EncodeDispatchKernel {
 
     template <typename WalkerType, typename InterfaceDescriptorType>
     static void overrideDefaultValues(WalkerType &walkerCmd, InterfaceDescriptorType &interfaceDescriptor);
+    template <typename WalkerType>
+    static void encodeWalkerPostSyncFields(WalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs);
 };
 
 template <typename GfxFamily>

shared/source/command_container/command_encoder_bdw_and_later.inl

@@ -278,6 +278,15 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
 
     auto threadGroupCount = cmd.getThreadGroupIdXDimension() * cmd.getThreadGroupIdYDimension() * cmd.getThreadGroupIdZDimension();
     EncodeDispatchKernel<Family>::encodeThreadGroupDispatch(idd, *args.device, hwInfo, threadGroupDims, threadGroupCount, kernelDescriptor.kernelAttributes.numGrfRequired, numThreadsPerThreadGroup, cmd);
+    EncodeWalkerArgs walkerArgs{
+        KernelExecutionType::defaultType,
+        args.requiresSystemMemoryFence(),
+        kernelDescriptor,
+        args.requiredDispatchWalkOrder,
+        args.additionalSizeParam,
+        args.device->getDeviceInfo().maxFrontEndThreads};
+    EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(rootDeviceEnvironment, cmd, walkerArgs);
+    EncodeDispatchKernel<Family>::encodeWalkerPostSyncFields(cmd, walkerArgs);
 
     memcpy_s(iddPtr, sizeof(idd), &idd, sizeof(idd));
 
@@ -404,6 +413,10 @@ template <typename Family>
 template <typename WalkerType>
 inline void EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(const RootDeviceEnvironment &rootDeviceEnvironment, WalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs) {}
 
+template <typename Family>
+template <typename WalkerType>
+inline void NEO::EncodeDispatchKernel<Family>::encodeWalkerPostSyncFields(WalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs) {}
+
 template <typename Family>
 template <typename InterfaceDescriptorType>
 void EncodeDispatchKernel<Family>::setupPreferredSlmSize(InterfaceDescriptorType *pInterfaceDescriptor, const RootDeviceEnvironment &rootDeviceEnvironment, const uint32_t threadsPerThreadGroup, uint32_t slmTotalSize, SlmPolicy slmPolicy) {}

shared/source/command_container/command_encoder_enablers.inl

@@ -32,6 +32,7 @@ template void NEO::EncodeDispatchKernel<Family>::forceComputeWalkerPostSyncFlush
 template void NEO::EncodeDispatchKernel<Family>::setWalkerRegionSettings<Family::DefaultWalkerType>(Family::DefaultWalkerType &walkerCmd, const HardwareInfo &hwInfo, uint32_t partitionCount,
                                                                                                     uint32_t workgroupSize, uint32_t maxWgCountPerTile, bool requiredWalkOrder);
 template void NEO::EncodeDispatchKernel<Family>::overrideDefaultValues<Family::DefaultWalkerType, Family::INTERFACE_DESCRIPTOR_DATA>(Family::DefaultWalkerType &walkerCmd, Family::INTERFACE_DESCRIPTOR_DATA &interfaceDescriptor);
+template void NEO::EncodeDispatchKernel<Family>::encodeWalkerPostSyncFields<Family::DefaultWalkerType>(Family::DefaultWalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs);
 
 template struct NEO::EncodeStates<Family>;
 template struct NEO::EncodeMath<Family>;

shared/source/command_container/command_encoder_xehp_and_later.inl

@@ -412,6 +412,7 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
         args.additionalSizeParam,
         args.device->getDeviceInfo().maxFrontEndThreads};
     EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(rootDeviceEnvironment, walkerCmd, walkerArgs);
+    EncodeDispatchKernel<Family>::encodeWalkerPostSyncFields(walkerCmd, walkerArgs);
 
     EncodeDispatchKernel<Family>::overrideDefaultValues(walkerCmd, idd);
 
@@ -1235,4 +1236,16 @@ void EncodeDispatchKernel<Family>::encodeThreadGroupDispatch(InterfaceDescriptor
     }
 }
 
+template <typename Family>
+template <typename WalkerType>
+void EncodeDispatchKernel<Family>::encodeWalkerPostSyncFields(WalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs) {
+    auto programGlobalFenceAsPostSyncOperationInComputeWalker = walkerArgs.requiredSystemFence;
+    int32_t overrideProgramSystemMemoryFence = debugManager.flags.ProgramGlobalFenceAsPostSyncOperationInComputeWalker.get();
+    if (overrideProgramSystemMemoryFence != -1) {
+        programGlobalFenceAsPostSyncOperationInComputeWalker = !!overrideProgramSystemMemoryFence;
+    }
+    auto &postSyncData = walkerCmd.getPostSync();
+    postSyncData.setSystemMemoryFenceRequest(programGlobalFenceAsPostSyncOperationInComputeWalker);
+}
+
 } // namespace NEO

shared/source/xe2_hpg_core/command_encoder_xe2_hpg_core.cpp

@@ -232,14 +232,6 @@ void EncodeSurfaceState<Family>::disableCompressionFlags(R_SURFACE_STATE *surfac
 template <>
 template <typename WalkerType>
 void EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(const RootDeviceEnvironment &rootDeviceEnvironment, WalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs) {
-    auto programGlobalFenceAsPostSyncOperationInComputeWalker = walkerArgs.requiredSystemFence;
-    int32_t overrideProgramSystemMemoryFence = debugManager.flags.ProgramGlobalFenceAsPostSyncOperationInComputeWalker.get();
-    if (overrideProgramSystemMemoryFence != -1) {
-        programGlobalFenceAsPostSyncOperationInComputeWalker = !!overrideProgramSystemMemoryFence;
-    }
-    auto &postSyncData = walkerCmd.getPostSync();
-    postSyncData.setSystemMemoryFenceRequest(programGlobalFenceAsPostSyncOperationInComputeWalker);
-
     bool computeDispatchAllWalkerEnable = walkerArgs.kernelExecutionType == KernelExecutionType::concurrent;
     int32_t overrideComputeDispatchAllWalkerEnable = debugManager.flags.ComputeDispatchAllWalkerEnableInComputeWalker.get();
     if (overrideComputeDispatchAllWalkerEnable != -1) {

shared/source/xe_hpc_core/command_encoder_xe_hpc_core.cpp

@@ -163,17 +163,6 @@ void EncodeDispatchKernel<Family>::programBarrierEnable(INTERFACE_DESCRIPTOR_DAT
 template <>
 template <typename WalkerType>
 void EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(const RootDeviceEnvironment &rootDeviceEnvironment, WalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs) {
-    const auto &productHelper = rootDeviceEnvironment.getHelper<ProductHelper>();
-    auto &hwInfo = *rootDeviceEnvironment.getHardwareInfo();
-    auto programGlobalFenceAsPostSyncOperationInComputeWalker = productHelper.isGlobalFenceInCommandStreamRequired(hwInfo) &&
-                                                                walkerArgs.requiredSystemFence;
-    int32_t overrideProgramSystemMemoryFence = debugManager.flags.ProgramGlobalFenceAsPostSyncOperationInComputeWalker.get();
-    if (overrideProgramSystemMemoryFence != -1) {
-        programGlobalFenceAsPostSyncOperationInComputeWalker = !!overrideProgramSystemMemoryFence;
-    }
-    auto &postSyncData = walkerCmd.getPostSync();
-    postSyncData.setSystemMemoryFenceRequest(programGlobalFenceAsPostSyncOperationInComputeWalker);
-
     int32_t overrideDispatchAllWalkerEnableInComputeWalker = debugManager.flags.ComputeDispatchAllWalkerEnableInComputeWalker.get();
     if (overrideDispatchAllWalkerEnableInComputeWalker != -1) {
         walkerCmd.setComputeDispatchAllWalkerEnable(overrideDispatchAllWalkerEnableInComputeWalker);

shared/source/xe_hpg_core/command_encoder_xe_hpg_core.cpp

@@ -58,8 +58,11 @@ void EncodeDispatchKernel<Family>::programBarrierEnable(INTERFACE_DESCRIPTOR_DAT
 
 template <>
 template <typename WalkerType>
-void EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(const RootDeviceEnvironment &rootDeviceEnvironment, WalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs) {
+void EncodeDispatchKernel<Family>::encodeAdditionalWalkerFields(const RootDeviceEnvironment &rootDeviceEnvironment, WalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs) {}
-}
+
+template <>
+template <typename WalkerType>
+void EncodeDispatchKernel<Family>::encodeWalkerPostSyncFields(WalkerType &walkerCmd, const EncodeWalkerArgs &walkerArgs) {}
 
 template <>
 void EncodeComputeMode<Family>::programComputeModeCommand(LinearStream &csr, StateComputeModeProperties &properties, const RootDeviceEnvironment &rootDeviceEnvironment) {

shared/test/unit_test/xe_hpc_core/pvc/dispatch_walker_tests_pvc.cpp

@@ -27,7 +27,7 @@ PVCTEST_F(WalkerDispatchTestsPvc, givenPvcWhenEncodeAdditionalWalkerFieldsThenPo
         int32_t programGlobalFenceAsPostSyncOperationInComputeWalker;
         bool expectSystemMemoryFenceRequest;
     } testInputs[] = {
-        {0x0, -1, false},
+        {0x0, -1, true},
         {0x3, -1, true},
         {0x0, 0, false},
         {0x3, 0, false},
@@ -53,7 +53,7 @@ PVCTEST_F(WalkerDispatchTestsPvc, givenPvcWhenEncodeAdditionalWalkerFieldsThenPo
                 testInput.programGlobalFenceAsPostSyncOperationInComputeWalker);
 
             postSyncData.setSystemMemoryFenceRequest(false);
-            EncodeDispatchKernel<FamilyType>::encodeAdditionalWalkerFields(rootDeviceEnvironment, walkerCmd, walkerArgs);
+            EncodeDispatchKernel<FamilyType>::encodeWalkerPostSyncFields(walkerCmd, walkerArgs);
             EXPECT_EQ(testInput.expectSystemMemoryFenceRequest, postSyncData.getSystemMemoryFenceRequest());
         }
     }
@@ -75,7 +75,7 @@ PVCTEST_F(WalkerDispatchTestsPvc, givenPvcSupportsSystemMemoryFenceWhenNoSystemF
         hwInfo.platform.usDeviceID = deviceId;
 
         postSyncData.setSystemMemoryFenceRequest(true);
-        EncodeDispatchKernel<FamilyType>::encodeAdditionalWalkerFields(rootDeviceEnvironment, walkerCmd, walkerArgs);
+        EncodeDispatchKernel<FamilyType>::encodeWalkerPostSyncFields(walkerCmd, walkerArgs);
         EXPECT_FALSE(postSyncData.getSystemMemoryFenceRequest());
     }
 }

shared/test/unit_test/xe_hpc_core/pvc/test_encode_dispatch_kernel_pvc.cpp

@@ -128,39 +128,3 @@ PVCTEST_F(CommandEncodeStatesTestPvc, GivenVariousSlmTotalSizesWhenSetPreferredS
 
     verifyPreferredSlmValues<FamilyType>(valuesToTest, pDevice->getRootDeviceEnvironment());
 }
-
-PVCTEST_F(EncodeKernelPvcTest, givenDefaultSettingForFenceAsPostSyncOperationInComputeWalkerWhenEnqueueKernelIsCalledThenDoNotGenerateFenceCommands) {
-    using DefaultWalkerType = typename FamilyType::DefaultWalkerType;
-    using MI_MEM_FENCE = typename FamilyType::MI_MEM_FENCE;
-
-    DebugManagerStateRestore restore;
-    debugManager.flags.ProgramGlobalFenceAsPostSyncOperationInComputeWalker.set(-1);
-
-    auto &hwInfo = *pDevice->getRootDeviceEnvironment().getMutableHardwareInfo();
-    auto &productHelper = pDevice->getProductHelper();
-
-    hwInfo.platform.usDeviceID = pvcXlDeviceIds[0];
-    VariableBackup<unsigned short> hwRevId{&hwInfo.platform.usRevId};
-    hwRevId = productHelper.getHwRevIdFromStepping(REVISION_A0, hwInfo);
-
-    uint32_t dims[] = {1, 1, 1};
-    std::unique_ptr<MockDispatchKernelEncoder> dispatchInterface(new MockDispatchKernelEncoder());
-    dispatchInterface->getCrossThreadDataSizeResult = 0u;
-
-    bool requiresUncachedMocs = false;
-    EncodeDispatchKernelArgs dispatchArgs = createDefaultDispatchKernelArgs(pDevice, dispatchInterface.get(), dims, requiresUncachedMocs);
-    dispatchArgs.isKernelUsingSystemAllocation = true;
-    dispatchArgs.isHostScopeSignalEvent = true;
-
-    EncodeDispatchKernel<FamilyType>::template encode<DefaultWalkerType>(*cmdContainer.get(), dispatchArgs);
-
-    GenCmdList commands;
-    CmdParse<FamilyType>::parseCommandBuffer(commands, ptrOffset(cmdContainer->getCommandStream()->getCpuBase(), 0), cmdContainer->getCommandStream()->getUsed());
-
-    auto itor = find<DefaultWalkerType *>(commands.begin(), commands.end());
-    ASSERT_NE(itor, commands.end());
-
-    auto walkerCmd = genCmdCast<DefaultWalkerType *>(*itor);
-    auto &postSyncData = walkerCmd->getPostSync();
-    EXPECT_FALSE(postSyncData.getSystemMemoryFenceRequest());
-}