performance: do not enable v2 algorithm in certain scenarios

- do not enable when SubSliceCount != MaxSubSlicesSupported

Signed-off-by: Michal Mrozek <michal.mrozek@intel.com>
Related-To: NEO-6989
Source: a129c29

Author: MichalMrozek

shared/source/xe_hpc_core/command_encoder_xe_hpc_core.cpp

@@ -43,7 +43,8 @@ void EncodeDispatchKernel<Family>::adjustInterfaceDescriptorData(InterfaceDescri
         if (DebugManager.flags.AdjustThreadGroupDispatchSize.get() != -1) {
             adjustTGDispatchSize = !!DebugManager.flags.AdjustThreadGroupDispatchSize.get();
         }
-        auto algorithmVersion = 2u;
+        // apply v2 algorithm only for parts where MaxSubSlicesSupported is equal to SubSliceCount
+        auto algorithmVersion = hwInfo.gtSystemInfo.MaxSubSlicesSupported == hwInfo.gtSystemInfo.SubSliceCount ? 2 : 1;
         if (DebugManager.flags.ForceThreadGroupDispatchSizeAlgorithm.get() != -1) {
             algorithmVersion = DebugManager.flags.ForceThreadGroupDispatchSizeAlgorithm.get();
         }

shared/test/unit_test/xe_hpc_core/test_encode_xe_hpc_core.cpp

@@ -819,6 +819,31 @@ XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenVariousDispatchParamtersWhenAl
     EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_2, iddArg.getThreadGroupDispatchSize());
 }
 
+XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenDualSubSliceCountNotEqualToMaxSubsliceCounteWhenTgDispatchSizeIsSelectedThenAlgorithmV1IsUsed) {
+    using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
+    using WALKER_TYPE = typename FamilyType::WALKER_TYPE;
+    WALKER_TYPE walkerCmd{};
+    const auto &productHelper = pDevice->getProductHelper();
+    auto mutableHwInfo = pDevice->getRootDeviceEnvironment().getMutableHardwareInfo();
+    mutableHwInfo->gtSystemInfo.MaxSubSlicesSupported = 64u;
+    mutableHwInfo->gtSystemInfo.SubSliceCount = 32u;
+    mutableHwInfo->gtSystemInfo.ThreadCount = 2048u;
+    auto hwInfo = pDevice->getHardwareInfo();
+
+    hwInfo.platform.usRevId = productHelper.getHwRevIdFromStepping(REVISION_B, hwInfo);
+    uint32_t numGrf = GrfConfig::DefaultGrfNumber;
+
+    INTERFACE_DESCRIPTOR_DATA iddArg = FamilyType::cmdInitInterfaceDescriptorData;
+
+    iddArg.setNumberOfThreadsInGpgpuThreadGroup(1u);
+    numGrf = GrfConfig::DefaultGrfNumber;
+    walkerCmd.setThreadGroupIdXDimension(256);
+    walkerCmd.setThreadGroupIdYDimension(1);
+    walkerCmd.setThreadGroupIdZDimension(1);
+    EncodeDispatchKernel<FamilyType>::adjustInterfaceDescriptorData(iddArg, *pDevice, hwInfo, 256u, numGrf, walkerCmd);
+    EXPECT_EQ(INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_1, iddArg.getThreadGroupDispatchSize());
+}
+
 XE_HPC_CORETEST_F(EncodeKernelXeHpcCoreTest, givenNumberOfThreadsInThreadGroupAndDebugFlagDisabledWhenCallingAdjustInterfaceDescriptorDataThenThreadGroupDispatchSizeIsDefault) {
     using INTERFACE_DESCRIPTOR_DATA = typename FamilyType::INTERFACE_DESCRIPTOR_DATA;
     using WALKER_TYPE = typename FamilyType::WALKER_TYPE;