From 6f58ec9d2c45843bbac0efa7ebebaa2252a82925 Mon Sep 17 00:00:00 2001
From: "Mrozek, Michal" <michal.mrozek@intel.com>
Date: Wed, 4 Oct 2023 15:55:40 +0000
Subject: [PATCH] fix: set walker params prior to adjusting interface
 descriptor

Signed-off-by: Mrozek, Michal <michal.mrozek@intel.com>
---
 .../hardware_interface_bdw_and_later.inl      | 10 +++----
 .../hardware_interface_xehp_and_later.inl     |  7 ++---
 .../xe_hpc_core/enqueue_tests_xe_hpc_core.cpp | 27 +++++++++++++++++++
 3 files changed, 36 insertions(+), 8 deletions(-)

diff --git a/opencl/source/command_queue/hardware_interface_bdw_and_later.inl b/opencl/source/command_queue/hardware_interface_bdw_and_later.inl
index 9d55cb60c5..91971f3059 100644
--- a/opencl/source/command_queue/hardware_interface_bdw_and_later.inl
+++ b/opencl/source/command_queue/hardware_interface_bdw_and_later.inl
@@ -77,6 +77,11 @@ inline void HardwareInterface<GfxFamily>::programWalker(
     auto isCcsUsed = EngineHelpers::isCcs(commandQueue.getGpgpuEngine().osContext->getEngineType());
     auto kernelUsesLocalIds = HardwareCommandsHelper<GfxFamily>::kernelUsesLocalIds(kernel);
 
+    GpgpuWalkerHelper<GfxFamily>::setGpgpuWalkerThreadData(&walkerCmd, kernel.getKernelInfo().kernelDescriptor,
+                                                           globalOffsets, startWorkGroups,
+                                                           numWorkGroups, walkerArgs.localWorkSizes, simd, dim,
+                                                           false, false, 0u);
+
     HardwareCommandsHelper<GfxFamily>::sendIndirectState(
         commandStream,
         dsh,
@@ -95,11 +100,6 @@ inline void HardwareInterface<GfxFamily>::programWalker(
         kernelUsesLocalIds,
         commandQueue.getDevice());
 
-    GpgpuWalkerHelper<GfxFamily>::setGpgpuWalkerThreadData(&walkerCmd, kernel.getKernelInfo().kernelDescriptor,
-                                                           globalOffsets, startWorkGroups,
-                                                           numWorkGroups, walkerArgs.localWorkSizes, simd, dim,
-                                                           false, false, 0u);
-
     EncodeWalkerArgs encodeWalkerArgs{kernel.getExecutionType(), false, kernel.getKernelInfo().kernelDescriptor};
     EncodeDispatchKernel<GfxFamily>::encodeAdditionalWalkerFields(rootDeviceEnvironment, walkerCmd, encodeWalkerArgs);
     *walkerCmdBuf = walkerCmd;
diff --git a/opencl/source/command_queue/hardware_interface_xehp_and_later.inl b/opencl/source/command_queue/hardware_interface_xehp_and_later.inl
index 3315b70b3a..164ace4151 100644
--- a/opencl/source/command_queue/hardware_interface_xehp_and_later.inl
+++ b/opencl/source/command_queue/hardware_interface_xehp_and_later.inl
@@ -96,6 +96,10 @@ inline void HardwareInterface<GfxFamily>::programWalker(
         EncodeMemoryPrefetch<GfxFamily>::programMemoryPrefetch(commandStream, *kernelAllocation, kernelInfo.heapInfo.kernelHeapSize, 0, rootDeviceEnvironment);
     }
 
+    GpgpuWalkerHelper<GfxFamily>::setGpgpuWalkerThreadData(&walkerCmd, kernelInfo.kernelDescriptor, globalOffsets, startWorkGroups,
+                                                           numWorkGroups, walkerArgs.localWorkSizes, simd, dim,
+                                                           localIdsGenerationByRuntime, inlineDataProgrammingRequired, requiredWalkOrder);
+
     HardwareCommandsHelper<GfxFamily>::sendIndirectState(
         commandStream,
         dsh,
@@ -114,9 +118,6 @@ inline void HardwareInterface<GfxFamily>::programWalker(
         localIdsGenerationByRuntime,
         commandQueue.getDevice());
 
-    GpgpuWalkerHelper<GfxFamily>::setGpgpuWalkerThreadData(&walkerCmd, kernelInfo.kernelDescriptor, globalOffsets, startWorkGroups,
-                                                           numWorkGroups, walkerArgs.localWorkSizes, simd, dim,
-                                                           localIdsGenerationByRuntime, inlineDataProgrammingRequired, requiredWalkOrder);
     bool kernelSystemAllocation = false;
     if (kernel.isBuiltIn) {
         kernelSystemAllocation = kernel.getDestinationAllocationInSystemMemory();
diff --git a/opencl/test/unit_test/xe_hpc_core/enqueue_tests_xe_hpc_core.cpp b/opencl/test/unit_test/xe_hpc_core/enqueue_tests_xe_hpc_core.cpp
index e2db11771f..7878256eb4 100644
--- a/opencl/test/unit_test/xe_hpc_core/enqueue_tests_xe_hpc_core.cpp
+++ b/opencl/test/unit_test/xe_hpc_core/enqueue_tests_xe_hpc_core.cpp
@@ -103,6 +103,33 @@ XE_HPC_CORETEST_F(MemoryPrefetchTestsXeHpcCore, givenPrefetchEnabledWhenEstimati
 
 using ProgramWalkerTestsXeHpcCore = EnqueueFixtureXeHpcCore;
 
+XE_HPC_CORETEST_F(ProgramWalkerTestsXeHpcCore, givenProperThreadGroupSizesWhenWalkerIsProgrammedThenThreadGroupDispatchSizeIsProperlyProgrammed) {
+    using COMPUTE_WALKER = typename FamilyType::COMPUTE_WALKER;
+
+    auto commandQueue = createCommandQueue<FamilyType>();
+    auto &commandStream = commandQueue->getCS(1024);
+
+    auto &heap = commandQueue->getIndirectHeap(IndirectHeap::Type::DYNAMIC_STATE, 1);
+    size_t workSize[] = {1, 1, 1};
+    Vec3<size_t> wgInfo = {1024, 1, 1};
+
+    HardwareInterfaceWalkerArgs walkerArgs = createHardwareInterfaceWalkerArgs(workSize, wgInfo, PreemptionMode::Disabled);
+
+    {
+        HardwareInterface<FamilyType>::programWalker(commandStream, *mockKernel->mockKernel, *commandQueue,
+                                                     heap, heap, heap, dispatchInfo, walkerArgs);
+        HardwareParse hwParse;
+        hwParse.parseCommands<FamilyType>(commandStream, 0);
+        auto itorWalker = find<COMPUTE_WALKER *>(hwParse.cmdList.begin(), hwParse.cmdList.end());
+        EXPECT_NE(hwParse.cmdList.end(), itorWalker);
+        auto walkerCmd = genCmdCast<COMPUTE_WALKER *>(*itorWalker);
+        EXPECT_NE(nullptr, walkerCmd);
+        auto &idd = walkerCmd->getInterfaceDescriptor();
+
+        EXPECT_EQ(FamilyType::INTERFACE_DESCRIPTOR_DATA::THREAD_GROUP_DISPATCH_SIZE_TG_SIZE_8, idd.getThreadGroupDispatchSize());
+    }
+}
+
 XE_HPC_CORETEST_F(ProgramWalkerTestsXeHpcCore, givenDebugVariableSetWhenProgrammingWalkerThenSetL3Prefetch) {
     using COMPUTE_WALKER = typename FamilyType::COMPUTE_WALKER;