Refactor and modularize walker partition code

Related-To: NEO-6244


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

opencl/test/unit_test/command_queue/CMakeLists.txt

@@ -88,7 +88,6 @@ set(IGDRCL_SRCS_tests_command_queue
     ${CMAKE_CURRENT_SOURCE_DIR}/ooq_task_tests.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/read_write_buffer_cpu_copy.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/sync_buffer_handler_tests.cpp
-    ${CMAKE_CURRENT_SOURCE_DIR}/walker_partition_tests_xehp_and_later.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/work_group_size_tests.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/zero_size_enqueue_tests.cpp
 )
@@ -98,7 +97,10 @@ if(TESTS_XEHP_AND_LATER)
        ${CMAKE_CURRENT_SOURCE_DIR}/dispatch_walker_tests_xehp_and_later.cpp
        ${CMAKE_CURRENT_SOURCE_DIR}/enqueue_media_kernel_xehp_and_later.cpp
        ${CMAKE_CURRENT_SOURCE_DIR}/enqueue_resource_barier_tests_xehp_and_later.cpp
-       ${CMAKE_CURRENT_SOURCE_DIR}/walker_partition_tests_xehp_and_later.cpp
+       ${CMAKE_CURRENT_SOURCE_DIR}/walker_partition_fixture_xehp_and_later.cpp
+       ${CMAKE_CURRENT_SOURCE_DIR}/walker_partition_fixture_xehp_and_later.h
+       ${CMAKE_CURRENT_SOURCE_DIR}/walker_partition_tests_xehp_and_later_1.cpp
+       ${CMAKE_CURRENT_SOURCE_DIR}/walker_partition_tests_xehp_and_later_2.cpp
   )
 endif()
 

opencl/test/unit_test/command_queue/dispatch_walker_tests_xehp_and_later.cpp

@@ -1072,8 +1072,6 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, XeHPAndLaterDispatchWalkerBasicTest, whenWalkerPart
     MockClDevice *device = deviceFactory.rootDevices[0];
     MockContext context{device};
 
-    auto synchronizeBeforeExecution = false;
-    auto staticPartitioning = false;
     auto cmdQ = std::make_unique<MockCommandQueueHw<FamilyType>>(&context, device, nullptr);
     auto &csr = cmdQ->getUltCommandStreamReceiver();
 
@@ -1087,35 +1085,39 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, XeHPAndLaterDispatchWalkerBasicTest, whenWalkerPart
     DispatchInfo dispatchInfo{};
     dispatchInfo.setNumberOfWorkgroups({32, 1, 1});
 
-    synchronizeBeforeExecution = false;
+    WalkerPartition::WalkerPartitionArgs testArgs = {};
+    testArgs.initializeWparidRegister = true;
+    testArgs.crossTileAtomicSynchronization = true;
+    testArgs.usePipeControlStall = true;
+    testArgs.partitionCount = 2u;
+    testArgs.tileCount = static_cast<uint32_t>(device->getDeviceBitfield().count());
+
     DebugManager.flags.SynchronizeWalkerInWparidMode.set(0);
-    staticPartitioning = false;
+    testArgs.staticPartitioning = false;
+    testArgs.synchronizeBeforeExecution = false;
     csr.staticWorkPartitioningEnabled = false;
-    auto partitionSize = WalkerPartition::estimateSpaceRequiredInCommandBuffer<FamilyType>(false, 16u, synchronizeBeforeExecution, false, staticPartitioning, false);
+    auto partitionSize = WalkerPartition::estimateSpaceRequiredInCommandBuffer<FamilyType>(testArgs);
     auto returnedSize = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_NDRANGE_KERNEL, false, false, *cmdQ.get(), kernel->mockKernel, dispatchInfo);
     EXPECT_EQ(returnedSize, partitionSize + baseSize);
 
-    synchronizeBeforeExecution = false;
-    DebugManager.flags.SynchronizeWalkerInWparidMode.set(0);
-    staticPartitioning = true;
+    testArgs.staticPartitioning = true;
     csr.staticWorkPartitioningEnabled = true;
-    partitionSize = WalkerPartition::estimateSpaceRequiredInCommandBuffer<FamilyType>(false, 16u, synchronizeBeforeExecution, false, staticPartitioning, false);
+    partitionSize = WalkerPartition::estimateSpaceRequiredInCommandBuffer<FamilyType>(testArgs);
     returnedSize = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_NDRANGE_KERNEL, false, false, *cmdQ.get(), kernel->mockKernel, dispatchInfo);
     EXPECT_EQ(returnedSize, partitionSize + baseSize);
 
-    synchronizeBeforeExecution = true;
     DebugManager.flags.SynchronizeWalkerInWparidMode.set(1);
-    staticPartitioning = false;
+    testArgs.synchronizeBeforeExecution = true;
+    testArgs.staticPartitioning = false;
     csr.staticWorkPartitioningEnabled = false;
-    partitionSize = WalkerPartition::estimateSpaceRequiredInCommandBuffer<FamilyType>(false, 16u, synchronizeBeforeExecution, false, staticPartitioning, false);
+    partitionSize = WalkerPartition::estimateSpaceRequiredInCommandBuffer<FamilyType>(testArgs);
     returnedSize = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_NDRANGE_KERNEL, false, false, *cmdQ.get(), kernel->mockKernel, dispatchInfo);
     EXPECT_EQ(returnedSize, partitionSize + baseSize);
 
-    synchronizeBeforeExecution = true;
-    DebugManager.flags.SynchronizeWalkerInWparidMode.set(1);
-    staticPartitioning = true;
+    testArgs.synchronizeBeforeExecution = true;
+    testArgs.staticPartitioning = true;
     csr.staticWorkPartitioningEnabled = true;
-    partitionSize = WalkerPartition::estimateSpaceRequiredInCommandBuffer<FamilyType>(false, 16u, synchronizeBeforeExecution, false, staticPartitioning, false);
+    partitionSize = WalkerPartition::estimateSpaceRequiredInCommandBuffer<FamilyType>(testArgs);
     returnedSize = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_NDRANGE_KERNEL, false, false, *cmdQ.get(), kernel->mockKernel, dispatchInfo);
     EXPECT_EQ(returnedSize, partitionSize + baseSize);
 }
@@ -1167,7 +1169,14 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, XeHPAndLaterDispatchWalkerBasicTest, whenQueueIsMul
                     HardwareCommandsHelper<FamilyType>::getSizeRequiredCS() +
                     EncodeMemoryPrefetch<FamilyType>::getSizeForMemoryPrefetch(kernel->kernelInfo.heapInfo.KernelHeapSize);
 
-    auto partitionSize = WalkerPartition::estimateSpaceRequiredInCommandBuffer<FamilyType>(false, 16u, false, false, false, false);
+    WalkerPartition::WalkerPartitionArgs testArgs = {};
+    testArgs.initializeWparidRegister = true;
+    testArgs.usePipeControlStall = true;
+    testArgs.crossTileAtomicSynchronization = true;
+    testArgs.partitionCount = 16u;
+    testArgs.tileCount = static_cast<uint32_t>(device->getDeviceBitfield().count());
+
+    auto partitionSize = WalkerPartition::estimateSpaceRequiredInCommandBuffer<FamilyType>(testArgs);
 
     DispatchInfo dispatchInfo{};
     dispatchInfo.setNumberOfWorkgroups({32, 1, 1});

opencl/test/unit_test/command_queue/walker_partition_fixture_xehp_and_later.cpp

@@ -0,0 +1,25 @@
+/*
+ * Copyright (C) 2021 Intel Corporation
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ */
+
+#include "opencl/test/unit_test/command_queue/walker_partition_fixture_xehp_and_later.h"
+
+void WalkerPartitionTests::SetUp() {
+    cmdBufferAddress = cmdBuffer;
+
+    testArgs.synchronizeBeforeExecution = false;
+    testArgs.nativeCrossTileAtomicSync = false;
+    testArgs.initializeWparidRegister = true;
+    testArgs.usePipeControlStall = true;
+    testArgs.crossTileAtomicSynchronization = true;
+}
+
+void WalkerPartitionTests::TearDown() {
+    auto initialCommandBufferPointer = cmdBuffer;
+    if (checkForProperCmdBufferAddressOffset) {
+        EXPECT_EQ(ptrDiff(cmdBufferAddress, initialCommandBufferPointer), totalBytesProgrammed);
+    }
+}

opencl/test/unit_test/command_queue/walker_partition_fixture_xehp_and_later.h

@@ -0,0 +1,38 @@
+/*
+ * Copyright (C) 2021 Intel Corporation
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ */
+
+#include "shared/source/command_container/walker_partition_xehp_and_later.h"
+#include "shared/test/common/cmd_parse/gen_cmd_parse.h"
+#include "shared/test/common/helpers/debug_manager_state_restore.h"
+#include "shared/test/common/helpers/unit_test_helper.h"
+
+#include "test.h"
+
+using namespace WalkerPartition;
+
+struct WalkerPartitionTests : public ::testing::Test {
+    void SetUp() override;
+
+    void TearDown() override;
+
+    template <typename GfxFamily>
+    auto createWalker(uint64_t postSyncAddress) {
+        WalkerPartition::COMPUTE_WALKER<GfxFamily> walker;
+        walker = GfxFamily::cmdInitGpgpuWalker;
+        walker.setPartitionType(COMPUTE_WALKER<GfxFamily>::PARTITION_TYPE::PARTITION_TYPE_X);
+        auto &postSync = walker.getPostSync();
+        postSync.setOperation(POSTSYNC_DATA<GfxFamily>::OPERATION::OPERATION_WRITE_TIMESTAMP);
+        postSync.setDestinationAddress(postSyncAddress);
+        return walker;
+    }
+
+    char cmdBuffer[4096u];
+    WalkerPartition::WalkerPartitionArgs testArgs = {};
+    void *cmdBufferAddress = nullptr;
+    uint32_t totalBytesProgrammed = 0u;
+    bool checkForProperCmdBufferAddressOffset = true;
+};

opencl/test/unit_test/command_stream/implicit_scaling_ocl_tests.cpp

@@ -14,3 +14,11 @@ using namespace NEO;
 TEST(ImplicitScalingApiTests, givenOpenClApiUsedThenSupportEnabled) {
     EXPECT_TRUE(ImplicitScaling::apiSupport);
 }
+
+TEST(ImplicitScalingApiTests, givenOpenClApiUsedThenSemaphoreProgrammingRequiredIsFalse) {
+    EXPECT_FALSE(ImplicitScaling::semaphoreProgrammingRequired);
+}
+
+TEST(ImplicitScalingApiTests, givenOpenClApiUsedThenCrossTileAtomicSynchronization) {
+    EXPECT_TRUE(ImplicitScaling::crossTileAtomicSynchronization);
+}

opencl/test/unit_test/test_files/igdrcl.config

@@ -266,8 +266,8 @@ ForceWorkgroupSize1x1x1 = -1
 ForceThreadGroupDispatchSize = -1
 ForceStatelessL1CachingPolicy = -1
 ForceMemoryBankIndexOverride = -1
-ExperimentalSynchronizeWithSemaphores = -1
-ExperimentalForceCrossAtomicSynchronization = -1
+SynchronizeWithSemaphores = -1
+UseCrossAtomicSynchronization = -1
 EnableStatelessCompression = -1
 EnableMultiTileCompression = -1
 EnablePrivateScratchSlot1 = -1
@@ -313,7 +313,7 @@ OverrideUseKmdWaitFunction = -1
 EnableCacheFlushAfterWalkerForAllQueues = -1
 Force32BitDriverSupport = -1
 OverrideCmdQueueSynchronousMode = -1
-ExperimentalUseAtomicsForNativeSectionCleanup = -1
+UseAtomicsForNativeSectionCleanup = -1
 HBMSizePerTileInGigabytes = 0
 OverrideSystolicPipelineSelect = -1
 OverrideSystolicInComputeWalker = -1
@@ -324,6 +324,9 @@ DoNotFreeResources = 0
 OverrideGmmResourceUsageField = -1
 LogAllocationType = 0
 ProgramAdditionalPipeControlBeforeStateComputeModeCommand = 0
+ProgramNativeCleanup = -1
+WparidRegisterProgramming = -1
+UsePipeControlAfterPartitionedWalker = -1
 OverrideBufferSuitableForRenderCompression = -1
 AllowMixingRegularAndCooperativeKernels = 0
 AllowPatchingVfeStateInCommandLists = 0