performance: Use lock pointer copy with sfence for dc flush mitigation

Resolves: NEO-12898

Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>

opencl/source/mem_obj/buffer.cpp

@@ -24,6 +24,7 @@
 #include "shared/source/memory_manager/memory_operations_handler.h"
 #include "shared/source/memory_manager/migration_sync_data.h"
 #include "shared/source/os_interface/os_interface.h"
+#include "shared/source/utilities/cpuintrinsics.h"
 
 #include "opencl/source/cl_device/cl_device.h"
 #include "opencl/source/command_queue/command_queue.h"
@@ -212,8 +213,7 @@ bool inline copyHostPointer(Buffer *buffer,
                                 size <= Buffer::maxBufferSizeForCopyOnCpu &&
                                 isCompressionEnabled == false &&
                                 productHelper.getLocalMemoryAccessMode(hwInfo) != LocalMemoryAccessMode::cpuAccessDisallowed &&
-                                isLockable &&
-                                !isGpuCopyRequiredForDcFlushMitigation;
+                                isLockable;
 
         if (debugManager.flags.CopyHostPtrOnCpu.get() != -1) {
             copyOnCpuAllowed = debugManager.flags.CopyHostPtrOnCpu.get() == 1;
@@ -222,6 +222,11 @@ bool inline copyHostPointer(Buffer *buffer,
             memory->setAubWritable(true, GraphicsAllocation::defaultBank);
             memory->setTbxWritable(true, GraphicsAllocation::defaultBank);
             memcpy_s(ptrOffset(lockedPointer, buffer->getOffset()), size, hostPtr, size);
+
+            if (isGpuCopyRequiredForDcFlushMitigation) {
+                CpuIntrinsics::sfence();
+            }
+
             return true;
         } else {
             auto blitMemoryToAllocationResult = BlitOperationResult::unsupported;

opencl/test/unit_test/libult/command_queue_ult.cpp

@@ -20,7 +20,7 @@ bool CommandQueue::isTimestampWaitEnabled() {
 }
 
 bool checkIsGpuCopyRequiredForDcFlushMitigation(AllocationType type) {
-    return false;
+    return ultHwConfig.useGpuCopyForDcFlushMitigation;
 }
 
 } // namespace NEO

opencl/test/unit_test/mem_obj/buffer_tests.cpp

@@ -599,26 +599,22 @@ TEST(Buffer, givenClMemCopyHostPointerPassedToBufferCreateWhenAllocationIsNotInS
     }
 }
 
-TEST(Buffer, givenDcFlushMitigationWhenCreateBufferCopyHostptrThenUseBlitterCopy) {
-    DebugManagerStateRestore restorer;
-    debugManager.flags.AllowDcFlush.set(0);
-    ExecutionEnvironment *executionEnvironment = MockClDevice::prepareExecutionEnvironment(defaultHwInfo.get(), 0u);
-    executionEnvironment->rootDeviceEnvironments[0]->getMutableHardwareInfo()->capabilityTable.blitterOperationsSupported = true;
+namespace CpuIntrinsicsTests {
+extern std::atomic<uint32_t> sfenceCounter;
+} // namespace CpuIntrinsicsTests
 
+TEST(Buffer, givenDcFlushMitigationWhenCreateBufferCopyHostptrThenUseLockPointerCopyWithSfence) {
+    ExecutionEnvironment *executionEnvironment = MockClDevice::prepareExecutionEnvironment(defaultHwInfo.get(), 0u);
     auto productHelper = executionEnvironment->rootDeviceEnvironments[0]->productHelper.get();
-    if (!(productHelper->isBlitterFullySupported(*defaultHwInfo) && productHelper->isDcFlushMitigated())) {
+    if (!productHelper->isDcFlushMitigated()) {
         GTEST_SKIP();
     }
 
-    auto blitterCalled = 0u;
-    auto mockBlitMemoryToAllocation = [&](const NEO::Device &device, NEO::GraphicsAllocation *memory, size_t offset, const void *hostPtr,
-                                          Vec3<size_t> size) -> NEO::BlitOperationResult {
-        memcpy(memory->getUnderlyingBuffer(), hostPtr, size.x);
-        blitterCalled++;
-        return BlitOperationResult::success;
-    };
-    VariableBackup<NEO::BlitHelperFunctions::BlitMemoryToAllocationFunc> blitMemoryToAllocationFuncBackup(
-        &NEO::BlitHelperFunctions::blitMemoryToAllocation, mockBlitMemoryToAllocation);
+    VariableBackup<UltHwConfig> backup(&ultHwConfig);
+    ultHwConfig.useGpuCopyForDcFlushMitigation = true;
+
+    DebugManagerStateRestore restorer;
+    debugManager.flags.AllowDcFlush.set(0);
 
     auto *memoryManager = new MockMemoryManagerFailFirstAllocation(*executionEnvironment);
     executionEnvironment->memoryManager.reset(memoryManager);
@@ -626,6 +622,7 @@ TEST(Buffer, givenDcFlushMitigationWhenCreateBufferCopyHostptrThenUseBlitterCopy
     auto device = std::make_unique<MockClDevice>(MockDevice::create<MockDevice>(executionEnvironment, 0));
 
     MockContext ctx(device.get());
+    CpuIntrinsicsTests::sfenceCounter.store(0u);
 
     cl_int retVal = 0;
     cl_mem_flags flags = CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR;
@@ -634,7 +631,8 @@ TEST(Buffer, givenDcFlushMitigationWhenCreateBufferCopyHostptrThenUseBlitterCopy
     std::unique_ptr<Buffer> buffer(Buffer::create(&ctx, flags, sizeof(memory), memory, retVal));
 
     ASSERT_NE(nullptr, buffer.get());
-    EXPECT_EQ(blitterCalled, 1u);
+    EXPECT_EQ(1u, CpuIntrinsicsTests::sfenceCounter.load());
+    CpuIntrinsicsTests::sfenceCounter.store(0u);
 }
 
 TEST(Buffer, givenPropertiesWithClDeviceHandleListKHRWhenCreateBufferThenCorrectBufferIsSet) {

shared/test/common/base_ult_config_listener.cpp

@@ -50,7 +50,7 @@ void BaseUltConfigListener::OnTestEnd(const ::testing::TestInfo &) {
 
     // Ensure that global state is restored
     UltHwConfig expectedState{};
-    static_assert(sizeof(UltHwConfig) == (16 * sizeof(bool) + sizeof(const char *)), ""); // Ensure that there is no internal padding
+    static_assert(sizeof(UltHwConfig) == (17 * sizeof(bool) + sizeof(const char *)) + sizeof(UltHwConfig::padding), ""); // Ensure that there is no internal padding
     EXPECT_EQ(0, memcmp(&expectedState, &ultHwConfig, sizeof(UltHwConfig)));
 
     EXPECT_EQ(0, memcmp(&referencedHwInfo.platform, &defaultHwInfo->platform, sizeof(PLATFORM)));

shared/test/common/helpers/ult_hw_config.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2020-2023 Intel Corporation
+ * Copyright (C) 2020-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -16,6 +16,7 @@ struct UltHwConfig {
     bool useWaitForTimestamps = false;
     bool useBlitSplit = false;
     bool useFirstSubmissionInitDevice = false;
+    bool useGpuCopyForDcFlushMitigation = false;
 
     bool csrFailInitDirectSubmission = false;
     bool csrBaseCallDirectSubmissionAvailable = false;
@@ -28,6 +29,8 @@ struct UltHwConfig {
     bool csrCreatePreemptionReturnValue = true;
     bool reserved = false;
 
+    char padding[7];
+
     const char *aubTestName = nullptr;
 };