Add support for zero-copy r/w buffer

Change-Id: Ie9f3f2211d107eb338bd97692d36e9c7d7a0feab
2025-12-25 21:42:53 +08:00 · 2018-01-11 15:42:55 +01:00
parent 79e4b2c104
commit 2c2bbbcdbb
13 changed files with 694 additions and 17 deletions
--- a/unit_tests/command_queue/enqueue_read_buffer_event_tests.cpp
+++ b/unit_tests/command_queue/enqueue_read_buffer_event_tests.cpp
@@ -22,6 +22,7 @@

 #include "runtime/command_queue/command_queue.h"
 #include "runtime/event/event.h"
+#include "unit_tests/helpers/debug_manager_state_restore.h"
 #include "unit_tests/fixtures/hello_world_fixture.h"
 #include "unit_tests/fixtures/buffer_fixture.h"
 #include "gtest/gtest.h"
@@ -108,3 +109,165 @@ TEST_F(EnqueueReadBuffer, eventReturnedShouldBeMaxOfInputEventsAndCmdQPlus1) {

    delete pEvent;
 }
+TEST_F(EnqueueReadBuffer, givenInOrderQueueAndEnabledSupportCpuCopiesAndDstPtrEqualSrcPtrWithEventsWhenReadBufferIsExecutedThenTaskLevelShouldNotBeIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnReadBuffer.set(true);
+    cl_int retVal = CL_SUCCESS;
+    uint32_t taskLevelCmdQ = 17;
+    pCmdQ->taskLevel = taskLevelCmdQ;
+
+    uint32_t taskLevelEvent1 = 8;
+    uint32_t taskLevelEvent2 = 19;
+    Event event1(pCmdQ, CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent1, 4);
+    Event event2(pCmdQ, CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent2, 10);
+
+    cl_bool blockingRead = CL_TRUE;
+    size_t size = sizeof(cl_float);
+    cl_event eventWaitList[] =
+        {
+            &event1,
+            &event2};
+    cl_uint numEventsInWaitList = sizeof(eventWaitList) / sizeof(eventWaitList[0]);
+    cl_event event = nullptr;
+    auto srcBuffer = std::unique_ptr<Buffer>(BufferHelper<>::create());
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    retVal = pCmdQ->enqueueReadBuffer(srcBuffer.get(),
+                                      blockingRead,
+                                      0,
+                                      size,
+                                      ptr,
+                                      numEventsInWaitList,
+                                      eventWaitList,
+                                      &event);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    ASSERT_NE(nullptr, event);
+
+    auto pEvent = (Event *)event;
+    EXPECT_EQ(19u, pEvent->taskLevel);
+    EXPECT_EQ(17u, pCmdQ->taskLevel);
+
+    pEvent->release();
+}
+TEST_F(EnqueueReadBuffer, givenOutOfOrderQueueAndEnabledSupportCpuCopiesAndDstPtrEqualSrcPtrWithEventsWhenReadBufferIsExecutedThenTaskLevelShouldNotBeIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnReadBuffer.set(true);
+    std::unique_ptr<CommandQueue> pCmdOOQ(createCommandQueue(pDevice, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE));
+    cl_int retVal = CL_SUCCESS;
+    uint32_t taskLevelCmdQ = 17;
+    pCmdOOQ->taskLevel = taskLevelCmdQ;
+
+    uint32_t taskLevelEvent1 = 8;
+    uint32_t taskLevelEvent2 = 19;
+    Event event1(pCmdOOQ.get(), CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent1, 4);
+    Event event2(pCmdOOQ.get(), CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent2, 10);
+
+    cl_bool blockingRead = CL_TRUE;
+    size_t size = sizeof(cl_float);
+    cl_event eventWaitList[] =
+        {
+            &event1,
+            &event2};
+    cl_uint numEventsInWaitList = sizeof(eventWaitList) / sizeof(eventWaitList[0]);
+    cl_event event = nullptr;
+    auto srcBuffer = std::unique_ptr<Buffer>(BufferHelper<>::create());
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    retVal = pCmdOOQ->enqueueReadBuffer(srcBuffer.get(),
+                                        blockingRead,
+                                        0,
+                                        size,
+                                        ptr,
+                                        numEventsInWaitList,
+                                        eventWaitList,
+                                        &event);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    ASSERT_NE(nullptr, event);
+
+    auto pEvent = (Event *)event;
+    EXPECT_EQ(19u, pEvent->taskLevel);
+    EXPECT_EQ(17u, pCmdOOQ->taskLevel);
+
+    pEvent->release();
+}
+TEST_F(EnqueueReadBuffer, givenInOrderQueueAndDisabledSupportCpuCopiesAndDstPtrEqualSrcPtrWithEventsWhenReadBufferIsExecutedThenTaskLevelShouldNotBeIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnReadBuffer.set(false);
+    cl_int retVal = CL_SUCCESS;
+    uint32_t taskLevelCmdQ = 17;
+    pCmdQ->taskLevel = taskLevelCmdQ;
+
+    uint32_t taskLevelEvent1 = 8;
+    uint32_t taskLevelEvent2 = 19;
+    Event event1(pCmdQ, CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent1, 4);
+    Event event2(pCmdQ, CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent2, 10);
+
+    cl_bool blockingRead = CL_TRUE;
+    size_t size = sizeof(cl_float);
+    cl_event eventWaitList[] =
+        {
+            &event1,
+            &event2};
+    cl_uint numEventsInWaitList = sizeof(eventWaitList) / sizeof(eventWaitList[0]);
+    cl_event event = nullptr;
+    auto srcBuffer = std::unique_ptr<Buffer>(BufferHelper<>::create());
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    retVal = pCmdQ->enqueueReadBuffer(srcBuffer.get(),
+                                      blockingRead,
+                                      0,
+                                      size,
+                                      ptr,
+                                      numEventsInWaitList,
+                                      eventWaitList,
+                                      &event);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    ASSERT_NE(nullptr, event);
+
+    auto pEvent = (Event *)event;
+    EXPECT_EQ(19u, pEvent->taskLevel);
+    EXPECT_EQ(19u, pCmdQ->taskLevel);
+
+    pEvent->release();
+}
+TEST_F(EnqueueReadBuffer, givenOutOfOrderQueueAndDisabledSupportCpuCopiesAndDstPtrEqualSrcPtrWithEventsWhenReadBufferIsExecutedThenTaskLevelShouldNotBeIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnReadBuffer.set(false);
+    std::unique_ptr<CommandQueue> pCmdOOQ(createCommandQueue(pDevice, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE));
+    cl_int retVal = CL_SUCCESS;
+    uint32_t taskLevelCmdQ = 17;
+    pCmdOOQ->taskLevel = taskLevelCmdQ;
+
+    uint32_t taskLevelEvent1 = 8;
+    uint32_t taskLevelEvent2 = 19;
+    Event event1(pCmdOOQ.get(), CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent1, 4);
+    Event event2(pCmdOOQ.get(), CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent2, 10);
+
+    cl_bool blockingRead = CL_TRUE;
+    size_t size = sizeof(cl_float);
+    cl_event eventWaitList[] =
+        {
+            &event1,
+            &event2};
+    cl_uint numEventsInWaitList = sizeof(eventWaitList) / sizeof(eventWaitList[0]);
+    cl_event event = nullptr;
+    auto srcBuffer = std::unique_ptr<Buffer>(BufferHelper<>::create());
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    retVal = pCmdOOQ->enqueueReadBuffer(srcBuffer.get(),
+                                        blockingRead,
+                                        0,
+                                        size,
+                                        ptr,
+                                        numEventsInWaitList,
+                                        eventWaitList,
+                                        &event);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    ASSERT_NE(nullptr, event);
+
+    auto pEvent = (Event *)event;
+    EXPECT_EQ(19u, pEvent->taskLevel);
+    EXPECT_EQ(19u, pCmdOOQ->taskLevel);
+
+    pEvent->release();
+}
--- a/unit_tests/command_queue/enqueue_read_buffer_tests.cpp
+++ b/unit_tests/command_queue/enqueue_read_buffer_tests.cpp
@@ -26,6 +26,7 @@
 #include "runtime/helpers/dispatch_info.h"
 #include "unit_tests/command_queue/enqueue_fixture.h"
 #include "unit_tests/command_queue/enqueue_read_buffer_fixture.h"
+#include "unit_tests/helpers/debug_manager_state_restore.h"
 #include "test.h"

 using namespace OCLRT;
@@ -406,3 +407,78 @@ HWTEST_F(EnqueueReadBufferTypeTest, givenNotAlignedPointerAndAlignedSizeWhenRead
    EXPECT_EQ(CacheSettings::l3CacheOn, csr.latestSentStatelessMocsConfig);
    EXPECT_FALSE(csr.disableL3Cache);
 }
+
+HWTEST_F(EnqueueReadBufferTypeTest, givenOOQWithEnabledSupportCpuCopiesAndDstPtrEqualSrcPtrWhenReadBufferIsExecutedThenTaskLevelNotIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnReadBuffer.set(true);
+    cl_int retVal = CL_SUCCESS;
+    std::unique_ptr<CommandQueue> pCmdOOQ(createCommandQueue(pDevice, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE));
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    EXPECT_EQ(retVal, CL_SUCCESS);
+    retVal = pCmdOOQ->enqueueReadBuffer(srcBuffer,
+                                        CL_FALSE,
+                                        0,
+                                        MemoryConstants::cacheLineSize,
+                                        ptr,
+                                        0,
+                                        nullptr,
+                                        nullptr);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    EXPECT_EQ(pCmdOOQ->taskLevel, 0u);
+}
+HWTEST_F(EnqueueReadBufferTypeTest, givenOOQWithDisabledSupportCpuCopiesAndDstPtrEqualSrcPtrWhenReadBufferIsExecutedThenTaskLevelNotIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnReadBuffer.set(false);
+    cl_int retVal = CL_SUCCESS;
+    std::unique_ptr<CommandQueue> pCmdOOQ(createCommandQueue(pDevice, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE));
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    EXPECT_EQ(retVal, CL_SUCCESS);
+    retVal = pCmdOOQ->enqueueReadBuffer(srcBuffer,
+                                        CL_FALSE,
+                                        0,
+                                        MemoryConstants::cacheLineSize,
+                                        ptr,
+                                        0,
+                                        nullptr,
+                                        nullptr);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    EXPECT_EQ(pCmdOOQ->taskLevel, 0u);
+}
+HWTEST_F(EnqueueReadBufferTypeTest, givenInOrderQueueAndEnabledSupportCpuCopiesAndDstPtrEqualSrcPtrWhenReadBufferIsExecutedThenTaskLevelShouldNotBeIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnReadBuffer.set(true);
+    cl_int retVal = CL_SUCCESS;
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    EXPECT_EQ(retVal, CL_SUCCESS);
+    retVal = pCmdQ->enqueueReadBuffer(srcBuffer,
+                                      CL_FALSE,
+                                      0,
+                                      MemoryConstants::cacheLineSize,
+                                      ptr,
+                                      0,
+                                      nullptr,
+                                      nullptr);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    EXPECT_EQ(pCmdQ->taskLevel, 0u);
+}
+HWTEST_F(EnqueueReadBufferTypeTest, givenInOrderQueueAndDisabledSupportCpuCopiesAndDstPtrEqualSrcPtrWhenReadBufferIsExecutedThenTaskLevelShouldNotBeIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnReadBuffer.set(false);
+    cl_int retVal = CL_SUCCESS;
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    EXPECT_EQ(retVal, CL_SUCCESS);
+    retVal = pCmdQ->enqueueReadBuffer(srcBuffer,
+                                      CL_FALSE,
+                                      0,
+                                      MemoryConstants::cacheLineSize,
+                                      ptr,
+                                      0,
+                                      nullptr,
+                                      nullptr);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    EXPECT_EQ(pCmdQ->taskLevel, 0u);
+}
--- a/unit_tests/command_queue/enqueue_write_buffer_event_tests.cpp
+++ b/unit_tests/command_queue/enqueue_write_buffer_event_tests.cpp
@@ -23,6 +23,7 @@
 #include "runtime/command_queue/command_queue.h"
 #include "runtime/event/event.h"
 #include "unit_tests/command_queue/buffer_operations_fixture.h"
+#include "unit_tests/helpers/debug_manager_state_restore.h"
 #include "unit_tests/fixtures/buffer_fixture.h"
 #include "gtest/gtest.h"
 #include <memory>
@@ -107,4 +108,167 @@ TEST_F(EnqueueWriteBufferTypeTest, eventReturnedShouldBeMaxOfInputEventsAndCmdQP
    EXPECT_LE(19u, pEvent->taskLevel);

    delete pEvent;
+}
+
+TEST_F(EnqueueWriteBufferTypeTest, givenInOrderQueueAndEnabledSupportCpuCopiesAndDstPtrEqualSrcPtrWithEventsWhenWriteBufferIsExecutedThenTaskLevelShouldNotBeIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnWriteBuffer.set(true);
+    cl_int retVal = CL_SUCCESS;
+    uint32_t taskLevelCmdQ = 17;
+    pCmdQ->taskLevel = taskLevelCmdQ;
+
+    uint32_t taskLevelEvent1 = 8;
+    uint32_t taskLevelEvent2 = 19;
+    Event event1(pCmdQ, CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent1, 4);
+    Event event2(pCmdQ, CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent2, 10);
+
+    cl_bool blockingRead = CL_TRUE;
+    size_t size = sizeof(cl_float);
+    cl_event eventWaitList[] =
+        {
+            &event1,
+            &event2};
+    cl_uint numEventsInWaitList = sizeof(eventWaitList) / sizeof(eventWaitList[0]);
+    cl_event event = nullptr;
+    auto srcBuffer = std::unique_ptr<Buffer>(BufferHelper<>::create());
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    retVal = pCmdQ->enqueueWriteBuffer(srcBuffer.get(),
+                                       blockingRead,
+                                       0,
+                                       size,
+                                       ptr,
+                                       numEventsInWaitList,
+                                       eventWaitList,
+                                       &event);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    ASSERT_NE(nullptr, event);
+
+    auto pEvent = (Event *)event;
+    EXPECT_EQ(19u, pEvent->taskLevel);
+    EXPECT_EQ(17u, pCmdQ->taskLevel);
+
+    pEvent->release();
+}
+TEST_F(EnqueueWriteBufferTypeTest, givenOutOfOrderQueueAndEnabledSupportCpuCopiesAndDstPtrEqualSrcPtrWithEventsWhenWriteBufferIsExecutedThenTaskLevelShouldNotBeIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnWriteBuffer.set(true);
+    std::unique_ptr<CommandQueue> pCmdOOQ(createCommandQueue(pDevice, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE));
+    cl_int retVal = CL_SUCCESS;
+    uint32_t taskLevelCmdQ = 17;
+    pCmdOOQ->taskLevel = taskLevelCmdQ;
+
+    uint32_t taskLevelEvent1 = 8;
+    uint32_t taskLevelEvent2 = 19;
+    Event event1(pCmdOOQ.get(), CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent1, 4);
+    Event event2(pCmdOOQ.get(), CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent2, 10);
+
+    cl_bool blockingRead = CL_TRUE;
+    size_t size = sizeof(cl_float);
+    cl_event eventWaitList[] =
+        {
+            &event1,
+            &event2};
+    cl_uint numEventsInWaitList = sizeof(eventWaitList) / sizeof(eventWaitList[0]);
+    cl_event event = nullptr;
+    auto srcBuffer = std::unique_ptr<Buffer>(BufferHelper<>::create());
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    retVal = pCmdOOQ->enqueueWriteBuffer(srcBuffer.get(),
+                                         blockingRead,
+                                         0,
+                                         size,
+                                         ptr,
+                                         numEventsInWaitList,
+                                         eventWaitList,
+                                         &event);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    ASSERT_NE(nullptr, event);
+
+    auto pEvent = (Event *)event;
+    EXPECT_EQ(19u, pEvent->taskLevel);
+    EXPECT_EQ(17u, pCmdOOQ->taskLevel);
+
+    pEvent->release();
+}
+TEST_F(EnqueueWriteBufferTypeTest, givenInOrderQueueAndDisabledSupportCpuCopiesAndDstPtrEqualSrcPtrWithEventsWhenWriteBufferIsExecutedThenTaskLevelShouldNotBeIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnWriteBuffer.set(false);
+    cl_int retVal = CL_SUCCESS;
+    uint32_t taskLevelCmdQ = 17;
+    pCmdQ->taskLevel = taskLevelCmdQ;
+
+    uint32_t taskLevelEvent1 = 8;
+    uint32_t taskLevelEvent2 = 19;
+    Event event1(pCmdQ, CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent1, 4);
+    Event event2(pCmdQ, CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent2, 10);
+
+    cl_bool blockingRead = CL_TRUE;
+    size_t size = sizeof(cl_float);
+    cl_event eventWaitList[] =
+        {
+            &event1,
+            &event2};
+    cl_uint numEventsInWaitList = sizeof(eventWaitList) / sizeof(eventWaitList[0]);
+    cl_event event = nullptr;
+    auto srcBuffer = std::unique_ptr<Buffer>(BufferHelper<>::create());
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    retVal = pCmdQ->enqueueWriteBuffer(srcBuffer.get(),
+                                       blockingRead,
+                                       0,
+                                       size,
+                                       ptr,
+                                       numEventsInWaitList,
+                                       eventWaitList,
+                                       &event);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    ASSERT_NE(nullptr, event);
+
+    auto pEvent = (Event *)event;
+    EXPECT_EQ(19u, pEvent->taskLevel);
+    EXPECT_EQ(19u, pCmdQ->taskLevel);
+
+    pEvent->release();
+}
+TEST_F(EnqueueWriteBufferTypeTest, givenOutOfOrderQueueAndDisabledSupportCpuCopiesAndDstPtrEqualSrcPtrWithEventsWhenWriteBufferIsExecutedThenTaskLevelShouldNotBeIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnWriteBuffer.set(false);
+    std::unique_ptr<CommandQueue> pCmdOOQ(createCommandQueue(pDevice, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE));
+    cl_int retVal = CL_SUCCESS;
+    uint32_t taskLevelCmdQ = 17;
+    pCmdOOQ->taskLevel = taskLevelCmdQ;
+
+    uint32_t taskLevelEvent1 = 8;
+    uint32_t taskLevelEvent2 = 19;
+    Event event1(pCmdOOQ.get(), CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent1, 4);
+    Event event2(pCmdOOQ.get(), CL_COMMAND_NDRANGE_KERNEL, taskLevelEvent2, 10);
+
+    cl_bool blockingRead = CL_TRUE;
+    size_t size = sizeof(cl_float);
+    cl_event eventWaitList[] =
+        {
+            &event1,
+            &event2};
+    cl_uint numEventsInWaitList = sizeof(eventWaitList) / sizeof(eventWaitList[0]);
+    cl_event event = nullptr;
+    auto srcBuffer = std::unique_ptr<Buffer>(BufferHelper<>::create());
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    retVal = pCmdOOQ->enqueueWriteBuffer(srcBuffer.get(),
+                                         blockingRead,
+                                         0,
+                                         size,
+                                         ptr,
+                                         numEventsInWaitList,
+                                         eventWaitList,
+                                         &event);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    ASSERT_NE(nullptr, event);
+
+    auto pEvent = (Event *)event;
+    EXPECT_EQ(19u, pEvent->taskLevel);
+    EXPECT_EQ(19u, pCmdOOQ->taskLevel);
+
+    pEvent->release();
 }
--- a/unit_tests/command_queue/enqueue_write_buffer_tests.cpp
+++ b/unit_tests/command_queue/enqueue_write_buffer_tests.cpp
@@ -25,6 +25,7 @@
 #include "runtime/gen_common/reg_configs.h"
 #include "runtime/helpers/dispatch_info.h"
 #include "unit_tests/command_queue/enqueue_fixture.h"
+#include "unit_tests/helpers/debug_manager_state_restore.h"
 #include "test.h"

 using namespace OCLRT;
@@ -325,3 +326,77 @@ HWTEST_F(EnqueueWriteBufferTypeTest, MediaVFEState) {
    // Generically validate this command
    FamilyType::PARSE::template validateCommand<MEDIA_VFE_STATE *>(cmdList.begin(), itorCmd);
 }
+HWTEST_F(EnqueueWriteBufferTypeTest, givenOOQWithEnabledSupportCpuCopiesAndDstPtrEqualSrcPtrWhenWriteBufferIsExecutedThenTaskLevelNotIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnWriteBuffer.set(true);
+    cl_int retVal = CL_SUCCESS;
+    std::unique_ptr<CommandQueue> pCmdOOQ(createCommandQueue(pDevice, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE));
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    EXPECT_EQ(retVal, CL_SUCCESS);
+    retVal = pCmdOOQ->enqueueWriteBuffer(srcBuffer,
+                                         CL_FALSE,
+                                         0,
+                                         MemoryConstants::cacheLineSize,
+                                         ptr,
+                                         0,
+                                         nullptr,
+                                         nullptr);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    EXPECT_EQ(pCmdOOQ->taskLevel, 0u);
+}
+HWTEST_F(EnqueueWriteBufferTypeTest, givenOOQWithDisabledSupportCpuCopiesAndDstPtrEqualSrcPtrWhenWriteBufferIsExecutedThenTaskLevelNotIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnWriteBuffer.set(false);
+    cl_int retVal = CL_SUCCESS;
+    std::unique_ptr<CommandQueue> pCmdOOQ(createCommandQueue(pDevice, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE));
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    EXPECT_EQ(retVal, CL_SUCCESS);
+    retVal = pCmdOOQ->enqueueWriteBuffer(srcBuffer,
+                                         CL_FALSE,
+                                         0,
+                                         MemoryConstants::cacheLineSize,
+                                         ptr,
+                                         0,
+                                         nullptr,
+                                         nullptr);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    EXPECT_EQ(pCmdOOQ->taskLevel, 0u);
+}
+HWTEST_F(EnqueueWriteBufferTypeTest, givenInOrderQueueAndEnabledSupportCpuCopiesAndDstPtrEqualSrcPtrWhenWriteBufferIsExecutedThenTaskLevelShouldNotBeIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnWriteBuffer.set(true);
+    cl_int retVal = CL_SUCCESS;
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    EXPECT_EQ(retVal, CL_SUCCESS);
+    retVal = pCmdQ->enqueueWriteBuffer(srcBuffer,
+                                       CL_FALSE,
+                                       0,
+                                       MemoryConstants::cacheLineSize,
+                                       ptr,
+                                       0,
+                                       nullptr,
+                                       nullptr);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    EXPECT_EQ(pCmdQ->taskLevel, 0u);
+}
+HWTEST_F(EnqueueWriteBufferTypeTest, givenInOrderQueueAndDisabledSupportCpuCopiesAndDstPtrEqualSrcPtrWhenWriteBufferIsExecutedThenTaskLevelShouldNotBeIncreased) {
+    DebugManagerStateRestore dbgRestore;
+    DebugManager.flags.DoCpuCopyOnWriteBuffer.set(false);
+    cl_int retVal = CL_SUCCESS;
+    void *ptr = srcBuffer->getCpuAddressForMemoryTransfer();
+    EXPECT_EQ(retVal, CL_SUCCESS);
+    retVal = pCmdQ->enqueueWriteBuffer(srcBuffer,
+                                       CL_FALSE,
+                                       0,
+                                       MemoryConstants::cacheLineSize,
+                                       ptr,
+                                       0,
+                                       nullptr,
+                                       nullptr);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    EXPECT_EQ(pCmdQ->taskLevel, 0u);
+}