diff --git a/opencl/source/command_queue/command_queue_hw.h b/opencl/source/command_queue/command_queue_hw.h
index f5ddce7418..885f03d173 100644
--- a/opencl/source/command_queue/command_queue_hw.h
+++ b/opencl/source/command_queue/command_queue_hw.h
@@ -452,7 +452,7 @@ class CommandQueueHw : public CommandQueue {
     LinearStream *obtainCommandStream(const CsrDependencies &csrDependencies, bool blitEnqueue, bool blockedQueue,
                                       const MultiDispatchInfo &multiDispatchInfo, const EventsRequest &eventsRequest,
                                       std::unique_ptr<KernelOperation> &blockedCommandsData,
-                                      Surface **surfaces, size_t numSurfaces) {
+                                      Surface **surfaces, size_t numSurfaces, bool isMarkerWithProfiling) {
         LinearStream *commandStream = nullptr;
 
         bool profilingRequired = (this->isProfilingEnabled() && eventsRequest.outEvent);
@@ -469,7 +469,7 @@ class CommandQueueHw : public CommandQueue {
             blockedCommandsData = std::make_unique<KernelOperation>(commandStream, *gpgpuCsr.getInternalAllocationStorage());
         } else {
             commandStream = &getCommandStream<GfxFamily, commandType>(*this, csrDependencies, profilingRequired, perfCountersRequired,
-                                                                      blitEnqueue, multiDispatchInfo, surfaces, numSurfaces);
+                                                                      blitEnqueue, multiDispatchInfo, surfaces, numSurfaces, isMarkerWithProfiling);
         }
         return commandStream;
     }
diff --git a/opencl/source/command_queue/enqueue_common.h b/opencl/source/command_queue/enqueue_common.h
index 562bae52af..bff95d23b7 100644
--- a/opencl/source/command_queue/enqueue_common.h
+++ b/opencl/source/command_queue/enqueue_common.h
@@ -228,7 +228,7 @@ void CommandQueueHw<GfxFamily>::enqueueHandler(Surface **surfacesForResidency,
     }
 
     auto &commandStream = *obtainCommandStream<commandType>(csrDeps, false, blockQueue, multiDispatchInfo, eventsRequest,
-                                                            blockedCommandsData, surfacesForResidency, numSurfaceForResidency);
+                                                            blockedCommandsData, surfacesForResidency, numSurfaceForResidency, isMarkerWithProfiling);
     auto commandStreamStart = commandStream.getUsed();
 
     if (this->context->getRootDeviceIndices().size() > 1) {
@@ -1177,7 +1177,7 @@ void CommandQueueHw<GfxFamily>::enqueueBlit(const MultiDispatchInfo &multiDispat
     LinearStream *gpgpuCommandStream = {};
     size_t gpgpuCommandStreamStart = {};
     if (isGpgpuSubmissionForBcsRequired(blockQueue)) {
-        gpgpuCommandStream = obtainCommandStream<cmdType>(csrDeps, true, blockQueue, multiDispatchInfo, eventsRequest, blockedCommandsData, nullptr, 0);
+        gpgpuCommandStream = obtainCommandStream<cmdType>(csrDeps, true, blockQueue, multiDispatchInfo, eventsRequest, blockedCommandsData, nullptr, 0, false);
         gpgpuCommandStreamStart = gpgpuCommandStream->getUsed();
     }
 
diff --git a/opencl/source/command_queue/gpgpu_walker.h b/opencl/source/command_queue/gpgpu_walker.h
index b433472cbe..e1355d2ca8 100644
--- a/opencl/source/command_queue/gpgpu_walker.h
+++ b/opencl/source/command_queue/gpgpu_walker.h
@@ -167,7 +167,7 @@ class GpgpuWalkerHelper {
 template <typename GfxFamily>
 struct EnqueueOperation {
     using PIPE_CONTROL = typename GfxFamily::PIPE_CONTROL;
-    static size_t getTotalSizeRequiredCS(uint32_t eventType, const CsrDependencies &csrDeps, bool reserveProfilingCmdsSpace, bool reservePerfCounters, bool blitEnqueue, CommandQueue &commandQueue, const MultiDispatchInfo &multiDispatchInfo);
+    static size_t getTotalSizeRequiredCS(uint32_t eventType, const CsrDependencies &csrDeps, bool reserveProfilingCmdsSpace, bool reservePerfCounters, bool blitEnqueue, CommandQueue &commandQueue, const MultiDispatchInfo &multiDispatchInfo, bool isMarkerWithProfiling);
     static size_t getSizeRequiredCS(uint32_t cmdType, bool reserveProfilingCmdsSpace, bool reservePerfCounters, CommandQueue &commandQueue, const Kernel *pKernel, const DispatchInfo &dispatchInfo);
     static size_t getSizeRequiredForTimestampPacketWrite();
     static size_t getSizeForCacheFlushAfterWalkerCommands(const Kernel &kernel, const CommandQueue &commandQueue);
@@ -180,8 +180,8 @@ struct EnqueueOperation {
 template <typename GfxFamily, uint32_t eventType>
 LinearStream &getCommandStream(CommandQueue &commandQueue, const CsrDependencies &csrDeps, bool reserveProfilingCmdsSpace,
                                bool reservePerfCounterCmdsSpace, bool blitEnqueue, const MultiDispatchInfo &multiDispatchInfo,
-                               Surface **surfaces, size_t numSurfaces) {
-    size_t expectedSizeCS = EnqueueOperation<GfxFamily>::getTotalSizeRequiredCS(eventType, csrDeps, reserveProfilingCmdsSpace, reservePerfCounterCmdsSpace, blitEnqueue, commandQueue, multiDispatchInfo);
+                               Surface **surfaces, size_t numSurfaces, bool isMarkerWithProfiling) {
+    size_t expectedSizeCS = EnqueueOperation<GfxFamily>::getTotalSizeRequiredCS(eventType, csrDeps, reserveProfilingCmdsSpace, reservePerfCounterCmdsSpace, blitEnqueue, commandQueue, multiDispatchInfo, isMarkerWithProfiling);
     return commandQueue.getCS(expectedSizeCS);
 }
 
diff --git a/opencl/source/command_queue/gpgpu_walker_base.inl b/opencl/source/command_queue/gpgpu_walker_base.inl
index 249cf372ff..0669e9ee2d 100644
--- a/opencl/source/command_queue/gpgpu_walker_base.inl
+++ b/opencl/source/command_queue/gpgpu_walker_base.inl
@@ -6,6 +6,7 @@
  */
 
 #pragma once
+#include "shared/source/command_container/command_encoder.h"
 #include "shared/source/command_stream/command_stream_receiver.h"
 #include "shared/source/helpers/aligned_memory.h"
 #include "shared/source/helpers/debug_helpers.h"
@@ -177,7 +178,7 @@ size_t GpgpuWalkerHelper<GfxFamily>::getSizeForWaDisableRccRhwoOptimization(cons
 }
 
 template <typename GfxFamily>
-size_t EnqueueOperation<GfxFamily>::getTotalSizeRequiredCS(uint32_t eventType, const CsrDependencies &csrDeps, bool reserveProfilingCmdsSpace, bool reservePerfCounters, bool blitEnqueue, CommandQueue &commandQueue, const MultiDispatchInfo &multiDispatchInfo) {
+size_t EnqueueOperation<GfxFamily>::getTotalSizeRequiredCS(uint32_t eventType, const CsrDependencies &csrDeps, bool reserveProfilingCmdsSpace, bool reservePerfCounters, bool blitEnqueue, CommandQueue &commandQueue, const MultiDispatchInfo &multiDispatchInfo, bool isMarkerWithProfiling) {
     size_t expectedSizeCS = 0;
     auto &hwInfo = commandQueue.getDevice().getHardwareInfo();
     auto &commandQueueHw = static_cast<CommandQueueHw<GfxFamily> &>(commandQueue);
@@ -205,8 +206,15 @@ size_t EnqueueOperation<GfxFamily>::getTotalSizeRequiredCS(uint32_t eventType, c
     if (commandQueue.getGpgpuCommandStreamReceiver().peekTimestampPacketWriteEnabled()) {
         expectedSizeCS += TimestampPacketHelper::getRequiredCmdStreamSize<GfxFamily>(csrDeps);
         expectedSizeCS += EnqueueOperation<GfxFamily>::getSizeRequiredForTimestampPacketWrite();
+        if (isMarkerWithProfiling) {
+            expectedSizeCS += 4 * EncodeStoreMMIO<GfxFamily>::size;
+        }
+    } else if (isMarkerWithProfiling) {
+        expectedSizeCS += 2 * MemorySynchronizationCommands<GfxFamily>::getSizeForSinglePipeControl();
+        if (!HwHelper::get(hwInfo.platform.eRenderCoreFamily).useOnlyGlobalTimestamps()) {
+            expectedSizeCS += 2 * EncodeStoreMMIO<GfxFamily>::size;
+        }
     }
-
     if (multiDispatchInfo.peekMainKernel()) {
         expectedSizeCS += EnqueueOperation<GfxFamily>::getSizeForCacheFlushAfterWalkerCommands(*multiDispatchInfo.peekMainKernel(), commandQueue);
     }
diff --git a/opencl/test/unit_test/command_queue/dispatch_walker_tests.cpp b/opencl/test/unit_test/command_queue/dispatch_walker_tests.cpp
index 9cf418acf0..446dfa5e9f 100644
--- a/opencl/test/unit_test/command_queue/dispatch_walker_tests.cpp
+++ b/opencl/test/unit_test/command_queue/dispatch_walker_tests.cpp
@@ -829,7 +829,7 @@ HWTEST_F(DispatchWalkerTest, givenBlockedEnqueueWhenObtainingCommandStreamThenAl
     EventsRequest eventsRequest(0, nullptr, nullptr);
     auto cmdStream = mockCmdQ.template obtainCommandStream<CL_COMMAND_NDRANGE_KERNEL>(csrDependencies, false, true,
                                                                                       multiDispatchInfo, eventsRequest, blockedKernelData,
-                                                                                      nullptr, 0u);
+                                                                                      nullptr, 0u, false);
 
     EXPECT_EQ(expectedSizeCS, cmdStream->getMaxAvailableSpace());
     EXPECT_EQ(expectedSizeCSAllocation, cmdStream->getGraphicsAllocation()->getUnderlyingBufferSize());
diff --git a/opencl/test/unit_test/command_queue/enqueue_kernel_1_tests.cpp b/opencl/test/unit_test/command_queue/enqueue_kernel_1_tests.cpp
index f261e6e062..8b58766635 100644
--- a/opencl/test/unit_test/command_queue/enqueue_kernel_1_tests.cpp
+++ b/opencl/test/unit_test/command_queue/enqueue_kernel_1_tests.cpp
@@ -1711,10 +1711,10 @@ HWTEST_F(PauseOnGpuTests, givenGpuScratchWriteEnabledWhenEstimatingCommandStream
     dispatchInfo.setKernel(mockKernel.mockKernel);
     multiDispatchInfo.push(dispatchInfo);
 
-    auto baseCommandStreamSize = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_NDRANGE_KERNEL, {}, false, false, false, *pCmdQ, multiDispatchInfo);
+    auto baseCommandStreamSize = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_NDRANGE_KERNEL, {}, false, false, false, *pCmdQ, multiDispatchInfo, false);
     DebugManager.flags.GpuScratchRegWriteAfterWalker.set(1);
 
-    auto extendedCommandStreamSize = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_NDRANGE_KERNEL, {}, false, false, false, *pCmdQ, multiDispatchInfo);
+    auto extendedCommandStreamSize = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_NDRANGE_KERNEL, {}, false, false, false, *pCmdQ, multiDispatchInfo, false);
 
     EXPECT_EQ(baseCommandStreamSize + sizeof(typename FamilyType::MI_LOAD_REGISTER_IMM), extendedCommandStreamSize);
 }
diff --git a/opencl/test/unit_test/command_queue/get_size_required_buffer_tests.cpp b/opencl/test/unit_test/command_queue/get_size_required_buffer_tests.cpp
index 5eb6cbecce..7d25dc784f 100644
--- a/opencl/test/unit_test/command_queue/get_size_required_buffer_tests.cpp
+++ b/opencl/test/unit_test/command_queue/get_size_required_buffer_tests.cpp
@@ -98,7 +98,7 @@ HWTEST_F(GetSizeRequiredBufferTest, WhenFillingBufferThenHeapsAndCommandBufferCo
     auto usedAfterSSH = ssh.getUsed();
 
     auto expectedSizeCS = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_FILL_BUFFER, CsrDependencies(), false, false,
-                                                                               false, *pCmdQ, multiDispatchInfo);
+                                                                               false, *pCmdQ, multiDispatchInfo, false);
     auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredDSH(multiDispatchInfo);
     auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredIOH(multiDispatchInfo);
     auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredSSH(multiDispatchInfo);
@@ -151,7 +151,7 @@ HWTEST_F(GetSizeRequiredBufferTest, WhenCopyingBufferThenHeapsAndCommandBufferCo
     auto usedAfterSSH = ssh.getUsed();
 
     auto expectedSizeCS = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_COPY_BUFFER, CsrDependencies(), false, false,
-                                                                               false, *pCmdQ, multiDispatchInfo);
+                                                                               false, *pCmdQ, multiDispatchInfo, false);
     auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredDSH(multiDispatchInfo);
     auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredIOH(multiDispatchInfo);
     auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredSSH(multiDispatchInfo);
@@ -205,7 +205,7 @@ HWTEST_F(GetSizeRequiredBufferTest, WhenReadingBufferNonBlockingThenHeapsAndComm
     auto usedAfterSSH = ssh.getUsed();
 
     auto expectedSizeCS = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_READ_BUFFER, CsrDependencies(), false, false,
-                                                                               false, *pCmdQ, multiDispatchInfo);
+                                                                               false, *pCmdQ, multiDispatchInfo, false);
     auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredDSH(multiDispatchInfo);
     auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredIOH(multiDispatchInfo);
     auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredSSH(multiDispatchInfo);
@@ -260,7 +260,7 @@ HWTEST_F(GetSizeRequiredBufferTest, WhenReadingBufferBlockingThenThenHeapsAndCom
     auto usedAfterSSH = ssh.getUsed();
 
     auto expectedSizeCS = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_READ_BUFFER, CsrDependencies(), false, false,
-                                                                               false, *pCmdQ, multiDispatchInfo);
+                                                                               false, *pCmdQ, multiDispatchInfo, false);
     auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredDSH(multiDispatchInfo);
     auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredIOH(multiDispatchInfo);
     auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredSSH(multiDispatchInfo);
@@ -315,7 +315,7 @@ HWTEST_F(GetSizeRequiredBufferTest, WhenWritingBufferNonBlockingThenHeapsAndComm
     auto usedAfterSSH = ssh.getUsed();
 
     auto expectedSizeCS = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_WRITE_BUFFER, CsrDependencies(), false, false,
-                                                                               false, *pCmdQ, multiDispatchInfo);
+                                                                               false, *pCmdQ, multiDispatchInfo, false);
     auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredDSH(multiDispatchInfo);
     auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredIOH(multiDispatchInfo);
     auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredSSH(multiDispatchInfo);
@@ -367,7 +367,7 @@ HWTEST_F(GetSizeRequiredBufferTest, WhenWritingBufferBlockingThenHeapsAndCommand
     auto usedAfterSSH = ssh.getUsed();
 
     auto expectedSizeCS = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_WRITE_BUFFER, CsrDependencies(), false, false,
-                                                                               false, *pCmdQ, multiDispatchInfo);
+                                                                               false, *pCmdQ, multiDispatchInfo, false);
     auto expectedSizeDSH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredDSH(multiDispatchInfo);
     auto expectedSizeIOH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredIOH(multiDispatchInfo);
     auto expectedSizeSSH = HardwareCommandsHelper<FamilyType>::getTotalSizeRequiredSSH(multiDispatchInfo);
diff --git a/opencl/test/unit_test/execution_model/parent_kernel_dispatch_tests.cpp b/opencl/test/unit_test/execution_model/parent_kernel_dispatch_tests.cpp
index d5b81b8449..306c2cfa6a 100644
--- a/opencl/test/unit_test/execution_model/parent_kernel_dispatch_tests.cpp
+++ b/opencl/test/unit_test/execution_model/parent_kernel_dispatch_tests.cpp
@@ -223,7 +223,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, ParentKernelCommandStreamFixture, GivenDispatchInfoW
     size_t totalKernelSize = alignUp(numOfKernels * size, MemoryConstants::pageSize);
 
     LinearStream &commandStream = getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*pCmdQ, CsrDependencies(), false, false,
-                                                                                          false, multiDispatchInfo, nullptr, 0);
+                                                                                          false, multiDispatchInfo, nullptr, 0, false);
 
     EXPECT_LT(totalKernelSize, commandStream.getMaxAvailableSpace());
 
diff --git a/opencl/test/unit_test/execution_model/scheduler_dispatch_tests.cpp b/opencl/test/unit_test/execution_model/scheduler_dispatch_tests.cpp
index 3c832d3817..3e0a6bbf69 100644
--- a/opencl/test/unit_test/execution_model/scheduler_dispatch_tests.cpp
+++ b/opencl/test/unit_test/execution_model/scheduler_dispatch_tests.cpp
@@ -58,7 +58,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, ExecutionModelSchedulerFixture, WhenDispatchingSched
     MultiDispatchInfo multiDispatchinfo(&scheduler);
     LinearStream &commandStream = getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*pCmdQ, CsrDependencies(),
                                                                                           false, false, false, multiDispatchinfo,
-                                                                                          nullptr, 0);
+                                                                                          nullptr, 0, false);
     pCmdQ->getIndirectHeap(IndirectHeap::SURFACE_STATE, minRequiredSizeForSchedulerSSH);
 
     GpgpuWalkerHelper<FamilyType>::dispatchScheduler(
@@ -174,7 +174,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, ExecutionModelSchedulerFixture, WhenDispatchingSched
 
     MultiDispatchInfo multiDispatchinfo(&scheduler);
     getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*pCmdQ, CsrDependencies(), false, false, false, multiDispatchinfo,
-                                                            nullptr, 0);
+                                                            nullptr, 0, false);
     pCmdQ->getIndirectHeap(IndirectHeap::SURFACE_STATE, minRequiredSizeForSchedulerSSH);
 
     GpgpuWalkerHelper<FamilyType>::dispatchScheduler(
@@ -209,7 +209,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, ParentKernelCommandQueueFixture, GivenEarlyReturnSet
     MultiDispatchInfo multiDispatchinfo(&scheduler);
     LinearStream &commandStream = getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*pCmdQ, CsrDependencies(),
                                                                                           false, false, false, multiDispatchinfo,
-                                                                                          nullptr, 0);
+                                                                                          nullptr, 0, false);
     pCmdQ->getIndirectHeap(IndirectHeap::SURFACE_STATE, minRequiredSizeForSchedulerSSH);
 
     GpgpuWalkerHelper<FamilyType>::dispatchScheduler(
diff --git a/opencl/test/unit_test/gen8/scheduler_dispatch_tests_gen8.cpp b/opencl/test/unit_test/gen8/scheduler_dispatch_tests_gen8.cpp
index b2c3d8473e..54ccefffc9 100644
--- a/opencl/test/unit_test/gen8/scheduler_dispatch_tests_gen8.cpp
+++ b/opencl/test/unit_test/gen8/scheduler_dispatch_tests_gen8.cpp
@@ -37,7 +37,7 @@ BDWTEST_F(BdwSchedulerTest, givenCallToDispatchSchedulerWhenPipeControlWithCSSta
         MultiDispatchInfo multiDispatchinfo(&scheduler);
         LinearStream &commandStream = getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*pCmdQ, CsrDependencies(),
                                                                                               false, false, false, multiDispatchinfo,
-                                                                                              nullptr, 0);
+                                                                                              nullptr, 0, false);
         pCmdQ->getIndirectHeap(IndirectHeap::SURFACE_STATE, minRequiredSizeForSchedulerSSH);
 
         GpgpuWalkerHelper<FamilyType>::dispatchScheduler(
diff --git a/opencl/test/unit_test/helpers/timestamp_packet_1_tests.cpp b/opencl/test/unit_test/helpers/timestamp_packet_1_tests.cpp
index 777a259115..6e698a02c7 100644
--- a/opencl/test/unit_test/helpers/timestamp_packet_1_tests.cpp
+++ b/opencl/test/unit_test/helpers/timestamp_packet_1_tests.cpp
@@ -349,11 +349,11 @@ HWCMDTEST_F(IGFX_GEN8_CORE, TimestampPacketTests, givenTimestampPacketWriteEnabl
     auto mockCmdQHw = std::make_unique<MockCommandQueueHw<FamilyType>>(context, device.get(), nullptr);
 
     device->getUltCommandStreamReceiver<FamilyType>().timestampPacketWriteEnabled = false;
-    getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*mockCmdQHw, CsrDependencies(), false, false, false, multiDispatchInfo, nullptr, 0);
+    getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*mockCmdQHw, CsrDependencies(), false, false, false, multiDispatchInfo, nullptr, 0, false);
     auto sizeWithDisabled = mockCmdQHw->requestedCmdStreamSize;
 
     device->getUltCommandStreamReceiver<FamilyType>().timestampPacketWriteEnabled = true;
-    getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*mockCmdQHw, CsrDependencies(), false, false, false, multiDispatchInfo, nullptr, 0);
+    getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*mockCmdQHw, CsrDependencies(), false, false, false, multiDispatchInfo, nullptr, 0, false);
     auto sizeWithEnabled = mockCmdQHw->requestedCmdStreamSize;
 
     auto extendedSize = sizeWithDisabled + sizeof(typename FamilyType::PIPE_CONTROL);
@@ -367,7 +367,7 @@ HWTEST_F(TimestampPacketTests, givenTimestampPacketWriteEnabledAndOoqWhenEstimat
 
     device->getUltCommandStreamReceiver<FamilyType>().timestampPacketWriteEnabled = false;
     getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*mockCmdQHw, CsrDependencies(), false, false,
-                                                            false, multiDispatchInfo, nullptr, 0);
+                                                            false, multiDispatchInfo, nullptr, 0, false);
     auto sizeWithDisabled = mockCmdQHw->requestedCmdStreamSize;
 
     device->getUltCommandStreamReceiver<FamilyType>().timestampPacketWriteEnabled = true;
@@ -397,7 +397,7 @@ HWTEST_F(TimestampPacketTests, givenTimestampPacketWriteEnabledAndOoqWhenEstimat
     eventsRequest.fillCsrDependenciesForTimestampPacketContainer(
         csrDeps, device->getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OnCsr);
 
-    getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*mockCmdQHw, csrDeps, false, false, false, multiDispatchInfo, nullptr, 0);
+    getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*mockCmdQHw, csrDeps, false, false, false, multiDispatchInfo, nullptr, 0, false);
     auto sizeWithEnabled = mockCmdQHw->requestedCmdStreamSize;
 
     size_t sizeForNodeDependency = 0;
@@ -417,7 +417,7 @@ HWTEST_F(TimestampPacketTests, givenTimestampPacketWriteEnabledWhenEstimatingStr
     auto mockCmdQHw = std::make_unique<MockCommandQueueHw<FamilyType>>(context, device.get(), nullptr);
 
     device->getUltCommandStreamReceiver<FamilyType>().timestampPacketWriteEnabled = false;
-    getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*mockCmdQHw, CsrDependencies(), false, false, false, multiDispatchInfo, nullptr, 0);
+    getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*mockCmdQHw, CsrDependencies(), false, false, false, multiDispatchInfo, nullptr, 0, false);
     auto sizeWithDisabled = mockCmdQHw->requestedCmdStreamSize;
 
     device->getUltCommandStreamReceiver<FamilyType>().timestampPacketWriteEnabled = true;
@@ -446,7 +446,7 @@ HWTEST_F(TimestampPacketTests, givenTimestampPacketWriteEnabledWhenEstimatingStr
     CsrDependencies csrDeps;
     eventsRequest.fillCsrDependenciesForTimestampPacketContainer(csrDeps, device->getGpgpuCommandStreamReceiver(), CsrDependencies::DependenciesType::OnCsr);
 
-    getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*mockCmdQHw, csrDeps, false, false, false, multiDispatchInfo, nullptr, 0);
+    getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*mockCmdQHw, csrDeps, false, false, false, multiDispatchInfo, nullptr, 0, false);
     auto sizeWithEnabled = mockCmdQHw->requestedCmdStreamSize;
 
     size_t sizeForNodeDependency = 0;
diff --git a/opencl/test/unit_test/mem_obj/buffer_bcs_tests.cpp b/opencl/test/unit_test/mem_obj/buffer_bcs_tests.cpp
index 5cd4898b18..4a5400308c 100644
--- a/opencl/test/unit_test/mem_obj/buffer_bcs_tests.cpp
+++ b/opencl/test/unit_test/mem_obj/buffer_bcs_tests.cpp
@@ -790,11 +790,11 @@ HWTEST_TEMPLATED_F(BcsBufferTests, givenBufferOperationWithoutKernelWhenEstimati
     auto &hwInfo = cmdQ->getDevice().getHardwareInfo();
 
     auto readBufferCmdsSize = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_READ_BUFFER, csrDependencies, false, false,
-                                                                                   true, *cmdQ, multiDispatchInfo);
+                                                                                   true, *cmdQ, multiDispatchInfo, false);
     auto writeBufferCmdsSize = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_WRITE_BUFFER, csrDependencies, false, false,
-                                                                                    true, *cmdQ, multiDispatchInfo);
+                                                                                    true, *cmdQ, multiDispatchInfo, false);
     auto copyBufferCmdsSize = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_COPY_BUFFER, csrDependencies, false, false,
-                                                                                   true, *cmdQ, multiDispatchInfo);
+                                                                                   true, *cmdQ, multiDispatchInfo, false);
     auto expectedSize = TimestampPacketHelper::getRequiredCmdStreamSizeForNodeDependencyWithBlitEnqueue<FamilyType>();
 
     if (cmdQ->isCacheFlushForBcsRequired()) {
diff --git a/opencl/test/unit_test/profiling/profiling_tests.cpp b/opencl/test/unit_test/profiling/profiling_tests.cpp
index 5174704dbf..9f74b94b66 100644
--- a/opencl/test/unit_test/profiling/profiling_tests.cpp
+++ b/opencl/test/unit_test/profiling/profiling_tests.cpp
@@ -71,13 +71,13 @@ HWCMDTEST_F(IGFX_GEN8_CORE, ProfilingTests, GivenCommandQueueWithProfilingAndFor
 
     MultiDispatchInfo multiDispatchInfo(&kernel);
     auto &commandStreamNDRangeKernel = getCommandStream<FamilyType, CL_COMMAND_NDRANGE_KERNEL>(*pCmdQ, CsrDependencies(), true, false, false,
-                                                                                               multiDispatchInfo, nullptr, 0);
+                                                                                               multiDispatchInfo, nullptr, 0, false);
     auto expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_NDRANGE_KERNEL, true, false, *pCmdQ, &kernel, {});
     EXPECT_GE(expectedSizeCS, requiredSize);
     EXPECT_GE(commandStreamNDRangeKernel.getAvailableSpace(), requiredSize);
 
     auto &commandStreamTask = getCommandStream<FamilyType, CL_COMMAND_TASK>(*pCmdQ, CsrDependencies(), true, false, false,
-                                                                            multiDispatchInfo, nullptr, 0);
+                                                                            multiDispatchInfo, nullptr, 0, false);
     expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_TASK, true, false, *pCmdQ, &kernel, {});
     EXPECT_GE(expectedSizeCS, requiredSize);
     EXPECT_GE(commandStreamTask.getAvailableSpace(), requiredSize);
@@ -93,13 +93,13 @@ HWTEST_F(ProfilingTests, GivenCommandQueueWithProfilingAndForWorkloadWithNoKerne
     MultiDispatchInfo multiDispatchInfo(nullptr);
     auto &commandStreamMigrateMemObjects = getCommandStream<FamilyType, CL_COMMAND_MIGRATE_MEM_OBJECTS>(*pCmdQ, CsrDependencies(),
                                                                                                         true, false, false,
-                                                                                                        multiDispatchInfo, nullptr, 0);
+                                                                                                        multiDispatchInfo, nullptr, 0, false);
     auto expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_MIGRATE_MEM_OBJECTS, true, false, *pCmdQ, nullptr, {});
     EXPECT_GE(expectedSizeCS, requiredSize);
     EXPECT_GE(commandStreamMigrateMemObjects.getAvailableSpace(), requiredSize);
 
     auto &commandStreamMarker = getCommandStream<FamilyType, CL_COMMAND_MARKER>(*pCmdQ, CsrDependencies(), true,
-                                                                                false, false, multiDispatchInfo, nullptr, 0);
+                                                                                false, false, multiDispatchInfo, nullptr, 0, false);
     expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_MARKER, true, false, *pCmdQ, nullptr, {});
     EXPECT_GE(expectedSizeCS, requiredSize);
     EXPECT_GE(commandStreamMarker.getAvailableSpace(), requiredSize);
@@ -121,9 +121,9 @@ HWCMDTEST_F(IGFX_GEN8_CORE, ProfilingTests, GivenCommandQueueWithProfilingAndFor
     multiDispatchInfo.push(dispatchInfo);
     multiDispatchInfo.push(dispatchInfo);
     auto &commandStreamTask = getCommandStream<FamilyType, CL_COMMAND_TASK>(*pCmdQ, CsrDependencies(), true, false, false,
-                                                                            multiDispatchInfo, nullptr, 0);
+                                                                            multiDispatchInfo, nullptr, 0, false);
     auto expectedSizeCS = EnqueueOperation<FamilyType>::getTotalSizeRequiredCS(CL_COMMAND_TASK, CsrDependencies(), true, false,
-                                                                               false, *pCmdQ, multiDispatchInfo);
+                                                                               false, *pCmdQ, multiDispatchInfo, false);
     EXPECT_GE(expectedSizeCS, requiredSize);
     EXPECT_GE(commandStreamTask.getAvailableSpace(), requiredSize);
 }
@@ -713,13 +713,13 @@ HWTEST_F(ProfilingWithPerfCountersTests, GivenCommandQueueWithProfilingPerfCount
     MultiDispatchInfo multiDispatchInfo(nullptr);
     auto &commandStreamMigrateMemObjects = getCommandStream<FamilyType, CL_COMMAND_MIGRATE_MEM_OBJECTS>(*pCmdQ, CsrDependencies(),
                                                                                                         true, true, false, multiDispatchInfo,
-                                                                                                        nullptr, 0);
+                                                                                                        nullptr, 0, false);
     auto expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_MIGRATE_MEM_OBJECTS, true, true, *pCmdQ, nullptr, {});
     EXPECT_GE(expectedSizeCS, requiredSize);
     EXPECT_GE(commandStreamMigrateMemObjects.getAvailableSpace(), requiredSize);
 
     auto &commandStreamMarker = getCommandStream<FamilyType, CL_COMMAND_MARKER>(*pCmdQ, CsrDependencies(), true, true, false,
-                                                                                multiDispatchInfo, nullptr, 0);
+                                                                                multiDispatchInfo, nullptr, 0, false);
     expectedSizeCS = EnqueueOperation<FamilyType>::getSizeRequiredCS(CL_COMMAND_MARKER, true, true, *pCmdQ, nullptr, {});
     EXPECT_GE(expectedSizeCS, requiredSize);
     EXPECT_GE(commandStreamMarker.getAvailableSpace(), requiredSize);