Move generic command stream receiver files to shared

Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
2025-12-23 11:03:02 +08:00 · 2021-09-22 22:03:07 +00:00
parent 479035c839
commit f8867e0b97
98 changed files with 246 additions and 159 deletions
--- a/level_zero/core/source/gen11/enable_family_full_l0_gen11.cpp
+++ b/level_zero/core/source/gen11/enable_family_full_l0_gen11.cpp
@@ -5,11 +5,11 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/helpers/populate_factory.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/source/mem_obj/buffer.h"

 #include "level_zero/core/source/helpers/l0_populate_factory.h"
--- a/level_zero/core/source/gen12lp/enable_family_full_l0_gen12lp.cpp
+++ b/level_zero/core/source/gen12lp/enable_family_full_l0_gen12lp.cpp
@@ -5,11 +5,11 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/helpers/populate_factory.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/source/mem_obj/buffer.h"

 #include "level_zero/core/source/helpers/l0_populate_factory.h"
--- a/level_zero/core/source/gen8/enable_family_full_l0_gen8.cpp
+++ b/level_zero/core/source/gen8/enable_family_full_l0_gen8.cpp
@@ -5,11 +5,11 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/helpers/populate_factory.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/source/mem_obj/buffer.h"

 namespace NEO {
--- a/level_zero/core/source/gen9/enable_family_full_l0_gen9.cpp
+++ b/level_zero/core/source/gen9/enable_family_full_l0_gen9.cpp
@@ -5,11 +5,11 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/helpers/populate_factory.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/source/mem_obj/buffer.h"

 #include "level_zero/core/source/helpers/l0_populate_factory.h"
--- a/level_zero/core/source/helpers/api_specific_config_l0.cpp
+++ b/level_zero/core/source/helpers/api_specific_config_l0.cpp
@@ -8,6 +8,8 @@
 #include "shared/source/debug_settings/debug_settings_manager.h"
 #include "shared/source/helpers/api_specific_config.h"

+#include "level_zero/core/source/compiler_interface/l0_reg_path.h"
+
 namespace NEO {
 bool ApiSpecificConfig::isStatelessCompressionSupported() {
    return false;
@@ -37,4 +39,8 @@ uint64_t ApiSpecificConfig::getReducedMaxAllocSize(uint64_t maxAllocSize) {
    return maxAllocSize;
 }

+const char *ApiSpecificConfig::getRegistryPath() {
+    return L0::registryPath;
+}
+
 } // namespace NEO
--- a/level_zero/core/source/xe_hp_core/enable_family_full_l0_xe_hp_core.cpp
+++ b/level_zero/core/source/xe_hp_core/enable_family_full_l0_xe_hp_core.cpp
@@ -5,11 +5,11 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/helpers/populate_factory.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/source/mem_obj/buffer.h"

 #include "level_zero/core/source/helpers/l0_populate_factory.h"
--- a/level_zero/core/test/unit_tests/sources/helper/api_specific_config_l0_tests.cpp
+++ b/level_zero/core/test/unit_tests/sources/helper/api_specific_config_l0_tests.cpp
@@ -7,6 +7,8 @@

 #include "shared/source/helpers/api_specific_config.h"

+#include "level_zero/core/source/compiler_interface/l0_reg_path.h"
+
 #include "gtest/gtest.h"

 namespace NEO {
@@ -31,4 +33,7 @@ TEST(ApiSpecificConfigL0Tests, givenMaxAllocSizeWhenGettingReducedMaxAllocSizeTh
    EXPECT_EQ(1024u, ApiSpecificConfig::getReducedMaxAllocSize(1024));
 }

+TEST(ApiSpecificConfigL0Tests, WhenGettingRegistryPathThenL0RegistryPathIsReturned) {
+    EXPECT_STREQ(L0::registryPath, ApiSpecificConfig::getRegistryPath());
+}
 } // namespace NEO
--- a/opencl/source/command_stream/CMakeLists.txt
+++ b/opencl/source/command_stream/CMakeLists.txt
@@ -7,27 +7,8 @@
 set(RUNTIME_SRCS_COMMAND_STREAM
    ${CMAKE_CURRENT_SOURCE_DIR}/CMakeLists.txt
    ${CMAKE_CURRENT_SOURCE_DIR}${BRANCH_DIR_SUFFIX}per_dss_backed_buffer.cpp
-    ${CMAKE_CURRENT_SOURCE_DIR}/aub_command_stream_receiver_hw.h
-    ${CMAKE_CURRENT_SOURCE_DIR}/aub_command_stream_receiver_hw_base.inl
-    ${CMAKE_CURRENT_SOURCE_DIR}/aub_command_stream_receiver_hw_bdw_and_later.inl
-    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_simulated_common_hw.h
-    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_simulated_common_hw_base.inl
-    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_simulated_common_hw_bdw_and_later.inl
-    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_with_aub_dump.h
-    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_with_aub_dump.inl
-    ${CMAKE_CURRENT_SOURCE_DIR}/create_command_stream_impl.cpp
-    ${CMAKE_CURRENT_SOURCE_DIR}/create_command_stream_impl.h
-    ${CMAKE_CURRENT_SOURCE_DIR}/definitions/command_stream_receiver_simulated_hw.h
 )

-if(SUPPORT_XEHP_AND_LATER)
-  list(APPEND RUNTIME_SRCS_COMMAND_STREAM
-       ${CMAKE_CURRENT_SOURCE_DIR}/aub_command_stream_receiver_hw_xehp_and_later.inl
-       ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_simulated_common_hw_xehp_and_later.inl
-       ${CMAKE_CURRENT_SOURCE_DIR}/tbx_command_stream_receiver_xehp_and_later.inl
-  )
-endif()
-
 target_sources(${NEO_STATIC_LIB_NAME} PRIVATE ${RUNTIME_SRCS_COMMAND_STREAM})
 set_property(GLOBAL PROPERTY RUNTIME_SRCS_COMMAND_STREAM ${RUNTIME_SRCS_COMMAND_STREAM})
 add_subdirectories()
--- a/opencl/source/dll/create_command_stream.cpp
+++ b/opencl/source/dll/create_command_stream.cpp
@@ -6,13 +6,12 @@
 */

 #include "shared/source/command_stream/aub_command_stream_receiver.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
+#include "shared/source/command_stream/create_command_stream_impl.h"
 #include "shared/source/command_stream/device_command_stream.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver.h"
 #include "shared/source/helpers/hw_info.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
-#include "opencl/source/command_stream/create_command_stream_impl.h"
-
 namespace NEO {

 CommandStreamReceiver *createCommandStream(ExecutionEnvironment &executionEnvironment,
--- a/opencl/source/dll/get_devices.cpp
+++ b/opencl/source/dll/get_devices.cpp
@@ -5,11 +5,10 @@
 *
 */

+#include "shared/source/command_stream/create_command_stream_impl.h"
 #include "shared/source/execution_environment/execution_environment.h"
 #include "shared/source/execution_environment/root_device_environment.h"

-#include "opencl/source/command_stream/create_command_stream_impl.h"
-
 namespace NEO {

 bool prepareDeviceEnvironments(ExecutionEnvironment &executionEnvironment) {
--- a/opencl/source/gen11/aub_command_stream_receiver_gen11.cpp
+++ b/opencl/source/gen11/aub_command_stream_receiver_gen11.cpp
@@ -5,12 +5,11 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw_bdw_and_later.inl"
 #include "shared/source/helpers/array_count.h"
 #include "shared/source/helpers/populate_factory.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
-#include "opencl/source/command_stream/aub_command_stream_receiver_hw_bdw_and_later.inl"
-
 namespace NEO {

 typedef ICLFamily Family;
--- a/opencl/source/gen11/command_stream_receiver_simulated_common_hw_gen11.cpp
+++ b/opencl/source/gen11/command_stream_receiver_simulated_common_hw_gen11.cpp
@@ -5,7 +5,7 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_simulated_common_hw_bdw_and_later.inl"
+#include "shared/source/command_stream/command_stream_receiver_simulated_common_hw_bdw_and_later.inl"

 namespace NEO {
 typedef ICLFamily Family;
--- a/opencl/source/gen11/enable_family_full_ocl_gen11.cpp
+++ b/opencl/source/gen11/enable_family_full_ocl_gen11.cpp
@@ -5,12 +5,12 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/helpers/populate_factory.h"

 #include "opencl/source/command_queue/command_queue_hw.h"
-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/source/device_queue/device_queue_hw.h"
 #include "opencl/source/helpers/cl_hw_helper.h"
 #include "opencl/source/mem_obj/buffer.h"
--- a/opencl/source/gen11/linux/command_stream_receiver_gen11.cpp
+++ b/opencl/source/gen11/linux/command_stream_receiver_gen11.cpp
@@ -5,7 +5,8 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+
 #include "opencl/source/os_interface/linux/device_command_stream.inl"
 #include "opencl/source/os_interface/linux/drm_command_stream.inl"
 #include "opencl/source/os_interface/linux/drm_command_stream_bdw_and_later.inl"
--- a/opencl/source/gen11/tbx_command_stream_receiver_gen11.cpp
+++ b/opencl/source/gen11/tbx_command_stream_receiver_gen11.cpp
@@ -5,14 +5,13 @@
 *
 */

+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.inl"
 #include "shared/source/gen11/hw_cmds.h"
 #include "shared/source/helpers/array_count.h"
 #include "shared/source/helpers/populate_factory.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
-
 namespace NEO {
 typedef ICLFamily Family;
 static auto gfxCore = IGFX_GEN11_CORE;
--- a/opencl/source/gen11/windows/command_stream_receiver_gen11.cpp
+++ b/opencl/source/gen11/windows/command_stream_receiver_gen11.cpp
@@ -5,7 +5,8 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+
 #include "opencl/source/os_interface/windows/device_command_stream.inl"
 #include "opencl/source/os_interface/windows/wddm_device_command_stream.inl"

--- a/opencl/source/gen12lp/aub_command_stream_receiver_gen12lp.cpp
+++ b/opencl/source/gen12lp/aub_command_stream_receiver_gen12lp.cpp
@@ -6,13 +6,12 @@
 */

 #include "shared/source/aub_mem_dump/aub_alloc_dump.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw_bdw_and_later.inl"
 #include "shared/source/helpers/array_count.h"
 #include "shared/source/helpers/populate_factory.h"
 #include "shared/source/memory_manager/memory_pool.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
-#include "opencl/source/command_stream/aub_command_stream_receiver_hw_bdw_and_later.inl"
-
 namespace NEO {
 typedef TGLLPFamily Family;
 static auto gfxCore = IGFX_GEN12LP_CORE;
--- a/opencl/source/gen12lp/command_stream_receiver_simulated_common_hw_gen12lp.cpp
+++ b/opencl/source/gen12lp/command_stream_receiver_simulated_common_hw_gen12lp.cpp
@@ -5,10 +5,9 @@
 *
 */

+#include "shared/source/command_stream/command_stream_receiver_simulated_common_hw_bdw_and_later.inl"
 #include "shared/source/gen12lp/helpers_gen12lp.h"

-#include "opencl/source/command_stream/command_stream_receiver_simulated_common_hw_bdw_and_later.inl"
-
 namespace NEO {
 typedef TGLLPFamily Family;

--- a/opencl/source/gen12lp/enable_family_full_ocl_gen12lp.cpp
+++ b/opencl/source/gen12lp/enable_family_full_ocl_gen12lp.cpp
@@ -5,12 +5,12 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/helpers/populate_factory.h"

 #include "opencl/source/command_queue/command_queue_hw.h"
-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/source/device_queue/device_queue_hw.h"
 #include "opencl/source/helpers/cl_hw_helper.h"
 #include "opencl/source/mem_obj/buffer.h"
--- a/opencl/source/gen12lp/linux/command_stream_receiver_gen12lp.cpp
+++ b/opencl/source/gen12lp/linux/command_stream_receiver_gen12lp.cpp
@@ -5,7 +5,8 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+
 #include "opencl/source/os_interface/linux/device_command_stream.inl"
 #include "opencl/source/os_interface/linux/drm_command_stream.inl"
 #include "opencl/source/os_interface/linux/drm_command_stream_bdw_and_later.inl"
--- a/opencl/source/gen12lp/tbx_command_stream_receiver_gen12lp.cpp
+++ b/opencl/source/gen12lp/tbx_command_stream_receiver_gen12lp.cpp
@@ -5,6 +5,7 @@
 *
 */

+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.inl"
 #include "shared/source/gen12lp/hw_cmds.h"
@@ -12,8 +13,6 @@
 #include "shared/source/helpers/populate_factory.h"
 #include "shared/source/memory_manager/memory_pool.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
-
 #include "aub_mem_dump.h"

 namespace NEO {
--- a/opencl/source/gen12lp/windows/command_stream_receiver_gen12lp.cpp
+++ b/opencl/source/gen12lp/windows/command_stream_receiver_gen12lp.cpp
@@ -1,11 +1,12 @@
 /*
- * Copyright (C) 2019-2020 Intel Corporation
+ * Copyright (C) 2019-2021 Intel Corporation
 *
 * SPDX-License-Identifier: MIT
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+
 #include "opencl/source/os_interface/windows/device_command_stream.inl"
 #include "opencl/source/os_interface/windows/wddm_device_command_stream.inl"

--- a/opencl/source/gen8/aub_command_stream_receiver_gen8.cpp
+++ b/opencl/source/gen8/aub_command_stream_receiver_gen8.cpp
@@ -5,12 +5,11 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw_bdw_and_later.inl"
 #include "shared/source/helpers/array_count.h"
 #include "shared/source/helpers/populate_factory.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
-#include "opencl/source/command_stream/aub_command_stream_receiver_hw_bdw_and_later.inl"
-
 namespace NEO {

 typedef BDWFamily Family;
--- a/opencl/source/gen8/command_stream_receiver_simulated_common_hw_gen8.cpp
+++ b/opencl/source/gen8/command_stream_receiver_simulated_common_hw_gen8.cpp
@@ -5,7 +5,7 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_simulated_common_hw_bdw_and_later.inl"
+#include "shared/source/command_stream/command_stream_receiver_simulated_common_hw_bdw_and_later.inl"

 namespace NEO {
 typedef BDWFamily Family;
--- a/opencl/source/gen8/enable_family_full_ocl_gen8.cpp
+++ b/opencl/source/gen8/enable_family_full_ocl_gen8.cpp
@@ -5,12 +5,12 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/helpers/populate_factory.h"

 #include "opencl/source/command_queue/command_queue_hw.h"
-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/source/device_queue/device_queue_hw.h"
 #include "opencl/source/event/perf_counter.h"
 #include "opencl/source/helpers/cl_hw_helper.h"
--- a/opencl/source/gen8/linux/command_stream_receiver_gen8.cpp
+++ b/opencl/source/gen8/linux/command_stream_receiver_gen8.cpp
@@ -5,7 +5,8 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+
 #include "opencl/source/os_interface/linux/device_command_stream.inl"
 #include "opencl/source/os_interface/linux/drm_command_stream.inl"
 #include "opencl/source/os_interface/linux/drm_command_stream_bdw_and_later.inl"
--- a/opencl/source/gen8/tbx_command_stream_receiver_gen8.cpp
+++ b/opencl/source/gen8/tbx_command_stream_receiver_gen8.cpp
@@ -5,13 +5,12 @@
 *
 */

+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.inl"
 #include "shared/source/gen8/hw_cmds.h"
 #include "shared/source/helpers/array_count.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
-
 namespace NEO {

 typedef BDWFamily Family;
--- a/opencl/source/gen8/windows/command_stream_receiver_gen8.cpp
+++ b/opencl/source/gen8/windows/command_stream_receiver_gen8.cpp
@@ -5,7 +5,8 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+
 #include "opencl/source/os_interface/windows/device_command_stream.inl"
 #include "opencl/source/os_interface/windows/wddm_device_command_stream.inl"

--- a/opencl/source/gen9/aub_command_stream_receiver_gen9.cpp
+++ b/opencl/source/gen9/aub_command_stream_receiver_gen9.cpp
@@ -5,12 +5,11 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw_bdw_and_later.inl"
 #include "shared/source/helpers/array_count.h"
 #include "shared/source/helpers/populate_factory.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
-#include "opencl/source/command_stream/aub_command_stream_receiver_hw_bdw_and_later.inl"
-
 namespace NEO {

 typedef SKLFamily Family;
--- a/opencl/source/gen9/command_stream_receiver_simulated_common_hw_gen9.cpp
+++ b/opencl/source/gen9/command_stream_receiver_simulated_common_hw_gen9.cpp
@@ -5,7 +5,7 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_simulated_common_hw_bdw_and_later.inl"
+#include "shared/source/command_stream/command_stream_receiver_simulated_common_hw_bdw_and_later.inl"

 namespace NEO {
 typedef SKLFamily Family;
--- a/opencl/source/gen9/enable_family_full_ocl_gen9.cpp
+++ b/opencl/source/gen9/enable_family_full_ocl_gen9.cpp
@@ -5,12 +5,12 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/helpers/populate_factory.h"

 #include "opencl/source/command_queue/command_queue_hw.h"
-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/source/device_queue/device_queue_hw.h"
 #include "opencl/source/event/perf_counter.h"
 #include "opencl/source/helpers/cl_hw_helper.h"
--- a/opencl/source/gen9/linux/command_stream_receiver_gen9.cpp
+++ b/opencl/source/gen9/linux/command_stream_receiver_gen9.cpp
@@ -5,7 +5,8 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+
 #include "opencl/source/os_interface/linux/device_command_stream.inl"
 #include "opencl/source/os_interface/linux/drm_command_stream.inl"
 #include "opencl/source/os_interface/linux/drm_command_stream_bdw_and_later.inl"
--- a/opencl/source/gen9/tbx_command_stream_receiver_gen9.cpp
+++ b/opencl/source/gen9/tbx_command_stream_receiver_gen9.cpp
@@ -5,14 +5,13 @@
 *
 */

+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.inl"
 #include "shared/source/gen9/hw_cmds.h"
 #include "shared/source/helpers/array_count.h"
 #include "shared/source/helpers/populate_factory.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
-
 namespace NEO {

 typedef SKLFamily Family;
--- a/opencl/source/gen9/windows/command_stream_receiver_gen9.cpp
+++ b/opencl/source/gen9/windows/command_stream_receiver_gen9.cpp
@@ -5,7 +5,8 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+
 #include "opencl/source/os_interface/windows/device_command_stream.inl"
 #include "opencl/source/os_interface/windows/wddm_device_command_stream.inl"

--- a/opencl/source/helpers/CMakeLists.txt
+++ b/opencl/source/helpers/CMakeLists.txt
@@ -30,8 +30,6 @@ set(RUNTIME_SRCS_HELPERS_BASE
    ${CMAKE_CURRENT_SOURCE_DIR}/hardware_commands_helper.h
    ${CMAKE_CURRENT_SOURCE_DIR}/hardware_commands_helper_base.inl
    ${CMAKE_CURRENT_SOURCE_DIR}/hardware_commands_helper_bdw_and_later.inl
-    ${CMAKE_CURRENT_SOURCE_DIR}/hardware_context_controller.cpp
-    ${CMAKE_CURRENT_SOURCE_DIR}/hardware_context_controller.h
    ${CMAKE_CURRENT_SOURCE_DIR}/helper_options.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/implicit_scaling_ocl.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}${BRANCH_DIR_SUFFIX}memory_properties_helpers.cpp
--- a/opencl/source/helpers/api_specific_config_ocl.cpp
+++ b/opencl/source/helpers/api_specific_config_ocl.cpp
@@ -8,6 +8,8 @@
 #include "shared/source/debug_settings/debug_settings_manager.h"
 #include "shared/source/helpers/api_specific_config.h"

+#include "opencl/source/os_interface/ocl_reg_path.h"
+
 namespace NEO {
 bool ApiSpecificConfig::isStatelessCompressionSupported() {
    return true;
@@ -37,4 +39,7 @@ uint64_t ApiSpecificConfig::getReducedMaxAllocSize(uint64_t maxAllocSize) {
    return maxAllocSize / 2;
 }

+const char *ApiSpecificConfig::getRegistryPath() {
+    return oclRegPath;
+}
 } // namespace NEO
--- a/opencl/source/helpers/hardware_context_controller.cpp
+++ b/opencl/source/helpers/hardware_context_controller.cpp
@@ -5,7 +5,7 @@
 *
 */

-#include "opencl/source/helpers/hardware_context_controller.h"
+#include "shared/source/helpers/hardware_context_controller.h"

 #include "shared/source/helpers/constants.h"
 #include "shared/source/os_interface/os_context.h"
--- a/opencl/source/helpers/windows/gmm_callbacks_tgllp_and_later.inl
+++ b/opencl/source/helpers/windows/gmm_callbacks_tgllp_and_later.inl
@@ -5,12 +5,12 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "shared/source/helpers/hw_helper.h"
 #include "shared/source/helpers/windows/gmm_callbacks.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "opencl/source/os_interface/windows/wddm_device_command_stream.h"

 namespace NEO {
--- a/opencl/source/os_interface/linux/device_command_stream.inl
+++ b/opencl/source/os_interface/linux/device_command_stream.inl
@@ -5,11 +5,11 @@
 *
 */

+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "shared/source/command_stream/device_command_stream.h"
 #include "shared/source/debug_settings/debug_settings_manager.h"
 #include "shared/source/helpers/api_specific_config.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "opencl/source/os_interface/linux/drm_command_stream.h"

 namespace NEO {
--- a/opencl/source/os_interface/windows/device_command_stream.inl
+++ b/opencl/source/os_interface/windows/device_command_stream.inl
@@ -9,10 +9,10 @@
 // Current order must be preserved due to two versions of igfxfmid.h
 #pragma warning(push)
 #pragma warning(disable : 4005)
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "shared/source/command_stream/device_command_stream.h"
 #include "shared/source/helpers/api_specific_config.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "opencl/source/os_interface/windows/wddm_device_command_stream.h"

 #include "hw_cmds.h"
--- a/opencl/source/xe_hp_core/aub_command_stream_receiver_xe_hp_core.cpp
+++ b/opencl/source/xe_hp_core/aub_command_stream_receiver_xe_hp_core.cpp
@@ -5,11 +5,10 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw_xehp_and_later.inl"
 #include "shared/source/helpers/array_count.h"
 #include "shared/source/helpers/populate_factory.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw_xehp_and_later.inl"
-
 namespace NEO {

 typedef XeHpFamily Family;
--- a/opencl/source/xe_hp_core/command_stream_receiver_simulated_common_hw_xe_hp_core.cpp
+++ b/opencl/source/xe_hp_core/command_stream_receiver_simulated_common_hw_xe_hp_core.cpp
@@ -5,7 +5,7 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_simulated_common_hw_xehp_and_later.inl"
+#include "shared/source/command_stream/command_stream_receiver_simulated_common_hw_xehp_and_later.inl"

 namespace NEO {
 typedef XeHpFamily Family;
--- a/opencl/source/xe_hp_core/enable_family_full_ocl_xe_hp_core.cpp
+++ b/opencl/source/xe_hp_core/enable_family_full_ocl_xe_hp_core.cpp
@@ -5,12 +5,12 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/helpers/populate_factory.h"

 #include "opencl/source/command_queue/command_queue_hw.h"
-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/source/device_queue/device_queue_hw.h"
 #include "opencl/source/helpers/cl_hw_helper.h"
 #include "opencl/source/mem_obj/buffer.h"
--- a/opencl/source/xe_hp_core/linux/command_stream_receiver_xe_hp_core.cpp
+++ b/opencl/source/xe_hp_core/linux/command_stream_receiver_xe_hp_core.cpp
@@ -5,7 +5,8 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+
 #include "opencl/source/os_interface/linux/device_command_stream.inl"
 #include "opencl/source/os_interface/linux/drm_command_stream.inl"
 #include "opencl/source/os_interface/linux/drm_command_stream_xehp_and_later.inl"
--- a/opencl/source/xe_hp_core/tbx_command_stream_receiver_xe_hp_core.cpp
+++ b/opencl/source/xe_hp_core/tbx_command_stream_receiver_xe_hp_core.cpp
@@ -5,6 +5,7 @@
 *
 */

+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.inl"
 #include "shared/source/helpers/array_count.h"
@@ -13,8 +14,6 @@
 #include "shared/source/memory_manager/memory_pool.h"
 #include "shared/source/xe_hp_core/hw_cmds.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
-
 namespace NEO {
 typedef XeHpFamily Family;
 static auto gfxCore = IGFX_XE_HP_CORE;
@@ -26,7 +25,7 @@ void populateFactoryTable<TbxCommandStreamReceiverHw<Family>>() {
    tbxCommandStreamReceiverFactory[gfxCore] = TbxCommandStreamReceiverHw<Family>::create;
 }

-#include "opencl/source/command_stream/tbx_command_stream_receiver_xehp_and_later.inl"
+#include "shared/source/command_stream/tbx_command_stream_receiver_xehp_and_later.inl"

 template class TbxCommandStreamReceiverHw<Family>;
 template class CommandStreamReceiverWithAUBDump<TbxCommandStreamReceiverHw<Family>>;
--- a/opencl/source/xe_hp_core/windows/command_stream_receiver_xe_hp_core.cpp
+++ b/opencl/source/xe_hp_core/windows/command_stream_receiver_xe_hp_core.cpp
@@ -5,7 +5,8 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
+
 #include "opencl/source/os_interface/windows/device_command_stream.inl"
 #include "opencl/source/os_interface/windows/wddm_device_command_stream.inl"

--- a/opencl/test/unit_test/aub/aub_helper_tests.cpp
+++ b/opencl/test/unit_test/aub/aub_helper_tests.cpp
@@ -7,12 +7,12 @@

 #include "shared/source/aub/aub_helper.h"
 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/helpers/basic_math.h"
 #include "shared/source/helpers/hw_helper.h"
 #include "shared/source/tbx/tbx_proto.h"
 #include "shared/test/common/helpers/debug_manager_state_restore.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/test/unit_test/fixtures/cl_device_fixture.h"
 #include "opencl/test/unit_test/mocks/mock_lrca_helper.h"
 #include "test.h"
--- a/opencl/test/unit_test/aub_tests/command_stream/aub_command_stream_fixture.h
+++ b/opencl/test/unit_test/aub_tests/command_stream/aub_command_stream_fixture.h
@@ -7,14 +7,14 @@

 #pragma once
 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/memory_manager/internal_allocation_storage.h"
 #include "shared/source/memory_manager/memory_banks.h"
 #include "shared/source/os_interface/os_context.h"
 #include "shared/test/unit_test/tests_configuration.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "opencl/test/unit_test/command_stream/command_stream_fixture.h"
 #include "opencl/test/unit_test/mocks/mock_allocation_properties.h"

--- a/opencl/test/unit_test/aub_tests/command_stream/aub_command_stream_tests.cpp
+++ b/opencl/test/unit_test/aub_tests/command_stream/aub_command_stream_tests.cpp
@@ -5,12 +5,12 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/helpers/ptr_math.h"
 #include "shared/source/os_interface/os_context.h"
 #include "shared/test/common/cmd_parse/gen_cmd_parse.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/test/unit_test/command_queue/command_queue_fixture.h"
 #include "opencl/test/unit_test/fixtures/cl_device_fixture.h"
 #include "test.h"
--- a/opencl/test/unit_test/aub_tests/command_stream/aub_mem_dump_tests.h
+++ b/opencl/test/unit_test/aub_tests/command_stream/aub_mem_dump_tests.h
@@ -6,13 +6,13 @@
 */

 #pragma once
+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/device/device.h"
 #include "shared/source/helpers/aligned_memory.h"
 #include "shared/source/helpers/api_specific_config.h"
 #include "shared/source/helpers/ptr_math.h"
 #include "shared/source/os_interface/hw_info_config.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "test.h"

 #include "aub_mapper.h"
--- a/opencl/test/unit_test/aub_tests/command_stream/aub_mi_atomic_tests.cpp
+++ b/opencl/test/unit_test/aub_tests/command_stream/aub_mi_atomic_tests.cpp
@@ -5,10 +5,10 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/test/common/helpers/dispatch_flags_helper.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/test/unit_test/aub_tests/fixtures/aub_fixture.h"
 #include "test.h"

--- a/opencl/test/unit_test/aub_tests/fixtures/aub_fixture.h
+++ b/opencl/test/unit_test/aub_tests/fixtures/aub_fixture.h
@@ -7,6 +7,8 @@

 #pragma once
 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/helpers/api_specific_config.h"
 #include "shared/source/helpers/hw_helper.h"
@@ -15,8 +17,6 @@
 #include "shared/test/common/mocks/mock_memory_operations_handler.h"
 #include "shared/test/unit_test/tests_configuration.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "opencl/source/platform/platform.h"
 #include "opencl/test/unit_test/command_queue/command_queue_fixture.h"
 #include "opencl/test/unit_test/mocks/mock_cl_device.h"
--- a/opencl/test/unit_test/command_stream/aub_command_stream_receiver_1_tests.cpp
+++ b/opencl/test/unit_test/command_stream/aub_command_stream_receiver_1_tests.cpp
@@ -6,6 +6,7 @@
 */

 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
+#include "shared/source/helpers/hardware_context_controller.h"
 #include "shared/source/helpers/hw_helper.h"
 #include "shared/source/os_interface/os_context.h"
 #include "shared/test/common/helpers/debug_manager_state_restore.h"
@@ -16,7 +17,6 @@
 #include "shared/test/common/mocks/mock_graphics_allocation.h"
 #include "shared/test/unit_test/fixtures/mock_aub_center_fixture.h"

-#include "opencl/source/helpers/hardware_context_controller.h"
 #include "opencl/test/unit_test/fixtures/aub_command_stream_receiver_fixture.h"
 #include "opencl/test/unit_test/fixtures/cl_device_fixture.h"
 #include "opencl/test/unit_test/libult/ult_aub_command_stream_receiver.h"
--- a/opencl/test/unit_test/command_stream/aub_command_stream_receiver_2_tests.cpp
+++ b/opencl/test/unit_test/command_stream/aub_command_stream_receiver_2_tests.cpp
@@ -7,6 +7,7 @@

 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
 #include "shared/source/helpers/engine_node_helper.h"
+#include "shared/source/helpers/hardware_context_controller.h"
 #include "shared/source/helpers/hw_helper.h"
 #include "shared/source/os_interface/os_context.h"
 #include "shared/test/common/helpers/debug_manager_state_restore.h"
@@ -18,7 +19,6 @@
 #include "shared/test/common/mocks/ult_device_factory.h"
 #include "shared/test/unit_test/fixtures/mock_aub_center_fixture.h"

-#include "opencl/source/helpers/hardware_context_controller.h"
 #include "opencl/source/mem_obj/mem_obj_helper.h"
 #include "opencl/source/platform/platform.h"
 #include "opencl/test/unit_test/fixtures/aub_command_stream_receiver_fixture.h"
--- a/opencl/test/unit_test/command_stream/aub_command_stream_receiver_3_tests.cpp
+++ b/opencl/test/unit_test/command_stream/aub_command_stream_receiver_3_tests.cpp
@@ -6,11 +6,11 @@
 */

 #include "shared/source/aub_mem_dump/aub_alloc_dump.h"
+#include "shared/source/helpers/hardware_context_controller.h"
 #include "shared/test/common/helpers/debug_manager_state_restore.h"
 #include "shared/test/common/helpers/engine_descriptor_helper.h"
 #include "shared/test/common/mocks/mock_aub_manager.h"

-#include "opencl/source/helpers/hardware_context_controller.h"
 #include "opencl/test/unit_test/fixtures/aub_command_stream_receiver_fixture.h"
 #include "opencl/test/unit_test/mocks/mock_aub_csr.h"
 #include "opencl/test/unit_test/mocks/mock_memory_manager.h"
--- a/opencl/test/unit_test/command_stream/aub_command_stream_receiver_tests_xehp_and_later.cpp
+++ b/opencl/test/unit_test/command_stream/aub_command_stream_receiver_tests_xehp_and_later.cpp
@@ -6,6 +6,7 @@
 */

 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/helpers/engine_node_helper.h"
 #include "shared/source/memory_manager/memory_banks.h"
 #include "shared/source/memory_manager/memory_pool.h"
@@ -17,7 +18,6 @@
 #include "shared/test/common/mocks/mock_execution_environment.h"
 #include "shared/test/common/mocks/mock_graphics_allocation.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/source/helpers/memory_properties_helpers.h"
 #include "opencl/source/mem_obj/buffer.h"
 #include "opencl/test/unit_test/fixtures/cl_device_fixture.h"
--- a/opencl/test/unit_test/command_stream/aub_file_stream_tests.cpp
+++ b/opencl/test/unit_test/command_stream/aub_file_stream_tests.cpp
@@ -6,6 +6,9 @@
 */

 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
+#include "shared/source/helpers/hardware_context_controller.h"
+#include "shared/source/helpers/neo_driver_version.h"
 #include "shared/source/os_interface/os_context.h"
 #include "shared/test/common/helpers/debug_manager_state_restore.h"
 #include "shared/test/common/helpers/engine_descriptor_helper.h"
@@ -15,9 +18,6 @@
 #include "shared/test/common/mocks/mock_execution_environment.h"
 #include "shared/test/common/mocks/mock_graphics_allocation.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
-#include "opencl/source/helpers/hardware_context_controller.h"
-#include "opencl/source/helpers/neo_driver_version.h"
 #include "opencl/test/unit_test/fixtures/aub_command_stream_receiver_fixture.h"
 #include "opencl/test/unit_test/fixtures/cl_device_fixture.h"
 #include "opencl/test/unit_test/mocks/mock_aub_csr.h"
--- a/opencl/test/unit_test/command_stream/command_stream_receiver_simulated_common_hw_tests_xehp_and_later.cpp
+++ b/opencl/test/unit_test/command_stream/command_stream_receiver_simulated_common_hw_tests_xehp_and_later.cpp
@@ -6,10 +6,10 @@
 */

 #include "shared/source/aub/aub_helper.h"
+#include "shared/source/command_stream/command_stream_receiver_simulated_common_hw.h"
+#include "shared/source/helpers/hardware_context_controller.h"
 #include "shared/test/common/helpers/debug_manager_state_restore.h"

-#include "opencl/source/command_stream/command_stream_receiver_simulated_common_hw.h"
-#include "opencl/source/helpers/hardware_context_controller.h"
 #include "opencl/test/unit_test/fixtures/cl_device_fixture.h"
 #include "opencl/test/unit_test/mocks/mock_aub_stream.h"
 #include "opencl/test/unit_test/mocks/mock_csr_simulated_common_hw.h"
--- a/opencl/test/unit_test/command_stream/command_stream_receiver_simulated_tests.cpp
+++ b/opencl/test/unit_test/command_stream/command_stream_receiver_simulated_tests.cpp
@@ -5,7 +5,9 @@
 *
 */

+#include "shared/source/command_stream/command_stream_receiver_simulated_hw.h"
 #include "shared/source/helpers/array_count.h"
+#include "shared/source/helpers/hardware_context_controller.h"
 #include "shared/source/memory_manager/memory_pool.h"
 #include "shared/source/memory_manager/os_agnostic_memory_manager.h"
 #include "shared/source/os_interface/os_context.h"
@@ -14,8 +16,6 @@
 #include "shared/test/common/mocks/mock_aub_manager.h"
 #include "shared/test/common/mocks/mock_graphics_allocation.h"

-#include "opencl/source/command_stream/definitions/command_stream_receiver_simulated_hw.h"
-#include "opencl/source/helpers/hardware_context_controller.h"
 #include "opencl/test/unit_test/helpers/hw_helper_tests.h"
 #include "opencl/test/unit_test/mocks/mock_csr_simulated_common_hw.h"
 #include "opencl/test/unit_test/mocks/mock_gmm.h"
--- a/opencl/test/unit_test/command_stream/command_stream_receiver_tests.cpp
+++ b/opencl/test/unit_test/command_stream/command_stream_receiver_tests.cpp
@@ -6,6 +6,7 @@
 */

 #include "shared/source/command_stream/command_stream_receiver.h"
+#include "shared/source/command_stream/command_stream_receiver_simulated_hw.h"
 #include "shared/source/command_stream/linear_stream.h"
 #include "shared/source/command_stream/preemption.h"
 #include "shared/source/command_stream/scratch_space_controller.h"
@@ -30,7 +31,6 @@
 #include "shared/test/common/test_macros/test_checks_shared.h"
 #include "shared/test/unit_test/direct_submission/direct_submission_controller_mock.h"

-#include "opencl/source/command_stream/definitions/command_stream_receiver_simulated_hw.h"
 #include "opencl/source/mem_obj/buffer.h"
 #include "opencl/source/platform/platform.h"
 #include "opencl/test/unit_test/fixtures/cl_device_fixture.h"
--- a/opencl/test/unit_test/command_stream/command_stream_receiver_with_aub_dump_tests.cpp
+++ b/opencl/test/unit_test/command_stream/command_stream_receiver_with_aub_dump_tests.cpp
@@ -5,6 +5,9 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.inl"
 #include "shared/source/command_stream/preemption.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/execution_environment/execution_environment.h"
@@ -20,9 +23,6 @@
 #include "shared/test/common/mocks/mock_execution_environment.h"
 #include "shared/test/unit_test/fixtures/mock_aub_center_fixture.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.inl"
 #include "opencl/source/helpers/dispatch_info.h"
 #include "opencl/source/platform/platform.h"
 #include "opencl/test/unit_test/libult/ult_command_stream_receiver.h"
--- a/opencl/test/unit_test/command_stream/create_command_stream_receiver_tests.cpp
+++ b/opencl/test/unit_test/command_stream/create_command_stream_receiver_tests.cpp
@@ -6,6 +6,7 @@
 */

 #include "shared/source/command_stream/command_stream_receiver.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "shared/source/execution_environment/execution_environment.h"
 #include "shared/source/memory_manager/os_agnostic_memory_manager.h"
 #include "shared/test/common/helpers/debug_manager_state_restore.h"
@@ -14,7 +15,6 @@
 #include "shared/test/common/libult/create_command_stream.h"
 #include "shared/test/unit_test/fixtures/mock_aub_center_fixture.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "opencl/test/unit_test/helpers/execution_environment_helper.h"
 #include "test.h"

--- a/opencl/test/unit_test/command_stream/tbx_command_stream_tests.cpp
+++ b/opencl/test/unit_test/command_stream/tbx_command_stream_tests.cpp
@@ -8,9 +8,11 @@
 #include "shared/source/aub_mem_dump/aub_alloc_dump.h"
 #include "shared/source/command_stream/aub_command_stream_receiver.h"
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver_hw.h"
 #include "shared/source/debug_settings/debug_settings_manager.h"
 #include "shared/source/helpers/engine_node_helper.h"
+#include "shared/source/helpers/hardware_context_controller.h"
 #include "shared/source/helpers/hw_helper.h"
 #include "shared/source/helpers/ptr_math.h"
 #include "shared/source/memory_manager/memory_banks.h"
@@ -26,8 +28,6 @@
 #include "shared/test/common/mocks/mock_tbx_csr.h"
 #include "shared/test/unit_test/fixtures/mock_aub_center_fixture.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
-#include "opencl/source/helpers/hardware_context_controller.h"
 #include "opencl/source/mem_obj/buffer.h"
 #include "opencl/source/mem_obj/mem_obj.h"
 #include "opencl/source/platform/platform.h"
--- a/opencl/test/unit_test/gen12lp/aub_command_stream_receiver_tests_gen12lp.inl
+++ b/opencl/test/unit_test/gen12lp/aub_command_stream_receiver_tests_gen12lp.inl
@@ -6,12 +6,12 @@
 */

 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/os_interface/os_context.h"
 #include "shared/test/common/helpers/debug_manager_state_restore.h"
 #include "shared/test/common/helpers/engine_descriptor_helper.h"
 #include "shared/test/common/mocks/mock_graphics_allocation.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/test/unit_test/fixtures/cl_device_fixture.h"
 #include "opencl/test/unit_test/helpers/hw_helper_tests.h"
 #include "opencl/test/unit_test/mocks/mock_aub_csr.h"
--- a/opencl/test/unit_test/gen12lp/command_stream_receiver_simulated_common_hw_tests_gen12lp.inl
+++ b/opencl/test/unit_test/gen12lp/command_stream_receiver_simulated_common_hw_tests_gen12lp.inl
@@ -5,9 +5,9 @@
 *
 */

+#include "shared/source/command_stream/command_stream_receiver_simulated_hw.h"
 #include "shared/test/common/helpers/debug_manager_state_restore.h"

-#include "opencl/source/command_stream/definitions/command_stream_receiver_simulated_hw.h"
 #include "opencl/test/unit_test/fixtures/cl_device_fixture.h"
 #include "opencl/test/unit_test/mocks/mock_aub_stream.h"
 #include "test.h"
--- a/opencl/test/unit_test/gen12lp/windows/gmm_callbacks_tests_gen12lp.cpp
+++ b/opencl/test/unit_test/gen12lp/windows/gmm_callbacks_tests_gen12lp.cpp
@@ -5,6 +5,8 @@
 *
 */

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "shared/source/command_stream/linear_stream.h"
 #include "shared/source/execution_environment/execution_environment.h"
 #include "shared/source/helpers/hw_helper.h"
@@ -13,8 +15,6 @@
 #include "shared/test/common/helpers/debug_manager_state_restore.h"
 #include "shared/test/common/helpers/default_hw_info.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "opencl/source/os_interface/windows/wddm_device_command_stream.h"
 #include "opencl/source/platform/platform.h"
 #include "opencl/test/unit_test/helpers/execution_environment_helper.h"
--- a/opencl/test/unit_test/helpers/api_specific_config_ocl_tests.cpp
+++ b/opencl/test/unit_test/helpers/api_specific_config_ocl_tests.cpp
@@ -7,6 +7,8 @@

 #include "shared/source/helpers/api_specific_config.h"

+#include "opencl/source/os_interface/ocl_reg_path.h"
+
 #include "gtest/gtest.h"

 namespace NEO {
@@ -31,4 +33,7 @@ TEST(ApiSpecificConfigOclTests, givenMaxAllocSizeWhenGettingReducedMaxAllocSizeT
    EXPECT_EQ(512u, ApiSpecificConfig::getReducedMaxAllocSize(1024));
 }

+TEST(ApiSpecificConfigOclTests, WhenGettingRegistryPathThenOclRegistryPathIsReturned) {
+    EXPECT_STREQ(oclRegPath, ApiSpecificConfig::getRegistryPath());
+}
 } // namespace NEO
--- a/opencl/test/unit_test/libult/ult_aub_command_stream_receiver.h
+++ b/opencl/test/unit_test/libult/ult_aub_command_stream_receiver.h
@@ -6,12 +6,11 @@
 */

 #pragma once
+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/execution_environment/execution_environment.h"
 #include "shared/source/memory_manager/os_agnostic_memory_manager.h"
 #include "shared/test/common/mocks/mock_experimental_command_buffer.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
-
 namespace NEO {

 template <typename GfxFamily>
--- a/opencl/test/unit_test/mocks/mock_aub_csr.h
+++ b/opencl/test/unit_test/mocks/mock_aub_csr.h
@@ -7,6 +7,7 @@

 #pragma once

+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/preemption.h"
 #include "shared/source/execution_environment/execution_environment.h"
 #include "shared/source/execution_environment/root_device_environment.h"
@@ -15,7 +16,6 @@
 #include "shared/test/common/helpers/dispatch_flags_helper.h"
 #include "shared/test/common/helpers/engine_descriptor_helper.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "opencl/source/platform/platform.h"
 #include "opencl/test/unit_test/mocks/mock_allocation_properties.h"

--- a/opencl/test/unit_test/mocks/mock_csr_simulated_common_hw.h
+++ b/opencl/test/unit_test/mocks/mock_csr_simulated_common_hw.h
@@ -7,7 +7,7 @@

 #pragma once

-#include "opencl/source/command_stream/definitions/command_stream_receiver_simulated_hw.h"
+#include "shared/source/command_stream/command_stream_receiver_simulated_hw.h"

 namespace NEO {

--- a/opencl/test/unit_test/mocks/ult_cl_device_factory.cpp
+++ b/opencl/test/unit_test/mocks/ult_cl_device_factory.cpp
@@ -7,11 +7,11 @@

 #include "opencl/test/unit_test/mocks/ult_cl_device_factory.h"

+#include "shared/source/command_stream/create_command_stream_impl.h"
 #include "shared/source/os_interface/device_factory.h"
 #include "shared/test/common/helpers/ult_hw_config.h"
 #include "shared/test/common/mocks/ult_device_factory.h"

-#include "opencl/source/command_stream/create_command_stream_impl.h"
 #include "opencl/test/unit_test/mocks/mock_cl_device.h"
 #include "opencl/test/unit_test/mocks/mock_memory_manager.h"

--- a/opencl/test/unit_test/os_interface/windows/device_command_stream_tests.cpp
+++ b/opencl/test/unit_test/os_interface/windows/device_command_stream_tests.cpp
@@ -7,6 +7,7 @@

 #include "shared/source/command_stream/aub_command_stream_receiver.h"
 #include "shared/source/command_stream/command_stream_receiver.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "shared/source/command_stream/device_command_stream.h"
 #include "shared/source/command_stream/linear_stream.h"
 #include "shared/source/command_stream/preemption.h"
@@ -33,7 +34,6 @@
 #include "shared/test/unit_test/fixtures/mock_aub_center_fixture.h"
 #include "shared/test/unit_test/os_interface/windows/mock_gdi_interface.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "opencl/source/mem_obj/buffer.h"
 #include "opencl/source/os_interface/windows/wddm_device_command_stream.h"
 #include "opencl/source/platform/platform.h"
--- a/opencl/test/unit_test/perf_tests/fixtures/device_fixture.cpp
+++ b/opencl/test/unit_test/perf_tests/fixtures/device_fixture.cpp
@@ -7,7 +7,7 @@

 #include "device_fixture.h"

-#include "opencl/source/command_stream/command_stream_receiver.h"
+#include "shared/source/command_stream/command_stream_receiver.h"

 #include "gtest/gtest.h"

--- a/shared/source/command_stream/CMakeLists.txt
+++ b/shared/source/command_stream/CMakeLists.txt
@@ -8,6 +8,9 @@ set(NEO_CORE_COMMAND_STREAM
    ${CMAKE_CURRENT_SOURCE_DIR}/CMakeLists.txt
    ${CMAKE_CURRENT_SOURCE_DIR}/aub_command_stream_receiver.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/aub_command_stream_receiver.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/aub_command_stream_receiver_hw.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/aub_command_stream_receiver_hw_base.inl
+    ${CMAKE_CURRENT_SOURCE_DIR}/aub_command_stream_receiver_hw_bdw_and_later.inl
    ${CMAKE_CURRENT_SOURCE_DIR}/aub_subcapture_status.h
    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver.h
@@ -15,6 +18,14 @@ set(NEO_CORE_COMMAND_STREAM
    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_hw_base.inl
    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_hw_bdw_and_later.inl
    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_hw_tgllp_and_later.inl
+    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_simulated_common_hw.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_simulated_common_hw_base.inl
+    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_simulated_common_hw_bdw_and_later.inl
+    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_simulated_hw.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_with_aub_dump.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_with_aub_dump.inl
+    ${CMAKE_CURRENT_SOURCE_DIR}/create_command_stream_impl.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/create_command_stream_impl.h
    ${CMAKE_CURRENT_SOURCE_DIR}/csr_definitions.h
    ${CMAKE_CURRENT_SOURCE_DIR}/csr_deps.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/csr_deps.h
@@ -50,10 +61,13 @@ set(NEO_CORE_COMMAND_STREAM

 if(SUPPORT_XEHP_AND_LATER)
  list(APPEND NEO_CORE_COMMAND_STREAM
+       ${CMAKE_CURRENT_SOURCE_DIR}/aub_command_stream_receiver_hw_xehp_and_later.inl
       ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_hw_xehp_and_later.inl
+       ${CMAKE_CURRENT_SOURCE_DIR}/command_stream_receiver_simulated_common_hw_xehp_and_later.inl
       ${CMAKE_CURRENT_SOURCE_DIR}/preemption_xehp_and_later.inl
       ${CMAKE_CURRENT_SOURCE_DIR}/scratch_space_controller_xehp_and_later.cpp
       ${CMAKE_CURRENT_SOURCE_DIR}/scratch_space_controller_xehp_and_later.h
+       ${CMAKE_CURRENT_SOURCE_DIR}/tbx_command_stream_receiver_xehp_and_later.inl
  )
 endif()

--- a/shared/source/command_stream/aub_command_stream_receiver_hw.h
+++ b/shared/source/command_stream/aub_command_stream_receiver_hw.h
@@ -8,14 +8,13 @@
 #pragma once
 #include "shared/source/aub/aub_center.h"
 #include "shared/source/command_stream/aub_command_stream_receiver.h"
+#include "shared/source/command_stream/command_stream_receiver_simulated_hw.h"
 #include "shared/source/helpers/array_count.h"
 #include "shared/source/memory_manager/os_agnostic_memory_manager.h"
 #include "shared/source/memory_manager/page_table.h"
 #include "shared/source/memory_manager/physical_address_allocator.h"
 #include "shared/source/utilities/spinlock.h"

-#include "opencl/source/command_stream/definitions/command_stream_receiver_simulated_hw.h"
-
 #include "aub_mapper.h"

 namespace NEO {
--- a/shared/source/command_stream/aub_command_stream_receiver_hw_base.inl
+++ b/shared/source/command_stream/aub_command_stream_receiver_hw_base.inl
@@ -11,15 +11,19 @@
 #include "shared/source/aub_mem_dump/aub_alloc_dump.h"
 #include "shared/source/aub_mem_dump/aub_alloc_dump.inl"
 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw.h"
 #include "shared/source/command_stream/command_stream_receiver.h"
 #include "shared/source/debug_settings/debug_settings_manager.h"
 #include "shared/source/execution_environment/execution_environment.h"
 #include "shared/source/execution_environment/root_device_environment.h"
 #include "shared/source/helpers/aligned_memory.h"
+#include "shared/source/helpers/api_specific_config.h"
 #include "shared/source/helpers/constants.h"
 #include "shared/source/helpers/debug_helpers.h"
 #include "shared/source/helpers/engine_node_helper.h"
+#include "shared/source/helpers/hardware_context_controller.h"
 #include "shared/source/helpers/hash.h"
+#include "shared/source/helpers/neo_driver_version.h"
 #include "shared/source/helpers/ptr_math.h"
 #include "shared/source/helpers/string.h"
 #include "shared/source/memory_manager/graphics_allocation.h"
@@ -28,12 +32,6 @@
 #include "shared/source/os_interface/hw_info_config.h"
 #include "shared/source/os_interface/os_context.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw.h"
-#include "opencl/source/helpers/hardware_context_controller.h"
-#include "opencl/source/helpers/neo_driver_version.h"
-#include "opencl/source/os_interface/ocl_reg_path.h"
-
-#include "driver_version.h"
 #include "third_party/aub_stream/headers/aub_manager.h"
 #include "third_party/aub_stream/headers/aubstream.h"

@@ -57,7 +55,7 @@ AUBCommandStreamReceiverHw<GfxFamily>::AUBCommandStreamReceiverHw(const std::str

    auto subCaptureCommon = aubCenter->getSubCaptureCommon();
    UNRECOVERABLE_IF(nullptr == subCaptureCommon);
-    subCaptureManager = std::make_unique<AubSubCaptureManager>(fileName, *subCaptureCommon, oclRegPath);
+    subCaptureManager = std::make_unique<AubSubCaptureManager>(fileName, *subCaptureCommon, ApiSpecificConfig::getRegistryPath());

    aubManager = aubCenter->getAubManager();

--- a/shared/source/command_stream/aub_command_stream_receiver_hw_bdw_and_later.inl
+++ b/shared/source/command_stream/aub_command_stream_receiver_hw_bdw_and_later.inl
@@ -5,7 +5,7 @@
 *
 */

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw_base.inl"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw_base.inl"

 namespace NEO {

--- a/shared/source/command_stream/aub_command_stream_receiver_hw_xehp_and_later.inl
+++ b/shared/source/command_stream/aub_command_stream_receiver_hw_xehp_and_later.inl
@@ -6,10 +6,9 @@
 */

 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
+#include "shared/source/command_stream/aub_command_stream_receiver_hw_base.inl"
 #include "shared/source/helpers/engine_node_helper.h"

-#include "opencl/source/command_stream/aub_command_stream_receiver_hw_base.inl"
-
 namespace NEO {

 template <typename GfxFamily>
--- a/shared/source/command_stream/command_stream_receiver_simulated_common_hw.h
+++ b/shared/source/command_stream/command_stream_receiver_simulated_common_hw.h
--- a/shared/source/command_stream/command_stream_receiver_simulated_common_hw_base.inl
+++ b/shared/source/command_stream/command_stream_receiver_simulated_common_hw_base.inl
@@ -7,17 +7,16 @@

 #include "shared/source/aub/aub_helper.h"
 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
+#include "shared/source/command_stream/command_stream_receiver_simulated_common_hw.h"
 #include "shared/source/debug_settings/debug_settings_manager.h"
 #include "shared/source/gmm_helper/gmm.h"
 #include "shared/source/gmm_helper/gmm_helper.h"
 #include "shared/source/gmm_helper/resource_info.h"
+#include "shared/source/helpers/hardware_context_controller.h"
 #include "shared/source/memory_manager/address_mapper.h"
 #include "shared/source/memory_manager/memory_manager.h"
 #include "shared/source/os_interface/os_context.h"

-#include "opencl/source/command_stream/command_stream_receiver_simulated_common_hw.h"
-#include "opencl/source/helpers/hardware_context_controller.h"
-
 #include "third_party/aub_stream/headers/aub_manager.h"

 namespace NEO {
--- a/shared/source/command_stream/command_stream_receiver_simulated_common_hw_bdw_and_later.inl
+++ b/shared/source/command_stream/command_stream_receiver_simulated_common_hw_bdw_and_later.inl
@@ -5,7 +5,7 @@
 *
 */

-#include "opencl/source/command_stream/command_stream_receiver_simulated_common_hw_base.inl"
+#include "shared/source/command_stream/command_stream_receiver_simulated_common_hw_base.inl"

 namespace NEO {

--- a/shared/source/command_stream/command_stream_receiver_simulated_common_hw_xehp_and_later.inl
+++ b/shared/source/command_stream/command_stream_receiver_simulated_common_hw_xehp_and_later.inl
@@ -6,14 +6,13 @@
 */

 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
+#include "shared/source/command_stream/command_stream_receiver_simulated_common_hw_base.inl"
 #include "shared/source/debug_settings/debug_settings_manager.h"
 #include "shared/source/helpers/hw_helper.h"
 #include "shared/source/memory_manager/memory_banks.h"
 #include "shared/source/memory_manager/memory_pool.h"
 #include "shared/source/memory_manager/physical_address_allocator.h"

-#include "opencl/source/command_stream/command_stream_receiver_simulated_common_hw_base.inl"
-
 namespace NEO {

 template <typename GfxFamily>
--- a/opencl/source/command_stream/definitions/command_stream_receiver_simulated_hw.h
+++ b/opencl/source/command_stream/definitions/command_stream_receiver_simulated_hw.h
@@ -7,7 +7,9 @@

 #pragma once
 #include "shared/source/aub/aub_helper.h"
+#include "shared/source/command_stream/command_stream_receiver_simulated_common_hw.h"
 #include "shared/source/helpers/debug_helpers.h"
+#include "shared/source/helpers/hardware_context_controller.h"
 #include "shared/source/helpers/hw_helper.h"
 #include "shared/source/memory_manager/graphics_allocation.h"
 #include "shared/source/memory_manager/memory_banks.h"
@@ -15,9 +17,6 @@
 #include "shared/source/memory_manager/physical_address_allocator.h"
 #include "shared/source/os_interface/os_context.h"

-#include "opencl/source/command_stream/command_stream_receiver_simulated_common_hw.h"
-#include "opencl/source/helpers/hardware_context_controller.h"
-
 #include "aub_mem_dump.h"
 #include "third_party/aub_stream/headers/allocation_params.h"
 #include "third_party/aub_stream/headers/aub_manager.h"
--- a/shared/source/command_stream/command_stream_receiver_with_aub_dump.h
+++ b/shared/source/command_stream/command_stream_receiver_with_aub_dump.h
--- a/shared/source/command_stream/command_stream_receiver_with_aub_dump.inl
+++ b/shared/source/command_stream/command_stream_receiver_with_aub_dump.inl
@@ -7,11 +7,10 @@

 #include "shared/source/aub/aub_center.h"
 #include "shared/source/command_stream/aub_command_stream_receiver.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "shared/source/execution_environment/execution_environment.h"
 #include "shared/source/execution_environment/root_device_environment.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
-
 namespace NEO {

 extern CommandStreamReceiverCreateFunc commandStreamReceiverFactory[2 * IGFX_MAX_CORE];
--- a/shared/source/command_stream/create_command_stream_impl.cpp
+++ b/shared/source/command_stream/create_command_stream_impl.cpp
@@ -6,14 +6,13 @@
 */

 #include "shared/source/command_stream/aub_command_stream_receiver.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver.h"
 #include "shared/source/execution_environment/execution_environment.h"
 #include "shared/source/execution_environment/root_device_environment.h"
 #include "shared/source/helpers/api_specific_config.h"
 #include "shared/source/os_interface/device_factory.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
-
 namespace NEO {

 extern CommandStreamReceiverCreateFunc commandStreamReceiverFactory[2 * IGFX_MAX_CORE];
--- a/shared/source/command_stream/create_command_stream_impl.h
+++ b/shared/source/command_stream/create_command_stream_impl.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2020 Intel Corporation
+ * Copyright (C) 2018-2021 Intel Corporation
 *
 * SPDX-License-Identifier: MIT
 *
--- a/shared/source/command_stream/tbx_command_stream_receiver_hw.h
+++ b/shared/source/command_stream/tbx_command_stream_receiver_hw.h
@@ -6,13 +6,12 @@
 */

 #pragma once
+#include "shared/source/command_stream/command_stream_receiver_simulated_hw.h"
 #include "shared/source/command_stream/tbx_command_stream_receiver.h"
 #include "shared/source/memory_manager/address_mapper.h"
 #include "shared/source/memory_manager/os_agnostic_memory_manager.h"
 #include "shared/source/memory_manager/page_table.h"

-#include "opencl/source/command_stream/definitions/command_stream_receiver_simulated_hw.h"
-
 #include "aub_mapper.h"

 #include <set>
--- a/shared/source/command_stream/tbx_command_stream_receiver_hw.inl
+++ b/shared/source/command_stream/tbx_command_stream_receiver_hw.inl
@@ -11,13 +11,16 @@
 #include "shared/source/aub_mem_dump/aub_alloc_dump.inl"
 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
 #include "shared/source/command_stream/aub_command_stream_receiver.h"
+#include "shared/source/command_stream/command_stream_receiver_with_aub_dump.h"
 #include "shared/source/debug_settings/debug_settings_manager.h"
 #include "shared/source/execution_environment/execution_environment.h"
 #include "shared/source/execution_environment/root_device_environment.h"
 #include "shared/source/helpers/aligned_memory.h"
+#include "shared/source/helpers/api_specific_config.h"
 #include "shared/source/helpers/constants.h"
 #include "shared/source/helpers/debug_helpers.h"
 #include "shared/source/helpers/engine_node_helper.h"
+#include "shared/source/helpers/hardware_context_controller.h"
 #include "shared/source/helpers/hw_helper.h"
 #include "shared/source/helpers/populate_factory.h"
 #include "shared/source/helpers/ptr_math.h"
@@ -27,10 +30,6 @@
 #include "shared/source/os_interface/hw_info_config.h"
 #include "shared/source/os_interface/os_context.h"

-#include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
-#include "opencl/source/helpers/hardware_context_controller.h"
-#include "opencl/source/os_interface/ocl_reg_path.h"
-
 #include <cstring>

 namespace NEO {
@@ -180,7 +179,7 @@ CommandStreamReceiver *TbxCommandStreamReceiverHw<GfxFamily>::create(const std::
        UNRECOVERABLE_IF(nullptr == subCaptureCommon);

        if (subCaptureCommon->subCaptureMode > AubSubCaptureManager::SubCaptureMode::Off) {
-            csr->subCaptureManager = std::make_unique<AubSubCaptureManager>(fullName, *subCaptureCommon, oclRegPath);
+            csr->subCaptureManager = std::make_unique<AubSubCaptureManager>(fullName, *subCaptureCommon, ApiSpecificConfig::getRegistryPath());
        }

        if (csr->aubManager) {
--- a/shared/source/command_stream/tbx_command_stream_receiver_xehp_and_later.inl
+++ b/shared/source/command_stream/tbx_command_stream_receiver_xehp_and_later.inl
--- a/shared/source/gen12lp/helpers_gen12lp_dg1.cpp
+++ b/shared/source/gen12lp/helpers_gen12lp_dg1.cpp
@@ -7,10 +7,9 @@

 #include "shared/source/aub_mem_dump/page_table_entry_bits.h"
 #include "shared/source/command_stream/command_stream_receiver.h"
+#include "shared/source/command_stream/command_stream_receiver_simulated_common_hw.h"
 #include "shared/source/gen12lp/helpers_gen12lp.h"

-#include "opencl/source/command_stream/command_stream_receiver_simulated_common_hw.h"
-
 namespace NEO {
 namespace Gen12LPHelpers {

--- a/shared/source/helpers/CMakeLists.txt
+++ b/shared/source/helpers/CMakeLists.txt
@@ -50,6 +50,8 @@ set(NEO_CORE_HELPERS
    ${CMAKE_CURRENT_SOURCE_DIR}/flush_stamp.cpp
    ${CMAKE_CURRENT_SOURCE_DIR}/flush_stamp.h
    ${CMAKE_CURRENT_SOURCE_DIR}/get_info.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/hardware_context_controller.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/hardware_context_controller.h
    ${CMAKE_CURRENT_SOURCE_DIR}/hash.h
    ${CMAKE_CURRENT_SOURCE_DIR}/heap_assigner.h
    ${CMAKE_CURRENT_SOURCE_DIR}/heap_assigner.cpp
@@ -75,6 +77,7 @@ set(NEO_CORE_HELPERS
    ${CMAKE_CURRENT_SOURCE_DIR}/local_id_gen.inl
    ${CMAKE_CURRENT_SOURCE_DIR}/local_id_gen_special.inl
    ${CMAKE_CURRENT_SOURCE_DIR}/local_id_gen_sse4.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/neo_driver_version.h
    ${CMAKE_CURRENT_SOURCE_DIR}/non_copyable_or_moveable.h
    ${CMAKE_CURRENT_SOURCE_DIR}/options.h
    ${CMAKE_CURRENT_SOURCE_DIR}/pause_on_gpu_properties.h
--- a/shared/source/helpers/api_specific_config.h
+++ b/shared/source/helpers/api_specific_config.h
@@ -19,6 +19,7 @@ struct ApiSpecificConfig {
    static ApiType getApiType();
    static std::string getName();
    static uint64_t getReducedMaxAllocSize(uint64_t maxAllocSize);
+    static const char *getRegistryPath();

    static std::string getAubPrefixForSpecificApi() {
        return (getName() + "_");
--- a/shared/source/helpers/hardware_context_controller.cpp
+++ b/shared/source/helpers/hardware_context_controller.cpp
@@ -0,0 +1,84 @@
+/*
+ * Copyright (C) 2019-2021 Intel Corporation
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ */
+
+#include "shared/source/helpers/hardware_context_controller.h"
+
+#include "shared/source/helpers/constants.h"
+#include "shared/source/os_interface/os_context.h"
+
+#include "aub_mem_dump.h"
+#include "third_party/aub_stream/headers/allocation_params.h"
+using namespace NEO;
+
+HardwareContextController::HardwareContextController(aub_stream::AubManager &aubManager, OsContext &osContext, uint32_t flags) {
+    auto deviceBitfield = osContext.getDeviceBitfield();
+    for (uint32_t deviceIndex = 0; deviceIndex < deviceBitfield.size(); deviceIndex++) {
+        if (deviceBitfield.test(deviceIndex)) {
+            hardwareContexts.emplace_back(aubManager.createHardwareContext(deviceIndex, osContext.getEngineType(), flags));
+        }
+    }
+}
+
+void HardwareContextController::initialize() {
+    for (auto &hardwareContext : hardwareContexts) {
+        hardwareContext->initialize();
+    }
+}
+
+void HardwareContextController::pollForCompletion() {
+    for (auto &hardwareContext : hardwareContexts) {
+        hardwareContext->pollForCompletion();
+    }
+}
+
+void HardwareContextController::expectMemory(uint64_t gfxAddress, const void *srcAddress, size_t length, uint32_t compareOperation) {
+    for (auto &hardwareContext : hardwareContexts) {
+        hardwareContext->expectMemory(gfxAddress, srcAddress, length, compareOperation);
+    }
+}
+
+void HardwareContextController::submit(uint64_t batchBufferGpuAddress, const void *batchBuffer, size_t batchBufferSize,
+                                       uint32_t memoryBank, uint64_t entryBits, bool overrideRingHead) {
+    for (auto &hardwareContext : hardwareContexts) {
+        hardwareContext->submitBatchBuffer(batchBufferGpuAddress, overrideRingHead);
+    }
+}
+
+void HardwareContextController::writeMemory(aub_stream::AllocationParams &allocationParams) {
+    if (hardwareContexts.size() == 1u) {
+        hardwareContexts.at(0)->writeMemory2(allocationParams);
+        return;
+    }
+
+    aub_stream::AllocationParams clonedParams = allocationParams;
+
+    for (auto bankId = 0u; bankId < hardwareContexts.size(); bankId++) {
+        auto &hardwareContext = hardwareContexts.at(bankId);
+
+        auto selectedBank = allocationParams.memoryBanks & (1 << bankId);
+        UNRECOVERABLE_IF(selectedBank == 0);
+
+        clonedParams.memoryBanks = selectedBank;
+        hardwareContext->writeMemory2(clonedParams);
+    }
+}
+
+void HardwareContextController::writeMMIO(uint32_t offset, uint32_t value) {
+    hardwareContexts[0]->writeMMIO(offset, value);
+}
+
+void HardwareContextController::dumpBufferBIN(uint64_t gfxAddress, size_t size) {
+    hardwareContexts[0]->dumpBufferBIN(gfxAddress, size);
+}
+
+void HardwareContextController::dumpSurface(const aub_stream::SurfaceInfo &surfaceInfo) {
+    hardwareContexts[0]->dumpSurface(surfaceInfo);
+}
+
+void HardwareContextController::readMemory(uint64_t gfxAddress, void *memory, size_t size, uint32_t memoryBanks, size_t pageSize) {
+    hardwareContexts[0]->readMemory(gfxAddress, memory, size, memoryBanks, pageSize);
+}
--- a/shared/source/helpers/hardware_context_controller.h
+++ b/shared/source/helpers/hardware_context_controller.h
--- a/shared/source/helpers/neo_driver_version.h
+++ b/shared/source/helpers/neo_driver_version.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2019-2020 Intel Corporation
+ * Copyright (C) 2019-2021 Intel Corporation
 *
 * SPDX-License-Identifier: MIT
 *
--- a/shared/test/common/helpers/api_specific_config_shared_tests.cpp
+++ b/shared/test/common/helpers/api_specific_config_shared_tests.cpp
@@ -33,5 +33,7 @@ uint64_t ApiSpecificConfig::getReducedMaxAllocSize(uint64_t maxAllocSize) {
 std::string ApiSpecificConfig::getName() {
    return "shared";
 }
-
+const char *ApiSpecificConfig::getRegistryPath() {
+    return "";
+}
 } // namespace NEO
--- a/shared/test/common/libult/create_command_stream.cpp
+++ b/shared/test/common/libult/create_command_stream.cpp
@@ -7,13 +7,12 @@

 #include "shared/test/common/libult/create_command_stream.h"

+#include "shared/source/command_stream/create_command_stream_impl.h"
 #include "shared/source/execution_environment/root_device_environment.h"
 #include "shared/test/common/helpers/default_hw_info.h"
 #include "shared/test/common/helpers/ult_hw_config.h"
 #include "shared/test/common/mocks/ult_device_factory.h"

-#include "opencl/source/command_stream/create_command_stream_impl.h"
-
 #include <cassert>

 namespace NEO {