Add implementation

Author: againull

cmake/FetchLevelZero.cmake

@@ -7,6 +7,8 @@ set(UR_LEVEL_ZERO_LOADER_LIBRARY "" CACHE FILEPATH "Path of the Level Zero Loade
 set(UR_LEVEL_ZERO_INCLUDE_DIR "" CACHE FILEPATH "Directory containing the Level Zero Headers")
 set(UR_LEVEL_ZERO_LOADER_REPO "" CACHE STRING "Github repo to get the Level Zero loader sources from")
 set(UR_LEVEL_ZERO_LOADER_TAG "" CACHE STRING " GIT tag of the Level Loader taken from github repo")
+set(UR_COMPUTE_RUNTIME_REPO "" CACHE STRING "Github repo to get the compute runtime sources from")
+set(UR_COMPUTE_RUNTIME_TAG "" CACHE STRING " GIT tag of the compute runtime taken from github repo")
 
 # Copy Level Zero loader/headers locally to the build to avoid leaking their path.
 set(LEVEL_ZERO_COPY_DIR ${CMAKE_CURRENT_BINARY_DIR}/level_zero_loader)
@@ -87,8 +89,31 @@ target_link_libraries(LevelZeroLoader
     INTERFACE "${LEVEL_ZERO_LIB_NAME}"
 )
 
+file(GLOB LEVEL_ZERO_LOADER_API_HEADERS "${LEVEL_ZERO_INCLUDE_DIR}/*.h")
+file(COPY ${LEVEL_ZERO_LOADER_API_HEADERS} DESTINATION ${LEVEL_ZERO_INCLUDE_DIR}/level_zero)
 add_library(LevelZeroLoader-Headers INTERFACE)
 target_include_directories(LevelZeroLoader-Headers
-    INTERFACE "$<BUILD_INTERFACE:${LEVEL_ZERO_INCLUDE_DIR}>"
+    INTERFACE "$<BUILD_INTERFACE:${LEVEL_ZERO_INCLUDE_DIR};${LEVEL_ZERO_INCLUDE_DIR}/level_zero>"
+              "$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>"
+)
+
+if (UR_COMPUTE_RUNTIME_REPO STREQUAL "")
+set(UR_COMPUTE_RUNTIME_REPO "https://github.com/intel/compute-runtime.git")
+endif()
+if (UR_COMPUTE_RUNTIME_TAG STREQUAL "")
+set(UR_COMPUTE_RUNTIME_TAG 24.39.31294.12)
+endif()
+include(FetchContent)
+# Sparse fetch only the dir with level zero headers to avoid pulling in the entire compute-runtime.
+FetchContentSparse_Declare(compute-runtime-level-zero-headers ${UR_COMPUTE_RUNTIME_REPO} "${UR_COMPUTE_RUNTIME_TAG}" "level_zero/include")
+FetchContent_GetProperties(compute-runtime-level-zero-headers)
+if(NOT compute-runtime-level-zero-headers_POPULATED)
+  FetchContent_Populate(compute-runtime-level-zero-headers)
+endif()
+add_library(ComputeRuntimeLevelZero-Headers INTERFACE)
+set(COMPUTE_RUNTIME_LEVEL_ZERO_INCLUDE "${compute-runtime-level-zero-headers_SOURCE_DIR}/../..")
+message(STATUS "Level Zero Adapter: Using Level Zero headers from ${COMPUTE_RUNTIME_LEVEL_ZERO_INCLUDE}")
+target_include_directories(ComputeRuntimeLevelZero-Headers
+    INTERFACE "$<BUILD_INTERFACE:${COMPUTE_RUNTIME_LEVEL_ZERO_INCLUDE}>"
               "$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>"
 )

source/adapters/cuda/device.cpp

@@ -1085,6 +1085,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
   case UR_DEVICE_INFO_GPU_SUBSLICES_PER_SLICE:
   case UR_DEVICE_INFO_GPU_EU_COUNT_PER_SUBSLICE:
   case UR_DEVICE_INFO_GPU_HW_THREADS_PER_EU:
+  case UR_DEVICE_INFO_2D_BLOCK_ARRAY_CAPABILITIES_EXP:
     return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
 
   case UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP:

source/adapters/hip/device.cpp

@@ -902,6 +902,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
   case UR_DEVICE_INFO_BFLOAT16:
   case UR_DEVICE_INFO_IL_VERSION:
   case UR_DEVICE_INFO_ASYNC_BARRIER:
+  case UR_DEVICE_INFO_2D_BLOCK_ARRAY_CAPABILITIES_EXP:
     return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
   case UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP: {
     int DriverVersion = 0;

source/adapters/level_zero/CMakeLists.txt

@@ -109,6 +109,7 @@ if(UR_BUILD_ADAPTER_L0)
         ${PROJECT_NAME}::umf
         LevelZeroLoader
         LevelZeroLoader-Headers
+        ComputeRuntimeLevelZero-Headers
     )
 
     target_include_directories(ur_adapter_level_zero PRIVATE
@@ -202,6 +203,7 @@ if(UR_BUILD_ADAPTER_L0_V2)
         ${PROJECT_NAME}::umf
         LevelZeroLoader
         LevelZeroLoader-Headers
+        ComputeRuntimeLevelZero-Headers
     )
 
     target_include_directories(ur_adapter_level_zero_v2 PRIVATE

source/adapters/level_zero/common.cpp

@@ -11,6 +11,7 @@
 #include "common.hpp"
 #include "logger/ur_logger.hpp"
 #include "usm.hpp"
+#include <level_zero/include/ze_intel_gpu.h>
 
 ur_result_t ze2urResult(ze_result_t ZeResult) {
   if (ZeResult == ZE_RESULT_SUCCESS)
@@ -330,6 +331,14 @@ template <> zes_structure_type_t getZesStructureType<zes_mem_properties_t>() {
   return ZES_STRUCTURE_TYPE_MEM_PROPERTIES;
 }
 
+#ifdef ZE_INTEL_DEVICE_BLOCK_ARRAY_EXP_NAME
+template <>
+ze_structure_type_t
+getZeStructureType<ze_intel_device_block_array_exp_properties_t>() {
+  return ZE_INTEL_DEVICE_BLOCK_ARRAY_EXP_PROPERTIES;
+}
+#endif // ZE_INTEL_DEVICE_BLOCK_ARRAY_EXP_NAME
+
 // Global variables for ZER_EXT_RESULT_ADAPTER_SPECIFIC_ERROR
 thread_local ur_result_t ErrorMessageCode = UR_RESULT_SUCCESS;
 thread_local char ErrorMessage[MaxMessageSize];

source/adapters/level_zero/device.cpp

@@ -1155,6 +1155,30 @@ ur_result_t urDeviceGetInfo(
     return ReturnValue(false);
   case UR_DEVICE_INFO_GLOBAL_VARIABLE_SUPPORT:
     return ReturnValue(true);
+  case UR_DEVICE_INFO_2D_BLOCK_ARRAY_CAPABILITIES_EXP: {
+#ifdef ZE_INTEL_DEVICE_BLOCK_ARRAY_EXP_NAME
+    const auto ZeDeviceBlockArrayFlags =
+        Device->ZeDeviceBlockArrayProperties->flags;
+
+    auto supportsFlags =
+        [&](ze_intel_device_block_array_exp_flags_t RequiredFlags) {
+          return (ZeDeviceBlockArrayFlags & RequiredFlags) == RequiredFlags;
+        };
+
+    ur_exp_device_2d_block_array_capability_flags_t BlockArrayCapabilities = 0;
+    if (supportsFlags(ZE_INTEL_DEVICE_EXP_FLAG_2D_BLOCK_LOAD)) {
+      BlockArrayCapabilities |=
+          UR_EXP_DEVICE_2D_BLOCK_ARRAY_CAPABILITY_FLAG_LOAD;
+    }
+    if (supportsFlags(ZE_INTEL_DEVICE_EXP_FLAG_2D_BLOCK_STORE)) {
+      BlockArrayCapabilities |=
+          UR_EXP_DEVICE_2D_BLOCK_ARRAY_CAPABILITY_FLAG_STORE;
+    }
+    return ReturnValue(BlockArrayCapabilities);
+#else
+    return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
+#endif
+  }
   default:
     logger::error("Unsupported ParamName in urGetDeviceInfo");
     logger::error("ParamNameParamName={}(0x{})", ParamName,
@@ -1563,6 +1587,17 @@ ur_result_t ur_device_handle_t_::initialize(int SubSubDeviceOrdinal,
         ZE_CALL_NOCHECK(zeDeviceGetProperties, (ZeDevice, &P));
       };
 
+#ifdef ZE_INTEL_DEVICE_BLOCK_ARRAY_EXP_NAME
+  ZeDeviceBlockArrayProperties.Compute =
+      [ZeDevice](
+          ZeStruct<ze_intel_device_block_array_exp_properties_t> &Properties) {
+        ze_device_properties_t P;
+        P.stype = ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES;
+        P.pNext = &Properties;
+        ZE_CALL_NOCHECK(zeDeviceGetProperties, (ZeDevice, &P));
+      };
+#endif // ZE_INTEL_DEVICE_BLOCK_ARRAY_EXP_NAME
+
   ImmCommandListUsed = this->useImmediateCommandLists();
 
   uint32_t numQueueGroups = 0;

source/adapters/level_zero/device.hpp

@@ -18,14 +18,14 @@
 #include <unordered_map>
 #include <vector>
 
+#include "adapters/level_zero/platform.hpp"
+#include "common.hpp"
+#include <level_zero/include/ze_intel_gpu.h>
 #include <ur/ur.hpp>
 #include <ur_ddi.h>
 #include <ze_api.h>
 #include <zes_api.h>
 
-#include "adapters/level_zero/platform.hpp"
-#include "common.hpp"
-
 enum EventsScope {
   // All events are created host-visible.
   AllHostVisible,
@@ -223,6 +223,10 @@ struct ur_device_handle_t_ : _ur_object {
   ZeCache<struct ze_global_memsize> ZeGlobalMemSize;
   ZeCache<ZeStruct<ze_mutable_command_list_exp_properties_t>>
       ZeDeviceMutableCmdListsProperties;
+#ifdef ZE_INTEL_DEVICE_BLOCK_ARRAY_EXP_NAME
+  ZeCache<ZeStruct<ze_intel_device_block_array_exp_properties_t>>
+      ZeDeviceBlockArrayProperties;
+#endif // ZE_INTEL_DEVICE_BLOCK_ARRAY_EXP_NAME
 
   // Map device bindless image offset to corresponding host image handle.
   std::unordered_map<ur_exp_image_native_handle_t, ze_image_handle_t>

source/adapters/native_cpu/device.cpp

@@ -366,6 +366,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
   case UR_DEVICE_INFO_MEMORY_CLOCK_RATE:
   case UR_DEVICE_INFO_MEMORY_BUS_WIDTH:
     return UR_RESULT_ERROR_INVALID_VALUE;
+  case UR_DEVICE_INFO_2D_BLOCK_ARRAY_CAPABILITIES_EXP:
+    return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
   case UR_DEVICE_INFO_ATOMIC_FENCE_ORDER_CAPABILITIES: {
     // Currently for Native CPU fences are implemented using OCK
     // builtins, so we have different capabilities than atomic operations

source/adapters/opencl/device.cpp

@@ -1067,7 +1067,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
   case UR_DEVICE_INFO_GLOBAL_MEM_FREE:
   case UR_DEVICE_INFO_MEMORY_CLOCK_RATE:
   case UR_DEVICE_INFO_MEMORY_BUS_WIDTH:
-  case UR_DEVICE_INFO_ASYNC_BARRIER: {
+  case UR_DEVICE_INFO_ASYNC_BARRIER:
+  case UR_DEVICE_INFO_2D_BLOCK_ARRAY_CAPABILITIES_EXP: {
     return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
   }
 

test/adapters/level_zero/CMakeLists.txt

@@ -29,6 +29,7 @@ if(UR_BUILD_ADAPTER_L0)
         target_link_libraries(test-adapter-level_zero PRIVATE
             LevelZeroLoader
             LevelZeroLoader-Headers
+            ComputeRuntimeLevelZero-Headers
         )
 
         target_include_directories(test-adapter-level_zero PRIVATE