OpenCL adapter implementation and update other adapters

Author: againull

source/adapters/cuda/device.cpp

@@ -1087,9 +1087,10 @@ 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_2D_BLOCK_ARRAY_CAPABILITIES_EXP:
+    return ReturnValue(
+        static_cast<ur_exp_device_2d_block_array_capability_flags_t>(0));
   case UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP:
   case UR_DEVICE_INFO_COMMAND_BUFFER_EVENT_SUPPORT_EXP:
     return ReturnValue(true);

source/adapters/hip/device.cpp

@@ -904,8 +904,10 @@ 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_2D_BLOCK_ARRAY_CAPABILITIES_EXP:
+    return ReturnValue(
+        static_cast<ur_exp_device_2d_block_array_capability_flags_t>(0));
   case UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP: {
     int DriverVersion = 0;
     UR_CHECK_ERROR(hipDriverGetVersion(&DriverVersion));

source/adapters/native_cpu/device.cpp

@@ -367,7 +367,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
   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;
+    return ReturnValue(
+        static_cast<ur_exp_device_2d_block_array_capability_flags_t>(0));
   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

@@ -1094,10 +1094,20 @@ 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:
   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_2D_BLOCK_ARRAY_CAPABILITIES_EXP: {
+    bool Is2DBlockIOSupported = false;
+    if (cl_adapter::checkDeviceExtensions(
+            cl_adapter::cast<cl_device_id>(hDevice),
+            {"cl_intel_subgroup_2d_block_io"},
+            Is2DBlockIOSupported) != UR_RESULT_SUCCESS ||
+        !Is2DBlockIOSupported) {
+      return ReturnValue(
+          static_cast<ur_exp_device_2d_block_array_capability_flags_t>(0));
+    }
+    return ReturnValue(UR_EXP_DEVICE_2D_BLOCK_ARRAY_CAPABILITY_FLAG_LOAD |
+                       UR_EXP_DEVICE_2D_BLOCK_ARRAY_CAPABILITY_FLAG_STORE);
   }
-
   case UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP: {
     cl_device_id Dev = cl_adapter::cast<cl_device_id>(hDevice);
     size_t ExtSize = 0;