[UR][Cuda] Implement intel device info extensions for HW_THREADS_PER_EU and EU_SIMD_WIDTH on Cuda

Author: Georgi Mirazchiyski

source/adapters/cuda/device.cpp

@@ -1082,13 +1082,32 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
   case UR_DEVICE_INFO_COMPOSITE_DEVICE:
   case UR_DEVICE_INFO_MAX_READ_WRITE_IMAGE_ARGS:
   case UR_DEVICE_INFO_GPU_EU_COUNT:
-  case UR_DEVICE_INFO_GPU_EU_SIMD_WIDTH:
   case UR_DEVICE_INFO_GPU_EU_SLICES:
   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:
     return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
 
+  case UR_DEVICE_INFO_GPU_HW_THREADS_PER_EU: {
+    int MaxHwThreads{0}; // TODO: Cache this as a ur_device_handle_t_ member
+    UR_CHECK_ERROR(cuDeviceGetAttribute(
+        &MaxHwThreads, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR,
+        hDevice->get()));
+    detail::ur::assertion(MaxHwThreads >= 0);
+    return ReturnValue(static_cast<size_t>(MaxHwThreads));
+  }
+  case UR_DEVICE_INFO_GPU_EU_SIMD_WIDTH: {
+    int MaxHwThreads{0}; // TODO: Cache this as a ur_device_handle_t_ member
+    UR_CHECK_ERROR(cuDeviceGetAttribute(
+        &MaxHwThreads, CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR,
+        hDevice->get()));
+    detail::ur::assertion(MaxHwThreads >= 0);
+    int WarpSize{0}; // TODO: Cache this as a ur_device_handle_t_ member
+    UR_CHECK_ERROR(cuDeviceGetAttribute(
+        &WarpSize, CU_DEVICE_ATTRIBUTE_WARP_SIZE, hDevice->get()));
+    detail::ur::assertion(WarpSize >= 0);
+    return ReturnValue(static_cast<uint32_t>(MaxHwThreads) / WarpSize);
+  }
+
   case UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP:
   case UR_DEVICE_INFO_COMMAND_BUFFER_EVENT_SUPPORT_EXP:
     return ReturnValue(true);