Fix UR_DEVICE_INFO_SUB_GROUP_SIZES_INTEL size confusion

The spec says it's an array of uint32_t, but CUDA, HIP, OpenCL, and
LevelZero adapters were returning arrays of size_t. As a result, we had
extra dummy zeroes in the output. Native CPU adapter was correct.

Author: al42and

source/adapters/cuda/device.cpp

@@ -270,7 +270,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
     int WarpSize = 0;
     UR_CHECK_ERROR(cuDeviceGetAttribute(
         &WarpSize, CU_DEVICE_ATTRIBUTE_WARP_SIZE, hDevice->get()));
-    size_t Sizes[1] = {static_cast<size_t>(WarpSize)};
+    uint32_t Sizes[1] = {static_cast<uint32_t>(WarpSize)};
     return ReturnValue(Sizes, 1);
   }
   case UR_DEVICE_INFO_MAX_CLOCK_FREQUENCY: {

source/adapters/hip/device.cpp

@@ -177,7 +177,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
     int WarpSize = 0;
     UR_CHECK_ERROR(hipDeviceGetAttribute(&WarpSize, hipDeviceAttributeWarpSize,
                                          hDevice->get()));
-    size_t Sizes[1] = {static_cast<size_t>(WarpSize)};
+    uint32_t Sizes[1] = {static_cast<uint32_t>(WarpSize)};
     return ReturnValue(Sizes, 1);
   }
   case UR_DEVICE_INFO_MAX_CLOCK_FREQUENCY: {

source/adapters/level_zero/device.cpp

@@ -626,11 +626,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(
     return ReturnValue(static_cast<ur_bool_t>(false));
   }
   case UR_DEVICE_INFO_SUB_GROUP_SIZES_INTEL: {
-    // ze_device_compute_properties.subGroupSizes is in uint32_t whereas the
-    // expected return is size_t datatype. size_t can be 8 bytes of data.
-    return ReturnValue.template operator()<size_t>(
-        Device->ZeDeviceComputeProperties->subGroupSizes,
-        Device->ZeDeviceComputeProperties->numSubGroupSizes);
+    return ReturnValue(Device->ZeDeviceComputeProperties->subGroupSizes,
+                       Device->ZeDeviceComputeProperties->numSubGroupSizes);
   }
   case UR_DEVICE_INFO_IL_VERSION: {
     // Set to a space separated list of IL version strings of the form

source/adapters/opencl/device.cpp

@@ -910,7 +910,6 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
   case UR_EXT_DEVICE_INFO_OPENCL_C_VERSION:
   case UR_DEVICE_INFO_BUILT_IN_KERNELS:
   case UR_DEVICE_INFO_MAX_WORK_ITEM_SIZES:
-  case UR_DEVICE_INFO_SUB_GROUP_SIZES_INTEL:
   case UR_DEVICE_INFO_IP_VERSION: {
     /* We can just use the OpenCL outputs because the sizes of OpenCL types
      * are the same as UR.
@@ -929,6 +928,19 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
 
     return UR_RESULT_SUCCESS;
   }
+  case UR_DEVICE_INFO_SUB_GROUP_SIZES_INTEL: {
+    // Have to convert size_t to uint32_t
+    size_t SubGroupSizesSize = 0;
+    CL_RETURN_ON_FAILURE(
+        clGetDeviceInfo(cl_adapter::cast<cl_device_id>(hDevice), CLPropName, 0,
+                        nullptr, &SubGroupSizesSize));
+    std::vector<size_t> SubGroupSizes(SubGroupSizesSize / sizeof(size_t));
+    CL_RETURN_ON_FAILURE(
+        clGetDeviceInfo(cl_adapter::cast<cl_device_id>(hDevice), CLPropName,
+                        SubGroupSizesSize, SubGroupSizes.data(), nullptr));
+    return ReturnValue.template operator()<uint32_t>(SubGroupSizes.data(),
+                                                     SubGroupSizes.size());
+  }
   case UR_DEVICE_INFO_EXTENSIONS: {
     cl_device_id Dev = cl_adapter::cast<cl_device_id>(hDevice);
     size_t ExtSize = 0;