Use urEnqueueNativeCommandExp if supported

Don't branch on backend, use UR entry point if it is supported. Also add
equivalent enum to PI.

Author: Hugh Delaney

sycl/include/sycl/detail/pi.h

@@ -513,6 +513,8 @@ typedef enum {
 
   // Virtual memory support
   PI_EXT_ONEAPI_DEVICE_INFO_SUPPORTS_VIRTUAL_MEM = 0x2011E,
+  // Native enqueue
+  PI_EXT_ONEAPI_DEVICE_INFO_ENQUEUE_NATIVE_COMMAND_SUPPORT = 0x2011F,
 } _pi_device_info;
 
 typedef enum {

sycl/plugins/unified_runtime/pi2ur.hpp

@@ -1325,6 +1325,9 @@ inline pi_result piDeviceGetInfo(pi_device Device, pi_device_info ParamName,
     PI_TO_UR_MAP_DEVICE_INFO(
         PI_EXT_ONEAPI_DEVICE_INFO_TIMESTAMP_RECORDING_SUPPORT,
         UR_DEVICE_INFO_TIMESTAMP_RECORDING_SUPPORT_EXP)
+    PI_TO_UR_MAP_DEVICE_INFO(
+        PI_EXT_ONEAPI_DEVICE_INFO_ENQUEUE_NATIVE_COMMAND_SUPPORT,
+        UR_DEVICE_INFO_ENQUEUE_NATIVE_COMMAND_SUPPORT_EXP)
     PI_TO_UR_MAP_DEVICE_INFO(PI_EXT_INTEL_DEVICE_INFO_ESIMD_SUPPORT,
                              UR_DEVICE_INFO_ESIMD_SUPPORT)
     PI_TO_UR_MAP_DEVICE_INFO(PI_EXT_ONEAPI_DEVICE_INFO_COMPONENT_DEVICES,

sycl/source/detail/scheduler/commands.cpp

@@ -420,8 +420,13 @@ class DispatchHostTask {
                           HostTask.MQueue->getContextImplPtr()};
         // TODO: should all the backends that support this entry point use this
         // for host task?
-        if (IH.get_backend() == backend::ext_oneapi_cuda ||
-            IH.get_backend() == backend::ext_oneapi_hip) {
+        auto &Queue = HostTask.MQueue;
+        bool NativeCommandSupport = false;
+        Queue->getPlugin()->call<PiApiKind::piDeviceGetInfo>(
+            detail::getSyclObjImpl(Queue->get_device())->getHandleRef(),
+            PI_EXT_ONEAPI_DEVICE_INFO_ENQUEUE_NATIVE_COMMAND_SUPPORT,
+            sizeof(NativeCommandSupport), &NativeCommandSupport, nullptr);
+        if (NativeCommandSupport) {
           EnqueueNativeCommandData CustomOpData{
               IH, HostTask.MHostTask->MInteropTask};
 
@@ -432,11 +437,9 @@ class DispatchHostTask {
           //
           // This entry point is needed in order to migrate memory across
           // devices in the same context for CUDA and HIP backends
-          HostTask.MQueue->getPlugin()
-              ->call<PiApiKind::piextEnqueueNativeCommand>(
-                  HostTask.MQueue->getHandleRef(), InteropFreeFunc,
-                  &CustomOpData, MReqPiMem.size(), MReqPiMem.data(), 0, nullptr,
-                  nullptr);
+          Queue->getPlugin()->call<PiApiKind::piextEnqueueNativeCommand>(
+              HostTask.MQueue->getHandleRef(), InteropFreeFunc, &CustomOpData,
+              MReqPiMem.size(), MReqPiMem.data(), 0, nullptr, nullptr);
         } else {
           HostTask.MHostTask->call(MThisCmd->MEvent->getHostProfilingInfo(),
                                    IH);