[SYCL][CUDA] atomic_ref.fetch_add used for fp64 reduction if device.has(atomic64)

sycl/include/CL/sycl/aspects.hpp

@@ -39,7 +39,9 @@ enum class aspect {
   ext_intel_gpu_eu_count_per_subslice = 23,
   ext_intel_max_mem_bandwidth = 24,
   ext_intel_mem_channel = 25,
-  atomic64 = 26
+  usm_atomic_host_allocations = 26,
+  usm_atomic_shared_allocations = 27,
+  atomic64 = 28
 };
 
 } // namespace sycl

sycl/source/detail/device_info.hpp

@@ -234,24 +234,23 @@ template <> struct get_device_info<bool, info::device::queue_profiling> {
 };
 
 // Specialization for atomic64 that is necessary because
-// PI_DEVICE_INFO_ATOMIC_64 isn't implemented for backend other than cuda.
-// TODO the if-statement can be removed when the other backends support
-// PI_DEVICE_INFO_ATOMIC_64.
+// PI_DEVICE_INFO_ATOMIC_64 is currently only implemented for the cuda backend.
 template <> struct get_device_info<bool, info::device::atomic64> {
   static bool get(RT::PiDevice dev, const plugin &Plugin) {
 
     bool result = false;
 
-    platform plt =
-        get_device_info<platform, info::device::platform>::get(dev, Plugin);
-
-    if (plt.get_backend() == backend::cuda) {
-      Plugin.call<PiApiKind::piDeviceGetInfo>(
-          dev, pi::cast<RT::PiDeviceInfo>(info::device::atomic64),
-          sizeof(result), &result, nullptr);
+    RT::PiResult Err = Plugin.call_nocheck<PiApiKind::piDeviceGetInfo>(
+        dev, pi::cast<RT::PiDeviceInfo>(info::device::atomic64), sizeof(result),
+        &result, nullptr);
+    if (Err == PI_INVALID_VALUE) {
+      std::cout
+          << "The Plugin Interface has returned an error:\n  The value of "
+             "the atomic64 device aspect is unknown, Setting atomic64 "
+             "= false.\n\n";
     }
 
-    return (result);
+    return result;
   }
 };