Add condition to ensure value type is trivially copyable to call redu…

…ce-then-scan

Signed-off-by: Matthew Michel <matthew.michel@intel.com>

include/oneapi/dpl/pstl/hetero/algorithm_ranges_impl_hetero.h

@@ -944,26 +944,28 @@ __pattern_reduce_by_segment(__hetero_tag<_BackendTag> __tag, _ExecutionPolicy&&
         return 1;
     }
 
+    using __diff_type = oneapi::dpl::__internal::__difference_t<_Range1>;
+    using __key_type = oneapi::dpl::__internal::__value_t<_Range1>;
+    using __val_type = oneapi::dpl::__internal::__value_t<_Range2>;
 #if _ONEDPL_BACKEND_SYCL
     // We would normally dispatch to the parallel implementation which would make the decision to invoke
     // reduce-then-scan. However, since the fallback is implemented at the ranges level we must choose
     // whether or not to use reduce-then-scan here.
-    if (oneapi::dpl::__par_backend_hetero::__is_gpu_with_sg_32(__exec))
+    if constexpr (std::is_trivially_copyable_v<__val_type>)
     {
-        auto __res = oneapi::dpl::__par_backend_hetero::__parallel_reduce_by_segment_reduce_then_scan(
-            _BackendTag{}, std::forward<_ExecutionPolicy>(__exec), std::forward<_Range1>(__keys),
-            std::forward<_Range2>(__values), std::forward<_Range3>(__out_keys), std::forward<_Range4>(__out_values),
-            __binary_pred, __binary_op);
-        __res.wait();
-        // Because our init type ends up being tuple<std::size_t, ValType>, return the first component which is the write index. Add 1 to return the
-        // past-the-end iterator pair of segmented reduction.
-        return std::get<0>(__res.get()) + 1;
+        if (oneapi::dpl::__par_backend_hetero::__is_gpu_with_sg_32(__exec))
+        {
+            auto __res = oneapi::dpl::__par_backend_hetero::__parallel_reduce_by_segment_reduce_then_scan(
+                _BackendTag{}, std::forward<_ExecutionPolicy>(__exec), std::forward<_Range1>(__keys),
+                std::forward<_Range2>(__values), std::forward<_Range3>(__out_keys), std::forward<_Range4>(__out_values),
+                __binary_pred, __binary_op);
+            __res.wait();
+            // Because our init type ends up being tuple<std::size_t, ValType>, return the first component which is the write index. Add 1 to return the
+            // past-the-end iterator pair of segmented reduction.
+            return std::get<0>(__res.get()) + 1;
+        }
     }
 #endif
-    using __diff_type = oneapi::dpl::__internal::__difference_t<_Range1>;
-    using __key_type = oneapi::dpl::__internal::__value_t<_Range1>;
-    using __val_type = oneapi::dpl::__internal::__value_t<_Range2>;
-
     // Round 1: reduce with extra indices added to avoid long segments
     // TODO: At threshold points check if the key is equal to the key at the previous threshold point, indicating a long sequence.
     // Skip a round of copy_if and reduces if there are none.