[oneDPL] Fix performance issue in __serial_merge (#2022)

Author: SergeyKopienko

include/oneapi/dpl/pstl/hetero/dpcpp/parallel_backend_sycl_merge.h

@@ -19,8 +19,9 @@
 #include <limits>    // std::numeric_limits
 #include <cassert>   // assert
 #include <cstdint>   // std::uint8_t, ...
-#include <utility>   // std::make_pair, std::forward
+#include <utility>   // std::make_pair, std::forward, std::declval
 #include <algorithm> // std::min, std::lower_bound
+#include <type_traits> // std::void_t, std::true_type, std::false_type
 
 #include "sycl_defs.h"
 #include "parallel_backend_sycl_utils.h"
@@ -130,6 +131,21 @@ __find_start_point(const _Rng1& __rng1, const _Index __rng1_from, _Index __rng1_
     return _split_point_t<_Index>{*__res, __index_sum - *__res + 1};
 }
 
+template <typename _Rng1DataType, typename _Rng2DataType, typename = void>
+struct __can_use_ternary_op : std::false_type
+{
+};
+
+template <typename _Rng1DataType, typename _Rng2DataType>
+struct __can_use_ternary_op<_Rng1DataType, _Rng2DataType,
+                            std::void_t<decltype(true ? std::declval<_Rng1DataType>() : std::declval<_Rng2DataType>())>>
+    : std::true_type
+{
+};
+
+template <typename _Rng1DataType, typename _Rng2DataType>
+constexpr static bool __can_use_ternary_op_v = __can_use_ternary_op<_Rng1DataType, _Rng2DataType>::value;
+
 // Do serial merge of the data from rng1 (starting from start1) and rng2 (starting from start2) and writing
 // to rng3 (starting from start3) in 'chunk' steps, but do not exceed the total size of the sequences (n1 and n2)
 template <typename _Rng1, typename _Rng2, typename _Rng3, typename _Index, typename _Compare>
@@ -156,11 +172,23 @@ __serial_merge(const _Rng1& __rng1, const _Rng2& __rng2, _Rng3& __rng3, const _I
         // One of __rng1_idx_less_n1 and __rng2_idx_less_n2 should be true here
         // because 1) we should fill output data with elements from one of the input ranges
         // 2) we calculate __rng3_idx_end as std::min<_Index>(__rng1_size + __rng2_size, __chunk).
-        if (__rng1_idx_less_n1 && __rng2_idx_less_n2 && __comp(__rng2[__rng2_idx], __rng1[__rng1_idx]) ||
-            !__rng1_idx_less_n1)
-            __rng3[__rng3_idx] = __rng2[__rng2_idx++];
+        if constexpr (__can_use_ternary_op_v<decltype(__rng1[__rng1_idx]), decltype(__rng2[__rng2_idx])>)
+        {
+            // This implementation is required for performance optimization
+            __rng3[__rng3_idx] = (!__rng1_idx_less_n1 || __rng1_idx_less_n1 && __rng2_idx_less_n2 &&
+                                                             __comp(__rng2[__rng2_idx], __rng1[__rng1_idx]))
+                                     ? __rng2[__rng2_idx++]
+                                     : __rng1[__rng1_idx++];
+        }
         else
-            __rng3[__rng3_idx] = __rng1[__rng1_idx++];
+        {
+            // TODO required to understand why the usual if-else is slower then ternary operator
+            if (!__rng1_idx_less_n1 ||
+                __rng1_idx_less_n1 && __rng2_idx_less_n2 && __comp(__rng2[__rng2_idx], __rng1[__rng1_idx]))
+                __rng3[__rng3_idx] = __rng2[__rng2_idx++];
+            else
+                __rng3[__rng3_idx] = __rng1[__rng1_idx++];
+        }
     }
 }