Fast path for experimental::row::equality

cpp/include/cudf/table/experimental/row_operators.cuh

@@ -245,6 +245,16 @@ using optional_dremel_view = thrust::optional<detail::dremel_device_view const>;
  * second letter in both words is the first non-equal letter, and `a < b`, thus
  * `aac < abb`.
  *
+ * @note: The operator overloads in sub-class `element_comparator` are templated via the
+ *        `type_dispatcher` to help select an overload instance for each column in a table.
+ *        So, `cudf::is_nested<Element>` will return `true` if the table has nested-type columns,
+ *        but it will be a runtime error if template parameter `has_nested_columns != true`.
+ *
+ * @tparam has_nested_columns compile-time optimization for primitive types.
+ *         This template parameter is to be used by the developer by querying
+ *         `cudf::detail::has_nested_columns(input)`. `true` compiles operator
+ *          overloads for nested types, while `false` only compiles operator
+ *          overloads for primitive types.
  * @tparam Nullate A cudf::nullate type describing whether to check for nulls.
  * @tparam PhysicalElementComparator A relational comparator functor that compares individual values
  * rather than logical elements, defaults to `NaN` aware relational comparator that evaluates `NaN`
@@ -857,6 +867,16 @@ class self_comparator {
    *
    * `F(i,j)` returns true if and only if row `i` compares lexicographically less than row `j`.
    *
+   * @note: The operator overloads in sub-class `element_comparator` are templated via the
+   *        `type_dispatcher` to help select an overload instance for each column in a table.
+   *        So, `cudf::is_nested<Element>` will return `true` if the table has nested-type columns,
+   *        but it will be a runtime error if template parameter `has_nested_columns != true`.
+   *
+   * @tparam has_nested_columns compile-time optimization for primitive types.
+   *         This template parameter is to be used by the developer by querying
+   *         `cudf::detail::has_nested_columns(input)`. `true` compiles operator
+   *         overloads for nested types, while `false` only compiles operator
+   *         overloads for primitive types.
    * @tparam Nullate A cudf::nullate type describing whether to check for nulls.
    * @tparam PhysicalElementComparator A relational comparator functor that compares individual
    * values rather than logical elements, defaults to `NaN` aware relational comparator that
@@ -1009,6 +1029,16 @@ class two_table_comparator {
    * only if row `i` of the right table compares lexicographically less than row
    * `j` of the left table.
    *
+   * @note: The operator overloads in sub-class `element_comparator` are templated via the
+   *        `type_dispatcher` to help select an overload instance for each column in a table.
+   *        So, `cudf::is_nested<Element>` will return `true` if the table has nested-type columns,
+   *        but it will be a runtime error if template parameter `has_nested_columns != true`.
+   *
+   * @tparam has_nested_columns compile-time optimization for primitive types.
+   *         This template parameter is to be used by the developer by querying
+   *         `cudf::detail::has_nested_columns(input)`. `true` compiles operator
+   *         overloads for nested types, while `false` only compiles operator
+   *         overloads for primitive types.
    * @tparam Nullate A cudf::nullate type describing whether to check for nulls.
    * @tparam PhysicalElementComparator A relational comparator functor that compares individual
    * values rather than logical elements, defaults to `NaN` aware relational comparator that
@@ -1131,11 +1161,22 @@ struct nan_equal_physical_equality_comparator {
  * returns false, representing unequal rows. If the rows are compared without mismatched elements,
  * the rows are equal.
  *
+ * @note: The operator overloads in sub-class `element_comparator` are templated via the
+ *        `type_dispatcher` to help select an overload instance for each column in a table.
+ *        So, `cudf::is_nested<Element>` will return `true` if the table has nested-type columns,
+ *        but it will be a runtime error if template parameter `has_nested_columns != true`.
+ *
+ * @tparam has_nested_columns compile-time optimization for primitive types.
+ *         This template parameter is to be used by the developer by querying
+ *         `cudf::detail::has_nested_columns(input)`. `true` compiles operator
+ *          overloads for nested types, while `false` only compiles operator
+ *          overloads for primitive types.
  * @tparam Nullate A cudf::nullate type describing whether to check for nulls.
  * @tparam PhysicalEqualityComparator A equality comparator functor that compares individual values
  * rather than logical elements, defaults to a comparator for which `NaN == NaN`.
  */
-template <typename Nullate,
+template <bool has_nested_columns,
+          typename Nullate,
           typename PhysicalEqualityComparator = nan_equal_physical_equality_comparator>
 class device_row_comparator {
   friend class self_comparator;       ///< Allow self_comparator to access private members
@@ -1246,14 +1287,14 @@ class device_row_comparator {
 
     template <typename Element,
               CUDF_ENABLE_IF(not cudf::is_equality_comparable<Element, Element>() and
-                             not cudf::is_nested<Element>()),
+                             (not has_nested_columns or not cudf::is_nested<Element>())),
               typename... Args>
     __device__ bool operator()(Args...)
     {
       CUDF_UNREACHABLE("Attempted to compare elements of uncomparable types.");
     }
 
-    template <typename Element, CUDF_ENABLE_IF(cudf::is_nested<Element>())>
+    template <typename Element, CUDF_ENABLE_IF(has_nested_columns and cudf::is_nested<Element>())>
     __device__ bool operator()(size_type const lhs_element_index,
                                size_type const rhs_element_index) const noexcept
     {
@@ -1437,6 +1478,16 @@ class self_comparator {
    *
    * `F(i,j)` returns true if and only if row `i` compares equal to row `j`.
    *
+   * @note: The operator overloads in sub-class `element_comparator` are templated via the
+   *        `type_dispatcher` to help select an overload instance for each column in a table.
+   *        So, `cudf::is_nested<Element>` will return `true` if the table has nested-type columns,
+   *        but it will be a runtime error if template parameter `has_nested_columns != true`.
+   *
+   * @tparam has_nested_columns compile-time optimization for primitive types.
+   *         This template parameter is to be used by the developer by querying
+   *         `cudf::detail::has_nested_columns(input)`. `true` compiles operator
+   *          overloads for nested types, while `false` only compiles operator
+   *          overloads for primitive types.
    * @tparam Nullate A cudf::nullate type describing whether to check for nulls.
    * @tparam PhysicalEqualityComparator A equality comparator functor that compares individual
    * values rather than logical elements, defaults to a comparator for which `NaN == NaN`.
@@ -1445,13 +1496,15 @@ class self_comparator {
    * @param comparator Physical element equality comparison functor.
    * @return A binary callable object
    */
-  template <typename Nullate,
+  template <bool has_nested_columns,
+            typename Nullate,
             typename PhysicalEqualityComparator = nan_equal_physical_equality_comparator>
   auto equal_to(Nullate nullate                       = {},
                 null_equality nulls_are_equal         = null_equality::EQUAL,
                 PhysicalEqualityComparator comparator = {}) const noexcept
   {
-    return device_row_comparator{nullate, *d_t, *d_t, nulls_are_equal, comparator};
+    return device_row_comparator<has_nested_columns, Nullate, PhysicalEqualityComparator>{
+      nullate, *d_t, *d_t, nulls_are_equal, comparator};
   }
 
  private:
@@ -1539,6 +1592,16 @@ class two_table_comparator {
    * Similarly, `F(rhs_index_type i, lhs_index_type j)` returns true if and only if row `i` of the
    * right table compares equal to row `j` of the left table.
    *
+   * @note: The operator overloads in sub-class `element_comparator` are templated via the
+   *        `type_dispatcher` to help select an overload instance for each column in a table.
+   *        So, `cudf::is_nested<Element>` will return `true` if the table has nested-type columns,
+   *        but it will be a runtime error if template parameter `has_nested_columns != true`.
+   *
+   * @tparam has_nested_columns compile-time optimization for primitive types.
+   *         This template parameter is to be used by the developer by querying
+   *         `cudf::detail::has_nested_columns(input)`. `true` compiles operator
+   *          overloads for nested types, while `false` only compiles operator
+   *          overloads for primitive types.
    * @tparam Nullate A cudf::nullate type describing whether to check for nulls.
    * @tparam PhysicalEqualityComparator A equality comparator functor that compares individual
    * values rather than logical elements, defaults to a `NaN == NaN` equality comparator.
@@ -1547,14 +1610,16 @@ class two_table_comparator {
    * @param comparator Physical element equality comparison functor.
    * @return A binary callable object
    */
-  template <typename Nullate,
+  template <bool has_nested_columns,
+            typename Nullate,
             typename PhysicalEqualityComparator = nan_equal_physical_equality_comparator>
   auto equal_to(Nullate nullate                       = {},
                 null_equality nulls_are_equal         = null_equality::EQUAL,
                 PhysicalEqualityComparator comparator = {}) const noexcept
   {
     return strong_index_comparator_adapter{
-      device_row_comparator(nullate, *d_left_table, *d_right_table, nulls_are_equal, comparator)};
+      device_row_comparator<has_nested_columns, Nullate, PhysicalEqualityComparator>(
+        nullate, *d_left_table, *d_right_table, nulls_are_equal, comparator)};
   }
 
  private:

cpp/src/binaryop/compiled/struct_binary_ops.cuh

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022, NVIDIA CORPORATION.
+ * Copyright (c) 2022-2023, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -106,6 +106,35 @@ void apply_struct_binary_op(mutable_column_view& out,
   }
 }
 
+template <typename OptionalIteratorType, typename DeviceComparatorType>
+struct struct_equality_functor {
+  struct_equality_functor(OptionalIteratorType optional_iter_,
+                          bool is_lhs_scalar_,
+                          bool is_rhs_scalar_,
+                          bool preserve_output_,
+                          DeviceComparatorType device_comparator_)
+    : optional_iter(optional_iter_),
+      is_lhs_scalar(is_lhs_scalar_),
+      is_rhs_scalar(is_rhs_scalar_),
+      preserve_output(preserve_output_),
+      device_comparator(device_comparator_)
+  {
+  }
+
+  auto __device__ operator()(size_type i)
+  {
+    auto lhs = cudf::experimental::row::lhs_index_type{is_lhs_scalar ? 0 : i};
+    auto rhs = cudf::experimental::row::rhs_index_type{is_rhs_scalar ? 0 : i};
+    return optional_iter[i].has_value() and (device_comparator(lhs, rhs) == preserve_output);
+  }
+
+  OptionalIteratorType optional_iter;
+  bool is_lhs_scalar;
+  bool is_rhs_scalar;
+  bool preserve_output;
+  DeviceComparatorType device_comparator;
+};
+
 template <typename PhysicalEqualityComparator =
             cudf::experimental::row::equality::physical_equality_comparator>
 void apply_struct_equality_op(mutable_column_view& out,
@@ -125,26 +154,37 @@ void apply_struct_equality_op(mutable_column_view& out,
   auto trhs = table_view{{rhs}};
   auto table_comparator =
     cudf::experimental::row::equality::two_table_comparator{tlhs, trhs, stream};
-  auto device_comparator =
-    table_comparator.equal_to(nullate::DYNAMIC{has_nested_nulls(tlhs) || has_nested_nulls(trhs)},
-                              null_equality::EQUAL,
-                              comparator);
 
   auto outd = column_device_view::create(out, stream);
   auto optional_iter =
     cudf::detail::make_optional_iterator<bool>(*outd, nullate::DYNAMIC{out.has_nulls()});
-  thrust::tabulate(rmm::exec_policy(stream),
-                   out.begin<bool>(),
-                   out.end<bool>(),
-                   [optional_iter,
-                    is_lhs_scalar,
-                    is_rhs_scalar,
-                    preserve_output = (op != binary_operator::NOT_EQUAL),
-                    device_comparator] __device__(size_type i) {
-                     auto lhs = cudf::experimental::row::lhs_index_type{is_lhs_scalar ? 0 : i};
-                     auto rhs = cudf::experimental::row::rhs_index_type{is_rhs_scalar ? 0 : i};
-                     return optional_iter[i].has_value() and
-                            (device_comparator(lhs, rhs) == preserve_output);
-                   });
+
+  auto const comparator_helper = [&](auto const device_comparator) {
+    thrust::tabulate(rmm::exec_policy(stream),
+                     out.begin<bool>(),
+                     out.end<bool>(),
+                     struct_equality_functor<decltype(optional_iter), decltype(device_comparator)>(
+                       optional_iter,
+                       is_lhs_scalar,
+                       is_rhs_scalar,
+                       op != binary_operator::NOT_EQUAL,
+                       device_comparator));
+  };
+
+  if (cudf::detail::has_nested_columns(tlhs) or cudf::detail::has_nested_columns(trhs)) {
+    auto device_comparator = table_comparator.equal_to<true>(
+      nullate::DYNAMIC{has_nested_nulls(tlhs) || has_nested_nulls(trhs)},
+      null_equality::EQUAL,
+      comparator);
+
+    comparator_helper(device_comparator);
+  } else {
+    auto device_comparator = table_comparator.equal_to<false>(
+      nullate::DYNAMIC{has_nested_nulls(tlhs) || has_nested_nulls(trhs)},
+      null_equality::EQUAL,
+      comparator);
+
+    comparator_helper(device_comparator);
+  }
 }
 }  // namespace cudf::binops::compiled::detail