Implement mixed equality/conditional semi/anti joins

cpp/CMakeLists.txt

@@ -322,11 +322,16 @@ add_library(
   src/jit/parser.cpp
   src/jit/type.cpp
   src/join/conditional_join.cu
-  src/join/mixed_join.cu
   src/join/cross_join.cu
   src/join/hash_join.cu
   src/join/join.cu
   src/join/join_utils.cu
+  src/join/mixed_join.cu
+  src/join/mixed_join_kernels.cu
+  src/join/mixed_join_kernels_semi.cu
+  src/join/mixed_join_semi.cu
+  src/join/mixed_join_size_kernels.cu
+  src/join/mixed_join_size_kernels_semi.cu
   src/join/semi_join.cu
   src/lists/contains.cu
   src/lists/combine/concatenate_list_elements.cu

cpp/include/cudf/join.hpp

@@ -1037,6 +1037,109 @@ mixed_full_join(
   std::optional<std::pair<std::size_t, device_span<size_type const>>> output_size_data = {},
   rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
 
+/**
+ * @brief Returns an index vector corresponding to all rows in the left tables
+ * where the columns of the equality table are equal and the predicate
+ * evaluates to true on the conditional tables.
+ *
+ * If the provided predicate returns NULL for a pair of rows (left, right), the
+ * left row is not included in the output. It is the user's responsiblity to
+ * choose a suitable compare_nulls value AND use appropriate null-safe
+ * operators in the expression.
+ *
+ * If the provided output size or per-row counts are incorrect, behavior is undefined.
+ *
+ * @code{.pseudo}
+ * left_equality: {{0, 1, 2}}
+ * right_equality: {{1, 2, 3}}
+ * left_conditional: {{4, 4, 4}}
+ * right_conditional: {{3, 4, 5}}
+ * Expression: Left.Column_0 > Right.Column_0
+ * Result: {1}
+ * @endcode
+ *
+ * @throw cudf::logic_error If the binary predicate outputs a non-boolean result.
+ * @throw cudf::logic_error If the number of rows in left_equality and left_conditional do not
+ * match.
+ * @throw cudf::logic_error If the number of rows in right_equality and right_conditional do not
+ * match.
+ *
+ * @param left_equality The left table used for the equality join.
+ * @param right_equality The right table used for the equality join.
+ * @param left_conditional The left table used for the conditional join.
+ * @param right_conditional The right table used for the conditional join.
+ * @param binary_predicate The condition on which to join.
+ * @param compare_nulls Whether or not null values join to each other or not.
+ * @param output_size_data An optional pair of values indicating the exact output size and the
+ * number of matches for each row in the larger of the two input tables, left or right (may be
+ * precomputed using the corresponding mixed_full_join_size API).
+ * @param mr Device memory resource used to allocate the returned table and columns' device memory
+ *
+ * @return A pair of vectors [`left_indices`, `right_indices`] that can be used to construct
+ * the result of performing a mixed full join between the four input tables.
+ */
+std::unique_ptr<rmm::device_uvector<size_type>> mixed_left_semi_join(
+  table_view const& left_equality,
+  table_view const& right_equality,
+  table_view const& left_conditional,
+  table_view const& right_conditional,
+  ast::expression const& binary_predicate,
+  null_equality compare_nulls = null_equality::EQUAL,
+  std::optional<std::pair<std::size_t, device_span<size_type const>>> output_size_data = {},
+  rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
+
+/**
+ * @brief Returns an index vector corresponding to all rows in the left tables
+ * for which there is no row in the right tables where the columns of the
+ * equality table are equal and the predicate evaluates to true on the
+ * conditional tables.
+ *
+ * If the provided predicate returns NULL for a pair of rows (left, right), the
+ * left row is not included in the output. It is the user's responsiblity to
+ * choose a suitable compare_nulls value AND use appropriate null-safe
+ * operators in the expression.
+ *
+ * If the provided output size or per-row counts are incorrect, behavior is undefined.
+ *
+ * @code{.pseudo}
+ * left_equality: {{0, 1, 2}}
+ * right_equality: {{1, 2, 3}}
+ * left_conditional: {{4, 4, 4}}
+ * right_conditional: {{3, 4, 5}}
+ * Expression: Left.Column_0 > Right.Column_0
+ * Result: {0, 2}
+ * @endcode
+ *
+ * @throw cudf::logic_error If the binary predicate outputs a non-boolean result.
+ * @throw cudf::logic_error If the number of rows in left_equality and left_conditional do not
+ * match.
+ * @throw cudf::logic_error If the number of rows in right_equality and right_conditional do not
+ * match.
+ *
+ * @param left_equality The left table used for the equality join.
+ * @param right_equality The right table used for the equality join.
+ * @param left_conditional The left table used for the conditional join.
+ * @param right_conditional The right table used for the conditional join.
+ * @param binary_predicate The condition on which to join.
+ * @param compare_nulls Whether or not null values join to each other or not.
+ * @param output_size_data An optional pair of values indicating the exact output size and the
+ * number of matches for each row in the larger of the two input tables, left or right (may be
+ * precomputed using the corresponding mixed_full_join_size API).
+ * @param mr Device memory resource used to allocate the returned table and columns' device memory
+ *
+ * @return A pair of vectors [`left_indices`, `right_indices`] that can be used to construct
+ * the result of performing a mixed full join between the four input tables.
+ */
+std::unique_ptr<rmm::device_uvector<size_type>> mixed_left_anti_join(
+  table_view const& left_equality,
+  table_view const& right_equality,
+  table_view const& left_conditional,
+  table_view const& right_conditional,
+  ast::expression const& binary_predicate,
+  null_equality compare_nulls = null_equality::EQUAL,
+  std::optional<std::pair<std::size_t, device_span<size_type const>>> output_size_data = {},
+  rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
+
 /**
  * @brief Returns the exact number of matches (rows) when performing a
  * mixed inner join between the specified tables where the columns of the
@@ -1123,6 +1226,90 @@ std::pair<std::size_t, std::unique_ptr<rmm::device_uvector<size_type>>> mixed_le
   null_equality compare_nulls         = null_equality::EQUAL,
   rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
 
+/**
+ * @brief Returns the exact number of matches (rows) when performing a mixed
+ * left semi join between the specified tables where the columns of the
+ * equality table are equal and the predicate evaluates to true on the
+ * conditional tables.
+ *
+ * If the provided predicate returns NULL for a pair of rows (left, right),
+ * that pair is not included in the output. It is the user's responsiblity to
+ * choose a suitable compare_nulls value AND use appropriate null-safe
+ * operators in the expression.
+ *
+ * @throw cudf::logic_error If the binary predicate outputs a non-boolean result.
+ * @throw cudf::logic_error If the number of rows in left_equality and left_conditional do not
+ * match.
+ * @throw cudf::logic_error If the number of rows in right_equality and right_conditional do not
+ * match.
+ *
+ * @param left_equality The left table used for the equality join.
+ * @param right_equality The right table used for the equality join.
+ * @param left_conditional The left table used for the conditional join.
+ * @param right_conditional The right table used for the conditional join.
+ * @param binary_predicate The condition on which to join.
+ * @param compare_nulls Whether or not null values join to each other or not.
+ * @param output_size An optional pair of values indicating the exact output size and the number of
+ * matches for each row in the larger of the two input tables, left or right (may be precomputed
+ * using the corresponding mixed_inner_join_size API).
+ * @param mr Device memory resource used to allocate the returned table and columns' device memory
+ *
+ * @return A pair containing the size that would result from performing the
+ * requested join and the number of matches for each row in one of the two
+ * tables. Which of the two tables is an implementation detail and should not
+ * be relied upon, simply passed to the corresponding `mixed_left_join` API as
+ * is.
+ */
+std::pair<std::size_t, std::unique_ptr<rmm::device_uvector<size_type>>> mixed_left_semi_join_size(
+  table_view const& left_equality,
+  table_view const& right_equality,
+  table_view const& left_conditional,
+  table_view const& right_conditional,
+  ast::expression const& binary_predicate,
+  null_equality compare_nulls         = null_equality::EQUAL,
+  rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
+
+/**
+ * @brief Returns the exact number of matches (rows) when performing a mixed
+ * left anti join between the specified tables.
+ *
+ * If the provided predicate returns NULL for a pair of rows (left, right),
+ * that pair is not included in the output. It is the user's responsiblity to
+ * choose a suitable compare_nulls value AND use appropriate null-safe
+ * operators in the expression.
+ *
+ * @throw cudf::logic_error If the binary predicate outputs a non-boolean result.
+ * @throw cudf::logic_error If the number of rows in left_equality and left_conditional do not
+ * match.
+ * @throw cudf::logic_error If the number of rows in right_equality and right_conditional do not
+ * match.
+ *
+ * @param left_equality The left table used for the equality join.
+ * @param right_equality The right table used for the equality join.
+ * @param left_conditional The left table used for the conditional join.
+ * @param right_conditional The right table used for the conditional join.
+ * @param binary_predicate The condition on which to join.
+ * @param compare_nulls Whether or not null values join to each other or not.
+ * @param output_size An optional pair of values indicating the exact output size and the number of
+ * matches for each row in the larger of the two input tables, left or right (may be precomputed
+ * using the corresponding mixed_inner_join_size API).
+ * @param mr Device memory resource used to allocate the returned table and columns' device memory
+ *
+ * @return A pair containing the size that would result from performing the
+ * requested join and the number of matches for each row in one of the two
+ * tables. Which of the two tables is an implementation detail and should not
+ * be relied upon, simply passed to the corresponding `mixed_left_join` API as
+ * is.
+ */
+std::pair<std::size_t, std::unique_ptr<rmm::device_uvector<size_type>>> mixed_left_anti_join_size(
+  table_view const& left_equality,
+  table_view const& right_equality,
+  table_view const& left_conditional,
+  table_view const& right_conditional,
+  ast::expression const& binary_predicate,
+  null_equality compare_nulls         = null_equality::EQUAL,
+  rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
+
 /**
  * @brief Returns the exact number of matches (rows) when performing a
  * conditional inner join between the specified tables where the predicate

cpp/src/join/conditional_join.cu

@@ -110,7 +110,6 @@ conditional_join(table_view const& left,
   } else {
     // Allocate storage for the counter used to get the size of the join output
     rmm::device_scalar<std::size_t> size(0, stream, mr);
-    CHECK_CUDA(stream.value());
     if (has_nulls) {
       compute_conditional_join_output_size<DEFAULT_JOIN_BLOCK_SIZE, true>
         <<<config.num_blocks, config.num_threads_per_block, shmem_size_per_block, stream.value()>>>(
@@ -260,7 +259,6 @@ std::size_t compute_conditional_join_output_size(table_view const& left,
 
   // Allocate storage for the counter used to get the size of the join output
   rmm::device_scalar<std::size_t> size(0, stream, mr);
-  CHECK_CUDA(stream.value());
 
   // Determine number of output rows without actually building the output to simply
   // find what the size of the output will be.

cpp/src/join/join_common_utils.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -25,6 +25,7 @@
 
 #include <rmm/mr/device/polymorphic_allocator.hpp>
 
+#include <cuco/static_map.cuh>
 #include <cuco/static_multimap.cuh>
 
 #include <limits>
@@ -60,6 +61,9 @@ using mixed_multimap_type = cuco::static_multimap<hash_value_type,
                                                   hash_table_allocator_type,
                                                   cuco::double_hashing<1, hash_type, hash_type>>;
 
+using semi_map_type = cuco::
+  static_map<hash_value_type, size_type, cuda::thread_scope_device, hash_table_allocator_type>;
+
 using row_hash = cudf::row_hasher<default_hash, cudf::nullate::DYNAMIC>;
 
 using row_equality = cudf::row_equality_comparator<cudf::nullate::DYNAMIC>;