Implement all methods of groupby rank aggregation in libcudf, python

cpp/benchmarks/CMakeLists.txt

@@ -194,10 +194,18 @@ ConfigureBench(FILL_BENCH filling/repeat.cpp)
 # ##################################################################################################
 # * groupby benchmark -----------------------------------------------------------------------------
 ConfigureBench(
-  GROUPBY_BENCH groupby/group_sum.cu groupby/group_nth.cu groupby/group_shift.cu
-  groupby/group_struct.cu groupby/group_no_requests.cu groupby/group_scan.cu
+  GROUPBY_BENCH
+  groupby/group_sum.cu
+  groupby/group_nth.cu
+  groupby/group_shift.cu
+  groupby/group_struct.cu
+  groupby/group_no_requests.cu
+  groupby/group_scan.cu
+  groupby/group_rank_benchmark.cu
 )
 
+ConfigureNVBench(GROUPBY_NVBENCH groupby/group_rank_benchmark.cu)
+
 # ##################################################################################################
 # * hashing benchmark -----------------------------------------------------------------------------
 ConfigureBench(HASHING_BENCH hashing/hash.cpp hashing/partition.cpp)

cpp/benchmarks/groupby/group_rank_benchmark.cu

@@ -0,0 +1,109 @@
+/*
+ * Copyright (c) 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.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include <benchmarks/common/generate_input.hpp>
+#include <benchmarks/fixture/rmm_pool_raii.hpp>
+#include <benchmarks/synchronization/synchronization.hpp>
+
+#include <cudf/groupby.hpp>
+#include <cudf/sorting.hpp>
+#include <cudf/table/table_view.hpp>
+#include <cudf/types.hpp>
+
+#include <nvbench/nvbench.cuh>
+
+template <cudf::rank_method method>
+static void nvbench_groupby_rank(nvbench::state& state,
+                                 nvbench::type_list<nvbench::enum_type<method>>)
+{
+  using namespace cudf;
+  using type           = int64_t;
+  constexpr auto dtype = type_to_id<int64_t>();
+  cudf::rmm_pool_raii pool_raii;
+
+  bool const is_sorted              = state.get_int64("is_sorted");
+  cudf::size_type const column_size = state.get_int64("data_size");
+  constexpr int num_groups          = 100;
+
+  data_profile profile;
+  profile.set_null_frequency(std::nullopt);
+  profile.set_cardinality(0);
+  profile.set_distribution_params<type>(dtype, distribution_id::UNIFORM, 0, num_groups);
+
+  auto source_table = create_random_table({dtype, dtype}, row_count{column_size}, profile);
+
+  // values to be pre-sorted too for groupby rank
+  if (is_sorted) source_table = cudf::sort(*source_table);
+
+  table_view keys{{source_table->view().column(0)}};
+  column_view order_by{source_table->view().column(1)};
+
+  auto agg = cudf::make_rank_aggregation<groupby_scan_aggregation>(method);
+  std::vector<groupby::scan_request> requests;
+  requests.emplace_back(groupby::scan_request());
+  requests[0].values = order_by;
+  requests[0].aggregations.push_back(std::move(agg));
+
+  groupby::groupby gb_obj(keys, null_policy::EXCLUDE, is_sorted ? sorted::YES : sorted::NO);
+
+  state.exec(nvbench::exec_tag::sync, [&](nvbench::launch& launch) {
+    rmm::cuda_stream_view stream_view{launch.get_stream()};
+    // groupby scan uses sort implementation
+    auto result = gb_obj.scan(requests);
+  });
+}
+
+enum class rank_method : int32_t {};
+
+NVBENCH_DECLARE_ENUM_TYPE_STRINGS(
+  cudf::rank_method,
+  [](cudf::rank_method value) {
+    switch (value) {
+      case cudf::rank_method::FIRST: return "FIRST";
+      case cudf::rank_method::AVERAGE: return "AVERAGE";
+      case cudf::rank_method::MIN: return "MIN";
+      case cudf::rank_method::MAX: return "MAX";
+      case cudf::rank_method::DENSE: return "DENSE";
+      default: return "unknown";
+    }
+  },
+  [](cudf::rank_method value) {
+    switch (value) {
+      case cudf::rank_method::FIRST: return "cudf::rank_method::FIRST";
+      case cudf::rank_method::AVERAGE: return "cudf::rank_method::AVERAGE";
+      case cudf::rank_method::MIN: return "cudf::rank_method::MIN";
+      case cudf::rank_method::MAX: return "cudf::rank_method::MAX";
+      case cudf::rank_method::DENSE: return "cudf::rank_method::DENSE";
+      default: return "unknown";
+    }
+  })
+
+using methods = nvbench::enum_type_list<cudf::rank_method::AVERAGE,
+                                        cudf::rank_method::DENSE,
+                                        cudf::rank_method::FIRST,
+                                        cudf::rank_method::MAX,
+                                        cudf::rank_method::MIN>;
+
+NVBENCH_BENCH_TYPES(nvbench_groupby_rank, NVBENCH_TYPE_AXES(methods))
+  .set_type_axes_names({"rank_method"})
+  .set_name("groupby_rank")
+  .add_int64_axis("data_size",
+                  {
+                    1000000,    // 1M
+                    10000000,   // 10M
+                    100000000,  // 100M
+                  })
+
+  .add_int64_axis("is_sorted", {0, 1});

cpp/include/cudf/aggregation.hpp

@@ -43,6 +43,32 @@ namespace detail {
 class simple_aggregations_collector;
 class aggregation_finalizer;
 }  // namespace detail
+
+/**
+ * @brief Tie-breaker method to use for ranking the column.
+ *
+ * @see cudf::make_rank_aggregation for more details.
+ * @ingroup column_sort
+ */
+enum class rank_method : int32_t {
+  FIRST,    ///< stable sort order ranking (no ties)
+  AVERAGE,  ///< mean of first in the group
+  MIN,      ///< min of first in the group
+  MAX,      ///< max of first in the group
+  DENSE     ///< rank always increases by 1 between groups
+};
+
+/**
+ * @brief Whether returned rank should be percentage or not and
+ *  mention the type of percentage normalization.
+ *
+ */
+enum class rank_percentage : int32_t {
+  NONE,             ///< rank
+  ZERO_NORMALIZED,  ///< rank / count
+  ONE_NORMALIZED    ///< (rank - 1) / (count - 1)
+};
+
 /**
  * @brief Abstract base class for specifying the desired aggregation in an
  * `aggregation_request`.
@@ -77,9 +103,7 @@ class aggregation {
     NUNIQUE,         ///< count number of unique elements
     NTH_ELEMENT,     ///< get the nth element
     ROW_NUMBER,      ///< get row-number of current index (relative to rolling window)
-    RANK,            ///< get rank       of current index
-    DENSE_RANK,      ///< get dense rank of current index
-    PERCENT_RANK,    ///< get percent (i.e. fractional) rank of current index
+    RANK,            ///< get rank of current index
     COLLECT_LIST,    ///< collect values into a list
     COLLECT_SET,     ///< collect values into a list without duplicate entries
     LEAD,            ///< window function, accesses row at specified offset following current row
@@ -323,20 +347,24 @@ std::unique_ptr<Base> make_row_number_aggregation();
 /**
  * @brief Factory to create a RANK aggregation
  *
- * `RANK` returns a non-nullable column of size_type "ranks": the number of rows preceding or
- * equal to the current row plus one. As a result, ranks are not unique and gaps will appear in
- * the ranking sequence.
+ * `RANK` returns a column of size_type or double "ranks" (see note 3 below for how the
+ * data type is determined) for a given rank method and column order.
+ * If nulls are excluded, the rank will be null for those rows, otherwise a non-nullable column is
+ * returned. Double precision column is returned only when percentage!=NONE and when rank method is
+ * average.
  *
  * This aggregation only works with "scan" algorithms. The input column into the group or
  * ungrouped scan is an orderby column that orders the rows that the aggregate function ranks.
  * If rows are ordered by more than one column, the orderby input column should be a struct
  * column containing the ordering columns.
  *
  * Note:
- *  1. This method requires that the rows are presorted by the group keys and order_by columns.
- *  2. `RANK` aggregations will return a fully valid column regardless of null_handling policy
- *     specified in the scan.
- *  3. `RANK` aggregations are not compatible with exclusive scans.
+ *  1. This method could work faster with the rows that are presorted by the group keys and order_by
+ *     columns. Though groupby object does not require order_by column to be sorted, groupby rank
+ *     scan aggregation does require the order_by column to be sorted if the keys are sorted.
+ *  2. `RANK` aggregations are not compatible with exclusive scans.
+ *  3. All rank methods except AVERAGE method and percentage!=NONE returns size_type column.
+ *     For AVERAGE method and percentage!=NONE, the return type is double column.
  *
  * @code{.pseudo}
  * Example: Consider a motor-racing statistics dataset, containing the following columns:
@@ -362,123 +390,37 @@ std::unique_ptr<Base> make_row_number_aggregation();
  * A grouped rank aggregation scan with:
  *   groupby column      : venue
  *   input orderby column: time
- * Produces the following rank column:
- * {   1,     2,     3,     3,     5,      1,     2,     2,     4,     5}
- * (This corresponds to the following grouping and `driver` rows:)
- * { "HAM", "LEC", "BOT", "NOR", "RIC",  "RIC", "NOR", "BOT", "LEC", "PER" }
- *   <----------silverstone----------->|<-------------monza-------------->
- * @endcode
- */
-template <typename Base = aggregation>
-std::unique_ptr<Base> make_rank_aggregation();
-
-/**
- * @brief Factory to create a DENSE_RANK aggregation
- *
- * `DENSE_RANK` returns a non-nullable column of size_type "dense ranks": the preceding unique
- * value's rank plus one. As a result, ranks are not unique but there are no gaps in the ranking
- * sequence (unlike RANK aggregations).
- *
- * This aggregation only works with "scan" algorithms. The input column into the group or
- * ungrouped scan is an orderby column that orders the rows that the aggregate function ranks.
- * If rows are ordered by more than one column, the orderby input column should be a struct
- * column containing the ordering columns.
- *
- * Note:
- *  1. This method requires that the rows are presorted by the group keys and order_by columns.
- *  2. `DENSE_RANK` aggregations will return a fully valid column regardless of null_handling
- *     policy specified in the scan.
- *  3. `DENSE_RANK` aggregations are not compatible with exclusive scans.
- *
- * @code{.pseudo}
- * Example: Consider a motor-racing statistics dataset, containing the following columns:
- *   1. venue:  (STRING) Location of the race event
- *   2. driver: (STRING) Name of the car driver (abbreviated to 3 characters)
- *   3. time:   (INT32)  Time taken to complete the circuit
- *
- * For the following presorted data:
+ * Produces the following rank column for each methods:
+ * first:   {   1,     2,     3,     4,     5,      1,     2,     3,     4,     5}
+ * average: {   1,     2,   3.5,   3.5,     5,      1,   2.5,   2.5,     4,     5}
+ * min:     {   1,     2,     3,     3,     5,      1,     2,     2,     4,     5}
+ * max:     {   1,     2,     4,     4,     5,      1,     3,     3,     4,     5}
+ * dense:   {   1,     2,     3,     3,     4,      1,     2,     2,     3,     4}
+ * This corresponds to the following grouping and `driver` rows:
+ *          { "HAM", "LEC", "BOT", "NOR", "RIC",  "RIC", "NOR", "BOT", "LEC", "PER" }
+ *            <----------silverstone----------->|<-------------monza-------------->
+ *
+ * min rank for each percentage types:
+ * NONE:             {   1,      2,     3,     3,     5,      1,     2,     2,     4,     5 }
+ * ZERO_NORMALIZED : { 0.16,  0.33,  0.50,  0.50,  0.83,   0.16,  0.33,  0.33,  0.66,  0.83 }
+ * ONE_NORMALIZED:   { 0.00,  0.25,  0.50,  0.50,  1.00,   0.00,  0.25,  0.25,  0.75,  1.00 }
+ * where count corresponds to the number of rows in the group. @see cudf::rank_percentage
  *
- *  [ //      venue,           driver,           time
- *    {   "silverstone",  "HAM" ("hamilton"),   15823},
- *    {   "silverstone",  "LEC" ("leclerc"),    15827},
- *    {   "silverstone",  "BOT" ("bottas"),     15834},  // <-- Tied for 3rd place.
- *    {   "silverstone",  "NOR" ("norris"),     15834},  // <-- Tied for 3rd place.
- *    {   "silverstone",  "RIC" ("ricciardo"),  15905},
- *    {      "monza",     "RIC" ("ricciardo"),  12154},
- *    {      "monza",     "NOR" ("norris"),     12156},  // <-- Tied for 2nd place.
- *    {      "monza",     "BOT" ("bottas"),     12156},  // <-- Tied for 2nd place.
- *    {      "monza",     "LEC" ("leclerc"),    12201},
- *    {      "monza",     "PER" ("perez"),      12203}
- *  ]
- *
- * A grouped dense rank aggregation scan with:
- *   groupby column      : venue
- *   input orderby column: time
- * Produces the following dense rank column:
- * {   1,     2,     3,     3,     4,      1,     2,     2,     3,     4}
- * (This corresponds to the following grouping and `driver` rows:)
- * { "HAM", "LEC", "BOT", "NOR", "RIC",  "RIC", "NOR", "BOT", "LEC", "PER" }
- *   <----------silverstone----------->|<-------------monza-------------->
  * @endcode
- */
-template <typename Base = aggregation>
-std::unique_ptr<Base> make_dense_rank_aggregation();
-
-/**
- * @brief Factory to create a PERCENT_RANK aggregation
  *
- * `PERCENT_RANK` returns a non-nullable column of double precision "fractional" ranks.
- * For row index `i`, the percent rank of row `i` is defined as:
- *   percent_rank = (rank - 1) / (group_row_count - 1)
- * where,
- *   1. rank is the `RANK` of the row within the group
- *   2. group_row_count is the number of rows in the group
- *
- * This aggregation only works with "scan" algorithms. The input to the grouped or
- * ungrouped scan is an orderby column that orders the rows that the aggregate function ranks.
- * If rows are ordered by more than one column, the orderby input column should be a struct
- * column containing the ordering columns.
- *
- * Note:
- *  1. This method requires that the rows are presorted by the group keys and order_by columns.
- *  2. `PERCENT_RANK` aggregations will return a fully valid column regardless of null_handling
- *     policy specified in the scan.
- *  3. `PERCENT_RANK` aggregations are not compatible with exclusive scans.
- *
- * @code{.pseudo}
- * Example: Consider a motor-racing statistics dataset, containing the following columns:
- *   1. venue:  (STRING) Location of the race event
- *   2. driver: (STRING) Name of the car driver (abbreviated to 3 characters)
- *   3. time:   (INT32)  Time taken to complete the circuit
- *
- * For the following presorted data:
- *
- *  [ //      venue,           driver,           time
- *    {   "silverstone",  "HAM" ("hamilton"),   15823},
- *    {   "silverstone",  "LEC" ("leclerc"),    15827},
- *    {   "silverstone",  "BOT" ("bottas"),     15834},  // <-- Tied for 3rd place.
- *    {   "silverstone",  "NOR" ("norris"),     15834},  // <-- Tied for 3rd place.
- *    {   "silverstone",  "RIC" ("ricciardo"),  15905},
- *    {      "monza",     "RIC" ("ricciardo"),  12154},
- *    {      "monza",     "NOR" ("norris"),     12156},  // <-- Tied for 2nd place.
- *    {      "monza",     "BOT" ("bottas"),     12156},  // <-- Tied for 2nd place.
- *    {      "monza",     "LEC" ("leclerc"),    12201},
- *    {      "monza",     "PER" ("perez"),      12203}
- *  ]
- *
- * A grouped percent rank aggregation scan with:
- *   groupby column      : venue
- *   input orderby column: time
- * Produces the following percent rank column:
- * { 0.00,  0.25,  0.50,  0.50,  1.00,   0.00,  0.25,  0.25,  0.75,  1.00 }
- *
- * (This corresponds to the following grouping and `driver` rows:)
- * { "HAM", "LEC", "BOT", "NOR", "RIC",  "RIC", "NOR", "BOT", "LEC", "PER" }
- *   <----------silverstone----------->|<-------------monza-------------->
- * @endcode
+ * @param method The ranking method used for tie breaking (same values).
+ * @param column_order The desired sort order for ranking
+ * @param null_handling  flag to include nulls during ranking. If nulls are not included,
+ * the corresponding rank will be null.
+ * @param null_precedence The desired order of null compared to other elements for column
+ * @param percentage enum to denote the type of conversion of ranks to percentage in range (0,1]
  */
 template <typename Base = aggregation>
-std::unique_ptr<Base> make_percent_rank_aggregation();
+std::unique_ptr<Base> make_rank_aggregation(rank_method method,
+                                            order column_order         = order::ASCENDING,
+                                            null_policy null_handling  = null_policy::EXCLUDE,
+                                            null_order null_precedence = null_order::AFTER,
+                                            rank_percentage percentage = rank_percentage::NONE);
 
 /**
  * @brief Factory to create a COLLECT_LIST aggregation