Add compound aggregations to cudf::segmented_reduce

conda/recipes/libcudf/meta.yaml

@@ -109,6 +109,7 @@ outputs:
         - test -f $PREFIX/include/cudf/detail/scan.hpp
         - test -f $PREFIX/include/cudf/detail/scatter.hpp
         - test -f $PREFIX/include/cudf/detail/search.hpp
+        - test -f $PREFIX/include/cudf/detail/segmented_reduction_functions.hpp
         - test -f $PREFIX/include/cudf/detail/sequence.hpp
         - test -f $PREFIX/include/cudf/detail/sorting.hpp
         - test -f $PREFIX/include/cudf/detail/stream_compaction.hpp

cpp/CMakeLists.txt

@@ -440,13 +440,18 @@ add_library(
   src/reductions/scan/scan.cpp
   src/reductions/scan/scan_exclusive.cu
   src/reductions/scan/scan_inclusive.cu
-  src/reductions/segmented_all.cu
-  src/reductions/segmented_any.cu
-  src/reductions/segmented_max.cu
-  src/reductions/segmented_min.cu
-  src/reductions/segmented_product.cu
-  src/reductions/segmented_reductions.cpp
-  src/reductions/segmented_sum.cu
+  src/reductions/segmented/all.cu
+  src/reductions/segmented/any.cu
+  src/reductions/segmented/max.cu
+  src/reductions/segmented/mean.cu
+  src/reductions/segmented/min.cu
+  src/reductions/segmented/product.cu
+  src/reductions/segmented/reductions.cpp
+  src/reductions/segmented/std.cu
+  src/reductions/segmented/sum.cu
+  src/reductions/segmented/sum_of_squares.cu
+  src/reductions/segmented/update_validity.cu
+  src/reductions/segmented/var.cu
   src/reductions/std.cu
   src/reductions/sum.cu
   src/reductions/sum_of_squares.cu

cpp/include/cudf/detail/aggregation/aggregation.hpp

@@ -292,7 +292,9 @@ class all_aggregation final : public reduce_aggregation, public segmented_reduce
 /**
  * @brief Derived class for specifying a sum_of_squares aggregation
  */
-class sum_of_squares_aggregation final : public groupby_aggregation, public reduce_aggregation {
+class sum_of_squares_aggregation final : public groupby_aggregation,
+                                         public reduce_aggregation,
+                                         public segmented_reduce_aggregation {
  public:
   sum_of_squares_aggregation() : aggregation(SUM_OF_SQUARES) {}
 
@@ -313,7 +315,8 @@ class sum_of_squares_aggregation final : public groupby_aggregation, public redu
  */
 class mean_aggregation final : public rolling_aggregation,
                                public groupby_aggregation,
-                               public reduce_aggregation {
+                               public reduce_aggregation,
+                               public segmented_reduce_aggregation {
  public:
   mean_aggregation() : aggregation(MEAN) {}
 
@@ -353,7 +356,8 @@ class m2_aggregation : public groupby_aggregation {
  */
 class std_var_aggregation : public rolling_aggregation,
                             public groupby_aggregation,
-                            public reduce_aggregation {
+                            public reduce_aggregation,
+                            public segmented_reduce_aggregation {
  public:
   size_type _ddof;  ///< Delta degrees of freedom
 

cpp/include/cudf/detail/reduction.cuh

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2022, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2023, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -16,7 +16,7 @@
 
 #pragma once
 
-#include "reduction_operators.cuh"
+#include <cudf/detail/reduction_operators.cuh>
 
 #include <cudf/column/column_factories.hpp>
 #include <cudf/utilities/type_dispatcher.hpp>
@@ -27,7 +27,6 @@
 #include <rmm/exec_policy.hpp>
 
 #include <cub/device/device_reduce.cuh>
-#include <cub/device/device_segmented_reduce.cuh>
 
 #include <thrust/for_each.h>
 #include <thrust/iterator/iterator_traits.h>
@@ -229,92 +228,6 @@ std::unique_ptr<scalar> reduce(InputIterator d_in,
   return std::unique_ptr<scalar>(result);
 }
 
-/**
- * @brief Compute the specified simple reduction over each of the segments in the
- * input range of elements.
- *
- * @tparam InputIterator    the input column iterator
- * @tparam OffsetIterator   the offset column iterator
- * @tparam OutputIterator   the output column iterator
- * @tparam BinaryOp         the device binary operator used to reduce
- * @tparam OutputType       the output type of reduction
- *
- * @param[in] d_in          the begin iterator to input
- * @param[in] d_offset_begin the begin iterator to offset
- * @param[in] d_offset_end  the end iterator to offset. Note: This is
- * num_segments+1 elements past `d_offset_begin`.
- * @param[out] d_out        the begin iterator to output
- * @param[in] binary_op     the reduction operator
- * @param[in] identity      the identity element of the reduction operator
- * @param[in] initial_value Initial value of the reduction
- * @param[in] stream        CUDA stream used for device memory operations and kernel launches
- *
- */
-template <typename InputIterator,
-          typename OffsetIterator,
-          typename OutputIterator,
-          typename BinaryOp,
-          typename OutputType = typename thrust::iterator_value<OutputIterator>::type,
-          typename std::enable_if_t<is_fixed_width<OutputType>() &&
-                                    !cudf::is_fixed_point<OutputType>()>* = nullptr>
-void segmented_reduce(InputIterator d_in,
-                      OffsetIterator d_offset_begin,
-                      OffsetIterator d_offset_end,
-                      OutputIterator d_out,
-                      BinaryOp binary_op,
-                      OutputType initial_value,
-                      rmm::cuda_stream_view stream)
-{
-  auto const num_segments = static_cast<size_type>(std::distance(d_offset_begin, d_offset_end)) - 1;
-
-  // Allocate temporary storage
-  rmm::device_buffer d_temp_storage;
-  size_t temp_storage_bytes = 0;
-  cub::DeviceSegmentedReduce::Reduce(d_temp_storage.data(),
-                                     temp_storage_bytes,
-                                     d_in,
-                                     d_out,
-                                     num_segments,
-                                     d_offset_begin,
-                                     d_offset_begin + 1,
-                                     binary_op,
-                                     initial_value,
-                                     stream.value());
-  d_temp_storage = rmm::device_buffer{temp_storage_bytes, stream};
-
-  // Run reduction
-  cub::DeviceSegmentedReduce::Reduce(d_temp_storage.data(),
-                                     temp_storage_bytes,
-                                     d_in,
-                                     d_out,
-                                     num_segments,
-                                     d_offset_begin,
-                                     d_offset_begin + 1,
-                                     binary_op,
-                                     initial_value,
-                                     stream.value());
-}
-
-template <typename InputIterator,
-          typename OffsetIterator,
-          typename OutputIterator,
-          typename BinaryOp,
-          typename OutputType = typename thrust::iterator_value<OutputIterator>::type,
-          typename std::enable_if_t<!(is_fixed_width<OutputType>() &&
-                                      !cudf::is_fixed_point<OutputType>())>* = nullptr>
-void segmented_reduce(InputIterator,
-                      OffsetIterator,
-                      OffsetIterator,
-                      OutputIterator,
-                      BinaryOp,
-                      OutputType,
-                      rmm::cuda_stream_view)
-{
-  CUDF_FAIL(
-    "Unsupported data types called on segmented_reduce. Only numeric and chrono types are "
-    "supported.");
-}
-
 }  // namespace detail
 }  // namespace reduction
 }  // namespace cudf

cpp/include/cudf/detail/reduction_functions.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2022, NVIDIA CORPORATION.
+ * Copyright (c) 2019-2023, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -338,171 +338,5 @@ std::unique_ptr<scalar> merge_sets(
   rmm::cuda_stream_view stream,
   rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
 
-/**
- * @brief Compute sum of each segment in input column.
- *
- * If an input segment is empty, the segment result is null.
- *
- * @throw cudf::logic_error if input column type is not convertible to `output_dtype`.
- * @throw cudf::logic_error if `output_dtype` is not an arithmetic type.
- *
- * @param col Input column to compute sum
- * @param offsets Indices to identify segment boundaries
- * @param output_dtype Data type of return type and typecast elements of input column
- * @param null_handling If `null_policy::INCLUDE`, all elements in a segment must be valid for the
- * reduced value to be valid. If `null_policy::EXCLUDE`, the reduced value is valid if any element
- * in the segment is valid.
- * @param init Initial value of each sum
- * @param stream CUDA stream used for device memory operations and kernel launches
- * @param mr Device memory resource used to allocate the returned column's device memory
- * @return Sums of segments in type `output_dtype`
- */
-std::unique_ptr<column> segmented_sum(
-  column_view const& col,
-  device_span<size_type const> offsets,
-  data_type const output_dtype,
-  null_policy null_handling,
-  std::optional<std::reference_wrapper<scalar const>> init,
-  rmm::cuda_stream_view stream,
-  rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
-
-/**
- * @brief Computes product of each segment in input column.
- *
- * If an input segment is empty, the segment result is null.
- *
- * @throw cudf::logic_error if input column type is not convertible to `output_dtype`.
- * @throw cudf::logic_error if `output_dtype` is not an arithmetic type.
- *
- * @param col Input column to compute product
- * @param offsets Indices to identify segment boundaries
- * @param output_dtype data type of return type and typecast elements of input column
- * @param null_handling If `null_policy::INCLUDE`, all elements in a segment must be valid for the
- * reduced value to be valid. If `null_policy::EXCLUDE`, the reduced value is valid if any element
- * in the segment is valid.
- * @param init Initial value of each product
- * @param stream CUDA stream used for device memory operations and kernel launches
- * @param mr Device memory resource used to allocate the returned scalar's device memory
- * @return Product as scalar of type `output_dtype`
- */
-std::unique_ptr<column> segmented_product(
-  column_view const& col,
-  device_span<size_type const> offsets,
-  data_type const output_dtype,
-  null_policy null_handling,
-  std::optional<std::reference_wrapper<scalar const>> init,
-  rmm::cuda_stream_view stream,
-  rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
-
-/**
- * @brief Compute minimum of each segment in input column.
- *
- * If an input segment is empty, the segment result is null.
- *
- * @throw cudf::logic_error if input column type is convertible to `output_dtype`.
- *
- * @param col Input column to compute minimum
- * @param offsets Indices to identify segment boundaries
- * @param output_dtype Data type of return type and typecast elements of input column
- * @param null_handling If `null_policy::INCLUDE`, all elements in a segment must be valid for the
- * reduced value to be valid. If `null_policy::EXCLUDE`, the reduced value is valid if any element
- * in the segment is valid.
- * @param init Initial value of each minimum
- * @param stream CUDA stream used for device memory operations and kernel launches
- * @param mr Device memory resource used to allocate the returned scalar's device memory
- * @return Minimums of segments in type `output_dtype`
- */
-std::unique_ptr<column> segmented_min(
-  column_view const& col,
-  device_span<size_type const> offsets,
-  data_type const output_dtype,
-  null_policy null_handling,
-  std::optional<std::reference_wrapper<scalar const>> init,
-  rmm::cuda_stream_view stream,
-  rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
-
-/**
- * @brief Compute maximum of each segment in input column.
- *
- * If an input segment is empty, the segment result is null.
- *
- * @throw cudf::logic_error if input column type is convertible to `output_dtype`.
- *
- * @param col Input column to compute maximum
- * @param offsets Indices to identify segment boundaries
- * @param output_dtype Data type of return type and typecast elements of input column
- * @param null_handling If `null_policy::INCLUDE`, all elements in a segment must be valid for the
- * reduced value to be valid. If `null_policy::EXCLUDE`, the reduced value is valid if any element
- * in the segment is valid.
- * @param init Initial value of each maximum
- * @param stream CUDA stream used for device memory operations and kernel launches
- * @param mr Device memory resource used to allocate the returned scalar's device memory
- * @return Maximums of segments in type `output_dtype`
- */
-std::unique_ptr<column> segmented_max(
-  column_view const& col,
-  device_span<size_type const> offsets,
-  data_type const output_dtype,
-  null_policy null_handling,
-  std::optional<std::reference_wrapper<scalar const>> init,
-  rmm::cuda_stream_view stream,
-  rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
-
-/**
- * @brief Compute if any of the values in the segment are true when typecasted to bool.
- *
- * If an input segment is empty, the segment result is null.
- *
- * @throw cudf::logic_error if input column type is not convertible to bool.
- * @throw cudf::logic_error if `output_dtype` is not bool8.
- *
- * @param col Input column to compute any
- * @param offsets Indices to identify segment boundaries
- * @param output_dtype Data type of return type and typecast elements of input column
- * @param null_handling If `null_policy::INCLUDE`, all elements in a segment must be valid for the
- * reduced value to be valid. If `null_policy::EXCLUDE`, the reduced value is valid if any element
- * in the segment is valid.
- * @param init Initial value of each any
- * @param stream CUDA stream used for device memory operations and kernel launches
- * @param mr Device memory resource used to allocate the returned scalar's device memory
- * @return Column of bool8 for the results of the segments
- */
-std::unique_ptr<column> segmented_any(
-  column_view const& col,
-  device_span<size_type const> offsets,
-  data_type const output_dtype,
-  null_policy null_handling,
-  std::optional<std::reference_wrapper<scalar const>> init,
-  rmm::cuda_stream_view stream,
-  rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
-
-/**
- * @brief Compute if all of the values in the segment are true when typecasted to bool.
- *
- * If an input segment is empty, the segment result is null.
- *
- * @throw cudf::logic_error if input column type is not convertible to bool.
- * @throw cudf::logic_error if `output_dtype` is not bool8.
- *
- * @param col Input column to compute all
- * @param offsets Indices to identify segment boundaries
- * @param output_dtype Data type of return type and typecast elements of input column
- * @param null_handling If `null_policy::INCLUDE`, all elements in a segment must be valid for the
- * reduced value to be valid. If `null_policy::EXCLUDE`, the reduced value is valid if any element
- * in the segment is valid.
- * @param init Initial value of each all
- * @param stream CUDA stream used for device memory operations and kernel launches
- * @param mr Device memory resource used to allocate the returned scalar's device memory
- * @return Column of bool8 for the results of the segments
- */
-std::unique_ptr<column> segmented_all(
-  column_view const& col,
-  device_span<size_type const> offsets,
-  data_type const output_dtype,
-  null_policy null_handling,
-  std::optional<std::reference_wrapper<scalar const>> init,
-  rmm::cuda_stream_view stream,
-  rmm::mr::device_memory_resource* mr = rmm::mr::get_current_device_resource());
-
 }  // namespace reduction
 }  // namespace cudf