Reduce libcudf binary size by trimming instantiations

cpp/CMakeLists.txt

@@ -943,18 +943,16 @@ add_library(
   src/reductions/segmented/nunique.cu
   src/reductions/segmented/product.cu
   src/reductions/segmented/reductions.cpp
-  src/reductions/segmented/std.cu
+  src/reductions/segmented/std_var.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/std_var.cu
   src/reductions/sum.cu
   src/reductions/sum_of_squares.cu
   src/reductions/sum_overflow.cu
   src/reductions/unique_count.cu
   src/reductions/unique_count_column.cu
-  src/reductions/var.cu
   src/replace/clamp.cu
   src/replace/nans.cu
   src/replace/nulls.cu

cpp/include/cudf/strings/detail/copy_if_else.cuh

@@ -78,7 +78,7 @@ std::unique_ptr<cudf::column> copy_if_else(StringIterLeft lhs_begin,
                     });
 
   // convert vector into strings column
-  auto result = make_strings_column(indices.begin(), indices.end(), stream, mr);
+  auto result = cudf::make_strings_column(indices, stream, mr);
   result->set_null_mask(std::move(null_mask), null_count);
   return result;
 }

cpp/include/cudf/strings/detail/strings_children.cuh

@@ -14,8 +14,10 @@
 #include <cudf/strings/detail/utilities.hpp>
 #include <cudf/strings/utilities.hpp>
 #include <cudf/utilities/default_stream.hpp>
+#include <cudf/utilities/export.hpp>
 #include <cudf/utilities/memory_resource.hpp>
 #include <cudf/utilities/prefetch.hpp>
+#include <cudf/utilities/span.hpp>
 
 #include <rmm/exec_policy.hpp>
 
@@ -31,6 +33,21 @@ namespace cudf {
 namespace strings {
 namespace detail {
 
+/**
+ * @brief Create an offsets column from already-materialized string sizes.
+ *
+ * This overload centralizes the common size_type input case so callers do not each
+ * instantiate the same CUB scan kernels.
+ *
+ * @param sizes The per-string byte sizes
+ * @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 Offsets column and total bytes
+ * @throw std::overflow_error if the output exceeds the column size limit
+ */
+CUDF_EXPORT std::pair<std::unique_ptr<column>, int64_t> make_offsets_child_column(
+  device_span<size_type const> sizes, cuda::stream_ref stream, rmm::device_async_resource_ref mr);
+
 template <typename Iter>
 struct string_offsets_fn {
   Iter _begin;
@@ -242,8 +259,8 @@ auto make_strings_children(SizeAndExecuteFunction size_and_exec_fn,
   for_each_fn(size_and_exec_fn);
 
   // Convert the sizes to offsets
-  auto [offsets_column, bytes] = cudf::strings::detail::make_offsets_child_column(
-    output_sizes.begin(), output_sizes.end(), stream, mr);
+  auto [offsets_column, bytes] =
+    cudf::strings::detail::make_offsets_child_column(output_sizes, stream, mr);
   size_and_exec_fn.d_offsets =
     cudf::detail::offsetalator_factory::make_input_iterator(offsets_column->view());
 

cpp/src/column/column_factories.cu

@@ -72,7 +72,7 @@ std::unique_ptr<cudf::column> column_from_scalar_dispatch::operator()<cudf::stri
     indices.begin(),
     indices.end(),
     row_value);
-  return cudf::strings::detail::make_strings_column(indices.begin(), indices.end(), stream, mr);
+  return cudf::make_strings_column(indices, stream, mr);
 }
 
 template <>

cpp/src/interop/from_arrow_device.cu

@@ -194,8 +194,7 @@ dispatch_tuple_t dispatch_from_arrow_device::operator()<cudf::string_view>(
                       d_indices.begin(),
                       binary_view_to_string_index_pair{d_items, d_ptrs, d_mask, skip_mask});
     // gather strings into output column
-    auto out_col =
-      cudf::strings::detail::make_strings_column(d_indices.begin(), d_indices.end(), stream, mr);
+    auto out_col = cudf::make_strings_column(d_indices, stream, mr);
     owned.emplace_back(std::move(out_col));
     stream.synchronize();
     return std::make_tuple<column_view, owned_columns_t>(owned.front()->view(), std::move(owned));

cpp/src/interop/from_arrow_host_strings.cu

@@ -114,7 +114,7 @@ std::unique_ptr<column> from_arrow_stringview(ArrowSchemaView const* schema,
     });
 
   stream.synchronize();
-  return cudf::strings::detail::make_strings_column(d_indices.begin(), d_indices.end(), stream, mr);
+  return cudf::make_strings_column(d_indices, stream, mr);
 }
 
 }  // namespace

cpp/src/io/parquet/experimental/variant_extract.cu

@@ -922,7 +922,7 @@ std::unique_ptr<column> get_variant_field(column_view const& variant_column,
 
   // Convert sizes to offsets
   auto [offsets_column, total_bytes] =
-    cudf::strings::detail::make_offsets_child_column(d_sizes.begin(), d_sizes.end(), stream, mr);
+    cudf::strings::detail::make_offsets_child_column(d_sizes, stream, mr);
   CUDF_EXPECTS(total_bytes <= std::numeric_limits<size_type>::max(),
                "VARIANT extracted bytes exceed cudf size_type limit",
                std::overflow_error);

cpp/src/io/utilities/data_casting.cu

@@ -857,9 +857,8 @@ static std::unique_ptr<column> parse_string(string_view_pair_it str_tuples,
     CUDF_CUDA_TRY(cudaGetLastError());
   }
 
-  auto [offsets, bytes] =
-    cudf::strings::detail::make_offsets_child_column(sizes.begin(), sizes.end(), stream, mr);
-  auto d_offsets = cudf::detail::offsetalator_factory::make_input_iterator(offsets->view());
+  auto [offsets, bytes] = cudf::strings::detail::make_offsets_child_column(sizes, stream, mr);
+  auto d_offsets        = cudf::detail::offsetalator_factory::make_input_iterator(offsets->view());
 
   // CHARS column
   rmm::device_uvector<char> chars(bytes, stream, mr);

cpp/src/json/json_path.cu

@@ -1,5 +1,5 @@
 /*
- * SPDX-FileCopyrightText: Copyright (c) 2021-2026, NVIDIA CORPORATION.
+ * SPDX-FileCopyrightText: Copyright (c) 2021-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
  * SPDX-License-Identifier: Apache-2.0
  */
 
@@ -1018,8 +1018,7 @@ std::unique_ptr<cudf::column> get_json_object(cudf::strings_column_view const& c
   CUDF_CUDA_TRY(cudaGetLastError());
 
   // convert sizes to offsets
-  auto [offsets, output_size] =
-    cudf::strings::detail::make_offsets_child_column(sizes.begin(), sizes.end(), stream, mr);
+  auto [offsets, output_size] = cudf::strings::detail::make_offsets_child_column(sizes, stream, mr);
   d_offsets = cudf::detail::offsetalator_factory::make_input_iterator(offsets->view());
 
   // allocate output string column

cpp/src/lists/interleave_columns.cu

@@ -202,7 +202,7 @@ struct interleave_list_entries_impl<T, std::enable_if_t<std::is_same_v<T, cudf::
                        cuda::counting_iterator<size_type>{0},
                        num_output_lists,
                        comp_fn);
-    return cudf::strings::detail::make_strings_column(indices.begin(), indices.end(), stream, mr);
+    return cudf::make_strings_column(indices, stream, mr);
   }
 };
 

cpp/src/reductions/segmented/std.cu → cpp/src/reductions/segmented/std_var.cu

@@ -14,6 +14,9 @@ namespace cudf {
 namespace reduction {
 namespace detail {
 
+// segmented_variance is intentionally co-located with segmented_standard_deviation in this
+// translation unit. Both reductions use the same var_std intermediate and segmented CUB reduction
+// shape; keeping them together avoids emitting duplicate device kernel instantiations.
 std::unique_ptr<cudf::column> segmented_standard_deviation(column_view const& col,
                                                            device_span<size_type const> offsets,
                                                            cudf::data_type const output_dtype,
@@ -27,6 +30,19 @@ std::unique_ptr<cudf::column> segmented_standard_deviation(column_view const& co
     col.type(), reducer(), col, offsets, output_dtype, null_handling, ddof, stream, mr);
 }
 
+std::unique_ptr<cudf::column> segmented_variance(column_view const& col,
+                                                 device_span<size_type const> offsets,
+                                                 cudf::data_type const output_dtype,
+                                                 null_policy null_handling,
+                                                 size_type ddof,
+                                                 cuda::stream_ref stream,
+                                                 rmm::device_async_resource_ref mr)
+{
+  using reducer = compound::detail::compound_segmented_dispatcher<op::variance>;
+  return cudf::type_dispatcher(
+    col.type(), reducer(), col, offsets, output_dtype, null_handling, ddof, stream, mr);
+}
+
 }  // namespace detail
 }  // namespace reduction
 }  // namespace cudf

cpp/src/reductions/std.cu → cpp/src/reductions/std_var.cu

@@ -15,6 +15,9 @@ namespace cudf {
 namespace reduction {
 namespace detail {
 
+// variance is intentionally co-located with standard_deviation in this translation unit. Both
+// reductions use the same var_std intermediate and CUB reduction shape; keeping them together
+// avoids emitting duplicate device kernel instantiations.
 std::unique_ptr<cudf::scalar> standard_deviation(column_view const& col,
                                                  cudf::data_type const output_dtype,
                                                  size_type ddof,
@@ -27,6 +30,18 @@ std::unique_ptr<cudf::scalar> standard_deviation(column_view const& col,
   return cudf::type_dispatcher(col_type, reducer(), col, output_dtype, ddof, stream, mr);
 }
 
+std::unique_ptr<cudf::scalar> variance(column_view const& col,
+                                       cudf::data_type const output_dtype,
+                                       size_type ddof,
+                                       cuda::stream_ref stream,
+                                       rmm::device_async_resource_ref mr)
+{
+  using reducer = compound::detail::element_type_dispatcher<op::variance>;
+  auto col_type =
+    cudf::is_dictionary(col.type()) ? dictionary_column_view(col).keys().type() : col.type();
+  return cudf::type_dispatcher(col_type, reducer(), col, output_dtype, ddof, stream, mr);
+}
+
 }  // namespace detail
 }  // namespace reduction
 }  // namespace cudf

cpp/src/replace/clamp.cu

@@ -92,7 +92,7 @@ std::unique_ptr<cudf::column> clamp_string_column(strings_column_view const& inp
                     indices.begin(),
                     fn);
 
-  return cudf::strings::detail::make_strings_column(indices.begin(), indices.end(), stream, mr);
+  return cudf::make_strings_column(indices, stream, mr);
 }
 
 template <typename T, typename OptionalIterator, typename ReplaceIterator>

cpp/src/reshape/interleave_columns.cu

@@ -176,7 +176,7 @@ struct interleave_columns_impl<T, std::enable_if_t<std::is_same_v<T, cudf::strin
                       indices.begin(),
                       interleave_strings_fn{*d_table});
 
-    return cudf::strings::detail::make_strings_column(indices.begin(), indices.end(), stream, mr);
+    return cudf::make_strings_column(indices, stream, mr);
   }
 };
 

cpp/src/strings/case.cu

@@ -455,7 +455,7 @@ std::unique_ptr<column> convert_case(strings_column_view const& input,
         ccfn, *d_strings, sizes.data());
     CUDF_CUDA_TRY(cudaGetLastError());
     // convert sizes to offsets
-    return cudf::strings::detail::make_offsets_child_column(sizes.begin(), sizes.end(), stream, mr);
+    return cudf::strings::detail::make_offsets_child_column(sizes, stream, mr);
   }();
 
   // build sub-offsets

cpp/src/strings/extract/extract.cu

@@ -165,7 +165,7 @@ std::unique_ptr<column> extract_single(strings_column_view const& input,
   launch_transform_kernel(
     extract_single_fn{*d_strings, group}, *d_prog, indices.data(), input.size(), stream);
 
-  return make_strings_column(indices.begin(), indices.end(), stream, mr);
+  return cudf::make_strings_column(indices, stream, mr);
 }
 
 }  // namespace detail

cpp/src/strings/extract/extract_all.cu

@@ -136,7 +136,7 @@ std::unique_ptr<column> extract_all_record(strings_column_view const& input,
   launch_for_each_kernel(
     extract_fn{*d_strings, d_offsets, indices.data()}, *d_prog, strings_count, stream);
 
-  auto strings_output = make_strings_column(indices.begin(), indices.end(), stream, mr);
+  auto strings_output = cudf::make_strings_column(indices, stream, mr);
 
   // Build the lists column from the offsets and the strings.
   return make_lists_column(

cpp/src/strings/filling/fill.cu

@@ -68,7 +68,7 @@ std::unique_ptr<column> fill(strings_column_view const& input,
                     indices.begin(),
                     fn);
 
-  return make_strings_column(indices.begin(), indices.end(), stream, mr);
+  return cudf::make_strings_column(indices, stream, mr);
 }
 
 }  // namespace detail

cpp/src/strings/merge/merge.cu

@@ -55,7 +55,7 @@ std::unique_ptr<column> merge(strings_column_view const& lhs,
                     });
 
   // convert vector into strings column
-  return make_strings_column(indices.begin(), indices.end(), stream, mr);
+  return cudf::make_strings_column(indices, stream, mr);
 }
 
 }  // namespace detail

cpp/src/strings/positions.cu

@@ -61,8 +61,7 @@ std::unique_ptr<column> create_offsets_from_positions(strings_column_view const&
     });
 
   // finally, convert the counts into offsets
-  return std::get<0>(
-    cudf::strings::detail::make_offsets_child_column(counts.begin(), counts.end(), stream, mr));
+  return std::get<0>(cudf::strings::detail::make_offsets_child_column(counts, stream, mr));
 }
 
 }  // namespace cudf::strings::detail

cpp/src/strings/replace/find_replace.cu

@@ -69,7 +69,7 @@ std::unique_ptr<cudf::column> find_and_replace_all(
                     indices.begin(),
                     find_replace_fn{*d_input, *d_values_to_replace, *d_replacements});
 
-  return make_strings_column(indices.begin(), indices.end(), stream, mr);
+  return cudf::make_strings_column(indices, stream, mr);
 }
 
 }  // namespace detail

cpp/src/strings/replace/multi.cu

@@ -401,12 +401,11 @@ std::unique_ptr<column> replace_character_parallel(strings_column_view const& in
     });
 
   // use this utility to gather the string parts into a contiguous chars column
-  auto chars      = make_strings_column(indices.begin(), indices.end(), stream, mr);
+  auto chars      = cudf::make_strings_column(indices, stream, mr);
   auto chars_data = chars->release().data;
 
   // create offsets from the sizes
-  offsets = std::get<0>(
-    cudf::strings::detail::make_offsets_child_column(counts.begin(), counts.end(), stream, mr));
+  offsets = std::get<0>(cudf::strings::detail::make_offsets_child_column(counts, stream, mr));
 
   // build the strings columns from the chars and offsets
   return make_strings_column(strings_count,

cpp/src/strings/replace/replace.cu

@@ -345,12 +345,11 @@ std::unique_ptr<column> replace_character_parallel(strings_column_view const& in
     });
 
   // use this utility to gather the string parts into a contiguous chars column
-  auto chars      = make_strings_column(indices.begin(), indices.end(), stream, mr);
+  auto chars      = cudf::make_strings_column(indices, stream, mr);
   auto chars_data = chars->release().data;
 
   // create offsets from the sizes
-  offsets = std::get<0>(
-    cudf::strings::detail::make_offsets_child_column(counts.begin(), counts.end(), stream, mr));
+  offsets = std::get<0>(cudf::strings::detail::make_offsets_child_column(counts, stream, mr));
 
   // build the strings columns from the chars and offsets
   return make_strings_column(strings_count,

cpp/src/strings/search/findall.cu

@@ -153,7 +153,7 @@ std::unique_ptr<column> findall(strings_column_view const& input,
     });
   }();
 
-  auto strings_output = make_strings_column(indices.begin(), indices.end(), stream, mr);
+  auto strings_output = cudf::make_strings_column(indices, stream, mr);
 
   // Build the lists column from the offsets and the strings
   return make_lists_column(input.size(),