Large strings support in cudf::concatenate (#15195)

Enables `cudf::concatenate` to create and return a large strings column (offsets are INT64).

This also introduces the `LIBCUDF_LARGE_STRINGS_ENABLED` environment variable and utilities around it.
One internal utility checks the value so appropriate logic can either throw an overflow exception or build INT64 offsets as appropriate. 

The `cudf::test::large_strings_enabler` is introduced to set/unset the env var for individual tests are needed.
A follow on PR will attempt to consolidate these kinds of tests with a specialized test fixture using this utility class.

Authors:
  - David Wendt (https://github.com/davidwendt)

Approvers:
  - Bradley Dice (https://github.com/bdice)
  - Mike Wilson (https://github.com/hyperbolic2346)

URL: #15195

Author: davidwendt

cpp/include/cudf/strings/detail/utilities.hpp

@@ -27,6 +27,24 @@ namespace cudf {
 namespace strings {
 namespace detail {
 
+/**
+ * @brief Create an offsets column to be a child of a strings column
+ *
+ * This will return the properly typed column to be filled in by the caller
+ * given the number of bytes to address.
+ *
+ * @param chars_bytes Number of bytes for the chars in the strings column
+ * @param count Number of elements for the offsets column.
+ *              This is the number of rows in the parent strings column +1.
+ * @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 The offsets child column for a strings column
+ */
+std::unique_ptr<column> create_offsets_child_column(int64_t chars_bytes,
+                                                    size_type count,
+                                                    rmm::cuda_stream_view stream,
+                                                    rmm::mr::device_memory_resource* mr);
+
 /**
  * @brief Creates a string_view vector from a strings column.
  *
@@ -52,6 +70,15 @@ rmm::device_uvector<string_view> create_string_vector_from_column(
  */
 int64_t get_offset64_threshold();
 
+/**
+ * @brief Checks if large strings is enabled
+ *
+ * This checks the setting in the environment variable LIBCUDF_LARGE_STRINGS_ENABLED.
+ *
+ * @return true if large strings are supported
+ */
+bool is_large_strings_enabled();
+
 /**
  * @brief Return a normalized offset value from a strings offsets column
  *

cpp/include/cudf_test/column_utilities.hpp

@@ -210,6 +210,29 @@ template <>
 std::pair<thrust::host_vector<std::string>, std::vector<bitmask_type>> to_host(column_view c);
 //! @endcond
 
+/**
+ * @brief For enabling large strings testing in specific tests
+ */
+struct large_strings_enabler {
+  /**
+   * @brief Create large strings enable object
+   *
+   * @param default_enable Default enables large strings support
+   */
+  large_strings_enabler(bool default_enable = true);
+  ~large_strings_enabler();
+
+  /**
+   * @brief Enable large strings support
+   */
+  void enable();
+
+  /**
+   * @brief Disable large strings support
+   */
+  void disable();
+};
+
 }  // namespace cudf::test
 
 // Macros for showing line of failure.
@@ -242,3 +265,5 @@ std::pair<thrust::host_vector<std::string>, std::vector<bitmask_type>> to_host(c
     SCOPED_TRACE(" <--  line of failure\n");                        \
     cudf::test::detail::expect_equal_buffers(lhs, rhs, size_bytes); \
   } while (0)
+
+#define CUDF_TEST_ENABLE_LARGE_STRINGS() cudf::test::large_strings_enabler ls___

cpp/src/strings/copying/concatenate.cu

@@ -220,9 +220,6 @@ std::unique_ptr<column> concatenate(host_span<column_view const> columns,
   CUDF_EXPECTS(offsets_count <= static_cast<std::size_t>(std::numeric_limits<size_type>::max()),
                "total number of strings exceeds the column size limit",
                std::overflow_error);
-  CUDF_EXPECTS(total_bytes <= static_cast<std::size_t>(std::numeric_limits<size_type>::max()),
-               "total size of strings exceeds the column size limit",
-               std::overflow_error);
 
   bool const has_nulls =
     std::any_of(columns.begin(), columns.end(), [](auto const& col) { return col.has_nulls(); });
@@ -232,8 +229,7 @@ std::unique_ptr<column> concatenate(host_span<column_view const> columns,
   auto d_new_chars = output_chars.data();
 
   // create output offsets column
-  auto offsets_column = make_numeric_column(
-    data_type{type_id::INT32}, offsets_count, mask_state::UNALLOCATED, stream, mr);
+  auto offsets_column = create_offsets_child_column(total_bytes, offsets_count, stream, mr);
   auto itr_new_offsets =
     cudf::detail::offsetalator_factory::make_output_iterator(offsets_column->mutable_view());
 

cpp/src/strings/utilities.cu

@@ -22,6 +22,7 @@
 #include <cudf/detail/get_value.cuh>
 #include <cudf/strings/detail/char_tables.hpp>
 #include <cudf/strings/detail/utilities.cuh>
+#include <cudf/strings/detail/utilities.hpp>
 #include <cudf/utilities/error.hpp>
 
 #include <rmm/cuda_stream_view.hpp>
@@ -31,6 +32,9 @@
 #include <thrust/iterator/counting_iterator.h>
 #include <thrust/transform.h>
 
+#include <cstdlib>
+#include <string>
+
 namespace cudf {
 namespace strings {
 namespace detail {
@@ -65,6 +69,27 @@ rmm::device_uvector<string_view> create_string_vector_from_column(
   return strings_vector;
 }
 
+/**
+ * @copydoc cudf::strings::detail::create_offsets_child_column
+ */
+std::unique_ptr<column> create_offsets_child_column(int64_t chars_bytes,
+                                                    size_type count,
+                                                    rmm::cuda_stream_view stream,
+                                                    rmm::mr::device_memory_resource* mr)
+{
+  auto const threshold = get_offset64_threshold();
+  if (!is_large_strings_enabled()) {
+    CUDF_EXPECTS(
+      chars_bytes < threshold, "Size of output exceeds the column size limit", std::overflow_error);
+  }
+  return make_numeric_column(
+    chars_bytes < threshold ? data_type{type_id::INT32} : data_type{type_id::INT64},
+    count,
+    mask_state::UNALLOCATED,
+    stream,
+    mr);
+}
+
 namespace {
 // The device variables are created here to avoid using a singleton that may cause issues
 // with RMM initialize/finalize. See PR #3159 for details on this approach.
@@ -123,13 +148,19 @@ special_case_mapping const* get_special_case_mapping_table()
 
 int64_t get_offset64_threshold()
 {
-  auto const threshold  = std::getenv("LIBCUDF_LARGE_STRINGS_THRESHOLD");
-  std::size_t const rtn = threshold != nullptr ? std::atol(threshold) : 0;
+  auto const threshold = std::getenv("LIBCUDF_LARGE_STRINGS_THRESHOLD");
+  int64_t const rtn    = threshold != nullptr ? std::atol(threshold) : 0L;
   return (rtn > 0 && rtn < std::numeric_limits<int32_t>::max())
            ? rtn
            : std::numeric_limits<int32_t>::max();
 }
 
+bool is_large_strings_enabled()
+{
+  auto const env = std::getenv("LIBCUDF_LARGE_STRINGS_ENABLED");
+  return env != nullptr && std::string(env) == "1";
+}
+
 int64_t get_offset_value(cudf::column_view const& offsets,
                          size_type index,
                          rmm::cuda_stream_view stream)

cpp/tests/copying/concatenate_tests.cpp

@@ -32,6 +32,8 @@
 #include <cudf/table/table.hpp>
 #include <cudf/utilities/default_stream.hpp>
 
+#include <thrust/iterator/constant_iterator.h>
+
 #include <numeric>
 #include <stdexcept>
 #include <string>
@@ -164,37 +166,6 @@ TEST_F(StringColumnTest, ConcatenateColumnView)
   CUDF_TEST_EXPECT_COLUMNS_EQUAL(*results, expected);
 }
 
-TEST_F(StringColumnTest, ConcatenateColumnViewLarge)
-{
-  // Test large concatenate, causes out of bound device memory errors if kernel
-  // indexing is not int64_t.
-  // 1.5GB bytes, 5k columns
-  constexpr size_t num_strings        = 10000;
-  constexpr size_t string_length      = 150000;
-  constexpr size_t strings_per_column = 2;
-  constexpr size_t num_columns        = num_strings / strings_per_column;
-
-  std::vector<std::string> strings;
-  std::vector<char const*> h_strings;
-  std::vector<cudf::test::strings_column_wrapper> strings_column_wrappers;
-  std::vector<cudf::column_view> strings_columns;
-
-  std::string s(string_length, 'a');
-  for (size_t i = 0; i < num_strings; ++i)
-    h_strings.push_back(s.data());
-
-  for (size_t i = 0; i < num_columns; ++i)
-    strings_column_wrappers.push_back(cudf::test::strings_column_wrapper(
-      h_strings.data() + i * strings_per_column, h_strings.data() + (i + 1) * strings_per_column));
-  for (auto& wrapper : strings_column_wrappers)
-    strings_columns.push_back(wrapper);
-
-  auto results = cudf::concatenate(strings_columns);
-
-  cudf::test::strings_column_wrapper expected(h_strings.begin(), h_strings.end());
-  CUDF_TEST_EXPECT_COLUMNS_EQUAL(*results, expected);
-}
-
 TEST_F(StringColumnTest, ConcatenateManyColumns)
 {
   std::vector<char const*> h_strings{
@@ -226,6 +197,49 @@ TEST_F(StringColumnTest, ConcatenateTooLarge)
   EXPECT_THROW(cudf::concatenate(input_cols), std::overflow_error);
 }
 
+TEST_F(StringColumnTest, ConcatenateLargeStrings)
+{
+  CUDF_TEST_ENABLE_LARGE_STRINGS();
+  auto itr = thrust::constant_iterator<std::string_view>(
+    "abcdefghijklmnopqrstuvwxyABCDEFGHIJKLMNOPQRSTUVWXY");                // 50 bytes
+  auto input = cudf::test::strings_column_wrapper(itr, itr + 5'000'000);  // 250MB
+  auto view  = cudf::column_view(input);
+  std::vector<cudf::column_view> input_cols;
+  std::vector<cudf::size_type> splits;
+  int const multiplier = 10;
+  for (int i = 0; i < multiplier; ++i) {  // 2500MB > 2GB
+    input_cols.push_back(view);
+    splits.push_back(view.size() * (i + 1));
+  }
+  splits.pop_back();  // remove last entry
+  auto result = cudf::concatenate(input_cols);
+  auto sv     = cudf::strings_column_view(result->view());
+  EXPECT_EQ(sv.size(), view.size() * multiplier);
+  EXPECT_EQ(sv.offsets().type(), cudf::data_type{cudf::type_id::INT64});
+
+  // verify results in sections
+  auto sliced = cudf::split(result->view(), splits);
+  for (auto c : sliced) {
+    CUDF_TEST_EXPECT_COLUMNS_EQUIVALENT(c, input);
+  }
+
+  // also test with large strings column as input
+  {
+    input_cols.clear();
+    input_cols.push_back(input);           // regular column
+    input_cols.push_back(result->view());  // large column
+    result = cudf::concatenate(input_cols);
+    sv     = cudf::strings_column_view(result->view());
+    EXPECT_EQ(sv.size(), view.size() * (multiplier + 1));
+    EXPECT_EQ(sv.offsets().type(), cudf::data_type{cudf::type_id::INT64});
+    splits.push_back(view.size() * multiplier);
+    sliced = cudf::split(result->view(), splits);
+    for (auto c : sliced) {
+      CUDF_TEST_EXPECT_COLUMNS_EQUIVALENT(c, input);
+    }
+  }
+}
+
 struct TableTest : public cudf::test::BaseFixture {};
 
 TEST_F(TableTest, ConcatenateTables)

cpp/tests/utilities/column_utilities.cu

@@ -1011,5 +1011,16 @@ std::pair<thrust::host_vector<std::string>, std::vector<bitmask_type>> to_host(c
   return {std::move(host_data), bitmask_to_host(c)};
 }
 
+large_strings_enabler::large_strings_enabler(bool default_enable)
+{
+  default_enable ? enable() : disable();
+}
+
+large_strings_enabler::~large_strings_enabler() { disable(); }
+
+void large_strings_enabler::enable() { setenv("LIBCUDF_LARGE_STRINGS_ENABLED", "1", 1); }
+
+void large_strings_enabler::disable() { setenv("LIBCUDF_LARGE_STRINGS_ENABLED", "0", 1); }
+
 }  // namespace test
 }  // namespace cudf