Refactor pinned memory vector and ORC+Parquet writers (rapidsai#13206)

Currently, `cudf::detail::pinned_allocator` is used in various places to implement a pinned host vector. This standardizes such usage, removing `cudf::detail::pinned_allocator` from the usage sites and replacing its usage by a standard `cudf::detail::pinned_host_vector` instead.

Some small changes are also made for ORC/Parquet writer classes, replacing `bool _single_write_mode` by `SingleWriteMode _single_write_mode`.

This is needed for rapidsai#13076.

Authors:
  - Nghia Truong (https://github.com/ttnghia)

Approvers:
  - AJ Schmidt (https://github.com/ajschmidt8)
  - Vukasin Milovanovic (https://github.com/vuule)
  - Yunsong Wang (https://github.com/PointKernel)

URL: rapidsai#13206

conda/recipes/libcudf/meta.yaml

@@ -139,7 +139,7 @@ outputs:
         - test -f $PREFIX/include/cudf/detail/utilities/integer_utils.hpp
         - test -f $PREFIX/include/cudf/detail/utilities/linked_column.hpp
         - test -f $PREFIX/include/cudf/detail/utilities/logger.hpp
-        - test -f $PREFIX/include/cudf/detail/utilities/pinned_allocator.hpp
+        - test -f $PREFIX/include/cudf/detail/utilities/pinned_host_vector.hpp
         - test -f $PREFIX/include/cudf/detail/utilities/vector_factories.hpp
         - test -f $PREFIX/include/cudf/detail/utilities/visitor_overload.hpp
         - test -f $PREFIX/include/cudf/dictionary/detail/concatenate.hpp

cpp/benchmarks/io/text/multibyte_split.cpp

@@ -23,7 +23,7 @@
 #include <cudf_test/file_utilities.hpp>
 
 #include <cudf/column/column_factories.hpp>
-#include <cudf/detail/utilities/pinned_allocator.hpp>
+#include <cudf/detail/utilities/pinned_host_vector.hpp>
 #include <cudf/detail/utilities/vector_factories.hpp>
 #include <cudf/io/text/data_chunk_source_factories.hpp>
 #include <cudf/io/text/detail/bgzip_utils.hpp>
@@ -33,7 +33,6 @@
 #include <cudf/types.hpp>
 #include <cudf/utilities/default_stream.hpp>
 
-#include <thrust/host_vector.h>
 #include <thrust/transform.h>
 
 #include <nvbench/nvbench.cuh>
@@ -136,7 +135,7 @@ static void bench_multibyte_split(nvbench::state& state,
   std::unique_ptr<cudf::io::datasource> datasource;
   auto device_input      = create_random_input(file_size_approx, delim_factor, 0.05, delim);
   auto host_input        = std::vector<char>{};
-  auto host_pinned_input = thrust::host_vector<char, cudf::detail::pinned_allocator<char>>{};
+  auto host_pinned_input = cudf::detail::pinned_host_vector<char>{};
 
   if (source_type != data_chunk_source_type::device &&
       source_type != data_chunk_source_type::host_pinned) {

cpp/include/cudf/detail/utilities/pinned_allocator.hpp → cpp/include/cudf/detail/utilities/pinned_host_vector.hpp

@@ -1,5 +1,5 @@
 /*
- *  Copyright 2008-2022 NVIDIA Corporation
+ *  Copyright 2008-2023 NVIDIA Corporation
  *
  *  Licensed under the Apache License, Version 2.0 (the "License");
  *  you may not use this file except in compliance with the License.
@@ -22,6 +22,8 @@
 
 #include <cudf/utilities/error.hpp>
 
+#include <thrust/host_vector.h>
+
 namespace cudf::detail {
 
 /*! \p pinned_allocator is a CUDA-specific host memory allocator
@@ -199,4 +201,11 @@ class pinned_allocator {
     return !operator==(x);
   }
 };
+
+/**
+ * @brief A vector class with pinned host memory allocator
+ */
+template <typename T>
+using pinned_host_vector = thrust::host_vector<T, pinned_allocator<T>>;
+
 }  // namespace cudf::detail

cpp/include/cudf/io/detail/orc.hpp

@@ -96,7 +96,7 @@ class writer {
    */
   explicit writer(std::unique_ptr<cudf::io::data_sink> sink,
                   orc_writer_options const& options,
-                  SingleWriteMode mode,
+                  single_write_mode mode,
                   rmm::cuda_stream_view stream);
 
   /**
@@ -109,7 +109,7 @@ class writer {
    */
   explicit writer(std::unique_ptr<cudf::io::data_sink> sink,
                   chunked_orc_writer_options const& options,
-                  SingleWriteMode mode,
+                  single_write_mode mode,
                   rmm::cuda_stream_view stream);
 
   /**

cpp/include/cudf/io/detail/parquet.hpp

@@ -156,7 +156,7 @@ class writer {
    */
   explicit writer(std::vector<std::unique_ptr<data_sink>> sinks,
                   parquet_writer_options const& options,
-                  SingleWriteMode mode,
+                  single_write_mode mode,
                   rmm::cuda_stream_view stream);
 
   /**
@@ -171,7 +171,7 @@ class writer {
    */
   explicit writer(std::vector<std::unique_ptr<data_sink>> sinks,
                   chunked_parquet_writer_options const& options,
-                  SingleWriteMode mode,
+                  single_write_mode mode,
                   rmm::cuda_stream_view stream);
 
   /**

cpp/include/cudf/io/detail/utils.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021, NVIDIA CORPORATION.
+ * Copyright (c) 2021-2023, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -22,7 +22,7 @@ namespace detail {
 /**
  * @brief Whether writer writes in chunks or all at once
  */
-enum class SingleWriteMode : bool { YES, NO };
+enum class single_write_mode : bool { YES, NO };
 }  // namespace detail
 }  // namespace io
 }  // namespace cudf

cpp/src/io/functions.cpp

@@ -428,7 +428,7 @@ void write_orc(orc_writer_options const& options)
   CUDF_EXPECTS(sinks.size() == 1, "Multiple sinks not supported for ORC writing");
 
   auto writer = std::make_unique<detail_orc::writer>(
-    std::move(sinks[0]), options, io_detail::SingleWriteMode::YES, cudf::get_default_stream());
+    std::move(sinks[0]), options, io_detail::single_write_mode::YES, cudf::get_default_stream());
 
   writer->write(options.get_table());
 }
@@ -444,7 +444,7 @@ orc_chunked_writer::orc_chunked_writer(chunked_orc_writer_options const& options
   CUDF_EXPECTS(sinks.size() == 1, "Multiple sinks not supported for ORC writing");
 
   writer = std::make_unique<detail_orc::writer>(
-    std::move(sinks[0]), options, io_detail::SingleWriteMode::NO, cudf::get_default_stream());
+    std::move(sinks[0]), options, io_detail::single_write_mode::NO, cudf::get_default_stream());
 }
 
 /**
@@ -519,7 +519,7 @@ std::unique_ptr<std::vector<uint8_t>> write_parquet(parquet_writer_options const
 
   auto sinks  = make_datasinks(options.get_sink());
   auto writer = std::make_unique<detail_parquet::writer>(
-    std::move(sinks), options, io_detail::SingleWriteMode::YES, cudf::get_default_stream());
+    std::move(sinks), options, io_detail::single_write_mode::YES, cudf::get_default_stream());
 
   writer->write(options.get_table(), options.get_partitions());
 
@@ -575,7 +575,7 @@ parquet_chunked_writer::parquet_chunked_writer(chunked_parquet_writer_options co
   auto sinks = make_datasinks(options.get_sink());
 
   writer = std::make_unique<detail_parquet::writer>(
-    std::move(sinks), options, io_detail::SingleWriteMode::NO, cudf::get_default_stream());
+    std::move(sinks), options, io_detail::single_write_mode::NO, cudf::get_default_stream());
 }
 
 /**