libcudf C++ examples for Parquet I/O

This C++ example demonstrates using libcudf APIs to read and write Parquet files with different encodings and compression types.

Blog post that uses this code: https://developer.nvidia.com/blog/encoding-and-compression-guide-for-parquet-string-data-using-rapids/

The following encoding and compression types are demonstrated:

Encoding types: DEFAULT, DICTIONARY, PLAIN, DELTA_BINARY_PACKED, DELTA_LENGTH_BYTE_ARRAY, DELTA_BYTE_ARRAY
Compression types: NONE, AUTO, SNAPPY, LZ4, ZSTD

There are two examples included:

parquet_io.cpp Reads an input Parquet file, writes it back out using the specified encoding and compression (optionally with page statistics), reads the transcoded file, and validates that the data round-trips correctly. The write and read steps are timed.
parquet_io_multithreaded.cpp Reads one or more Parquet files (or directories of files) using multiple threads and a configurable I/O source type (FILEPATH, HOST_BUFFER, PINNED_BUFFER, DEVICE_BUFFER), optionally writing and validating the output.

Compile and execute

bash

# Configure project
cmake -S . -B build/
# Build
cmake --build build/ --parallel $PARALLEL_LEVEL
# Execute using the included example.parquet and default encoding/compression
build/parquet_io
# Execute with explicit arguments:
#   <input file> <output file> <encoding type> <compression type> <write page stats: yes/no>
build/parquet_io example.parquet output.parquet DELTA_BINARY_PACKED ZSTD
# Execute the multithreaded example:
#   <comma delimited list of dirs and/or files> <input multiplier> <io source type>
#   <number of times to read> <thread count> <write to temp output files and validate: yes/no>
build/parquet_io_multithreaded example.parquet

Pass -h or --help to either executable to print full usage information.

If your machine does not come with a pre-built libcudf binary, expect the first build to take some time, as it would build libcudf on the host machine. It may be sped up by configuring the proper PARALLEL_LEVEL number.