libstreamvbyte 0.3.8

A C++ implementation of StreamVByte, with Python bindings.

Explore Usage »

Report Bug · Request Feature

Table of Contents

1. About The Project
2. Getting Started
   • Installation
   • Usage
   • Example
3. Benchmark
4. Roadmap
5. Contributing
6. License
7. Reference
8. Contact

About The Project

libstreamvbyte is a C++ implementation of StreamVByte, with Python bindings using pybind11.

StreamVByte is an integer compression technique that use SIMD instructions (vectorization) to improve performance. The library is optimized with SSSE3 intrinsics, which are supported by most x86_64 processors. It uses sse2neon to translate SSSE3 intrinsics to NEON intrinsics for ARM processors. The library can also be used with other 32-bit architectures, although it will fall back to scalar implementations in those cases.

With libstreamvbyte, you can quickly and efficiently compress integer sequences, reducing the amount of storage space and network bandwidth required. The library is easy to use and integrates seamlessly with Python via pybind11 bindings. Whether you're working with large datasets or building a distributed computing system, libstreamvbyte can help you improve performance and reduce the resources needed to handle your data.

Currently supports Python 3.8+ on Windows, Linux (manylinux_2_17, musllinux_1_1) and macOS (universal2).

(back to top)

Getting Started

Installation

For Python

Install from PyPI using pip.

pip install libstreamvbyte

Or install from .whl file.

pip install "path/to/your/downloaded/whl"

To find appropriate .whl file, please visit releases.

For C++

You must have CMake installed on your system.

# clone the repo
git clone https://github.com/wst24365888/libstreamvbyte
cd libstreamvbyte

# build and install
cmake .
make
sudo make install

Usage

For Python

Import libstreamvbyte first.

import libstreamvbyte as svb

And here are the APIs.

# Encode an array of unsigned integers into a byte array.
encode(in_uint32: numpy.ndarray[numpy.uint32]) -> numpy.ndarray[numpy.uint8]

# Decode a byte array into an array of unsigned integers.
decode(in_uint8: numpy.ndarray[numpy.uint8], size: int) -> numpy.ndarray[numpy.uint32]

# Encode an array of signed integers into an array of unsigned integers.
zigzag_encode(in_int32: numpy.ndarray[numpy.int32]) -> numpy.ndarray[numpy.uint32]

# Decode an array of unsigned integers into an array of signed integers.
zigzag_decode(in_uint32: numpy.ndarray[numpy.uint32]) -> numpy.ndarray[numpy.int32]

# Check if the current wheel is a vectorized version.
is_vectorized_version() -> bool

For C++

Include streamvbyte.h first.

#include "streamvbyte.h"

For the APIs, please refer to include/streamvbyte.h.

Example

For Python

import libstreamvbyte as svb

N = 2**20 + 2

# type(original_data) == np.ndarray
# original_data.dtype == np.int32
original_data = np.random.randint(-2**31, 2**31, N, dtype=np.int32)

# type(compressed_bytes) == np.ndarray
# compressed_bytes.dtype == np.uint8
compressed_bytes = svb.encode(svb.zigzag_encode(original_data))

# type(recovered_data) == np.ndarray
# recovered_data.dtype == np.int32
recovered_data = svb.zigzag_decode(svb.decode(compressed_bytes, N))

For C++

#include "streamvbyte.h"

int main() {
    std::size_t N = (1 << 20) + 2;

    std::vector<int32_t> original_data(N);
    for (std::size_t i = 0; i < N; ++i) {
        original_data[i] = rand() - rand();
    }

    std::vector<uint8_t> compressed_bytes = streamvbyte::encode(streamvbyte::zigzag_encode(original_data));
    std::vector<int32_t> recovered_data = streamvbyte::zigzag_decode(streamvbyte::decode(compressed_bytes, N));

    return 0;
}

Compile it with linking to libstreamvbyte.

g++ -o example example.cpp -lstreamvbyte

(back to top)

Benchmark

OS: Linux 5.15.79.1-microsoft-standard-WSL2 x86_64
CPU: AMD Ryzen 5 3600 6-Core Processor (12) @ 3.600GHz

Run on (12 X 3593.26 MHz CPU s)
CPU Caches:
  L1 Data 32 KiB (x6)
  L1 Instruction 32 KiB (x6)
  L2 Unified 512 KiB (x6)
  L3 Unified 16384 KiB (x1)
Load Average: 0.81, 0.85, 0.69
-----------------------------------------------------------------------------------
Benchmark                              Time             CPU   Iterations Throughput
-----------------------------------------------------------------------------------
BM_memcpy/4096                       149 ns          149 ns      4688531 13.7122G/s
BM_memcpy/8192                       548 ns          548 ns      1275803 7.46783G/s
BM_memcpy/16384                     1139 ns         1138 ns       640835 7.19553G/s
BM_memcpy/32768                     2185 ns         2185 ns       320840 7.49932G/s
BM_memcpy/65536                     4921 ns         4921 ns       142703 6.65895G/s
BM_memcpy/131072                   10968 ns        10968 ns        63502 5.97511G/s
BM_memcpy/262144                   22465 ns        22465 ns        31134 5.83457G/s
BM_memcpy/524288                   45101 ns        45100 ns        15541 5.81245G/s
BM_memcpy/1048576                  91131 ns        91131 ns         7639 5.75314G/s
BM_streamvbyte_encode/4096          1222 ns         1222 ns       580855 1.67556G/s
BM_streamvbyte_encode/8192          2470 ns         2467 ns       282349 1.66064G/s
BM_streamvbyte_encode/16384         4945 ns         4945 ns       139671 1.65662G/s
BM_streamvbyte_encode/32768         9990 ns         9989 ns        70497 1.64017G/s
BM_streamvbyte_encode/65536        19853 ns        19853 ns        30963 1.65051G/s
BM_streamvbyte_encode/131072       39933 ns        39932 ns        17401 1.64118G/s
BM_streamvbyte_encode/262144       80563 ns        80562 ns         8193 1.62697G/s
BM_streamvbyte_encode/524288      160716 ns       160716 ns         4284  1.6311G/s
BM_streamvbyte_encode/1048576     319253 ns       319253 ns         1942 1.64223G/s
BM_streamvbyte_decode/4096           691 ns          691 ns      1040462 2.96191G/s
BM_streamvbyte_decode/8192          1341 ns         1341 ns       516979 3.05539G/s
BM_streamvbyte_decode/16384         2683 ns         2683 ns       261208 3.05359G/s
BM_streamvbyte_decode/32768         5348 ns         5348 ns       130319 3.06353G/s
BM_streamvbyte_decode/65536        10817 ns        10817 ns        64427 3.02936G/s
BM_streamvbyte_decode/131072       23207 ns        23207 ns        31546   2.824G/s
BM_streamvbyte_decode/262144       45746 ns        45746 ns        11291 2.86519G/s
BM_streamvbyte_decode/524288       88660 ns        88660 ns         7947 2.95673G/s
BM_streamvbyte_decode/1048576     178497 ns       178497 ns         3907 2.93724G/s
BM_zigzag_encode/4096                810 ns          810 ns       854076 2.52829G/s
BM_zigzag_encode/8192               1611 ns         1608 ns       433154   2.548G/s
BM_zigzag_encode/16384              3174 ns         3174 ns       219165 2.58084G/s
BM_zigzag_encode/32768              6457 ns         6457 ns       108415 2.53754G/s
BM_zigzag_encode/65536             12582 ns        12582 ns        54747 2.60432G/s
BM_zigzag_encode/131072            25243 ns        25243 ns        27802 2.59617G/s
BM_zigzag_encode/262144            50278 ns        50278 ns        13952 2.60693G/s
BM_zigzag_encode/524288           100563 ns       100562 ns         6932 2.60678G/s
BM_zigzag_encode/1048576          211846 ns       211845 ns         3222 2.47487G/s
BM_zigzag_decode/4096                675 ns          675 ns      1041044 3.03263G/s
BM_zigzag_decode/8192               1342 ns         1342 ns       523553 3.05196G/s
BM_zigzag_decode/16384              2643 ns         2643 ns       265497 3.09905G/s
BM_zigzag_decode/32768              5383 ns         5383 ns       130976 3.04377G/s
BM_zigzag_decode/65536             11474 ns        11474 ns        60817 2.85588G/s
BM_zigzag_decode/131072            21777 ns        21777 ns        32345 3.00944G/s
BM_zigzag_decode/262144            43477 ns        43478 ns        14387  3.0147G/s
BM_zigzag_decode/524288            86120 ns        86120 ns         8145 3.04393G/s
BM_zigzag_decode/1048576          173095 ns       173093 ns         4028 3.02894G/s

The unit of Throughput is GB/s (Giga Bytes per second).

Build Benchmarks from Source

cmake . \
    -DCMAKE_BUILD_TYPE=Release \
    -DBUILD_SHARED_LIBS=OFF \
    -DBUILD_PYBIND11=OFF \
    -DPRINT_BENCHMARK=OFF \
    -DBUILD_TESTS=ON \
    -DBUILD_BENCHMARKS=ON
make libstreamvbyte_benchmarks
./libstreamvbyte_benchmarks --benchmark_counters_tabular=true

(back to top)

Roadmap

• Zigzag encoding/decoding.
• Support ARM processors with NEON intrinsics.
• Differential coding (delta encoding/decoding).

See the open issues for a full list of proposed features (and known issues).

(back to top)

Contributing

Contributions are what make the open source community such an amazing place to learn, inspire, and create. Any contributions you make are greatly appreciated.

If you have a suggestion that would make this better, please fork the repo and create a pull request. You can also simply open an issue with the tag "enhancement". Don't forget to give the project a star! Thanks again!

1. Fork the Project
2. Create your Feature Branch (git checkout -b feat/amazing-feature)
3. Commit your Changes with Conventional Commits
4. Push to the Branch (git push origin feat/amazing-feature)
5. Open a Pull Request

(back to top)

License

Distributed under the MIT License. See LICENSE for more information.

(back to top)

Reference

• Daniel Lemire, Nathan Kurz, Christoph Rupp, Stream VByte: Faster Byte-Oriented Integer Compression, Information Processing Letters 130, 2018.

(back to top)

Contact

Author

• HSING-HAN, WU (Xyphuz)
  • Mail me: [e-mail]
  • About me: https://www.xyphuz.com
  • GitHub: https://github.com/wst24365888

Project Link

• https://github.com/wst24365888/libstreamvbyte

(back to top)