nvCOMP is a CUDA library that features generic compression interfaces to enable developers to use high-performance GPU compressors and decompressors in their applications.
Example benchmarking results and a brief description of each algorithm are available on the nvCOMP Developer Page.
From version 2.3 onwards, the compression / decompression source code will not be released.
- Cascaded, GDeflate, zStandard, Deflate, Gzip and Bitcomp decompressors can only operate on valid input data (data that was compressed using the same compressor). Other decompressors can sometimes detect errors in the compressed stream.
- Cascaded, zStandard and Bitcomp batched decompression C APIs cannot currently accept nullptr for actual_decompressed_bytes or device_statuses values. Deflate and Gzip cannot accept nullptr for device_statuses values.
- The Bitcomp low-level batched decompression function is not fully asynchronous.
- The high-level interface is not available for Deflate, zStandard or Gzip.
- Gzip low-level interface only provides decompression.
- You can download the appropriate built binary packages from the nvCOMP Developer Page
- Choose the package that corresponds to your CUDA toolkit version, operating system, and arch
- For example, on linux, the package includes
include/
nvcomp/ #nvcomp API headers
gdeflate/ #Gdeflate CPU library headers
lib/
libnvcomp.so
<Other nvcomp libraries that are used internally by nvcomp's APIs>
libnvcomp_gdeflate_cpu.so # CPU library for gdeflate
cmake/ <Package files to allow use through cmake>
bin/
<benchmark scripts>
- Pascal (sm60) or higher GPU architecture is required. Volta (sm70)+ GPU architecture is recommended for best results.
- To compile using nvCOMP as a dependency, you need a compiler with C++ 11 support (e.g. GCC 4.4, Clang 3.3, MSVC 2017).
- To use the packages provided for Linux, you'll need to compile translation units that will be linked against nvcomp using the old ABI, i.e. for GCC
-D_GLIBCXX_USE_CXX11_ABI=0
- To use the packages provided for Linux, you'll need to compile translation units that will be linked against nvcomp using the old ABI, i.e. for GCC
Please view the following guides for information on how to use the two APIs provided by the library. Each of these guides links to a compilable example for further reference.
GPU Benchmark source are included in the binary releases. Source code for the benchmarks is also provided here on Github to provide additional examples on how to use nvCOMP. For further information on how to execute the benchmarks, please view Benchmarks Page
We provide some examples of how you might use CPU compression and GPU decompression or vice versa for LZ4 GDeflate and Deflate. These require some external dependencies, namely:
- zlib for the GDeflate and Deflate CPU compression/decompression example (
zlib1g-devon debian based systems) - LZ4 for the LZ4 CPU compression example (
liblz4-devon debian based systems) - libdeflate for the Deflate CPU compression/decompression example (
libdeflate-devon debian based systems)
The CPU example executables are:
gdeflate_cpu_compression {-f <input_file>}
lz4_cpu_compression {-f <input_file>}
lz4_cpu_decompression {-f <input_file>}
deflate_cpu_compression {-a <0 libdeflate, 1 zlib_compress2, 2 zlib_deflate> -f <input_file>}
deflate_cpu_decompression {-a <0 libdeflate, 1 zlib_inflate> -f <input_file>}
gzip_gpu_decompression {-f <input_file>}
To build only the examples, you'll need cmake >= 3.18 and an nvcomp artifact. Then, you can follow the following steps from the top-level of your clone of nvCOMP from Github
cmake .. -DCMAKE_PREFIX_PATH=<path_to_nvcomp_install>
The path_to_nvcomp_install is the directory where you extracted the nvcomp artifact.
To compile the benchmarks too, you can add -DBUILD_BENCHMARKS=1, but note this is only provided for an additional example of building against the artifacts. The benchmarks are already provided in the artifact bin/ folder.