hipCOMP-CORE

Caution

This release is an early-access software technology preview. Running production workloads is not recommended.

hipCOMP CORE is a library for fast lossless compression/decompression on AMD GPUs.

The code is based on nvCOMP release branch branch-2.2. GPU implementations of algorithms like DEFLATE, GDEFLATE, ANS, and zSTD that have been later introduced as part of nvCOMP's proprietary releases are not part of this repo (yet).

Build From Source

HIP/AMD

Note

If you experience compiler errors related to cooperative_groups with ROCm versions <=6.0.X, additionally specify the CMake build option -D CG_WORKAROUND=1.

cd hipcomp-core/
mkdir build/
cd build/
CMAKE_PREFIX_PATH=/opt/rocm/lib/cmake cmake ../
# To build with tests, append `-D BUILD_TESTS=1`:
# CMAKE_PREFIX_PATH=/opt/rocm/lib/cmake cmake ../ -D BUILD_TESTS=1
make

HIP/CUDA

Like HIP/AMD but with additional -D CUDA_BACKEND=1 option:

cd hipcomp-core/
mkdir build/
cd build/
CMAKE_PREFIX_PATH=/opt/rocm/lib/cmake cmake ../ -D CUDA_BACKEND=1
# To build with tests, append `-D BUILD_TESTS=1`:
# CMAKE_PREFIX_PATH=/opt/rocm/lib/cmake cmake ../ -D BUILD_TESTS=1 -D CUDA_BACKEND=1
make

Debugging

To create debug builds append the following option:

-D CMAKE_BUILD_TYPE=Debug

Run tests

After completing the build, run:

cd hipcomp-core/
cd build/
make test

Tips: