Research Paper Snapshot
cuSZ-Hi

Installation

Run the following command to compile and execute cuSZ-Hi

git clone https://github.com/shixun404/cuSZ-Hi.git
cmake -S cuSZ-Hi -B cuSZ-Hi/build \  
    -D PSZ_BACKEND=cuda \
    -D PSZ_BUILD_EXAMPLES=off \
    -D CMAKE_CUDA_ARCHITECTURES="80" \
    -D CMAKE_BUILD_TYPE=Release
cmake --build cuSZ-Hi/build -- -j
cd cuSZ-Hi/build

Notes

Please use the corresponding CMAKE_CUDA_ARCHITECTURES for your hardware. For example, NVIDIA A100 should be "80", NVIDIA RTX 4090/6000 Ada should be "80;89", and NVIDIA H100 should be "90".

Compression

./cuszhi --report time,cr  -z -t f32 -m r2r --dim3 [DimX]x[DimY]x[DimZ] -e [REL_ERROR_BOUND] --predictor spline3 -i [input.data] -s [cr/tp];

Notes

• input.data is the binary input file;
• -t: f32 (float32 data) or d64 (double precision data, under development);
• DimX is the fastest dimension, DimZ is the slowest;
• -s: cr (Huffman-integrated lossless pipeline, slower but higher compression ratio) or tp (Huffman-free lossless pipeline, lower compression ratio but fast);
• Add "-a rd-first" if you need better rate-distortion rather than a maximized compression ratio under a fixed error bound.

Decompression

./cuszhi --report time -x -i [input.data.cusza] --compare input.data