szcompressor / vecSZ

An efficient error-bounded lossy compressor with SIMD vectorization support.

vecSZ

vecSZ is an implementation of the popular SZ lossy compressor optimized for efficient CPU performance using vectorization.

Use

Compilation

To compile vecSZ with default optimizations (shown below), just type make.

Default Optimizations

All options are shown with their default value.

Enable vector support   -  vector_support=AVX2
Enable -O3 Optimization -  optimize=True
Enable OpenMP           -  omp=False
Use -g for debugging    -  debug=False

To enable an alternate level of vector support, use vector_support=<None|AVX|AVX2|AVX512>.

For example, enable AVX-512 and OpenMP by typing make vector_support=AVX512 omp=True.

Basic Use

Type vecsz or vecsz -h for instant instructions. Also, a basic use vecSZ is given below.

./bin/vecsz -t f32 -m abs -e 1.0e-4.0 -i ./data/ex-cesm-CLDHGH -l 3600,1800 -z
            ------ ------ ----------- ------------------------ ------------  |
             dtype  mode  error bound      input datum file    low-to-high  zip

./bin/vecsz -i ./data/ex-cesm-CLDHGH.sz -x
            ---------------------------  |
                      sz archive        unzip

In this case, 1800-by-3600 (y-x order) CESM-ATM is listed in the demo set. Use -D cesm to specify the preset for demonstration. Type vecsz or vecsz -h to look up the presets.

vecsz -m abs -e 1e-4 -i ./data/ex-cesm-CLDHGH -D cesm -z

The following essential arguments are required,

• WORKFLOW: -z to compress; -x to decompress.
• CONFIG: -m to specify error control mode from abs (absolute) and r2r (relative to value range)
• CONFIG: -e to specify error bound
• CONFIG: -t to specify datatype from f32 (single-precision float) and f64 (double-precision float)
• INPUT: -i to specify input file
• INPUT: -D to specify demo dataset name or -l to dimensions
• OUTPUT: -o or --output to specify output path for both compression and decompression
• LOG: -V or --verbose to print CPU information

Limitations

• Current support for only x86 architectures (AVX/SSE support)
• Support for single-precision float data only (f32)