Fast Compression of Floating-Point Values with Exponent/Mantissa Shuffling
Abstract
We investigate fast compression algorithms for vectors of floating-point values, as these frequently appear in numerical computations. Our proposal bases on Snappy yet uses a modified data layout. Use of data chunks allows parallel thread-based compression without loss of compression quality. Our proposal beats previous proposals and standard compression algorithms in compression ratio and/or speed. The gain in compression ratio and speed can be used to shorten communication phases, e.g., between CPUs and accelerators.
Type
Publication
PARS-Mitteilungen