BitstreamZoo: A Benchmark Suite for Bitstream Workloads on Commodity CPUs and GPUs
Abstract
Bitwise-parallel workloads pack many independent data items into bitstreams, so a bitstream operation describes work over all stream positions at once. This execution style appears in application areas as varied as cryptography, text processing, and neural network inference, yet these areas have largely been studied using separate, domain-specific tools. Existing CPU and GPU benchmark suites do not represent packed bitstream computations in a common form. To address this gap, we present BitstreamZoo, an open benchmark suite that expresses 15 bit-level parallel workload families in a shared bitstream program format. We also build reference CPU and GPU backends that interpret this format directly. We characterize the programs by their operation mix, dependency structure, and memory behavior. The workloads span orders of magnitude in dependency depth and in the number of operations available at each dependency level. We also find that executing one bitstream operation at a time as the reference backends do materializes every intermediate stream and can leave substantial inter-operation parallelism unexploited. BitstreamZoo gives compiler and architecture researchers a reproducible workload set for bit-level parallel programs on CPUs and GPUs.