Implementation of an associative-computing model on GPU: a basic procedure library of the STAR language

The associative (content addressable) parallel processors of the SIMD type with vertical data processing are oriented on solving problems of non-numeric data processing. The simulation of such systems is described using an abstract SIMD-type model of a STAR machine. On the basis of this model, a number of efficient algorithms are developed to solve many graph problems. Since the associative architectures are not widely available, however, these algorithms cannot be used in practice. With advances in the production of GPU, the possibilities to implement the associative parallel models without significant loss of efficiency are increased. As the first stage in the implementation of the STAR-machine on GPU in the form of a CUDA library, specific data types and simple operations of the STAR language were developed. In this paper, we consider an efficient GPU implementation of the standard associative procedure library. The runtime of this implementation is compared with the runtime of similar procedures in the standard libraries (STL on CPU and CUDA thrust on GPU). We plan to use our library implementation to solve graph problems.