GPU Accelerated Indexing for High Order Tensors in Google Colab

摘要

Processing large volume of multi-dimensional data or high order tensors can be challenging in terms of limitless execution time. We can improve the processing time with the parallel computation on a suitable indexing scheme. There are many software and hardware implementations available to apply the parallelism. We use Graphics Processing Unit (GPU) as the parallel device considering the idea of General-Purpose GPU (GPGPU) computing. In this paper, we implement an Index Partitioning Algorithm (IPA) to speed-up the tensor indexing where the tensor indices are divided into smaller segments and executed concurrently. This algorithm ensures the load balancing by the equal distribution of the segments on GPU threads. We also embed a Scalable Tensor Structure (STS) to transform the n-dimensional index into a 2- dimensional index. This structure gives better performance by the efficient memory utilization and lesser index computation. We deploy our algorithms in the “Google Colab” platform for experimental purposes. Because of the current increasing rate of data creation, our algorithm is expected to contribute heavily in many research domains for large scale data processing.