high performance computing on graphics processing units: hgpu.org

In this paper, we propose a scheme to accelerate the Probabilistic Latent Semantic Indexing (PLSI), which is an automated document indexing method based on a statistical latent semantic model, exploiting the high parallelism of Graphics Processing Unit (GPU). Our proposal is composed of three techniques: the first one is to accelerate the Expectation-Maximization (EM) computation by applying GPU matrix-vector multiplication; the second one uses the same principles as the first method, but deals with the sparseness of co-occurrence of words and documents; and the third one is to use the concurrent kernel execution, which is available on NVIDIA Fermi architecture, in order to speed up the process. We compare the performance of the proposed scheme with the non-parallelized implementation. The results show that our method could be more than 100 times faster than the CPU-based implementation in our environment. By dealing with the sparseness of the data, we could not only process more documents and words using GPU, but we could also keep more data on the device memory so that we can avoid massive data copy transfer between the host and the device susceptible to reduce the execution performance.