Cross-platform and polyhedral programming for Nussinov RNA folding

Abstract

This article addresses the use of codes from polyhedral compilers with tiled and parallel code designed for CPU processors, automatically generated as source-to-source OpenMP for NVIDIA GPU graphics cards using CUDA. In previous publications, we demonstrated that it is possible to use large language models (LLM) to translate code, generate kernels, and correctly manage memory transfers between the host and the device without manual effort. Unfortunately, when the target architecture is not taken into account in detail, the performance of code designed for CPUs leaves much to be desired when running on GPUs. The architectural differences between these two platforms like cores, cache, and the dimensionality of computations require careful attention to performance portability. In this article, we address the Nussinov algorithm, a popular benchmark in bioinformatics, to achieve higher performance on the NVIDIA platform than automatically generated codes by LLM. Nussinov’s loop nests are a non-trivial kernel from the non-serial polyadic dynamic programming (NPDP) benchmark with non-uniform loops. We will utilize a polyhedral code framework that tiles and then manually modifies the most nested loop nest containing the majority of the computations, using the two-dimensional thread blocks. To accelerate the computations, shared memory within blocks is utilized. The resulting codes were tested on two modern NVIDIA devices for various RNA sequence lengths, compared to parallel and tiled CPU codes, and previously generated Nussinov’s GPU codes using LLMs. The correctness of these codes and their scalability were analyzed. Comparison to related approaches and future work are outlined.