Mixed-Tool Performance Analysis on Hybrid Multicore Architectures

Abstract

Abstract-This paper proposes a triangular solve algorithm with variable block size for graphics processing unit (GPU). By using diagonal blocks inversion with recursion, this algorithm works with tunable block size to achieve the best performance. Various methods are shown on how to make use of existing profiling tools to successfully measure and analyze performance of this algorithm. We use some of the most popular CPU and GPU profiling tools for their advantages and overcome their disadvantages with several new techniques to analyze the performance and relationship of different components of applications. With the presented methodologies, insight information is produced which helps to understand and tune the proposed algorithm and considerably improve the performance of the solver itself as well as the application using it.

From CUDA to OpenCL: Towards a performance-portable solution for multi-platform GPU programming

“…The performance of TRSM is dominated by GEMM [26]. Since GEMM is one of the most important kernels in linear algebra, we will focus on implementing and analyzing a fast OpenCL GEMM in the coming sections.…”