Seyrek matris-vektör çarpımları için performans iyileştirmeleri

Sparse matrix-vector multiplication (spMV) is a kernel operation in scientific com- putation. There exist problems where a matrix is repeatedly multiplied by many different vectors. For such problems, specializing the spMV code based on the matrix has the potential of producing significantly faster code. This, in fact, has been one of the motivational examples of program generation. Using program generation, spMV code can be unfolded fully to eliminate loop overheads as well as enable high-impact optimizations. In this work we focus on specialization of spMV by unfolding the code according to a given matrix. We provide an experimental evaluation of performance using 70 sparse matrices collected from real-world scientific computation domains. We present optimizations with which high-performant assembly code can be generated rapidly without having to generate source-level code and go through all the phases of a general-purpose compiler. We finally present how one of the optimizations we studied can be implemented as a code-transforming pass.