Seyrek matrıs-vektör çarpımının koşut zamanda özelleştirilmesı ve otomatık ayarlanması

Runtime specialization is used for optimizing programs based on partial information available only at runtime. In this thesis, we present a purpose-built compiler to quickly specialize Sparse Matrix-Vector Multiplication code for a particular matrix at runtime. There are several specialization methods and the best one depends both on the matrix and the platform. To avoid having to generate all the specialization variations, we use an autotuning approach to predict the best specializer for a given matrix. To this end, we define a set of matrix features for autotuning. Several of these features are unique to our work. We evaluate our system on two machines and show that our approach predicts either the best or the second best method in 91-96\% of the matrices. Predictions achieve average speedups that are very close to the speedups achievable when only the best methods are used. By using an efficient code generator and a carefully designed set of matrix features, we show the total runtime costs of autotuning and specialization can be amortized to bring performance benefits for many real-world cases.