CUSP is a GPU-accelerated library focusing on operations on sparse matrices (linear algebra operations, a system of equations solvers, etc.). Performance-wise, the most critical kernel in the CUSP library is SpMV (sparse matrix dense vector multiplication), which is used by the system of equation solvers. KTT is a framework for autotuning of CUDA and OpenCL kernels developed at ICS MUNI and FI MUNI.

The goal of the master thesis is the implementation of dynamic autotuning of SpMV into CUSP using KTT. First, the student analyzes CUSP architecture and finds a way to integrate KTT. Second, the student selects at least two formats of sparse matrices, analyzes literature to find out existing optimizations of SpMV with selected formats, and implements tunable code of SpMV, which contains selected published or even new optimizations. Third, the student evaluates the performance impacts of the implemented optimizations and compares the performance of the new CUSP with autotuning over the original CUSP implementation using a set of representative matrices.