WG211/M10Kamin

Run-time specialization for matrix multiplication by Sam Kamin

In this project, we - this is joint work with Baris Aktemur - are
addressing this question: Given a large but sparse matrix M,
what is the most efficient code we can generate to calculate Mv,
for dense matrix v? This question is of practical importance in
those applications where a single matrix is multiplied by numerous
vectors. In this talk, I will discuss our progress since the last
WG 2.11 meeting. I will explain various generation methods and
discuss timing results, as well as our ongoing efforts to find good
explanations for those timing results. I will also discuss further
steps needed to create a practical matrix-vector multiplication routine.