Abstract

The growing complexity of modern processors has made the generation of highly efficient code increasingly difficult. Manual code generation is very time consuming, but it is often the only choice since the code generated by today’s compiler technology often has much lower performance than the best hand-tuned codes. A promising code generation strategy, implemented by systems like ATLAS, FFTW, and SPIRAL, uses empirical search to find the parameter values of the implementation, such as the tile size and instruction schedules, that deliver near-optimal performance for a particular machine. However, this approach has only proven successful on scientific codes whose performance does not depend on the input data. In this thesis we study machine learning techniques that extend empirical search to the generation of algorithms whose performance depends on both the input characteristics and the architecture of the target machine. More specially, we target our study on sorting and recursive matrix-matrix multiplication, which are two fundamental algorithm problems. We observe that various sorting algorithms perform differently depending on input characteristics. We first study if it is possible to predict and select the best sorting algorithm for a specific input. We develop a machine-learning based technique to find the mapping from architectural features and input characteristics to the selection of best algorithm. The mapping is used at runtime to make selection of sorting algorithms. Experiments show that our approach always

Citations