Scientific Computing on Hybrid Architectures

Modern computer architectures, with multicore CPUs and GPUs or other accelerators, make stronger demands than ever on writers of scientific code. Normally, the most efficient program has to be written — using a substantial effort — by expert programmers for a certain application on a particular computer. This thesis deals with several algorithmic and technical approaches towards effectively satisfying the demand for high performance parallel scientific applications on hybrid computer architectures without incurring such a high cost in expert programmer time. Efficient programming is accomplished by writing performanceportable code where performance-critical functionality is provided either by an optimized library or by adaptively selecting which computational tasks that are executed on the CPU and the accelerator.