This collaborative project is developing technologies to enable students, scientists, and other non-expert developers to use computer languages that facilitate rapid prototyping, and yet still automatically convert such programs to have high performance. In this research, the PI and co-PIs focus on programs that operate over visual data, such as programs in computer graphics, computer vision, and visualization. Visual data is important because visual datasets are rapidly growing in size, due to the use of cell-phone cameras, photo and video sharing online, and in scientific and medical imaging. The intellectual merits are that specialized program optimizations are being developed specifically for visual computing and for languages that enable rapid prototyping, alongside techniques that allow the computer to automatically search through different candidate optimizations and choose the fastest one. The project's broader significance and importance are that it will make the writing of computer programs that operate over visual datasets more accessible to novice programmers, make visual computing more accessible to a broader audience, permit faster research and development over visual programs, and make such programs themselves be more efficient.
More specifically, this research program is producing translating compilers that are specialized to handle programs that compute over visual data. The group led by the PI is researching new compilers that translate code from dynamic languages into highly efficient code in a target language. Dynamic languages are defined as those with a very dynamic run-time model, for example, MATLAB, Python, and Javascript. The target language is a language such as C that permits implementation of highly efficient programs. This research framework incorporates ideas from compilers, graphics, computer vision, visual perception, and formal and natural languages. The research will make a number of key intellectual contributions. First, new domain-specific translations and optimizations for visual computing will be formalized into manual rules that can be applied to any input program. Second, the team will research a novel approach of automatically learning translations, instead of using manually-coded rules. This can take the form of learning translation "suggestions" from humans, who can interactively suggest better output code. Third, a new search process based on offline auto-tuning will be used to select the translations that result in the fastest program. The success of the project will be verified against a comprehensive test suite of programs from computer vision and graphics.
