gfursin
commented
6 years ago Actually, we tried DNN frameworks to classify compiler optimizations in our collaborative project with @raspberrypi (unified CK APIs and meta information allows us to quite easily plug in and evaluate different ML techniques). You can see results in this interactive CK report: http://cKnowledge.org/rpi-crowd-tuning . You can see shared workflows here: https://github.com/dividiti/ck-rpi-optimization .
Now, when we have enough statistics about GCC and LLVM distinct optimizations across different programs, datasets and hardware such as Android mobiles and RPi3 (see http://cKnowledge.org/repo), as well as CK-powered DNN frameworks, we can classify these optimization and then try to find relevant program, dataset and hardware features either manually (if they do not exist in the system - see our important example in https://arxiv.org/abs/1506.06256 or using DNN by feeding in program sources, intermediate representations, MILEPOST features, datasets, etc - see https://scholar.google.com/citations?view_op=view_citation&hl=en&user=IwcnpkwAAAAJ&citation_for_view=IwcnpkwAAAAJ:LkGwnXOMwfcC and https://scholar.google.com/citations?view_op=view_citation&hl=en&user=IwcnpkwAAAAJ&cstart=20&citation_for_view=IwcnpkwAAAAJ:KlAtU1dfN6UC for further details.
It can be a nice internship or GSOC project.