عنوان

Machine Learning in Compiler Optimization

Number

TL57869

.Language of Text, Soundtrack etc

انگلیسی

Title Proper

Machine Learning in Compiler Optimization

General Material Designation

[Thesis]

First Statement of Responsibility

Haj-Ali, Ameer

Subsequent Statement of Responsibility

Asanovic, Krste

Name of Publisher, Distributor, etc.

University of California, Berkeley

Date of Publication, Distribution, etc.

2020

Text of Note

110 p.

Dissertation or thesis details and type of degree

Ph.D.

Body granting the degree

University of California, Berkeley

Text preceding or following the note

2020

Text of Note

The end of Moore's law is driving the search for new techniques to improve system performance as applications continue to evolve rapidly and computing power demands continue to rise. One promising technique is to build more intelligent compilers. Compilers map high-level programs to lower-level primitives that run on hardware. During this process, compilers perform many complex optimizations to boost the performance of the generated code. These optimizations often require solving NP-Hard problems and dealing with an enormous search space. To overcome these challenges, compilers currently use hand-engineered heuristics that can achieve good but often far-from-optimal performance. Alternatively, software engineers resort to manually writing the optimizations for every section in the code, a burdensome process that requires prior experience and significantly increases the development time. In this thesis, novel approaches for automatically handling complex compiler optimization tasks are explored. End-to-end solutions using deep reinforcement learning and other machine learning algorithms are proposed. These solutions dramatically reduce the search time while capturing the code structure, different instructions, dependencies, and data structures to enable learning a sophisticated model that can better predict the actual performance cost and determine superior compiler optimizations. The proposed techniques can outperform existing state-of-the-art solutions while requiring shorter search time. Furthermore, unlike existing solutions, the deep reinforcement learning solutions are shown to generalize well to real benchmarks.

Subject Term

Computer engineering

Subject Term

Computer science

Subject Term

Electrical engineering

Haj-Ali, Ameer

Asanovic, Krste

University of California, Berkeley

Electronic name

p

[Thesis]

276903

a

Y