LLVM Hyderabad Meetup(In-person) | ML Driven Compiler Optimizations using IR2Vec

Register here (mandatory): https://forms.gle/1J6Dbjt1F2yt2UdX9

Abstract:
------------
It is well acknowledged that Machine Learning (ML) has been successful in improving on heuristics based compiler optimizations. Application of ML techniques for compiler optimizations has graduated from being an interesting academic exercise to being deployed in real-world compilers like LLVM. This research space has been laced with several program representation techniques ranging from feature-based to distributed program embeddings.
In this talk, we quickly begin with generic program embeddings and their utility in several applications. Specifically, we will talk about IR2Vec, our LLVM based program representation technique, which uses a representation learning based approach to build state-of-the-art scalable program encodings. We then discuss our experiences with applying variations of IR2Vec encodings to several classic downstream applications like device mapping (CPU vs. GPU), phase ordering (for code-size and execution-time), loop optimizations (distribution for vectorization) and backend optimizations (register allocation).

Speaker's Bio:
--------------------
Ramakrishna Upadrasta is an Associate Professor in Computer Science and Engineering Department in Indian Institute of Technology Hyderabad (IITH). He also holds an adjunct position in the Heritage Science and Technology department at IITH. He leads the Scalable Compilers for Heterogeneous Architectures group at IITH.
The group has contributed to program embeddings, generating efficient code for DNN programs, efficient cache miss calculations and tools that verify the correctness of parallel programs. The group members have also contributed to LLVM and the novel MLIR infrastructures, along with creating and maintaining novel indigenous infrastructures. Some of the accolades of the group include Google PhD Fellowship, PMRF fellowship, among others. The group has active collaborations with top research labs, academia and industries.

The publications, research areas and other details could be found at https://compilers.cse.iith.ac.in/, https://github.com/IITH-Compilers