Explore a 15-minute conference talk from ACM SIGPLAN on integrating data layout into compilers and code generators. Delve into the importance of data layout in achieving high-performance code, focusing on how organizing data according to desired access order can significantly reduce data movement. Learn about dlcomp, a code generator for tensor contraction computations that takes tensor computations in Einstein notation and data layout descriptions in sparse polyhedral format as inputs. Discover how dlcomp uses a combination of polyhedra scanning and synthesis to generate code that efficiently iterates over tensor layouts and matches nonzero elements. Gain insights into the potential for generalizing this approach beyond sparse tensors, enhancing your understanding of advanced compiler optimization techniques.
Overview
Syllabus
[CTSTA'23] Integrating Data Layout into Compilers and Code Generators
Taught by
ACM SIGPLAN