Tracking source locations in the Futhark compiler

9 months ago

Futhark is a programming language designed for writing fast programs, but sometimes programs are not fast, necessitating profiling.
Initially, Futhark had limited profiling support, but recent updates have improved it by adding human-readable reports from machine-readable data.
The profiler initially provided data with compiler-generated names, making it hard for users to relate to their original code.
To improve usability, the Futhark compiler now tracks provenance (source code origins) throughout compilation, even through aggressive optimizations like fusion and fission.
Provenance tracking piggybacks on existing certificate infrastructure in the compiler, ensuring accurate propagation of source information.
The solution focuses on tracking leaf scalar operations rather than higher-level constructs like maps, providing more useful profiling data.
Recent updates now show detailed source locations in profiling logs, helping users identify slow parts of their code.
Future improvements could include better machine-readable profiling data, heat maps of code performance, and suggestions for optimization.
Challenges remain in correlating GPU kernel performance with specific parts of the source code and providing actionable optimization advice.

Hasty Briefsbeta