Achieving High-Performance the Functional Way - A Functional Pearl on Expressing High...
summary
Functional languages RISE and ELEVATE rewrite high-level data-parallel programs into optimized low-level code, expressing tiling, fusion and loop-permutation as composable rewrite strategies rather than ad-hoc schedules. The system reproduces TVM/Halide-style optimizations across seven versions and performs about 45,000–62,000 rewrites per optimized variant, completing each optimization under two seconds. Users define custom strategies (e.g., map+reduce fusion, tile composition, loop perms) with well-defined semantics and generate competitive parallel code automatically.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 0:29Why HighPerformance
- 1:30Manual HighPerformance
- 2:27Decoupled HighPerformance
- 5:50The Functional Way
- 6:49Elevate
- 8:20Traversals
- 12:22Evaluation
related talks
24mVerified Tensor-Program Optimization Via High-Level Scheduling Rewrites
45mFP: The Good, the Bad and the Ugly
21mA Functional Approach to Synthesizing Routable Programmable Accelerators for Neural(…)
35mOutperforming Imperative with Pure Functional Languages
43mEverything You Always Wanted to Know About Functional Programming...
02h 50mFunctional Array Funhouse Intensive - Part 3 of 5
23mDeep Fusion for Efficient Nested Recursive Computations
27mDistributing and Parallelizing Non-canonical Loops