Navigating the Frontier of EDDO-Based Accelerators: Programming and Code(…)
summary
Argues that EDDO-style accelerators like IBM AIU trade hardware complexity for much harder software and compiler problems, then shows a concrete codegen fix: a polyhedral/affine substitution pass that rewrites address and variable definitions so four overlapping live ranges collapse to a single-register mapping on cores with no stack or multipliers. Demonstrates the pass on a triply nested loop with constant bounds and strides, explains when it applies to ML access patterns, and reports it succeeds when addresses are affine and loop counts are known. affine substitution
AI-written summary · about
chapters (10)
- 0:00Energy efficiency in computing
- 1:05Coupled vs decoupled access
- 2:31EDDO architectures intro
- 5:00Hardware simplification vs software
- 5:21EDDO programming challenges
- 7:27Code generation in simplified cores
- 8:30Register management example
- 12:44Affine live range reduction
- 16:58Optimizing special workloads
- 17:19Acknowledgments
related talks
24mMapping Sparse Applications to Accelerated Computing Systems
22mAlgorithm-Aware Hardware Optimization using E-Graph Rewriting: how should we marry(…)
29mReverse Automatic Differentiation for Accelerate (Extended Abstract)
15mCoarsening Optimization for Differentiable Programming
39mAsk Me Anything with Alex Aiken
37mDealing with aliasing using contracts
13mMind the Abstraction Gap: Bringing Equality Saturation to Real-World ML Compilers
30mA Compiler-First Planetary Compute Engine: Automatic differentiable and performance(…)