DigiNews

Tech Watch by Johan Denoyer

← Back to articles

Write your own kernel optimiser

Quality: 8/10 Relevance: 9/10

Summary

The article explains Cachet-based kernel optimization for matrix multiplication, showing how naive loops are transformed into cache-friendly schedules that reuse inputs and are verifiable by the compiler. It covers CPU and GPU strategies (blocking, packing, tiling, FMA, and parallelism), discusses attention-related reductions, and demonstrates cross-hardware benchmarking and reproducibility checks. It also discusses the role of semirings, proofs, and traffic budgets in ensuring correctness when regrouping floating-point reductions.

🚀 Service construit par Johan Denoyer