"Manual" Vectorization in 2025: When the Compiler Is Not Enough
summary
Examines manual SIMD/vectorization techniques on RISC-V and Arm using Eigen and OpenBLAS as case studies, comparing their performance to what compilers' auto-vectorization achieves. Analyzes why auto-vectorization falls short for AI/ML, AR/VR, CV, and HPC workloads, and shows performance and roofline plots to locate compute-bound vs memory-bound algorithms.
AI-written summary · about
chapters (10)
- 0:00Introduction to vectorization
- 5:47Matrix operations examples
- 9:42Performance problem example
- 12:33Memory hierarchy optimization
- 17:56Matrix extensions and speedup
- 21:55Theoretical speedup estimate
- 26:34Auto-vectorization vs manual
- 33:00Conclusions and recommendations
- 38:38Closing remarks
- 39:21Q&A
related talks
24mMoving to RISC-V Vector: A Practical Journey of AI Operator Optimization
13mUnleashing ML Processing Power Through RISC-V Vectors: Applications, Implementation and...
16mAccelerate HPC and AI Applications with RVV Auto-Vectorization
01h 01mIntroduction to Vector Parallelism
01h 45mPorting SIMD to RISC-V Vectors
19mGetting the Most out of the LLVM Auto Vectorizer for RISC-V Vectors (RVV)
17mExploring Static Code Generation and SIMD-Acceleration for Machine Learning on RISC-V
27mImproving Vectorization Heuristics in a Dynamic Compiler with Machine Learning Models