Accelerate HPC and AI Applications with RVV Auto-Vectorization
summary
Claims an LLVM-based compiler called ZCC that auto-vectorizes both inner and outer loops and uses pragmas to pick loop level. Benchmarks report 18% fewer dynamic instructions versus hand-written RVV intrinsics and 30% fewer dynamic instructions versus upstream LLVM on SPECInt 2006, about 10% cycle improvement on an out-of-order CPU. MobileNet runs 2.5× faster at vlen=128 and 7.7× at vlen=512 versus scalar, and shows 4–7.5× speedups compared to GCC across image models.
AI-written summary · about
chapters (4)
related talks
24mMoving to RISC-V Vector: A Practical Journey of AI Operator Optimization
19mGetting the Most out of the LLVM Auto Vectorizer for RISC-V Vectors (RVV)
16mImprovements to RISC-V Vector code generation in LLVM
27mImproving Vectorization Heuristics in a Dynamic Compiler with Machine Learning Models
25mDeep Dive: Accelerating Neural Networks using RVV and Open Standard Software
17mRecent Developments in Optimizing Compilers for RISC-V
19mAggregation Optimization for SIMD Everywhere from ARM Neon to RISC-V Vector and Crypto Extensions
20mBoosting Video Codec With RISC-V Vector Extension