Multi-tier Dynamic Vectorization for Translating GPU Optimizations into CPU Performance