Best Practices for Vectorization Getting Ready for Intel AVX 512 (2015) 计算机科学

Best Practices for Vectorization Getting ready for Intel:registered: Advanced Vector Extensions 512(Intel:registered: AVX-512)Manel FernándezIntel HPC Software Workshop Series 2015May 6th 2015, MunichThe need for SIMD vectorization Is the Intel:registered: Xeon PhiTM coprocessor right for me?Single thread (ST) performance is limited in today’s CPUs• Clock frequency constraints• Difficult to discover “near” Instruction level parallelism (ILP) by hardwareMore transistors dedicated to exploit “