VECTORIZATION OF OPERATIONS ON SMALL- DIMENSIONAL MATRICES FOR INTEL XEON PHI KNIGHTS LANDING PROCESSOR

• Main Authors: Leonid A. Benderskiy, Sergey A. Leshchev, Alexey A. Rybakov
• Format: Article
• Language: Russian
• Published: The Fund for Promotion of Internet media, IT education, human development «League Internet Media» 2018-03-01
• Series: Современные информационные технологии и IT-образование
• Subjects: Matrix operations; vectorization; KNL; AVX-512; intrinsic functions
• Online Access: http://sitito.cs.msu.ru/index.php/SITITO/article/view/343

The article is devoted to the vectorization of calculations for Intel Xeon Phi Knights Landing (KNL) processor. Small-dimensional matrices are considered as objects for optimization. These operations are wide common in calculation codes in various scopes of research, for example, in calculational fluid dynamics. KNL is the latter Intel Xeon Phi processor, it contains up to 72 calculational cores and allows running applications using massive parallelism. They implement wide range of opportunities for effective performance of supercomputer calculations. In particular, they support different memory and cluster modes. In many cases the compiler isn't able to create high-performance parallel vectorized execution code. This leads to performance losses. One of the reserves of improving code performance is the manual vectorization of the hot blocks of the code. This leads to the entire application acceleration. An important step in the program optimizing when using KNL processors is applying special 512-bit vector instructions that can significantly increase the speed of the execution code. Using of 512-bit vector instructions allows processing vectors consisting of 16 floating-point values. Special fused multiply-add instructions allow us to combine operations of componentwise multiplication and addition of these vectors. For simplification of the manual vectorization of the program code, special intrinsic functions are used. In fact these functions are just wrappers over the processor instructions. Vectorization of operations on matrices, performed with the intrinsic functions, made it possible to reduce the execution time of these operations in the range from 23% to 70% in comparison with the version compiled by the Intel compiler with the maximum level of optimization. The results received show additional hidden performance reserves of applications that can be obtained by manual optimization of the source code.