KernelFaRer

Well-crafted libraries deliver much higher performance than code generated by sophisticated application programmers using advanced optimizing compilers. When a code pattern for which a well-tuned library implementation exists is found in the source code of an application, the highest performing solution is to replace the pattern with a call to the library. Idiom-recognition solutions in the past either required pattern matching machinery that was outside of the compilation framework or provided a very brittle solution that would fail even for minor variants in the pattern source code. This article introduces Kernel Find & Replacer ( KernelFaRer ), an idiom recognizer implemented entirely in the existing LLVM compiler framework. The versatility of KernelFaRer is demonstrated by matching and replacing two linear algebra idioms, general matrix-matrix multiplication (GEMM), and symmetric rank-2k update (SYR2K). Both GEMM and SYR2K are used extensively in scientific computation, and GEMM is also a central building block for deep learning and computer graphics algorithms. The idiom recognition in KernelFaRer is much more robust than alternative solutions, has a much lower compilation overhead, and is fully integrated in the broadly used LLVM compilation tools. KernelFaRer replaces existing GEMM and SYR2K idioms with computations performed by BLAS, Eigen, MKL (Intel’s x86), ESSL (IBM’s PowerPC), and BLIS (AMD). Gains in performance that reach 2000× over hand-crafted source code compiled at the highest optimization level demonstrate that replacing application code with library call is a performant solution.

Back-end porting of FT_MX based on LLVM compilation architecture

The processor FT_MX is a high-performance chip independently developed by the National University of Defense Technology, with an innovative architecture and instruction set. LLVM architecture is a widely used and efficient open source compiler framework initiated by the University of Illinois. This paper introduces the basic architecture and functions of LLVM, analyzes the back-end migration mechanism of the architecture in detail, and gives the specific process of implementing FT_MX back-end migration, and realizes the support of LLVM architecture to the back-end of FT_MX processor.

Autotuning LLVM OptimizationPasses for Matrix Multiplication in Rust

The LLVM compiler framework transforms its Intermediate Representation (IR) to optimize code. These transformations are controlled by flags which interfere on metrics such as the execution time. Selecting flags to improve the execution time of a program is difficult, and requires expert knowledge. Autotuning methods can automate parts of this process and help understanding the underlying search spaces. This paper describes ongoing work, showing that LLVM flags can impact the execution time of a Rust matrix multiplication algorithm, and planning future autotuning experiments for flag selection.