Throttling Automatic Vectorization: When Less is More

SIMD vectors are widely adopted in modern general purpose processors as they can boost performance and energy efficiency for certain applications. Compiler-based automatic vectorization is one approach for generating codethat makes efficient use of the SIMD units, and has the benefit of avoiding hand development and platform-specific optimizations. The Superword-Level Parallelism (SLP) vectorization algorithm is the most well-known implementation of automatic vectorization when starting from straight-line scalar code, and is implemented in several major compilers. The existing SLP algorithm greedily packs scalar instructions into vectors starting from stores and traversing the data dependence graph upwards until it reaches loads or non-vectorizable instructions. Choosing whether to vectorize is a one-off decision for the whole graph that has been generated. This, however, is sub-optimal because the graph may contain code that is harmful to vectorization due to the need to move data from scalar registers into vectors. The decision does not consider the potential benefits of throttling the graph by removing this harmful code. In this work we propose asolution to overcome this limitation by introducing Throttled SLP (TSLP), a novel vectorization algorithm that finds the optimal graph to vectorize, forcing vectorization to stop earlier whenever this is beneficial. Our experiments show that TSLP improves performance across a number of kernels extractedfrom widely-used benchmark suites, decreasing execution time compared to SLP by 9% on average and up to 14% in the best case.

[1]  David H. Bailey,et al.  The Nas Parallel Benchmarks , 1991, Int. J. High Perform. Comput. Appl..

[2]  Albert Cohen,et al.  Vapor SIMD: Auto-vectorize once, run everywhere , 2011, International Symposium on Code Generation and Optimization (CGO 2011).

[3]  Peng Wu,et al.  Efficient SIMD code generation for runtime alignment and length conversion , 2005, International Symposium on Code Generation and Optimization.

[4]  Noah Treuhaft,et al.  Scalable Processors in the Billion-Transistor Era: IRAM , 1997, Computer.

[5]  Vikram S. Adve,et al.  LLVM: a compilation framework for lifelong program analysis & transformation , 2004, International Symposium on Code Generation and Optimization, 2004. CGO 2004..

[6]  Ken Kennedy,et al.  PFC: A Program to Convert Fortran to Parallel Form , 1982 .

[7]  Fred Weber,et al.  AMD 3DNow! technology: architecture and implementations , 1999, IEEE Micro.

[8]  Ayal Zaks,et al.  Auto-vectorization of interleaved data for SIMD , 2006, PLDI '06.

[9]  Richard Henderson,et al.  Multi-platform auto-vectorization , 2006, International Symposium on Code Generation and Optimization (CGO'06).

[10]  Michael Wolfe,et al.  High performance compilers for parallel computing , 1995 .

[11]  Ayal Zaks,et al.  Outer-loop vectorization - revisited for short SIMD architectures , 2008, 2008 International Conference on Parallel Architectures and Compilation Techniques (PACT).

[12]  Mahmut T. Kandemir,et al.  A compiler framework for extracting superword level parallelism , 2012, PLDI '12.

[13]  Michael Wolfe.  Vector Optimization vs. Vectorization , 1987, ICS.

[14]  Jaewook Shin,et al.  Superword-level parallelism in the presence of control flow , 2005, International Symposium on Code Generation and Optimization.

[15]  Richard M. Russell,et al.  The CRAY-1 computer system , 1978, CACM.

[16]  Paul S. Wang,et al.  Chains of recurrences—a method to expedite the evaluation of closed-form functions , 1994, ISSAC '94.

[17]  David A. Padua,et al.  Dependence graphs and compiler optimizations , 1981, POPL '81.

[18]  Vivek Sarkar,et al.  Efficient Selection of Vector Instructions Using Dynamic Programming , 2010, 2010 43rd Annual IEEE/ACM International Symposium on Microarchitecture.

[19]  Saman P. Amarasinghe,et al.  Exploiting superword level parallelism with multimedia instruction sets , 2000, PLDI '00.

[20]  Peng Wu,et al.  Vectorization for SIMD architectures with alignment constraints , 2004, PLDI '04.

[21]  Timothy M. Jones,et al.  PSLP: Padded SLP automatic vectorization , 2015, 2015 IEEE/ACM International Symposium on Code Generation and Optimization (CGO).

[22]  Wilfried Oed.  Cray Y-MP C90: System features and early benchmark results (Short communication) , 1992, Parallel Comput..

[23]  David A. Padua,et al.  An Evaluation of Vectorizing Compilers , 2011, 2011 International Conference on Parallel Architectures and Compilation Techniques.

[24]  Gang Ren,et al.  Optimizing data permutations for SIMD devices , 2006, PLDI '06.

[25]  Michael Wolfe,et al.  The KAP/S-1 : An Advanced Source-to-Source Vectorizer for the S-1 Mark IIa Supercomputer , 1986, ICPP.

[26]  Erik Lindholm,et al.  NVIDIA Tesla: A Unified Graphics and Computing Architecture , 2008, IEEE Micro.

[27]  Ken Kennedy,et al.  Automatic translation of FORTRAN programs to vector form , 1987, TOPL.

[28]  Scott A. Mahlke,et al.  SIMD defragmenter: efficient ILP realization on data-parallel architectures , 2012, ASPLOS XVII.

[29]  P. Sadayappan,et al.  Dynamic trace-based analysis of vectorization potential of applications , 2012, PLDI.