Research On SIMD Auto-vectorization Optimization Technologies

With the support for floating-point operations becoming more sophisticated, SIMDextensions are more widely used in high performance computing. SW-VEC is an automaticparallelizing compiler which is designed to support the SIMD infrastructure of SW1600. Itexploits the SIMD data parallelism based on a series of optimization techniques such as programanalysis and reording, so as to generate SIMD codes for SW1600. Considering the exisitance ofvarious applications, it is impossible to introduce a general vectorization to exploit all thepotential parallelism. However, we can achieve it by classifying the applications according totheir characteristic. This thesis carries studies some SIMD auto-vectorization technologies on topof the SW-VEC. The main contributions of this thesis include:1. An across basic block vectorization method. Existing vectorization approaches such asSLP exploit the parallelisim within a single basic block. When coming to loops including severalbasic blocks, exisitng across basic block methods either need specific SIMD sets or are justavailable to special programs. As a result, all of these methods cannot be widely used.Therefore,we present an across basic block vectorization method to enhance the ability ofvectorization exploitation for SW-VEC.2. A method aiming at the vectorization of unnormalizable loops. Using traditionaltechniques, some loops cannot be normalized. In order to make these unnormalizible loopsvectorizable, we transform the loops into normailizible first by applying an improved method. Ifit fails, an improved SLP based method is used to exploit the results after unroll-jaming. Thesemethods improve the ability of vector recognition for SW-VEC significantly.3. A source to source algorithm oriented to dynamic array and intermediate representationcontaining pseudo-register after vectorization.An information-preserved based source to sourcealgorithm is presented to deal with the case that intermediate representation of information isabsent. It can address the translation problems that errors often occur in the past.In experimental sections, we test the source to source algorithm based on informationpreservation and vectorization exploiting approaches with some benchmark suites respectively.when applying all the algorithms proposed in this study, it has been proven that SW-VEC isavailable and effective for the SIMD infrastructure of SW1600. The algorithms proposed in thisstudy have been implemented in SW-VEC system developed by our research group.