Resumen de SIMD@OpenMP: a programming model approach to leverage SIMD features
Diego Luis Caballero de Gea
SIMD instruction sets are a key feature in current general purpose and high performance architectures. SIMD instructions apply in parallel the same operation to a group of data, commonly known as vector. A single SIMD/vector instruction can, thus, replace a sequence of scalar instructions. Consequently, the number of instructions can be greatly reduced leading to improved execution times. However, SIMD instructions are not widely exploited by the vast majority of programmers. In many cases, taking advantage of these instructions relies on the compiler. Nevertheless, compilers struggle with the automatic vectorization of codes. Advanced programmers are then compelled to exploit SIMD units by hand, using low-level hardware-specific intrinsics. This approach is cumbersome, error prone and not portable across SIMD architectures. This thesis targets OpenMP to tackle the underuse of SIMD instructions from three main areas of the programming model: language constructions, compiler code optimizations and runtime algorithms. We choose the Intel Xeon Phi coprocessor (Knights Corner) and its 512-bit SIMD instruction set for our evaluation process. We make four contributions aimed at improving the exploitation of SIMD instructions in this scope. Our first contribution describes a compiler vectorization infrastructure suitable for OpenMP. This infrastructure targets for-loops and whole functions. We define a set of attributes for expressions that determine how the code is vectorized. Our vectorization infrastructure also implements support for several advanced vector features. This infrastructure is proven to be effective in the vectorization of complex codes and it is the basis upon which we build the following two contributions. The second contribution introduces a proposal to extend OpenMP 3.1 with SIMD parallelism. Essential parts of this work have become key features of the SIMD proposal included in OpenMP 4.0. We define the "simd" and "simd for" directives that allow programmers to describe SIMD parallelism of loops and whole functions. Furthermore, we propose a set of optional clauses that leads the compiler to generate a more efficient vector code. These SIMD extensions improve the programming efficiency when exploiting SIMD resources. Our evaluation on the Intel Xeon Phi coprocessor shows that our SIMD proposal allows the compiler to efficiently vectorize codes poorly or not vectorized automatically with the Intel C/C++ compiler. In the third contribution, we propose a vector code optimization that enhances overlapped vector loads. These vector loads redundantly read from memory scalar elements already loaded by other vector loads. Our vector code optimization improves the memory usage of these accesses by means of building a vector register cache and exploiting register-to-register instructions. Our proposal also includes a new clause (overlap) in the context of the SIMD extensions for OpenMP of our first contribution. This new clause allows enabling, disabling and tuning this optimization on demand. The last contribution tackles the exploitation of SIMD instructions in the OpenMP barrier and reduction primitives. We propose a new combined barrier and reduction tree scheme specifically designed to make the most of SIMD instructions. Our barrier algorithm takes advantage of simultaneous multi-threading technology (SMT) and it utilizes SIMD memory instructions in the synchronization process. The four contributions of this thesis are an important step in the direction of a more common and generalized use of SIMD instructions. Our work is having an outstanding impact on the whole OpenMP community, ranging from users of the programming model to compiler and runtime implementations. Our proposals in the context of OpenMP improves the programmability of the programming model, the overhead of runtime services and the execution time of applications by means of a better use of SIMD.
