Programmers invest extensive development effort to optimize a GPU program to achieve peak performance. Achieving this requires an efficient usage of global memory, and avoiding memory bandwidth underutilization. The OpenACC programming model has been introduced to tackle the accelerators programming complexity. However, this model’s coarse-grained control on a program can make the memory bandwidth utilization even worse compared to the version written in a native GPU languages such as CUDA. We propose an extension to OpenACC in order to reduce the traffic on the memory interconnection network, using a compression method on floating point numbers. We examine our method on six case studies, and achieve up to 1.36X speedup. / Graduate / 0544 / 0984 / ebads67@uvic.ca

GPU Computing promises very high performance per watt for highly-parallelizable workloads. Nowadays, there are various programming models developed to utilize the computational power of GPGPUs. Low-level programming models provide full control over GPU resources and allow programmers to achieve peak performance of the chip. In contrast, high-level programming models hide GPU-specific programming details and allow programmers to mainly express parallelism. Later, the compiler parses the parallelization notes and translates them to low-level programming models. This saves tremendous development effort and improves productivity, often achieved at the cost of sacrificing performance. In this dissertation, we investigate the limitations of high-level programming models in achieving a performance near to low-level models. Specifically, we study the performance and productivity gap between high-level OpenACC and low-level CUDA programming models and aim at reducing the performance gap, while maintaining the productivity advantages. We start this study by developing our in-house OpenACC compiler. Our compiler, called IPMACC, translates OpenACC for C to CUDA and uses the system compile to generate GPU binaries. We develop various micro-benchmarks to understand GPU structure and implement a more efficient OpenACC compiler. By using IPMACC, we evaluate the performance and productivity gap between a wide set of OpenACC and CUDA kernels. From our findings, we conclude that one of the major reasons behind the big performance gap between OpenACC and CUDA is CUDA’s flexibility in exploiting the GPU software-managed cache. Identifying this key benefit in low-level CUDA, we follow three effective paths in utilizing software-managed cache similar to CUDA, but at a lower development effort (e.g. using OpenACC instead). In the first path, we explore the possibility of employing existing OpenACC directives in utilizing software-managed cache. Specifically, the cache directive is devised in OpenACC API standard to allow the use of software-managed cache in GPUs. We introduce an efficient implementation of OpenACC cache directive that performs very close to CUDA. However, we show that the use of the cache directive is limited and the directive may not offer the full-functionality associated with the software-managed cache, as existing in CUDA. In the second path, we build on our observation on the limitations of the cache directive and propose a new OpenACC directive, called the fcw directive, to address the shortcomings of the cache directive, while maintaining OpenACC productivity advantages. We show that the fcw directive overcomes the cache directive limitations and narrows down the performance gap between CUDA and OpenACC significantly. In the third path, we propose fully-automated hardware/software approach, called TELEPORT, for software-managed cache programming. On the software side, TELEPORT statically analyzes CUDA kernels and identifies opportunities in utilizing the software-managed cache. The required information is passed to the GPU via API calls. Based on this information, on the hardware side, TELEPORT prefetches the data to the software-managed cache at runtime. We show that TELEPORT can improve performance by 32% on average, while lowering the development effort by 2.5X, compared to hand-written CUDA equivalent. / Graduate

As a consequence of the CPU and GPU's architectures advance, in the last years there was a raise of the number of parallel programming APIs for both devices. While OpenMP is used to make parallel programs for the CPU, CUDA and OpenACC are employed in the parallel processing in the GPU. In the programming for the GPU, CUDA presents a model based on functions that make the source code extensive and prone to errors, in addition to leading to low development productivity. OpenACC emerged aiming to solve these problems and to be an alternative to the utilization of CUDA. Similar to OpenMP, this API has policies that ease the development of parallel applications that run on the GPU only. To further increase performance and take advantage of the parallel aspects of both CPU and GPU, it is possible to develop hybrid algorithms that split the processing on the two devices. In that sense, the main objective of this work is to verify if the advantages that OpenACC introduces are also positively reflected on the hybrid programming using OpenMP, if compared to the OpenMP + CUDA model. A second objective of this work is to identify aspects of the two programming models that could limit the performance or on the applications' development. As a way to accomplish these goals, this work presents the development of three hybrid parallel algorithms that are based on the Rodinia's benchmark algorithms, namely, RNG, Hotspot and SRAD, using the hybrid models OpenMP + CUDA and OpenMP + OpenACC. In these algorithms, the CPU part of the code is programmed using OpenMP, while it's assigned for the CUDA and OpenACC the parallel processing on the GPU. After the execution of the hybrid algorithms, the performance, efficiency and the processing's splitting in each one of the devices were analyzed. It was verified, through the hybrid algorithms' runs, that, in the two proposed programming models it was possible to outperform the performance of a parallel application that runs on a single API and in only one of the devices. In addition to that, in the hybrid algorithms RNG and Hotspot, CUDA's performance was superior to that of OpenACC, while in the SRAD algorithm OpenACC was faster than CUDA. / Como consequência do avanço das arquiteturas de CPU e GPU, nos últimos anos houve um aumento no número de APIs de programação paralela para os dois dispositivos. Enquanto que OpenMP é utilizada no processamento paralelo em CPU, CUDA e OpenACC são empregadas no processamento paralelo em GPU. Na programação para GPU, CUDA apresenta um modelo baseado em funções que deixam o código fonte extenso e propenso a erros, além de acarretar uma baixa produtividade no desenvolvimento. Objetivando solucionar esses problemas e sendo uma alternativa à utilização de CUDA surgiu o OpenACC. Semelhante ao OpenMP, essa API disponibiliza diretivas que facilitam o desenvolvimento de aplicações paralelas, porém para execução em GPU. Para aumentar ainda mais o desempenho e tirar proveito da capacidade de paralelismo de CPU e GPU, é possível desenvolver algoritmos híbridos que dividam o processamento nos dois dispositivos. Nesse sentido, este trabalho objetiva verificar se as facilidades que o OpenACC introduz também refletem positivamente na programação híbrida com OpenMP, se comparado ao modelo OpenMP + CUDA. Além disso, o trabalho visa relatar as limitações nos dois modelos de programação híbrida que possam influenciar no desempenho ou no desenvolvimento de aplicações. Como forma de cumprir essas metas, este trabalho apresenta o desenvolvimento de três algoritmos paralelos híbridos baseados nos algoritmos do benchmark Rodinia, a saber, RNG, Hotspot e SRAD, utilizando os modelos híbridos OpenMP + CUDA e OpenMP + OpenACC. Nesses algoritmos é atribuída ao OpenMP a execução paralela em CPU, enquanto que CUDA e OpenACC são responsáveis pelo processamento paralelo em GPU. Após as execuções dos algoritmos híbridos foram analisados o desempenho, a eficiência e a divisão da execução em cada um dos dispositivos. Verificou-se através das execuções dos algoritmos híbridos que nos dois modelos de programação propostos foi possível superar o desempenho de uma aplicação paralela em uma única API, com execução em apenas um dos dispositivos. Além disso, nos algoritmos híbridos RNG e Hotspot o desempenho de CUDA foi superior ao desempenho de OpenACC, enquanto que no algoritmo SRAD a API OpenACC apresentou uma execução mais rápida, se comparada à API CUDA.

This thesis explores potential use of GPGPU libraries to accelerate k-Wave toolkit's acoustic wave propagation simulation. Firstly, the thesis researches and assesses available high level GPGPU libraries. Afterwards, an insight into k-Wave toolkit's current state of simulation acceleration is provided. Based on that, an approach to enhance currently available code for processors into a heterogeneous application, that is capable of being run on graphics card, is proposed. The outcome of this thesis is an application that can utilize graphics card. If graphics card is unavailable, a fallback into thread and SIMD based acceleration for processor is executed. The product of this thesis is then evaluated based on its performance, maintenance difficulty and usability.

Computational Fluid Dynamics (CFD) applications usually involve intensive computations, which can be accelerated through using open accelerators, especially GPUs due to their common use in the scientific computing community. In addition to code acceleration, it is important to ensure that the code and algorithm are implemented numerically correctly, which is called code verification. This dissertation focuses on accelerating research CFD codes on multi-CPUs/GPUs using MPI and OpenACC, as well as the code verification for turbulence model implementation using the method of manufactured solutions and code-to-code comparisons. First, a variety of performance optimizations both agnostic and specific to applications and platforms are developed in order to 1) improve the heterogeneous CPU/GPU compute utilization; 2) improve the memory bandwidth to the main memory; 3) reduce communication overhead between the CPU host and the GPU accelerator; and 4) reduce the tedious manual tuning work for GPU scheduling. Both finite difference and finite volume CFD codes and multiple platforms with different architectures are utilized to evaluate the performance optimizations used. A maximum speedup of over 70 is achieved on 16 V100 GPUs over 16 Xeon E5-2680v4 CPUs for multi-block test cases. In addition, systematic studies of code verification are performed for a second-order accurate finite volume research CFD code. Cross-term sinusoidal manufactured solutions are applied to verify the Spalart-Allmaras and k-omega SST model implementation, both in 2D and 3D. This dissertation shows that the spatial and temporal schemes are implemented numerically correctly. / Doctor of Philosophy / Computational Fluid Dynamics (CFD) is a numerical method to solve fluid problems, which usually requires a large amount of computations. A large CFD problem can be decomposed into smaller sub-problems which are stored in discrete memory locations and accelerated by a large number of compute units. In addition to code acceleration, it is important to ensure that the code and algorithm are implemented correctly, which is called code verification. This dissertation focuses on the CFD code acceleration as well as the code verification for turbulence model implementation. In this dissertation, multiple Graphic Processing Units (GPUs) are utilized to accelerate two CFD codes, considering that the GPU has high computational power and high memory bandwidth. A variety of optimizations are developed and applied to improve the performance of CFD codes on different parallel computing systems. The program execution time can be reduced significantly especially when multiple GPUs are used. In addition, code-to-code comparisons with some NASA CFD codes and the method of manufactured solutions are utilized to verify the correctness of a research CFD code.

High-level parallel programming approaches, such as OpenACC, have recently become popular in complex fluid dynamics research since they are cross-platform and easy to implement. OpenACC is a directive-based programming model that, unlike low-level programming models, abstracts the details of implementation on the GPU. Although OpenACC generally limits the performance of the GPU, this model significantly reduces the work required to port an existing code to any accelerator platform, including GPUs. The purpose of this research is twofold: to investigate the effectiveness of OpenACC in developing a portable and maintainable GPU-accelerated code, and to determine the capability of OpenACC to accelerate large, complex programs on the GPU. In both of these studies, the OpenACC implementation is optimized and extended to a multi-GPU implementation while maintaining a unified code base. OpenACC is shown as a viable option for GPU computing with CFD problems. In the first study, a CFD code that solves incompressible cavity flows is accelerated using OpenACC. Overlapping communication with computation improves performance for the multi-GPU implementation by up to 21%, achieving up to 400 times faster performance than a single CPU and 99% weak scalability efficiency with 32 GPUs. The second study ports the execution of a more complex CFD research code to the GPU using OpenACC. Challenges using OpenACC with modern Fortran are discussed. Three test cases are used to evaluate performance and scalability. The multi-GPU performance using 27 GPUs is up to 100 times faster than a single CPU and maintains a weak scalability efficiency of 95%. / Master of Science / The research and analysis performed in scientific computing today produces an ever-increasing demand for faster and more energy efficient performance. Parallel computing with supercomputers that use many central processing units (CPUs) is the current standard for satisfying these demands. The use of graphics processing units (GPUs) for scientific computing applications is an emerging technology that has gained a lot of popularity in the past decade. A single GPU can distribute the computations required by a program over thousands of processing units. This research investigates the effectiveness of a relatively new standard, called OpenACC, for offloading execution of a program to the GPU. The most widely used standards today are highly complex and require low-level, detailed knowledge of the GPU’s architecture. These issues significantly reduce the maintainability and portability of a program. OpenACC does not require rewriting a program for the GPU. Instead, the developer annotates regions of code to run on the GPU and only has to denote high-level information about how to parallelize the code. The results of this research found that even for a complex program that models air flows, using OpenACC to run the program on 27 GPUs increases performance by a factor of 100 over a single CPU and by a factor of 4 over 27 CPUs. Although higher performance is expected with other GPU programming standards, these results were accomplished with minimal change to the original program. Therefore, these results demonstrate the ability of OpenACC to improve performance while keeping the program maintainable and portable.

Directive-based programming of graphics processing units (GPUs) has recently appeared as a viable alternative to using specialized low-level languages such as CUDA C and OpenCL for general-purpose GPU programming. This technique, which uses directive or pragma statements to annotate source codes written in traditional high-level languages, is designed to permit a unified code base to serve multiple computational platforms and to simplify the transition of legacy codes to new architectures. This work analyzes the popular OpenACC programming standard, as implemented by the PGI compiler suite, in order to evaluate its utility and performance potential in computational fluid dynamics (CFD) applications. Of particular interest is the handling of stencil algorithms, which are an important component of finite-difference and finite-volume numerical methods. To this end, the process of applying the OpenACC Fortran API to a preexisting finite-difference CFD code is examined in detail, and all modifications that must be made to the original source in order to run efficiently on the GPU are noted. Optimization techniques for OpenACC are also explored, and it is demonstrated that tuning the code for a particular accelerator architecture can result in performance increases of over 30%. There are also some limitations and programming restrictions imposed by the API: it is observed that certain useful features of modern Fortran (2003/8) are effectively disabled within OpenACC regions. Finally, a combination of OpenACC and OpenMP directives is used to create a truly cross-platform Fortran code that can be compiled for either CPU or GPU hardware. The performance of the OpenACC code is measured on several contemporary NVIDIA GPU architectures, and a comparison is made between double and single precision arithmetic showing that if reduced precision can be tolerated, it can lead to significant speedups. To assess the performance gains relative to a typical CPU implementation, the execution time for a standard benchmark case (lid-driven cavity) is used as a reference. The OpenACC version is compared against the identical Fortran code recompiled to use OpenMP on multicore CPUs, as well as a highly-optimized C++ version of the code that utilizes hardware aware programming techniques to attain higher performance on the Intel Xeon platforms being tested. Low-level optimizations specific to these architectures are analyzed and it is observed that the stencil access pattern required by the structured-grid CFD code sometimes leads to performance degrading conflict misses in the hardware managed CPU caches. The GPU code, which primarily uses software managed caching, is found to be free from these issues. Overall, it is observed that the OpenACC GPU code compares favorably against even the best optimized CPU version: using a single NVIDIA K20x GPU, the Fortran+OpenACC code is seen to outperform the optimized C++ version by 20% and the Fortran+OpenMP version by more than 100% with both CPU codes running on a 16-core Xeon workstation. / Master of Science