Data Layout Optimization for Portable Performance

Sharma, Kamal; Karlin, Ian; Keasler, Jeff; McGraw, James R.; Sarkar, Vivek

doi:10.1007/978-3-662-48096-0_20

Cited by 9 publications

(8 citation statements)

References 10 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Therefore, the layout of the data in the memory is an important factor for performance. There are two commonly used layouts [31]: (1) Arrayof-Structures (AoS) layout, where the data associated with one element is in consecutive places in the array (and thus in memory) and (2) Structure-of-Arrays (SoA) where the components of elements are stored consecutively e.g. the first data component of the elements are in the beginning of the array followed by the second, etc.…”

Section: Array-of-structures (Aos) Vs Structure-of-arrays (Soa)mentioning

confidence: 99%

See 1 more Smart Citation

Locality optimized unstructured mesh algorithms on GPUs

Sulyok

Balogh

Reguly

et al. 2019

Journal of Parallel and Distributed Computing

View full text Add to dashboard Cite

Unstructured-mesh based numerical algorithms such as finite volume and finite element algorithms form an important class of applications for many scientific and engineering domains. The key difficulty in achieving higher performance from these applications is the indirect accesses that lead to data-races when parallelized. Current methods for handling such data-races lead to reduced parallelism and suboptimal performance. Particularly on modern many-core architectures, such as GPUs, that has increasing core/thread counts, reducing data movement and exploiting memory locality is vital for gaining good performance.In this work we present novel locality-exploiting optimizations for the efficient execution of unstructured-mesh algorithms on GPUs. Building on a two-layered coloring strategy for handling data races, we introduce novel reordering and partitioning techniques to further improve efficient execution. The new optimizations are then applied to several well established unstructuredmesh applications, investigating their performance on NVIDIA's latest P100 and V100 GPUs. We demonstrate significant speedups (1.1-1.75×) compared to the state-of-the-art. A range of performance metrics are benchmarked including runtime, memory transactions, achieved bandwidth performance, GPU occupancy and data reuse factors and are used to understand and explain the key factors impacting performance. The optimized algorithms are implemented as an open-source software library and we illustrate its use for improving performance of existing or new unstructured-mesh applications.

show abstract

Section: Array-of-structures (Aos) Vs Structure-of-arrays (Soa)mentioning

confidence: 99%

“…This leads to performance improvements of 292 GB/s versus 268 GB/s. Low register counts (28)(29)(30)(31)(32) and single-precision data types also resulted in achieving a higher occupancy on the GPU compared to Airfoil. This, we believe explains why performance appears to be independent of the block size as shown in the Figure 11.…”

Section: Volnamentioning

confidence: 99%

Locality optimized unstructured mesh algorithms on GPUs

Sulyok

Balogh

Reguly

et al. 2019

Journal of Parallel and Distributed Computing

View full text Add to dashboard Cite

show abstract

“…Annotations and specific data layout optimizations with compiler support has been proposed by Sharma et al [17]. The source-to-source transformation requires to describe in a separate file the desired array interleaving.…”

Section: Related Workmentioning

confidence: 99%

Rewriting System for Profile-Guided Data Layout Transformations on Binaries

Haine

Aumage²,

Barthou

2017

Lecture Notes in Computer Science

View full text Add to dashboard Cite

Abstract. Careful data layout design is crucial for achieving high performance. However exploring data layouts is time-consuming and errorprone, and assessing the impact of a layout transformation on performance is difficult without performing it. We propose to guide application programmers through data layout restructuring by providing a comprehensive multidimensional description of the initial layout, built from trace analysis, and then by giving a performance evaluation of the transformations tested and an expression of each transformed layout. The programmer can limit the exploration to layouts matching some patterns. We apply this method to two multithreaded applications. The performance prediction of multiple transformations matches within 5% the performance of hand-transformed layout code.

show abstract

“…Choosing between the Array of Structures and Structure of Arrays representations is a known challenge in practice [1,22], with either representation potentially advantageous, depending on usage.…”

Section: Array Of Struct Vs Struct Of Arraymentioning

confidence: 99%

Interactive data representation migration: exploiting program dependence to aid program transformation

Narasimhan

Reichenbach

Lawall

2017

Proceedings of the 2017 ACM SIGPLAN Workshop on Partial Evaluation and Program Manipulation

View full text Add to dashboard Cite

Data representation migration is a program transformation that involves changing the type of a particular data structure, and then updating all of the operations that somehow depend on that data structure according to the new type. Changing the data representation can provide benefits such as improving efficiency and improving the quality of the computed results. Performing such a transformation is challenging, because it requires applying data-type specific changes to code fragments that may be widely scattered throughout the source code, connected by dataflow dependencies. Refactoring systems are typically sensitive to dataflow dependencies, but are not programmable with respect to the features of particular data types. Existing program transformation languages provide the needed flexibility, but do not concisely support reasoning about dataflow dependencies. To address the needs of data representation migration, we propose a new approach to program transformation that relies on a notion of semantic dependency: every transformation step propagates the transformation process onward to code that somehow depends on the transformed code. Our approach provides a declarative transformation-specification language, for expressing typespecific transformation rules. We further provide scoped rules, a mechanism for guiding rule application, and tags, a device for simple program analysis within our framework, to enable more powerful program transformations. We have implemented a prototype transformation system based on these ideas for C and C++ code and evaluate it against three example specifications, including vectorization, transformation of integers to big integers, and transformation of array-of-structures data types to structure-of-arrays format. Our evaluation shows that our approach can improve program performance and the precision of the computed results, and that it scales to programs of up to 3700 lines.

show abstract

Data Layout Optimization for Portable Performance

Cited by 9 publications

References 10 publications

Locality optimized unstructured mesh algorithms on GPUs

Locality optimized unstructured mesh algorithms on GPUs

Rewriting System for Profile-Guided Data Layout Transformations on Binaries

Interactive data representation migration: exploiting program dependence to aid program transformation

Contact Info

Product

Resources

About