Fast and Effective Orchestration of Compiler Optimizations for Automatic Performance Tuning

Pan, Zhelong; Eigenmann, Rudolf

doi:10.1109/cgo.2006.38

Cited by 59 publications

(8 citation statements)

References 14 publications

(29 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Most similar to our approach is Grammatical Evolution (GE) [Ryan et al 1998]. Here, a grammar is used, giving the same flexibility as we do and also avoiding the problems of GP.…”

Section: Related Workmentioning

confidence: 98%

“…Many smart search heuristics have been developed [Cooper et al 2005;Kulkarni et al 2004;Pan and Eigenmann 2006;Triantafyllis et al 2003]. Cooper et al [1999Cooper et al [ , 2002Cooper et al [ , 2005 explore the optimisation space using hill climbing and genetic algorithms.…”

Section: Related Workmentioning

confidence: 99%

“…Some of the state-of-the-art machine learning tools, such as Support Vector Machines (SVMs), are very slow, sometimes by two orders of magnitude compared to simpler tools such as decision trees. In this work, then, we use C4.5 decision trees [Quinlan 1993], which are considerably faster. However, since like many search problems this could be trivially parallelised, the runtime of the slower tools could become tolerable.…”

Section: Machine Learningmentioning

confidence: 99%

See 2 more Smart Citations

Automatic feature generation for machine learning--based optimising compilation

Leather

Bonilla

O’Boyle

2014

ACM Trans. Archit. Code Optim.

View full text Add to dashboard Cite

Recent work has shown that machine learning can automate and in some cases outperform handcrafted compiler optimisations. Central to such an approach is that machine learning techniques typically rely upon summaries or features of the program. The quality of these features is critical to the accuracy of the resulting machine learned algorithm; no machine learning method will work well with poorly chosen features. However, due to the size and complexity of programs, theoretically there are an infinite number of potential features to choose from. The compiler writer now has to expend effort in choosing the best features from this space. This article develops a novel mechanism to automatically find those features that most improve the quality of the machine learned heuristic. The feature space is described by a grammar and is then searched with genetic programming and predictive modelling. We apply this technique to loop unrolling in GCC 4.3.1 and evaluate our approach on a Pentium 6. On a benchmark suite of 57 programs, GCCs hard-coded heuristic achieves only 3% of the maximum performance available, whereas a state-of-the-art machine learning approach with hand-coded features obtains 59%. Our feature generation technique is able to achieve 76% of the maximum available speedup, outperforming existing approaches.

show abstract

“…Most similar to our approach is Grammatical Evolution (GE) [Ryan et al 1998]. Here, a grammar is used, giving the same flexibility as we do and also avoiding the problems of GP.…”

Section: Related Workmentioning

confidence: 98%

Section: Related Workmentioning

confidence: 99%

Section: Machine Learningmentioning

confidence: 99%

See 1 more Smart Citation

Automatic feature generation for machine learning--based optimising compilation

Leather

Bonilla

O’Boyle

2014

ACM Trans. Archit. Code Optim.

View full text Add to dashboard Cite

show abstract

“…Thus, this approach can make it possible to generate/evaluate the entire search space and determine the optimal phase ordering solution. It is also interesting to note that this exhaustive solution to the phase ordering problem subsumes the related issue of phase selection, which deals with deciding what transformations to apply without considering their order [7,15,23]. Any phase sequence of any length from the phase order/selection search space can be mapped to a node in the DAG of Figure 1.…”

Section: Algorithm For Exhaustive Search Space Enumerationmentioning

confidence: 99%

“…Thus, increasingly, individual phases are designed with minimal regard to the phase ordering problem, and the issue of the most appropriate phase sequence is dealt with as an afterthought at the end of the compiler design process. Empirical search algorithms are commonly used to iterate over possible phase sequences or orderings for each input program, applying, evaluating, and comparing each sequence with all others to find the best one [6][7][8][9][10][11][12]. Unfortunately, optimization phase ordering/selection search spaces in current compilers have been reported to consist in excess of 15 32 [13], or 16 10 [14], or 2 60 [15] unique phase sequences, making exhaustive phase order searches highly time consuming and impractical.…”

Section: Introductionmentioning

confidence: 99%

Analyzing and addressing false interactions during compiler optimization phase ordering

Jantz

Kulkarni

2013

Softw. Pract. Exper.

View full text Add to dashboard Cite

Compiler optimization phase ordering is a fundamental, pervasive, and long-standing problem for optimizing compilers. This problem is caused by interacting optimization phases producing different codes when applied in different orders. Producing the best phase ordering code is very important in performance-oriented and cost-constrained domains, such as embedded systems. In this work, we analyze the causes of the phase ordering problem in our compiler, Very Portable Optimizer (VPO), and report our observations. We devise new techniques to eliminate, what we call, false phase interactions in our compiler. We find that reducing such false phase interactions significantly prunes the phase order search space. We also develop and study algorithms to find the best average performance that can be delivered by a single phase sequence over our benchmark set and discuss the challenges in resolving this important problem. Our results show that there is no single sequence in VPO that can achieve the optimal phase ordering performance across all functions.ANALYZING FALSE INTERACTIONS DURING OPTIMIZATION PHASE ORDERING 645 set. Therefore, in this work, we also explore this problem of automatically finding the best average performance that can be achieved by any single compiler phase sequence in our compiler and discuss the challenges in addressing this issue. Our study empirically shows the non-existence of a single phase sequence that can achieve optimal code for all functions.Although phase ordering is a pervasive issue across all or most compilers, we must necessarily restrict this study to a single compiler. The presence of an exhaustive framework for phase ordering research, along with our intimate knowledge of its internal organization, made VPO the best choice for us to conduct this research. However, we believe that our motivation and methodology for exploring and addressing phase interaction issues should be more generally applicable to other compilers. Most of all, we believe that even though different compilers may have different optimization phases, implementations, and phase interactions, this study shows the potential and benefit of understanding such interactions to guide future implementation decisions that minimize the phase ordering problem and generate higher-quality code in compilers. The major contributions of this research work are as follows:ANALYZING FALSE INTERACTIONS DURING OPTIMIZATION PHASE ORDERING 647 § In theory, compiler optimizations can undo changes made by preceding phases and introduce back-edges in the directed graph. However, for our set of 244 benchmark functions, the exhaustive search algorithm only produced 'acyclic' graphs (no back-edges), and so we call them DAGs in this paper. The acyclic nature of the graphs is not a requirement for the exhaustive search algorithm. ANALYZING FALSE INTERACTIONS DURING OPTIMIZATION PHASE ORDERING 651 ** Applying conservative copy propagation explicitly as a distinct reorderable increases the size of the search space by over 77%, on average, and reduces...

show abstract

Comprehensive cache performance tuning with a toolset

Tao

2010

Future Generation Computer Systems

View full text Add to dashboard Cite

Fast and Effective Orchestration of Compiler Optimizations for Automatic Performance Tuning

Abstract: Abstract

Cited by 59 publications

References 14 publications

Automatic feature generation for machine learning--based optimising compilation

Automatic feature generation for machine learning--based optimising compilation

Analyzing and addressing false interactions during compiler optimization phase ordering

Comprehensive cache performance tuning with a toolset

Contact Info

Product

Resources

About