Compiler Heuristics Research Articles

Compilation Forking: A Fast and Flexible Way of Generating Data for Compiler-Internal Machine Learning Tasks

Compiler optimization decisions are often based on hand-crafted heuristics centered around a few established benchmark suites. Alternatively, they can be learned from feature and performance data produced during compilation. However, data-driven compiler optimizations based on machine learning models require large sets of quality data for training in order to match or even outperform existing human-crafted heuristics. In static compilation setups, related work has addressed this problem with iterative compilation. However, a dynamic compiler may produce different data depending on dynamically-chosen compilation strategies, which aggravates the generation of comparable data. We propose compilation forking, a technique for generating consistent feature and performance data from arbitrary, dynamically-compiled programs. Different versions of program parts with the same profiling and compilation history are executed within single program runs to minimize noise stemming from dynamic compilation and the runtime environment. Our approach facilitates large-scale performance evaluations of compiler optimization decisions. Additionally, compilation forking supports creating domain-specific compilation strategies based on machine learning by providing the data for model training. We implemented compilation forking in the GraalVM compiler in a programming-language-agnostic way. To assess the quality of the generated data, we trained several machine learning models to replace compiler heuristics for loop-related optimizations. The trained models perform equally well to the highly-tuned compiler heuristics when comparing the geometric means of benchmark suite performances. Larger impacts on few single benchmarks range from speedups of 20% to slowdowns of 17%. The presented approach can be implemented in any dynamic compiler. We believe that it can help to analyze compilation decisions and leverage the use of machine learning into dynamic compilation.

MiCOMP

Recent compilers offer a vast number of multilayered optimizations targeting different code segments of an application. Choosing among these optimizations can significantly impact the performance of the code being optimized. The selection of the right set of compiler optimizations for a particular code segment is a very hard problem, but finding the best ordering of these optimizations adds further complexity. Finding the best ordering represents a long standing problem in compilation research, named the phase-ordering problem. The traditional approach of constructing compiler heuristics to solve this problem simply cannot cope with the enormous complexity of choosing the right ordering of optimizations for every code segment in an application. This article proposes an automatic optimization framework we call MiCOMP, which <u>Mi</u>tigates the <u>Com</u>piler <u>P</u>hase-ordering problem. We perform phase ordering of the optimizations in LLVM’s highest optimization level using optimization sub-sequences and machine learning. The idea is to cluster the optimization passes of LLVM’s O3 setting into different clusters to predict the speedup of a complete sequence of all the optimization clusters instead of having to deal with the ordering of more than 60 different individual optimizations. The predictive model uses (1) dynamic features, (2) an encoded version of the compiler sequence, and (3) an exploration heuristic to tackle the problem. Experimental results using the LLVM compiler framework and the Cbench suite show the effectiveness of the proposed clustering and encoding techniques to application-based reordering of passes, while using a number of predictive models. We perform statistical analysis on the results and compare against (1) random iterative compilation, (2) standard optimization levels, and (3) two recent prediction approaches. We show that MiCOMP’s iterative compilation using its sub-sequences can reach an average performance speedup of 1.31 (up to 1.51). Additionally, we demonstrate that MiCOMP’s prediction model outperforms the -O1, -O2, and -O3 optimization levels within using just a few predictions and reduces the prediction error rate down to only 5%. Overall, it achieves 90% of the available speedup by exploring less than 0.001% of the optimization space.

Directive-based GPU programming for computational fluid dynamics

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. In this work we analyze 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. We examine the process of applying the OpenACC Fortran API to a test CFD code that serves as a proxy for a full-scale research code developed at Virginia Tech; this test code is used to asses the performance improvements attainable for our CFD algorithm on common GPU platforms, as well as to determine the modifications that must be made to the original source code in order to run efficiently on the GPU. Performance is measured on several recent GPU architectures from NVIDIA and AMD (using both double and single precision arithmetic) and the accelerator code is benchmarked against a multithreaded CPU version constructed from the same Fortran source code using OpenMP directives. A single NVIDIA Kepler GPU card is found to perform approximately 20× faster than a single CPU core and more than 2× faster than a 16-core Xeon server. An analysis of optimization techniques for OpenACC reveals cases in which manual intervention by the programmer can improve accelerator performance by up to 30% over the default compiler heuristics, although these optimizations are relevant only for specific platforms. Additionally, the use of multiple accelerators with OpenACC is investigated, including an experimental high-level interface for multi-GPU programming that automates scheduling tasks across multiple devices. While the overall performance of the OpenACC code is found to be satisfactory, we also observe some significant limitations and restrictions imposed by the OpenACC API regarding certain useful features of modern Fortran (2003/8); these are sufficient for us to conclude that it would not be practical to apply OpenACC to our full research code at this time due to the amount of refactoring required.

Continuous learning of compiler heuristics

Optimizing programs to exploit the underlying hardware architecture is an important task. Much research has been done on enabling compilers to find the best set of code optimizations that can build the fastest and less resource-hungry executable for a given program. A common approach is iterative compilation, sometimes enriched by machine learning techniques. This provides good results, but requires extremely long compilation times and an initial training phase lasting even for days or weeks. We present long-term learning, a new algorithm that allows the compiler user to improve the performance of compiled programs with reduced compilation times with respect to iterative compilation, and without an initial training phase. Our algorithm does not just build good programs: it acquires knowledge every time a program is compiled and it uses such knowledge to learn compiler heuristics, without the need for an expert to manually define them. The heuristics are evolved during every compilation, by evaluating their effect on the generated programs. We present implementations of long-term learning on top of two different compilers, and experimental data gathered on multiple hardware configurations showing its effectiveness.

Evolving Compiler Heuristics to Manage Communication and Contention

As computer architectures become increasingly complex, hand-tuning compiler heuristics becomes increasingly tedious and time consuming for compiler developers. This paper presents a case study that uses a genetic algorithm to learn a compiler policy. The target policy implicitly balances communication and contention among processing elements of the TRIPS processor, a physically realized prototype chip. We learn specialized policies for individual programs as well as general policies that work well across all programs. We also employ a two-stage method that first classifies the code being compiled based on salient characteristics, and then chooses a specialized policy based on that classification.This work is particularly interesting for the AI community because it 1) emphasizes the need for increased collaboration between AI researchers and researchers from other branches of computer science and 2) discusses a machine learning setup where training on the custom hardware requires weeks of training, rather than the more typical minutes or hours.

Meta optimization

Compiler writers have crafted many heuristics over the years to approximately solve NP-hard problems efficiently. Finding a heuristic that performs well on a broad range of applications is a tedious and difficult process. This paper introduces Meta Optimization, a methodology for automatically fine-tuning compiler heuristics. Meta Optimization uses machine-learning techniques to automatically search the space of compiler heuristics. Our techniques reduce compiler design complexity by relieving compiler writers of the tedium of heuristic tuning. Our machine-learning system uses an evolutionary algorithm to automatically find effective compiler heuristics. We present promising experimental results. In one mode of operation Meta Optimization creates application-specific heuristics which often result in impressive speedups. For hyperblock formation, one optimization we present in this paper, we obtain an average speedup of 23% (up to 73%) for the applications in our suite. Furthermore, by evolving a compiler's heuristic over several benchmarks, we can create effective, general-purpose heuristics. The best general-purpose heuristic our system found for hyperblock formation improved performance by an average of 25% on our training set, and 9% on a completely unrelated test set. We demonstrate the efficacy of our techniques on three different optimizations in this paper: hyperblock formation, register allocation, and data prefetching.

Task Selection for the Multiscalar Architecture

The multiscalar architecture advocates a distributed processor organization and task-level speculation to exploit high degrees of instruction level parallelism (ILP) in sequential programs without impeding improvements in clock speeds. The main goal of this paper is to understand the key implications of the architectural features of distributed processor organization and task-level speculation for compiler task selection from the point of view of performance. We identify the fundamental performance issues to be: control flow speculation, data communication, data dependence speculation, load imbalance, and task overhead. We show that these issues are intimately related to a few key characteristics of tasks: task size, intertask control flow, and intertask data dependence. We describe compiler heuristics to select tasks with favorable characteristics. We report experimental results to show that the heuristics are successful in boosting overall performance by establishing larger ILP windows. We also present a breakdown of execution times to show that register wait, load imbalance, control flow squash, and conventional pipeline losses are significant for almost all the SPEC95 benchmarks.