Search

The escalating demand to migrate legacy software across different Instruction Set Architectures (ISAs) has driven the development of assembly-to-assembly translators to map between their respective assembly languages. However, the development of these tools requires substantial engineering effort. State-of-the-art approaches use lifting, a technique where source assembly code is translated to an architecture-independent intermediate representation (IR) (for example, the LLVM IR) and use a pre-existing compiler to recompile the IR to the target ISA. However, the hand-written rules these lifters employ are sensitive to the particular compiler and optimization level used to generate the code and require significant engineering effort to support each new ISA. We propose Forklift, the first neural lifter that learns how to translate assembly to LLVM IR using a token-level encoder-decoder Transformer. We show how to incrementally add support to new ISAs by fine tuning the assembly encoder and freezing the IR decoder, improving the overall accuracy and efficiency. We collect millions of parallel LLVM IR, x86, ARM, and RISC-V programs across compilers and optimization levels to train Forklift and set up an input/output-based accuracy harness. We evaluate Forklift on two challenging benchmark suites and translate 2.5x more x86 programs than a state-of-the-art hand-written lifter and 4.4x more x86 programs than GPT-4 as well as enabling translation from new ISAs.

Deep Neural Networks (DNNs) are extremely computationally demanding, which presents a large barrier to their deployment on resource-constrained devices. Since such devices are where many emerging deep learning applications lie (e.g., drones, vision-based medical technology), significant bodies of work from both the machine learning and systems communities have attempted to provide optimizations to accelerate DNNs. To help unify these two perspectives, in this paper we combine machine learning and systems techniques within the Deep Learning Acceleration Stack (DLAS), and demonstrate how these layers can be tightly dependent on each other with an across-stack perturbation study. We evaluate the impact on accuracy and inference time when varying different parameters of DLAS across two datasets, seven popular DNN architectures, four DNN compression techniques, three algorithmic primitives with sparse and dense variants, untuned and auto-scheduled code generation, and four hardware platforms. Our evaluation highlights how perturbations across DLAS parameters can cause significant variation and across-stack interactions. The highest level observation from our evaluation is that the model size, accuracy, and inference time are not guaranteed to be correlated. Overall we make 13 key observations, including that speedups provided by compression techniques are very hardware dependent, and that compiler auto-tuning can significantly alter what the best algorithm to use for a given configuration is. With DLAS, we aim to provide a reference framework to aid machine learning and systems practitioners in reasoning about the context in which their respective DNN acceleration solutions exist in. With our evaluation strongly motivating the need for co-design, we believe that DLAS can be a valuable concept for exploring the next generation of co-designed accelerated deep learning solutions.

MLIR is an emerging compiler infrastructure for modern hardware, but existing programs cannot take advantage of MLIR's high-performance compilation if they are described in lower-level general purpose languages. Consequently, to avoid programs needing to be rewritten manually, this has led to efforts to automatically raise lower-level to higher-level dialects in MLIR. However, current methods rely on manually-defined raising rules, which limit their applicability and make them challenging to maintain as MLIR dialects evolve. We present mlirSynth -- a novel approach which translates programs from lower-level MLIR dialects to high-level ones without manually defined rules. Instead, it uses available dialect definitions to construct a program space and searches it effectively using type constraints and equivalences. We demonstrate its effectiveness \revi{by raising C programs} to two distinct high-level MLIR dialects, which enables us to use existing high-level dialect specific compilation flows. On Polybench, we show a greater coverage than previous approaches, resulting in geomean speedups of 2.5x (Intel) and 3.4x (AMD) over state-of-the-art compilation flows for the C programming language. mlirSynth also enables retargetability to domain-specific accelerators, resulting in a geomean speedup of 21.6x on a TPU.

Coarse-grained reconfigurable arrays (CGRAs) are domain-specific devices promising both the flexibility of FPGAs and the performance of ASICs. However, with restricted domains comes a danger: designing chips that cannot accelerate enough current and future software to justify the hardware cost. We introduce FlexC, the first flexible CGRA compiler, which allows CGRAs to be adapted to operations they do not natively support. FlexC uses dataflow rewriting, replacing unsupported regions of code with equivalent operations that are supported by the CGRA. We use equality saturation, a technique enabling efficient exploration of a large space of rewrite rules, to effectively search through the program-space for supported programs. We applied FlexC to over 2,000 loop kernels, compiling to four different research CGRAs and 300 generated CGRAs and demonstrate a 2.2$\times$ increase in the number of loop kernels accelerated leading to 3$\times$ speedup compared to an Arm A5 CPU on kernels that would otherwise be unsupported by the accelerator.

This work outlines a time-domain numerical integration technique for linear hyperbolic partial differential equations sourced by distributions (Dirac $\delta$-functions and their derivatives). Such problems arise when studying binary black hole systems in the extreme mass ratio limit. We demonstrate that such source terms may be converted to effective domain-wide sources when discretized, and we introduce a class of time-steppers that directly account for these discontinuities in time integration. Moreover, our time-steppers are constructed to respect time reversal symmetry, a property that has been connected to conservation of physical quantities like energy and momentum in numerical simulations. To illustrate the utility of our method, we numerically study a distributionally-sourced wave equation that shares many features with the equations governing linear perturbations to black holes sourced by a point mass., Comment: 29 pages, 4 figures

Decompilation is a well-studied area with numerous high-quality tools available. These are frequently used for security tasks and to port legacy code. However, they regularly generate difficult-to-read programs and require a large amount of engineering effort to support new programming languages and ISAs. Recent interest in neural approaches has produced portable tools that generate readable code. However, to-date such techniques are usually restricted to synthetic programs without optimization, and no models have evaluated their portability. Furthermore, while the code generated may be more readable, it is usually incorrect. This paper presents SLaDe, a Small Language model Decompiler based on a sequence-to-sequence transformer trained over real-world code. We develop a novel tokenizer and exploit no-dropout training to produce high-quality code. We utilize type-inference to generate programs that are more readable and accurate than standard analytic and recent neural approaches. Unlike standard approaches, SLaDe can infer out-of-context types and unlike neural approaches, it generates correct code. We evaluate SLaDe on over 4,000 functions from ExeBench on two ISAs and at two optimizations levels. SLaDe is up to 6 times more accurate than Ghidra, a state-of-the-art, industrial-strength decompiler and up to 4 times more accurate than the large language model ChatGPT and generates significantly more readable code than both.

Dedicated tensor accelerators demonstrate the importance of linear algebra in modern applications. Such accelerators have the potential for impressive performance gains, but require programmers to rewrite code using vendor APIs - a barrier to wider scale adoption. Recent work overcomes this by matching and replacing patterns within code, but such approaches are fragile and fail to cope with the diversity of real-world codes. We develop ATC, a compiler that uses program synthesis to map regions of code to specific APIs. The mapping space that ATC explores is combinatorially large, requiring the development of program classification, dynamic analysis, variable constraint generation and lexical distance matching techniques to make it tractable. We apply ATC to real-world tensor and linear algebra codes and evaluate them against four state-of-the-art approaches. We accelerate between 2.6x and 7x more programs, leading to over an order of magnitude performance improvement., Comment: This is the author's version of the work. It is posted here for your personal use. Not for redistribution. The definitive Version of Record was published in Proceedings of the 32nd ACM SIGPLAN International Conference on Compiler Construction (CC '23), February 25-26, 2023, Montr\'eal, QC, Canada, https://doi.org/10.1145/3578360.3580262

The scheduled launch of the LISA Mission in the next decade has called attention to the gravitational self-force problem. Despite an extensive body of theoretical work, long-time numerical computations of gravitational waves from extreme-mass-ratio-inspirals remain challenging. This work proposes a class of numerical evolution schemes suitable to this problem based on Hermite integration. Their most important feature is time-reversal symmetry and unconditional stability, which enables these methods to preserve symplectic structure, energy, momentum and other Noether charges over long time periods. We apply Noether's theorem to the master fields of black hole perturbation theory on a hyperboloidal slice of Schwarzschild spacetime to show that there exist constants of evolution that numerical simulations must preserve. We demonstrate that time-symmetric integration schemes based on a 2-point Taylor expansion (such as Hermite integration) numerically conserve these quantities, unlike schemes based on a 1-point Taylor expansion (such as Runge-Kutta). This makes time-symmetric schemes ideal for long-time EMRI simulations., Comment: 43 pages, 8 figures

Deep learning has had a significant impact on many fields. Recently, code-to-code neural models have been used in code translation, code refinement and decompilation. However, the question of whether these models can automate compilation has yet to be investigated. In this work, we explore neural compilation, building and evaluating Transformer models that learn how to produce x86 assembler from C code. Although preliminary results are relatively weak, we make our data, models and code publicly available to encourage further research in this area., Comment: Published in AIPLANS 2021

Improving the performance of deep neural networks (DNNs) is important to both the compiler and neural architecture search (NAS) communities. Compilers apply program transformations in order to exploit hardware parallelism and memory hierarchy. However, legality concerns mean they fail to exploit the natural robustness of neural networks. In contrast, NAS techniques mutate networks by operations such as the grouping or bottlenecking of convolutions, exploiting the resilience of DNNs. In this work, we express such neural architecture operations as program transformations whose legality depends on a notion of representational capacity. This allows them to be combined with existing transformations into a unified optimization framework. This unification allows us to express existing NAS operations as combinations of simpler transformations. Crucially, it allows us to generate and explore new tensor convolutions. We prototyped the combined framework in TVM and were able to find optimizations across different DNNs, that significantly reduce inference time - over 3$\times$ in the majority of cases. Furthermore, our scheme dramatically reduces NAS search time. Code is available at~\href{https://github.com/jack-willturner/nas-as-program-transformation-exploration}{this https url}.

Compiler architects increasingly look to machine learning when building heuristics for compiler optimization. The promise of automatic heuristic design, freeing the compiler engineer from the complex interactions of program, architecture, and other optimizations, is alluring. However, most machine learning methods cannot replicate even the simplest of the abstract interpretations of data flow analysis that are critical to making good optimization decisions. This must change for machine learning to become the dominant technology in compiler heuristics. To this end, we propose ProGraML - Program Graphs for Machine Learning - a language-independent, portable representation of whole-program semantics for deep learning. To benchmark current and future learning techniques for compiler analyses we introduce an open dataset of 461k Intermediate Representation (IR) files for LLVM, covering five source programming languages, and 15.4M corresponding data flow results. We formulate data flow analysis as an MPNN and show that, using ProGraML, standard analyses can be learned, yielding improved performance on downstream compiler optimization tasks., Comment: 9 pages, plus appendices. arXiv admin note: text overlap with arXiv:2003.10536

This paper is concerned with synthesizing programs based on black-box oracles: we are interested in the case where there exists an executable implementation of a component or library, but its internal structure is unknown. We are provided with just an API or function signature, and aim to synthesize a program with equivalent behavior. To attack this problem, we detail Presyn: a program synthesizer designed for flexible interoperation with existing programs and compiler toolchains. Presyn uses high-level imperative control-flow structures and a pair of cooperating predictive models to efficiently narrow the space of potential programs. These models can be trained effectively on small corpora of synthesized examples. We evaluate Presyn against five leading program synthesizers on a collection of 112 synthesis benchmarks collated from previous studies and real-world software libraries. We show that Presyn is able to synthesize a wider range of programs than each of them with less human input. We demonstrate the application of our approach to real-world code and software engineering problems with two case studies: accelerator library porting and detection of duplicated library reimplementations., Comment: Accepted to GPCE 2020

Library migration is a challenging problem, where most existing approaches rely on prior knowledge. This can be, for example, information derived from changelogs or statistical models of API usage. This paper addresses a different API migration scenario where there is no prior knowledge of the target library. We have no historical changelogs and no access to its internal representation. To tackle this problem, this paper proposes a novel approach (M$^3$), where probabilistic program synthesis is used to semantically model the behavior of library functions. Then, we use an SMT-based code search engine to discover similar code in user applications. These discovered instances provide potential locations for API migrations. We evaluate our approach against 7 well-known libraries from varied application domains, learning correct implementations for 94 functions. Our approach is integrated with standard compiler tooling, and we use this integration to evaluate migration opportunities in 9 existing C/C++ applications with over 1MLoC. We discover over 7,000 instances of these functions, of which more than 2,000 represent migration opportunities., Comment: Accepted to ASE 2020

We present a generalized piecewise polytropic parameterization for the neutron-star equation of state using an ansatz that imposes continuity in not only pressure and energy density, but also in the speed of sound. The universe of candidate equations of state is shown to admit preferred dividing densities, determined by minimizing an error norm consisting of integral astrophysical observables. Generalized piecewise polytropes accurately reproduce astrophysical observables, such as mass, radius, tidal deformability and mode frequencies, as well as thermodynamic quantities, such as the adiabatic index. This makes the new EOS useful for Bayesian parameter estimation from gravitational waveforms. Moreover, since they are differentiable, generalized piecewise polytropes can improve pointwise convergence in numerical relativity simulations of neutron stars. Existing implementations of piecewise polytropes can easily accommodate this generalization., Comment: 16 pages, 7 figures, Version accepted to PRD

When deploying a deep neural network on constrained hardware, it is possible to replace the network's standard convolutions with grouped convolutions. This allows for substantial memory savings with minimal loss of accuracy. However, current implementations of grouped convolutions in modern deep learning frameworks are far from performing optimally in terms of speed. In this paper we propose Grouped Spatial Pack Convolutions (GSPC), a new implementation of grouped convolutions that outperforms existing solutions. We implement GSPC in TVM, which provides state-of-the-art performance on edge devices. We analyze a set of networks utilizing different types of grouped convolutions and evaluate their performance in terms of inference time on several edge devices. We observe that our new implementation scales well with the number of groups and provides the best inference times in all settings, improving the existing implementations of grouped convolutions in TVM, PyTorch and TensorFlow Lite by 3.4x, 8x and 4x on average respectively. Code is available at https://github.com/gecLAB/tvm-GSPC/, Comment: Camera ready version to be published at ASAP 2020 - The 31st IEEE International Conference on Application-specific Systems, Architectures and Processors. 8 pages, 6 figures

Symbolic holes are one of the fundamental building blocks of solver-aided and interactive programming. Unknown values can be soundly integrated into programs, and automated tools such as SAT solvers can be used to prove properties of programs containing them. However, supporting symbolic holes in a programming language is challenging; specifying interactions of holes with the type system and execution semantics requires careful design. This paper motivates and introduces the implementation of symbolic holes with unknown type to LLVM IR, a strongly-typed compiler intermediate language. We describe how such holes can be implemented safely by abstracting unsound and type-unsafe details behind a new primitive IR manipulation. Our implementation co-operates well with existing features such as type and dependency checking. Finally, we highlight potentially fruitful areas for investigation using our implementation., Comment: Accepted to TyDe 2020

Convolutional neural networks (CNNs) are used in many embedded applications, from industrial robotics and automation systems to biometric identification on mobile devices. State-of-the-art classification is typically achieved by large networks, which are prohibitively expensive to run on mobile and embedded devices with tightly constrained memory and energy budgets. We propose an approach for ahead-of-time domain specific optimization of CNN models, based on an integer linear programming (ILP) for selecting primitive operations to implement convolutional layers. We optimize the trade-off between execution time and memory consumption by: 1) attempting to minimize execution time across the whole network by selecting data layouts and primitive operations to implement each layer; and 2) allocating an appropriate workspace that reflects the upper bound of memory footprint per layer. These two optimization strategies can be used to run any CNN on any platform with a C compiler. Our evaluation with a range of popular ImageNet neural architectures (GoogleNet, AlexNet, VGG, ResNet and SqueezeNet) on the ARM Cortex-A15 yields speedups of 8x compared to a greedy algorithm based primitive selection, reduces memory requirement by 2.2x while sacrificing only 15% of inference time compared to a solver that considers inference time only. In addition, our optimization approach exposes a range of optimal points for different configurations across the Pareto frontier of memory and latency trade-off, which can be used under arbitrary system constraints.

Convolutional Neural Networks (CNN) are becoming a common presence in many applications and services, due to their superior recognition accuracy. They are increasingly being used on mobile devices, many times just by porting large models designed for server space, although several model compression techniques have been considered. One model compression technique intended to reduce computations is channel pruning. Mobile and embedded systems now have GPUs which are ideal for the parallel computations of neural networks and for their lower energy cost per operation. Specialized libraries perform these neural network computations through highly optimized routines. As we find in our experiments, these libraries are optimized for the most common network shapes, making uninstructed channel pruning inefficient. We evaluate higher level libraries, which analyze the input characteristics of a convolutional layer, based on which they produce optimized OpenCL (Arm Compute Library and TVM) and CUDA (cuDNN) code. However, in reality, these characteristics and subsequent choices intended for optimization can have the opposite effect. We show that a reduction in the number of convolutional channels, pruning 12% of the initial size, is in some cases detrimental to performance, leading to 2x slowdown. On the other hand, we also find examples where performance-aware pruning achieves the intended results, with performance speedups of 3x with cuDNN and above 10x with Arm Compute Library and TVM. Our findings expose the need for hardware-instructed neural network pruning., Comment: A copy of this was published in IISWC'19

Sparse linear algebra is central to many scientific programs, yet compilers fail to optimize it well. High-performance libraries are available, but adoption costs are significant. Moreover, libraries tie programs into vendor-specific software and hardware ecosystems, creating non-portable code. In this paper, we develop a new approach based on our specification Language for implementers of Linear Algebra Computations (LiLAC). Rather than requiring the application developer to (re)write every program for a given library, the burden is shifted to a one-off description by the library implementer. The LiLAC-enabled compiler uses this to insert appropriate library routines without source code changes. LiLAC provides automatic data marshaling, maintaining state between calls and minimizing data transfers. Appropriate places for library insertion are detected in compiler intermediate representation, independent of source languages. We evaluated on large-scale scientific applications written in FORTRAN; standard C/C++ and FORTRAN benchmarks; and C++ graph analytics kernels. Across heterogeneous platforms, applications and data sets we show speedups of 1.1$\times$ to over 10$\times$ without user intervention., Comment: Accepted to CC 2020

Recently, different machine learning methods have been introduced to tackle the challenging few-shot learning scenario that is, learning from a small labeled dataset related to a specific task. Common approaches have taken the form of meta-learning: learning to learn on the new problem given the old. Following the recognition that meta-learning is implementing learning in a multi-level model, we present a Bayesian treatment for the meta-learning inner loop through the use of deep kernels. As a result we can learn a kernel that transfers to new tasks; we call this Deep Kernel Transfer (DKT). This approach has many advantages: is straightforward to implement as a single optimizer, provides uncertainty quantification, and does not require estimation of task-specific parameters. We empirically demonstrate that DKT outperforms several state-of-the-art algorithms in few-shot classification, and is the state of the art for cross-domain adaptation and regression. We conclude that complex meta-learning routines can be replaced by a simpler Bayesian model without loss of accuracy., Comment: Advances in Neural Information Processing Systems (NeurIPS 2020, Spotlight)

Fast numerical libraries have been a cornerstone of scientific computing for decades, but this comes at a price. Programs may be tied to vendor specific software ecosystems resulting in polluted, non-portable code. As we enter an era of heterogeneous computing, there is an explosion in the number of accelerator libraries required to harness specialized hardware. We need a system that allows developers to exploit ever-changing accelerator libraries, without over-specializing their code. As we cannot know the behavior of future libraries ahead of time, this paper develops a scheme that assists developers in matching their code to new libraries, without requiring the source code for these libraries. Furthermore, it can recover equivalent code from programs that use existing libraries and automatically port them to new interfaces. It first uses program synthesis to determine the meaning of a library, then maps the synthesized description into generalized constraints which are used to search the program for replacement opportunities to present to the developer. We applied this approach to existing large applications from the scientific computing and deep learning domains. Using our approach, we show speedups ranging from 1.1$\times$ to over 10$\times$ on end to end performance when using accelerator libraries., Comment: Accepted to PACT 2019

Effective program synthesis requires a way to minimise the number of candidate programs being searched. A type signature, for example, places some small restrictions on the structure of potential candidates. We introduce and motivate a distilled program synthesis problem where a type signature is the only machine-readable information available, but does not sufficiently minimise the search space. To address this, we develop a system of property relations that can be used to flexibly encode and query information that was not previously available to the synthesiser. Our experience using these tools has been positive: by encoding simple properties and by using a minimal set of synthesis primitives, we have been able to synthesise complex programs in novel contexts, Comment: TyDe 2019 Extended Abstract

The desire to map neural networks to varying-capacity devices has led to the development of a wealth of compression techniques, many of which involve replacing standard convolutional blocks in a large network with cheap alternative blocks. However, not all blocks are created equally; for a required compute budget there may exist a potent combination of many different cheap blocks, though exhaustively searching for such a combination is prohibitively expensive. In this work, we develop BlockSwap: a fast algorithm for choosing networks with interleaved block types by passing a single minibatch of training data through randomly initialised networks and gauging their Fisher potential. These networks can then be used as students and distilled with the original large network as a teacher. We demonstrate the effectiveness of the chosen networks across CIFAR-10 and ImageNet for classification, and COCO for detection, and provide a comprehensive ablation study of our approach. BlockSwap quickly explores possible block configurations using a simple architecture ranking system, yielding highly competitive networks in orders of magnitude less time than most architecture search techniques (e.g. under 5 minutes on a single GPU for CIFAR-10). Code is available at https://github.com/BayesWatch/pytorch-blockswap., Comment: ICLR 2020