MLIR: Scaling Compiler Infrastructure for Domain Specific Computation

Lattner, Chris; Amini, Mehdi; Bondhugula, Uday; Cohen, Albert; Davis, Andy; Pienaar, J.; Riddle, River; Shpeisman, Tatiana; Vasilache, Nicolas; Zinenko, Oleksandr

doi:10.1109/cgo51591.2021.9370308

Cited by 194 publications

(66 citation statements)

References 42 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In this work, we introduce the Quantum Intermediate Representation for Optimization (QIRO), an IR for universal quantum computation that leverages MLIR [21] to support quantum-classical co-optimization. In contrast to existing IRs for quantum computing, we design our optimization dialect in a way such that data dependencies are explicit for both quantum and classical variables.…”

Section: Uantum Multi-level Irmentioning

confidence: 99%

See 1 more Smart Citation

QIRO: A Static Single Assignment-based Quantum Program Representation for Optimization

Ittah

Häner

Kliuchnikov

et al. 2022

ACM Transactions on Quantum Computing

View full text Add to dashboard Cite

We propose an IR for quantum computing that directly exposes quantum and classical data dependencies for the purpose of optimization. The Quantum Intermediate Representation for Optimization (QIRO) consists of two dialects, one input dialect and one that is specifically tailored to enable quantum-classical co-optimization. While the first employs a perhaps more intuitive memory-semantics (quantum operations act on qubits via side-effects), the latter uses value-semantics (operations consume and produce states) to integrate quantum dataflow in the IR’s Static Single Assignment (SSA) graph. Crucially, this allows for a host of optimizations that leverage dataflow analysis. We discuss how to map existing quantum programming languages to the input dialect and how to lower the resulting IR to the optimization dialect. We present a prototype implementation based on MLIR that includes several quantum-specific optimization passes. Our benchmarks show that significant improvements in resource requirements are possible even through static optimization. In contrast to circuit optimization at run time, this is achieved while incurring only a small constant overhead in compilation time, making this a compelling approach for quantum program optimization at application scale.

show abstract

Section: Uantum Multi-level Irmentioning

confidence: 99%

“…In contrast to the rigidity of LLVM, MLIR [21] provides an extensible representation with infrastructure for transformations, analysis, and debugging. There exists a variety of mechanisms to enable and manage extensibility in MLIR.…”

Section: Mlirmentioning

confidence: 99%

QIRO: A Static Single Assignment-based Quantum Program Representation for Optimization

Ittah

Häner

Kliuchnikov

et al. 2022

ACM Transactions on Quantum Computing

View full text Add to dashboard Cite

show abstract

“…To bridge the semantic gap between high-level language and low level Intermediate Representations (IRs), we leverage the MLIR framework. MLIR has been proposed for both reusability and extensibility [21] and allows intergration of multiple IRs with different level of semantics at the same time.…”

Section: B Multi-level Intermediate Representation (Mlir)mentioning

confidence: 99%

“…Maintaining all these compiler frameworks and porting each of them to any new architecture are challenging tasks, which may limit the scope of each language to a limited number of target architectures. The MLIR framework addresses this fragmentation problem by proposing a modular and reusable IR stack that sits in between the language representation and the architectural representation [21]. In this way, architecturalspecific operations and types can be encapsulated in specific IRs, while sharing common operations, types, and optimizations across languages and target architectures.…”

Section: B Multi-level Intermediate Representation (Mlir)mentioning

confidence: 99%

“…Also, these abstractions are generic enough to be used for future cost models and mappers. Our ecosystem leverages the recently introduced MLIR infrastructure [21] to integrate different high-level languages or frameworks, such as Tensorflow, PyTorch for ML, and COMET [22] for HPC, with mappers and cost models. To the best of our knowledge, Union is the first framework unifying multiple high-level frameworks for tensor computations, mappers, and cost models for spatial accelerators.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Union: A Unified HW-SW Co-Design Ecosystem in MLIR for Evaluating Tensor Operations on Spatial Accelerators

Jeong¹,

Kestor²,

Chatarasi³

et al. 2021

Preprint

View full text Add to dashboard Cite

To meet the extreme compute demands for deep learning across commercial and scientific applications, dataflow accelerators are becoming increasingly popular. While these "domain-specific" accelerators are not fully programmable like CPUs and GPUs, they retain varying levels of flexibility with respect to data orchestration, i.e., dataflow and tiling optimizations to enhance efficiency. There are several challenges when designing new algorithms and mapping approaches to execute the algorithms for a target problem on new hardware. Previous works have addressed these challenges individually. To address this challenge as a whole, in this work, we present a HW-SW codesign ecosystem for spatial accelerators called Union 1 within the popular MLIR compiler infrastructure. Our framework allows exploring different algorithms and their mappings on several accelerator cost models. Union also includes a plug-and-play library of accelerator cost models and mappers which can easily be extended. The algorithms and accelerator cost models are connected via a novel mapping abstraction that captures the map space of spatial accelerators which can be systematically pruned based on constraints from the hardware, workload, and mapper. We demonstrate the value of Union for the community with several case studies which examine offloading different tensor operations (CONV/GEMM/Tensor Contraction) on diverse accelerator architectures using different mapping schemes.

show abstract

Multigraph learning for parallelism discovery in sequential programs

Shen

Peng

et al. 2023

Concurrency and Computation

View full text Add to dashboard Cite

Summary Parallelization is an optimization technique that is playing an increasingly vital role in software applications. Discovering the potential parallelism in sequential programs is of primary importance for parallelization. To this end, various tools have been created to obtain parallelizable targets by analyzing the dependencies in sequential programs. However, the majority are constructed based on manually designed dependence analysis rules. Consequently, they must follow some constraints, which limit the scope of their program analysis capability. Furthermore, constructing a dependency analysis rule is a complicated and highly sophisticated endeavor. In this study, we exploit the fact that neural network models can learn the potential features of programs to tackle the parallelism discovery task in an end‐to‐end manner. Specifically, we developed a multigraph learning architecture on top of multiple abstract code representations in a complementary manner. In the architecture, we equip code representations with different neural network models in a targeted manner, including a deep convolutional neural network for analyzing both control and data flows and a novel neural network for learning the abstract syntax trees. Experimental results obtained on a common parallelism discovery dataset indicate that the developed multigraph learning architecture can learn potential parallelism patterns in sequential programs with high accuracy and efficiency.

show abstract

MLIR: Scaling Compiler Infrastructure for Domain Specific Computation

Cited by 194 publications

References 42 publications

QIRO: A Static Single Assignment-based Quantum Program Representation for Optimization

QIRO: A Static Single Assignment-based Quantum Program Representation for Optimization

Union: A Unified HW-SW Co-Design Ecosystem in MLIR for Evaluating Tensor Operations on Spatial Accelerators

Multigraph learning for parallelism discovery in sequential programs

Contact Info

Product

Resources

About