Heterogeneous parallel virtual machine: A portable program representation and compiler for performance and energy optimizations on heterogeneous parallel systems

Kotsifakou, Maria

Permalink

https://hdl.handle.net/2142/108496 Copy

Description

Title

Heterogeneous parallel virtual machine: A portable program representation and compiler for performance and energy optimizations on heterogeneous parallel systems

Author(s)

Kotsifakou, Maria

Issue Date

2020-07-15

Director of Research (if dissertation) or Advisor (if thesis)

Adve, Vikram S

Doctoral Committee Chair(s)

Adve, Vikram S

Committee Member(s)

• Adve, Sarita V
• Misailovic, Sasa
• Brooks, David M

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Virtual ISA
• Compiler
• Parallel IR
• Heterogeneous Systems
• GPU
• Vector SIMD
• Approximate Computing
• Deep Neural Networks
• DNNs
• Approximation Tuning

Abstract

"Programming heterogeneous parallel systems, such as the SoCs (System-on-Chip) on mobile and edge devices is extremely difficult; the diverse parallel hardware they contain exposes vastly different hardware instruction sets, parallelism models and memory systems. Moreover, a wide range of diverse hardware and software approximation techniques are available for applications targeting heterogeneous SoCs, further exacerbating the programmability challenges. In this thesis, we alleviate the programmability challenges of such systems using flexible compiler intermediate representation solutions, in order to benefit from the performance and superior energy efficiency of heterogeneous systems. First, we develop Heterogeneous Parallel Virtual Machine (HPVM), a parallel program representation for heterogeneous systems, designed to enable functional and performance portability across popular parallel hardware. HPVM is based on a hierarchical dataflow graph with side effects. HPVM successfully supports three important capabilities for programming heterogeneous systems: a compiler intermediate representation (IR), a virtual instruction set (ISA), and a basis for runtime scheduling. We use the HPVM representation to implement an HPVM prototype, defining the HPVM IR as an extension of the Low Level Virtual Machine (LLVM) IR. Our results show comparable performance with optimized OpenCL kernels for the target hardware from a single HPVM representation using translators from HPVM virtual ISA to native code, IR optimizations operating directly on the HPVM representation, and the capability for supporting flexible runtime scheduling schemes from a single HPVM representation. We extend HPVM to ApproxHPVM, introducing hardware-independent approximation metrics in the IR to enable maintaining accuracy information at the IR level and mapping of application-level end-to-end quality metrics to system level ""knobs"". The approximation metrics quantify the acceptable accuracy loss for individual computations. Application programmers only need to specify high-level, and end-to-end, quality metrics, instead of detailed parameters for individual approximation methods. The ApproxHPVM system then automatically tunes the accuracy requirements of individual computations and maps them to approximate hardware when possible. ApproxHPVM results show significant performance and energy improvements for popular deep learning benchmarks. Finally, we extend to ApproxHPVM to ApproxTuner, a compiler and runtime system for approximation. ApproxTuner extends ApproxHPVM with a wide range of hardware and software approximation techniques. It uses a three step approximation tuning strategy, a combination of development-time, install-time, and dynamic tuning. Our strategy ensures software portability, even though approximations have highly hardware-dependent performance, and enables efficient dynamic approximation tuning despite the expensive offline steps. ApproxTuner results show significant performance and energy improvements across 7 Deep Neural Networks and 3 image processing benchmarks, and ensures that high-level end-to-end quality specifications are satisfied during adaptive approximation tuning."

Graduation Semester

2020-08

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/108496

Copyright and License Information

Copyright 2020 Maria Kotsifakou

Owning Collections