Hashim Sharif

  • Advisor:
      • Vikram Adve (PhD), Vikram Adve and Sasa Misailovic (Postdoc)
  • Departments:
  • Areas of Expertise:
      • Compilers
      • Accuracy-aware Optimization
      • Static Analysis
  • Thesis Title:
      • ApproxHPVM: A Retargetable Compiler Framework for Accuracy-aware Optimizations
  • Thesis abstract:
      • With the increasing need for machine learning and data processing near the edge, compilers can play a crucial role in bridging the gap between the computational demands of emerging workloads and the limited computational capabilities of current edge systems. To enable energy-efficient computing at the edge, I present ApproxHPVM, a portable optimizing compiler and runtime system that allows flexible, optimized use of multiple software and hardware approximations in a unified, easy-to-use framework. ApproxHPVM takes an end-to-end quality constraint input from developers and automatically maps individual IR operations to software and hardware approximations. ApproxHPVM uses a hardware-agnostic approximation-tuning phase to do this translation that provides greater portability across heterogeneous hardware platforms. ApproxHPVM incorporates: (a) a compiler IR with high-level tensor operations b) a flexible and efficient approximation-tuning framework called ApproxTuner that selects approximation knobs on a per-operation level, c) backend transforms that facilitate compilation to heterogeneous hardware targets, and d) a runtime that dynamically tunes approximation knobs in response to system level changes such as slowdowns and changing accuracy constraints. We evaluate ApproxHPVM on 10 deep learning benchmarks and 5 image processing benchmarks. Leveraging approximation knobs on a specialized analog processing hardware accelerator with configurable voltage knobs, ApproxHPVM provides speedups ranging from 1-9x and energy reductions ranging from 1-11x. ApproxHPVM can map to lower precision compute and software approximations on GPU, providing a mean speedup of 2.1x and a mean energy reduction of 2x across benchmarks. The ApproxTuner autotuning framework uses novel tuning techniques that speed up tuning times by 12.8x compared to conventional empirical tuning while achieving comparable benefits. The ApproxTuner runtime can also dynamically tune approximation knobs to counteract slowdowns imposed by changing system conditions such as frequency modes.
  • Downloads:

    Contact information:
    hsharif3@illinois.edu