Chinese
English
Assistant Research Fellow/Professor  |  Hong, Ding-Yong  
 
contact
vita
education
experience
interests
activities
invited_talk
honors
publications
software
supervised
 
 
 
 
 
Software
 
 
Software Name:Efficient Video Captioning on Heterogeneous System Architectures
Inventors:Horng-Ruey Huang, Ding-Yong Hong, and Jan-Jan Wu
Country:
Date:2021/05–/
Abstract:

Video captioning is a core technology to many important applications, such as video question answering, storytelling through videos, and lip-reading. Video captioning employs a hybrid CNN+RNN model. Accelerating such a hybrid model on a heterogeneous system is challenging because (1) CNN and RNN exhibit very different computing behaviors, (2) data dependency exists between the CNN and RNN within a video frame and adjacent RNNs across video frames; (3) effective utilization of accelerator resources. The key contributions of this project are as follows:

  • An efficient coarse-grained and fine-grained scheduling scheme for mapping computation and devices.
  • A pipeline scheduling scheme for exploiting maximum parallelism.
  • Two capacity-guided scheduling schemes for improving GPU utilization and minimizing off-chip memory overhead of the Edge TPU.
  • 3.24x speedup with CPU+GPU over the GPU-only execution on the sever platform.
  • 54.9x speedup with CPU+EdgeTPU over the ARM CPU-only execution on the edge platform. (59 frames per second)

Demo:   ARM-Only (FPS ~1)    ARM+EdgeTPU (FPS ~60)

 
Software Name:Exploiting Parallelism via Dynamic Binary Translation
Inventors:Ding-Yong Hong and Jan-Jan Wu
Country:
Date:2016/01–/
Abstract:

This project aims to develop a dynamic binary translation software to transform application binaries among different parallel execution models. The goal is (1) for legacy guest binary to fully utilize host hardware’s parallelism, and (2) to achieve execution migration among heterogeneous architectures. Current results show performance speedups of 1.59X/2.82X with the SIMD loop transformation from ARM NEON to x86 AVX2/AVX512.

 
Software Name:HQEMU: An Efficient and Retargetable Dynamic Binary Translator on Multicores
Inventors:Ding-Yong Hong, Chun-Chen Hsu and Jan-Jan Wu
Country:
Date:2014/08–/
Abstract:

HQEMU is an efficient and retargetable dynamic binary translator on multicores. It integrates QEMU and LLVM as its building blocks—the enhanced QEMU acts as a fast translator with low translation overhead, and the optimization-intensive LLVM optimizer running on separate threads dynamically improves code for higher performance. With the hybrid QEMU+LLVM approach, HQEMU can achieve low translation overhead and good translated code quality.

The official HQEMU website: http://csl.iis.sinica.edu.tw/hqemu/
Demo 1: QEMU vs. HQEMU
Demo 2: HQEMU user mode with 32 threads
Demo 3: HQEMU system mode with 8 VCPUs

 
 
bg