How to efficiently process sparse kernels

In this post, related research around sparsity will be covered.

  • SCNN, ISCA’17
  • ExTensor, MICRO’19


Keywords for GPGPU Memory Subscription

In this post, I would like to explain some keywords related to GPGPU Memory Subscription.


  • Thrashing
  • CPU Physical Memory
  • GPU Physical Memory
  • Unified Virtual Memory
  • Interplay between Hardware Prefetcher and Page Eviction Policy in CPU-GPU Unified Virtual Memory, ISCA’19
  • A Framework for Memory Oversubscription Management in Graphics Processing Units, ASPLOS’19

19 Fall semester

From this semester, I am starting my PhD of Computer Science studies at Georgia Tech majoring in computer architecture.

I will be taking CS6290, High Performance Computer Architecture. With the lecture, I will also watch some videos from the computer architecture course taught by Onur Mutlu, ETH Zurich. The lectures are available at Lectures.

TED talk-An astronaut's story of curiosity, perspective and change | Leland Melvin

The TED talk, An astronaut’s story of curiosity, perspective and change, Leland Melvin, is available at youtube.

I believe there are a lot of perspectives in the world, and it may not be possible to know all of them. What kinds of perpectives that I should have, or what kinds of perspectives are right(wrong)?

At least, it is trivial that to know more perspectives in the world, you need experience more beside what you are already familiar with.

Anaconda commands

In this post, I will explain some useful commands in Anaconda. Using Anaconda, you can manage various environments conveniently.


  • Make new environment with the specified name.

    conda create --name NAME python=3.6

  • Show all environments available.

    conda env list

  • Activate an environment with the specified name.

    source activate NAME

image captioning

For this winter, I will be working on image captioning.

To do list

  • Paper 1
    • Bottom-Up and Top-Down Attention for Image Captioning and Visual Question Answering
    • Motivation
      • While top-down visual attention mechanisms are used widely in image cationing and VQA, bottom-up mechanisms were not that famous.
      • The bottom-up mechanism (based on Faster R-CNN) will be able to provide more various image regions.
  • Paper 2
    • Recurrent fusion network for image captioning
    • Motivation
      • In image captioning, encoder-decoder frameworks are being used widely.
      • Existing frameworks only use one kind of CNNs. This limits the performance of whole framework to the performance of the base CNN.
  • Reference source code
    • Code
    • From ./model/, check how topdown core class and attention class work.

cs231n Stanford

Stanford cs231n 2017 lectures are available at Lectures.