ITK/Release 4/GPU Acceleration/Tcon-2010-11-22

Attendees

• Won Ki - Harvard University
• Joe Stem - NVIDIA
• Kimberly Powell - NVIDIA
• Dennis Sessanna - NVIDIA
• Luis Ibanez - Kitware Inc.

Topics

Overview

• Quick summary of ITKv4 effort (Luis Ibanez)
• Summary of ITK-GPU approach (Won Ki)

Questions

• Level of abstraction ?
  • Joe suggests to look at OpenCV
    • Expose the interactions with the GPU
    • Most GPU programmers do things synchronously (so they unfortunately do too many data transfers, and don't get full benefit from the GPU).
• Joe asked for typical Use Cases
  • We listed:
    • Radiology : 100Mb per image (512x512x200)
    • Microscopy : 10Gb
    • Video : 10Mb images, 30~100 frames per second.
• CUDA vs OpenCL ?
  • Joe answers
  • OpenCL is better for asynchronous multi-GPU programming.
  • Reasons for using CUDA over OpenCL
    • Tedious API in OpenCL
    • Large collection of CUDA existing libraries
    • Performance optimization may be harder in OpenCL
  • Luis asked about vendor's commitment to OpenCL (for next 5 ~ 10 years)
    • Joe answers
      • NVIDIA supports the OpenCL standard (so, what is in OpenCL will be supported in NVidia cards)
      • Some third party vendors are doing CUDA for NVidia platforms and OpenCL for other platforms (splitting the effort)
      • Translation from CUDA to OpenCL is straight forward. (having a dual implementation may be lower than twice the effort)
      • Other options CUDA x86 compiler coming up (commercial product)
      • Translator from CUDA to OpenCL (need to find links to it)