PathFinder on CUDA

1
PathFinder on CUDA

• James Doyle
• Advisor Prof. John MacCormick

2
Background

• Pathfinder is an image segmentation algorithm,
  uses superpixels
• Ill be working on
  improving performance.

3
My Focus Performance

• With greater speed, PathFinder could be used in
  real-time video processing
• Performance increases will not come from improved
  code
• Instead, we'll use

4
CUDA

• Compute Unified Device Architecture
• C-based programming environment
• Allows developers to utilize GPU for any
  processing tasks

5
Changes to Project

• Clarification not a complete reimplementation
• Instead, use Java Native Interface (JNI)
• Allows programmer to call non-Java code from java
  applications, should allow us to outsource to
  CUDA and the GPU

6
Potential Problems Update

• Hardware/Lab setup
• Hello, world.
• Speed of memory access
• Working with C

7
Time Analysis

• Conducted analysis of PathFinder to investigate
  potential performance gains from CUDA

8
Time Analysis (cont)

• Used conservative improvement estimate

9
Next Steps

• Utilize JNI to improve one segment of algorithm
• Image transpose
• Create or adapt pre-existing CUDA code for other
  segments
• Copy the image
• Calculate edge strengths
• Calculate grid paths
• Tests!

10
Literature Review

• Research into superpixel algorithms
• GPU processing research
• A couple specific to GPU and computer vision
• CUDA implementations in research