University of Pennsylvania, CIS 565: GPU Programming and Architecture, Project 1 - Flocking

• Jacky Lu
  • LinkedIn

README

Result:

• Naive Implementation (5,000 Boids & 10,648 Cells)

• Uniform Grid Implementation With Scattered Data [27 Neighboring Cells Search] (5,000 Boids & 10,648 Cells)

• Uniform Grid Implementation With Scattered Data [8 Neighboring Cells Search] (5,000 Boids & 10,648 Cells)

• Uniform Grid Implementation With Coherent Data [27 Neighboring Cells Search] (5,000 Boids & 10,648 Cells)

• Uniform Grid Implementation With Coherent Data [8 Neighboring Cells Search] (5,000 Boids & 10,648 Cells)

• Uniform Grid Implementation With Coherent Data [8 Neighboring Cells Search] (20,000,000 Boids & 64,964,808 Cells)

• P.S. I wasn't able to perform much performance analysis due to the lack of time because I have been trying to find the bug that crashes my program for the entire weekend. However, as a result of the lengthy debugging process, I was able to have a lot of practice and hands-on experience with the Nsight debugger. I find the Nsight debugger very helpful. By tracking global buffer's data in Warp Watch, I was able to locate the unusual data pattern in dev_particleArrayIndices, and ultimately found out that I freed the buffers' device memories at the wrong place in the program.