Post-PASI training: Week 4

Syllabus

Class 9 (August 29th):

Lecture 9 – Slides

• Challenges of GPU implementations in CFD codes
• The role of sparse matrices in CFD:
  • Importance of implicit schemes
• Explanation of the 2D implicit heat transfer using Crank Nicolson:
  • Getting from the mesh to the matrix

Lab 7 (August 30th):

Lab 7 – Slides

Lab7 – code

FD_2D_cusp.cu (shared memory version)

FD_2D_cusp_global.cu (global memory version)

(See global memory version first, it’s easier to understand)

• Implementation of 2D implicit heat transfer
• Use 2D stencil codes and Cusp

• References:
  • Layton S., Krishnan A. and Barba L. cuIBM: A GPU Accelerated Immersed Boundary Method. ParCFD\’11.
  • Cohen, J. and Molemaker, J. A Fast Double Precision CFD Code using CUDA. Parallel CFD, 2009.

Class 10 (August 31st):

Lecture 10 – Slides

• N-body simulation with GPU
• Example of applications:
  • Astrophysics
  • Surface reconstruction with RBF
  • Others (MD)

• Revisit optimizations:
  • Shared memory use and tiling
  • Loop unrolling and other optimization techniques

Lab 8 (September 1st):

Lab 8 – Slides

lab8_nbody.cu

nbody_gpu.cu (N body tiled version)

nbody_gpu_2.cu (N body with N not a multiple of threads per block)

nbody_optimized.cu (N body with bigger tiles)

• Implementation of a fast N-body simulation

• References:
  • Nyland, Harris, Prins. “Fast N-body Simulation with CUDA”. GPU Gems 3, Chapter 31.