Lecture 23

Learning objectives

After this class, you should be able to:

1. Given an algorithm or code, identify possibility of performance penalty due thread divergence and change the algorithm or code to reduce its performance impact.
2. Describe how the arithmetic on gpu.cs.fsu.edu differs from the IEEE standard, and compare the level of support with that on the PS3 and with SSE.

Reading assignment

1. GPU-6 on Blackboard, under the "course library" tab.

Exercises and review questions

• Exercises and review questions on current lecture's material

  1. Write CUDA code with thread divergence and check PTX code to see if predicated execution is used.

  2. Compare the difference in results of some mathematical operations using fast math on CUDA, the PS3, and a conventional processor, using single and double precision on the latter. Post your answers on the discussion board.

• Preparation for the next lecture
  1. None. Work on optimizing the code for your group project.

---

Last modified: 21 Apr 2010