Lecture 13

Learning objectives

After this class, you should be able to:

1. Given an algorithm, derive its cache complexity under the ideal cache model.
2. Given nested loops, reorder them to improve cache performance.
3. Given a problem, develop a cache aware algorithm for it.
4. Given a problem, develop a cache oblivious algorithm for it.

Reading assignment

1. Read the document on cache aware and cache oblivious algorithms under the course library tab on Blackboard.

Exercises and review questions

• Exercises and review questions on current lecture's material

  1. Note the small difference in loop order, and some related optimizations, between Lec13/CAMM1D.c and Lec13/CAMMRowMaj.c. Make other changes to your matrix multiplication code and report the performance on the discussion board.

  2. Search the following assembly codes: Lec13/CAMM1D.s, Lec13/CAMMRowMaj.s, and Lec13/MMBLAS.s for the SSE SIMD instruction for single precision multiplication: mulps. Explain why the first code has worse performance than the second one and why the second one has worse performance than the first. (Note: You can disassemble executables on Linux using objdum -d, in case you wish to perform similar analysis on other executables.)

• Preparation for the next lecture
  1. Change the number of threads in the Lec13/CAMM_omp.c code and report, on the discussion board, the performance as the number of threads and the matrix size change.

---

Last modified: 22 Feb 2010