Project #1: High Performance Dense Matrix Multiplication

Objectives

• code generation techniques
• cache oblivious algorithms
• cache aware techniques
• analysis techniques for HPC
• empirical analysis methods

Methods

• The non-optimal naive matrix-matrix multiply
• A blocked cache aware matrix-matrix multiply as described in the Cache Oblivious Paper
• An optimized cache oblivious algorithm with generated optimized base cases and a planner that optimizes the recursive assembly of the base cases.

Measure the performance of these algorithms on any candidate architecture. Describe your measurement methodology: How many measurements did you take? What was the variation in the measurements? What architecture did you use? What properties did this architecture have and does it match your model? Use statistical techniques if necessary.

In your analysis, provide the asymptotic analysis of the algorithms work and cache complexity. Compare these models to your analysis. Explain the behavior you see, to the extent that it is applicable, using the models developed in the analysis.

Group Project

What do you turn in?

• A research report documenting your work. Include references where necessary. This report should document: analysis of algorithms, software techniques used, experimental methods and results, comparison of analysis to experimental results. A hardcopy of this report will be turned in during class time.
• The software you created. Include all source files and makefiles and an instruction file on how to compile and run. Do not include binaries or .o files! Use tar to create a single file and send this file as an attachment to luke@cse.msstate.edu. The project source code is due by 2pm on the due date.

Due Date:

Examples: