Motivation: an example

 

• We will begin by looking at matrix multiplication.
• We will study the performance of a double-precision floating point matrix multiply.
• We will investigate the performance of various versions in order to determine what transformations a compiler should be applying.

• Initial version
• Architectural details
• Memory system
• Performance
• Accumulate into a scalar
• Interchange loops
• ...
• Optimising memory accesses
• Blocking (a.k.a. ``tiling'')
• ...
• Performance of blocked version
• Summary
• Moral
• ...