Speedup

• Speedup

• Unrolling

• Cache Blocking

There are many ways to speedup a program. We’ll use matrix operations as the classical example. This section will not cover parallel processing.

Unrolling

Optimize the use of independent instruction to accumulate as much binary size (registers) as possible.

result = 0
for i in range(0, len(arr), 4):
    result += arr[i]  
    result += arr[i+1]
    result += arr[i+2]
    result += arr[i+3]

Cache Blocking

Optimize temporal locality of cache by changing the way we access data.

• For NxN matrix operations one method is to split the matrices into sub-matrix