Tutorial 4: Performance optimizations for CPU-based applications

Have feedback?Share it anonymously

Graph View

Section A: tutorial questions
Question 1: user CPU time
Question 2: MIPS vs. execution time
Question 3: speedup & theoretical maximum speedup
Question 4: effects of different compilations
Section B: optimization exercise
Part 1: matrix multiplication
Introducing the problem scenario: matrix multiplication
Tutorial structure
Performance metric
Maximum achievable FLOP/s
Version 1: optimize memory access pattern
Version 2: increase instruction-level parallelism (ILP)
CPU architecture
Optimization
Version 3: vector instructions
How to vectorize?
Version 4: improve data reuse
Memory hierarchy
Optimization
Part 2.1: histogram
Introducing the problem scenario: histogram
Tutorial structure
Performance metric
Maximum achievable bandwidth
Parallel histogram
Solution: parallel histogram
Work distribution
#pragma omp critical versus #pragma omp atomic
Private histogram
Performance comparison
Part 2.2: histogram with scatter
Introducing the problem scenario: histogram with scatter
Tutorial structure
Performance metric
Solution: histogram with scatter
Example
Part 2.3: radix sort
Introducing the problem scenario: radix sort
Tutorial structure
Radix sort
Parallel binary radix sort
Conclusion

Backlinks