GPU-based Parallel Implementation of Swarm Intelligence Algorithms

Chapter 5

Implementation Considerations

Contents

5.1 Float-Point  57

5.1.1 Instabilities   57

5.1.2 Performance   58

5.2 Memory Accesses  58

5.2.1 Device Memory   59

5.2.2 Global Memory   59

5.2.3 Shared Memory   60

5.2.4 Read-Only Memory   61

5.3 Random Number Generation  62

5.4 Branch Divergence  62

5.5 Occupancy  62

5.6 Summary  62

5.1 Float-Point

Float-point (FP) arithmetic is ubiquitous in computer systems. However, it has been known that it is erroneous to compute with floating-point numbers and operations due to the finite precision.

5.1.1 Instabilities

FP suffers some well-known numerical instabilities [65]:

• Rounding generates an error.

• Non associativity. For instance, (a + b) + c and a + (b + c) are not necessarily identical.

•