Performance Guidelines

PowerVR Graphics Cores are capable of efficiently executing a range of parallel processing algorithms. To achieve the best performance, it is important to follow the guidelines throughout this section as closely as possible. These guidelines are based on four main strategies:

• Execute Tasks on the Most Appropriate Processor

• Minimise Bandwidth Usage

• Maximise Utilisation and Occupancy

• Maximise Instruction Throughput

For all these strategies, it is always a good idea to profile applications to find out where the bottlenecks in processing are. This will work as a guide as to where to target optimisation efforts and can be used to determine the results of any given optimisation. The PowerVR SDK provides PVRTune as a real-time hardware profiler which supports the visualisation of how the Graphics Core is being used.

Method	OpenCL	OpenGL	OpenGL ES	RenderScript
Create Allocations with Correct USAGE Flags	✗	✗	✗	✓
Create Shared Memory Objects with CL-ALLOC-HOST-PTR	✓	✗	✗	✗
Create Buffers with the Correct USAGE Hints	✗	✓	✓	✗
Use Zero-Copy Paths if Possible	✓	✓	✓	✗
Use Shared/Local Memory Sensibly	✓	✓	✓	✗
Avoid Barrier Directly after Local or Constant Memory Access	✓	✓	✓	✗
Use Built-ins for Type Conversions	✓	✓	✓	✗