Introduction to the PowerVR Compute Development Recommendations

The PowerVR family of Graphics Cores from Imagination Technologies is designed to perform both compute and graphics tasks. Its programmable core architecture allows for extremely efficient and high-performance compute execution.

Document Overview

This document describes the recommended usage guidelines for achieving optimal performance when using compute on Imagination’s PowerVR Graphics Cores. Most of the guide describes constructs and patterns that directly emerge from the PowerVR cores architecture. This enables the developer to make the correct decisions irrespective of the preferred approach as far as APIs and programming languages are concerned. Additionally, several specific details are included for OpenCL, OpenGL, and OpenGL ES Compute. These details have been written against the following API versions:

• OpenGL ES 3.1

• OpenGL 4.3

• OpenCL 1.x and 2.0

This document assumes that the reader has a good working knowledge of at least one of these APIs. It is also assumed that the reader has worked through the examples of the relevant PowerVR Compute SDK.

After reading this document, readers should have a solid understanding of how compute works on PowerVR Graphics Cores, as well as a good understanding of how to develop efficient and well-optimised code for these devices.

This document also provides an optimisation strategy quick reference sheet.

Glossary

Throughout this document, extensive use is made of terminology identified in the table below.

Common terms

Term	Also referred to as	Description
USC (Unified Shading Cluster)	Shading Cluster, Shading Unit, Execution Unit	A semi-autonomous part of the Graphics Core that can typically execute an entire work-group. Other large parts like Texture Units can be shared among USCs.
Core	Processor, Graphics Core	An almost completely autonomous part of the Graphics Core. Typically, a collection of USCs and possibly supporting hardware such as texture units.
Task	Thread Group, Warp, Wavefront	The native grouping of threads that a USC executes. Number of threads varies per PowerVR Core.

API-specific terms

Term	OpenGL and OpenGL ES	OpenCL	RenderScript
Kernel / shader	Compute Shader	Kernel	Kernel
Thread	Thread, shader instance	Work-item	Kernel invocation
Work-group	Work-group	Work-group	N/A
Shared memory	shared variables	local memory	N/A
Image	Texture, image	Image	Image
Constant, constant memory	const / uniform variable, uniform block, uniform buffer	constant memory	Constants
Private memory	Local variables, temporaries	Variables, private memory	Local variables, temporaries
Dataset	Dispatch size, dataset	Global work, ND Range	Dataset

Architecture Types

Architecture Type	Core
Type 1	6/6XT, 7/7XT, 8XE/XEP, 9XE/XM
Type 2	AXE, AXM, BXE, BXM, BXS (64 and Smaller)
Type 3	9XTP
Type 4	AXT/BXT, BXS (256 and Larger)