Heterogeneous Computing with OpenCL, 2nd Edition

Chapter 9

OpenCL Case StudyHistogram

Introduction

This chapter discusses specific optimizations for a memory-bound kernel. The kernel we choose for this chapter is an image histogram operation. The source data for this operation is an 8-bit-per-pixel image with a target of counting into each of 256 32-bit histogram bins.

The principle of the histogram algorithm is to perform the following operation over each element of the image:

 for(many input values) {

   histogram[value]++;

 }

This algorithm performs many scattered read-modify-write accesses into a small histogram data structure. On a CPU, this application will use the cache, although with a high rate of reuse of elements. On a GPU, these accesses will be resolved in global memory, which will ...