In the section Histogram calculation in Chapter 3, Buffers and Image Objects, we discussed about the naive implementation of histogram computation of an image. We read an input image file and pass the pixel buffer to the OpenCL device to compute the histogram of the image. By now you must have observed that this implementation is not so optimized which involves sequential reads. In this section we will try to optimize this implementation by making use of atomic_inc OpenCL C built-in and make use of coalesced reads and writes to the global and local memory. Take a look at the following kernel:
#define BIN_SIZE 256
#define ELEMENTS_TO_PROCESS 256
__kernel
void histogram_kernel(__global const uint* data,
__global uint* binResultR,
__global uint* binResultG,
__global uint* binResultB)
{
__local int sharedArrayR[BIN_SIZE];
__local int sharedArrayG[BIN_SIZE];
__local int sharedArrayB[BIN_SIZE...