OpenCL template matching with larger template is slower than OpenCV CPU version - opencl

Question

I am new to opencl, now I am workin on a optimisation of template matching with OpenCL. I have done some experiements with smaller templates and found that the my OpenCL implementation is faster than OpenCV's CPU version. But in this particular case the template size is really big(2048x2048) and the original image size is ( 3072x3072), OpenCV cpu implementation(137 seconds) is far ahead ahead of OpenCL( 2000 seconds ). Kindly suggest some way to optimise my code showed below.
void __kernel corrln(global const unsigned char* ref_image, global const
unsigned char* template, global float* corrln )
{
const uint Width = get_global_size(0);
const int2 pos = {get_global_id(0), get_global_id(1)};
float sum = 0;
for(int y = pos.y; y < 2048; y++ )
{
for(int x =pos.x; x < 2048; x++ )
{
const int2 xy = { x, y };
const int2 txy = { x - pos.x, y - pos.y };
sum += ref_image[index(xy, Width)] * template[index(txy,
2048)];
}
}
corrln[index(pos, Width)]= sum;
}

Answer 1

Considering your ref_image to be of reasonable size less than 2048 (say, 1024x1024), and ND size to be equal to ref_image size, every WI (Work Item) is doing different amount of calculations.
WI with pos.x == 0 & pos.y == 0 is doing 2048 * 2048 = 4M calculations within 2 loops, WI with pos.x == 1023 & pos.y == 1023 is doing 1024 * 1024M calculations within 2 loops. That's too much job for single WI.
Try to dice up this task in such fashion that every WI will do some reasonable fixed amount of calculations. Say, for 1st column of ref_image, do multiple launches of kernel, each of which will process 16 columns to the right and calculate & accumulate corrln array and then go to 2nd column, etc.
Kernel might look like this (just for illustration!!!):
void __kernel corrln(
global const unsigned char* ref_image,
global const unsigned char* template,
global float* corrln )
{
const uint Width = get_global_size(0);
const int2 pos = {get_global_id(0), get_global_id(1)};
uchar16 ref = vload16(index(xy, Width), ref_image);
uchar16 tpl = vload16(index(xy, Width), template);
float sum = corrln[index(pos, Width)] + dot(ref, tpl);
corrln[index(pos, Width)]= sum;
}