opencl workitem run parallel - optimization

Question

asking about speed or optimize the code
the kernel for sobel edge detection for gray img
When I run the program without any process only show input video and output(same as input) the frame per secounds fps=70 but when process down to 20 (process using GPU kernel for sobel)
Does anyone have an idea of how to speed up this code? I used local memory instead of global memory but the change is small.
How can I make all work items process the image?
sobel kernel
__kernel void hello_kernel(const __global uchar *input, __global uchar *output,const uint width,const uint height)
{
int x = get_global_id(0);
int y = get_global_id(1);
int index = width * y + x;
float a,b,c,d,e,f,g,h,i;
float8 v;
float sobelX = 0;
float sobelY = 0;
//if(index > width && index < (height*width)-width && (index % width-1) > 0 && (index % width-1) < width-1){
a = input[index-1-width] * -1.0f;
b =input[index-0-width] * 0.0f;
c = input[index+1-width] * +1.0f;
d = input[index-1] * -2.0f;
e = input[index-0] * 0.0f;
f = input[index+1] * +2.0f;
g = input[index-1+width] * -1.0f;
h = input[index-0+width] * 0.0f;
i = input[index+1+width] * +1.0f;
sobelX = a+b+c+d+e+f+g+h+i;
a = input[index-1-width] * -1.0f;
b = input[index-0-width] * -2.0f;
c = input[index+1-width] * -1.0f;
d = input[index-1] * 0.0f;
e = input[index-0] * 0.0f;
f = input[index+1] * 0.0f;
g = input[index-1+width] * +1.0f;
h = input[index-0+width] * +2.0f;
i = input[index+1+width] * +1.0f;
sobelY = a+b+c+d+e+f+g+h+i;
output[index] = sqrt(pow(sobelX,2) + pow(sobelY,2));
}