在openCL中优化分支for循环

Question

I have a kernel code something like this

__kernel fn(***){
    //X,Y would be image cordinates
    int x = get_global_id(0);
    int y = get_global_id(1);

    //Initialize pixel value
    int c =  -5 + x * dx;
    int d =  -5 + y * dy;

    int k=0;
    for(; k< 500; k++){
        //Perform Some Calculations using c and d
        //Most of the calculations happen here
        if(val > threshold)
            break;
    }
    //Write data based on k
    out[x*width+j] = k;
}

I've a feeling that as most of the calculations happens inside the for loop, and as the for loop creates a branch, some of the work items in a work group complete their kernel execution and wait for the entire work group to complete.

How can this be optimized if the output is based on the execution counter k?

Answer 1

The for loop will have a branch even if you remove that

if(val > threshold)
    break;

It will be generated by the compiler to see if the loop should be continued or not. Though we can remove the extra branch created inside the for loop.

k += static_cast<int>(val > threshold) * 500;

This will increase k by 500 if val > threshold and therefore quit the loop in the same branch that checks if k has reached the desired value, without an extra branch. Depending on how heavy the calculation inside the loop is, this may not matter.