__kernel void gpu_func(__global float* input, __global float* output)

     output[i] = input[i] * input[i];

__kernel

void gpu_func(__global float* input, __global float* output)

{

    int i = get_global_id(0);

    output[i] = input[i] * input[i];

     output[i] = input[i] * input[i];

__kernel void gpu_func(__global float* input, __global float* output)

     output = input * input;