__kernel void l2_normalize(__global uchar *src_ptr, uint src_stride_x, uint src_step_x, uint src_offset_first_element_in_bytes, __global uchar *sum_ptr, uint sum_stride_x, uint sum_step_x, uint sum_offset_first_element_in_bytes, __global uchar *dst_ptr, uint dst_stride_x, uint dst_step_x, uint dst_offset_first_element_in_bytes, DATA_TYPE epsilon)
This kernel performs reduction given an operation.