Created by: sandyhouse

For prelu: f(x) = x if x > 0 else a*x, where a is a learnable parameter. so df(x) = 1 if x > 0 else a.

Todo：

• Replace the macro definition of CUDA_NUM_THREADS with the constant PADDLE_CUDA_NUM_THREADS;
• Remove the PADDLE prefix of the PADDLE_GET_BLOCKS function or define the function as a global function.