When OpenCL device gets a kernel from command queue, the kernel should be target machine code, not IR(intermediate representation), right? otherwise, If it’s IR, device needs to support runtime to decode IR and generate machine code, which will impact performance.