This avoids unecessary complexity. Also makes the GPUImmediate threadsafe by using a threadlocal imm variable.