I don't think you can do it. If you look at the timing you have to be accurate to 100nS and there simply isn't enough time in a 16MHz processor to do that, it turns out to be about two clock cycles. Even in machine code the overhead of reading data and producing those two diffrent waveforms times is too tight.
I think you will have to have a diffrent processor, one that runs faster. Or a very fancy piece of hardware.