This is a bad idea for different reasons. I'll short-list them and then go into details.
Your distribution isn't random. X and Y values are constrained to screensize, accordingly different devices have different possible randomness distributions, which is a "bad thing"™
User input is not random. "Guess a random number between 1 and 10" is not random and neither is "type on the screen randomly". Humans are really bad at faking randomness.
The x and y value constraints are a bad thing in the first place. Smaller devices have less entropy to generate, larger devices have repetition when the x and y coordinates "repeat", since you truncate any value larger than 8 bits.
Luckily screen-sizes are usually a nice multiple of \$256\times256\$ "tiles". The actually useful entropy you can currently extract accordingly is in one such tile.
Unfortunately this also means that "randomly" tapping similar areas in these tiles (as probably many non-technical people would do) will produce somewhat predictable bytes.
This directly brings us to the next point. The "Blue Seven Phenomenon". There's a metric ton of academic articles about the subject, but the bottom line is: Humans are bad at pretending to be random.
The brain is trained evolutionarily (and throughout our whole lives) to recognize patterns. True randomness doesn't actually exhibit those patterns. Any patterns we see in randomness are just ... imagination. Well not quite, but there's no actual meaning behind them.
The problem is when you ask users to provide "random" input, their brain will simply not be able to provide it, and instead you'll get a more or less pseudorandom pattern.
Long story short, you'll not receive something that I'd deem cryptographically secure.