Please don't go inventing your own cryptography when it can be avoided.

Scrypt did the work to write up an extensive analysis of their effort, and it was reviewed by many people at great cost. And even then the result has not been totally criticism free (http://eprint.iacr.org/2013/525).

With a quick glance at the shape of your implementation, it could have trivially failed to be memory hard at all had you appended instead of prepending. I expect that it's actually spending all of its time shuffling around strings, and that a gpu implementation of it would still end up being a zillion times faster...