Quoting Bruce Schneier:
Anyone can design a cipher that he himself cannot break. This is why you should uniformly distrust amateur cryptography, and why you should only use published algorithms that have withstood broad cryptanalysis. All cryptographers know this, but non-cryptographers do not. And this is why we repeatedly see bad amateur cryptography in fielded systems.
'Rolling your own crypto' means you designing a new algorithm that you then proceed to use without either:
By implementing a known and trusted process or algorithm like pbkdf2, you have some level of guarantee that it has been audited and will not keel over and die the minute someone prods the algorithm itself. By using a widely accepted standard, you have some level of assurance that the algorithm is not going to be flawed and you only need to ensure that your implementation is secure. This is not rolling your own crypto. This is you implementing it and if your implementation is correct and you follow the guidelines, you're golden.
Further reading that might help: Schneier on the subject