In my experience, "don't roll your own crypto" primarily means "do not reimplement a well-known algorithm, because you are likely to get it wrong and likely to leak side channels out the wazoo". I think most people have realized by now that inventing their own algorithm is a terrible idea. (Although certainly there are people who haven't realized that!)
I think that most people mean inventing own primitives by "don't roll your own crypto". Side channel attacks on software implementations on software implementations of cryptography are certainly practical, but often not significant for overall security of the whole system (as either there are simpler attack vectors or the system has to be already partially compromised for the attack to be possible).
Amounts of home-grown algorithms (pseudo-"asymmetric" algorithms optimized for simple hardware or small messages), misused algorithms (RC4 without IV, mostly), otherwise secure protocols implemented with parties transposed around (many proprietary RFID and other authentication token protocols authenticate reader against token, not other way around) and plain weirdness in many even newly designed commercial cryptosystems today is still more significant that any kind of sidechannels caused by careless implementation.
Implementing a well known algorithm can be a very good learning experience! I implemented AES-256 for my college project this semester. Learned a whole lot about Cryptography that I never know before. :)
