Using high-level, no-brainer parallel constructs is kosher (like a pre-built, concurrent data structures lib built by odd people like Doug Lea), similar to the advice about cryptography. Playing with the atoms yourself is dangerous. You can understand the concepts well, be contentious in your coding and build something the right person can drive a truck through.

How do you verify correctness? How do you test? A bug might only present one in ten billion loops, perhaps when you push the cores really hard. Or fail on the new processors with the slightly different barrier/cache/coherency semantics. Or fail in ways that send you down a million heisenbug gullies. Without any proper tools or diagnostics (yet).