Almost all the kernel time is spent on the lyra part (which is essentially Blake on a bigger internal structure). It is slow because it doesn't fit in the registers nor in local ram; and the access is random, not sequential.
So if you want to make it GPU friendly, the only way is removing the "lyra" from "lyra2re". What remains is something similar to a subset of x11.
maybe is the time for worldwide contest for GPU oriented algo, with significant sequential memory access, utilising minimum 1GB memory.