Anyone who has written a non-trivial GPU kernel has hit the same wall. You define a struct in C++ and pass it to a kernel. Six months later, somebody adds a member, forgets to update the matching layout on the device…
