This might be a moot point as I haven’t yet tried to implement it (I’m working to get it on the CPU first), but I’m wondering about it. I have a derived data type with, say, 40 elements that I’d like to put into constant memory on my GPU. What I’m wondering is whether I can/should use SEQUENCE with the TYPE or not?
As I recall, SEQUENCE tells the compiler to put the TYPE members in successive memory locations. For a GPU this seems like a useful/good thing, but does it mean anything for GPU TYPEs with device/constant attributes and, even if it does, will it matter with constant memory since it’s a broadcast (in my code) and so memory order may not matter. (This might be more important for device TYPEs.)
Any advice on using TYPEs with GPUs?