I'm still a bit confused between data-parallel and shared address space. It seems that data-parallel does sometimes involve loads and stores on a shared address space (like in the example on slide 30). Is the only difference in the synchronization, as in forall handles the assignment and synchronization while in shared address space each "thread" is responsible for finding its own assignment (through something like threadId)?
This comment was marked helpful 0 times.
kayvonf
Great question! I wouldn't pull hairs about the differences because many real programming frameworks are combinations of these ideas, but there are some clear differences in how one thinks about how to express a program in each model. What might those be? One similarity is that both the data-parallel ("streaming") and shared address space models, in the simple form I talked about here, adopt a single global shared address space.
I'm still a bit confused between data-parallel and shared address space. It seems that data-parallel does sometimes involve loads and stores on a shared address space (like in the example on slide 30). Is the only difference in the synchronization, as in forall handles the assignment and synchronization while in shared address space each "thread" is responsible for finding its own assignment (through something like threadId)?
This comment was marked helpful 0 times.
Great question! I wouldn't pull hairs about the differences because many real programming frameworks are combinations of these ideas, but there are some clear differences in how one thinks about how to express a program in each model. What might those be? One similarity is that both the data-parallel ("streaming") and shared address space models, in the simple form I talked about here, adopt a single global shared address space.
This comment was marked helpful 0 times.