Parallella Community

Is this limitation actually an architectural thing, or is it just how things work on the Parallella in particular? From what I gather, all core-local memory is greater than 0x00100000 on the Parallella, so that's not an issue, but this might not necessarily be the case for other systems, such as a simulator.

If this is indeed a board-specific thing, it should probably be removed from the architecture manual and added to the Parallella reference instead.

The sentence should probably be rewritten to say that the address used for TESTSET must be the "global" address and not the local address. For example, a core would access its own memory starting at address 0x with a regular load store, but if it wanted to do a TESTSET on its own memory it would have to use the complete memory address (as in..0x824000000)
This applies to the Epiphany memory architecture and will likely not change going forward.

Does that make sense?

Andreas

It makes sense, but surprises me a bit.

Won't this mean that any core intending to do atomic operations with this instruction will have to be recompiled for every type of board it's going to run on? This doesn't seem too future-proof. :(

What's the reason that a regular address can't be supported?

Forgot to add: This actually complicates things a *lot* for me, because so far my simulator does not have a concept of host-side addresses; it only speaks Epiphany addresses. This has vastly simplified the implementation, and having to add translation from host-side addresses to Epiphany-side addresses would make me a sad panda. I can do it, but it seems very overkill for this one instruction.

Not sure, but I think we are misaligned here...I am not talking about the host side address,I am referring to the Epiphany address map with up to 4095 Epiphany cores, each one with its own unique address range. It's just that the TESTSET instructions cannot be used on addresses starting at 0x00000000.(see Figure 5)

Ooooh. Sorry. It's 5am and I probably shouldn't be programming at this hour. :)

So before I say anything else, what it means is just that cores placed below 0x00100000 cannot have TESTSET used on their memory, right?

Go to sleep:-) The minimum memory space per core in the Epiphany is 1MB (0x00100000). There is really no way to have a core at address 0x00000000.

Ah, right! This all makes sense now.

That said, maybe this belongs in the chapter about the memory map? The description of TESTSET is a surprising place to find this stuff. ;)