Can Epiphany cores access unmapped memory?

[alexrp]

[notzed]

I'd still like to know if the 32MB ram window (as is currently defined; i know it isn't fixed-for-all-time) is the only memory the epiphany can see. Or can it access anything in the whole xGB range willy nilly including memory mapped i/o?

alexrp: a paged virtual memory system means just that. 'N' could be stored on disk and re-used by another process at any time, the only guarantee is that M only needs to appear to exist when P accesses it. You can't and shouldn't have any expectation that it would work as you describe. If M is bigger than a page then N can't even be expected to be contiguous.

There is a dedicated shared contiguous region for this purpose, for a reason.

I don't know whether the shared block is also mapped with no caching - if it is then what you describe would not even work properly even if the memory was guaranteed to stay in the same place without additional low-level operations to manage the cache or page table properties.

[ysapir]

@notzed, as I mentioned in the earlier reply - the Epiphany can access any of the addressable 4GB on the system.

[notzed]

[alexrp]

So, in summary, while the Epiphany can access all of the system memory, there is no reliable way to sanely expose the rest of the system memory to the Epiphany when the host OS is multitasking (such as Linux)?

[ysapir]

The problem is not necessarily multitasking, but rather using virtual memory space via a MMU and/or paging (swap) files.

[notzed]

The assigned block of memory should be able to be changed in size - the way memory is assigned to a typical integrated graphics card. This is pretty normal for co-processors on SOCs.

I guess an mmu could be implemented in the fpga at some point in the future? I think as reads contain the source-core address it could even conceivably be per-core. Adds a fair deal of complication at every level, performance implications, and would need a kernel driver.

[ed2k]