Re: Vector Memory Ordering


Bill Huffman
 


On 9/4/20 10:25 PM, David Horner wrote:
EXTERNAL MAIL


On 2020-09-04 1:18 p.m., Bill Huffman wrote:

I think from this morning, we are considering:

  1. Ordered scatters are done truly in order

I tentatively agree *if* "truly in order" means "as if writes were the equivalent scalar writes of the register elements processed in order from 0 to vl" .

However, RVWMO should be assumed and no further strengthening of "order" should be stipulated.

None overlapping element writes are allowed in RVWMO global memory order to freely pass each other.

Yes, that's what I intended.

      Bill


I would however prefer to weaken the constraint

as #528 recommends (in addition to a mnemonic change):

...  VSUXEI  ... stipulate that the net result of the operation allows multiple outcomes visible to the local hart; essentially depending upon which element is last written to overlapping memory locations.

...  [VSXEI] to stipulate that the data written to memory will be as if the active elements were first copied to an XLEN length buffer ... processing the elements in the order from 0 to vl, and then only the affected bytes written to memory in whatever order RVVWMO allows.

...

The ideas are that

  1. rather than suggesting VSUXEI is the exception to the rule, it indicates that the ordered VSOXEI is a more constrained alternative to the default processor model that allows concurrent execution wherever possible and
  2. VSOXEI makes no guarantee on what is globally visible than what the RVVWMO model allows
  3. VSUXEI is still constrained by the RVVWMO rules.


Join tech-vector-ext@lists.riscv.org to automatically receive all group messages.