- Vector Memory Ordering
Re: Vector Memory Ordering
toggle quoted messageShow quoted text
Bill, great lists.
can we start building a testcase list for these situations and others? maybe a github doc? i am sure these will be documented but I don't want to lose bullet lists like this and have to derive them from a doc later if we can avoid it.
Allen, does the compliance group have a proposed way (or format) to define testcases and input data sets? If not should we?
general note, I think we need to specifically add debug to the definition of done. I can imagine things like watchpoints affecting a lot of extensions in interesting ways.
On Fri, Sep 4, 2020 at 10:18 AM Bill Huffman <huffman@...
I think from this morning, we are considering:
- Ordered scatters are done truly in order
- Strided stores that overlap (including segmented ones) will trap as illegal
- All other vector loads and stores do their memory accesses in arbitrary order.
- A vector load that accesses the same location multiple times is free to use the same loaded value for any subset of results
- All loads with vector sources must use a different register for the destination than any source (including mask).
- Maybe a vector load may access the memory location corresponding to a given element multiple times (for exception handling)??
A few of the consequences of this are:
- A gather with repeated elements can access the higher numbered elements first and lower ones later
- A vector memory access where multiple elements match watchpoint criteria can trap on any of the multiple elements, regardless of watchpoint requirements on order
- A stride-0 load accessing an "incrementing" location can see a result with larger values at lower element numbers than smaller values
- When vector loads or stores access an "ordered channel" the elements will still be accessed in arbitrary order
- Strided loads, gathers, and unordered scatters to non-idempotent regions will not behave as might be expected.
- A stride-0 store to a FIFO will trap
- A stride-0 load to a FIFO will pop an arbitrary number of entries from the FIFO (from 1 to more than vl) and elements are distributed in an arbitrary way in the result.
- A non-idempotent memory location accessed by a vector load may be accessed multiple times.
We need to be sure software is OK with these characteristics as "ordered channels" and non-idempotent regions can't be known at compile time. Even strides can't always be known at compile time. Will this plan reduce the amount
of auto-vectorization that can be done?
Exception reporting still has issues:
- Unless stores can be done multiple times, there is a need to save some representation of what stores have and have not been done.
- For loads and stores, watchpoints can happen more than once without some representation of what elements are complete.
- There may need to be a way to report a watchpoint on one element but restart on an earlier element
- If loads have to do this exception reporting as well, do we forbid loads to happen more than once for each element? Does that help anything if we do?
I'd like to see us relax the ordering of gathers and unordered scatters with younger instructions in some way. If we don't, younger scalar memory accesses will stall for some time as comparisons are much more difficult than
for unit stride or even strided accesses.
Join firstname.lastname@example.org to automatically receive all group messages.