search

riscv / riscv-v-spec

Working draft of the proposed RISC-V V vector extension
https://jira.riscv.org/browse/RVG-122
Creative Commons Attribution 4.0 International
956 stars 271 forks source link

Question about the vl when using element groups. #931

Closed JerryShih closed 10 months ago

JerryShih commented 10 months ago

From the doc: https://github.com/riscv/riscv-v-spec/blob/e100d4321d7ded558846c32590f9efb0887515da/element_groups.adoc?plain=1#L70

The _element_ _group_ _size_ (EGS) is the number of elements in one group, and must be a power-of-two (POT).
EGSMAX is the largest EGS supported by the implementation.

I think the EGSMAX is equal to VLMAX. What's the counterexample for EGSMAX!=VLMAX?

When AVL > VLMAX, the value of `vl` must be set to either VLMAX or a positive integer multiple of EGSMAX.

Assume EGSMAX==VLMAX. I think the vl will be VLMAX when AVL > VLMAX.

JerryShih commented 10 months ago

cc @kasanovic

aswaterman commented 10 months ago

As you quoted, "EGSMAX is the largest EGS supported by the implementation." The implementation chooses the value of EGSMAX for a given vtype setting, so of course it can be less than VLMAX.

JerryShih commented 10 months ago

Assume VLEN=128, EGW=128, EGS=4 and EEW=32 for the specific instruction.

Could the implementation have the EGSMAX<16 for LMUL=4 case? Why just make EGSMAX=16?

aswaterman commented 10 months ago

Consider the case that an implementation doesn't support any instructions that use element groups. Then obviously it doesn't make sense to have EGSMAX>1, so why describe EGSMAX as being 16?

From an implementation perspective, the rationale is that you normally need to collect all of the elements of a group together to process them. Supporting very large element groups can therefore become expensive.

Anyway, I think the question is answered: it's an implementation parameter.