[Libre-soc-isa] [Bug 1057] split out all int/fp min/max ops into their own RFC ls013

bugzilla-daemon at libre-soc.org bugzilla-daemon at libre-soc.org
Thu Apr 20 13:58:36 BST 2023


--- Comment #17 from Luke Kenneth Casson Leighton <lkcl at lkcl.net> ---
(In reply to Jacob Lifshay from comment #13)
> vector minuw is useful as a replacement for the sequence:
> u64x4 -> u32x4 -> minu -> u64x4
> or other min-at-half-width ops.

yes that looks terrible.

> note we need half-width compare anyway for vector cmp[l]w, 

click. yes. ok am in.  so that's a 3-bit Mode.

(In reply to Jacob Lifshay from comment #12)

> I just pointed out arm has 32/64-bit scalar min/max, 

missed that sorry

> if it's that much of an
> issue we could declare min[u]w/max[u]w unvectorizable 

blech no.

> but imho that's a
> terrible idea, we should have both scalar (what I think is the main
> justification) and svp64 vector min[u]w/max[u]w (just to be orthogonal).

yes. orthogonal "wins" evrry time.

You are receiving this mail because:
You are on the CC list for the bug.

More information about the Libre-SOC-ISA mailing list