[Libre-soc-dev] we need two XLENs
programmerjake at gmail.com
Fri Sep 9 02:49:44 BST 2022
On Thu, Sep 8, 2022, 17:52 Luke Kenneth Casson Leighton <lkcl at lkcl.net>
> On Fri, Sep 9, 2022 at 1:27 AM Jacob Lifshay <programmerjake at gmail.com>
> > reading through this, it occurred to me that operations can have
> > different source and destination elwidths, therefore just one XLEN
> > variable isn't enough to express the full functionality required of
> > scalar operations when svp64-prefixed.
> yyeah i was hoping to avoid that: we'll find out soon enough
> when implementing elwidths in the simulator.
> what i *think* (hope) will happen is that results can be post-truncated
> to DLEN, particularly as a post-analysis phase (saturate, CA, CA32).
> thus - i *hope* - that XLEN is sufficient.
imho it's a waay better idea to have the two variables, a post-analysis
isn't sufficient because of things like combined fp arithmetic and fp
format conversion, (e.g. fsqrt with srcelwid=64 and destelwid=f16) where
rounding correctly requires fsqrt to know what the actual destination
format is, if it just rounds to f64 and then rounds again to f16, it'll
give the wrong result in some cases.
More information about the Libre-soc-dev