[libre-riscv-dev] [isa-dev] Re: FP reciprocal sqrt extension proposal
hendrik at topoi.pooq.com
Thu Aug 8 18:45:36 BST 2019
On Thu, Aug 08, 2019 at 09:53:57AM -0700, Allen Baum wrote:
> CO rdic is great for extremely low gate count implementations. But
> for highest performance, there are likely much better ways.
It may be useful to provide both imlementations in out codebase, and
use conditional execution to choose which to put in a particular chip
design. This leaves a choice between gate count and speed, which may
be done differently for different applications.
Except, of course, that we have limited time and resources to produce
our first working designs.
I'd suggest for now we choose one: whichever is the simplest, easiest
to implement and has a reasonable chance of meeting out goals.
But stay modular enough that we could swap it out for another if we
find we need to.
> > On Aug 8, 2019, at 9:01 AM, 'MitchAlsup' via RISC-V ISA Dev <isa-dev at groups.riscv.org> wrote:
> >> On Thursday, August 8, 2019 at 1:46:39 AM UTC-5, lkcl wrote:
> >> Just a thought: should this be merged into Ztrans? A case could be nade either way.
> >> * The Libre RISCV FPDIV pipeline also computes SQRT and RSQRT
> >> * Some OTFC pipeline designs do RSQRT "for free"
> >> * However, if doing RSQRT chances are high it will be part of a design that needs EXP, LOG etc anyway.
> >> * CORDIC can do a ton of algorithms including SQRT, RSQRT, LOG, SIN etc.
> >> Thoughts?
> >> L.
> > CORDIC is SLOW
> > The Moto CORDIC unit took 80-cycles or more
> > The Intel CORDIC was similarly slow
> > Modern SQRT/RSQRT hardware will deliver results in 23-30 cycles.
> > --
> > You received this message because you are subscribed to the Google Groups "RISC-V ISA Dev" group.
> > To unsubscribe from this group and stop receiving emails from it, send an email to isa-dev+unsubscribe at groups.riscv.org.
> > To view this discussion on the web visit https://groups.google.com/a/groups.riscv.org/d/msgid/isa-dev/13bb0fc5-dbb4-4e44-bce8-96058684042e%40groups.riscv.org.
> libre-riscv-dev mailing list
> libre-riscv-dev at lists.libre-riscv.org
More information about the libre-riscv-dev