> > VIS gives you 64-bit registers, but not all the operations you need
> > for ECDL (if I remember correctly, and/or/not/xor are provided, but no
> > shifts).
>
> Right...
>
> > Besides, isn't it the case that all UltraSPARCs are SPARC V9
> > compliant, i.e. already have 64-bit general-purpose registers and
> > 64-bit integer operations? If so, just compiling the 64-bit ECDL
> > client will give excellent performance already without the VIS stuff.
>
> That's the case! I guess the only gain could be achieved
> by carefully (read manually) scheduling the code so that one
> could run two multiplies at the same time: one in the int
> pipe and the other in the graph/FP pipe. Good luck!
This was one of the ideas triggering the thought about VIS. You have a lot
more free registers to use, and by scheduling the instructions by hand could
maybe generate perfect code, i.e. with smallest latency possible for the
given problem. I have made good experience with this with some other complex
algorithm, so the thought came up to use VIS here...
But, with the need to use memory to exchange values between the integer and
the VIS/FPU registers, and the lack of shifts in the VIS instructions, the
whole story makes less sense to me now. Thanks for you replys anyway, I will
continue calculating points the usual way...
Eddie C. Dost
ecd@skynet.be
This archive was generated by hypermail 2b29 : Mon Jan 17 2000 - 20:38:29 MET