Re: Requiring the EDP in 754 (or 1788.1 for that matter)
> On 27 Jan 2016, at 14:27, Michel Hack <mhack@xxxxxxx> wrote:
>
>
> Now, the 754-2018 revision is going well, so perhaps we can revisit the
> initial restraint. I urge those who care to join the 754 working group.
Hopefully, I have already done this, but if not, can you forward the following for me,
Michael?
> In the meantime I take every opportunity to remind my colleagues at IBM
> who are involved in processor design of the properties and advantages of
> Complete Arithmetic (as we had 30 years ago for HFP, easier due to the
> narrower exponent range of the old S/360 format) -- but I have no direct
> influence.
I’ve been wondering about the ARM-NEON instructions and how to utilise them
as a long accumulator. If we assume that they were treated as 32x32 bit,
then the following four operations could be implemented:
EDP_ZERO vmov.I32Q{0-15} #0
!! Might be a single instruction; might need replication
EDP_SAVE <addr> !! A sequence of instructions to move vector registers to memory (slow)
EDP_RESTORE <addr> !! A sequence of instructions to move memory to vector registers (fast)
Of course the missing one is:
EDP <addr_x> <addr_y>
!! MUF r0, <addr_x>, <addr_y>
!! split r0 into a signed mantissa, and an exponent
!! .. and then the magic bit:
!! AN INTEGER ADD to the 1024 bit vector in Q{0-15}.
Like you, I have no influence over ARM’s engineers, but a suggestion
that we should have a vector add/sub for 1024 bit accumulators, would
permit very fast EDP — as well as fast GMP, with all that that means for
crypto-applications.
I’ll suggest it to them..
Dave
>
> Michel