[libre-riscv-dev] [isa-dev] Re: FP transcendentals (trigonometry, root/exp/log) proposal

[Allen Baum]

Actually, it doesn't matter if its more accurate or less; the problem is
that it's different,
That difference probably doesn't matter for some programs, but unless
you're only running some very small number of programs where you might be
able to tell if it matters  - then the difference matters.

On Sat, Sep 14, 2019 at 3:25 PM lkcl <luke.leighton at gmail.com> wrote:

> On Saturday, September 14, 2019 at 10:50:15 PM UTC+1, MitchAlsup wrote:
>
>
>> When you get a multi-generational customer, which has an existing
>> application;
>> the first thing that customer will do is to run a sanity check of the new
>> HW implementation using a
>> "reference" benchmark. When any bit of the output of that program
>> mismatches any bit from the
>> new implementation, the multi-generational customer will not be able to
>> rationalize that this new
>> HW is acceptable or not. This leads to various consternation,..... This
>> happened to me at GOULD
>> S.E.L circa 1982 when the new, faster, HW had more accurate FP
>> implementation.
>>
>
> rright.  ok.  so let me just check.  it's not the accuracy setting that
> you're objecting to / red-flag-raising, it's the "if less accurate is set
> you're allowed to be *more* accurate", is that correct?
>
> i am reminded of a message by hendrik [allo similarly old person, told ya
> your input was valuable :) ] on iibre-riscv-dev a couple months back about
> how Ahmdahl got similar customer complaints, and had to have an OS patch
> providing *EXACTLY* bit-for-bit the brain-dead broken FP accuracy of the
> mainframe that Ahmdahl were copying.
>
> is this what you're referring to, Mitch?
>
> l.
>
>>