Re: Will someone make a formal motion? Re: mid-rad, inf-sup, a caution...

[John Pryce <j.d.pryce@xxxxxxxxxxxx>]

Nate & all

On 13 May 2010, at 02:19, Nate Hayes wrote:
> That's a fair summary on the cache issue. Please excuse me, but in this context I found John's comment very shocking!
> 
> The recently discussed conversions for mid-rad (ala Vienna and Arnolds comment about the x+-r constructor) are good enough for P1788; I gave the computer graphics and non-754 mid-rad encoding examples to support this idea!!!
> 
> My only lingering concern is that we should not prevent (whenever possible) internal implementations of mid-rad. I've continued to be a little confused on this point, because I hear what I think are conflicting messages.

Ah, I think I understand. Nate, you were objecting to the two words "hence bandwidth" in my email. My comment about "saving less than 50%" was *purely* about memory saving. I am well aware of the drastic effects cache spillage can have, which I have observed mainly in standard 1- to 4-core machines doing intensive sparse linear algebra, when a size parameter passes a threshold value. I agree that in highly parallel work, Amdahl's law might magnify the effect hugely, very much dependent on the application. I withdraw those two words.

On the substantive matter of mid-rad support there seems to be violent agreement.

John