Re: [Algorithms] General purpose task parallel threading approach

[Sebastian S. <seb...@gm...>]

On Mon, Apr 13, 2009 at 7:37 PM, Tony Cox <to...@mi...> wrote:

>  " that any parallel loop with a static call graph could be turned into a
> 16 wide vector version mechanically."
>
>
>
> "Vector complete" is also a more or less irrelevant property for real world
> problems, since the "static call graph" constraint is typically only
> satisfied by trivial code. Simple shader type things meet the requirement,
> complex ones with conditionals don't, and good luck with anything resembling
> game logic.
>
>
>
> Don't get me wrong, I think LRB is doing some really interesting things,
> but it's by no means anything like a silver bullet.
>

Yeah definitely, you'll probably only find suitable data parallelism very
close to the leaves of the call graph, not in the big bulky middle, so you
clearly need an efficient general purpose CPU to run most of the code.
It is a lot more flexible then we're used to at dealing with those rare data
parallel bits though. Of course, it remains unknown how it will stack up
against whatever Nvidia/AMD will have out at the time of its release, as
they're both getting more and more general purpose too. Whatever happens I'm
sure it'll be an interesting couple of years ahead!

-- 
Sebastian Sylvan
+44(0)7857-300802
UIN: 44640862