From: Bill Mark [billmark@graphics.stanford.edu]
Sent: Wednesday, October 10, 2001 11:47 PM
To: Bill Dally
Cc: Ian Buck; lance@leland.Stanford.EDU; ujk@leland.Stanford.EDU;
Billmark@lambert.stanford.edu; pmattson@leland.Stanford.EDU;
serebrin@stanford.edu; merez@stanford.edu; jowens@graphics.stanford.edu;
horowitz@chroma.stanford.edu; hanrahan@graphics.stanford.edu
Subject: Re: Brook - multiple input and output streams

On Wed, 10 Oct 2001, Bill Dally wrote:

>...
> The conditional and unconditional << and >> operators in KernelC are pretty
> useful.  You should consider adopting this model in Brook - where the
> kernel is 'active' for the duration of the stream - rather than a single
> element - and can explicitly read from inputs and write to outputs.
> ...

"Persistent kernels" are powerful, but allowing them does present
some challenges, especially from a programming-model point of view.

What happens when you have more than one processing unit operating on
the stream?  There are duplicate copies of the persistent state,
possibly with different values.  On Imagine, you have to add explicit
communication between the processors to do something reasonable in
this situation.  My impression is that this communication code is
often designed for a particular number of processors.  It's not clear
to me that there's any simple way to hide the number of processors.

If the language provided the "illusion" of one processor, the compiler
might be able to insert the appropriate communication, but I suspect
that it would then be very easy for the programmer to create code that
would run very slowly when the number of stream processors is high.
We should look at this question in more detail; my instinct may be
wrong here.

Allowing communication between processing units also seems to require
that the processors be SIMD.  SIMD's ability to simply perform
communication of this type is perhaps its greatest advantage over
MIMD.

Bill M.