John Zabroski wrote:

> From OO to FPGA: Fitting Round Objects into Square Hardware? [1]
> was one of the interesting talks I sat in on when I attended SPLASH.

Thanks for the link!

> I was primarily interested in attending because of VPRI's long-range
> mission and the speculation that FPGA hardware will be fundamental
> to its top-to-bottom, side-to-side late-bound approach.  Ian was asked
> about compiling FONC stuff to FPGAs late 2009 [2], and he replied by
> saying FONC loves FPGA concepts but has no FPGA experts and so
> has no concrete plans for using FPGA at the moment [3].

Chuck Thacker has been doing interesting (but so far conventional)
things with FPGAs and his work is listed as part of the VPRI efforts in
the anual reports. So I would say there will eventually be results.

The group I am working with, called the "Reconfigurable Computing
Laboratory", is full of FPGA experts and publishes almost exclusively in
FPGA conferences.

> Jens Palsberg gave the talk. He proclaimed at the end that this will
> be one of the most important open, hard problems for the next 10
> years.

In his case, the bulk of the complexity is hidden inside the AutoPilot
tool created by a third party. I was very shocked that it worked so well
for him. Particularly impressive was getting the Richards benchmark to
fit into a reasonably sized FPGA.

Most of the state of the art in C-to-hardware is very disappointing. My
own proposal is that if the hardware and software versions of objects
can talk to each other, then you could use adaptive compilation to
convert just the tiny part of the application that has the most effect
on the performance (the "hot spot") into hardware and run the bulk of
the code on a processor.

> He said that there is a lot of redundancy between the front-end and
> back-end tools and we really need to collapse the two intermediate
> representations used into a single form, because too much information
> has to get reconstructed when you use a subset of the C language as
> an IR.

True, and every tool has a certain view of the hardware which might not
match an OO model that well. If you look at the Bluespec
(http://www.bluespec.com/) hardware description language, for example,
it deals with very different kinds of objects than Verilog (though it
gets compiled into that) or VHDL.
 
> Interestingly, Jens alluded during his talk that it took him 3 years to
> get this work published and he had to refine it to satisfy reviewers.

My own paper on adaptive compilation for FPGAs was rejected, but I have
to agree with the reviewer's complaints. Some here might have seen it
even so due to friends showing it to friends.

I don't like the premise that FPGAs as they are now are a given and the
software will have to deal with it, however. This related research into
a different architecture is very interesting, for example:

http://rala.cba.mit.edu/

http://fab.cba.mit.edu/classes/MIT/961.09/04.27/RALA/rala.ppt

-- Jecel


_______________________________________________
fonc mailing list
[email protected]
http://vpri.org/mailman/listinfo/fonc

Reply via email to