[Open-graphics] Center for High-Performance Reconfigurable Computing

[Timothy Miller]

On 10/15/06, Nick S-A <nova at macintoshclub.com> wrote:
> > We're trying to convert
> > functions to hardware, and exposing some of the underlying semantics
> > is an okay thing to do.
> > To learn a new language, the performance gain would have to be very
> > serious.
> >
> > I'm skipping the GPU entirely and proposing that we convert to FPGA
> > logic.
>
>
> Given these three constraints, I agree that a Limited C->Verilog
> converter is probably the best option.
> I didn't realize that we were just using FPGA Logic, but I guess that
> makes sense because it is more efficient, and more configurable, than
> going through a GPU. However, I still think we should focus more on
> getting the GPU working, and this has the nice side-effect of letting
> us use BrookGPU. Once we get the GPU working, the OHF should have
> enough followers to support any number of languages that people feel
> like implementing.

Well, the Open Graphics GPU isn't programmable in the way that the
latest high-end graphics chips are.  Our design goals are to get the
maximum result from a minimum of logic, and that requires us to
implement only the OpenGL fixed-function fragment pipeline.

This is another reason I'm skipping the GPU for high-performance computing.

That doesn't stop us from developing a programmable shader later, but
my perspective on using GPUs for HPC is that there's a better way, and
I'd rather work on THAT.

>
> As a side note, isn't this all that SystemC is? I have never even
> looked at SystemC, but given the name it seems like this is the goal.

http://en.wikipedia.org/wiki/SystemC

Looks like an HDL-ish set of libraries for C++.  With the right
restrictions, I imagine you could make it synthesizable.  It appears
to be more powerful than a regular HDL.  And I'm guessing that it's
probably harder to synthesize.  There are probably aspects of this
that we would want to imitate.