Do you have examples (or simulation) of systolic implementation of the
QR-RLS algorithm in Adaptive Filter Theory (Haykin's book)? I found it
difficult to understand just by reading!
Reply by Fred Marshall●June 17, 20052005-06-17
"tradewind" <email@example.com> wrote in message
> Do you have examples (or simulation) of systolic implementation of the
> QR-RLS algorithm in Adaptive Filter Theory (Haykin's book)? I found it
> difficult to understand just by reading!
I don't have the book and I'm not an expert on QR-RLS algo.
Does it help to know that a systolic implementation is a multiprocessor
The idea is that interim results "pump" through a chain of processors that
are each dedicated to one part of the process. It was a bandwagon the
mid-1980's. I don't think I've heard much about it since.
Note: in those days getting high throughput processing often seemed to be a
matter of using more processors and brought with it all manner of
communication and memory access issues. Some board manufacturers seem to
have done a pretty good job of making multiprocessor boards with useful
communication and memory access architectures.
I worked on a project at Honeywell that provided back-end 8-processor
TMS320C30 accelerator boards for the Intel iPSC2 "hypercube" (Multibus II).
We sold an accelerated iPSC2 to the Navy. Otherwise we were too invisible
in the market to make a commercial splash of it. And, I'm sure the market
for specialized boards like this is limited (even with a more main-stream
bus interface). Maybe Al would comment.
Our next attempt was a 16-processor TMS320C40 with two communication
mechanisms: a bus and the COM ports in a mesh. The original was on a
15"x15" board with daughterboards for the processors or board-edge I/O (up
to 20 of them). The eventual idea was to use chip-on-board with the whole
thing on a liquid cooled SEM E card (roughly 4"x5"). We won the Navy
contract to build it but COTS took over right then and the project died
(appropriately so). I still wonder to this day what the programming
protocols would have been for it - in terms of which communication mechanism
would be used for what (e.g. data vs. control / bus vs. COM port mesh).
In those days of designing multiprocessor systems we liked to believe that
Mercury got their idea for their RACE architecture from us.
We also manufactured a couple of Warp machines for DARPA in the early 80's.
That was another multiprocessor architecture.
This was the end of my pursuits in processors that might have lent
themselves to systolic approaches.
After losing out to the Pentium in competition with our TMS320C80 boards at
PDI, I ventured into doing image processing using quad Pentiums -
multithreaded software, etc..... It might have worked out but the Intel
server platforms necessary didn't seem to match a more traditionally
"embedded" application space. But this was certainly parallel processing of
the more typical variety and definitely not intended to be systolic
Of course, in *these* days it still comes up. Witness the multicore
processors that are coming to market for PCs, etc. Any more, I'm not sure
what a systolic approach really buys.... and maybe never did know for sure!
There was a lot of DARPA hype in the 80's.... The devil is in the details.