These problems have arisen from the unfortunate fact that they're easy to
fix, though it's not very efficient use of the silicon. Since there's
plenty of it, not many of us are concerned about that any more. Since
there's a pipeline register in every CLB, you might as well build your
circuits as synchronous ones.
See my comments below, please.
Dick
-----Original Message-----
From: Tony Duell <ard_at_p850ug1.demon.co.uk>
To: Discussion re-collecting of classic computers
<classiccmp_at_u.washington.edu>
Date: Monday, August 30, 1999 7:28 PM
Subject: Re: FPGAs and PDP-11's
>> In times pretty much gone by, one ran the risk of having the FPGA pinout
>> change due to rerouting. Todays tools generally support pin-locking,
i.e.
>> firmly binding the signal to a predefined pin, shuffling other resources
>> instead of reassigning pins when rerouting a circuit. This can, however,
>> impact timing. READ THE FINE PRINT!
>
>Any FPGA tool that doesn't let you assign pins yourself is fundamentally
>broken IMHO. The whole point of an FPGA is that you can make changes to
>the logic (and thus fix bugs) without having to redesign the PCB.
>
Choosing the pins at the outset isn't as important to me as choosing to have
them remain what they become after the first route.
>
>Every FPGA tool I've used has let you assign a pin number to the I/O
points.
>
>However, that's not what I was refering to. I was talking about the fact
>that when you make a change to the circuit, the compiler will probably
>re-route internal signals. Adding an output counts as a change, of course.
>
>And when it re-routes signals, said signals will change in timing. And
>when that happens, you have the possibility, in a badly designed circuit,
>of getting glitches.
>
I'd make two points about these comments. They're right, of course, seen in
the proper light, but they're really not the case at all. Taking the second
one first, synchronous circuits do not produce glitches. By definition, the
outputs are clocked at the same rate as the inputs, which are, if properly
designed, clocked so as to let the signals settle between clocks, there can
be no glitches. Outputs are clocked, not gated. It's possible to build
circuits which have glitches, but not recommended. With a synchronous
circuit, the only concern, and it is significant, is the clock skew.
Careful planning and cautious design can minimize/mitigate this risk.
Additionally, though my experience with this phenomenon is limited, it was
considered advisable to let the software determine the pinouts on the first
pass through the routing process, thereby making the timing constraints the
primary priority for the router to manage. Subsequent routes, particularly
changes to other parts of the circuits than the critical pinouts, can be
allowed to be automatically assigned as well later, but the important pins
held constant, or locked. Since these routing algorithms do strange things,
it's still possible you won't get the pin assignment you prefer, your odds
are improved in this way.
>
>[....]
>
>> >I've done this many times (in fact, IMHO it's the _only_ way to see what
>> >an FPGA circuit is really doing), but a word of warning :
>> >
>> >When you add the extra outputs, you recompile the circuit and probably
>> >change the routing of some signals. Now, these signals will therefore
>> >suffer different routing delays which means that glitches may move
about,
>> >appear 'from nowhere' or vanish.
>
Glitches can only occur when transient signals are propagated through a
combinatorial stage. If circuits are pipelined correctly (meaning for the
circuit in particular, not according to some cookbook) decoders' inputs will
be registered, and their outputs will also be registered, hence, they can't
generate glitches. Routing delays must be considered, and the skew of the
clock can, therefore engender routing races, but these must be taken into
account when computing the system clock rate. The clock must be chosen such
that it doesn't allow such glitching. Metastability can also become a risk
when different segments of a circuit are clocked with different generations
of the same clock or with entirely different ones. Mitigation for those
hazards must also be provided.
>
>> >Yes, I know that a well-designed FPGA circuit won't have problems with
>> >routing delays. But if you're new to FPGA design, especially if you've
>> >done a lot with TTL, you won't expect your wires to give sigificant
>> >(longer than switching time of a gate) delays.
>
>-tony
>
Received on Tue Aug 31 1999 - 00:49:39 BST