URL
https://opencores.org/ocsvn/socgen/socgen/trunk
Subversion Repositories socgen
[/] [socgen/] [trunk/] [doc/] [src/] [guides/] [reset_sys_design.html] - Rev 20
Go to most recent revision | Compare with Previous | Blame | View Log
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN">
<html>
<head>
<meta http-equiv="CONTENT-TYPE" content="text/html; charset=utf-8">
<title>start</title>
<meta name="GENERATOR" content="OpenOffice.org 3.0 (Linux)">
<meta name="CREATED" content="0;0">
<meta name="CHANGED" content="20100309;9305300">
<meta name="KEYWORDS" content="start">
<meta name="Info 3" content="">
<meta name="Info 4" content="">
<meta name="date" content="2008-01-08T12:01:41-0500">
<meta name="robots" content="index,follow">
<style type="text/css">
<!--
H3.western { font-family: "Albany", sans-serif }
H3.cjk { font-family: "HG Mincho Light J" }
H3.ctl { font-family: "Arial Unicode MS" }
-->
</style>
</head>
<body dir="ltr" lang="en-US">
<div id="toc__header" dir="ltr">
<p><br>
</p>
</div>
<h1><a name="socgen_project"></a>Design considerations for Reset
Systems<br>
</h1>
<p><br>
<br>
</p>
<p>In a world as fast moving as the semiconductor industry it is
essential that all designers continuously update their knowledge as the
technology changes. It is very easy to become complacent and then
suddenly discover that the techniques that have served you for many
years no longer work. <br>
<br>
</p>
<p>This paper was written to explore some of the mistakes that reset
system designers have made over the years and why they are no longer
true.<br>
</p>
<h3 class="western"><br>
</h3>
<h3 class="western"></h3>
<h3 class="western">Do we really need a reset system? <br>
</h3>
<p>Actually you don't. It is a good design practice to ensure that
there are no dead end states in your logic and that any state will
eventually lead into a valid operating mode. For many years designs
were simple and robust enough that they would function even if they
were enabled without a reset. Then along came embedded processors and
the world became much more complex. I have seen some pretty audacious
attempts to create a watchdog to detect and restart a lost system
but the best that they can do is to improve the odds that the system
will recover. None of them were 100%.<br>
</p>
<p>It is possible that you do not need to reset all the storage
elements in a design. In many cases the data is reloaded shortly before
it is needed and it doesn't care what it was before that time. Some
designers will leave certain storage elements off of the power on reset
because it has no effect on the operation.<br>
</p>
<p>BUT.<br>
</p>
<p>There was a mathematician named Fermat who came up with a theorem
that eventually became known as Fermat's last theorem. It was a simple
little equation that worked in every test case that they threw at
it and they threw a lot of test cases at it. But it took over 350 years
before someone could prove that it would really work in all cases.</p>
<p><br>
If you allow your designers the option to leave storage elements off
the power on reset system then they will come up with these wonderful
little designs that appear to work and they will work in any test case
that you throw at it. But it will take you FOREVER to fully verify that
it will work in all cases.<br>
</p>
<p>You do not need a power on reset system for your logic to work. You
need it in order to verify that your logic works. It takes
longer to verify a design than it does to create it and not providing a
100% known startup condition will make the verification effort that
much harder. All storage elements must be on a reset if only for test
and verification purposes. If you have logic that must function during
a power up reset then put it on a special reset that is only active in
test mode.<br>
</p>
<br>
<h3 class="western">All components must come out of reset on exactly
the same clock<br>
</h3>
Thats true, or at least it was back in the 60's. Back then every
component would come out of reset and start "componenting". The reset
system acted like a conductor so that everybody started on the
same beat. Those types of systems are rare today. Most major chips have
one or more microprocessors in side so components come out of reset
only to sit there waiting for the cpu to configure them and get them
started. It doesn't matter what cycle you come out of reset on as
long as you are up and ready before someone else asks you to do
something. <br>
<br>
This has led to two prong approach to reset system design. The
majority of the chip is on a large slow reset distribution
tree that doesn't even try to get everybody reset on
the same cycle. Then you have a second smaller and
faster tree that only resets the cpu and anything else that can
initiate activity. The fast reset is delayed long enough to
ensure that the slow reset is finished before starting the cpu. In
modern designs this can be a significant number of clock cycles. I have
seen repairable memories where you had to hold off starting the cpu for
3000 clocks to ensure that any repair would be finished before the cpu
started.<br>
<br>
<br>
<h3 class="western">You must design an asynchronous reset system<br>
</h3>
Absolutely. Most of the time your mission mode requirements will
dictate that the power on reset system works even in the absence of
clock. If it doesn't then the test engineer will require that all pads
must respond to an async reset in case a board is built missing it's
clock. Asynchronous reset design is essential. A power up monitor
will drive the reset input active as the power is ramping
up. You will not have a clock at this time so the reset system must be
able to work without one.<br>
<br>
<br>
<br>
<h3 class="western">You must design your logic using synchronous design
methods<br>
</h3>
Absolutely. Todays chips are huge. The only way that you can close
timing on a large design is if everyone follows strict synchronous
design rules. The mistake that many of todays designers make is
that they think that because they have to design an asynchronous reset
system that they get an exemption from following the rules for
synchronous design. Sorry guys, it not one or the other its BOTH.
You have to design a asynchronous reset system but you cannot use any
flip flops with an asynchronous reset port.<br>
<br>
The funny thing is that synchronous design methodology is quite
capable of creating an asynchronous reset system and
will actually give you a smaller and faster design that
either of the traditional async only or sync only solutions.<br>
<br>
<br>
<h3 class="western">Don't worry about making the reset system
testable.The test engineer has a tool that will fix any problem in the
back end<br>
</h3>
That used to be true. The first thing a vendor does when they get a net
list is to run a full drc that looks for dft issues. If anybody has any
signals crossing between the async reset port on a flipflop and
either a D or a Q port then it flags it as a violation. So you can
either send it back to the customer and wait a week for them to find
it, fix it, and resynthesizes or you can eco in a test mux at the flop
and have it fixed in 5 minutes. Everyone took the easy way out.<br>
<br>
But then along came Logic equivalence checking (LEC). The final
routed net list will be sent back and compared with the customers
golden net list and all of these ecos will show up in the report.
Now somebody has check out each and every item in the report
before you can release the masks. It now becomes easier for the
customer to find and fix these errors before synthesis than it is to
deal with thousands of lec errors.<br>
<br>
Besides with the newer processes the days when you could eco in a small
tweak on a routed net list and not have it break something are fast
disappearing. You will eco the rtl code and then re-synthesis and
reroute.<br>
<br>
<br>
<br>
<br>
<br>
<h3 class="western">You must use a sync_reset pragma if you design a
synchronous reset system<br>
</h3>
I cringe whenever I hear someone say this. If you do a
synchronous reset design then you will find that your gate simulations
will not run. Many of your flipflops will never reset to a known value.
They will get a valid clock and the reset in the block will be valid
but synthesis will have combined the reset logic in with the mission
mode logic and it will be distributed throughout the logic cone feeding
the D input. It also uses the flops current state in order to compute
the next state. It creates a situation where if the flop has a 0
or 1 in it then the logic will compute the next state as 0 when reset
is active. However if the flop is unknown as it is at power up then
verilog is unable to figure out the correct next state and it remains
at x.<br>
<br>
This is a simulation only issue as flops in real silicon will always
resolve to a valid state.<br>
<br>
Tool vendors created the sync_reset pragma so that you could tell the
tool not to combine the reset logic with the mission mode logic. You
place it at the very tip of the logic cone and it will remain there in
gates.<br>
<br>
So whats wrong with that?<br>
<br>
The synthesis tool will make a list of all signals that enter the logic
cone along with the relative time it enters before the next clock
edge. If it finds a early arriving signal entering the cone
closer to the tip than a late arriving one then it will try
to remap the logic and swap them so that the late
arrival can move closer to tip. Ideally the
latest signals are moved towards the tip and the early ones are moved
to the rear. <br>
<br>
The reset from a properly designed distribution tree and the feedback
signal from the flop that you are working on will always be two of the
earliest signals. They will get pushed up away from the tip of the cone
simply to make room for the mission critical late signals. This is a
good thing, you want this to happen. <br>
<br>
The problem is that designers think that they must prove that the reset
system works in gate sims. Verilog is a great tool when every node is
in a known state but it is lousy when dealing with unknowns. There are
times like this when it is possible to resolve a X into a known
value and it can't. There are also times when it will resolve an
X to a known value when it shouldn't. The only way to use verilog is to
start with everything in a known state and stop it when anything goes
X. That means theres a problem and nothing downstream from that X
can be trusted.<br>
<br>
<br>
You do not prove your reset system design in gates sims. You prove the
design in rtl sims and use LEC to prove that gates matches the design
that works. Then you use initial statements to force all flops to
a known state at startup and use gates sims to prove that everything
else works. Verilog gates is the wrong tool to use to verify the reset
system.<br>
<br>
You never use the sync_reset pragma unless you really like big slow
designs.<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<h3 class="western">Doing a synchronous reset design adds logic in the
D pathway and will slow down the design<br>
</h3>
<br>
Wrong. Adding logic in the critical path will slow down the design.
Adding it into a non-critical path simply reduces slack in that path.
If you put the reset logic at the very tip of the logic cone then you
are adding it into the critical path and the synthesis tool will move
it up the cone until it is in a safe location.<br>
<br>
Adding a synchronous reset system doesn't really add much logic to the
design. The tools will first locate any mission mode logic that also
forces the flop into the reset state and it will piggyback the reset
system with that logic. You don't add gates , you bump a gate up to add
an extra input.<br>
<br>
<br>
<br>
<h3 class="western">A component must perform reset in one clock
cycle<br>
</h3>
The power on reset is really a slow operation. A typical system
could see:<br>
<br>
<ul>
<li>Ramp time for power rails</li>
<li>clock start up time</li>
<li>pll lock time</li>
</ul>
You are looking at activity that is measured in the milliseconds on a
system clock that is measured in the nanoseconds. Performing a reset in
one clock cycle requires adding logic to every single flipflop<br>
for no good reason. A designer should only add reset logic as a last
resort. The prefered method is to use the existing mission mode logic
to perform the reset. If you have a computational block with a fifty
stage deep pipeline then reset should force it's inputs to 0 and open
all the gates so that every flipflop will be flushed out in 50 clocks.
Better yet would be to have the block feeding your input force it's
output to all 0's during reset.<br>
<br>
Every design should spec a multicycle reset and give the designers the
freedom to reset any way they want as long as it's finished by the end
of the reset pulse.<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<div id="toc__header" dir="ltr">
<p><br>
</p>
</div>
<h1><a name="socgen_project"></a>How to design the Reset System<br>
</h1>
<br>
<br>
<h3 class="western">1) Write a mission statement<br>
</h3>
The first step in any design task is to write a statement that sums up
what the thing you are designing will do. This is important
because everything after this point must be traceable back to this
statement. The statement will tell you what
steps you must follow. Anything that you cannot trace back to something
in the mission statement is not part of the design<br>
<br>
<br>
The mission statement for the reset system is:<br>
<br>
<br>
<br>
The reset system will force all the nodes in a system or subsystem into
a known good state while a reset trigger is active.<br>
<br>
<br>
<br>
<h3 class="western">2) Define the reset triggers<br>
</h3>
We must now make a list of all the events that will cause us to reset
all or part of the system. Our list is:<br>
<br>
<ul>
<li>The design has a power monitor chip that provides a low signal
when the supply rails have not been above the limit for a long
enough period of time</li>
</ul>
<ul>
<li>The design has a soft reset block that can reset any subblock if
its reset flop is set to 1.</li>
</ul>
<ul>
<li>The clocks must run during reset but the divider has a special
reset input for simulation and testing</li>
</ul>
<ul>
<li>The design has ieee 1149.1 test logic with a active low trst* pin.</li>
</ul>
<ul>
<li>The reset signal has a metastable filter to sync it with
the clock.<br>
</li>
</ul>
The last is important because some designers will forget that the
filtered output is actually it's own seperate reset domain<br>
<br>
<br>
<br>
<br>
<h3 class="western">3) Define a known good state<br>
</h3>
We now look at every storage element in the design and define a
safe state for each element of either 1 or 0. Don't cares are not
allowed. If you cannot pick a value then one will be assigned for you.
This task is best performed after the system and board designers have
defined the known good state for the PCA. They will define the
state for all of the pads, the ic design team must define the states
for all internal nodes.<br>
<br>
<h3 class="western">4) Assign storage elements to triggers<br>
</h3>
Once we have a list of all storage element we list any and all triggers
that will force them into a safe mode.<br>
A typical list would list all the flipflops in timer module u12.r567
would be reset by:<br>
<br>
<ul>
<li>an active high on soft reset bit #23</li>
</ul>
<ul>
<li>an active low on the power monitor input</li>
</ul>
<ul>
<li>an active low on the simulation/test reset</li>
</ul>
<ul>
<li>an active low on the output of the metastable filter</li>
</ul>
<br>
The jtag reset is not included because it doesnt reset the timer.
Once this step is complete it will provide a map for the reset
distribution tree that you will need. The best way to distribute the
reset over a large design is to use what is called a "synchronous reset
tree".<br>
<br>
<br>
<br>
<br>
<br>
<h3 class="western">5) Select between synchronous or asynchronous reset
system<br>
</h3>
At this point it is easy to see if we need a synchronous or
asynchronous reset system. If your trigger is asynchronous then you
must design a asynchronous reset system. If the trigger can occur
without a clock then you must be able to reset the system without a
clock.<br>
<br>
If the trigger is synchronous then you may design a synchronous reset
system or you may also choose to design a asynchronous one. The
async one will enter reset one cycle before the sync one
but they will both exit on at the same time. <br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<br>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
<p><br>
<br>
</p>
</body>
</html>
Go to most recent revision | Compare with Previous | Blame | View Log