Dear PICsters,

A couple of times in the last few weeks some people have been
asking how to go about getting a decent period reading both for
short and long periods.

I recently had to devise a method to do exactly that, and I
thought that others on the list might be interested in what I
came up with.

It is often useful to discuss methods by actually applying them
to an example. That makes it easier to follow, because then
you can actually attach numbers to the concepts and see what
is happening.

Let's say that we are accumulating a period count into a 16 bit
count register that consists of COUNT_LOW and COUNT_HIGH.
Let's further assume that we have an interrupt routine that
occurs once each millisecond. That would give us a maximum
period count of 65,535 milliseconds. The granularity of the
count is 1 millisecond, so an error of 1 count at the high
end (65,535) is .0015%.  At the LOW end, an error of 1
count could represent almost 100% error.

As you can see, the real problem occurs when the count is
low. Slight changes in the period can represent BIG changes
in, say, speed. When the period is small it helps a lot if
instead of just finding the period we find the AVERAGE
period based on multiple periods. The shorter the period,
the MORE samples we should include in our average.

The method that I devised uses a switchover point. Any period
longer than a certain value is handled regularly. Periods
that are shorter than a certain minimum value are accumulated
until an integer number of short periods has been accumulated,
and then the average value is dtermined.

To simplify things a bit, it is useful if the minimum count
can be expressed by a binary number that has
only a single "1" in its representation. This means we would
pick a switchover value of 1024, 2048, 4096, 8192, 16384, etc.

Upon initialization, and at the end of a successful period count,
COUNT_LOW and COUNT_HIGH are cleared to zero. In addition, a
DIVISOR register is also cleared to zero. A bit flag is initialized
to the current level of the "PERIOD port bit". This is used later
to determine edge changes.

Inside the interrupt routine the period value is incremented
once each millisecond. Then the proper PERIOD port bit is tested to
determine if the period has terminated. If the period has
terminated, then the DIVISOR register is incremented.

Then we check the high order bits that represent the numbers
greater than or equal to 1024. If they are ALL "0", then
we have not yet reached our minimum count, so we "fake" the
system by immediately setting things up as if a new
period is beginning, but leaving the COUNT_LOW, COUNT_HIGH,
and DIVISOR registers as-is.

Eventually the count reaches 1024 or more. If the 16 bit
counter is now divided by the DIVISOR value, then you will obtain
the AVERAGE period value. This average value will have less
tendency to bobble around.

In this technique the system has a switchover at the minimum
count point. Long periods are measured "normally", and periods
below the minimum will be "averaged". This method is quite good
at reducing digit bobble.

*********
Another method is to keep accumulating periods until either the
period count exceeds some arbitrary value (such as 1024), with a
DIVISOR value of 1, -OR- until the DIVISOR value reaches a
value of 10. In this case you set a flag that says the number
is a multiple of ten. Process the number as if it was normal,
but DISPLAY it shifted. (The decimal point position effectively
performs the equivalent divide by ten function). If your main aim
is simply to reduce bobble (instead of improve resolution),
then you can simply ignore the digit that would be to the right
of the decimal point.

You could extend this concept to handle even shorter periods
by extending the DIVISOR value to 100.

Fr. Tom McGahee

--
http://www.piclist.com hint: The PICList is archived three different
ways.  See http://www.piclist.com/#archives for details.