Go here to sign up for The Embedded Muse.
The Embedded Muse Logo The Embedded Muse
Issue Number 382, September 16, 2019
Copyright 2019 The Ganssle Group

Editor: Jack Ganssle, jack@ganssle.com

   Jack Ganssle, Editor of The Embedded Muse

You may redistribute this newsletter for non-commercial purposes. For commercial use contact jack@ganssle.com. To subscribe or unsubscribe go here or drop Jack an email.

Editor's Notes

Express Logic

Over 400 companies and more than 7000 engineers have benefited from my Better Firmware Faster seminar held on-site, at their companies. Want to crank up your productivity and decrease shipped bugs? Spend a day with me learning how to debug your development processes.

Attendees have blogged about the seminar, for example, here, here and here.

Quotes and Thoughts

Management has three basic tasks to perform:

  1. Establish the requirements that employees are to meet
  2. Supply the wherewithal that the employees need in order to meet those requirements
  3. Spend all its time encouraging and helping the employees to meet those requirements

- Philip Crosby

Tools and Tips

SEGGER Embedded Studio The leading cross platform IDE

Please submit clever ideas or thoughts about tools, techniques and resources you love or hate. Here are the tool reviews submitted in the past.

Brian Rosenthal wrote: This android app (https://play.google.com/store/apps/details?id=com.yatrim.stlinkp) allows you to plug your ST-LINK/V2 into your phone and download a new program to an STM32 device.  It requires an OTG adapter, but beats having to drag your laptop out to the floor for a quick update!

In the last Muse I ran my top ten reasons for project failure. Chris Gates added:

Missing from your top ten list is "Lack of Systems Engineering", or even "systems awareness."

I have seen so many projects where siloed development teams create devices & assemblies that will never be able to work together.
To my way of thinking an "Interface Control Document", is the most critical document in any development.  Interfaces are too brittle to leave up to chance.

Freebies and Discounts

Dan Neuburger won August's Siglent SDS2304X 300 MHz oscilloscope.

September's giveaway is a DS213 miniature oscilloscope that I reviewed in Muse 374. About the size of an iPhone it has two analog and two digital channels, plus an arbitrary waveform generator. At 15 MHz (advertised) it's not a high-performance instrument, but given the form factor, is pretty cool.

Mini DSO

Enter via this link.

Ultra-Low Power Systems

At last month's Embedded Systems Conference there were two themes that seemed to run through the show: autonomous driving (not surprising as the ESC was merged with "Drive" - an automotive electronics event), and ultra-low power embedded systems.

I still see wild claims being made about many years of operation from small batteries like coin cells. These all suppose the systems are sleeping most of the time, of course, but too many issues are glossed over. Five years ago I wrote a series of articles about this, now collected here. Over an 18 month period I discharged about 100 cells to study their behavior in ultra-low power applications, and ran numerous other experiments to learn what one can reasonably expect.

Here are the key takeaways:

  • Sleep current is almost irrelevant (i.e., there is nearly no practical difference between an MCU that sleeps at 1 µA and another at 10 nA).
  • What matters is time spent sleeping...
  • ... and contributions of other leaks and drains
  • The internal resistance of coin cells increases hugely as the battery is discharged.
  • The vendors suggest waking up at full speed to get work done and quickly go back to sleep. They're wrong.
  • Don't use the on-board brown-out reset circuits.
  • Droop will trim another 10% or so of battery capacity.
  • From a practical standpoint, it's impossible to use a capacitor's time constant to boost Vdd for even a handful of ms.
  • Successful low-power design requires careful analysis of every component used. Even poor decoupling capacitor selection will drain a battery in short order.
  • I doubt anyone will get 10 years from a coin cell even with the most careful design practices.

My report covers hardware and software considerations for building these sorts of systems.


Raw data tends to be pretty raw. The real world is a noisy place; everything from distant lightning strikes to Johnson noise corrupts what we'd hope would be pristine data read from an ADC.

It's common to average some number of ADC readings to reduce the noise and better approximate the actual data we're trying to measure.

Averaged garbage is still garbage. See the discussion in Muse 373 about how it's important to ensure the data fits the physics. Don't apply a "does this data make any sense at all" check and you'll get the 737 Max disaster.

Crazy temperature reading

Software fail: No common sense mitigation applied.

Averaging reduces Gaussian noise by the square root of the number of samples. Ten improves the signal-to-noise ration by about a factor of three; 100 samples yields an order of magnitude improvement.

There are two problems with averaging: It's slow (though read on for a better algorithm), and the average presumes all of the data, no matter how old, should contribute equally to the result. Sometimes a more complex filter will better fit the data.

When I went to college we learned about convolutions in, of all places, circuit theory. The definition is:

" />

The math baffled me and I couldn't see any reason we'd use such a thing. And yet in my day job we were massaging data using the Savitzky-Golay algorithm to fit data to a polynomial (see Muse 254), which, though I didn't realize it at the time, is nothing more than a convolution.

It turns out convolutions are simplicity itself disguised in academic rigor meant to torment university Juniors. Suppose you're reading an ADC and averaging the most recent N samples. That's the same as multiplying the entire universe of input data by zero, except for the most recent N, which are multiplied by 1 and averaged. In math-speak, the input data is convolved with a function whose value is zero except for the most recent N samples which are 1. The data slides through the N-points of the average.

But does that make sense? It supposes that old data is just as representative as the latest readings. This is fine in many circumstances, say for data that changes slowly. It will slow the systems response and smear rapid changes.

There are many alternatives, the most common, perhaps, is the weighted moving average. Multiply the most recent reading by, say, 5, the next most recent by 4, and so on. This is the Finite Impulse Response filter (FIR), which smooths data using only recent readings. Another option is the Infinite Impulse Response (IIR) filter, which includes all of the data collected to date. An example might use a weight of 5 for the most recent, 2.5 for the next, 1.25 after that, and so on.

With simple averages one sums the data and divides by the number of samples. The weighted moving average's divisor is the sum of the various weights.

So, how does one implement moving and weighted moving averages in code? The simplest approach is to save a circular buffer of the most-recent N data points and do a multiply-accumulate of the entire buffer each time a result is needed. That's computationally expensive.

A better approach for the simple moving average is to compute:

AN+1 = (XN+1 + N * AN)/(N+1)

Where AN+1 is the new moving average, N is the number of samples, XN+1 is the latest reading, and AN is the previous average. This is a simple, fast, calculation and one does not need to store old data. (For a proof of this non-intuitive result see this).

(The division can sometimes be eliminated, if absolute units aren't needed).

Convolutions get more interesting in picture processing where pixels in two dimensions, say a 3x3 square, are manipulated to sharpen or defocus images.

A $50 Vector Network Analyzer

Pete Friedrichs offered a review of an inexpensive vector network analyzer. Pete is the fellow who wrote a couple of fascinating books about building electronic projects... from scratch. Like making your own capacitors and vacuum tubes. His review:

The NanoVNA is a two-port vector network analyzer, 50 kHz to 900 MHz. This thing will compute/plot return loss, swr, complex impedance, filter sweeps, smith charts, and more.

The whole instrument is about the size of an Altoids tin. The display is a color touchscreen, additional input is made through a tactile rocker switch. The device is rechargeable and contains its own lithium battery which is recharged through a USB C port.

Speaking of USB, the device presents itself to a PC as a comm port. There are a set of instructions that allow the unit to be controlled and queried through plaintext messages. People have already written PC software to control and interrogate the device so as to produce large graphs and perform additional functions like time domain reflectometry. At least one person has written a Labview front end for it.

There is a fairly robust user group of this device as evidenced by youtube videos and a user's group on google.

Here's the most incredible part…the instrument costs about 50 BUCKS!

No, it's not the same as a $XX,XXX dollar Agilent VNA, but it does much of what a "real" VNA will do and it's three orders of magnitude cheaper. In my experience the instrument you have is ALWAYS better than the instrument you'd like but can't afford.

In fact, more than one user, with access to a "real" VNA for comparison has reported reasonable agreement between their high-end instrument and the NanoVNA.

I purchased one recently and have had endless fun with it. I run Linux, and on my boxes, it shows up as ttyACM0. I have successfully controlled the device through minicomm, and I'm now contemplating writing some software of my own.


I submit the above link as an example, not necessarily an endorsement. Actually,  the NanoVNA is available all over the Internet from multiple sources. There are at least three versions. The "original" version is claimed to be the best. Details of what comprises the "original" vs "clone" versions—along with modifications that can improve the clones—can be found in the google user group.

This is one of those gadgets that is too cheap NOT to buy. In my opinion it's been worth the price of admission just for the hands-on learning.

On Comments

Many, many people commented on commenting function headers from the last Muse. A common theme: Comments are bad because they are often incorrect. As the code changes, the comments don't. We've all been cursed with code like that. My take is that as professionals we are obligated to maintain our comments. To abdicate that responsibility is like the woodworker who never sharpens his tools. Sure, the blade will still cut, leaving a rough surface behind, while filling the shop with smoke. But with enough post-cut sanding he can make something that is somewhat presentable, usually, slowly, with plenty of four-letter words peppering the work.

Here are some highlights from readers. Martin Buchanan wrote:

Regarding not putting comments on obvious function headers, I call this the Get Foo issue. Object-oriented code is often full of getter and setter method functions that do nothing but unadorned reads and writes (and hopefully some constraint checking for the setters). So you then have a GetFoo method with a header that says "Get Foo." My suggested approach is that if you are using documentation generators that generate documentation from comments in the code, then go ahead and comment everything, even what is obvious. If you are not using such documentation generators, put some rules in your coding standards regarding obvious constructs that do not need comments, then omit those comments.

Tom Mazowiesky commented:

I think it depends on the function and how complicated it is.  When you have a simple function that uses descriptive naming of the function and the variables, it doesn't require a lot of commenting.  When the function is fairly complex, involves lots of calculations, if's, switch's, etc, then you need something up front to describe what it's supposed to do.  

I know some guys who can tell you what the code they wrote two years ago does in detail, but my brain doesn't work like that.  I need the comments I wrote to figure out what's going on a couple of weeks later.  Once I finish it, I basically toss the info to work on another problem.  I also think we need to document code thoroughly enough to allow someone else to maintain it.  I've worked on projects/products that were produced many years and needed ongoing support.  Comments certainly help with that.  

I think comments are especially required when you do something that is odd.  I remember working on a DSP project where I needed to move a variable from one kind of memory directly to another memory and the CPU did not have an instruction for that.  So I wrote some self modifying code to do it.  Had to be done because the application had to run extremely fast, and to use an intermediate register or memory wouldn't get the job done in time.  Even though it was all done in one line of assembler code, I wrote a couple of paragraphs describing the situation and what was done in detail.

One of the young engineers we hired would name all variables and routines with AAAAA, AAAAB, AAAAC, etc.  This came from one of her professors in school.  She kept a separate set of notes explaining what each function did and what each variable did (was an assembler project).  

She did a great job writing the code (was for a test system) as it worked properly and she got it done in a reasonable time frame.  It was however difficult to follow since you needed the separate doc to figure out what was going on.  Fortunately we didn't have to do much work on it after she did the initial code.  We showed her how we did code, using names that meant something, and she had no problem converting over.  But I always wondered where this idea came from and it's rationale.  

John Carter and others had somewhat dissenting opinions:

A similar conversation went by on Lobste.rs


My take was and is that Rusty Is Right....


On a scale of 10, needing to read the comments to get it right, is a low 3 out of 10 in terms of Good Design.

Sometimes you need to go there, but always strive for something better. I think  the claim that we need more comments is dubious.
I have bashed my head literally daily against thousands of comments that are…

  • Content free - The comment says less than the function signature, less precisely, less concisely and more ambiguously.
  • Inaccurate - I usually set my editor colouring to minimize the visual impact of comments. I debug code, not comments.

Sure, dealing with a well commented code base is a pleasure. But it is a complete rarity. They are jewels in a vast pile of commenting mush.
Things to do before asking for more comments...

  • Using meaningful variable names and function names. (Yup, suffers from the same problem as comments, in that what the function does may change. But it tends to slap you in the face every time you look at it.)
  • Using short simple functions whose intent and functionality is simple and obvious. (Use Humble Function / Humble Class pattern at the higher levels)
  • Use pre-condition, post-condition and invariant (especially class invariant) asserts. Think of them as "executable comments" that are proven to still be valid.
  • Use unit testing where each test case sets up and checks and tears down a single behaviour. Think of it as executable documentation. Make sure your "executable documentation" is readable and easy to understand.
  • If you comment, remember names are about "What", comments are about "Why".

I have come to define "Good Design" as "How little I need to read and understand to make a beneficial change".
In the realm of multi-megaline code bases this is a fairly compelling criteria, as I have met no one who can keep such a monster in their head.

Caron Williams suggested:

Something I stumbled over by accident recently was quite educational. I remembered writing a set of 'libraries' some 15 years ago which were applicable to a current project. I never made formal libraries - just copied the c and h files into the project. Both platforms were Microchip so I expected hardware related issues to be predictable, minor, and easily addressed. And so it turned out to be, even though it was written for a PIC18 and the new Target was a PIC24.

The problem was that even though I had written all the functions myself, and adequately commented them - or so I thought - I could not understand how they worked from a simple read-through. It was clear what they did - just not how.

That led to the realization that as well as explaining what functions are for, or do, the comments should also address how the function does it.

I know if you take this too far you'll end up with:
I = l + 1;     // add one to l

  which is silly for many reasons. But if you can't clearly see how a piece of code works on a quick read through the comments are insufficient or inappropriate. Difficult, I know, when it's fresh in your mind, but try re-reading your old code and see if it passes muster.

If you don't understand how it works then adapting to a new platform—or indeed, editing for any reason—is fraught with danger.

A pet peeve, which is related, is code (often vendor supplied) which has several levels of nested #defines. This is done in an attempt to achieve platform independence. These nested defines are simple synonyms, macros, and combinations of the two. It makes sense to the manufacturer to have a single code base to maintain across chip families, but unraveling it to find out how it works is tedious. And, although the preprocessor rules are well established, there's a niggling feeling that a complex set of nested defines might take a new preprocessor version into uncharted—or untested—territory.

Many people wrote to chastise me about my yearning for an IDE that supports in-code WYSIWYG editing, like we take for granted with word processors. The common theme was that we already have these, in terms of editors that support keyword highlighting, comments in different colors, etc. While those are great features, my dream goes further: I want to embed pictures in source code. People complain that comments can get too long. Well, as the saying goes, a picture is worth a thousand words. How about tables (not crummy ASCII-formatted versions that get scrambled with changing indentation settings)? Maybe even complex objects like spreadsheets.

It's easy to see how an external tool could strip this stuff and present the compiler with nothing more than the printable ASCII text, but that would be a kludge. I'd sure like to see the next C standard support XML source files, or something similar, so the compilers themselves can digest files rich with visual information. Clarity is our goal, and sometimes pure text is a forest whose trees hide what could be presented in obvious ways.

Tony Whitley had similar thoughts::

I think I mentioned this before but I didn't limit myself to rich text, I'd like to see diagrams and images in source code comments ('scope traces is just one example, even scans of hand-drawn diagrams would help) and hyperlinks to documentation, particularly external documents (for example a 3GPP spec. that defines what a particular bit of code is implementing).  Also, there must be cases where embedded media would be a valuable part of understanding what a program is doing.

It's simple enough, write your code in some kind of smart HTML-based editor then pre-process it to filter the <code>... </code> sections then strip the HTML out of that before sending on to the compiler.  Another benefit is that we could get rid of // /* # comment markers from source code and comments could be in boxes that float around the page (anchored in some way to the code they refer to) rather than sit artificially spaced out (necessitating a load of fiddling to restore the layout if the code or comment is edited).

Different way to comment code

Not a very imaginative example!


Let me know if you’re hiring embedded engineers. No recruiters please, and I reserve the right to edit ads to fit the format and intent of this newsletter. Please keep it to 100 words. There is no charge for a job ad.

Joke For The Week

These jokes are archived here.

From Scott Anonymous, a real story:

Here's another story about code reviews, in light of the comments about 'figuring out what the code is supposed to do', you need comments or external documentation as others have said. In one company I was at, an engineer submitted code for review just an hour or two before the review was scheduled. There was no documentation, and only five comments in the 1 KLOC file. One comment was the less-than-helpful:

//#include <math.h>      // fix this

As one of my co-workers said "I can't review this. This first function has nothing says what it's supposed to do. I don't know if it's a good function for displaying letters on the screen or a bad square root function." Yes, the function name was something like "put_str", but that could have been wrong too! No documentation, remember?

About The Embedded Muse

The Embedded Muse is Jack Ganssle's newsletter. Send complaints, comments, and contributions to me at jack@ganssle.com.

The Embedded Muse is supported by The Ganssle Group, whose mission is to help embedded folks get better products to market faster.