I've been working for the last three days on interfacing the DSP board with a PIC.

My first approach was a serial interface, transferring one bit every left/right sample period, or about 96 kb/s. I thought that would be adequate data speed, but after thinking about it more, I realized that it is marginal. By the time you allow for synchronization and other overhead, the net throughput is about 2500 32-bit words per second, or 5000 16-bit words/sec. If we use a 320 x 240 pixel display, and if we desire 10 to 12 refreshes per second, there isn't much margin left in a 96 kb/s serial link.

The DSP development board has a memory-mapped 4-bit I/O interface, so yesterday and today have been devoted to starting a 4-bit interface. At the moment, I can read 4-bit data from the DSP into an 18F4620 PIC, and output it over an RS232 serial link. The RS232 serial link won't run fast enough to keep up with the DSP-PIC transfer, so I'm only sampling the data. Over the next week or so, I plan to get the DSP-PIC data transfer operating in both directions. This will provide a platform for commanding the DSP via computer keyboard, through the PIC. Still lots of work to do, but it's making some sense now.
 

Clock 1 is the strobe signal, with valid data on the trailing edge (positive to ground transition). The four data lines are OUT-0 through OUT-3.

The line +FourBus is what Intronix calls an "interpreter" or data analyzer. I've set it to convert the logic levels of the four out lines to a hex digit. Other options are binary or decimal.

This is much easier than debugging data flow using even a digital storage oscilloscope.

I've finished the manuscript on ferrite rod antennas and posted a printed copy and CD-ROM to the QEX editor I've been dealing with. Assuming it passes editorial muster, you may read it around the end of the year.

I recently purchased an HP3456A digital multimeter with GPIB interface. This instrument has 4-wire ohm capability and I received the Kelvin clips to make use of that feature a couple days ago. The Kelvin clips came with 0.1% resistors of 1K, 10K, 100K and 1M nominal value to allow one to test the ohmmeter accuracy.

Before I present the results, there may be some who are not familiar with 4-wire ohm measurements or know what a Kelvin clip is.

Let's start with the familiar 2-wire ohmmeter. A simplified version is shown below. The microammeter is calibrated in ohms, but in fact is simply measures current through the circuit, with the scale being a form of analog computer solution of Ohm's law, marked to show the approximate value of Runknown, the resistor being tested.

In fact, a two-wire ohmmeter has many frailties. For example, the resistance between the test lead and the part being measured is unlikely to be the same as when you adjusted the zero setting. And, the battery voltage may not be the same either.

Lord Kelvin, when faced with the problem of making accurate resistance measurements in the 19th century, invented a wonderfully clever solution to these problems. Lord Kelvin viewed the conventional ohmmeter circuit as really two individual circuits; one to supply a known current to the resistor being  tested and the second to measure the voltage across  the resistor being tested. Ohm's law thus allows the unknown resistor to be calculated via R = E/I.

This scarcely seems like an improvement, you may think. We've gone from a simple circuit with one meter to one requiring both a voltmeter and an ammeter. Surely that reduces the accuracy in half you ask. 

However, look at the advantages of this arrangement. First, the resistance of the test lead and the internal parts of the ohmmeter are effectively removed from the computation. We don't care what these are (shown as Rinstrument in the diagram), as our computation only requires knowledge of the current through the series circuit of Rinstrument and Runknown. The voltmeter's impedance is assumed to be sufficiently high that its burden current does not cause measurement error. And, since the voltmeter draws negligible current, the resistance in its leads and clip lead connections will not materially affect the resulting measurement accuracy.

A Kelvin clip provides a way of easily making the 4-wire connection to the unknown resistance. The photo below should show you how a Kelvin clip is constructed.

Now, if you are measuring a 1 MΩ resistor, a 2-wire ohmmeter will be more than adequate, as the imperfections of internal resistance variation and clip lead changes are much smaller than the resistor under test. But, suppose you've been asked to measure a 0.05 Ω current shunt resistor. How can you accurately measure it, without worrying about test lead resistance variation? Or, suppose you have a printed circuit board with a shorted track, with a total resistance of 0.1 Ω. How do you locate the actual short point?

A 4-wire ohmmeter with Kelvin clips make both jobs easy. To demonstrate this, I've measured a 0.5 ohm, 1% 5 watt current sense resistor with the HP 3456A in 4-wire ohms mode with Kelvin clips.

As a matter of fact, it's possible to see a major difference in measured resistance by simply moving the Kelvin clips to the end of the resistor.

Here's the result of the 0.1% resistance measurement, by the way.
 

Of course, with these relatively high value resistors, 2-wire and 4-wire resistance readings will be close, as the lead errors are generally in the 0.1 to 0.2 ohm range.

I'm trying to keep several balls in the air at once for the next few weeks. In addition to the DSP work, I've agreed to write several articles for possible publication, so I'll be busy. 

The first article covers a few week's work I did in 2000 building and testing ferrite loop antennas. At the time, I drafted 90% of an article on my results and set it aside until a few days ago. Whether it has improved by aging, like a fine wine, remains to be seen.

This afternoon, I added a software low pass filter to the DSP IF section I've been working with. As the block diagram below illustrates, the result isn't far from a complete IF chain. The major missing part is an AGC system. And, of course, all the frequencies, bandwidths and the like are hard coded. Plus, there's no digital interface to the outside world. All are projects for the next weeks.

And, before I add too many more features, I will restructure the code. It's in the "throw it together and see if it works" form now, and the longer I wait to clean it up, the more difficult the job will be. Moreover, the code's structure mimics an analog IF and detector strip. It does not, for example, take advantage of I & Q sampling and corresponding image rejection, folding spectrum at 0 Hz and the like.

But, before moving to these exotic things, my grounding in the nuts and bolts of the TMS320VC33 coding must improve.
 

I'll jump from 2007 to 1947 in this posting. Whilst looking for something else in a storage cabinet, I ran across a 1947 vintage Cornell-Dubilier CDA-5 decade capacitance box, from 100 μμF to 0.01μF, in 100 μμF steps. None of this pF or nF stuff in 1947. Nope, there were "micky mikes" and "micky-micky mikes" and don't you forget it.

I don't remember when I acquired the decade capacitance box, but I think I've had it 25 or 30 years and recall that there was a problem on some ranges when I picked it up at a swap and shop. Desirous of a break from punching assembler code in the TMS320VC33 DSP, I measured the decade capacitance box on all ranges with a digital capacitance meter. Based on the readings, I deduced that both sides of the range switches had one bad capacitor each, a bad 0.002 μF on the large step side and a bad 400 μμF capacitor on the small step side.

I replaced both mica capacitors with modern parts, a 2000 pF polystyrene and a hand selected 390 pF dipped silver mica (measuredat 398 pF) and I hope it will be good for another 60 years.

I put the defective parts in a plastic bag inside the case for any future owners that might be interested in historic restoration.

I'm seeing signs of progress with my DSP work. This morning, I made a simple IF/product detector work in software. The arrangement is:

• 15 KHz IF bandpass filter with the following parameters:

133 tap, Parks-McClellan FIR filter
Center frequency: 15 KHz
Bandwidth: 200 Hz
Transition bandwidth: 2.2 KHz (shape factor 1:11)
Stopband attenuation: 80 dB
Passband ripple: 0.5 dB

• Numerical oscillator BFO at 14 KHz, computed on-the-fly by vector rotation
• Product detector via multiplication of BFO and post-filter signal values

The product detector/mixer is simply a line of code that multiplies the numerical value of the filter output times the numerical value of the BFO signal. Of course, it also needs a low pass filter (to be added later) to knock out the signal and BFO feed through, but I can't hear 14 KHz so my ears provide the necessary filtering. (Product detectors are really mixers, although for historical reasons they are called product detectors.)

As I tune a signal generator through the 15 KHz input frequency range, there's a nice, clean and stable beat note in headphones connected to the Development Board's DAC audio output jack.

Pretty nifty! All those things I learned many years ago about the product of two sine waves being their sum and difference are true. (Not that I had any doubt, mind you, but it's still nice to see direct verification.)

Here's the code required to implement the product detector/mixer. (This is not efficient code, but rather a kludge to demonstrate the concept works.)

The first two lines load the current value of the BFO and post-filtered input signal into the TMS320VC33's R1 and R0 registers, as these two numbers have been computed by other code and their values placed in the V1 and OUT_L floating point variables. The third line is the actual mixer, multiplying the values held in registers R0 and R1, saving the product in R0. The fourth line scales the product held in R0 to a range the University DSK's D/A converter can handle, via a scale factor previously defined in floating point variable scale.

The 15 KHz filter response looks good. I can't verify the 80 dB design stopband as my HP3465A voltmeter's noise floor is about 500 uV, thus limiting me to about 65 dB dynamic measurement range in this configuration.

Thus, the measured 0.5 dB bandwidth of the filter is very close to the 200 Hz design target, but the 3 dB bandwidth is about 600 Hz.

One more point and I'll wrap up this discussion. Spectrum analyzers normally use a Gaussian filter response, with a 3 dB : 60 dB bandwidth ratio of 1:15 or so. Hence the above design, although not a true Gaussian filter, is more than acceptable as a 600 Hz spectrum analyzer filter shape. If anything, the skirts are a bit too sharp. (As to why spectrum analyzers use Gaussian filters, it's to accurately pass amplitude information in the minimum time. I've covered the topic in more detail in my Z90 Filter Design article in May/June 2007 QEX.)
 

Yesterday, I made some output power measurements on my K2/100 transceiver to verify wattmeter performance data. To generate the two-tone input signal (700 Hz and 1900 Hz) I used a Telulex digital synthesized function generator in two-tone mode. I've used this generator in the past and other than broadband noise being a bit higher than I might like, it's been a good performer at a bargain price.

Yesterday, however, I noticed that the LP100 wattmeter data was bouncing around, and looking at the transceiver's output signal with an oscilloscope, it was apparent why. The two-tone signal peaks were not the same level from peak to peak. Larry's LP100 wattmeter captures 100 or more readings per second, necessary in order to read power from a single 60 wpm dot or to read voice peaks on SSB. It was apparent that the LP100 was accurately reflecting wildly differing peak power values from one envelope cycle to another.

The oscilloscope capture below shows how a two-tone envelop should look, or at least almost how it should look. There's still a slight difference amongst the peaks, amounting to about 0.1 graticule division. At the worst, yesterday's peaks differed nearly one full graticule division.

The problem, of course, was that 60 Hz was mixed with the 700 and 1900 Hz tones, in effect modulating the two test tones. I had recently rearranged the AC power arrangement in this corner of my basement shop, and apparently there was enough difference in ground potential amongst the SG-100, the TDS-340 digital oscilloscope and the Astron 20M DC power supply connected to the K2, to cause a ground loop and thus inject undesired 60 Hz signal into the K2's microphone input.
 

However, connecting the transformer was a bit of a Rube Goldberg effort with clip leads, adapters and wires over the test bench, so this morning I decided to make a more permanent audio isolation transformer arrangement. the "iso-dapter", as illustrated below.

There's nothing complicated about the design. It's a 600 ohm audio  transformer with one set of windings connected in parallel to four different connectors and the second windings connected to four more connectors. Either side may be used as an input and the other as an output. The particular connectors used represent the ones I normally use around the shop and shack. (I have a cable made up with an 8-pin microphone connector for Kenwood and Elecraft on one end and a mono 3.5 mm phone plug on the other end.)

The transformer is actually a split 600 ohm (four 150 ohm windings) transformer, but I've jumpered the windings  together to provide 1:1 impedance. Note also that I've used an inexpensive Radio Shack plastic box so the input and output ground references float with respect to each other, as otherwise the purpose of the transformer isolation would be defeated.

The common mode isolation could be improved with a screened transformer, i.e., one with an electrostatic shield between the primary and secondary windings, with the shield being connected to the ground on either the input or output. This would reduce stray capacitive coupling across the transformer windings and further isolate the input and output connections. I measure 104 pF capacitance between the input and output ground pins. At 60 Hz, this corresponds to 25 MΩ capacitive reactance.

Added features could include a toggle switch to go between normal, open and short on the output side. Or, more connectors could be added, perhaps with two transformers so that left and right stereo channels may be separately maintained.

Here's a frequency response plot of the Iso-dapter:

However, a high quality 600 audio transformer, such as a Western Electric 119C coil, will provide a much flatter response.

By the way, Elecraft's model 2T two-tone generator avoids the ground loop problem, if it is powered from a 9 V battery, as intended. Using AC powered general purpose test equipment carries the risk of ground loop hum. If found, an audio isolation transformer should solve the problem.

I mentioned that the LP100 will accurately read power and SWR with a single dot at the K2's maximum internal keyer speed. Here's an oscilloscope capture of the dot. The power reading was accurately displayed, despite the dot's duration being only about 15 milliseconds.
 

I've uploaded a revised Z100 Operating Manual this morning, version 1.3, dated 14 June 2007. The changes are relatively small and if you have already assembled your Z100, the new manual is not worth downloading and printing. Kits shipped today include version 1.3.

The Z100 manual is available by clicking here, or via the Z100 page or the Documents page.

I received a proof copy this morning of a QST Technical Correspondence letter I've written, to be published in August 2007 edition. The note is titled "MEASURING MOTIONAL PARAMETERS OF A QUARTZ CRYSTAL" and corrects an error in Wes Hayward's excellent book Experimental Methods in RF Design, co-authored with Rick Campbell, KK7B, and Bob Larkin, W7UPA, relating to G3UUR's oscillator frequency shift method of measuring crystal motional parameters. In essence, the equation provided in EMRD omits  the crystal's holder capacitance and hence causes the calculated motional capacitance to be 15-20% low. The corrected equation is:

This equation, along with sample measurements using it, can be found at my Crystal Motional Parameter notes, available at the Documents page, or by clicking the link above.

If you own EMRD and have not already done so, please check Wes's corrections page. The ARRL did a particularly poor job of editing and producing the book and Wes has provided an extensive errata at http://users.easystreet.com/w7zoi/em0.htm .

Yesterday evening, I  threw together a quick Liberty Basic program to control my Telulex SG-100 function generator (now a Berkeley Nucleonics Corporation model 625) over the RS232 serial port and my HP 3465A digital voltmeter over the GPIB port (via a Prologix version 4.2 USB-to-GPIB interface unit). The program steps the SG-100 (sinewave output) from a specified starting frequency to a specified end frequency in specified frequency increments. Each frequency step is saved to a data file, along with the 3465A's reading. The data can then be easily imported to a spread sheet, or a graphing program for further analysis. (I use OriginLab's Corp. Origin v. 7.5 scientific and engineering plot program.).  It's a lot easier to run this program than to write down the response values and keyboard them into the computer.

I've verified that the SG-100 and HP 3465A track over the range 100 Hz - 50 KHz within a small fraction of a dB.

Sample graphs are below. You may compare the first plot to the one I made with hand entering values yesterday morning.  The automated data has many more data points and avoids keyboard entry errors.

If you wish to experiment with digital filter designs, http://www.dsptutor.freeuk.com/ has a very nicely implemented interactive design applet. You specify the center frequency, transition band and other parameters and the applet computes the filter coefficients and plots the filter response. If you are at all interested in digital filtering, I recommend visiting the site and experimenting with the filter applets.

After several days of running in circles, I have a simple low pass filter running on the TMS320VC33.

I've mentioned the spotty documentation provided by TI. That extends to TI's Burr-Brown division's PCM3003 data sheet. The PCM3003 is a stereo CODEC, i.e., it has a left and right channel A/D converter and a left and right channel D/A converter, all with 20-bit resolution. The University SDK board has a PCM3003 CODEC installed as the primary I/O device.

Most data sheets discuss the expected data output format, i.e., the coded output corresponding to a specific voltage level input. Microchip's data sheets, for example, are very good in this regard. The PCM3003 data sheet, however, only discusses how the serial data is outputted. After a couple of days of strange results that I attributed to problems with my filter algorithm, I started at the beginning, and looked at the PCM3003's data output for specific voltage level inputs. Since the board capacitive couples input and output, this required a low frequency square wave input signal, and capturing the data output with a digital storage oscilloscope.

The rest of the story is that the A/D uses 2's complement output for voltages below Vcc/2, but I had assumed that since the data sheet was silent, the output format was similar to the internal A/D in a PIC Microcontroller, which is linear, from 0 ... 1023 for a 10-bit A/D.

The PCM3003's comparison circuitry is biased to Vcc/2. Input voltages above this are output as 0...524288 (2¹⁹)  Voltages below Vcc/2 are output in 2's complement form (20-bit length), so that 1 LSB below Vcc/2 is -1, but -1 is encoded as 1048575, or in hex 0xFFFFF.  The TMS320VC33 is a 32-bit device, so it reads 0xFFFFF not as -1, but rather as 1048575. The result is very strange looking output waveforms. After realizing the problem, it was simple to fix the code to convert the PCM3003's format to a form the DSP is happy with.

My test code is a very simple low pass filter that uses a 15 length moving average (boxcar) filter. The response is a sin(x)/x form, easily seen in the frequency domain data I measured this morning.
 

Channel 1's frequency response is about 22 KHz, as it goes through filtering in both the A/D and D/A steps. Hence it's rise and fall time is also degraded from the function generator's sub-microsecond specification.

The DSP's sample frequency is 48.8 KHz, or 20.5 μs. The full 15 moving average period is thus 15 x 20.5 = 307 μs. The digital oscilloscope (Tektronix TDS430) computes rise and fall time with the traditional 90% - 10% points, so we expect it to show a rise/fall of 0.8 x 307 μs, or 245 μs, close enough to the TDS430's 246 μs measurement.

I've returned from an interesting two day visit with Mike, W4XN, and his wife in Charlottesville, VA, about 100 miles from Clifton. Mike's basement lab looks like an HP showroom, and among other things, we had a chance to measure the phase noise on several different types of HP signal generators, using his 8566A spectrum analyzer and an 8 MHz crystal notch filter I designed and made. I've added his data for the following generators to my Oscillator Noise Measurements page.

I consider myself fortunate to have made Mike's acquaintance recently.

As planned, the last two days were also devoted to working with the TMS320VC33 DSK. My plan was to complete working code to read the two A/D converters (stereo, one left channel and one right channel, in a common IC), buffer the results in a circular array and output the buffered values to the D/A converter.  I finished that code this afternoon. Here is a sample result.
 

The 3.22 ms delay is a combination of two elements. First, the buffer delay. This example uses a 127 element buffer array, with a sample rate of 48.8 ks/s. The buffer delay is  thus 127 / 48800 = 2.602 ms.

There is thus an extra delay of about 620 μs. This delay corresponds to zero wait, and is comprised of digital filter delay in both the D/A and A/D.
 

In working with the TMS320VC33 University DSK DSP kit, I've tried to set small end-of-day goals. Yesterday's was to have a functioning circular buffer. Today's was to have the digital-to-analog (DAC) portion of the PCM3003 CODEC output a waveform that I had programmed in as a data table.

This is how direct digital synthesis, or DDS, works, albeit at audio frequencies. The programmed data table is for a triangle waveform and here's the output waveform:
 

Tomorrow's goal will be to work more with an output waveform table.

Here is the output waveform with the dual update fixed. It's 1 KHz, as planned.
 

Although I'm becoming more accustomed to TI's assembler and DSK3DW "integrated development environment," my original opinion that these tools offer woefully underwhelming stability, features and user friendliness, is reinforced daily.

Another day in my basement lair working on DSP programming. After considerable headscratching, I've managed to make a functioning circular buffer work in TMS320VC33 assembler.

This might not seem like much of an accomplishment for 8 hours of work, but it was, due to a combination of my failing to carefully read TI's documentation and also assuming that some sample code was in fact correct.

A circular buffer, by the way, is an area in RAM memory that is addressed in a wrap-around or "circular" fashion. You might have 32 bytes of memory in which to store inbound data, for example,. What do you do when the 33rd byte arrives? Depending on the purpose of your code, you might stop inbound data, or take some other action. But, it's a common event in digital signal processing to take the 33rd byte (the newest data) and save it where the oldest data was before. Hence, the newest byte overwrites the oldest byte and the next byte that arrives overwites the next oldest byte, etc.  Thus, the data addressing wraps around the end of the reserved memory. One can also think of this as a buffer array with the head and tail connected, thus forming a circle. Hence the common terminology of a "circular array."

Since DSPs deal with circular arrays so often, they have special instructions to make their use efficient. For example, the TMS320VC33's hardware can keep track of where the next data element is to be stored, providing you have properly set up the size and location of the memory reserved for the circular array. There's where my problem came in. It turns out that you have to reserve more memory than the size of the array. I won't go into details, but if the array is 31 elements long, you must reserve 32 spaces. (The TMS320VC33 uses 32-bit memory, so the elements are not 8-bit byte width.) And, the array memory space must be reserved on an address boundary that is "even" in binary based on the reserved size.

An example is that the 31 element array requires 32 memory spaces and the reserved memory must start on an address such that it is an even multiple of 32. A starting address of, e.g., 48 won't work. Rather, the starting address must be 32, 64, 96, etc.

I misread TI's documentation to say that the reserved space must be ≥ the array size, when in fact it is a strict >. Hence my 32 element circular array with 32 reserved data elements starting on a 32 multiple boundary failed. To make it work, either the length must be changed to 31 or less or if the length is to be 32, then 64 memory locations must be reserved and the starting address must be at a multiple of 64.

After sorting that out—and misreading the same text at least a dozen times—my circular buffer is working as it should.

Circular buffers are very useful in implementing digital filters, as a common filter methodology multiplies past samples by a series of coefficients to produce low pass, high pass, bandpass, etc. responses. This is all very elegantly explained in Steven Smith's book The Scientist and Engineer's Guide to Digital Signal Processing, which he has made available for free download at http://www.dspguide.com/pdfbook.htm. If you have not already done so, you should read it.

Here's a photo of TI's University DSK printed circuit board.
 

As usual, I've archived the May 2007 updates to a separate page, viewable by clicking here or via the links at the top of this page.

Yesterday was devoted to TMS320VC33 DSP work. I had purchased two used Dell SX260 computer at the end of 2006, with the thought of using one in my radio room to update the elderly 266 MHz laptop machine I use there now. The SX260's are essentially a laptop board in a small table top case, but with a generous complement of physical connections, a real hardware parallel port, a real hardware serial port, 6 USB ports, serial keyboard, serial mouse and a 100 MHz Ethernet connection, audio I/O and a VGA video board supporting up to 1600 x 1200 resolution. I decided to use this second Dell computer for my DSP programming and connected it yesterday. I purchased a 20" ViewSonic monitor, VP2030b, supporting 1600 x 1200 resolution, to use with it.  It's much easier to use TI's software programming software with a large screen, as I can open multiple windows and still read the text.  TI's software is still buggy, and klunky, but it's now legible.

By yesterday evening, I had a simple loopback program running on the TMS320VC33. It converts analog audio input to a digital format using the on-board Burr-Brown/TI PCM3003 CODEC. The digital data is then echoed back to the PCM3003's D/A section and output as an analog signal. (CODEC is the acronym for coder-decoder and describes a module that performs both encoding [A/D] and decoding [D/A].) TI refers to these functions as AIC, standing for "Analog Interface Circuit."

Using a company-standard term, AIC, instead of the generic CODEC, is another example of TI going out of its way to complicate things. Understanding TI's documentation first requires you to master its terminology.