Nerdy Texts of Analog and Embedded Systems Wizardry

Analog is sexy, we all agree, right? Embedded systems on the other hand, are full of lots of unglamorous problems. Filesystems, say. Inherently un-sexy.

But I think component variation is maybe the best, most un-sexy problem there ever was. The unsexy cherry on the diet sundae. Like, you HAVE to solve it if you are making lots of something or that thing as a population will suck, even though the one on your bench always ruled.

All the pots in WTPA2 are these custom Taiwan Alpha jobbies. There are two values, 10kA and 100kA. The VCO uses one of the 10kAs as a coarse control, and it sets the voltage into a current sink which in turn sets the frequency. I’d been messing with the op amps in this circuit to try and get some performance improvements and “all of a sudden” one of the DUTs didn’t work correctly. At first I figured it was the opamp change, but after a lot of measurement and desoldering and component testing, it turned out one of the 10k pots was really 11.4k. This was a greater than 10% variation!

I’d built a margin in for error, but this was above it, and the current sink was getting too high of a voltage. I tested a dozen pots or so from the bin, and all of them were much less off. Still, since one was off, probably another one could be as well. It could even have been a result of the soldering process. I actually bothered to do a DC simulation at this point (using qucs) and fiddled with the component values until they were all as off as I could imagine them possibly being, and then resized the scaling resistor that sets the upper range of the VCO. It was a really crappy annoying unsatisfying solution, because it means that MOST of the units will be operating at a slower maximum clock than they need to. But that one in twelve or one in 100 will work correctly. Serves me right for getting the cheap pots, but there you go. Margin. Component variation.

Least Sexy Problem Evar.

TB

So, after getting back to client work for a minute, I decided to try and nail the clock pulse shaping circuit problem with a more viable solution than throwing in an extra $5 op amp.
The problem with the original pulse shaper circuit was simply that it was designed with a function generator and not a 20 cent opamp in a RC oscillator. The idea was sound (I think) but the values were not.

The real problem is that the square-to-pulse converter has to shape two different clocks — it’s always driving the same IRQ pin, but it can be hooked up to WTPA2′s 4046 based VCO, or the LM358 based on a user switch. The 4046 is HC logic, and has really square edges. The LM358′s edges are not square, and their slew rate seems frequency dependent also. So, you could optimize components for one or the other, but not both. I did some bench tests to figure out what I needed to change to get this right.

Check it. Here’s the rising edge of the output from the VCO:

And the corresponding output from the pulse shaping network:

Since we aren’t changing the VCO, this is what we’re gonna call “normal”. The top trace shows a risetime of about 0.1uS (scope is 0.1uS/div, 2v/div) which is quite fast (50V/uS in opamp terms). The ringing here probably has to do with the long ground connection on my probe, and it doesn’t hurt anything except my pride. The bottom trace (the output from the pulse shaper) shows a clean low pulse which is about 6uS long total (2uS/div)
Now, here’s the LM358:

And the corresponding output from the pulse shaping network:

Waaay different! This is the LM358 at its best incidentally — tested at low oscillator frequencies. At higher clock frequencies it slews even more slowly.
The top trace is 10uS/div, and shows a rise time of about 25uS (it’s 60uS with the clock cranked up to 25kHz). Annoyingly, it has that characteristic LM358 style crossover mess. AND it only gets up to about 4v. The rise time is really what matters though, and it is orders of magnitude slower than the 74hc4046. The bottom trace shows the output of the pulse shaper, trying but not quite making it. That dip never makes zero volts and might last 0.25uS. This doesn’t consistently trigger our interrupt-on-change IRQ.

So, the question was what to do. I tested a TLV2462 opamp (my goto op amp for embedded stuff, made by TI, a tank) and it performed equivalently to the 4046, and the pulses worked great. It’s slew rate was rated at 1.6V/uS, which is about 5 times faster than the LM358′s 0.3 V/uS. So it was faster, but not by orders of magnitude. If I could find an opamp which cost about the same as the LM358 and had a better slew rate, that seemed appealing rather than trying to hack up a circuit on 300 already-fabbed boards. The question was how fast we needed to go.

I settled on three opamps for the test: The Microchip MCP6002 (0.6V/us), the Microchip MCP602 (2.3V/uS) and the Texas Instruments TLC272 (5.3V/uS). A few days later I had them all from Digikey. I tested the MCP6002 first, since it was the cheapest. (0.27 at quantity, as opposed to the LM358′s 0.20) Surprise surprise! It worked great on the first try.
Although I didn’t measure the rise time, it looked clean on a scope. The ouptut from the pulse shaper was 6-7uS which is as good as (and more importantly in line with) the logic chip in the VCO. This was also consistent with the TLV2462.

In conclusion, the cheapest and easiest way to solve this problem is (I think) to eat 0.27 per kit and throw in another opamp. Further, the results are interesting because they show that above a certain rise time, performance remains the same. My guess is that there’s a knee point in that filter, and as long as the dominant frequency of the edge is above it, we’re good to go. In this case, a clean 0.6V/uS output was enough to trigger the shaper reliably.

Now that the results are consistent and I’m in tweak mode anyway, I’ll probably try and get those pulse times down by half or so, just in case the ISR gets faster.

Analog is fun, yo.
xoxox
TB

All right, so now that Cory’s biz is done and the art world is safe again, I can get back to God’s Work, by which I mean making samples that sound like farts. That’s right, WTPA2!
WTPA2 has been promised now for like a million years. I’m shooting for actually having it ready by the end of June for Bent Festival.

In that spirit, I dug out my old prototype. There’s a lot wrong with it. I found most of the hardware bugs way back when, and I added another input for a separate pitch control to the second sample bank. The idea was to use the spare op-amp to make an RC oscillator and use it to clock the second sample bank and use the main oscillator to clock the first bank. Clock sources could then be switched or interchanged in hard or software.

Problem is, the only uncommitted pins left that can trigger an IRQ are interrupt-on-change pins. That means that I can’t only trigger on a rising or falling edge — the ISR will trigger on BOTH. That means a 10kHz square wave will trigger 20k interrupts a second. I could make the clock half as fast I guess, but that seems like it will confuse people. For the time being I dealt with it by checking the state of the pin in the ISR, but that’s lame too. It means we vector away from mainline code twice as often as we need to. So I came up with this:

It’s a pulse shaper. It takes a clock input, and regardless of duty cycle, spits out a low-going pulse on every rising clock edge. The diode and cap here are responsible for separating out the edges, and the transistor squares them up again (more or less). Hooked up to the function generator (Agilent 33120A, 50ohm out) I can get a nice 0.5uS low going pulse really consistently! I can use this to trigger an interrupt, and the pulse will ALWAYS rise again during the ISR (the fastest ISR in WTPA2 is like 9uS). Then at the end of the ISR I can clear the interrupt flag. Viola, rising edge interrupts with a couple cents worth of hardware! I’ve rolled this and some other hardware changes into the next proto revision and will be ordering it soon.

So the first iteration of WTPA2 has some dumbass mistakes — bus problems during flashing hardware (missing pullups), some switch latch goofiness, and turns out all those RC filters in the encoder datasheet really ARE a good idea. However, once all the traces got cut and the little merce-resistors got in place, the thing works great. The VCO is spot on. More importantly, so is the FLASH MEMORY! The SST flash kinda sucks in that it’s not fancy and requires you to manage erasing-before-writing and demands paying attention to buffering and stuff, but you can totally turn off WTPA2 and turn it back on and keep playing with that perfect burp sound you made.

Also, equally excitingly, the ISR has gotten A LOT FASTER — this proto recorded and played back just fine at 45kHz. A lot of this had to do with taking some very good suggestions from Olivier over at Mutable Instruments (of the Shruthi-1 fame) who is a great programmer and shamed me out of much laziness in my code.
As if that wasn’t enough, I finally licked the lion’s share of the noise sources that plagued WTPA1. I’d always been really careful about analog signal routing, but I’d been pretty cavalier about ignoring the hell out of some of the “Analog Noise Cancelling Techniques” in the Atmega datasheet. Turns out I traced most of the noise back to on-chip activity which had to do with reading and writing to the SRAM (toggling GPIOs) while the ADC conversion was taking place. I moved some of the accesses around and that NAILED it. Like, totally duh!

Also, re: the picture — I bought one of those Saleae Logic analyzers the moment they started supporting Linux because it seemed like a cool toy. But it’s actually really useful and works great! In addition to actually seeing what’s going on over the bus (as attached here) it’s REALLY handy for timing ISRs. Like, you toggle a pin high when you vector, and then low again when you exit. I always did this with a scope, but the logic analyzer is great because it records a lot of them and you can analyze variation, see what happens between several different calls, use many channels, etc etc.

Anyhoo, WTPA2 had an exciting week. It may take a break for a minute as I have a really busy winter coming up, but still, good time.
xo
TB

So this weekend marked the last screwing around (breadboarded, at least) with the VCO design for WTPA2. I knew that I could get the frequency range I needed out of a 4046 based VCO with a current sink in place of a frequency setting resistor but I was a little worried about temperature drift and very low frequency performance. Just for shits, I gave the circuit above a shot. It replaces the current sink with a standard two-transistor current mirror. The two devices above are crappily thermally coupled with a piece of heat shrink.

What I found with this is that temperature performance wasn’t too much better. It was still pretty easy to get this VCO to go crazy with the heat gun, although it was possible to get it to go crazy BI-DIRECTIONALLY based on the direction you were slathering on the sweat. Cute, but not useful.
One thing that WAS cool is that the current source compliance was great! The control voltage could get a lot higher than it could with the one-transistor sink before saturation, I assume because of the lack of emitter resistors in the circuit. Still, that was like a consolation prize.

So I caved and built this:

Here’s the schematic for it:

It’s what all the docs I saw suggested originally (a closed loop current sink where the op-amp compensated the temp drift of the transistor) and it totally works and does what it’s supposed to. The only annoying thing is that at VERY low currents the device is non-linear, but what are you gonna do. At a hundred Hz sampling rate all samples sound like farts anyway.

The CV has to be kept below about 2.2v to keep the transistor from saturating (the VCO goes nuts when this happens) but you can fix this with a divider on the input. The range with these components is about 0-20kHz with a CV from 0 to 2.1v or so.

Now I gotta figure out what to do with that other half of the 358. I sort of don’t want to pass it audio, because it’s a recipe for coupling noise into the circuit. We’ll see….

This week I’m working on the board layout for the first WTPA2 revision. I expect the next post here will be about that.
xo
TB