Nerdy Texts of Analog and Embedded Systems Wizardry

No pretty pictures today. The hex for WTPA is up to >10kB, which is quite possibly the most code I’ve ever crammed onto an 8-bit part.

The feature list is full and then some, and the ISRs are beginning to suffer speed wise. When the sampler is cranking out multitimbral stuff at high rates you can hear the audio bog down, which I guess is ideologically in line with the “crusty first” aesthetic, but it bugs the anal programmer in me.

In the waning hours before Bent this means going through all the pristine, general ISRs and doing what I can to save cycles, avoid divides, replace conditional branches with jump tables, etc etc. This usually means the code gets ugly, at least for me.

Tomorrow (this morning, I guess) the proper board Rev comes in — that means I’ve got to make a new WTPA and take pretty pictures, and write the manual. Lots of stuff to do before Saturday…

Oh, and did I mention the limited edition video-synthesis daughter board I’m making for WTPA? Come see it in New York at Secret Project Robot on the 25th.

Oh Snap! There’s the goofy silkscreen I promised last time. But don’t worry, it gets better :-)

So I’ve done my best with the above Still-Life-With-Nerd collage to try and convey, in a meaningful visual way, all the cool stuff that’s happened with WTPA since last week. One really perceptually-difficult-to-convey-yet-great thing is that I’ve put my proper rent-paying clients on hold for the time being to make room for this project full time. It’s scary, but it feels really good! And as rebellion goes I suppose it beats buying a Camaro.

Outside of that, the BOARDS ARE HERE! Sort of.

The bad news is, I found a copper bug (my fault) pretty much right away. With as many hardware revisions as I’ve done this was really embarrassing. It led to a natural dilemma — it was a small bug that could be easily fixed with an exacto knife and some jumper wires, so should I you pass that fix off to you, the kit builder, or should I pony up for an ENTIRE NEW SET of boards so that they’d be perfect?

I debated, then reordered them. This also allowed me to fix a half a dozen little cosmetic and noise-floor / routing things that I wasn’t happy with anyway. I wasn’t nuts about dropping another grand on boards, but I’d be less happy knowing that the final kits were going out half-assed. You can’t pay too much for pride :-)

Silver Lining: This now allows me to sell bare boards, which a ton of people have asked for. Originally I hadn’t planned to do this, but now I can sell the flawed boards (for cheap) with an errata sheet to those among you who are exceptional broke asses and not mess up the math and parts counts for the complete kits. It’s a small consolation, but it’s better than making them drink coasters.
Cosmetically, the boards are flawless! I got a not-quite-perfect board from PCBCart before this (for another job) and let them hear about it, and they bent over backwards to make this one dead-on.

Taiwan Alpha also came through — the pots (all 2000 of them) showed up from China early and perfect.

Then there was Dorkbot which was a blast. Above you can see me in front of a page of C code (looks like the ISR) apparently casting some kind of spell.

People geeked out, I gave away some free PCBs, I laughed, I cried, it was better than Cats.

Finally, social business done, it was back to the Fortress of Solitude with some espresso and a dream. Two great things have come out of this so far:

One:

WTPA now has a banked sample system! This means that instead of holding a loop, WTPA can hold an arbitrary number of loops (theoretically, anyway) and can do anything it is able to do to any or all of those loops independently and all at the same time.
Practically speaking, I’ve pinned that “arbitrary” number at 2 :-). There was a lot that went into this! A huge portion of the code had to be re-written — basically all the audio and memory handling parts. The audio system now grabs data from the ADC and passes it to however many “virtual samplers” for recording, and then sums together whatever output they have before putting it back onto the DAC. Better still, the different banks can use different clock sources! Meaning that you can be triggering events with midi and controlling pitch arbitrarily one one bank, and twiddling knobs and generally being a caveman on the other, with a totally independent samples.
By far the hardest part though was having to wrap my head around implementing a memory manager (ever think you’d have to write malloc? Me neither) which was quite challenging. To be honest, I copped out a little, and it is because of this that WTPA uses only two banks for now.

All this cool functionality also has a totally new (and much more intuitive, I think) menu system. Finally, I’m in the midst of getting rid of “modality” in WTPA so that there’s no such thing as “MIDI mode” and “Manual Mode”. Anything you can do with WTPA you can always just do, without mucking around in a menu.

Two:

I crunched the money numbers! As of today, I know exactly how much all this crap has cost me, and therefore how much these magic beans will cost YOU, dear reader. I’m not spoiling the surprise, but I will say that although I have definitely screwed up some estimates in my time, I was pleasantly surprised with how this one came out. And you should be too.

I’ll keep you all posted on the nerdy revelations better for this last week. Don’t expect total coherence or an Infinite Jest command of grammar, but I got you. Oh, and I think I’m going to LA to talk about WTPA at the end of the month, so if you’re on the West Coast come say hi! More details on that to come.

Xoxoxo,

TMB

Online in 2k9, my ducks.

So sometimes, contrary to common blogline belief there are times when I’m quite busy even when I’m not posting. This New Year / Holiday period was one of them. There are some big advances on the WTPA front:

First, all circuits and modes are totally functional. The resampling problem I was having had two roots. First (and lesser of the two) the input impedance of the AVR’s A/D converter is actually a lot lower than I ever realized. In the datasheet Atmel recommends that you don’t drive the ADC with anything less than a 10k impedance but I actually found that an input resistance as small as 470 Ohms still
produced a noticeably smaller waveform at the ADC input pin than driving the pin directly from the output of an opamp. This was easy (though annoying) to fix — it requires another op-amp section and a redesign of the summing amp into the resampling ADC input. The only unfortunate thing about this is that it will drive the cost of the final device up for you, dear customer, and this kit is already, as they say in France, tres cher.
Still, I made a decision a long time ago that when I design things for me (as opposed to the toy world) I will always err on the side of badassery rather than cost, because I can. And because everybody wants to be a baller, right?

The other (more significant) problem was a Read-Modify-Write issue in the audio ISR. OMGWTFBBQ, was it hard to find! When is a NOP not just a NOP? When it lets you sync an output port you’ve just written to with its input latch. Weird, weird bugs from that one, to be sure. And fixing bugs in my code is, like, totally free!

In other important news I bothered to scope out the speed of the audio IRQs and compare them to the old WTPA. As I had hoped, there was mondo improvement on that front:

• WTPA 0.9, Time in IRQ:
• Recording / Playback: 22-24uS
• WTPA 0.95, Time in IRQ:
• Recording: 5.8uS
• Playback: 7.5uS
• Resampling/Overdubbing: 9.0uS

This means that we could now sample at 44kHz (CD quality — remember CDs?) without breaking a sweat if we wanted to! Of course, I’m not sure how much good that would do us since the ADC on the AVR is already “not-so-accurate” at the paltry 24kHz I’ve got it set to now. But hey, the headroom is there if you all want to get your overclock on.
Practically this means that the processor now actually has some time to get some work done other than audio, which is good.

So the final hardware related issue in this animal is the jitter generator. The way it was originally designed the jitter generator used a PRBS white noise source to mess up the clock signal, and it worked! Sort of.
It became apparent that when the “jitter” control was at maximum, the knee frequency of the filter on the white noise became the dominant interrupt frequency. In English: The fastest noise waveform actually determined the clock frequency in a totally deterministic sounding way.
Since this frequency was different than the clock frequency, you would hear a “jump” in sample rates when adjusting the jitter. This sucked. Right now I’m trying to get the jitter to mess up clocks of all frequencies equally well without inserting a dominant frequency of its own. Meaning a jittered-up 600Hz clock will still have a frequency of 600Hz, it will simply have a totally random duty cycle. (You know, like peak-to-peak displacement, like Wikipedia says under “Jitter”).
After fiddling around with another analog solution (pictured above) and then with some circuit sketches using AND gates and JK flipflops I realized that the best way to do this (everybody all together now) was IN SOFTWARE. I can’t decide if it’s sad or not that it always seems to be that way. Anyhoo, once I’m done with that it’s time for new boards.
Oh, and my birthday is coming up in a little less than a week. Send me naked pictures, or LOLCATs or something. Xoxoxox.

Yesterday was a good day for sampler technology!

Three really important things got done; two are improvements on the old sampler:

First: the serial link between the MCUs is now totally robust and fast. As I suspected the real problem with the serial before was the internal RC oscillator on the helper MCU. For now it’s been replaced with a 7.3728MHz crystal, but Digikey just rolled up with some 18.432MHz crystals this morning which I’ll be using instead, I think (these particular and seemingly-weird frequencies are exact integer multiples of common UART frequencies, and allow 0 baud rate error).

There were a couple other little forehead-smacking moments where I got bit by casting errors and other general programming bugs but they weren’t really bad once the UART worked. Like for instance — if you declare a variable “theByte” as an unsigned char, set it equal to ~’p’, and then later check for equality with ~’p’, it comes back false. Know why? Know how to fix it? I do now. :-)

Second: It makes audio! Since I already put a picture of a noble sawtooth wave up a million years ago and since the old sampler definitely did this just fine, this is less exciting. BUT it does mean a couple things. The new DAC works, and the analog sections (mostly) work and all the volume pots are the right way around. The VCO works, and the jitter generator — Oh boy does it work! All these need tweaking but they’re all rocking solid.

Third: I spent a long long time fiddling with the PCB and tracking down noise demons. This, I am happy to say, was a big success and a _huge_ improvement on the old sampler!

Pictured above are some example waveforms. On the left is a picture of the output of the old sampler with the preamp gain, through level, DAC level, and master volume all the way up. You probably can’t read the V/div knob on the scope, but this waveform is unmistakably clock noise which has capacitively coupled into the audio path and has been amplified such that it is ~220mV at the output. Yikes! This is really bad. Those of you who’ve heard the original WTPA know that this whine is one of its not-so-good-akshully characteristics.

The photo on the right is a noise picture of the new sampler after a little screwing around by me. Under full-gain, worst-case noise conditions, the new sampler has approximately 4mV of clock feedthrough. This is 34.8dB of improvement!

There are a couple noise sources in the WTPA family. One source of LF noise is the LEDs pulling current spikes from the supply when they turn on. This is at its worst when running from a not-so-low impedance supply (like a 9v battery or worse an STK500 through the ISP header) and did manage to find its way into the audio but it wasn’t terrible. The best fix for it was to improve filtering at the op-amp reference terminals (the VDD/2 level at the non-inverting terminal of a summing amp, say) since the noise that got there got amplified by the circuit’s gain. This helped for sure.

By far the worst noise source is clock bleedthrough, although the new design also shows bleedthrough from the white-noise generator. Improving this was a little trickier. A couple things made a big difference. Varying the gain on the preamp as opposed to keeping the gain fixed (and high) and attenuating its input level was one. Re-doing the design with better ground and signal routing mattered A LOT, although I still had to rip up a trace with an exacto — one poorly-thought out connection added nearly 100mV to the output noise here. Filtering the analog references better helped deal with this noise a little, too.

Finally, I think a ground plane (or two) are in order for the final hardware revision. I also think associating the VCO and Jitter Generator with the digital supply lines (or at least isolating them from the quiet amplifier analog supply) would be a really good idea. I might even use a choke input to the quiet analog supply. We’ll see. Either way, this statistic is already A LOT better than it was, and I’m excited.

Next: Sampling / RAM / throughput rate test, and more fun analog tweaking.

The lone and level sands stretch far away from this nasty Nas, I’m afraid:

Not that you can tell from gazing on this benighted silicon Ozymandias, but no pretty waveforms will happen tonight. It’s not a total loss; there’s some badassery that came through and TON of code, but since I said I’d post today here’s the breakdown:

The body of the code is done for both MCUs, at least at first blush, and the ISRs are now based on the parallel interfaces. The LEDs look crazy hot. Messy little resistors and whatnot are dutifully appearing on the board. The square waves are almost square.

The bad news is that all the old timers who I didn’t pay attention to were right — UART communication on internal RC oscillators is dodgy at best. The real (and unforseen) challenge of the last 16 hours has been massaging the serial link between the two MCUs enough that it works. It limps along now but it sucks. My studied conclusion is that the second MCU really needs its own crystal, and it might as well be at a UART-friendly frequency.
I probably have some laying around somewhere, and will dig them (and exacto blades) out this weekend.

Other fun bugs and riders of the nitpick train include:

The AVR toolchain doesn’t seem to be very excited about the 48p/88p/168p MCUs, and although you can fool it, doing so raises the “bad idea” error flag. Further, the STK500 seems to only want to program the old-ass m168s I dug out to rectify this problem ONE TIME, before getting totally weird. This is a good one: These parts are socketed in the STK, mind, and the fuse bits remain totally unchanged through this whole process. You put code onto the part just fine and boom, it becomes unprogrammable. I did this THREE times. Furthermore, reading the Vtarg on the STK500 shows 6+ Volts, and the device times out when you try to communicate with it. Say what?
I started shopping for a new STK, until further fiddling revealed you can a.) physically pull the RESET pin to ground on the socketed target device using an alligator clip (or stream of foul language) which allows it program just fine OR you can socket and connect to another WORKING part, and move the Vtarg down to 3v, whereupon the old “broken” part works just fine again. Gremlin central.

Writing the code was, thankfully, pretty straightforward and I discovered a kickass feature in avr-libc: ISR_ALIASOF() — this reduced the accumulated ISR code to half the size — thus doubling its beauty.
Finally, a point of consideration: Naming an electonic device on your website “Where’s The Party At” makes for no small amount of schoolgirlish giggling when looking at your keyword referrals in Google Analytics. In keeping, I have decided to name my next piece “Hot Hipster Tang For Altbros”.